I h
.335.“
:3...’.: A..

i ll; 3 2

. ..m.
.3233;

 

 

 

..c

I
.7

?,

 

IIIllfli'l‘lljflfllllfllLiflllllljflllll w

This is to certify that the
dissertation entitled
IPCA: An Intelligent Control Architecture
Based on the Generic Task Approach to
Knowledge-Based Systems

presented by
David Bruce Decker

has been accepted towards fulfillment
of the requirements for

Ph. D . degree in Computer Science

Mwljor professor

DMe November 13, 1995

 

MSU i: an Aflirmatiw Action/Equal Opportunity Institution 0- 12771

 

 

 

 

 

 

LIBRARY
Michigan State
University

 

 

 

PLACE IN RETURN BOX to romovo this chookom from your record.
TO AVOID FINES mum on or bdoro dot. duo.

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

:
moan” -* " '1 "r,

 

IPCA: An Intelligent Control Architecture Based on the

Generic Task Approach to Knowledge-Based Systems

By

David Bruce Decker

A DISSERTATION

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY

Department of Computer Science

1995

   

 

ABSTRACT

IPCA: An Intelligent Control Architecture Based on the
Generic Task Approach to Knowledge-Based Systems

By

David Bruce Decker

A knowledge-based architecture for the control of complex dynamic processes is pre-
sented. The control architecture is rooted in the Generic Task (GT) approach to know]-
edge-based systems. The architecture extends the GT—based DSPL framework for routine
design and planning by providing a complete architecture supporting knowledge-based
control system synthesis and execution.

The Intelligent Process Control Architecture (IPCA) incorporates three primary
components: a Blueprint Generator; a Plan Assembler; and a real—time reactive Plan
Execution Monitor. The Blueprint Generator produces a specification for a complex pro-
cess control plan which is designed for a specific execution environment. The blueprint
includes information about overall control policies and processing goals but not their spe-
cific algorithmic implementations. The Plan Assembler implements each blueprint by
combining pre-compiled plan and control method fragments. The selected fragments are

assembled to form an executable process control plan. Process control plans are intended

 

for use with the Plan Execution Monitor for execution within the environment to be con-
trolled. The Plan Execution Monitor incrementally determines appropriate fragments
from the process control plan based on sensory input gathered from real-time process
monitoring, reactively selecting an appropriate path through the process control plan
structure. As a result of fragment execution, processing goals are modified and process
parameters (control outputs) are changed, thereby facilitating control of the process.

The usefulness of the approach is demonstrated by applying IPCA to the control of a
microwave-based composite material fabricator. Experimental results are presented for
selected polymeric composite structures which were cured using IPCA. These experi-
ments demonstrate that IPCA effectively controls the fabricator, resulting in high-quality,
high-speed composites processing.

IPCA was also tested in conjunction with a simulated control system interface in a sec-
ond series of experiments. These experiments were used to demonstrate intelligent
response to various hypothetical conditions which are difficult or impossible to create in
the laboratory. The experiments also demonstrated the breadth of knowledge-based

domain coverage included in the IPCA prototype.

 

ACKNOWLEDGMENTS

I would like to acknowledge the contributions of several faculty from the Michigan State
University Department of Computer Science: Dr. William F. Punch; Dr. Jon Sticklen; and
Dr. George Stockman. Their guidance, support and encouragement were essential in
allowing the completion of this work. I would also like to thank Dr. Martin C. Hawley of
the MSU Department of Chemical Engineering for providing the domain which has
enabled a real-world test environment for this control architecture. I would also like to
thank Dr. James K. McDowell for his many contributions and advice.

Financial support was provided by the Composite Materials and Structutres Center at
Michigan State University and by the MSU Department of Computer Science. A signifi-
cant portion of the computer hardware used during this research was supplied by AppleTM
Computer Corporation.

Several graduate students were helpful through discussions and providing tools
enabling the implementation:, including Kurt Patzer, Ahmed Kamel and Robert Hawkins.
A special thanks goes to Valerie Adegbite and Yunchang Qiu for constructing the experi-
mental microwave cavity, implementing the conventional control system and assisting
with many of the fabrication experiments.

I would like to thank my family for their love, support and understanding in this long

and sometimes trying endevor: Mr. Nestor J. Decker Jr.; Mrs. Marilyn A. Decker; Mr.

iv

 

Nestor (Tory) J. Decker III; and Mrs. Gertrude Shier.
Finally, I would like to acknowedge the contributions of two of my closest friends:
John Bauman and Keith Pyne. Their friendship provided much of the support I needed

while completing this work. Thanks also go to Keith for proofreading the final copy.

 

TABLE OF CONTENTS

LIST OF TABLES ................................................. x
LIST OF FIGURES ............................................... xi
CHAPTER I ...................................................... 1
1.1 Motivation and Problem Introduction ............................. 1
1.2 Approach .................................................. 3
1.3 Outline .................................................... 5
CHAPTER II ..................................................... 9
2.1 Related Work and Literature Review ............................. 9
2.1.1 The Contributions of Control Systems Theory ................ 11

2.1.2 The Contributions of Knowledge-based Planning .............. 12

2.1.3 The Contributions of Task-Specific Architectures .............. 15

2.1.4 The Contributions of Knowledge-based Design ............... 17

2.2 Research Approach and Architectural Introduction ................. 18
2.3 Expected Results and Evaluation Criteria ........................ 21
CHAPTER III .................................................... 24
3.1 Domain Introduction ......................................... 24
3.2 Related Domain Research .................................... 25
3.2.1 Composite Materials Research at Michigan State University ..... 25

3.2.2 Electromagnetic Processing of Polymer Composites ........... 27

3.2.3 Research in Advanced Computing Techniques ............... 29

3.3 Conventional Composite Processing ............................ 31

vi

 

3.4 Microwave-based Processing ................................. 32

3.5 The Experimental Single-mode Resonant Cavity .................. 38
3.6 Extemal Controller Overview .................................. 40
CHAPTER IV ................................................... 43
4.1 Theory of Operation ......................................... 43
4.2 Controller State-Space Complexity Analysis ...................... 45
4.3 Knowledge Level Analysis ................................... 47
4.4 IPCA Task Descriptions ...................................... 49
4.5 GT Approach .............................................. 51
4.5.1 DSPL and the Blueprint Generator ......................... 52

4.5.2 Sponsor-Selector Mechanisms ............................ 57
CHAPTER V .................................................... 61
5.1 Architectural Overview ....................................... 61
5.2 The Plan Generator ......................................... 63
5.2.1 The Blueprint Generator ................................. 64

5.2.2 The Plan Assembler and Fragment Librarian ................. 65

5.2.3 Fragments ............................................ 67
5.2.3.1 Fragment Library Content ............................ 69

5.2.3.2 Adaptive Fragments ................................ 71

5.2.4 Process Control Plan Format ............................. 72

5.2.5 IPCA Goals ........................................... 74

5.3 The Plan Execution Monitor ................................... 76
5.3.1 The Execution Unit ..................................... 76

5.3.2 The Observation Unit ................................... 77

5.3.3 The Control Interface Unit ................................ 79

5.4 Extemal Controllers ......................................... 79
5.5 The Simulated Control Unit ................................... 81
CHAPTER VI ................................................... 83
6.1 Experimental Scope and Objectives ............................ 83

vii

 

6.1.1 C-Series Experiments ................................... 83

6.1.2 F-Series Experiments ................................... 84
6.2 Experimental Preliminaries: F-Series ........................... 85
6.2.1 Sample Geometry and Composition ........................ 85
6.2.2 Temperature Probe Placement ............................ 85
6.2.3 Electromagnetic Modes .................................. 86
6.2.4 Heating Profiles: F-Series ................................ 89
6.2.4.1 Series F1 ......................................... 89
6.2.4.2 Series F2 ......................................... 92
6.2.4.3 Series F3 ......................................... 95
6.2.4.4 Series F4 ......................................... 97

6.2.5 Mode Maps ........................................... 99
6.2.5.1 Series F1 ........................................ 102
6.2.5.2 Series F2 ........................................ 104
6.2.5.3 Series F3 ........................................ 106
6.2.5.4 Series F4 ........................................ 106

6.2.6 Mechanical Limitations ................................. 107
6.3 Evaluation Criteria ......................................... 111
CHAPTER VII .................................................. 114
7.1 Experimental Results and Discussion .......................... 114
7.2 C-Series .............................................. 114
7.2.1 Series C1 ......................................... 114

7.2.2 Series C2 ......................................... 116

7.2.3 Series 03 ......................................... 118

7.3 F-Series .............................................. 119
7.3.1 Series F1 ......................................... 119

7.3.2 Series F2 ......................................... 124

7.3.3 Series F3 ......................................... 129

7.3.4 Series F4 ......................................... 133

7.4 Analysis ................................................. 136

viii

 

CHAPTER VIII ................................................. 142

8.1 Conclusions .............................................. 142
8.2 Future Work .............................................. 144
APPENDIX A
Glossary ...................................................... 145
APPENDIX B
IPCA User’s Manual ............................................ 152
APPENDIX C
IPCA Software Utilities .......................................... 173
APPENDIX D
External Controller and IPCA Interface Protocol ..................... 178
APPENDIX E
Domain Background ............................................ 188
APPENDIX F
Extending the Capabilities of IPCA: A Case Study ................... 200
BIBLIOGRAPHY ................................................ 205

 

 

LIST OF TABLES
Table 5.1 - Typical Blueprint Format .................................. 65
Table 5.2 - Control Method Fragments in the IPCA Fragment Library ........ 70
Table 5.3 - Plan Fragments in the IPCA Fragment Library ................. 71
Table 6.1 - Modes Used for Fabrication Series F1 - F4 ................... 88
Table 6.2 - Qualitative Part States for Series F1-F3 ..................... 101
Table 6.3 - Qualitative Part States for Series F4 ........................ 102
Table 7.1 - Summary of Experimental Series C1 ....................... 115
Table 7.2 - Summary of Experimental Series 02 ....................... 117
Table 7.3 - Summary of Experimental Series CS ....................... 118
Table 7.4 - Summary of Experimental Series F1 ....................... 120
Table 7.5 - Summary of Experimental Series F2 ....................... 125
Table 7.6 - Summary of Experimental Series F3 ....................... 130
Table 7.7 - Summary of Experimental Series F4 ....................... 133
Table F.1 - Common IPCA Extensions ............................... 201

 

 

 

LIST OF FIGURES

Figure 2.1 - Overall IPCA System Architecture .......................... 19
Figure 3.1 - Relationships Between CMSC, MMPI and the NSF Center ...... 26
Figure 3.2 - Typical Autoclave Curing Profile ........................... 32
Figure 3.3 - Composite Layup for Microwave Processing .................. 36
Figure 3.4 - Teflon Mold for Microwave Processing ...................... 37
Figure 3.5 - Experimental Single-Mode Microwave Cavity ................. 39
Figure 4.1 - Structure of a Continuous-time Control System ............... 43
Figure 4.2 - Structure of a Digital Control System ........................ 44
Figure 4.3 - IPCA Task Structure .................................... 50
Figure 4.4 - DSPL Design Agents ................................... 53
Figure 4.5 - Structure of IPCA Design Specialists, Plans and Tasks ......... 56
Figure 4.6 - Sponsor-Selector Structure ............................... 58
Figure 4.7 - Match Group Table Structure ............................. 59
Figure 5.1 - System Architecture ..................................... 62
Figure 5.2 - Typical Process Control Plan Structure ...................... 67
Figure 5.3 - Example Fragment from MDEmergencyPolicyA ............... 73
Figure 5.4 - Communication with an External Controller ................... 81
Figure 6.1 - Fiber-Optic Temperature Probe Placement ................... 86
Figure 6.2 - Heating Profile: F1 (Mode 1) .............................. 90
Figure 6.3 - Heating Profile: F1 (Mode 2) .............................. 90
Figure 6.4 - Heating Profile: F1 (Mode 3) .............................. 91
Figure 6.5 - Heating Profile: F1 (Mode 4) .............................. 91
Figure 6.6 - Heating Profile: F1 (Mode 5) .............................. 92
Figure 6.7 - Heating Profile: F2 (Mode 1) .............................. 93

 

Figure 6.8 - Heating Profile: F2 (Mode 2) .............................. 93

Figure 6.9 - Heating Profile: F2 (Mode 3) .............................. 94
Figure 6.10 - Heating Profile: F2 (Mode 4) ............................. 94
Figure 6.11 - Heating Profile: F3 (Mode 1) ............................. 95
Figure 6.12 - Heating Profile: F3 (Mode 2) ............................. 96
Figure 6.13 - Heating Profile: F3 (Mode 3) ............................. 96
Figure 6.14 - Heating Profile: F3 (Mode 4) ............................. 97
Figure 6.15 - Heating Profile: F4 (Mode 1) ............................. 98
Figure 6.16- Heating Profile: F4 (Mode 2) .............................. 98
Figure 6.17 - Heating Profile: F4 (Mode 3) ............................. 99
Figure 6.18 - Mode Map Surface: F1 (initial) ........................... 103
Figure 6.19 - Mode Map Surface: F1 (final) ........................... 104
Figure 6.20 - Mode Map Surface: F2 (initial) ........................... 105
Figure 6.21 - Mode Map Surface: F2 (final) ........................... 105
Figure 6.22 - Mode Map Surface: F3 ................................ 106
Figure 6.23 - Mode Map Surface: F4 ................................ 107
Figure 6.24 - Average Mode Switch Times: F1 ......................... 109
Figure 6.25 - Average Mode Switch Times: F2 ......................... 110
Figure 6.26 - Average Mode Switch Times: F3 ......................... 110
Figure 7.1 - Thermal Profile (F1.2) .................................. 121
Figure 7.2 - Thermal Profile (F15) .................................. 121
Figure 7.3 - Surface Scans (F1.2— F1.8) .............................. 123
Figure 7.4 - Thermal Profile (F2.1) .................................. 126
Figure 7.5 - Thermal Profile (F2.3) .................................. 127
Figure 7.6 - Thermal Profile (F24) .................................. 128
Figure 7.7 - Surface Scans (F2.1 - F28) ............................. 129
Figure 7.8 - Thermal Profile (F32) .................................. 131
Figure 7.9 - Surface Scans (F3.1 - F3.3) ............................. 132
Figure 7.10 - Thermal Profile (F4.1) ................................. 135
Figure 7.11 - Thermal Profile (F4.4) ................................. 135
Figure 7.12 - Surface Scans (F4.1 - F4.4) ............................ 136

 

Figure 7.13 - Hold Time, Gradient and Heating Rate Relationships ......... 138

Figure 3.1 - Starting IPCA using the VisuaIWorks Finder ................. 153
Figure 8.2 - Launcher Interface .................................... 154
Figure 3.3 - Blueprint Generator Interface ............................ 154
Figure 8.4 - Blueprint Generator Problem-Solver (PS) Menu ............. 155
Figure 8.5 - Blueprint Generator Browse Menu ........................ 156
Figure 8.6 - Blueprint Generator Cases Menu ......................... 156
Figure 8.7 - Specifying Design Type ................................. 157
Figure 3.8 - Specifying Design Cases ............................... 158
Figure 8.9 - Reviewing the Selected Fabrication Model .................. 159
Figure 8.10 - Blueprint Generator Rough Design Complete ............... 159
Figure 8.11 - Blueprint Generator Design Complete .................... 160
Figure 8.12 - Reviewing Generated Blueprints ........................ 161
Figure 8.13 - Plan Assembler Interface .............................. 162
Figure 8.14 - Fragment Librarian Interface ............................ 163
Figure 8.15 - Completed Plan Assembly ............................. 164
Figure 3.16 - Execution Monitor Interface ............................. 165
Figure 3.17 - Execution Unit Interface ............................... 166
Figure 8.18 - Sponsor-Selector Match Group (SSMG) Browser ............ 167
Figure 8.19 - Observation Unit Interface .............................. 168
Figure 8.20 - Parameter Source Browser ............................. 169
Figure 8.21 - Control Interface Unit Implementation ..................... 170
Figure 8.22 - Simulated Control Unit (CU) Interface ..................... 171
Figure 3.23 - Part Graphics Interface ................................ 172
Figure 0.1 - Log Translator Interface ................................ 174
Figure 0.2 - Log Translator File Specification .......................... 175
Figure 0.3 - Mode Map Generator Interface ........................... 177
Figure 0.1 - External Controller Interface ............................. 179
Figure 0.2 - Microwave Equipment and Configuration .................. 180
Figure 0.3 - Experimental Microwave Power Delivery Wiring .............. 181
Figure E.1 - Partial Taxonomy of Reinforcement Types .................. 190

xiii

 

Figure E.2 - Partial Taxonomy of Matrix Types ......................... 191
Figure E.3 - Chemical Crosslinking Behavior of Therrnoset Composite ...... 193
Figure E.4 - Chemical Structure of High-Performance Thermoplastic ....... 195

xiv

 

CHAPTER I

1.1 Motivation and Problem Introduction

The problem addressed by this research involves producing control algorithms which are
suited for controlling specific physical systems. Controlling real-world phenomena often
demands the successful integration of knowledge-based planning techniques and numeri-
cal control methods. There are many approaches from which to choose. Which techniques
and approaches might be appropriate for the control of a particular real-world system or
process? How can selected planning and control algorithms be generated and combined to
form an appropriate control system for a given process? How can reasonable processing
goals be determined for a particular process? Can knowledge about the process and
domain help to guide the selection of appropriate goals and control strategies?

Although portions of the overall task have been extensively studied in the fields of
control systems theory, planning and knowledge-based design, most approaches concen-
trate on only one or two of these areas. Within the domain of modern control systems the-
ory, for example, physical systems are often assumed to be fixed and well-defined so that
they may be modeled mathematically [Gopal 93]. The use of such simplifying assump-
tions can lead directly to elegant and efficient control algorithms based on well—estab-

lished development methodologies. But these systems are almost exclusively hand-crafted

 

2

and they cannot plan a complex set of actions to accomplish high-level goals. Moreover,
classical control systems design methodologies may fail altogether when one attempts to
model poorly-understood processes [Boyd 91].

Planning systems, as opposed to traditional control models, are much more effective at
the task of reformulating and pursuing goals and in reacting to unforeseen complications
[Winston 92]. Planners are capable of creating and revising complex plans designed to
accomplish a large set of potentially interacting or conflicting goals. Many planning sys-
tems can apply their plans in real-time [Musliner 95]. For example, planning systems used
in robotics and autonomous vehicle research cope well with extremely complex environ-
ments [Firby 88]. But planning systems also have drawbacks. Classical (static) planners
do not work well in dynamic environments since the state of the world can change inde-
pendently of applying selected actions from a plan [Fikes 71]. As a result, much of the
planning research is not useful in “real-world” environments. Reactive planning
approaches attempt to address this problem, but delays encountered in formulating and
fixing plans during their execution may slow such systems to an unacceptable degree.
Universal planners create plans which are size-exponential in the number of sensed
parameters because they directly map all possible world-states into appropriate actions
and goal re-forrnulations [Schoppers 89].

Another limitation of many planning systems is that well-understood subproblems of
an overall planning and control task cannot be effectively addressed by many control algo-
rithms since the planner must reason about operator interactions and potential goal con-
flicts. Most importantly, reactive planners must be provided with goals (which may need

to be revised as replanning occurs) but goals are usually treated as either static or some-

 

 

how provided from outside the planner.

The forte of knowledge-based design systems is their ability to synthesize artifacts
which can meet specifications and satisfy design constraints [Coyne 90]. For example,
most design frameworks could support a hypothetical problem-solver which produced
control system designs given appropriate knowledge, specifications and design con-
straints. Such design systems, however, are almost exclusively intended solely for the pro-
duction of artifacts. Thus, typical design systems could support the task of producing a
control system design, but the computational activity of constructing and running the
resulting artifact would be seen as a problem exterior to that of the design framework.

It seems reasonable to posit that a successful approach to creating control algorithms
for the management of dynamic physical processes might combine the strengths of mod-
ern control theory and knowledge-based design and planning. The architecture described
here, the Intelligent Process Control Architecture (IPCA), is based upon precisely that

combination of approaches.

1 .2 Approach

The overall problem-solving task addressed by IPCA involves constructing executable
control systems from declarative domain specifications. Task-specific frameworks are
used to specify appropriate knowledge structures and control strategies for the major ele-
ments of the overall task. The frameworks used are based on the Generic Task (GT)
approach to knowledge-based systems, a task-specific approach to knowledge-based sys-
tem analysis and construction [Chandrasekaran 86]. The IPCA approach considers the
overall problem-solving task of creating and applying appropriate control algorithms to be

composed of three major subtasks:

 

0 Subtask 1 - Determination of suitable control system features and pro-
cessing goals based on design specifications, constraints and domain-
specific knowledge of the system to be controlled.

0 Subtask 2 - Automated implementation of control system features and
goals by incorporating selected pre-compiled planning and control
algorithms into an overall process control plan.

0 Subtask 3 - Reactive execution of process control plans in a real-time
environment capable of supporting control algorithm selection and
goal maintenance.

It will be shown that Subtask 1 is a routine design and planning activity and that it is a
suitable candidate for implementation under the task-specific knowledge-based DSPL
framework for routine design and planning [Brown 89]. The task involves the creation of
an overall structure for the control system and specification of the overall processing
parameters and goals to be achieved by the system.

Subtask 2 involves the selection and assembly of appropriate goals, plan fragments
(pro-compiled static plans, heuristics and/or dynamic planning algorithms) and control
method fragments (pm-implemented, tested control algorithms) to implement process
control plans meeting the criteria generated by Subtask 1. Knowledge regarding the
applicability of each fragment (the conditions under which the fragment may be useful) is
included in the resulting process control plan implementation. A knowledge-based selec-
tion mechanism is utilized to provide a framework for supporting real-time fragment
selection based on situated applicability. The sponsor—selector system for real-time frag-
ment selection is described in Chapter 4.

Subtask 3 involves facilitating the execution of the resulting process control plan.

This task involves providing mechanisms for real-time fragment selection, data abstrac-

tion, goal maintenance and fragment library access. Communication with sensors, actua-

 

tors and external control systems is also supported.

Together, the subtasks combine knowledge about various planning/control strategies
and when each is appropriate, knowledge about the process to be controlled, and knowl-
edge about the composition and physical characteristics of the material to be processed.
The task-specific frameworks provide knowledge representations and control strategies
which are appropriate for each of the subtasks. The IPCA architecture integrates the three
subtasks into a comprehensive problem-solving framework which provides automatic
construction and execution of control systems based on declarative domain specifications.

A prototype system has been constructed based on the IPCA architecture. The proto-
type is based on a particular real-world domain, that of microwave-based composite mate-
rial fabrication. The prototype has been tested under a variety of fabrication conditions to
prove the viability and flexibility of the IPCA approach. The details of the architecture,
the underlying information-processing theory, the prototype implementation, the selected
domain, the experiments conducted and the results obtained are summarized in the chap-

ters to follow.

1.3 Outline

Chapter 2 provides background material including the research approach, an overview of
IPCA system architecture, expected results and evaluation criteria. The literature review
and related work section discusses task—specific architectures (including an introduction to
the Generic Task approach to knowledge-based systems), classical and reactive planning
systems, gives a brief overview of control theory, and introduces real-time knowledge-
based control systems. The strengths and uniqueness of the IPCA research approach and

contributions to the areas of AI and composite materials processing (the domain) are sum-

 

marized. The overall structure of the IPCA architecture is briefly outlined in the architec-
tural overview. Finally, the expected results and criteria for evaluation are presented.

Chapter 3 presents domain background regarding composite materials and fabrication
technologies. A survey of processing technologies is provided, with emphasis on a novel
single-mode microwave fabrication technique pioneered at Michigan State University.
Related work in the area of high-speed low-cost composite materials processing is pre-
sented to provide a domain context for the IPCA research. A prototype single-mode labo-
ratory-scale cavity and the associated external control system are described.

Chapter 4 includes theoretical background necessary for introducing the IPCA archi-
tecture. A complexity analysis of the overall IPCA information-processing task is pre-
sented. A knowledge-level analysis is then offered, followed by task descriptions of the
major elements which comprise the IPCA architecture. The following sections describe
the Generic Task (GT) approach to knowledge-based systems, including descriptions of
sponsor-selector mechanisms and an introduction to the DSPL language for routine
design.

Chapter 5 contains a detailed description of the IPCA system architecture. The Blue-
print Generator, Plan Assembler, Fragment Librarian and Plan Execution Monitor mod-
ules are described. The organization of knowledge within each module and the format of
information exchanged between modules are specified. The information processing
responsibilities and implementation of each module are discussed. Finally, the mechanism
for data exchange between the Plan Execution Monitor and the External Control system is
outlined.

Chapter 6 contains introductory and supporting material describing the experimental

 

laboratory-scale microwave-based fabrication setup. The initial section details the experi-
mental scope and objectives for each experimental series. A preliminary considerations
section describes sample composition, geometry, orientation and process instrumentation
common to all the fabrication experiments. The electromagnetic modes selected for pro-
cessing are given, together with heating profiles and mode maps for each experimental
series. The mechanical limitations of the single-mode cavity are summarized. The evalua-
tion criteria for the experiments is detailed in the final section.

Chapter 7 includes experimental data and analysis taken from material fabrication runs
in the laboratory. Summaries for each experimental series are given. A summary and
evaluation of experimental results and analysis is also presented for each series. Exten-
sions to the approach are discussed in terms of potential applicability to alternate compos-
ite materials processing technologies (including variable-frequency tuning) and within a
broader context of controlling dynamic physical processes outside the domain of compos-
ite materials processing.

Chapter 8 includes conclusions and a section addressing future work.

Appendix A contains a glossary of technical terms used in the dissertation.

Appendix B includes a user-level guide for the IPCA implementation, providing infor-
mation regarding the operation of all central IPCA software features, including user-inter-
face descriptions for all software components.

Appendix C contains a brief summary of the software utilities used in conjunction
with IPCA.

Appendix D contains a description of the External Controller and the protocol specifi-

cation used to support network-based communications with external control systems, sen-

 

sors and actuators.

Appendix E contains background information regarding the domain of composite
material fabrication.

Appendix F contains a case study involving the addition of a new control strategy
(specifically, a new type of mode-switching control fragment) to the IPCA system. Con-
siderations for adding capabilities to IPCA are presented. The considerations surrounding

Plan Generator maintenance and fragment library maintenance are investigated.

 

CHAPTER II

2.1 Related Work and Literature Review

The research presented here builds on the strengths of selected approaches from four
related areas and combines them into a unique comprehensive control architecture. The

work which is foundational to IPCA arises from:

0 Control systems theory
0 Knowledge-based planning systems
0 Task-specific knowledge-based system architectures

0 Knowledge-based routine design

Following a brief introduction which explains the manner in which the four founda-
tional areas contribute to the IPCA architecture, each of the areas will be introduced in the
order listed. A brief literature overview of each area will be presented.

Although modern control system development methodologies are effective, there are
many unrealized opportunities for exploiting knowledge to assist in the design, implemen-
tation, testing and tuning of such systems [Boyd 91]. Some complex control systems, for
example, are designed to support interaction between low-level control and higher-level

supervisory planning systems [Amen 89]. At present, the design, implementation, testing

 

10

and tuning of such systems is largely ad hoc. IPCA demonstrates that advanced computing
techniques can be used to provide tools to assist the designer with much of the control sys-
tem design and implementation process. IPCA reduces development time, promotes reus-
ability of control and planning algorithms, heuristics and approaches, and provides a
framework to capture knowledge about planning and control.

The knowledge-based IPCA architecture utilizes experience gained from control the-
ory by providing a standard set of implemented and tested control method fragments,
available for potential use in a variety of control system applications. Similarly, plan
fragments provide IPCA with a set of useful planning algorithms. The set of fragments
is easily extended to allow new control approaches, planning algorithms, techniques and
heuristics to be incorporated into IPCA. Fragments are intended for use and reuse in
many different combinations and configurations. They are used to produce a variety of
complex, effective algorithms for various controlled systems. By providing mechanisms
which intelligently select and assemble appropriate fragments into complete control sys-
tem solutions, IPCA assists the control engineering task by exploiting knowledge about
control and planning.

Because the IPCA architecture builds on solutions provided by modern approaches to
state-based control, control theory provides an important framework for a portion the
architecture and forms the basis for the algorithmic content of the control method frag-
ments. Control theory is therefore introduced in section 2.1.1. Similarly, knowledge-based
planning has an important role to play in the creation of effective plan fragments within
the IPCA architecture. Knowledge-based planning is therefore reviewed in section 2.1.2.

The knowledge-based problem-solving frameworks upon which IPCA is based are

 

11

task-specific. Task-specific approaches to knowledge-based systems are therefore
described in section 2.1.3. Finally, the contributions from the area of knowledge-based

design are treated in section 2.1.4.

2.1.1 The Contributions of Control Systems Theory
Classical linear control systems development methodologies call for a careful analysis and
characterization of a particular physical system, leading to an initial estimation of the
aspects of the system which must be monitored and controlled [Newton 57]. Many differ-
ent approaches to control may be considered during the development process, the
mechanics of which may include the construction of Bode plots, root-locus computations
and consideration of control processor characteristics such as whether proportional (P),
proportional-integrative (PI), proportional-derivative (PD) or proportional-integral-deriv-
ative (PID) control laws might be most effective [Boyd 91]. Any such approach can be
selected to form the basis for a control system architecture. Appropriate sensors and actu-
ators are selected and, in some cases, a model of the system is developed which specifies
outputs (actuators) as a function of inputs (sensors). The resulting model is sometimes
based on ordinary or partial differential equations. A control system implementation is
then developed and tested. The completed implementation is often tuned for optimal per-
formance. Performance may be evaluated against a number of criteria including regula-
tion, response, robustness, sensitivity and margin.

Although modern state-based linear control system design follows a similar develop-
ment methodology [Chen 87], the form of the model (i.e. the control laws) may be very
different. State-based control systems exploit the availability of digital control processors

which are capable of implementing more sophisticated control laws than could older vac-

 

12

uum tube-based or transistor-based control processors of the classical control systems era.
With modern control system designs, discrete control laws may approximate the continu-
ous forms covered under classical linear control systems theory or they may be much
more complex [Brogan 91]. Some of the more interesting possibilities include linear qua-
dratic regulators, estimated state feedback controllers and linear quadratic Gaussian con-
trollers [Astrom 90].

Advances in control theory continue to add a variety of new control approaches. Con-
trol approaches based on neural networks and fuzzy logic have appeared. Adaptive con-
trol algorithms have become an entire area of study. All these control approaches are
useful in particular circumstances, as well as many other types of control algorithms
which were not explicitly mentioned here. The import is that the IPCA architecture uti-
lizes control theory as a tool to provide a useful and appropriate repertory of control algo-
rithms from which process control plans may be assembled. Control algorithms employed
in the current IPCA implementation include single and multiple setpoint control, state-
based control and proportional control. These algorithms are contained in the control
method fragments and, taken together, fragments implement the control laws required to
control the processes of interest. IPCA is designed to be extensible so that new control

algorithms may be easily added at any time.

2.1.2 The Contributions of Knowledge-based Planning

Just as control theory contributes to the construction of control method fragments, knowl-
edge-based planning research contributes to the construction of IPCA’s plan fragments.
Unlike control method fragments, which focus on controlling process parameters directly,

plan fragments control process strategies through the manipulation of goals. In order to be

 

13

effective at the task of planning and reformulating process goals, the plan fragments may
draw upon the wealth of research which has been conducted in knowledge-based plan-
ning.

As briefly introduced in chapter 1, there are many types of knowledge-based planning
systems. All of them, however, share certain characteristics. All planning systems attempt
to generate sets of actions which will satisfy goals, given an appropriate formal descrip-
tion of the world.

Static planners, such as STRIPS, provided an effective approach to solving many
types of planning problems in static worlds [Fikes 71]. Such planners rest on the assump-
tion that the world will not change unless acted upon by the planner, but, unfortunately,
many real-world environments do not enjoy this property. Many early planning systems
such as STRIPS assumed a linear (completely ordered) plan structure. With the develop-
ment of NOAH, Earl Sacerdoti reformulated the problem of planning from world states to
one of a search through plan space and allowed the generation and manipulation of par-
tially ordered plans [Sacerdoti 74]. From this development, it became clear that plans
could be partially ordered and that it was often advantageous to defer placing planning
actions into a total order. Hence, many successful non-linear (partial order) planning
approaches appeared [Weld 95].

These encouraging developments led planning researchers in many directions. Some
developed reactive planning systems, which fixed plans only as they interacted with their
respective real-world environments [Allen 90]. These systems became popular in robotics
and autonomous vehicle research, since they could rapidly react to many unexpected con-

ditions without the need for detailed or extensive preplanning [Bihari 89]. Some

 

14

approaches were based on incremental planning, using least commitment strategies to plan
only as far ahead as required [Dean 88]. Other approaches called for pre-computation of
so-called universal plans, which function as pre-compiled reactive plan caches [Schop-
pers 87]. Many ideas developed out of the reactive planning community, including the
somewhat controversial theory of situated action [Brooks 86].

The RAPS (Reactive Action PackageS) system allows control over complex dynamic
processes by bundling reactive actions into “packages,” each of which competes for
resources until explicit goals are achieved. [Firby 88]. The RAPS system is similar to a
parallel blackboard architecture. The system provides a facility for defining and maintain-
ing goals (a blackboard), and each RAP (agent) competes for execution time in an attempt
to satisfy specific goals. The execution of IPCA process control plans is similar, but
IPCA does permit several actions to execute in parallel.

The Procedural Reasoning System (PRS) [Georgeff 87] is also similar to the reactive
execution environment of IPCA. The PRS system is composed of a database of beliefs
about the state of the world (using first-order predicate calculus), a set of goals to be
achieved, a set of knowledge areas (plans, KAs) which are presently available, an inten-
tion structure containing those plans which have been chosen for potential execution, and
an interpreter which manipulates all these components and selects KAs for execution.
KAs consist of an invocation condition and a procedural body. Unlike IPCA, PRS allows
plans to execute in parallel. Overall, the execution mechanisms of RAPS, PRS and IPCA
are similar, although RAPS and PRS are somewhat more sophisticated. However, IPCA
differs from both RAPS and PRS by providing an additional mechanism which supports

the automatic generation of plans.

 

. fl. -4 “-‘r-Hs

15

Advances in knowledge-based planning continue to produce a variety of new planning
approaches. Each of the approaches described here is useful in particular circumstances,
but there are also many other approaches to knowledge-based planning which were not
explicitly mentioned. The import is that the IPCA architecture utilizes knowledge-based
planning as a basis for providing a useful and appropriate repertory of planning algorithms

from which process control plans may be assembled.

2.1.3 The Contributions of Task-Specific Architectures

Task-specific architectures (TSAs) adopt the position that knowledge representation and
modeling should not be arbitrary “ad hoc” processes [Chandrasekaran 86]. Rather, TSA
approaches to knowledge-based systems hypothesize the existence of commonalities
between certain classes of tasks that can be exploited when performing analysis, modeling
and knowledge acquisition. The commonalities or “shared properties” of the tasks must
be such that they are meaningful across a large set of systems (i.e. they must be somewhat
generic), yet specific enough to offer a reasonable exploitation of constraints during analy-
sis and modeling (i.e. they must offer leverage during problem-solving). This “grain size”
problem is one of the major ways in which common TSA approaches differ.

Grain size has been the focus of much debate, especially between groups working on
different TSA approaches. But even within specific TSA approaches, there are problems
with achieving agreement on what constitutes a reasonable (and disjoint) set of problem-
solving mechanisms [Chandrasekaran 86].

Some of the common task-specific approaches to knowledge-level modelling include
Role-limiting Methods (i.e. Mole) [Marcus 88], Components of Expertise [Steels 90],

KADS [Vlfrelinga 92], and Generic Tasks [Chandrasekaran 86].

 

16

The Mole system (based on role-limiting methods) is used for knowledge acquisition
and explanation for cover-and-differentiate tasks often associated with diagnosis. The
grain size of role-limiting methods is quite small, and the resulting models are very well
specified. In fact, the models can be machine-translated into sets of rules for directly pro-
ducing the desired behavior.

Components of Expertise, a componential framework directed at the knowledge and
knowledge-use levels, attempts to combine four viewpoints into a complete methodology
for expert system design. The components of expertise include tasks/subtasks, domain
models, case models and problem-solving methods. The objective of the work is to create
a “handbook” which can assist by mapping task features into expert system solutions
[Steels 90].

KADS uses a “componential methodology” to construct layered models of tasks,
including distinct layers for domain, inference and task knowledge. KADS serves as a
knowledge engineering methodology by guiding the modification of an interpretation
model chosen by the designer. However, following the KADS methodology results in the
production of a high-level design that cannot be used to automatically produce executable
code. KADS is therefore primarily a tool for knowledge-acquisition and task analysis, not
for supporting implementation directly.

The Generic Task (GT) approach to knowledge-based systems differs in several
ways from other TSA approaches. First, the grain size is comparatively large. Common
tasks are directly supported by specific tools designed to help with problem-solving.
There is considerable debate over the number and nature of the Generic Tasks, yet several

tools have emerged to assist with common tasks such as hierarchical classification (CSRL)

 

17

[Bylander 86], functional reasoning (FM) [Sticklen 87] and routine design (DSPL)
[Brown 89]. These tools have been used to solve a variety of problems from many
domains [Chandrasekaran 86]. Each tool contributes knowledge structures and control
strategies appropriate to the task, thus providing an underlying theory of commonality
between tasks which are similar and easing system development by providing tools to
assist with problem-solving. The GT methodology will play a central role in the imple-

mentation of several key problem-solving components of IPCA.

2.1.4 The Contributions of Knowledge-based Design

One GT tool, the Design Structures and Plans Language (DSPL), makes a central contri-
bution to IPCA by providing a problem-solving mechanism in which to implement one of
the components in the architecture. The problem-solving task of selecting appropriate
fragments and goals to form the overall control plan implementation has been assigned to
a DSPL-based component. That is, the synthesis of lists of fragments and goals which are
suitable for inclusion in the final, generated control plan is treated by IPCA as a “routine
design activity.”

In general, the knowledge-based design process is concerned with the timely produc-
tion of designs which meet specific criteria [Coyne 90]. The resulting design is an abstrac-
tion from which an artifact can be constructed. In particular, routine or “class 3” design is
concerned with a particular subclass of the overall design problem [Brown 89]. Routine
design is concerned only with the production of designs which do not require “innova-
tion.” Some requirements for GT-based class 3 design problem-solving include the exist-
ence of effective problem-decompositions for the task, the availability of compiled design

plans for the lowest-level “component” parts of the design subtasks and explicit knowl-

 

18

edge about how to handle design component failures. It is hypothesized that much of
design falls under the type III designation [Brown 89].

A DSPL problem-solver contains a complex problem-solving mechanism which is
described in detail in chapter 4. DSPL supports a hierarchical decomposition of an overall
routine design task into a series of smaller design tasks, including suggested plans to
accomplish each of those tasks and knowledge to recognize when design failure of a sub-
component might occur. The problem—solver selects and attempts to follow design plans
to accomplish each of the subtasks in turn until the overall design is complete. This pro-
cess normally involves backtracking whenever failures occur, i.e. when one or more of the
design constraints are not satisfied. If it is determined that the constraints cannot be met
with any of the available design plans, the design effort ends in failure and either the con-
straints must be relaxed or one or more new design plans must be created which might
prove more successful.

The underlying problem-solving framework of GT-based class 3 design is one of state-
space search in a large but finite problem space [Newell 80]. Although the overall process
of routine design is theoretically a constraint-satisfaction problem within such a space,
conventional approaches to constraint satisfaction are not efficient at solving large, practi-
cal, complex problems in design. Propose and revise strategies can potentially cause the
entire state-space to be searched while other methods, such as hill-climbing, can result in
problems with local optima [Brown 89]. DSPL seeks to avoid these pitfalls by providing

appropriate, structured knowledge to guide the design process.

2.2 Research Approach and Architectural Introduction

The IPCA approach to intelligent control combines some of the strengths of task-specific

 

l9

architectures (in particular, the GT approach to knowledge-based systems), the DSPL lan-
guage for routine design, a knowledge-based sponsor-selector mechanism [Punch 89],
classical and reactive planning approaches and control theory [Decker 94]. The resulting
architecture is made possible by contributions from all of these areas.

The IPCA architecture provides a unique combination of design, control and planning
activities and has been tested in a practical, real—world domain to determine its effective-
ness. The domain chosen involves the control of a microwave-based fabrication process
for composite materials, though the IPCA architecture is not restricted to this domain.

The overall information processing task addressed by IPCA involves the construction
of planning and control algorithms for complex (process control) tasks.

I— Proc—es-s_
Control Plum

‘ uep 1
Q
Execution
Monitor-In

Plan Generation

 

 

 

 

 

 

 

 

 

Figure 2.1 - Overall IPCA System Architecture

As shown in figure 2.1, the overall problem is separated into three components:

0 The blueprint generation task
- The plan assembly task

0 The plan execution monitoring task

 

20

Although the context of composite material fabrication will be assumed in order to
describe the specific functions of the components in greater detail, the IPCA architecture
can be applied to many other process control domains.

The blueprint generation task is one of routine (class 3) design and is accomplished
by a hierarchical collection of specialists. Each specialist contains a set of design plans
addressing a particular sub-task of the overall design problem. Knowledge about the pro-
cess to be controlled is used in conjunction with knowledge about the applicability of par-
ticular control and planning approaches to synthesize an applicable plan blueprint for a
specific environment. The plan blueprint specifies a collection of plan and control method
fragments (strategies, policies, approaches, algorithms) and appropriate goals and control
information for a particular environment. The Blueprint Generator contains explicit,
structured knowledge which is used together with more specific process information con-
tained in a fabrication model. Each fabrication (domain) model contains specific informa-
tion relating to a particular process. For example, in the microwave-based composite
material fabrication process, the fabrication models include a description of part geometry,
physical characteristics of the material to be processed, and the availability and configura-
tion of processing hardware, actuators, and sensors.

The plan assembly task involves retrieving the appropriate plan and control method
fragments referenced in the plan blueprint and assembling them into a hierarchical struc-
ture based on knowledge about their applicability. The resulting structure (which is spe-
cific to a single process, material and geometry) is called the process control plan.

The plan execution monitoring task involves the application of the process control

plan, based on information gathered through real-time process monitoring. One fragment

 

21

from the process control plan is chosen for execution during each cycle, based on esti-
mated utility in the current process state. The selected fragment is executed, thereby pro-
ducing a series of control outputs and goal state changes which effect the process. The
execution cycle then repeats after all sensed information (and thus, the estimated process
state) is updated. The same process control plan is used throughout the execution process,
so process control plans must typically contain a variety of control methods, plans and
strategies for a variety of problem—state contingencies.

Chapter 4 contains a task-level analysis of each of the three highest-level problem-

solving tasks. The architectural details of IPCA are reviewed in chapter 5.

2.3 Expected Results and Evaluation Criteria

Expected results and evaluation criteria are divided into two areas: those expected within
the area of knowledge-based systems and those expected in the domain.

\Vrth regard to knowledge-based systems research, particularly within the GT
approach to knowledge-based systems, this research is expected to offer the following
contributions:

0 Analysis, architecture and implementation of a complete knowledge-
based system including process control plan generation, assembly and
execution.

0 An extension of the GT approach to knowledge-based design and
planning through the incorporation of a knowledge representation
and control strategy for the synthesis and execution of process control

plans.

0 A structured framework for representing knowledge regarding com-
posite materials processing.

0 Broad applicability and ease of knowledge transfer to other process
control domains.

 

22

An intelligent control system built on the IPCA architecture has been implemented in
VisualWorks SmallTalkT" and integrated with an existing implementation of DSPL
[Kamel 94]. The resulting implementation has met certain subjective criteria, including
responsiveness, ease of use and interface consistency. It will be shown that the implemen-
tation demonstrates the viability of the IPCA architecture and validates the appropriate-
ness of the architecture in the chosen domain.

Within the domain of tuned single-mode microwave-based polymeric composite mate-
rial fabrication, the research is expected to produce the following contributions:

0 A facility to support the automated synthesis and execution of discrete
process control plans yielding composite structures of acceptable
quality within a rapid timeframe.

0 Knowledge-based frameworks for effectively representing domain-
specific knowledge of single-mode microwave-based composite mate-
rials processing.

. Breadth and depth of domain coverage sufficient to demonstrate the
applicability and effectiveness of IPCA in this domain.

0 A reduction in the average thermal gradient within the material during
curing beyond that achieved through the use of previous control
approaches.

0 Theoretical proof-of-concept for rapid, controlled material heating.

0 Exploration of the limitations of mechanical mode tuning.

0 Ability to support additional fabrication technologies including vari-
able-frequency microwave tuning.

The initial implementation supports one polymeric matrix material and three geomet-

ric structures as a proof-of-principle. Quality judgements of the resulting composite struc-

tures are based largely upon physical examination and analysis of temperature gradients

 

23

measured during fabrication (see chapter 7). IPCA showed a significant reduction in the
measured thermal gradient when compared to fixed-mode microwave fabrication tech-
niques. It is further demonstrated that a reduction in thermal gradient has been achieved
when compared to static “complementary” mode switching approaches, and that the gradi-
ent is ultimately limited by the mechanical response of the mode switching apparatus. A
degree of breadth in domain coverage is also demonstrated. Finally, a foundation is
offered for the ability to conduct rapid materials processing using higher power and vari-

able-frequency tuning.

 

CHAPTER III

3.1 Domain Introduction

The effectiveness of the IPCA prototype has been demonstrated by applying the imple-
mentation to a series of related real-world process control problems from a particular
domain. All experimental evaluations conducted with IPCA address the domain of poly-
meric composite material fabrication. In this section, the domain characteristics which are
responsible for the suitability of IPCA to the chosen domain are examined.

The majority of composite materials are composed of two constituents. The matrix
material is used to support and protect the reinforcement. There are several classes of
matrix, including polymeric, metal and ceramic types. Similarly, there are several types of
reinforcements, based on shape (including fibers, particulates and whiskers) and on mate-
rial (such as glass or carbon). Polymeric carbon-fiber reinforced composite fabrication
has been used as a focus for demonstrating the effectiveness of the IPCA architecture.
The process to be controlled is based on a particular class of polymer composites known
as thermoset polymeric composites. For additional detail regarding composite materials,
refer to appendix E.

Fabricating polymeric thermoset composite structures is accomplished by initiating

and maintaining a chemical crosslinking reaction through the application of heat and/or

24

 

25

pressure. Most commonly, composite structures are fabricated or “cured" in autoclaves
under heat and pressure, and these approaches are briefly introduced in section 3.3. How-
ever, the processing technology used for demonstrating IPCA uses microwave energy to
initiate and maintain the exothermic crosslinking reaction. Microwave processing has a

number of desirable features which are described in detail in section 3.4.

3.2 Related Domain Research

Research into low-cost, high-speed composites processing addresses a wide range of dis-
ciplines, technologies and approaches, including high-speed processing, basic materials
research and the use of composites as structural components. The related research forms a
domain context which extends well beyond intelligent control of microwave-based fabri-
cation for polymeric thermoset composites. For completeness, an overview of the sur-
rounding domain context is provided. This section discusses some of the related research
into composite materials and structures and the issues involved with low-cost, high-speed
materials processing. Closely related research conducted at the Composite Materials and

Structures Center (CMSC) at Michigan State University is presented.

3.2.1 Composite Materials Research at Michigan State University

The CMSC is a cooperative multi-disciplinary research center for composite materials and
processing research, which also serves as a technology transfer agent to the industrial sec-
tor. The work conducted at the CMSC includes research sponsored by the NSF Center on
Low-Cost, High-Speed Polymer Composites Processing and the Michigan Materials and
Processing Institute (MMPI). The relationship between CMSC, MMPI and the NSF cen-

ter is demonstrated in figure 3.1.

 

26

 

 

 

 

 

 

 

 

  
    
 

 

 

 

 

 

 

 

 

 

 

 

Composite Materials Michigan Materials
and Stnictures and Processing
Center (CMSC) Institute (MMPI)
. . - Recycling and
dvanced Engineering NSF Center for Waste Mana ement
aterials Lab Experiment Low-cost, High-speed g ‘
Station Polymer Composites Joining and
Processin Attachment
Interfaces Liquid Advanced Recyclable Novel
and Joining Molding Computing and Renewable Processing
I Materials and Thermal

 

 

 

Management
I PCA

Figure 3.1 - Relationships Between CMSC, MMPI and the NSF Center

MMPI is a non-profit institute focused on polymer composites for the durable goods
industries. The board of directors and executive committee oversee four technical com-
mittees. Technical committee 1 (TCOMl) addresses Design and Performance Issues in
Composites. The NSF Center is considered to be TCOM2. TCOM3 addresses Recycling
and Waste Management. TCOM4 is Technology Transfer. Each committee has a board
which oversees the projects.

The NSF center is focussed on increasing the competitiveness of US-based industry
and encouraging local economic development through a federal/state/industry partnership.
The NSF center is focused on basic and applied research which will help to provide high-
quality polymer composites at high-speed and low-cost for consumer goods applications.
The technical thrust areas for the NSF center include novel processing and thermal man-
agement, liquid molding, recyclable and renewable materials, advanced computing tech-

niques for composite design and processing, and interfaces and joining. The IPCA

 

27

research is included under the thrust of advanced computing techniques.

Although IPCA research represents one of the important activities within the advanced
computing thrust, there are many other research projects which are closely related. The
following two sections detail some of the other research approaches in electromagnetic
processing and advanced computing techniques and also delineate relationships to the

work done with IPCA.

3.2.2 Electromagnetic Processing of Polymer Composites

Much of the research into high-speed polymer processing has focussed on the utilization
of electromagnetic fields for rapid material heating. Electromagnetic processing forms an
important basis for the research conducted with IPCA.

Research in electromagnetic processing includes work on cavity characterization and
automation detailed in [Adegbite 95]. This work formed one of the starting points for the
work with IPCA and also influenced the development of the IPCA architecture. The cav-
ity described in chapter 3 was built while the IPCA research was in progress, and the cav-
ity was characterized as IPCA was in the later stages of development. The goals of this
related work were to automate the single-mode microwave process studied previously in
[Jow 88] and to study the high-speed processing capabilities of the automated microwave
fabricator.

The microwave cavity described in detail in chapter 3 was built and outfitted with an
external control system as one part of the work described in [Adegbite 95]. In addition,
the electromagnetic modes of the resulting cavity were empirically characterized through
a series of experiments and the results were compared with theoretical EM field models

[Adegbite 95]. A closely-related variant of the conventional control system developed and

 

28

described in [Adegbite 95] was developed and used in conjunction with the IPCA proto-
type. This conventional control system is described in chapter 5 and appendix D. The con-
ventional control system was designed to perform control and sensing functions which are
specific to the controlled (domain) process. As described in section 5.4, IPCA cannot
function without an External Controller. It is therefore important to emphasize that the
conventional control system forms an integral portion of the IPCA prototype, and that the
related research on cavity characterization and closed-loop control described in [Adegbite
95] influenced and complemented the work done with IPCA. This interdisciplinary
research model has been used extensively and with great success within MMPI.

The mode switching research of [Fellows 91] is also closely related to the work con-
ducted with IPCA. Fellow’s research explored a technique known as “mode switching”
which involves alternating between two or more resonant modes (see section 3.4) to help
ensure uniform heating. The work was extended to allow for the processing of complex,
non-planar shapes, and showed an encouraging improvement over the previous micro-
wave-based processing. However, the approach required the manual generation of heating
profiles for each mode to be considered, which proved difficult and time-consuming. The
approach also relied upon pre-defined static thresholds for controlling the selection of
modes. A variety of materials could not be processed without changes to the controller
software, nor could a variety of microwave fabricators be supported without creating mul-
tiple control programs. Although the work demonstrated the viability of the mode switch-
ing approach, it also pointed toward a need for a more sophisticated support system, for
composite material fabrication. One of Fellow’s observations was that more sophisticated

control algorithms were needed to effectively control material processing using mode

 

29

switching techniques [Fellows 91]. The quantitative results obtained in [Fellows 91] are
valuable, and are later used as a basis for evaluating the laboratory fabrication results

obtained in this research.

3.2.3 Research in Advanced Computing Techniques

A variety of advanced computing techniques for polymer composites processing have
been explored and are related to the research which eventually produced the IPCA proto-
type. The research projects described in this section share the goal of using advanced
computing techniques to assist with the design and manufacture of composite structures in
some manner.

The research described in [Kamel 94] resulted in a knowledge-based design system
which assists a human designer with the task of generating multiple composite material
selections (designs) based on specific design requirements. The underlying system archi-
tecture is based on a tool developed to support routine design activities [Brown 89]. The
research showed that knowledge-based material selection was possible, and that knowl-
edge-based material selection was a feasible approach to a difficult problem. In addition,
Kamel made an important contribution in extending the underlying routine design frame-
work to generate multiple designs. As with many MMPI research efforts, Kamel’s work
made contributions in the domain of composite material design and also within the knowl-
edge-based systems community.

The system described in [Kamel 94] was subsequently extended to cover a wider
range of materials as a result of work done by [Lenz 94]. The resulting third-generation
decision support system (COMADE) was not an architectural extension, but information

regarding many alternate materials was included to provide broader knowledge-based

30

support for material selection. Support for thermoplastic matrices, expanded thermoset
coverage and chopped-fiber matrix materials were included in the COMADE effort, prov-
ing out the extensibility of the knowledge-based design approach.

The approach of developing advanced computing architectures and subsequently
extending their capabilities within their respective domains has proven successful within
the NSF Center. This approach forms the working model for many of the projects within
the Center, and allows the interdisciplinary development of advanced computing architec-
tures and small-scale prototypes (such as IPCA). The approach also provides extensible
computing frameworks which are typically used to represent additional knowledge from
domain experts to increase their scope and their usefulness.

Other intelligent systems for decision support have been developed, including a sec-
ond-generation fabrication technology selection system described in [May 94]. This fab-
rication technology selector is used to recommend appropriate processing technologies
when given a set of design requirements including material types, economic factors, pro-
cess throughput and part geometry. The system uses a knowledge-based hierarchical
selection mechanism to choose an appropriate fabrication method or multiple methods.

The ultimate goal of the Advanced Computing research thrust involves combining the
intelligent support systems into an automated end-to-end support system suite for compos-
ite material selection, part design, fabrication technology selection and fabrication control
[McDowell 93]. The IPCA research contributes to this goal by providing a framework for
effective knowledge-based fabrication control for a variety of composite materials and
geometries. Since it is a first generation system, the current IPCA prototype does not yet

support a wide variety of materials or processing technologies. However, the IPCA sys-

rmc;
nfilt
hasal

meta

3.3 (

firni

midi
wrdu
ittdba
drkci
mans

lpkd

31

tem can be extended to cover additional materials and processing technologies in a man-
ner very similar to the effort which ultimately produced COMADE. The IPCA prototype
has already been extended on a small scale, and those extensions are described in detail in

the case study supplied as appendix F.

3.3 Conventional Composite Processing

Therrnoset polymeric composites are typically cured in autoclaves under heat and pressure
[Loos 83]. Key process variables include time, temperature, heating rate, pressure and
cooling rate. The relationships between these variables are complex, but it is possible to
qualitatively characterize major interactions to a useful extent.

The curing profile specifies temperature and pressure to be applied to the material as a
function of time. Thermocouples and other temperature-sensing devices are commonly
used to monitor the internal temperatures in the material being cured in order to provide
feedback to either a human operator or a control system. Also directly measurable is the
dielectric constant of the material, which indicates ion mobility, and, indirectly, an approx-
imate state of the cure. Viscosity may also be measured through the use of rheometers. A

typical curing profile for an epoxy resin is shown in figure 3.2.

32

s
§

§

 

 

 

 

Temperature (degrees C) —

0 60 120 180 240
Time (minutes)

Figure 3.2 - Typical Autoclave Curing Profile

The typical autoclave curing profile depicted in figure 3.2 includes a ramp-up time for
allowing the material to come up to a desired processing temperature while minimizing
the formation of large voids, caused by trapped water vapor in the material. Pressure is
applied to attempt to consolidate the material and minimize void formation. The applica-
tion of heat and pressure continue until the curing cycle is complete. Often, the material is
postcured in an attempt to ensure that the crosslinking reaction is complete and the fin-

ished structure has the best material properties possible.

3.4 Microwave-based Processing

A number of alternate fabrication methods for polymer composite structures have been
investigated, including the microwave curing of thermoset polymers [Asmussen 87].
Some of the advantages of microwave curing are rapid, controlled heating, high energy

efficiency, low cost and high throughput [Wei 93]. It is for these reasons that microwave

33

processing technologies are being investigated as an energy-efficient means to produce
low-cost, high-quality composite structures at high speeds.

Microwave processing involves curing composite structures through interaction with
microwave-frequency electromagnetic (EM) fields. The applied EM field causes rapid
heating within the composite structure. The interaction between composite structures and
the applied EM field is highly complex, however, and dependant upon a number of fac-
tors, including the shape, volume and physical characteristics of the cavity used and the
characteristics of the composite structure. The geometry, mass distribution, placement
and dielectric constant of the composite structure all affect how the EM field will be per-
turbed. There are so many factors to account for that predicting a pattern of interaction
between a given EM field and a given composite structure remains an extremely difficult
and unsolved problem. However, some generalities may be stated. For example, “lossy”
reinforcement materials such as glass-fiber are distinguished by their higher dielectric
constants. Lossy materials therefore heat more slowly than lossless (e.g. carbon-fiber)
reinforced materials. This is true because lossy materials absorb less energy from the
applied EM field. The overall dielectric constant of the material thus plays a key role in
determining how the structure will interact with an applied EM field.

The cavity characteristics are also central in determining EM field interactions, since
the cavity behaves as a waveguide which focuses EM energy. Depending on the physical
characteristics of the cavity and the frequency of the applied EM field, microwave heating
may be single-mode or multi-mode in nature. The two types of heating differ in howthe
BM energy is focused.

The applied EM field used for multi-mode heating may be visualized as a linear com-

34

bination of several distinct EM “basis” vectors. The cavity is detuned in the sense that a
stable standing-wave pattern does not exist and the energy is “randomly” distributed. The
pattern is not actually random, of course, but depends upon so many factors that it is
impossible to predict without making many simplifying assumptions. This means that the
energy delivered through multi-mode heating is largely unfocused. Power absorption is
generally quite low and therefore multi-mode systems do not heat efficiently, although the
overall heating distribution may be more uniform than could be achieved through the use
of any single basis field vector. The main advantage to multi-mode heating is that it can be
used to heat many different types and shapes of materials with no need for any feedback,
tuning or focusing. For this reason, commercial microwave ovens use multi-mode heating
exclusively. Although multi-mode heating is not efficient, it can be used to heat a wide
range of materials to moderate temperatures fairly evenly.

Single-mode heating offers the ability to heat materials more rapidly and much more
selectively. When using a single-mode heating technique, a standing wave is set up in the
cavity so that maximum energy may be focused into the material. This approach has both
a major advantage and a disadvantage. The advantage is that the material heats rapidly
and, since efficiency is very high, that the material may be heated to high temperatures if
desired. The disadvantage is that all the EM energy is doing work on the material in a spe-
cific, static, and generally uneven fashion. This means that the pattern of heating within
the material is often uneven, since the applied EM field is more focused on some parts of
the material than on others. This is troublesome for composite materials manufacturing
because thermal gradients in a composite structure are undesirable since non-uniform

heating negatively impacts the physical properties of the finished structure. Single-mode

35

heating techniques produce efficient, rapid energy delivery, but power application and
mode selection must be carefully monitored and adjusted to minimize thermal gradients
and maintain specific material processing temperatures. IPCA is designed to specifically
address this problem by providing automated process control plan construction and execu-
tion.

Research on mode-switching for processing 3-dimensional low-loss polyester/glass

composite structures based on empirically-gathered heating data can be found in [Fellows
95]. Mode-switching techniques proved useful in this work for maintaining sample tem-
peratures between the onset of reaction and the onset of thermal degradation for complex
3-dimensional structures. Two or more “complimentary” modes were selected based on
heating experiment results with a cured sample. The complimentary modes were then
chosen during subsequent processing runs in an attempt to achieve uniform curing.
Mechanical testing confirmed that the properties of the resulting structures were not as
desirable as those achieved through autoclave curing. Best—case results from 12—ply 2-
dimensional polyester/glass samples and 60-minute cure cycles involved temperature gra-
dients of approximately 30 degrees C. Fellows concluded that variable-frequency tuning
would not only reduce these gradients further, but would allow the application of pressure
during processing.

A diagram of a sample composite layup is shown in figure 3.3. The layup includes a
hand-laid composite structure (the sample) enveloped by several layers of various films

and enclosed in a teflon mold.

36

Upper mold

Release film
Non-porous film
Porous film

Composite sample
Porous film
Bleeder cloth
Non-porous film
Release film

[‘ ] Pressureplate

F . _ , g r] Lower mold

Figure 3.3 - Composite Layup for Microwave Processing

 

 

 

 

 

 

The porous film on either side of the composite allows excess resin to flow away from
the sample while preventing adhesion. Excess resin flows through the porous film into
layers of bleeder cloth. One layer of bleeder cloth is included for every 3 plies in the sam-
ple. Under the bleeder cloth is a layer of non-porous film, designed to contain any excess
resin which is not absorbed by the bleeder cloth layers. A release film is used between the
non-porous film and the mold as a final guard against leakage of epoxy. The entire layup
is enclosed in a tefion mold, which ensures that the material will remain in the desired
shape during processing, while simultaneously providing access for temperature probes to
the composite structure within. The structure of a typical mold is shown in figure 3.4.
This mold was used in the fabrication of 1” square composite structures, detailed in chap-

ter 7. Probe placement and identification are given in chapter 6.

37

0.375" dia holes
0.5“ depth x 4

  
 
 
 

 

L

i

 

 

 

 

 

I
2.875" 1.375“ O + Top view
l ,
i
| 0.125" dia holes
. thru x 4
l

 

.\\ I .\‘
: 1.375” I I l ‘1'. Side view
- ‘ s-t

Figure 3.4 - Teflon Mold for Microwave Processing

 

There are several methods for creating and maintaining a particular EM mode within a
given cavity. If the cavity dimensions are assumed fixed, various modes can be selected
by changing the frequency (and therefore the wavelength) of the applied EM field until it
is an integral multiple of the cavity length. For example, if the cavity were 40cm long, the-
oretical (empty-cavity) modes would be present at wavelengths of 40cm (n=l), 20cm
(n=2), 13.33cm (n=3), 10 cm (n=4), and so on. The exact wavelengths required to pro-
duce standing waves in a loaded cavity would depend upon many other factors, as men-
tioned earlier. Nonetheless, it would be a simple matter to switch from one mode to
a“Other by changing frequencies. Although this approach would offer the ability to rap-

idly and efficiently change modes without altering the physical characteristics of the cav-

38

ity, it would require an extremely costly variable-frequency microwave power supply.

Many research budgets do not allow for such an expenditure, so a lower-cost cavity tuning

approach was developed.

3.5 The Experimental Single-mode Resonant Cavity

An experimental single-mode variable-length cavity has been developed for laboratory-
scale microwave-based composite materials processing [Adegbite 95]. The cavity fea-
tures a pair of motor-controlled actuators which alter the cavity length and the coupling
probe depth to facilitate the selection of specific EM field patterns (modes) in conjunction
with a fixed-frequency microwave power supply. Although this technique does not allow
instantaneous switching between modes, it has the advantage of low cost and is nearly as
flexible as a variable-frequency power source. However, there are limitations imposed by

the time required to switch between modes. These limitations are explored in chapter 6.

A diagram of the experimental cavity is shown in figure 3.5.

39

  
   
  

Cavity Tuning Mechanism

 

.......
r » wwwww

Position sensor .

Probe Tuning Mechanism

 

 

 

 

 

Probe

-
h-

" H | Sample Loading Mechanism

Figure 3.5 - Experimental Single-Mode Microwave Cavity

The experimental cavity is a hollow, cylindrical brass structure 17.78 cm in diameter
and 40.00 cm in length. Two brass shorting plates are used to seal the ends of the cavity.
The lower shorting plate is raised into position by a motor-controlled sample loading
mechanism and remains fixed in position until the sample is removed after curing. The
lower shorting plate serves only to support the sample and seal one end of the waveguide.
It does not move at any time during material processing. The upper shorting plate is also
equipped with a motor-controlled cavity tuning mechanism and is used to select particular
EM field patterns during processing. Each desired mode corresponds to a particular cavity

length (LC). In addition, a probe-tuning mechanism is used to modify the depth of the

4O

microwave coupling probe (PD). This adjustment is useful when fine-tuning the cavity

modes, since slight movements of the coupling probe can compensate for small variations
in sample placement, geometry and composition without the need to modify the cavity
length.
A number of sensors are included in the design of the cavity. A pair of precision 5K-
Ohm linear potentiometers are used as linear sensing elements for the cavity length and
probe depth positions. These sensors are employed to give feedback on the position of the
upper shorting plate and the microwave coupling probe. The incident and reflected power
are monitored through the use of a pair of power sensors and analog meters.

In addition to these dedicated sensors, a variety of invasive and non-invasive sensors
may be used to monitor the temperature of samples, including fiber-optic probes and non-
invasive IR sensors. Most experiments conducted with this equipment have employed a
Luxtron 755 four-channel fluoroptic temperature measuring system. Two fiuoroptic ther-
mometers (each with four temperature probes) can be used to monitor up to eight temper-
atures inside the cavity. Electrode—based dielectric sensors are not used since metal cannot
reside inside the cavity during processing, but alternative methods for estimating dielectric
constant of the material are under investigation.

Refer to appendix D for more detail about the External Controller and the equipment
setup. Chapter 6 contains a description of probe placement and configuration for the vari-

ous fabrication experiments.

3.6 External Controller Overview

In the current implementation, the External Controller preforms sensor monitoring, pre-

41

conditioning, low-level mode tuning, and other activities thus relieving the remainder of
the architecture of these responsibilities. The prototype External Controller is written in
the National Instruments LabVIEWT" development environment [National 91]. Lab-
VIEWT" provides graphical programming tools and an extensive “virtual instrument”
library for communication with a wide variety of laboratory instruments connected to an
IEEE-488 bus or attached to a National InstrumentsTM I/O interface. The External Con-
troller monitors sensors attached to the I/O interface including temperature, power and
position and is designed to maintain a tuning lock on whichever mode is selected by the
intelligent controller. Since the External Controller runs on a dedicated machine, it can
rapidly monitor reflected power and adjust the cavity length and probe depth positions to
keep the reflected power to a minimum as the dielectric constant of the sample changes
during fabrication. A PID control algorithm is employed to accomplish this task [Astrom
90].

Since the External Controller and IPCA run concurrently, the External Controller has
time to perform sample averaging and sensor preconditioning to reduce sensor noise. For
example, the External Controller can sample the AID converters on the I/O interface hun-
dreds of times per second. Spurious or noisy readings can be discarded while others are
averaged since the intelligent controller only needs the sensor readings once every 5-10

seconds.

The External Controller serves to provide a vital link between IPCA and the physical

sensors and actuators required to control most real-world processes. Although neither the
form nor the implementation of the External Controller is specified by the IPCA architec-

ture, the services which it typically provides are described. The interface protocol

42

between the External Controller and IPCA is described in appendix D. Additional detail

regarding the External Controller can also be found in appendix D.

CHAPTER IV

4.1 Theory of Operation

The overall problem-solving task addressed by IPCA involves creating a mapping from a
set of sensor values to an set of actuator values for use throughout process control plan
execution. The underlying nature of the process control task will therefore be examined.

An example diagram showing the control problem in continuous time is given in figure

4.1.

 

l' 1 ——>
Controller

1

Figure 4.1 - Structure of a Continuous-time Control System

 

 

 

 

"ii—T'—>

 

Referring to this continuous-time model, the controller maps from control inputs

Yinyq (sensed values representing the state of the process) and reference commands r1..rn

(describing desired states for the process) to produce appropriate control forces ul "up to

43

44

be exerted on the process. This relationship is demonstrated by the equation
i" -i i- i- 1\ \

” r10) ylitfl am)

500): ’2“) x y2(t) =9 “2“) ,Vueattzo)

 

 

 

 

 

 

 

 

 

 

((_r,,(t)_ LquL bup(t)_ ) j
which assumes a continuous (linear or non-linear) time-variant system. As stated in chap-
ter 2, classical control theory assumes linear, time-invariant systems. Therefore, it can
already be seen that the production of such a relation for non-linear, time-variant systems
will require more than classical control theory can offer. There are n-q continuous-valued
time-dependant equations required to generate the p control forces for any given value of
r. The complexity of producing this mapping for the continuous-time case will not be
investigated, since modern state-space based digital control has been selected to model
processes investigated by this research.

Modern digital controllers are designed as discrete-time systems using quantized
(sampled) inputs and outputs. Discrete-time process control systems have a general struc-

ture as depicted in figure 4.2.

 

..___,.
Clock
r IT l
mi“) —> tank) —> wit) 0-. y1(t)_..

‘ All.) . - Process - >—
Vd‘k’ —’ 69m) —> tiplt) a» id!) —>

t t

”(‘9 Inflk)

 

 

 

 

 

 

 

 

 

 

Figure 4.2 - Structure of a Digital Control System

45

The discrete-time digital controller samples the continuous process outputs y,(t) at reg-
ular time intervals to form discretized control inputs y,(tk). These inputs, together with
reference commands rim) are used by the controller to generate a set of discrete control
forces up(rk). The discrete control forces are then transformed into continuous values ui(t)
through the use of a D/A converter which holds the control force outputs constant between

conversions to maintain a continuous output signal. The controlled process reacts to the

control force outputs and is sampled again at the next discrete time interval.

4.2 Controller State-Space Complexity Analysis

Modern control theory makes a number of assumptions to simplify the task of creating an
effective digital control system. The relationship between input and output may be

expressed for the discrete-time case as
r '1 (’1‘) Y1(tk) “1(tk)

pm): ’20:.) x ”W = “2W ,V(keZ,k20)

\

 

 

 

 

 

 

 

 

(In (’1‘)‘ Ly? (tk)_l _“p(tk)d )
where the tk represent discrete time instances dependant on the sampling period, h. Thus,

the relationship
tk = k - h, ke Z

holds, and the sampling frequency (in samples per second, or Hz) is given by

l
Fs=72

Assuming that we are interested in controlling the time-variant process for Tf seconds, we

46

have

k = [1,?" = LTf'F-ri
Further, assuming each control input yi is quantized into Qy levels (using a log2(Qy) bit A/
D converter) and each of the reference commands r,- can be represented using Q, discrete

states (using log2(Q,) bits), then the size of the domain in the resulting control state-space

is given by

k.2(Q,-q-Q.-n)

and, assuming Qp quantization levels (logz (Qp) bit D/A converters) for the control forces

ui, the co-domain of the control space is of size

2(Q.-p)

Therefore, we find a set of

(Q,-q-Q,-n) 2(Q..-p) (Qy-q-Q,-n-Q.'p)

(k°2 )-( )=[_Tf-Fs_]-2
distinct mappings from the domain to the co-domain.

The overall information-processing task of IPCA involves the design, synthesis and
application of a selected mapping from among this set, using knowledge about the process
to be controlled. But even for the most modest control task, the size of this space is enor-
mous. For example, using only 4 process inputs and 4 control force outputs with 4 bit AID
and D/A converters, a sampling period of 1 second, and two 2-bit reference commands to
control a time-variant process for 2 seconds would yield

4..2..4. 17
L7,f.1:,s".2(Q,qQ,n Q. 1?) =L2°1J'2(2 42 2 2 4) =2(2 +1)~4.5x1039

possible control mappings from which to select.

Of course, IPCA does not represent process control plans as simple mappings. The

47

complexity analysis is intended only to show the size of the problem space. The method
by which IPCA uses knowledge to generate process control plans is described in the fol-

lowing sections.

4.3 Knowledge Level Analysis

The proposal of the knowledge level (KL) hypothesis has been extremely influential in
modern artificial intelligence research [Newell 82]. The KL hypothesis states that an
implementation-independent computer systems level (the knowledge level) exists imme-
diately above the symbol level. The symbol level, in turn, exists immediately above the
implementation level. Each computer systems level is associated with particular primitive
constructs and properties. For example, the gate level contains primitive elements such as
AND, OR and XOR logic gates, flip-flops, clocks, and so on. The elements of the register
transfer level (comprising registers, adders, latches, etc.) are constructed from the ele—
ments from the gate level. It is possible to discuss structures at the register transfer level
without including details about the next lower (gate) level.

Similarly, the elements of the knowledge-level are composed of primitives from the
symbol and implementation levels, making it possible to describe activities at the knowl-
edge level without committing to any particular implementation. As with the other levels,
the knowledge level is a distinct computer systems level complete with its own properties
and constructs. The KL hypothesis argues that a problem-solving analysis of a task can be
given without any reference to the particular implementation details of how knowledge is
represented (frames, rules, semantic networks, etc.). In fact, a knowledge level analysis of
the behavior of a system is based on an “external observer” standpoint, treating the system

as a black box. The KL analysis, then, attempts to explain the behavior of the system in

48

terms of what is does (observation), how it works (mechanization) and why it works the
way it does (rationalization) [Van deVelde 93].

Therefore, the following analysis of the overall problem-solving task addressed by
IPCA is given. The task analysis is from the viewpoint of an external observer, and avoids
discussion of the implementation details of task decomposition and knowledge representa-
tion.

At the knowledge level, the overall information-processing task addressed by IPCA
involves the generation of control forces which effectively manage a variety of complex
controlled systems based on high-level descriptions of those systems and real-time feed-
back from the systems. The specific requirements for the IPCA task are thus subject to the
nature of the input and output descriptions. These descriptions will therefore be made
more explicit, without a commitment to any representation, before the analysis proceeds.

As input, we observe that the IPCA task accepts a description of a particular system to
be controlled. This description is relatively high-level, i.e. it makes specific references to
the real-world equipment involved and information about how it is configured into an
overall processing system, but does not include any formal mathematical description of
the process. For the domain of microwave composite material fabrication, the input
description also contains information about the material to be processed, including the
material composition and geometry. Again, these descriptions are high-level, including the
names of the materials, for example, but not their chemical or structural descriptions. The
output requirements for the task are somewhat modest. The system needs to generate a
series of control force outputs which will effectively manage the controlled system in real-

time.

49

Using the given information about the nature of the input and output of the task, we
can describe some of the knowledge which must be present to support problem-solving.
Obviously, the task must contain knowledge which supports a mapping from the current
system configuration into a set of control force outputs. In order to generate meaningful
control force outputs, the system must contain knowledge about the physical sensors and
actuators detailed in the input description. Since the system is described in terms of physi-
cal equipment, materials and geometry, the task must utilize knowledge about the nature
of the specific configuration of these items, including the relationships between sensed
inputs and appropriate outputs. Knowledge about the characteristics of various materials
must be included.

A more detailed KL analysis cannot be performed without making assumptions about
how the system is structured and how knowledge is used inside the “black-box.” There-

fore, the overall task will be decomposed into three subtasks for further analysis.

4.4 IPCA Task Descriptions

The IPCA architecture views the overall problem-solving task as consisting of three dis-
tinct information-processing subtasks. As shown in figure 4.3, the tasks include synthesis
of appropriate process control plan specifications, assembly of process control plans from
compiled fragments and plan execution based on information gathered from real-time

process monitoring.

50

.4 ‘ h--i.i§}mm25 «R‘s-

 

information ' E33060“
.3 ..

 

~ - 1-111! museum. in:-

Figure 4.3 - IPCA Task Structure

The synthesis task involves using process specifications and compiled knowledge to
construct a control plan specification. Process specifications contain high-level descrip-
tions of equipment and materials, sensor types, placement and actuator configurations.
The control plan specification, called the plan blueprint, contains information about
which control and planning methods are applicable and how they relate to the associated
policies and goals required for process management. Knowledge required to synthesize

the plan blueprint includes:

0 Knowledge of appropriate fabrication goals for the process.

0 Knowledge of the applicability and utility of various planning and
control methods in terms of ability to achieve the goals provided.

0 Knowledge about how the methods could be used together to form
policies and processing strategies.

0 Knowledge regarding the physical properties of the material to be pro-
cessed.

51

The assembly task involves the construction of a complete, executable process con-
trol plan which incorporate the methods, goals and policies set out in the plan blueprint.
Knowledge required in order to accomplish this task includes:

0 Knowledge of the applicability and utility of each planning and con-
trol method in terms of ability to meet the goals provided.

0 Declarative procedural (executable) representations for each planning
method and control method.

0 Knowledge of dependencies among planning and control methods.

The execution task involves applying the process control plan while gathering and
reacting to process information in real-time. This task requires:
0 Knowledge about how to process raw sensor data into meaningful
abstract characterizations of the process state.

0 Knowledge about how to select appropriate planning and control
methods at appropriate times during execution.

0 A facility for supporting goal revision, thereby allowing planning
methods to alter the initial goals provided in the plan blueprint.

IPCA goals are described in detail in chapter 5.

4.5 GT Approach

As introduced in chapter 2, the GT approach to knowledge-based systems forms a basis
for IPCA. The architecture, however, also adds to the set of GT tools by providing an
extension to the routine design/planning framework of DSPL. The architectural imple-
mentation of IPCA adds knowledge-based support tools for plan assembly and execution

to the existing DSPL system. The full functionality of the routine design mechanism pro-

52

vided by DSPL is still available, but is enhanced with mechanisms designed to support
plan assembly and execution.

A second GT-style problem-solving mechanism is used in the IPCA architecture to
support the knowledge-based selection of appropriate plan alternatives during plan execu-
tion. A sponsor-selector mechanism will provide this support, using matchgroup tables to
yield appropriateness measures under various conditions. Both mechanisms are discussed

in detail in the following sections.

4.5.1 DSPL and the Blueprint Generator

As briefly introduced in section 2.2, the design structures and plans language (DSPL) pro-
vides a problem-solving mechanism for routine design and planning tasks [Brown 89].
Specific knowledge representations and control strategies are included to facilitate this
task-specific problem-solving approach for routine design problems. A diagram of the

major DSPL constructs is given in figure 4.4 below.

53

 

 

Specialist A
r I l l
Rough Plan B Detailed|Plan C Detailed Plan D
Speciallist E Tarifk L * Tas'k M Task N
Rough Plan F Step 0 Step Q Step S

Constraint P Constraint R
Task G — Step H
Task I — Step J
Constraint K

Figure 4.4 - DSPL Design Agents

Using DSPL terminology, an overall design task is hierarchically decomposed into a
number of sub-tasks, each of which is supported by a specific design specialist. Each spe-
cialist includes a list of one or more plans from which to select in order to design a specific
sub-component for which it is responsible. Specialists may also contain a list of con-
straints which are used to determine whether a particular plan was successful. Plans can
contain references to specialists, so that any portion of the design task can be decomposed
into many levels of sub-tasks if required. Plans can also contain direct references to tasks,
possibly interspersed with constraints. Each DSPL task, in turn, is a design agent which
contains a list of steps and constraints. Steps are used to set the value of a single design
variable, while constraints are used to determine whether steps were successful.

There are two types of plans: rough-design plans and detailed design plans. The DSPL
system makes a first pass through the entire specialist hierarchy in an attempt to complete

a rough design. During this rough design phase, each specialist tries to apply it’s rough-

54

design plans, tasks and steps in an attempt to meet all design constraints. Referring to fig-
ure 4.4, Rough Plan B and Rough Plan F would be utilized during the rough design phase.
The rough design phase is used to determine whether a successful design is possible.
This task is accomplished by determinng whether the critical design variables can be set in
such a manner so as not to violate important design constraints. If the rough design phase
indicates that a successful design is not possible, the detailed design phase is never
attempted. Referring to figure 4.4, if Constraint K is not met, the rough design will fail
since there are no alternate rough design plans from which to choose. This two-phase
design approach generally saves time since clearly impossible designs are not be
attempted in detail. However, if the top—level specialist ultimately reports success (mean-
ing all rough design constraints have been satisfied) the detailed design phase is initiated.
Having established critical design values during the rough design phase, a detailed
design phase is subsequently attempted during which each specialist selects from a set of
detailed design plans. The detailed design phase proceeds just as the rough design phase
did, with each specialist attempting to use the detailed design plans (from most to least
suitable) to design each sub—component in turn. Referring to figure 4.4, either Detailed
Plan C or Detailed Plan D could make an initial attempt to design the component for
which Specialist A is responsible. If the more suitable of the two plans should fail, the
other plan is then invoked (unless the plan sponsor indicates that it should be ruled out). If
the detailed design pass ends in success, meaning that all design constraints are satisfied,
the design is complete. Note (also from figure 4.4) that Detailed Plan D will always suc-
ceed since there are no constraints within any subordinate agents. Also, Detailed Plan C

will always attempt both Task L and Task M, requiring that both Constraint P and Con-

55

straint R be satisfied for success. For additional detail on the problem-solving mechanism
of DSPL, see [Brown 89].

A simplified structural representation of the DSPL specialist, plan, task hierarchy used
for a portion of the prototype IPCA implementation is shown in figure 4.5. Many of the
design agents have been omitted for clarity and only a single fabrication technology is
shown. The top—level specialist, called the “fabrication specialist,” is responsible for
choosing an appropriate processing technology and directing the overall process control
plan design through the selection of appropriate rough and detailed design plans. In this
example, only the microwave discrete processing design plans are shown. Additional
plans which might be appropriate for other processing technologies (such as autoclave
processing or pultrusion) would appear under the top-level specialist.

Each specialist is responsible for some element of the design. For example, the “MD
process structure” specialist is responsible for decisions regarding the policies which will
be used to implement mode switching and power control. The tasks under this specialist
choose the strategies which are most appropriate. For example, static, scheduled mode
switching is appropriate if temperature sensors are unavailable but the mode interactions
with the part (heating patterns) are well—understood. Similarly, reactive, adaptive mode
switching is appropriate if little is known about the heating patterns, but temperature infor-

mation is available.

56

1. Fabrication technology specialist
2. MD rough plan

3. MD process structure

4. MI) rough plan structure

5. Set process policies

6. Fabrication technology

7. Mode control policy

8. Power control policy

9. MD detail plan

10. MD detail specialist
ll.Thermo-power-timing plan
12. Thermo specialist

13. Thermo plan

14. T-deg temp

O

p—t
q

«s» o co
MGMIBB

15. Set T-deg temp

16. Max temp

17. Min temp

18. Power specialist

19. Power plan

20. Set recommended power
21. Static processing

22. Schedules

23. Processing time

24. Mode switch schedule
25. Power schedule

26. Reactive processing
2‘7.Completion events

28. Time limits

Figure 4.5 - Structure of IPCA Design Specialists, Plans and Tasks

The specialist hierarchy used within the Blueprint Generator coordinates the problem-
solving activities needed to fill out the detail in the blueprint, including knowledge about

the utility of various control methods for maintaining tuned modes, appropriate power lev-

57

els and schedules for movement between processing phases. The overall processing strat-
egy is selected by the fabrication specialist and subsequently influenced by information
about sensor availability, processing experience, initial material state, material type and
geometry. Other agents determine the EM field types which are best suited to the initial
material setting requirements and the criteria for making transitions between the process-
ing phases. Each design decision adds a key/value pair to the blueprint. For example, if
reactive, adaptive mode switching is chosen, the pair (ModeSwitchPolicy MWDiscrete-
ModeSwitchPolicyC) is added to the blueprint.

This naturally leads to the question of how knowledge about the appropriateness of
DSPL plans, tasks and steps is represented. This representational mechanism, called a

sponsor-selector system, is described in the next section.

4.5.2 Sponsor-Selector Mechanisms
IPCA uses a sponsor-selector mechanism to support knowledge about the applicability of
agents under various conditions [Punch 95]. The mechanism is used for two distinct pur-
poses within the architecture:
0 Within the DSPL problem-solver to represent the appropriateness of
specialists, plans, tasks and steps.
0 Within the generated process control plan to represent the appropriate-
ness of fragments.
For both of these purposes, the mechanism functions identically. A sponsor—selector
mechanism consists of a hierarchy of selectors, sponsors and selection items. Each selec-

tor may be associated with any number of sponsors, which are grouped together under the

58

selector. Each sponsor, however, can represent only a single selection item. In multi-level
sponsor-selector hierarchies, selection items are sometimes other selectors.

Figure 4.6 shows an example sponsor-selector structure containing two levels. The
top-level selector chooses from a set of three sponsors, one of which has another selector
as its selection item. This second-level selector, in turn, chooses from a set of three spon-
sors. Each of these sponsors has one associated selection item.

Selectors effectively choose from among rating values returned by the sponsors to
determine which selection items are most appropriate for use under the current problem-
solving conditions. The selectors also contain a mechanism for deciding which sponsor is
to receive preference in the event of a tie. In IPCA, this mechanism is implemented by
ordering the sponsors by priority, but many other implementations are possible, including

breaking ties randomly.

@ sponsor .+ ma:- sponsor . Im-
Sponsor Selector . Sponsor , m
sponsor . In:-

Figure 4.6 - Sponsor-Selector Structure

Each sponsor contains knowledge about the appropriateness of a single selection item.
Though it can be done in any appropriate manner, this implementation encodes knowledge

in a match-group (MG) table, as depicted in figure 4.7.

59

X<0 I Y>=0 I! z<1 il

 

true true true
=
true false *

# E
false true false
false "' *

L—————¥-————

l false false false

Figure 4.7 - Match Group Table Structure

lflllllf

 

 

 

 

 

 

 

 

An MG table consists of one row of column header expressions, followed by one or
more rows of Booleans and don’t-cares forming the body of the table. Every MG table
contains at least two columns. The right-most column always contains the result to be
returned if conditions cause the corresponding row to be matched. The set of possible val-
ues that these result values can assume is called a rating system, which will be discussed
in chapter 5.

When evaluating an MG table, the column header expression truth values are deter-
mined first (ignoring the last column, which is always labelled “Result”). Rows in the
table are examined to determine which row(s) match the values determined for the header

«*9,

expressions. Since don’t-cares may be used (indicated by a in figure 4.7), multiple ini-
tial row matches are possible. The most specific (fewest don’t-cares) matching row is
selected from this group and it’s result value is returned. For example, using the MG table
from figure 4.7, assume X=3, Y=1 and Z=2. The column-header expressions evaluate to

false, true, false. This pattern matches both rows four and five of the six-row table. Row

four, hOwever, is more specific because it contains fewer don’t-cares. Therefore, the result

60

for row four (the symbol “3”) is returned. Also, note that some values of X, Y and 2 will
cause no row to be matched. In such cases, a default result is returned.

Selection item characteristics are not restricted by the sponsor-selector architecture,
and in fact many strategies have been encoded through the use of this mechanism (e.g.,
DSPL plans [Brown 84], hypotheses [Punch 90], IPCA control methods [Decker 94],
etc.). In each case, the sponsors act as “local” measures of how appropriate their associ-
ated items are for achieving the current goals, while the selector takes the more “global”
view of choosing which of the fragments is the most appropriate under the given problem-

solving situation.

CHAPTER V

5.1 Architectural Overview

This chapter contains an architectural overview of IPCA. Implementation details are pro-
vided in appendix B.

The IPCA architecture exploits the benefits of the Generic Task methodology by inte-
grating established approaches to knowledge-based design and planning with real-time
plan execution monitoring. The plan generation component is based on the DSPL lan-
guage for routine design and planning [Brown 89]. The plan execution monitoring compo-
nent is based on another problem-solving mechanism termed a sponsor-selector system
[Punch 89].

Figure 5.1 shows the detailed system architecture in the context of controlling a micro-
wave composite curing process, the present laboratory system on which the IPCA imple-
mentation is being evaluated. Fabrication models are used by IPCA during a plan
generation step, wherein a blueprint is generated and a process control plan is subse-
quently assembled from fragments contained in a fragment library. The resulting pro—
cess control plan is then applied in real-time by a plan execution monitoring facility.
During this plan execution monitoring activity, the IPCA system interacts with the con-

trolled process, obtaining sensor inforrnation and producing control actions.

61

62

     

n

x_

g -> . Blueprint I -> i Plan Assembler ‘

em
Fragment Library

4
-
J

~. Control Method fifififitfiififnfio?
Fragments " aim“:
1w, ,

     

 

     
 

 

 

 

 

 

 

 

 

Figure 5.1 - System Architecture

Process control plan generation and execution, though integrated in the overall system
Operation, are activities supported by separate problem-solving modules, each with its par-
ticular knowledge representations and control strategies.

The Blueprint Generator creates a plan blueprint which specifies processing goals
and potentially applicable fragments to be used during fabrication. Every fabrication
l'Ilodel contains information about a specific process which is used to produce a unique
blLIeprint. Blueprints are specific to the process described in their corresponding fabrica-

tion model. Further detail is provided in section 5.2.1.

The Plan Assembler integrates the plan and control method fragments specified in the

63

blueprint and creates the final process control plan by retrieving from the fragment library
those methods called for in the blueprint. The Plan Assembler also creates sponsoring
mechanisms for the specified fragments, and integrates them into a seamless fabrication
plan structure. The Plan Assembler is described in detail in section 5.2.2.

The responsibilities of the Plan Execution Monitor are to gather sensor data in real-
time and to select appropriate fragment sequences based on that information. The Plan
Execution Monitor therefore consists of three interacting components: an observation unit
(for real-time data capture and analysis), an execution unit (for plan fragment selection
and execution) and a Control Interface Unit for data acquisition. The Observation Unit
within the plan execution monitor gathers sensor data using network interface services
provided by the Control Interface Unit. The Control Interface Unit communicates with an
External Controller through a network interface. A description of the Plan Execution Mon-
itor is given in section 5.3.

The External Controller is responsible for monitoring a collection of sensors which
measure various aspects of the material state during fabrication and for reporting these
values to the Control Interface Unit when requested. Currently, sensed parameters include
temperature, incident and reflected power and physical positioning information for the
motor-controlled cavity tuning mechanisms. The External Controller also communicates
control outputs directly to physical actuators. The roles of the External Controller are

described in section 5.4 and further detail is available in appendix D.

5.2 The Plan Generator

The Plan Generator is comprised of three interacting components. The Blueprint Gener-

ator is described first, followed by the Plan Assembler and the Fragment Librarian.

64

The information processing task of the plan generator is to produce process control plans
based on information about the process to be controlled (provided by fabrication models),
knowledge of various fabrication processes (contained within the blueprint generator) and

a library of pre-compiled fragments (contained within the fragment library).

5.2.1 The Blueprint Generator

The Blueprint Generator produces blueprints by utilizing knowledge incorporated in fab-
rication models and compiled domain-specific knowledge of composite material fabrica-
tion, thermodynamics and heat transfer. Blueprints (see table 5.1) consist of a set of
references to potentially applicable fragments and the conditions under which they can be
executed. As suggested by table 5.1, the blueprint is implemented as a data structure con-
taining a number of unordered key/value pairs.

The final process control plan may be completely reactive in situations where a full
complement of sensors is available but little or no experience has been gained in process-
ing similar structures or materials. Conversely, process control plans may also be com-
pletely static in cases where there are few or no sensors but previous processing
experience is available. Typically, however, process control plans contain plan fragments
of both types, and this information is reflected in the key/value pairs contained in each

blueprint.

65

 

 

 

 

 

 

 

 

Key Value Key Value
PartMass 3O Ply'I‘ype Crossply
Matrix DGEBPA Plies 12
Technology MicrowaveDiscrete ModeControlPolicy ReactiveAdaptive
PowerControlPol- Reactive Non- PhaseChangePolicy Event-TimeB ased
icy Adaptive
TargetTemp 155 TDegTemp 180
Modes 1-5 PreviousRuns 5
TargetGrad 5 HoldTime 45

 

 

 

 

 

 

Table 5.1 - Typical Blueprint Format

5.2.2 The Plan Assembler and Fragment Librarian

The Plan Assembler generates a sponsor for every fragment specified in the blueprint,
encoding in each sponsor the knowledge of when that fragment is appropriate. Each key/
value pair contained in the blueprint is examined and matched by the Plan Assembler so
that appropriate fragments are included in the final process control plan. However, note
that the resulting process control plan is not “flat” but consists of a hierarchy of fragments.
Thus the Plan Assembler arranges fragments into an appropriate hierarchical configura-
tion. The Plan Assembler generates a sponsor for each of the fragments retrieved by the
Fragment Librarian for inclusion in the process control plan, placing it into the process
control plan hierarchy under the appropriate selector. The resulting plan structure is in a
format utilizable by the knowledge-based selection system contained within the real-time
plan execution monitor. 9

All fragments are managed by a Fragment Librarian. The Fragment Librarian pro-

vides a mechanism for organizing plan and control method fragments into specific librar-

66

ies. The Fragment Librarian also provides library retrieval services to the Plan Assembler.
Other elements within the same fragment contain lists of various parameters used by the
fragment, dependencies on other fragments and methods for computing required parame-
ters. Additional information regarding fragments is given in section 5.2.3.

The Plan Assembler is designed to examine all the key/value pairs contained in a blue-
print and to request appropriate fragments from the Fragment Librarian. The Plan Assem-
bler combines the required fragments into a process control plan by sponsoring each
fragment and placing it within the process control plan hierarchy at an appropriate posi-
tion. The Plan Assembler utilizes information about dependencies between fragments to
ensure that all required fragments are included in the process control plan and that appro-
priate parameters are created and initialized.

It is important that all fragments referenced in the blueprint exist in the fragment
library at assembly time. In the event that the blueprint contains a key/value pair which is
not found in the fragment library, an error is generated and the assembly operation will
end in failure.

The typical (simplified) structure of a process control plan is shown in figure 5.2. The
process control plan consists of several structures including a root-level selector, some
number of fragment sponsors, and a fragment associated with each sponsor. Associated
with each group of sponsors is a single selector which can be used to resolve conflicts that
can arise when multiple fragments appear equally appropriate. The structure of the pro-
cess control plan reflects the nature of the control task in the composite material process-
ing domain. The control functions are decentralized, with, for example, power control and

mode switching responsibilities residing in different sponsors. This structure reflects both

67

the modularity of the IPCA fragment library and the nature of the controlled process.

          

lnitllizatlon Coasting Cooling

   
 
   
   

Mode Switching

     
   

A Selector B Selector

Method

Figure 5.2 - Typical Process Control Plan Structure

5.2.3 Fragments

Two types of fragments are contained in the fragment library: plan fragments and control
method fragments. The fragment types are identical except for their effects when exe-
cuted by the plan execution monitor. Plan fragments are distinguished by their direct
influence over processing goals, as detailed in section 5.2.5. When a plan fragment is exe-
cuted, processing goals may be directly modified as a result. Plan fragment execution

does not directly modify any other process parameters, such as outputs, although the mod-

68

ification of processing goals may subsequently influence particular control method frag—
ments to be executed, which might subsequently change control outputs. Control method
fragments directly control process outputs when executed but they do not modify goals.
The distinction between fragment types is made only for ontological convenience. There
are no differences in the way the fragments are stored, selected, sponsored or executed,
nor is there strict checking to ensure that an executing plan fragment does not modify pro-
cess outputs. Similarly, there is no guarantee that executing a particular control method
fragment will not indirectly modify processing goals. It is therefore possible to create frag-
ments which modify both processing goals and control outputs, although no such frag-
ments exist in the present implementation. This dichotomy was enforced to keep planning
and control methods distinct, which helps ensure greater generality and reusability.

There are few restrictions on the algorithms contained within fragments. The only
requirement is that fragments contain compilable SmallTalk source code. Thus, a control
method fragment could contain a fuzzy-logic-based control algorithm, another could con-
tain a complete neural network, a third could contain a PID control algorithm and so on.
Neither are there restrictions placed on combining fragments of many different types
within the same process control plan. Thus, some control method fragments could be
based on PD) algorithms while others could be based on simple setpoint control. Simi-
larly, there are few restrictions on plan fragments. Each plan fragment can be constructed
to act on processing goals using any available algorithm. Thus, a plan fragment could
contain a complete STRIPS planner implementation, a reactive planning algorithm, a sim-
ple heuristic and so on. This approach adds flexibility and ensures that process control

plans are not restricted to a single planning or control approach.

69

5.2.3.1 Fragment Library Content

Currently, the Fragment Librarian contains a variety of static, reactive and adaptive con-
trol and planning fragments which are used for microwave composites processing, includ-
ing single and multiple setpoint control methods for managing power, proportional control
methods, and a variety of qualitative state-based methods for mode selection, cure mainte-
nance and fabrication planning. The IPCA control method fragment content is summa-
rized in table 5.2.

The IPCA power control policies evolved as experience was gained from fabrication
experiments. The earlier algorithms such as single setpoint control were supplemented by
increasingly sophisticated versions for all but material heating applications, which still use
the single setpoint algorithm to ensure a preset power level. The various mode control
policies are used under different fabrication conditions. For example, plans for gathering
heating profiles for a new part utilize MDModeControlPolicyE, which provides a static,
timed schedule which cycles through each of the available modes. Standard reactive
mode switching, adaptive-reactive switching, timed static switching are all available and
their selection depends upon sensor availability, previous processing experience and other

factors.

7O

 

 

Library Name Fragment Name Notes
PowerControlPolicy MDPowerControlPolicyA Single setpoint (on/off)
power control used mainly
for heating trials

 

MDPowerControlPolicyB Multi-level setpoint con-
trol. Power sets back 10%
at each excursion.

 

MDPowerControlPolicyC Proportional power control

 

ModeControlPolicy MDModeControlPolicyA Static, timed mode switch-
ing - no probes required.

 

MDModeControlPolicyB Reactive mode switching
with hold time

 

MDModeControlPolicyC Reactive-adaptive mode
switching with bold time

MDModeControlPolicyB Reactive mode switching
using uniform energy strat-
egy, with hold time

MDModeControlPolicyB Used for heating trials -
cycles through all modes

 

 

 

 

 

EmergencyPolicy MDEmergencyPolicyA Shut down if any problems

 

 

 

Table 5.2 - Control Method Fragments in the lPCA Fragment Library

The IPCA fragment library also contains a number of useful planning methods for
microwave-based polymeric composite fabrication. The IPCA plan fragment content is

summarized in table 5.3.

l'i.

71

 

 

 

 

 

 

 

 

Library Name Fragment Name Notes
InitilizationPolicy MDInitilizationPolicyA MW discrete processing
Setup0001 Master (&Model-S) 3” square DGEBPA modes

for experimental series Fl
Setup0002 Master (&Mode1-4) 1” square DGEBPA modes
for experimental series F2
Setup0003 Master (&Mode1-4) 2” square DGEBPA modes
for experimental series F3
Setup0004 Master (&Model-3) 2” square DGEBPA com-
plimentary modes for
experimental series F4
FabricationTechnology MicrowaveDiscrete Fabrication planner for
fixed-time cure
PhaseChangePolicy MDPhaseChangePolicyA Maintains processing
goals (e.g. warmup com-
plete when MinProcess-
ingTemp reached)

 

 

 

 

 

Table 5.3 - Plan Fragments in the IPCA Fragment Library

SetupOOOI through Setup0004 contain fragments for a Master and associated mode
selection fragments, one per mode to be used during processing. The applicability of each
mode is included within the appropriate mode fragment. Since one fabrication technology
and material is currently supported by IPCA, there is one phase change policy (material
dependant) and one initialization policy (fabrication technology dependant) contained in

the plan fragment library.

5.2.3.2 Adaptive Fragments
IPCA contains a facility for supporting adaptive fragments. An access mechanism is pro-

vided to allow fragments to modify the contents of the fragment library during process

pr:

lhc'

72

control plan execution. The access mechanism is provided by a pair of messages to which
the fragment library will respond. One message provides query access to the matchgroup
tables within any fragment in the library. The second message allows matchgroup ele-
ments to be replaced with new values. Together, the messages provide read and write
execution-time access to the library contents.

An adaptive mode switching fragment has been developed and tested. This fragment
modifies the estimated applicability entries of mode fragments based on their performance
during execution. The MDModeSwitchPolicyB and MDModeSwitchPolicyC control
method fragments estimate the qualitative state of the sample and select a processing
mode which is most appropriate. Whenever a mode is selected (under the adaptive-reac-
tive mode switch policy, MDModeControlPolicyC) which causes an increase in the ther-
mal gradient, the rating for the mode in that qualitative state is weakened by the adaptive
mode switching algorithm. Similarly, if the gradient is reduced as a result of selecting a
particular mode, the rating for that mode is strengthened for the appropriate qualitative
state and mode. Through the use of this adaptive mode switching method, the applicabil-
ity of selected modes in various qualitative states are automatically refined during material
processing.

For detail regarding mode fragment generation, see appendix C. For details regarding
the adaptive mode switching fragment, see “MDModeControlPolicyC” in the lPCA frag-

ment library.

5.2.4 Process Control Plan Format
The process control plan takes the form of a pair of ASCII files created by the Blueprint

Generator. The plan file contains information about the process control plan structure,

73

match groups used for determining sponsor applicability and the SmallTalk code for all
the fragments included in the process control plan. The parameter file contains declara-
tions for the parameters used by the process control plan, including names, types, initial
values, mappings to physical hardware and methods for computing abstract parameter val-
ues. A sample process control plan and parameter file are given in appendix A.

One portion of a fragment from a generated process control plan is shown in textual
form in figure 5.3. A typical process control plan contains many fragments, most of which
have more elaborate control method algorithms than the one included this example. How-
ever, the example clearly shows the sponsor (emergencySponsor), the matchgroup table
(with column headers, wildcards and ratings), a default row match return value, and the

procedural body of the method.

"Sponsor for emergency conditions: applicable if part is too hot"
Sponsor: emergencySponsor.
Headers: ('TlSensor > ThermalDegradationTemp' 'TZSensor >
ThermalDegradationTemp' 'T3Sensor > ThermalDegradationTemp' 'T4Sensor >
ThermalDegradationTemp' 'PartTemp > ThermalDegradationTemp' 'Rating').
Contents: (

(true * " * "' 'APlus')

(* true "' * * 'APlus')

(* * true * " 'APlus')

(* * "' true * 'APlus')

(* * * * true 'APlus')

).

DefaultReturnValue: 'E‘.

Item: (method partTooHotMethod).

MethodName: partTooHotMethod.

Method: '

partTooHotMethod
Transcript show: "Thermal degradation limit exceeded".
PowerActuatorl := 0.

Figure 5.3 - Example Fragment from MDEmergencyPolicyA

is Sl
mat
is ii

the:

The

rual
to a
abli
but
para
talc

obla

inch
"For
Pen

Cling

5.2.5

74

The process control plan fragment shown switches off the microwave power source if
the sample temperature exceeds a preset limit. The value of ‘TherrnalDegradationTemp”
is set by the Blueprint Generator and included in the plan blueprint as a parameter. The
match group table indicates that the method is applicable (with a rating of “APlus” which
is the highest rating available) when the temperature of any part of the sample exceeds a
thermal degradation limit.

All four types of IPCA parameters are present in this simple control method fragment.
The parameter “TlSensor” is an input parameter. The values of all input parameters are
updated at the beginning of each execution cycle. Each input parameter is assigned a vir-
tual channel by the Blueprint Generator and Plan Assembler, which is eventually mapped
to a physical sensor. The parameter ‘ThennalDegradationTemp” is an example of a vari-
able parameter. These parameters have an initial value (set by the Blueprint Generator)
but the value does not change unless modified as the result of fragment execution. The
parameter “PartTemp” is an abstract parameter. Abstract parameters have their values re-
calculated before every execution cycle. Irnmediately after new input parameters are
obtained, the Observation Unit updates the values of all abstract parameters by invoking
methods included in the parameter file. SmallTalk code implementing these methods is
included in the parameter file by the Plan Assembler. Finally, the parameter
“PowerActuatorl” is an example of an output parameter. Output parameters, like input
parameters, are assigned to virtual channels. At the completion of each control cycle, the

current values of all output parameters are routed to the appropriate physical actuators. _

5.2.5 IPCA Goals

IPCA goals are defined and initialized by the Blueprint Generator and maintained by the

Cit
pa]
[in
by

m:

pat

dis

[BE
in

32

He
llal

Pia

Cre;

iles

75

various planning fragments found within the generated process control plans. IPCA goals
are not supported by any specialized framework within the architecture, but are simply
maintained by the Plan Execution Monitor in the same manner as all other process param-
eters. The Plan Execution Monitor provides a generic facility for maintaining all process
parameters, including goals. Therefore, there is no architectural or implementational dis-
tinction between goals and variable parameters. The distinction between goals (modified
by plan fragments) and variable parameters (modified by control method fragments) is
made for notational convenience.

Initial goals for processing are setup by the Blueprint Generator. For example, the
pair (ProcessingGoal InitialMaterialHeating) is added to the blueprint for all microwave
discrete processing by a particular step under the Thermo-power-timing plan (refer to fig-
ure 4.5). The pair (MinProcessingTemp 125) might also be added to a blueprint (this is
material dependant) by a step under the Thermo plan. Although both tuples are subse-
quently treated as variable parameters by IPCA, the first would generally be referred to as
a goal and the second as a setpoint.

When the resulting blueprint is subsequently assembled, the Plan Assembler creates a
variable parameter named ProcessingGoal and fills in an initial value of InitialMaterial-
Heating. Similarly, a variable parameter named MinProcessingTemp is created and is ini-
tialized to a value of 125. The plan fragments included in the assembled process control
plan will presumably modify the variable parameter ProcessingGoal during execution.

Before the assembled process control plan is executed, the Plan Execution Monitor
creates a variable parameters named ProcessingGoal and MinProcessingTemp and initial-

izes their values to InitialMaterialHeating and 125, respectively. During the execution of

76

the process control plan, a plan fragment is eventually selected when the average sample
temperature exceeds MinProcessingTemp. The purpose of the plan fragment is to watch
for the event which signals that the sample has reached a minimum processing tempera-
ture and to change ProcessingGoal to reflect this fact. When the selected plan fragment is
executed, it changes the value of ProcessingGoal to MaintainMinimalGradient, subse-

quently affecting the selection and operation of other plan and control method fragments.

5.3 The Plan Execution Monitor

The plan execution monitor is composed of three interacting components. The Execution
Unit described in section 5.3.1 selects the plan fragment sequence in real-time. The
Observation Unit described in section 5.3.2 processes sensed process information and
provides abstractions from sensor data gathered during real-time process monitoring. The
Control Interface Unit, which provides network access to the External Controller, is

described in section 5.3.3.

5.3.1 The Execution Unit
The Execution Unit applies the process control plan by selecting fragments for execution
based on data gathered by the Observation Unit. The Execution Unit must coordinate exe-
cution of the process control plan with sensing and control functions. The External Con-
troller relieves the execution unit of the need to supervise low-level control loops
involved with cavity tuning operations.

Upon initialization, the Execution Unit parses the selected process control plan and
instantiates the sponsor-selector mechanisms to be used during plan fragment selection.

The Execution Unit pre-compiles all the fragment code contained within the process con-

77

trol plan, so that fragments are ready to execute immediately upon selection.

Execution of the processes control plan is cyclic. At the beginning of each execution
cycle, the Execution Unit requests the latest sensed process parameters from the Observa-
tion Unit. The Execution Unit then evaluates all the sponsors of the root-level selector and
gathers their appropriateness measures for their associated fragments. The root-level
selector then chooses a winning fragment based on the sponsor values and, if necessary,
the selector priority list. The process control plan structure may be of considerable depth,
since the selected fragment may be another selector. The search through the process con-
trol plan proceeds best-first until a single plan or control method fragment is reached. The
winning fragment is then executed in a protected mode as a background process so that
any run-time errors which might result do not interfere with the execution of the overall
process control plan. The Execution Unit kills the background process if an error or time-
out occurs. The executing fragment may make adjustments to the controlled process
parameters or change the processing goals. In either case, the execution cycle repeats
immediately after the fragment finishes execution or a timeout or error occurs. Plan exe-

cution continues indefinitely until suspended by an executing fragment or by the operator.

5.3.2 The Observation Unit

The Observation Unit communicates directly with the Control Interface Unit to obtain
input parameter values. The Observation Unit also manages several other types of process
parameters and performs analysis during process control plan execution. Process parame-
ters are used in conjunction with the process control plan to effectively control the fabrica-
tion process through appropriate selection of sponsored plan and control method

fragments.

78

As introduced in section 5.2.4, there are four types of parameters maintained by the
Observation Unit: input, output, abstract and variable. Their relationships to the Observa-
tion Unit will now be explored in more detail.

Input parameters are those values sensed by the External Controller and subsequently
passed to the Observation Unit through the Control Interface Unit. Output parameters are
modified by control method fragments and are passed to the external control system
through the same network interface and Control Interface Unit services. Abstract param-
eters have their values calculated once per execution cycle according to computational
algorithms included in the process control plan parameter file. These values are not com-
municated to the Control Interface Unit or External Controller. Finally, variable parame-
ters are used to maintain all other values required for fragment execution. Typically, goals
and global variables are encoded as variable parameters and modified during fragment
execution. Variable parameters are used for many purposes, including plan goal mainte-
nance and providing global data structures for fragment use. A variable parameter can
contain any SmallTalk data type and is therefore a highly flexible and completely adapt-
able facility for persistent storage.

Before each execution cycle begins, the Observation Unit requests the most recent
sampled sensor data from the Control Interface Unit. The requested information is nor-
mally transmitted by the dedicated External Controller to the Control Interface Unit
through the services of a network interface. From these sensed parameters, the Observa-
tion Unit computes a set of abstract parameters by invoking specific computational algo-
rithms included in the process control plan. The specifications for the abstract parameters

required for a given process and the algorithms used to compute them are provided within

79

the process control plan parameter file. Some examples of typical abstract parameter com-
putations for the microwave fabrication process are elapsed time, average temperature and
thermal gradient. Abstract parameters are computed once per control cycle prior to initia-
tion of the selection process by the Plan Execution Monitor.

All sensed parameters are timestamped by the Observation Unit so that recency of data

can be determined during fragment execution.

5.3.3 The Control Interface Unit

The Control Interface Unit is used to provide a standard virtual interface between the Exe-
cution Unit, Observation Unit and the External Controller. All plan and control method
fragments are designed to address input and output parameters on virtual channels so that
they are not bound to the details of how physical equipment is configured or how I/O
channels are assigned. The Control Interface Unit maps virtual channel designations into
physical sensor and actuator channels needed to access the hardware.

As a secondary function, the Control Interface Unit provides data logging services that
allow process inputs and outputs to be captured for later playback or analysis. Also, the
Control Interface Unit provides switching between multiple physical hardware configura-
tions and the simulated control unit interface used for testing plan and control method

fragments. The simulated control unit is described in section 5.5 below.

5.4 External Controllers

An External Controller is used to provide facilities for connecting physical sensors, actu-
ators and processing equipment to the IPCA prototype. The architecture and implementa-

tion of external controllers are not dictated by IPCA, since the structure and duties of each

80

external controller are domain and process-specific. Many external controllers may be
used, potentially accommodating several different processes. Each external controller pro-
vides a conduit between the Control Interface and the processing hardware.

Although the internal design and implementation of external controllers are not speci-
fied by the RCA architecture, certain duties are expected of every external controller.
Each external controller is expected to monitor available physical sensors and to report
sensor values to the RCA Control Interface on a timely, periodic, synchronous basis.
Similarly, each external controller is expected to accept timely, periodic, synchronous
actuator outputs from the RCA Control Interface and to pass these values to the appropri-
ate physical actuators. These responsibilities are described in detail in appendix D, which
contains a protocol document detailing the mechanism of information exchange.

External controllers are generally designed to reduce the computational load on the
RCA Plan Execution Monitor by providing local, low-level control operations and by
providing sensor preconditioning facilities. The external controller implemented for use
with the RCA prototype contains these features, using sensor preconditioning to reduce
the effective sensor sampling rate and PID control algorithms to facilitate mode tuning
operations.

External controllers typically execute on dedicated hardware and communicate with
RCA through the use of a shared memory facility. In the present implementation, the
external controller executes on a Macintosh IIciTM and the RCA prototype executes on a
Macintosh PowerBookT" 180. A network interface provides mutual access to a shared
memory area, in this case, a RAM disk. Messages are posted and retrieved from this

shared medium to facilitate communications, as shown in figure 5.4. Additional details

Wi
pri

act

sin

Sell

lflpu

81

regarding the format of exchanged information may be found in appendix D.

Retrieve
Retrieve /

;’.,"i"”‘l’& ",1”

 

 

 

 

 

 

 

 

 

   

 

 

 

 

T Cy‘fi 3"?» 5;;
' t“

.,..t.‘

 

 

Figure 5.4 - Communication with an External Controller

5.5 The Simulated Control Unit

RCA includes a simulated control unit which allows the execution monitor to function
without need for an external controller. The simulated control unit is useful for testing
process control plan reactions to a variety of exceptional conditions (such as sensor and
actuator failures) which would be difficult or impossible to otherwise demonstrate. The
simulated control unit is also useful for verifying the proper operation of process control
plans before they are connected to physical equipment and control systems in laboratory
settings.

The simulated control unit allows process control plans to be executed while process

inputs and outputs are routed normally through the Control Interface Unit. From the view-

82

point of the Execution and Observation Units, the Control Interface Unit is connected to a
generic source of process data. The simulated control unit features 16 virtual sensors and
8 virtual actuators which can be used to simulate up to 16 input channels and 8 output

channels.

For implementation details regarding the simulated control unit, see appendix B.

CHAPTER VI

6.1 Experimental Scope and Objectives

The functionality, utility, domain applicability and performance of the RCA prototype has
been evaluated through a collection of experiments. The experiments are of two types: C-
series and F-series. Both types of experiments are introduced and an overview of the
scope and objectives of each experimental series is presented. The criteria used to evalu-

ate experiments from each series is presented in the final section.

6.1.1 C-Series Experiments

The C-series experiments are designed to demonstrate the breadth and capability of the
RCA prototype Plan Generator and Plan Assembler by using the Simulated Control Unit
described in section 5.5. C-series experiments are conducted by presenting specific fabri-
cation models for blueprint generation and plan assembly. The resulting blueprint and/or
process control plan are examined for fragment content. Process control plans may be
exercised on the simulated control unit to demonstrate capabilities under simulated condi-
tions. C—series experiments do not involve real-world processing hardware, but are

designed to demonstrate the following capabilities:

83

84

0 The Blueprint Generator and Plan Assembler generate executable pro-
cess control plan structures for various fabrication scenarios (series
C1).

0 The range of process control plans produced is sufficiently broad,
including reasonable material coverage (series C2).

0 Process control plans respond effectively to potentially dangerous or

hazardous conditions (such as emergency handling) which would be
difficult to observe in the laboratory (series C3).

6.1.2 F-Series Experiments
F-series experiments are designed to verify the viability of RCA prototype and the appli-
cability of the planning and control method fragments to real-world laboratory-scale com-
posite structure fabrication. The F-series experiments are used to demonstrate the
following capabilities:
0 Microwave-based fabrication processes can be effectively managed by
IPCA-generated process control plans.

0 Quality composite structures can be rapidly fabricated under a variety
of conditions.

0 Rapid, controlled heating is possible using small, low-mass samples.
0 Thermal gradients during processing are lower than those associated
with other control techniques. Gradients are dependant on a number

of factors which are investigated in detail.

- IPCA effectively communicates with sensors, actuators and external
control systems.

0 The adaptive control methods are useful and effective. They improve
their performance with experience.
These capabilities are collectively demonstrated through the use of four experimental

series featuring the same composite material, but different structural geometries.

85

6.2 Experimental Preliminaries: F-Series

This section addresses experimental considerations for the F—series experiments. The
sample composition, geometry, layup, sensor placement, electromagnetic mode selection
and characterization and mechanical limitations of experimental fabricator are presented

in detail.

6.2.1 Sample Geometry and Composition

The composite material used in all F-series experiments was a carbon-fiber reinforced Di-
Glycidyl Ester of Bis-Phenol A (DGEBPA) prepreg. All samples were hand-laid 12-ply
unidirectional squares. Seven 3-inch square 12-ply samples were processed in experi-
ments Fl.1 through F 1.7 from series Fl. One 3-inch square 24-ply sample was processed
in experiment F 1.8. Eight l-inch square 12-ply samples of a second geometry were pro-
cessed in experiments F2.1 through F210. Three 2-inch square 12—ply samples were pro-
cessed in experiments F3.l through F3.3. Four 2-inch square 12-ply samples were
processed in experiments F4.l through F4.4. In each of the series Fl through F4, heating

data was gathered using additional samples.

6.2.2 Temperature Probe Placement
All F-series experiments were conducted using a Luxtronm four-channel fiber-optic ther-
mometry system to measure sample temperatures at four sites during processing. The
probe channels 1 through 4 are designated as T1, T2, T3 and T4 for all experimental
series.

Each thin-section composite sample was processed using a 90 degree fiber orientation,

indicating that the fibers in the unidirectional layups were perpendicular to the coupling

86

probe axis. The fiber-optic temperature probes were inserted into glass capillary tubes
designed to protect the tips from physical damage during processing. The protective tubes
were inserted through 1/ ” holes in the TeflonTM mold and rested directly on the upper
release film (see layup diagram, chapter 3). Probe placement is the same for all experi-
mental series, with T1 in the center of the layup, surrounded by T2, T3 and T4. A diagram

of the fiber-optic temperature probe placement is shown in figure 6.1.

 

Fiber direction

0
e e
T2 T3

0
T1

 

Coupling probe

Figure 6.1 - Fiber-Optic Temperature Probe Placement

6.2.3 Electromagnetic Modes

Each fabrication experiment was conducted using a particular set of electromagnetic
modes from which the Execution Monitor can ultimately select. Electromagnetic process-
ing modes were initially identified using a cavity-sweeping procedure described in [Fel-

lows 91]. The cavity-sweeping procedure involved loading the cavity with an appropriate

pli

Hf

311i

87

sample (1”, 2” or 3” square) and applying a low-power microwave-frequency signal to the
coupling probe. The cavity length was changed while the reflected power was monitored
using a dual-channel oscilloscope. Potential modes were identified based on minimized
reflected power, indicating effective “tuned” EM coupling. In an effort to further mini—
mize the reflected power, the probe depth was swept after each suitable cavity length was
located. The cavity length and probe depth of each candidate mode were noted for further
investigation.

Merely knowing the location of a mode does not give information about how the
resulting EM field will interact with the sample. Since some modes are associated with
EM field patterns which “surround” or “miss” the sample (particularly for very small sam-
ples), these modes were deemed as poor candidates for composite materials processing.
Even modes having EM patterns which effectively heat the sample generally distribute
energy into the material in an uneven fashion. Since this information is critical to making
intelligent decisions about the utility of each mode, a series of heating experiments was
conducted on each candidate mode using higher power.

The modes selected for the part geometries used in series Fl through F4 are summa-
rized in table 6.1. The table includes a heating pattern for each mode, which lists the
probe sites in order of decreasing EM coupling effectiveness. For example, mode 1 from
series Fl tends to heat T2 the most, followed by T1, T3 and T4. The type of mode is also
provided, and designations of TE (for transverse electric) and TM (for transverse mag-
netic) modes are used together with the number of standing half-wavelengths in the x,y
and 2 planes to form the complete mode designation. For more information about EM

modes, see [Herrington 61].

The

N

Seri

mar

a1e:

88

Five modes were used during the Fl-series fabrication runs, while four modes were

used in series F2 and F3. Series F4 used three modes.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Series Mzde 1:35;: Lc(cm) Lp(mm) Heating pattern Notes

F 1 1 TEm 13.5 21.5 T2, T1, T3, T4 Fast; Edge/Center
2 TE,” 15.6 21.0 T3, T4, T1, T2 High Pr; Edge
3 TEm 18.5 21.0 T2, T1, T4, T3 Slow
4 mm 21.1 21.0 T4, T1, T3, T2 Fast; Edge/Center
5 T5213 23.1 33.0 T3, T1, T4, T2 Even; slow

F2 1 TF4“ 8.7 15.7 T1, T2, T3, T4 Very fast; low Pr
2 TEon 12.2 24.4 T1, T3, T4, T2 Highly favors T1
3 TF4” 18.2 30.8 T1, T3, T4, T2 Even; med. rate
4 TM013 20.6 12.9 T3, T1, T4, T2 Even; slow rate

F3 1 TE”: 22.6 30.0 T2, T1, T3, T4 Ring; fast heating
2 TE”: 17.2 27.2 T1, T2, T3, T4 Even
3 TMmz 14.3 30.6 T1, T2, T3, T4 Fast; T1,T2 close
4 TE,“ 9.0 30.5 T2, T1, T4, T3 Even

F4 1 TM“ 27.4 28.0 T4, T1, T2, T3 Highly favors T4
2 TE”: 22.9 28.0 T2, T4, T1, T3 Ring
3 mm 21.7 28.0 T4, T2, T3, T1 T2, T4 close

 

 

 

 

 

 

 

Table 6.1 - Modes Used for Fabrication Series F1 - F4

 

A total of 15 candidate modes were identified for series F l, 13 for series F2, 8 for
series F3 and 9 for series F4. The majority of these candidate modes were unsuitable for
material processing either because they did not heat the sample effectively or were located

at extreme cavity and/or probe positions which were difficult or impossible to tune due to

89

physical clearance problems with the mold.

6.2.4 Heating Profiles: F-Series

Heating profiles were created for each candidate mode by applying power while simulta-
neously recording information from the fiber-optic temperature probes. Each profile
reveals the manner in which EM energy is transferred into the sample, and thus reveals
information about the utility of the mode during processing. Information from the result-
ing heating profiles was encoded into RCA either manually (the method used for experi-
mental series F1) or through the use of the RCA mode map generator (used for
experimental series F2 through F4).

The heating profiles for the experimental series F1 through F4 are given in figures 6.2
through 6.17. One set of profiles is given for each experimental series, since the profiles
are dependant on sample geometry. Each profile reveals the behavior of each fiber-optic
probe during the heating experiments for the sample in the respective mode. The profiles
are generally divergent, as most modes preferentially heat one or more areas of the sam-

ple.

6.2.4.1 Series F1

Five modes were selected for use with the 3” square samples in the Fl-series experiments.
The heating profiles for modes 1 through 5 are given below in figures 6.2 through 6.6. The
time scales for these heating profiles vary because the average heating rates in each mode

are not identical.

90

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

150 .
125 +
G 100 ~
g» ” T“ ' +T1
3 . . e 0 0 . . +T2
a ' ' +T3
g —o—T4
I-
25
0 I i 1 fi fi
0 1 2 a 4
Time (min)
Figure 6.2 - Heating Profile: F1 (Mode 1)
150 -
125 .
8 100 1
g: —l—T1
3 75 - +T2
g, ; , , . ” .‘ . . . +T3
e e . . . 0 . —'O—
'2 50 0' ‘—. l- — a T4
-
25
o l i T i i i T l
o 1 2 3 4 5 6 7

Time (min)

 

Figure 6.3 - Heating Profile: F1 (Mode 2)

 

91

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

5

4
Time (min)

 

150 v
125 a
G 100 «
g» _ . +T1
3 75- .w ° " +T2
% .r ..°‘°"‘° +T3
,2 so gr. ' ‘0‘”
25 .
O i x f i i l
o 1 2 3 4 5 6
Time (mln)
Figure 6.4 - Heating Profile: F1 (Mode 3)
100.
75 -
8
g} +T1
3 ‘ +T2
‘1 +r3
E, —0—T4
.—

 

 

 

 

Figure 6.5 - Heating Profile: F1 (Mode 4)

 

 

92

 

 

 

 

 

150 ~

125 ~
8 100
9 +T1
*1 +T3
E -o—T4
.—

 

 

 

 

 

 

Tlme (mln)

 

Figure 6.6 - Heating Profile: F1 (Mode 5)

6.2.4.2 Series F2

Four modes were selected for use with the 1” square samples in the F2-series experiments

The heating profiles for modes 1 through 4 are given below in figures 6.7 through 6.10.

 

93

 

 

 

 

 

 

 

 

 

 

 

 

 

Time (min)

 

 

 

150 —
-‘
125
5 100
pi; +T1
3 75 +T2
g- +T3
-o—T4
,2 so
25 1
o . , E
o 1 2 3
Time (min)
Figure 6.7 - Heating Profile: F2 (Mode 1)
150 —
125 1
A . o'e-e~._..§
o 100 ~ ,g.’-’°fl '
if 13"” +T1
3 75 1 ’ «-'°"" +T2
% x... 1.,o". +T3
25 <
o r 1 n
o 1 2

 

Figure 6.8 - Heating Profile: F2 (Mode 2)

 

 

 

 

94

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

150 1
a
125 s ” ’
r ’ i
A / / - e . . 0
0 100 1 / - . . '
. / .
3' . ° +T1
3 75 ~ /7. ° . +T2
0' ’0 . +T3
§ 4/ —O—T4
,_ 50
25 1
o T f
o 1 2
Time (min)
Figure 6.9 - Heating Profile: F2 (Mode 3)
150 1
5' ’ ’
125 1 , ’ ’
g/ ’
8 100 1 . . . . , . ° °
:1 ' . - - ° ' +T1
Q / ‘ . e
3 75 1 /, ég’. . +T2
g. ;e". +T3
—o—r4
,2 501
25 1
o 1 1
o 1 2

Time (min)

 

Figure 6.10 - Heating Profile: F2 (Mode 4)

 

95

6.2.4.3 Series F3
Four modes were selected for use with the 2” square samples in the F3-series experiments.

The heating profiles for modes 1 through 4 are given below in figures 6.11 through 6.14.

 

 

 

 

 

 

 

 

 

 

125 1
100 7
8
g: 75 ‘ +T1
3 .. .. .- ' +T2
a J 1.,— ~ ....OOO... T3
50 1 ‘:-;.r".. ...00e0
25
O T i 7 T l l
O 1 2 3 4 5
Time (mln)

 

Figure 6.11 - Heating Profile: F3 (Mode 1)

96

 

 

125 7

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

8
3’ +T1
3 ‘ ‘ +r2
=- +T3
E -O—T4
.—

25 1

O i r f r T i r r n—
o 1 2 3 4 5 6 7 8 9
Time (min)
Figure 6.12 - Heating Profile: F3 (Mode 2)

125 1

100 1
8
g? 75 ‘ A _. . , +T1
v e—e'eJ°W.’. . ° . .
g- 0W 0 ° . T3
E 50 . 0 T4
.—

25 1

 

o 1 T . . T

0 1 2 3 4 5
Time (min)

 

 

Figure 6.13 - Heating Profile: F3 (Mode 3)

 

 

97

 

 

 

 

 

 

 

 

 

 

150 1
125 1
3 100 1
g” - 1‘ ‘ ' +T1
375 . ........eeeee+T2
3 .°:—?-f-‘T"°°°° °
, - ., . .. . T4
,2 50- D
25 1
o F i l
o 1 2 3 4
Time (min)
Figure 6.14 - Heating Profile: F3 (Mode 4)
6.2.4.4 Series F4

Three modes were selected for use with the 2” square samples in the F4-series experi-

ments. The heating profiles for modes 1 through 3 are given below in figures 6.15 through
6.17.

 

98

 

 

 

 
   
   

 

      
 
 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

150 "
125 ‘
8 100 A u! t'u‘“ "" t.‘
9 +T1
3 75 1 . +T2
g- 6!! an.‘ '3’". .0. + T 3
. ".0 M'.'¢'e. .— ,,
.2 50 .0 ' 57- '.‘..',w¢lvv,"'ve ,qu ,‘ ‘ -°_T4
25
o i i i r r i i r
0 1 2 3 4 5 6 7
Time (min)
Figure 6.15 - Heating Profile: F4 (Mode 1)
150 1
125 ‘
5 100 1
9 -."°° -I—T1
'U o
v 75 '1 le' ' +T2
O. . ._.._O . . 0’04. ‘ ‘ - +1.3
.‘ _ ,.-...e O.-.-.' ‘ ‘ . -- 1
.. 4.1-64.1w» +1..
25 1
0 a . , I
0 1 2 3 4

Time (min)

 

Figure 6.16— Heating Profile: F4 (Mode 2)

 

 

 

 

99

 

 

 

 

 

 

 

 

 

 

 

 

150 1
125 1
8
g. +T1
3 +T2
e +T3
§ —o—T4
'-
25 1
O r ‘T T i
o 1 2 3 4 5
Time (min)
Figure 6.17 - Heating Profile: F4 (Mode 3)
6.2.5 Mode Maps

Information about the applicability of each selected processing mode is contained in a
series of RCA structures referred to as mode maps. These structures are specific to the
geometry, placement and composition of the composite layup, hence there are distinct
mode maps for each experimental series F 1 through F4. The maps were manually gener-
ated for the initial experimental series F1. For subsequent experimental series F2 through
F4, the maps were generated automatically by the RCA mode map generation utility
described in appendix C. .
Each mode map is used with an associated qualitative part state, which determines

how the map is used to determine the ideal processing mode for a sample which is in a

100

given qualitative state. For example, if a particular sample is cold in the center, some
modes would be more appropriate candidates to warm the center portion than others (it
would be desirable to choose a mode which heated the center in order to reduce the ther-
mal gradient across the sample.) The mode map provides this information to RCA, given
a suitable description of the qualitative state of the sample.

For the experimental series F1 through F3, 2 l6-valued qualitative part state was used.

The sixteen qualitative part states are shown along the y-axis of the mode maps for the F1

through F3 series. The part states result from 24 possible binary states of the 4 tempera-
ture probe inputs when compared with the average sample temperature. For example, part
state 2 indicates that T4 is cold while Tl-T3 are hot. The sixteen possible part states are
depicted in table 6.2 below.

For the experimental series F4, 2 12-valued qualitative part state was used. The twelve
qualitative part states are shown along the y-axis of the F4-series mode map. The part
states result from 4x3 possible combinations of the hottest and coldest of 4 temperature
probes. There are 4 choices for the hottest probe, then 3 remaining choices for the coldest
probe, yielding 4x3=12 qualitative part states. For example, part state 1 indicates that T1
is the hottest probe and T2 is the coldest. The twelve possible part states are given in table

6.3 below.

51

SU]

101

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

51:; T1 T2 r3 r4
1 Hot Hot Hot Hot
2 Hot Hot Hot Cold
3 Hot Hot Cold Hot
4 Hot Hot Cold Cold
5 Hot Cold Hot Hot
6 Hot Cold Hot Cold
7 Hot Cold Cold Hot
8 Hot Cold Cold Cold
9 Cold Hot Hot Hot
10 Cold Hot Hot Cold
1 1 Cold Hot Cold Hot
12 Cold Hot Cold Cold
13 Cold Cold Hot Hot
14 Cold Cold Hot Cold
15 Cold Cold Cold Hot
16 Cold Cold Cold Cold

 

 

 

 

 

 

Table 6.2 - Qualitative Part States for Series F1-F3

The qualitative part state model was altered for series F4 on the basis of experimental
results from the F1 through F3 series, which indicated that the two measured temperature
extremes should be weighted more heavily than the two more moderate temperature mea-

surements in determining the state of the material.

6.2.!

6.18
Wail

DCSS

102

 

 

 

 

 

 

 

 

 

 

 

 

 

Part State Hot probe Cold probe
1 T1 T2
2 T1 T3
3 T1 T4
4 T2 T1
5 T2 T3
6 T2 T4
7 T3 T1
8 T3 T2
9 T3 T4
10 T4 T1
1 1 T4 T2
12 T4 T3

 

 

 

 

 

Table 6.3 - Qualitative Part States for Series F4

6.2.5.1 Series F1

The mode maps for experimental series F1 are shown in figure 6.18 and 6.19. Figure
6.18 depicts the initial, hand-generated surface before mode adaptation occurred. The
available modes 1 through 5 are shown along the x-axis. The z-axis shows appropriate-
ness ratings as a function of mode and qualitative part state. Higher ratings along the z-
axis indicate greater applicability. For example, there is a peak in the map at the point (S4,
8) which indicates that mode 4 is highly useful when the sample is in qualitative part state

8.

103

 

 

I10-12
l8-10
I6-8
04-6
l2-4

Rating I 0 2

 

 

 

 

 

 

 

Figure 6.18 - Mode Map Surface: F1 (initial)

As the F1 experimental series proceeded, the mode map gradually changed due to the
influence of the adaptive mode switching methods employed. The adaptive methods
changed the applicability measures for the sample in response to actual processing experi-

ence. The resulting “tuned” map is shown in figure 6.19.

 

Rating

 

 

 

 

 

 

Figure 6.19 - Mode Map Surface: F1 (final)

6.2.5.2 Series F2

The mode maps for experimental series F2 are given in figures 6.20 and 6.21. The initial
mode map shown in figure 6.20 was generated with the RCA mode map generation utility
from the heating profiles previously presented. Mode adaptation was again used to “fine-
tune” the map across the set of F2 experiments. The final map (after adaptation) is shown

in figure 6.21.

105

 

 

l8-9
l7-8
I6-7
l5-6
I4-5
l3-4
Raflng D2-3
I1-2
I0-1

 

 

 

31 Mode

 

Figure 6.20 - Mode Map Surface: F2 (initial)

 

 

 

 

 

$1 Mode

 

Figure 6.21 - Mode Map Surface: F2 (final)

 

 

6.2.5.3 Series F3

The mode map for experimental series F3 is given in figure 6.22. The mode map was
automatically generated by the RCA mode map generation utility based on heating pro-
files previously presented. Mode adaptation was not used during series F3, so the map
remained static and did not adjust with processing experience. Just as with series F1 and

F2, the l6-state qualitative part model was used.

 

 

I 10-12
I8-10
ins-a
n4-6
l2-4
Io-2

 

 

 

 

 

 

 

Figure 6.22 - Mode Map Surface: F3

6.2.5.4 Series F4
The mode map for experimental series F4 is given in figure 6.23. Mode adaptation was
not used during experimental series F4, but the 12-value qualitative part state model was

used. The F4 mode map was automatically generated by the RCA mode map generation

107

utility described in appendix C. The mode map is based on the heating profiles for the 2”

square 12-p1y DGEBPA sample described in the previous section.

 

 

I1o-12

1° ls-1o

8 aa-a

6 Rating 04-6
7'. 4 I2-4
" 2 I0-2
o

 

 

 

   

Part state

82

Si

 

 

 

Figure 6.23 - Mode Map Surface: F4

6.2.6 Mechanical Limitations

Mechanical limitations prevent instantaneous switching between processing modes since
the cavity and probe tuning mechanisms must be physically moved to new positions. This
is unfortunate from the perspective of microwave-based materials processing, since lower
gradients could be achieved if the switching delay could be reduced or eliminated. The
limitations of the cavity and probe tuning mechanisms and their impact on the experimen-

tal series F1 through F4 are explained in this section.

108

The delay in switching between modes (the mode switch time) has several compo-
nents:
0 Delay in transmitting mode-change directives to the External Control-
ler

0 Delay in receiving and processing mode-change directives at the Exter-
nal Controller

0 Delay in moving the cavity-tuning mechanism to an approximate new
location

0 Delay in moving the probe-tuning mechanism to an approximate new
location

0 Delay in tuning the requested mode at new cavity and probe locations

The first two components of the mode switch time are almost negligible. The delay in
transmitting the mode-change directive to the External Controller depends on the end-to-
end speed of the network link established between RCA and the External Controller, but
is generally far less than 1 second. Similarly, the time required for the External Controller
to receive and process the mode-change directive is generally less than 1 second, unless
the External Controller is busy with another task at the time.

The significant components of the mode switch time result from the need to move the
cavity and probe tuning mechanisms. This portion of the overall delay is called the seek
time. The seek time between two modes is directly proportional to the distance that the
cavity and probe tuning mechanisms are required to travel. Also, mode seeking takes
longer when the cavity-tuning mechanism is moved upward (against the force of gravity)
than when it is moved downward because the stepper-motors have limited torque and tend

to “miss” occasional steps when heavily loaded.

109

The other significant portion of the overall mode switch time results from the delay in
tuning the precise mode location once the cavity and probe mechanisms are located within
the correct general vicinity. Since the dielectric constant of the composite sample changes
during curing (and the mode locations thus shift), the modes cannot be simply selected by
establishing particular static cavity and probe positions. The External Controller must
locate (lock) the mode within a range of cavity and probe positions. The lock time can
vary, depending on the precision of the mode lock specified. The precision is specified
within the External Controller as a fraction of incident power which must be achieved
before the tuning attempt is aborted. Specifying a more precise mode lock will result in a
longer average lock time.

The average mode switch times for the F1 through F3 series were measured and the

results are shown in figures 6.24 through 6.26.

 

From mode

   

 

 

Mode switch time (seconds)

 

Figure 6.24 - Average Mode Switch Times: F1

110

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

8
O
E
E
2
ll.
Mode switch time (seconds)
Figure 6.25 - Average Mode Switch Times: F2
1
3
O
E
E
2
IL

 

 

 

 

Mode switch time (seconds)

 

 

 

Figure 6.26 - Average Mode Switch Times: F3

In all F—series experiments, an artificial hold time was used to obtain a trade-off

between heating rate and thermal gradient. The hold time is defined as the minimum time

111

between successfully tuning a particular mode and permitting a request for a different
mode. If this time is small, the mechanism switches between modes often, thus maintain-
ing a small gradient. However, under these conditions the cavity is not frequently in a
tuned state, thereby directing little energy into useful heating. Conversely, if the hold time

is large, the mechanism is usually in a tuned state and directs more energy into the sample.

6.3 Evaluation Criteria

The C1 series is designed to demonstrate the capabilities of the RCA Blueprint Generator
and Plan Assembler implementations through the use of selected fabrication models. The
criteria used to evaluate this series include completeness and correctness of the generated
process control plans, including the appropriateness of the selected plan fragments.

Series C2 demonstrates the breadth of domain coverage in the Blueprint Generator and
fragment library. Blueprints generated under the C2 series are examined for fragment and
goal content, demonstrating that a range of materials and fabrication scenarios are cov-
ered. Evaluation includes demonstrating variant fragment content in the generated blue-
prints and showing that the referenced fragments are contained within the fragment
library.

Series C3 is used to demonstrate process control plan behavior under various simu-
lated conditions. Evaluation criteria include appropriateness of response to simulated con-
ditions and reaction time.

Series F1 is used to show the appropriateness of the prototype implementation by
effectively controlling a real-world process. This series demonstrates that control behav-
ior improves as fragments are suitably modified, either through adaptive methods or

through direct modification. Evaluation criteria for series F1 include subjective quality

112

measures on the fabricated structures and minimization of recorded thermal gradients for
the 3” square layup. Subjective quality measures are based on a visual analysis of the
completed structure. Any charring, incomplete curing, delamination, voids, warping or
other surface faults are noted in the analysis and used to evaluate effectiveness of the con-
trol approach. The criteria for indirectly evaluating the quality of the cure involves an
examination of the thermal profile. Heating rates, extent of thermal gradient and power
delivered to the structure are used as evaluation criteria. Improvement is measured by
confirming a reduction in thermal gradients during fabrication and a reduction in the
defects on the completed structures.

Series F2 demonstrates that a second structure can be fabricated and is used to demon-
strate the trade-off between heating rapidly and maintaining a minimal thermal gradient.
Due to the low mass of the 1” square part used in series F2, much greater heating rates can
be obtained than with the 3” square part used in series F1. The limitations of the mechan-
ical tuning mechanisms are explored and a case is presented for scaling to higher power
microwave sources and larger samples. The benefits of using a variable-frequency power
supply in place of manual mode tuning are presented.

Series F3 demonstrates coverage for a third structural geometry: the 12-ply 2” square.
The experiments in this series are designed to demonstrate rapid heating with slightly
larger parts. Heating rates are not as great as in series F2, but controllability is greater due
to the larger size of the sample in comparison with the cavity diameter.

Series F4 shows how complimentary mode selection and mode-switching policy influ-
ence the quality of processing. The policy used for these runs involves selecting modes

based only on the two most extreme (hottest and coldest) sample temperatures. Unfortu-

113

nately, the mode sets used in the F-series experiments were not fully complimentary with
regard to their heating patterns. Ideally, one would prefer to utilize modes which compli-
ment each other to the largest possible degree in order to minimize thermal gradients. The
F4 series shows that a high-quality mode set can considerably reduce thermal gradients

during processing.

Sill
un.
be:
are

Sill

CHAPTER VII

7.1 Experimental Results and Discussion

The experimental results are summarized in this section. Results are presented using

tables and graphs. A brief discussion of results is presented with each experimental series.

7.2 C-Series

7.2.1 Series C1

The C1 series demonstrates that the RCA blueprint generator and fragment library has
sufficient breadth to generate process control plans for fabricating composite structures
under various conditions. These conditions include the generation of plans for gathering
heating profiles for candidate modes, fabricating composite materials without pre-gener-
ated mode maps, fixed-time curing and variations for high and low mass samples. An

summary of experimental series C1 is given in table 7.1.

114

115

 

 

 

 

 

 

 

 

 

 

 

Experiment Conditions Process control plan I fragment Fragment
content notes
C l .1 Normal plan, MDInitializationPolicyA Standard
DGEBPA, MDModeControlPolicyB Reactive
Carbon fiber, MDEmergencyPolicyA Standard
3” square, MDPowerControlPolicyB Multi-level
4 temp. probes, MDPhaseChangePolicyB Event-based
Balance gradi- MinProcessingTemp 120 deg. C
ent and heating MaxProcesssingTemp 140 deg. C
rate RecommendedProcessingPower 45 watts
HoldTime (before/after warmup) 60/30 sec.
C l .2 Heating profiles MDModeControlPolicyB Heating
DGEB PA ProcessingTime 50 min.
3” square ModeUsed 5
5 modes RecommendedProcessingPower 40 watts
C1.3 No mode maps MDModeControlPolicyC Reactive-
Adaptive
HoldTime 180 sec.
C1.4 Fixed-time MDPhaseChangePolicyA Time-based
Cl.5 Low mass RecommendedProcessingPower 30 watts
C1.6 High mass RecommendedProcessingPower 60 watts

 

 

Table 7.1 - Summary of Experimental Series C1

Experiment C l .l and C1.4 confirmed that RCA can generate complete process con-
trol plans. The plans were exercised against the simulated control unit to verify appropri-
ate fragments responses and timing. The C l .2 experiment demonstrated that RCA can
generate process control plans designed to automatically gather mode (heating profile)
data. This scenario was used to collect the mode map information for experimental series
F2 and F3. The C1.2 plan sequentially places the cavity into each mode from the candi-

date set, applies a preset power level (based on the mass of the sample), allows the mate-

116

rial to either heat to a minimum temperature (120 deg. C for DGEBPA-based samples) or
for ten minutes, and turns off the power to allow sample cooling. The procedure is
repeated until all modes have been tested.

Experiment C1.3 involves producing a process control plan for a sample with no avail-
able mode maps. A default set of maps is created, with all elements at an intermediate and
equal rating. A reactive-adaptive mode switch method is included which randomly
chooses modes (initially) and modifies the maps as processing proceeds.

Experiments C15 and C l .6 demonstrate that the blueprint generator can appropriately
scale recommended processing power with sample mass. The high-mass experiment C1.6
involved doubling the normal 30g. mass of a hypothetical 12-ply 3” DGEBPA sample to
60g by doubling the number of plies to 24. It was expected (and verified) that the Blue-
print Generator would modify the recommended processing power. Recommended pro-
cessing power (used after initial material heating is complete) was scaled upward from 45
to 60 watts during experiment C1.6. Similarly, the low-mass experiment C1.5 halved the
specified part mass to 15g. (6-ply) and the recommended processing power was scaled

downward from 45 to 30 watts.

7.2.2 Series C2

The C2 series demonstrates that the RCA Blueprint Generator and fragment library have
sufficient coverage to produce process control plans which are suitable for processing a
variety of sample sizes and compositions. The experiments C2.1 through C2.3 were con-
ducted before processing began in the laboratory for the F1 through F3 series. These
experiments show that the Blueprint Generator adjusts power levels in response to sample

size and mass. The recommended processing power for C1.2 was 45 watts, for C1.3 it

117

was 30 watts and for C2.3 it was 20 watts.

 

 

 

 

 

 

 

 

 

 

 

 

Experiment Conditions Fragment content Fragment notes
C2.l DGEBPA MinProcessingTemp 120
3” square TargetProcessingTemp 130
MaxProcessingTemp 140
ThermalDegradationTemp 160
RecommendedProcessingPower 45 watts
C2.2 DGEBPA RecommendedProcesingPower 40 watts
2” square
C2.3 DGEBPA RecommendedProcessingPower 35 watts
1” square
C2.4 PolyesterGlass MinProcessingTemp 180
3” square TargetProcessingTemp 190
MaxProcessingTemp 200
ThermalDegradationTemp 250
RecommendedProcessingPower 100 watts
C2.5 PolyesterGlass RecommendedProcessingPower 80 watts
2” square
C2.6 PolyesterGlass RecommendedProcessingPower 60 watts
1” square

 

Table 7.2 - Summary of Experimental Series 02

Experiments C2.4 through C2.6 show that the Blueprint Generator can set appropriate
processing and thermal degradation temperatures for various matrix materials. In C2.l
through C2.3, the target processing temperature for DGEBPA was set to 130 deg. C and
the thermal degradation temperature was set to 160 deg. C. The processing temperatures
for experiments C2.4 through C2.6 were set at 180 deg. C, while the power levels were

set to 100, 80 and 60 watts. The thermal degradation temperature was set to 250 deg. C.

118

7.2.3 Series C3
The C3 series demonstrates that RCA-generated process control plans can effectively
respond to a variety of unexpected inputs due to sensor failure, loss of control, mechanical

failure and thermal runaway.

 

 

 

 

 

 

 

Experiment Simulated RCA reaction
condition
C3.1 Overtemp: Shut off power next execution cycle; continue executing
TDeg temp. process control plan; return power to 80% after cooling
exceeded below desired processing temperature
C3.2 Stuck Lc Stop execution of process control plan next cycle; shut-
down signalling actuator or sensor failure
C3.3 Stuck Lp Stop execution of process control plan next cycle; shut-
down signalling actuator or sensor failure
C3.4 Low Pi Stop execution of process control plan next cycle; shut-
down assuming sensor or power supply failure
C3.5 High Pi Stop execution of process control plan next cycle; shut-
down assuming sensor or equipment failure
C3.6 High Pr Stop execution of process control plan next cycle; shut-
down assuming sensor or equipment failure

 

 

 

 

 

Table 7.3 - Summary of Experimental Series C3

During experiment C3.1, the simulated sample temperature was suddenly increased
beyond the thermal degradation limit. The emergency sponsor activated on the next exe-
cution cycle and shutdown the power completely.

Experiments C32 and C3.3 demonstrated that RCA-generated plans would not con-
tinue to apply control signals which could burn out control electronics or motors. The Lc

and Lp positioning potentiometer signals were frozen during a simulation. The control

119

signals called for a change of mode, but the mechanisms did not “move.” The emergency
sponsor was activated and the processing run was shut down.

For experiment C3.4 and C35, the Pi sensor reading was artificially maintained at 0 or
100, thus indicating sensor failure or power supply failure. In C3.5, the reflected power
sensor reading was maintained at 80% of incident power, indicating a dangerous safety
condition in which excessive power was being reflected back from the cavity. In both
experiments, the emergency sponsor was activated on the next control cycle, the equip-

ment was shut down and the execution of the plan was terminated.

7.3 F-Series
The F-series experiments were conducted in the laboratory, using the single-mode micro-
wave cavity already described. The results of selected experiments from the F-series are

presented in this section.

7.3.1 Series F1

Series F1 demonstrated that RCA can produce process control plans which effectively
control the fabrication of 3” 12-p1y unidirectional DGEBPA samples. The experiments
which formed the F1 series were the first real-world trials for the IPCA implementation.
The average thermal gradients were gradually reduced over time, due to the use of the
adaptive mode-switching methods. The standard deviation of the thermal gradient was
also reduced with processing experience. All the samples from this series were cured for
90 minutes (with the exception of F1.1). Processing run F 1.8 was performed on a 24-ply
sample. Sensor and actuator values were not recorded during processing run Fl.8, thus

statistics are not available for that particular experiment.

120

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Experiment (2:11:16) Egg :3" Grass,” 6:31;“ Gradsrddev
F1 .1 32 60 60.61 22.02 6.77 3.85

F l .2 90 60 155.96 76.27 51.78 16.01

F1.3 9O 60 150.01 58.64 42.90 11.37

F 1.4 90 60 135.53 58.44 28.78 8.42

F1.5 90 60 157.61 31.78 20.27 5.54

Fl .6 9O 90 152.22 69.92 51.17 6.59

F1.7 90 90 130.48 60.93 39.72 4.99

F1.8 90 90 n/a n/a n/a n/a

 

The early (figure 7.1) and late (figure 7.2) thermal profiles clearly show the reduced
thermal gradients during processing, mainly attributed to mode map adaptation. The aver-
age thermal gradient was reduced from over 51 degrees C in experiment F1.2 to under 21
degrees C in experiment F15. The gradients for experiments Fl.6 and F1.7 were some-
what greater due to an increase in the hold time from 60 to 90 seconds. The longer hold

time encouraged faster heating in the latter experiments, but increased the average and

Table 7.4 - Summary of Experimental Series F1

maximum gradients, and this is clearly reflected in the data of table 7.4.

Each F-series thermal profile depicts the temperatures at the four probe locations T1-
T4 in degrees Celsius (C), the incident power level Pi (measured in watts), and the current
processing mode. The mode number has been scaled by a factor of 10 to assist with Over-

all clarity and readability, so mode 1 is graphed with a value of 10, mode 2 with a value of

20, and mode 5 shows up as 50.

 

 

121

 

 

 

   

 

 

 

 

 

 

 

 

 

 

 

 

 

 

175 1
1501 ‘1 ~ u...‘
8 125 7
. 3 —T1
913100 : ______ _T2
35 ‘1—73
5,... ———Pi
'- 50 7 -—Mode
25
o I I I I 7 I I I I r I I I I I I I
0 5101520 253035 404550 556065 7075 8085
Time (min)
Figure 7.1 - Thermal Profile (F 1 .2)
175 1
150 1 1“ - .. -,
. v \ /
A 125 1 ‘r 4"
o ,3 1’ & —T1
333'; 100 1 . " -—T2
35 ,5» \fl -. . _."' ‘ A ‘“ _T3
E a {"7
.2
501
i
25

o .

 

 

 

 

0 5 101520253035 4045 505560 657075 8085

Time (min)

 

Figure 7.2 - Thermal Profile (F1.5)

 

 

122

Since reduced processing gradients result in higher quality composite structures with
superior mechanical properties, it is reasonable to state that the quality of the resultant
parts should have increased as the thermal gradients were minimized during subsequent
processing runs. Although material testing was not performed on the samples resulting
from the IPCA research, it can be said that lower thermal gradients yield higher-quality
composite structures. The question of how much the mechanical properties of the result-
ing materials improved is a question beyond the scope of this research.

Therefore, during the analysis of subsequent processing runs and in comparison with
previous results [Fellows 91], the thermal gradient will be taken as a quantitative indicator
of resultant part quality.

Figure 7 .3 depicts the parts resulting from experiments F1.2 through F1.8, left to right,
top to bottom. The part from F1 .2 is charred on the sides. This effect was reduced in F13
and HA, as the average gradients fell from 51.78 deg. C to 42.90 and 28.78 deg. C,
respectively. The least charring occurred in F1.5, with an average thermal gradient of just

20.27 deg. C. Note the uniform, dark surface and lack of discoloration in the F15 image.

123

 

Figure 7.3 - Surface Scans (F1.2- F1 .8)

The samples from Fl.6 and F1.7 are not charred, but they show some evidence of sur-
face irregularities, perhaps as a result of the more rapid material heating and slower mode-
switching response. Sample F1.8 also showed a uniform appearance and lack of charting,
demonstrating that variances in material thickness do not necessarily interfere with IPCA’s

ability to obtain a high-quality cure.

124

7.3.2 Series F2

Series F2 demonstrated that smaller composite structures can be heated to reaction tem-
perature rapidly (in comparison with the 3 inch squares used in experimental series F1)
and that there is a trade-off between heating quickly and heating evenly. One-inch square
12-ply unidirectional DGEBPA samples were used for all F2 series experiments. Average
gradients ranging from 4.76 to 37.3 degrees C were obtained. Heating rates from 0.90 to
45.97 degrees C per minute were also observed, which greatly exceeded the heating rates
obtained in the F1 series. Experiments F2.8 through F2.10 used a proportional power
control method while all other F2-series experiments used a fixed power as listed in table
7 .5.

A clear relationship between hold time, thermal gradient and heating rate may be con-
structed from the data in table 7.5, and the precise details of this trade-off will be explored
in section 7.4 concerning the quantitative analysis of experimental results.

The central qualitative conclusion drawn from the F2 series experiments is that longer
hold times create faster sample heating but larger thermal gradients. This is because
shorter hold times cause the cavity to spend more time in an detuned state (since it is
switching between modes more frequently) and that this mechanical thrashing behavior
causes the sample to absorb less (focused) microwave energy. If the hold time is very
short, IPCA is permitted to request mode changes very frequently, thus preventing the

cavity from remaining in a tuned state for a long period of time.

125

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Experiment :33 £131: 3):“: :3" Hizgflg 61:?8
(sec) (C/mm)

F2.] 8 60 20 126.95 12.68 37.30
F22 6.5 30 30 120.00 20.46 17.17
F2.3 8 15 30 67.81 0.90 4.76
F2.4 12 20 40 107.97 4.28 9.70
F2.5 1.5 30 50 120.29 45.97 22.16
F2.6 2.5 30 40 129.68 16.95 32.28
F2.7 8.5 20 50 131.03 11.61 33.06
F2.8 2 30 Prop 109.24 35.16 21.71
F2.9 7.5 30 Prop 82.24 1.63 12.09
F2.10 8.5 45 Prop 117.20 8.03 20.71

 

A demonstration of rapid heating during experiment F2.1 can be seen in figure 7.4.
The heating rate from experiment F2.1 was very good (compared to the F1 series results)
at a rate of 12.68 deg. C per minute, but the temperatures diverged to a 38.2 deg. C maxi-
mum gradient. The hold time was 60 seconds, allowing time for only 4 mode changes

during the processing run. The inability to switch between modes more rapidly had a pos-

Table 7.5 - Summary of Experimental Series F2

itive effect on the heating rate, but had a negative effect on the thermal gradient.

 

126

 

 

 

 

 

 

 

 

 

 

 

 

150 7
1254
8A —

. 8 100 T1
8’: T2
:1 —T4
5°- 50 - / -——Pi
"' ’ ——Mode

25
7‘7 ‘ ’ \J W
0 f , , r , a
0 1 2 3 4 5 6 7
Time (min)

 

Figure 7.4 - Thermal Profile (F2.1)

Conversely, during experiment F2.3 (shown in figure 7.5) the average gradient was
only 4.75 deg. C, but the average heating rate was less than 1 degree C per minute. The
hold time for F2.3 was 15 seconds. This experiment showed that switching between
modes very quickly leads to a very low average heating rate, especially when a low power
level (30 watts) was used.

Several other combinations of bold time and power level were explored in order to

provide data with which to characterize this time-gradient trade-off.

127

 

 

 

 

 

 

 

 

 

100 r
75 ‘
6A _
. 3 T1
3’: T2
is; ‘ ""—'"T4
3 \ —--Pl
0
"' N /\ / —Mode
25 r \ I T f
0 r r T T r r l
0 1 2 3 4 5 6 7 8

 

Time (min)

 

 

Figure 7.5 - Thermal Profile (F2.3)

A reasonable compromise between heating rate and gradient was realized in experi-
ment F2.4 (figure 7.6). The average gradient was only 9.7 deg C., yet the sample heated
effectively at a rate of 4.28 degrees C per minute. The hold time was 20 seconds and the
power used was 40 watts. 'IVvelve mode switch Operations occurred during the 12 minute
fabrication run, which produced both rapid material heating and low thermal gradient.

The F2 series demonstrated the trade-off between thermal gradient and hold time, and
it also demonstrated the feasibility of processing a second material geometry. Thus, the F2
series showed that IPCA could successfully process composite structures with different

geometries.

128

 

125‘

1001

751

(watts)

 

0|
0

Temp (deg. C)
Pl

 

 

 

 

25‘

 

 

 

O 1 2 3 4 5 6 7 8 9 1O 11
Time (min)

 

 

 

Figure 7.6 - Thermal Profile (F2.4)

The samples resulting from fabrication experiments F2.] through F2.8 are shown in
figure 7.7. The same sample was used for the three proportional power control experi-
ments F2.8 through F210, and that sample is shown in figure 7.7 as F2.8. This time-sav-
ing practice was made possible by the fact that short processing times were employed and
the experiments were carried out in succession, after allowing for the material to cool.

In contrast to the F1 series experiments, none of the samples from series F2 were
charred during processing. This result demonstrates a marked improvement in the effec-
tiveness of the power control fragment over the behavior observed during series F1. How-
ever, since the F2 series experiments were not intended to process the samples for
extended periods of time, the curing was incomplete. Thus, the lack of charring is also

partially explained by a shorter processing time.

129

 

Figure 7.7 - Surface Scans (F2.1 - F2.8)

7.3.3 Series F3

The F3 series demonstrates that a compromise can be achieved between the rapid but
somewhat uneven heating obtained in series F2 and the slow but uniform heating obtained
in series F1. The series also demonstrated that a third structural geometry could be pro-
cessed.

A series of 2” DGEBPA square 12-ply samples was used for series F3. The selection
of this planar geometry was driven by the availability of a mold and by heat transfer calcu-
lations which showed that a 2” square mass would heat effectively in the cavity.

The experimental summary for series F3 is shown in table 7.6. A hold time of 60 sec-

onds was used for experiments F31 and F32. The hold time selected for experiment F3.3

130

was 30 seconds. The average heating rate for the F3 series was 5.70 degrees C per minute,
while the average thermal gradient was 25.25 deg. C. As expected, the heating rates and
gradients were intermediate between those observed during the F l and F2-series fabrica-

tion experiments. A third structural geometry was successfully processed.

 

 

 

 

Experiment Time (min) Tmax (C) Gradmax (C) Gr adavg (C) Gradstddev
F3.] 19 58.42 45.55 22.72 13.86

F3.2 9 l 13.09 56.05 30.08 14.02

F33 11.5 124.23 37.29 22.96 7.73

 

 

 

 

 

 

 

 

Table 7.6 - Summary of Experimental Series F3

The thermal profile for experiment F3.2 is shown in figure 7.8, and was chosen to
illustrate a problem which prompted a change in the approaches to mode map generation
and mode selection used in series F4. Upon consultation with domain experts, it was
determined that the thermal behavior obtained in experimental series F1 through F3 dem-
onstrated little or no “crossing” behavior. More exactly, the thermal profiles tended to
confirm that one or more temperature probe locations consistently remained above the
average sample temperature, while others were always below average. This is undesir-
able, even if the thermal gradient is low, because a higher-quality cure would result from
an equal temperature—time integral for each temperature probe location.

Referring to experiment F32, for example (see figure 7.8), it can be easily observed
that T2 and T4 never reach a temperature at which reaction initiation can begin. For the
DGEBPA material used in this experiment, that temperature, called the “onset of reaction

temperature,” would be well-above 100 deg. C. Additionally, the temperature-time prod-

131

uct would ideally be equal throughout the sample for the duration of the processing run,

helping to ensure optimal mechanical properties in the resulting material. This practice

would also help ensure that samples were cured more uniformly.

These observations were used as a basis for modification to the mode-mapping and

mode-switching algorithms. The modified algorithms were subsequently used for experi-

mental series F4, and are described in the next section.

 

 

 

 

 

 

 

 

 

 

 

 

 

1501
125
8A —

. 8 100 T1
gha- T2
35 751 "'"""T3
fig. 501 -—Pi
" I \‘ ——Mode

25 r
o I I l I r
0 2 3 4 5 6 7 8
Time (min)
g

 

 

Figure 7.8 - Thermal Profile (F32)

The experimental series F1 through F3 utilized a mode-switching fragment designed

to strictly minimize the thermal gradient across the sample. Thus, the mode-switching

fragment was designed to select a new mode whenever the gradient exceeded a preset

threshold for the material and the difference between the temperature readings could be

reduced (subject, of course, to the hold time constraint).

132

The mode-switching fragment used for series F3 was modified for the F4 series exper-
iments. The F4 mode switching fragment was designed to consider only the highest and
lowest temperature probe readings in evaluating the qualitative part state. The modified
mode switching fragment then selects the processing mode which will minimize the ther-
mal gradient and also ensure maximum crossing behavior between the probe measure-
ments. This approach was to encourage equal temperature-time integrals throughout
material processing, but it also dictated a change in the manner in which mode maps are
created and used. These details are discussed in the next section describing experimental
series F4.

The scanned images of the resulting composite samples for the F3 series are shown in
figure 7.9. All samples were incompletely cured, due to the short processing times. No

charring was noted in any of the samples.

   

Figure 7.9 - Surface Scans (F3.1 - F3.3)

133

7.3.4 Series F4

The F4 series was designed to demonstrate that an alternate mode-switching policy could
improve resultant part quality and as an experiment to determine whether thermal gradi-
ents could be reduced further from the levels obtained in the F3 series experiments. An
alternate mode-switching strategy was employed. Unlike series F1 through F3, the mode
selection strategy used in series F4 estimates the qualitative part state based on only the
highest and lowest reported probe temperatures. This approach required the use of new
mode maps, described in section 6.2.5.

The experimental summary for series F4 is shown in table 7.7. A 60/30 second hold
time was used. The bold time was maintained at 60 seconds (to encourage rapid heating)
until the sample reached a minimum processing temperature. When the minimum pro-
cessing temperature was reached, the hold time was reduced to 30 seconds by the phase

change plan fragment included in the F4-series process control plans.

 

 

 

 

 

Experiment Time (min) '1"max (C) Gradmax (C) Gradavg (C) Gradstddev
F4.l 60 131.31 49.80 17.41 6.84
F4.2 60 136.82 38.52 16.30 7.16
F4.3 40 137.22 50.29 12.82 8.16
F4.4 40 138.67 33.20 9.47 4.50

 

 

 

 

 

 

 

 

Table 7.7 - Summary of Experimental Series F4

The fabrication results from the F4 series showed a significant improvement in terms
of both thermal gradients and heating rates over the series F1 through F3. For example,

the thermal profile for F4.1 is shown in figure 7 . 10. The temperature profiles for each tem-

134

perature probe show excellent crossing behavior, resulting in a nearly uniform average
temperature at each probe location.

Experimental results confirm that the modified F4 mode switching algorithm
decreased average thermal gradients in comparison with the gradients observed during
series F3. The average thermal gradient for experiment F4.4 was only 9.47 degrees, a
59% reduction from the smallest gradients obtained in the F3 series. All temperature
probes were maintained above the temperature required for the onset of reaction (the min-
imum processing temperature for the DGEBPA material), yet well below the thermal deg-
radation limit. Fabrication run F4.4 (shown in figure 7.11) gave the highest-quality results
obtained in this research.

In comparison with the results of previous research, the thermal gradients have been
reduced [Fellows 91] by approximately two-thirds. Quantitative comparisons with previ-

ous research results are provided in section 7.4.

 

135

 

 

 
 
  

 

 

 

 

 

 

 

 

 

 

150 -
125 ‘
a? 100 ‘ _T1 I
g._ 1 "—'""T4
,2 a 50 1 ““131
——Mode
1M4 7' 1 1 l I 'l
o U T I I I I I I I I I I I ‘ j
0 5 10 15 20 25 30 35 40 45 50 55 60
Time (mln)
Figure 7.10 - Thermal Profile (F4.1)
C’- 3 —T1
3’; ‘ T2
(EL """"T4
E ——Pi
,2 50 1
25 ‘
O I T f I I I I —r
0 5 10 15 20 25 30 35 40

Time (min)

 

 

 

 

 

Figure 7.11 - Thermal Profile (F4.4)

 

 

136

The samples from experimental series F4 are shown in figure 7.12. Each sample was
uniformly cured with no evidence of cham'ng or large voids. These samples represent the

highest-quality materials produced during any of the experimental fabrication runs.

Figure 7.12 - Surface Scans (F4.1 - F4.4)

7.4 Analysis

The C1 series demonstrated that IPCA is capable of producing a variety of plan types for
the fabrication and heating of composite materials. The capabilities of the Blueprint Gen-
erator and Plan Assembler were effectively demonstrated in this series.

The breadth of the Blueprint Generator and fragment library were demonstrated in the
C2 series. The series showed that two materials were supported (DGEBPA/Carbon and
Polyester/Glass) and that processing parameters were set appropriately for each material

type. The series also demonstrated that power levels were appropriately set for various

137

sample configurations (3” square, 2” square and 1” square unidirectional layups).

Series C3 demonstrated effective control under various hypothetical failure or hazard-
ous situations during processing. An appropriate action was taken in each of the C3
experiments to prevent damage to processing equipment and surroundings.

Experimental series F1 showed that IPCA can fabricate high-quality composite mate-
rials by effectively controlling the processing hardware. The performance improved with
processing experience, demonstrating the effectiveness of the adaptive mode-switching
fragments. Gradients were eventually brought below 22 degrees, which is a 26% reduc-
tion over previous automated control approaches exhibiting a 30 degree gradient [Fellows
91]. Although the F15 gradient was low, it could have been made lower with more effort.
As reflected in the reduced charring of the samples from F1.2 through F1.5, improving
fragment control algorithms (in this case, power control) has a direct payoff in resultant
cure quality. Additional refinements were made which eventually led to the F4 series
experiments and lower thermal gradients.

Series F2 demonstrated the trade-off between thermal gradient and average heating
rate. As shown in figure 7.13, there is a direct correlation between observed heating rate,

hold time and observed thermal gradient.

138

 

Value
A
O

   

 

lEg.—Htg. Rate (076$
03 IAvg. Grad (C)

u. ,5 El Hold time (sec)
Experiment “'

 

".
cu
LLN N

 

 

 

Figure 7.13 - Hold Time, Gradient and Heating Rate Relationships

The correlation between hold time and gradient can be seen from figure 7.13. As hold
time increases, so does the average gradient. Similarly, the average heating rate is directly
proportional to the hold time and inversely proportional to the average thermal gradient.
As hold time increases, so does the heating rate. Also evident from the figure is the reduc-
tion in thermal gradients with processing experience due to the action of the adaptive
mode-switching fragments.

The correlation coefficient between the average heating rates and average thermal gra-
dients was 0.78 across the entire set of F-series experiments, indicating a very strong cor-
relation between the two data series. The correlation coefficients between the selected

hold time and average observed thermal gradients were 0.62 across the entire F-series,

139

which also indicates a strong correlation between hold time and average observed thermal
gradients.

The F2 series demonstrates an inadequacy of the mechanical mode-switching appara-
tus. Since the average thermal gradient is directly proportional to the hold time, it is desir-
able to reduce the hold time to zero. As previously described, mechanical limitations
prohibit reducing the mechanical mode switching time to zero. These important results
point out the desirability of a variable frequency tuning approach free of mechanical limi-
tations. If the frequency of the microwave power supply could be varied under the control
of IPCA, the hold time could be reduced to essentially zero, thus affording the opportunity
to establish new lower limits on thermal gradients during processing. Modes would be
selected not by moving metal shorting plates and probes, but by shifting power supply fre-
quencies. This is a promising direction for future research in the microwave fabrication of
composite materials.

Series F3 demonstrated that IPCA can process parts of a third geometry (2” square).
The heating rates and average thermal gradients observed during the F3 series were inter-
mediate between those observed during the F1 (3” square) and F2 (1” square) series, thus
demonstrating that heating rate is dependant on part mass, given limited available pro-
cessing power. This result demonstrates the scalability of the fabrication technology in
terms of the ability to achieve a desired heating rate.

Discarding the median temperature probe measurements and basing the qualitative
part state only on the highest and lowest temperatures resulted in superior thermal gradi-
ents, compared with series F1 through F3. Once again, the quality of the IPCA control

method fragments were shown to have a direct effect on the quality of the resultant fabri-

140

cation run.

Series F4 demonstrated that thermal gradients can be reduced to very low levels when
high-quality complimentary modes are selected for processing and appropriate qualitative
part state models are used. The qualitative part state model used in the F4 series proved
very useful in reducing gradients below 10 degrees C. This level of performance repre-
sents approximately a 66% reduction over previous fabrication methods based on mode-
switching and comparable materials [Fellows 91], which produced gradients on the order
of 30 deg. C using two complimentary modes.

The original mode switching work [Fellows 91] was not based on motor—driven tuning
mechanisms, resulting a longer locking time since the actuators were moved manually.
Some of the factors contributing to the decrease in thermal gradients observed in the IPCA
research include the availability of an increased number of characterized modes from
which to select, a reliance on motor-driven actuators which facilitate more rapid mode
switching and the employment of more sophisticated adaptive mode-switching algo-
rithms.

The IPCA-supported fabrication process is also much easier to manage than previous
approaches, since IPCA automates the tasks of capturing heating profiles, mode map gen-
eration and fabrication control. This level of automation also tends to increase the avail-
ability of characterized processing modes at fabrication time, since thermal profiles and
mode maps do not have to be generated manually.

Collectively, the C-series and F-series experiments demonstrated the viability, applica-
bility and extensibility of the IPCA architecture. The experiments showed that IPCA can

effectively control processes during rapid material heating for a variety of part geometries

141

and can also minimize thermal gradients during fabrication through the selection of appro-

priate processing parameters and process control strategies.

CHAPTER VIII

8.1 Conclusions

The IPCA architecture has made contributions in two major research areas. In the area of
knowledge-based systems and task-specific architectures, the following contributions

have been identified:

0 A flexible, task-specific intelligent control architecture.

0 An extension to task-specific approaches to routine design & planning.

0 IPCA provides a unique, comprehensive planning and control archi-
tecture which combines many of the elements of design, process con-

trol and planning research.

0 IPCA provides a unique automated design, assembly and execution
monitoring mechanism.

0 IPCA provides an extensible mechanism for combining planning and
control techniques into a single, unified approach.

0 IPCA extends the DSPL language for routine design with sophisti-
cated planning and control capabilities.

Vlfrthin the domain of composite materials manufacturing, the following contributions

have been made:

0 A structured framework for representing knowledge regarding com-
posite materials processing.

142

143

0 An improvement over conventional control techniques for managing
single-mode microwave polymeric composites processing.

0 A prototype for supporting the automated synthesis and execution of
process control plans.

0 A facility has been provided for managing knowledge regarding com-
posite material fabrication technologies.

0 Reactive planning and monitoring facilities have been provided to
enable the production of high-quality composite structures using
mode-switching techniques.

0 The limitations of mechanical mode switching have been investigated.
Variable-frequency tuning is proposed as a promising future research
area.

0 Considerable reduction of thermal gradients over previous automatic
control approaches.

0 Demonstration of feasibility and applicability of the IPCA approach to
the control of high-speed polymer processing.

The IPCA architecture provides a framework for capturing knowledge about compos-
ite materials fabrication technologies. The Blueprint Generator contains knowledge about
many processes and their potential applicability. The system can help to choose appropri-
ate fabrication technologies given a list of processing requirements. Additionally, the
IPCA implementation provides the first intelligent, automated real-time reactive monitor-
ing facility for the microwave fabrication of composite structures.

The limitations of the IPCA architecture include dependance upon the routine design
framework on which the Blueprint Generator is based and the sequential nature of plan
and control method selection. The IPCA architecture is limited to use in applications for
which the DSPL-based Blueprint Generator can produce a suitable plan blueprint. This

class of applications extends beyond composite material fabrication into the more general

144

area of process control. Additionally, the architecture may not be well-suited to control-
ling simultaneous processes with few interacting goals. The IPCA plan execution moni-
toring facility is designed for process control applications in which goals may interact or
conflict and a central, cyclic evaluation mechanism is more efficient than multiple control

plans executing in parallel.

8.2 Future Work

The IPCA architecture and prototype could be extended in the following areas:

0 Addition of IPCA support for MDSPL multiple blueprint generation.

0 Interactive replanning capability between the Plan Execution Monitor
and Plan Generator.

0 Investigation of variable-frequency tuning and the associated perfor-
mance limitations.

0 Incorporation of increasingly advanced control method and planning
fragments.

0 Enhancing the depth and breadth of materials and processing technol-
ogy coverage.

APPENDIX A
Glossary

Abstract parameter
One of four types of parameters which may be included in a process control plan.
Abstract parameters have their values updated by the Observation Unit on each execu-
tion cycle, immediately prior to Execution Unit fragment selection. See chapter 5.

Blueprint
The non-executable design artifact produced by the Blueprint Generator. The blue-
print contains a specification of the control method fragments, plan fragments and
goals which are deemed useful in controlling a specific process. Blueprints are used by
the Plan Assembler to create executable process control plans.

Blueprint Generator

A central component of the IPCA architecture. The Blueprint Generator is imple-
mented as a DSPL problem solver. The task of the Blueprint Generator is to produce
blueprints from fabrication models. See chapter 5.

Composite material

A material composed of two of more constituents which do not co-mingle at the
molecular level. Composite materials generally consist of a matrix and a reinforce-
ment. See appendix E.

Constraint

A DSPL construct, used to determine whether a portion of an overall design was suc-
cessful. Constraints may be interspersed between plans, tasks or steps in a DSPL
implementation to control problem-solving. See chapter 4.

Control Interface Unit

An important portion of the IPCA architecture used to coordinate communications
with external control systems, sensors and actuators. The Observation Unit relies on
services provided by the Control Interface Unit to acquire sensor data. The Plan Exe-

145

146

cution Unit also relies on the Control Interface Unit for directing output to appropriate
physical actuators. See chapter 5.

Control method fragment
One of two types of fragments included in the fragment library. Control method frag-
ments may directly influence output parameters but cannot alter processing goals.
Many such fragments are typically included in a single process control plan. See
chapters 2, 4 and 5.

Crossply
A particular type of layup in which the fibers in alternating layers are non-parallel. A
crossply layup may be used to encourage more isotropic material properties than those
which might be achieved through the use of a unidirectional layup. See chapter 3 and
appendix E.

Delamination

A generally undesirable failure state of a composite material in which the layers (plies)
in the composite peel apart. See chapter 3.

DGEBPA

Diglycidyl esther of bisphenol A, a polymeric composite matrix material. This mate-
rial was used for the fabrication experiments reported in chapter 7. See chapter 3 and
appendix E.

DSPL

Design structures and plans language, a GT-based approach to routine design and
planning problems. A DSPL-based tool was used to implement the Blueprint Genera-
tor. See chapter 4.

EM field

An alternating vector field consisting of electric and magnetic components. An EM
field is maintained in the cavity during material processing to deliver energy into the
material. See chapter 3, 6.

Execution Unit

The execution unit is responsible for coordinating fragment selection and execution
within the IPCA architecture. See chapter 5.

External Controller
A low-level feedback control system used to manage mode tuning and data gathering
operations. IPCA interfaces with this external system through services provided by
the Control Interface Unit. See chapter 3 and appendix D.

147

Fabrication model

Used as input to direct the blueprint generation process, the fabrication model contains
specific information regarding the material to be processed and the availability and
configuration of processing equipment. See chapter 4.

Fiber

One type of reinforcement commonly used to strengthen polymeric composites.
Often, fibers are composed of carbon, graphite or glass. See also, reinforcement. See
appendix E.

Fragment

A generalized, reusable control method or planning algorithm which can be assembled
with other fragments to create many distinct control strategies when combined with
appropriate goals specified by the Blueprint Generator. The Plan Assembler assembles
fragments into executable process control plans. See also plan fragment, control
method fragment. See chapter 4.

Fragment library
The library of fragments which is managed by the Fragment Librarian.

Fragment Librarian

An important portion of the IPCA architecture responsible for the storage, retrieval
and maintenance of plan and control method fragments. See chapter 5.

Generic Task approach

A task-specific approach to knowledge-based systems used as a basis for the IPCA
architecture. See chapters 2 and 4.

Gradient (thermal)

A temperature differential which is undesirable during composite material processing.
Large gradients result in uneven heating and reduced properties in the finished com-
posite structure. See chapter 3.

GT approach

See generic task approach.

Input parameter
One of four types of parameters managed by the Observation Unit. Input parameters
are used to represent sensor values obtained during real-time execution monitoring.
Input parameters are updated by the Observation Unit (through communication with
the Control Interface Unit) at the start of every execution cycle. See chapter 5.

148

IPCA

Intelligent Process Control Architecture, the intelligent control architecture which is
described in this document. See chapters 1, 2 and 4.

Knowledge-level

A distinct problem-solving level proposed by Alan Newell, advocating an implemen-
tation-independent description of complex knowledge-based problem-solving tasks.
See chapter 4.

Labview
A visual programming language and development environment by National Instru-
ments. Designed to support instrument control, data collection and general informa-
tion processing. Used to implement the external control system. See chapter 3.

Layup
A method for preparing composite structures in which layers of material are (usually)
hand-laid into a specific form (such as a mold) in preparation for processing. See
chapter 3, appendix E.

Matchgroup table
A table of two or more columns and one or more rows, used by sponsors to indicate
the utility of a sponsored item under various conditions. In the IPCA architecture,
matchgroup tables may contain wildcards in addition to the usual binary values. The
last column in every matchgroup table contains ratings which are returned with a cor-
responding row matches. See chapter 4.

Matrix

One of the constituents of a composite material. The function of the matrix is to trans-
fer loads to the reinforcement and to protect the reinforcement from environmental
damage. Matrices are frequently polymeric, metal or ceramic. See appendix E.

Microwave-based composite material processing
An alternate to conventional autoclave-based composite material fabrication which
has the potential to offer advantages in terms of speed, material properties, energy effi-
ciency and controllability. See chapter 3.

Mode
A tuned, stable EM standing wave pattern within the single-mode cavity. Each mode
typically beats a given composite structure in a unique manner. By changing modes,
thermal gradients may be effectively controlled. See chapter 3, 6.

Observation Unit

149

The portion of the IPCA architecture which is responsible for monitoring sensors and
updating parameters based on the resultant values. See chapter 5.

Output parameter
One of four parameters managed by the Observation Unit. The values of all output
parameters are conveyed by the Observation Unit to the appropriate actuators through
the services of the Control Interface Unit at the completion of each execution cycle.
See chapter 5.

Plan (DSPL)
A DSPL construct used to specify the tasks which carry out a particular portion of the
overall design. Each specialist contains a list of plans which may be used to carry out
a portion of the overall design. See chapter 4.

Plan Assembler

The portion of the IPCA architecture responsible for assembling selected fragments
into complete, executable process control plans. The Plan Assembler requests speci-
fied fragments from the Fragment Librarian, sponsors them and integrates them into a
hierarchical sponsor-selector structure. See chapter 5.

Plan Execution Monitor

The IPCA architectural component responsible for process control plan execution and
real-time process monitoring. The Plan Execution Monitor is comprised of the Execu-
tion Unit, Observation Unit and Control Interface Unit. See chapters 4 and 5.

Plan fragment

One of two types of fragments managed by the Fragment Librarian. Plan fragments
may modify processing goals, but they do not directly control actuators or other con-
trol outputs. See also, fragment. See chapters 2, 4 and 5.

Plan generator

The plan generation subsystem of the IPCA architecture, consisting of the Blueprint
Generator and Plan Assembler. See chapter 5.

Ply
A single layer in a multi-layered composite structure. See chapter 3 and appendix E.

Process control plan .
The executable artifact produced by the Plan Assembler. The process control plan
contains fragments, processing goals and sponsor-selector structures used by the Exe-

cution Unit to choose appropriate fragments for execution in real-time. See chapters 4
and 5.

150

Reactive planning
An approach to knowledge-based planning which attempts to address the problems
associated with plan execution in dynamic environments. Reactive planners usually
support interleaved, incremental planning and plan execution in an attempt to provide
rapid response to unexpected events. See chapter 2.

Reinforcement

One of two composite material constituents. Reinforcements give composite materials
strength. Common reinforcements include fibers, particulates and whiskers. See
appendix E.

Selector

One of the components of a sponsor-selector mechanism. Selectors are used to group
and select among related sponsors. See chapter 4.

Single-mode tuned cavity
A particular type of waveguide (cavity) in which a variety of standing EM field pat-
terns may be produced and maintained. The interior volume of the prototype cavity

(described in chapter 3) is variable, thus supporting the mode tuning process. See
chapter 3, 6.

SmallTalk

A high-level object-oriented programming environment used to implement the IPCA

prototype.

Specialist (DSPL)
A DSPL construct used to coordinate problem-solving for a portion of the overall
design task. Each specialist contains a set of plans from which to select to accomplish
the design process for a particular subcomponent and is charged with the responsibil-
ity for successfully designing that portion of the overall artifact. See chapter 4.

Sponsor-selector system

A mechanism used to support knowledge-based selection by using sponsors, selectors
and matchgroup tables. Used by IPCA within the Blueprint Generator and in the pro-
cess control plan. See chapter 4.

Sponsor
One of the components of a sponsor-selector mechanism. Sponsors rate the applica-
bility of associated selection items (typically fragments) through the use of match-
group tables. See also sponsor-selector system. See chapter 4.

Static planning

151

An approach to knowledge-based planning suited for non-dynamic environments. See
chapter 2.

Step (DSPL)

The lowest-level action in DSPL, involving setting a single design parameter to a par-
ticular value. See chapter 4.

Task (DSPL)

A collection of steps and interspersed constraints used to support the design of a par-
ticular subcomponent of an overall design. See chapter 4.

Thermoset composite

A composite with a thermosetting (not thermoplastic) matrix. Thermosetting materi-
als undergo crosslinking reactions which are generally exothermic and non-reversible.
See appendix E.

Variable parameter

One of four types of parameters supported by the Observation Unit. Variable parame-
ter values are calculated during each execution cycle from methods included in the
process control plan. See chapter 5.

Void
An undesirable “bubble” of trapped water vapor within a cured composite material
which generally serves to adversely effect mechanical properties. See appendix E.

APPENDIX B
IPCA User’s Manual

The IPCA software prototype is described in this section. An overview of the features of
each architectural component implementation is presented in a chronological progression,
beginning with launching the system and proceeding through blueprint generation, assem-
bly and execution.

The IPCA software implementation will run on any platform supported by ParcPlace
VisualWorksTM release 1.0 and Objectworks\SmalltalkTM release 4.1 [ParcPlace 92]. Sup-
ported operating systems include Apple Macintosh“, Microsoft Windows“, IBM RS/
6000"“, and UNIX”. Although the MacintoshTM platform was used to generate all figures
and examples, the use and appearance of the software is nearly identical across all plat-
forms. Throughout this section, references to visual interface components (windows, but-

tons, selection items, menus) are given in italic bold.

Launching IPCA

The IPCA prototype is launched from the Objectworks\SmalltalkTM environment by using
the Finder tool as depicted in figure B.l. To launch IPCA, the class AC Launcher is
selected and the Start option is selected. This creates an instance of the AC Launcher

class, shown in figure B.2. This visual component is called the IPCA Launcher:

152

153

 

 

rig

 

 

llierv Class Canvas

 

Class Canvas

—I§

_

   
 
 

 

ACTra nslator

Assembleroanvas 3,;
Ca lculatorExamp le
ClCa nvas
ClassifyingSpecialistUl
ClassifyingSpecialisN/ithn
CodingAssista nt
Co lorCo nsta nts ' ' "

57
lBrorvse|| Start | fldd [Remove [E]

Figure 8.1 - Starting IPCA using the VisualWorks Finder

    
 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

The IPCA launcher interface is used to activate the Blueprint Generator, Plan Assembler
and Plan Execution Monitor. The launcher serves simply as a tool to facilitate the creation
of an instance of a selected software component. Normally, the components are launched
in sequence to generate, assemble and execute process control plans. However, the
launcher allows flexibility since multiple components of a particular type may be launched
simultaneously and in any order. Closing the launcher interface causes the entire IPCA
software system to terminate, as the launcher is the top-most component of the IPCA user

interface.

154

 

=li§_— IPCH Launcher E—T—E
. 1 Intelligent Control Syst em

 

 

7 intelligent Systems Laboratory A
' ‘ Michigan State University
- [c] 1994

 

 

 

 

Figure 8.2 - Launcher Interface

Blueprint Generation

Selecting the Blueprint option causes the top-level Blueprint Generator Interface
shown in figure 33 to appear (the interface is titled RD Interface because some of the
visual components of IPCA are shared with other SmallTalk-based tools which preceded

it, namely, the DSPL-based Routine Designer).

 

 

 

 

 

Eli-L— Bl] interface ;=_'l:

PS Browse cases
ACPlan

 

 

 

  
 

current case

 

 

 

 

 

 

Figure 8.3 - Blueprint Generator Interface

155

The drop-down PS menu shown in figure B.4 reveals options for running the Blueprint

Generator.

 

 

 

 

 

 

 

Eli=lln interface Elia
_ Brotv'se  j j cases ”
run PS K
, inspect PS
saveian
save PS

 

 

 

Specify ranking method

 

rename PS
destrofigfiPS

 

 

 

Figure 8.4 - Blueprint Generator Problem-Solver (PS) Menu

The Browse menu allows access and modification to the structure of the DSPL prob-
lem-solver as shown in figure B.5. Options are included to support adding, deleting,

renaming and moving specialists, plans, tasks, steps and constraints.

156

 

EEE BB ‘lntefiaceglfia ‘

 

 

 

  
 
 

Browse Specialist Hierarchy
Browse flgent Hierarchy
Browse Database

 

 

 

 

 

 

 

 

 

Figure 8.5 - Blueprint Generator Browse Menu

The Blueprint Generator Cases menu, shown in figure 86, allows access to IPCA

fabrication models through the use of a sophisticated database interface. It is through

these services that the fabrication models are maintained. The fabrication models are

referred to as cases throughout the Blueprint Generator Interface.

 

 

 

 

 

PS Browse

 

ACPlan

 

current case

 

 

 

[5E an Interface +1973

 

 

new case

load case from disk
inspectcase

save current case

 

 

 

 

 

 

destroy current case

 

 

E|

 

Figure 8.6 - Blueprint Generator Cases Menu

The process of blueprint generation begins with selecting the run PS option from the

157

PS menu. Immediately after selecting the run PS option, the menu shown in figure B.7
appears. The routine design tool upon which the Blueprint Generator is based can support
single or multiple design [Kamel 94]. The prototype Blueprint Generator is designed to

work only with the Single design option, so this option should always be selected from the

resulting menu.

 

 

_|

Multiple designs
Single design 5 —
current case F

__ BB Interface ;@§

 

 
 

 

 

 

 

   

 

 

 

 

 

 

Figure B.7 - Specifying Design Type

Upon selecting the Single design option, the menu shown in figure B.8 appears.
From this menu, it is possible to specify the case (fabrication model) which will be used
for blueprint generation. If the desired fabrication model (Setup0002 in this example) is
displayed, the Run Current Case option may be selected, otherwise the Run New Case

option is selected and a fabrication model is specified.

158

 

 

 

 

EEE_ BB Interface —[§:
' Browse cases' ‘

 

Bun Current Case K
Bun New Case

 

 

 

 

 

 

 

 

 

Figure 8.8 - Specifying Design Cases

Once a fabrication model has been specified, the model is displayed as in figure B.9.
Values for any of the parameters in the model may be changed by selecting the parameter
of interest. This gives an opportunity to change the inputs to the Blueprint Generator
without altering the stored copy of the fabrication model. The blueprint generation pro-

cess is started by selecting the Proceed option.

159

 

  
 
 

Plannerinput
7"”:at'ionPermitted =
Lafitmclt'uatibr
”'7 yLengthSensor I.
Crosslinkhgent ' ‘
CurlngAgent _ .. p _ ~ ._.
DlelectrlcSensor .. , _ ‘ '11!) input."_‘, 9 »
IneldentPowerSensor ij 3 7 I
InvasweSensorsPermtted _ ‘. 'Yes'
HetnxClass ‘

   
 

 

 

 

'ThermosetPolymerlc

 

 

 

 

Figure 8.9 - Reviewing the Selected Fabrication Model

Some time will elapse while the Blueprint Generator completes the rough blueprint
design. Generally, the rough design takes less than a minute to complete on most plat-

forms. When the rough design pass is complete, the display shown in figure 8.10 is pre-

sented.

 

Plough Design completed successfully. Continue with Design?
.1 1
Figure 8.10 - Blueprint Generator Rough Design Complete

Selecting the yes option will cause the Blueprint Generator to continue with the

detailed design. Selecting no will cause the Blueprint Generator to abort the blueprint

160

generation process. After selecting yes, another brief pause will be observed as the Blue-
print Generator generates the detailed blueprint. This process takes about a minute on
most platforms. When the detailed blueprint design is complete, the display shown in fig-
ure B.11 will appear, indicating a successful completion of the blueprint generation pro-

cess.

 

BB Interface
Browse cases

ACPIan

 

 

 

  

current case

 

 

 

Design com-pleted successfully.

 

Figure 8.11 - Blueprint Generator Design Complete

Selecting 0K will cause the display in figure B. 12 to appear.

161

 

"’Pla'nil or output
mf’tatmnSelected

 

'IiDllodeAdaptati‘onPohcyA' , ‘

,. .. 1 _ ' 'liDllodeContr01PolicyC'
PhaseChangePollcy . ‘. * 'IIDPhaseChangePoli0yB'
PowerControlPollcy - « 'liDPowerControlPolicyBg'

     

 

     

Cflgfii It

 

 

 

Figure 8.12 - Reviewing Generated Blueprints

The blueprint can then be reviewed. The steps which caused a particular value to be
set may be viewed by selecting a particular key/value pair from the list. Selecting File out
causes the completed blueprint to be written to disk. The Proceed option may then be
used to return to the top-level Blueprint Generator Interface shown in figure B3. The
Blueprint Generator Interface may then be closed and the assemble option selected from

the IPCA launcher.

Assembling Process Control Plans

Selecting the assemble option causes the top-level user interface for the Plan Assembler to
appear. The Assemble option invokes the assembly process, and the Library option

invokes the Fragment Librarian interface shown in figure B. 14.

162

 

 

 

- —_—-——————_— m
— _—n——_ ————_————_——
— _———————-—_ ——————-—————_——_ —
..___—— ————-——-——_——_
I" ~.r:\"'w«~mfi-~ra ufi‘m‘W-{frneL«arfrvvrwr-fi “a; «av-W14" Ix“ ’=".9-Ae-fv" ~ \‘Wfl-K‘.‘<‘i~»2*‘u rs” Hut-".1 21.-nun ”-1-. . ' ' "“I ~ 'se-41* 1"1‘"
"‘Idle'i‘“ 1' ‘ ' * 1 ~ »“ .
#5414 115:»:51'5: ': ...'. ‘. .
‘ _
—

 

 

 

©‘f‘FfeiIifiitriIf3. 7 “ I 2. QJ‘OInii‘IiI hm]. 3.91313 73- j f; O M3999! 9mm] _ 019??)
'.“.._f".’i -j- \ .O.on.t.putpm:tce:-.. 0.;rariaue;nammetee i;

 

 

I_.
'0'

© «Eflfimqniflo’ e5 "O CWW' Plan 59"“? O PW eler 39 me : i:

 

 

 

 

 

 

 

 

It [migrating “r‘jfl

 

 

 

Figure 8.13 - Plan Assembler Interface

In figure B.l4, the element Plan-Source of the fragment Master in the library
ModeSetsSetup0002 is displayed. This element contains source code for the selected
fragment, which can be edited and placed into the fragment library. All fragment changes
are either incorporated or rejected when closing the Plan Fragment Librarian interface.
To accept all changes and make them permanent, close the Plan Fragment Librarian

interface. To abort changes and leave the fragment library untouched, choose the Abort

163

option.

 

 

 

 

 

 

“Sponsorforsllrnode changes: 60 second minimum if
experience processing part, else 180 sec on d"
"minim um time permode."
Sponsor:modeSwitchSponsor.

Headers: ['[ProcessingPhase = #processin g] &
[PowerActuatorl > 0]” '[ElapsedTimelnMode > 60]&
[PreviousRuns > 0]' '[ElapsedTimelnMode > 180]&
[PreviousRuns < l]' '[PreviousModesSelected > 0]' 'Rating').
Contents:[

   

 

 

gal I“ 5;}

 

 

 

Figure 8.14 - Fragment Librarian Interface

The Assemble option (shown in figure B.13) causes the current blueprint to be assem-
bled. After assembly, the Plan Assembler interface appears as in figure B.15, showing
information about the assembled process control plan, including lists of fragments and
parameters which were included in the plan. The resulting process control plan and
parameter source can be viewed in the lower scrollable display area, along with informa-

tion about any errors or warnings produced during plan assembly.

164

 

 

 

 

 

 

 

ESE Plan Bssembler i———_—__T-:——__—————:__E§

3W"? e ,w .-

. \ , ,

 

wear-

 

EmergencyP olicy®MD Em erg encyP olicyA
FabricationSetup®Setup0002

Fabricati onTec hn oIogY@MicrowaveDiscrete
InitializationPolicy®MDlnitializationPoIicyA
ModeControlPolicy®MDModeControlPolicyC

 

 

 

 

 

@fragments 3‘: ‘Olnputparameters OAhstract parameters
; ' a; ': ~‘ ‘ , *' "0 Output parameters ' ~C) 1I'aria'ble.‘pai'a'r'neters
"°” This plan file was automatically generated" {}
”m bythe AC Plan Ass em bler." 1,

 

'_'0 email librarian Plan-Source elem ent"

“Sp ona ortor powerc ontrol“
Sponsor. 'powerControI S p one or.
(.‘ar's' "‘I (a knee] 1.. A. on <.-.' res-“1' a" '. G

O Errorsn'iarninos . G) Control plan source . 0 Parameter source

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Figure 8.15 - Completed Plan Assembly

Plan Execution

After plan assembly is complete, the Plan Assembler interface is closed and the Execute
Option is selected from the IPCA Launcher, causing the Execution Monitor interface
shown in figure B.16 to appear. The Execution Unit, Observation Unit, and the Control

Interface Unit are simultaneously launched from the Execution Monitor interface. Select-

165

ing the Go! option causes the Execution Unit, Observation Unit and Control Interface Unit
to launch immediately using default process control plan files. By using the Open and

Launch options, other process control plans can be selected instead of the default (newest)

instances.

 

Egg [Hewtion Monitor E‘ 21%
EH Status ‘ A , I

rfinmw ‘9- nah-2%)." mwm “'5‘ :3; J6": raging—.3: mist?) gin-v.1...

 

3435' saying. “.1qu .
s,

552. ‘$“i\~—"vafi‘&w~ ”Jew“ H.33— Am. 2:1 .wn-‘a v.5- -1 :J ,

 

 

 

 

Figure 8.16 - Execution Monitor Interface

Figure B.17 shows the Execution Unit interface. The number of execution cycles are
displayed, along with the cumulative number of sponsor evaluations and average execu-
tion cycle period. The current (winning) sponsor and bid are displayed. The Step option

is used to single-step execution while the Resume and Suspend options are used to start

and stop execution, respectively.

166

 

 

 

   

   

 

553"” Eflecution Unit E;I§§
Active 3 onsors ~ Cycles .. 3
..-_;Izélmti$pnnsormmwmsrzzzae 1
' ' ‘ ’ Evaluatlorrs
Cycle penod
IlOBBsec

 

 

Process control plan
"m This plan fil e was automatically 9 en stated
'7’” bythe AC Plan Assembler

 
  
  

"D efault II branan Plan-Source elem ent

 

 

 

 

 

Figure B.17 - Execution Unit Interface

The SSMG option invokes the Sponsor-Selector Match Group browser shown in fig-
ure B.18, so that the structure of the process control plan may be examined. The current
selector is displayed, together with a list of sponsors grouped under the selector, the match
group and plan or control method fragment. The >> and << controls are used to browse

through different levels in the process control plan hierarchy.

167

 

 

 

 

 

 

|§fi= Sponsor-Selector Match Group E—‘E’EI

§§!§££9!”.M--,..- .13.... mm. W... .--... _. ..
Sponsors [ordered by pn unity]-
em erg encySpons or ‘

initializationSpons or ,

 
   

 

  
 

 

 

 

  

noOpSponsor'

 

9.9.1.991! "3.8!"! Fl”. -.., -. 1. .....-....._ - , .-
Column header _

Elgarefiissnt <100 ; . ' '.'
Hatch group table

 

 

 

 

 

 

 

 

true * {false ‘ 4}
false r 1 * . true ‘
<:=I 14>

 

Fragment
'Jp‘owérCo'iitrélléléth0'0 "Q ' 1
powerControlMethod
"Simple on} off control of heating"
[PartT emp < Min Proc essin 9T em p] ifTrue:[
‘ . '[CgureExtent < 20] ifTrue:[

 

 

 

 

 

 

 

Figure 8.18 - Sponsor-Selector Match Group (SSMG) Browser

Figure B. 19 shows the Observation Unit interface. Input, output, abstract and variable

parameters are displayed with their most recent values and corresponding timestamps.

168

Tim estam p

 

.F'i.i'!’ili’Ii:tt:§li ~‘I|)|.|¥i)9

 

 

 

Figure 8.19 - Observation Unit Interface

The content of the parameter file can be displayed by selecting the Parameter source

option, as shown in figure B.20.

169

  

 

  
 

Parameter Source

    
  

" ‘”ThI3“paIamétersourc Viewa‘tol'iiefi :I
£“e’n"%r?dbyt‘§ Efilgnfiss” In?“

   

 

 

0b ReflectedPowerSensorInput 48] [T 1 Sensorlnput 8 LIL

 

 

 

Figure 8.20 - Parameter Source Browser

The Control Interface Unit is shown in figure B2]. The upper scrollable area,
labeled Source, displays a list of sensed physical input channels and their most recent val-
ues, as reported by the Source for sampled information. The source of sensed process
information may be a Simulated control unit (described in the next section), the Live
External Controller, or a previously-recorded data Log. Similarly, the Sink area displays
the most recent output channel designators and values, which can be directed at the Simu-
lated control unit, the Live control system, or Both. The checkboxes labeled Graphics
and Log CI activity to: are used to display the part graphics interface shown in figure B23

and to log all Control Interface Unit activity to a specified file, respectively.

170

 

 

 

 

 

 

 

 

©5qu _j‘j"‘ -

 

Sink

  

 

 

 

 

fiesamjuffiré'r'o III. cam-’1'  

El Graphics [:1 Log CI activityto: “

 

 

 

 

Figure 8.21 - Control Interface Unit Implementation

The Simulated CU (control unit), shown in figure B22, is used to provide a limited
facility for testing process control plans and fragments without the need for physical con—
trol system hardware. The interface appears when Simulator is selected as the Source for
the Control Interface Unit. The controls may be used to simulate any sensor values

required to test the behavior of any process control plan of interest.

171

 

'fi;—:————-== Simulated to :;E

 

 

 

 

 

Figure B.22 -- Simulated Control Unit (CU) Interface

During a simulation, process inputs are obtained directly from the simulated control
unit. By selecting Simulator as the Sink for the Control Interface Unit, output parameter
values may be monitored as well, rather than being sent to the External Controller. Only
the virtual channels which are utilized by the process control plan are displayed. Also,
note that the simulated control unit is not a simulation of the process, but a tool intended

for debugging process control plans and for permitting plan execution in the absence of

172

processing hardware.

The Part model graphics display is shown in figure B.23. This display appears when-
ever the Graphics box on the Control Interface is checked. The display is intended to
graphically illustrate the current state of the controlled process. In the upper portion of the
display, the positions of the upper shorting plate and probe are shown, together with a bar-
graph indicator for selected power level and a pair of simulated analog meters for monitor-
ing incident and reflected power. In the lower portion of the display, the temperature of
the part is shown though the use of color. Absolute and relative temperatures are dis-

played, with blue indicating the coldest regions and red indicating the warmest.

 

 

BEE Part model graphics E—F’J—E

   

 

 

 

 

 

 

 

 

 

 

.»~-.' .;..».'.~.t.\'..~... ,-. '-.‘....... ..‘ ..~.' . ... _...,, - . .,.-. m... ;_,.'. .1. >....- J3...“ L........

Relative

 

Hbsolute

 

 

 

 

Figure 8.23 - Part Graphics Interface

APPENDIX C
IPCA Software Utilities

The IPCA implementation is supported by a set of software utilities. A log translator util-
ity is used to translate logged Control Interface Unit files into a format suitable for import-
ing by major mathematical, statistical analysis, visualization or spreadsheet software
packages. A mode-map generation utility is used to produce mode-map fragments from
heating profile data sets collected through mode characterization (heating) experiments.
This utility allows heating profiles to be automatically incorporated into the IPCA frag-

ment library. Both utilities are described in this section.

Log 'h‘anslator

The log translator utility is used to prepare logged fabrication data for analysis. Control
Interface Unit data is recorded during fabrication whenever the Log CI activity option is
selected on the Control Interface window. Each Control Interface Unit log file contains a
complete summary of the input and output parameters for a specific fabrication experi-
ment. The values of all input and output parameters are logged at the completion of each
execution cycle, giving a partial record of the fabrication run. The values of abstract and
variable parameters are not logged in the current implementation of the Control Interface

Unit.

173

174

The log translator utility extracts sensor information from a specified Control Interface
Unit log file and translates it into tab-delimited ASCII format suitable for import to a num-
ber of statistical analysis, mathematical visualization, spreadsheet or database packages
for further analysis. The graphs included in chapter 6 (showing temperature profiles dur-
ing specific fabrication experiments) were produced by importing translated Control Inter-
face Unit log files into Microsoft Excel”.

The Log Thanslator interface shown in figure C.1 appears when the Visual-
works\SmalltalkTM Finder is used to Start an instance of the AC Danslator class. Select-

ing the DansIate option causes the dialog in figure C.2 to appear.

 

I.o Translator

Intelligent Systems Lab
Michigan State University

WI

 

 

Figure 0.1 - Log Translator Interface

The name and location of the file to be translated must be supplied. The translator util-
ity may take a minute or more to translate the file into tab-delimited ASCII format. A dia-
log showing a fianslation Complete message then appears. Selecting 0K causes the
interface to return to the state shown in figure C.l. Another file may then be translated, or

the interface may be closed.

175

 

Log Translator

 

Intelligent Systems Lab
Michigan State University

 

 

 

LocationzName of file to be translated?

__'t_

Figure 0.2 - Log Translator File Specification

The translated file name is based on the original log file name, but includes the suffix
“.xl". In the example shown, a file named “SamplePlanJogl .xl” would be produced, leav-
ing the original file untouched. The translated file appears on the same volume and in the

same folder as the original.

Mode Map Generator

The mode map generator utility is used to produce mode-switching control fragments
from previously logged heating information (typically, log files produced by the Execution
Monitor). The utility accepts tab-delimited (translated) input with any number of columns
and rows. Each row typically contains information regarding temperature and other sen-
sor values at a particular moment. The series of rows thus reveals a history of sensor val-
ues across some timespan.

The mode map generator functions by performing an analysis of the heating character-
istics of each mode included in the logged heating information. A set of temperature dif-

ferentials is first computed for each mode included in the logged heating information.

176

Each differential is computed by subtracting the initial temperature at each probe (upon
first entering the mode) from the final temperature for the same probe. The differentials
are then sorted, producing an ordered list. For example, the list might be (T 1 T3 T4 T2)
for a particular mode, say, mode 1, indicating that T1 heated the most, followed by T3, T4
and T2.

The list is then used in a matrix multiplication operation involving a representation of
all possible qualitative part states to create the resulting mode 1 map. The multiplication
weights each of the matches between the qualitative states and the position of each
selected probe in the list. For example, if a part is in the qualitative state (T2 T4 T3 T1),
T2 is hottest and T1 is coolest. This would be an ideal match for the differential observed
above, since Tl would be heated most and T2 would be heated least. The matchgroup
table row for this mode and qualitative state would be filled in with a very high rating.
However, the rating in the state row (Tl T3 T4 T2) would be quite low, because the mode
heats T1 the most and T2 the least. Therefore, the row for this mode and qualitative state
would be filled in with a very low rating.

The user interface works almost identically to that of the log translator. The column
numbers corresponding to the logged temperature information must be provided as shown
in figure C.3. If four temperatures were monitored, T5-T8 must be set to zero. For exam-
ple, figure C.3 shows that information on T1 is to be taken from column 3 in the input, T2
is found in column 4 and so on. The fixed/variable radio buttons are used to specify
whether a map for a particular mode is to be generated (all the data belongs to a particular
“fixed” mode) or whether the data includes information gathered from multiple (variable)

modes. If “variable” is selected, the column which specifies the mode number must be

177

included (column 7 in this example). If “fixed” is selected, the mode to which all the tem-
perature data pertains must be specified instead of a column number, so that the appropri-

ate numbered map may be produced.

 

ENE ModeMap Generator Eli

Intelligent Systems Lab
Michigan State University

Data Column
T1
T2
T3
T4
T5
T6
T7
T8

 

(9 Fixed [mode #1
0 Variable [col 1]

r——-‘I

H___._..__l

 

 

 

 

Figure 0.3 - Mode Map Generator Interface

After specifying data columns and mode type/column/number, the Generate button is
used to initiate the mode generation process. A file dialog box immediately appears and a
filename is specified for the data file. The data is read, processed and the resulting mode
maps are written to a file with the same name, in the same folder, but with a “.maps”
extension. The file contains the complete fragments for the appropriate mode(s). The

resulting mode map fragments may then be pasted into the fragment library.

APPENDIX D
External Controller and IPCA Interface Protocol

External Controller Detail

The External Controller is implemented in the National Instruments LabVIEWW develop-
ment environment. Responsibilities of the controller implementation include monitoring
and processing sensor information (temperature sensing, linear position sensing, incident
and reflected microwave power measurements) and actuator signals (power control and
mode control). Sensor information is pre-processed, filtered and summarized before it is
reported to IPCA. The External Controller also provides mode-tuning services. Any pro-
cessing modes requested by the IPCA system are tuned, maintained and locked. The con-
troller attempts to maintain a tuning lock on any requested mode by seeking a
predetermined probe and cavity shorting plate position and then monitoring and minimiz-
ing reflected power in a predetermined linear region of operation. The tuning lock is
accomplished and maintained through incremental probe and cavity length adjustments
designed to find the precise location of the mode and maintain it through changes in
dielectric properties of the composite. The External Controller interface is shown in figure

D.1.

178

179

'

 

6 Elle Edlt Operate Controls llllndours Tent

 

-mm--mm-“

Figure 0.1 - External Controller Interface

From figure D.2, it is clear that two computer systems are used for operating the cavity
and conuolling the fabrication process. One machine runs the intelligent control system
implementation while the external control system software executes on the second

machine. The two machines communicate through the use of a network cable.

180

 

 

 

 

 

Key:

1. Apple 14" color dbplay

2. Apple Mad“ PowerBook 180

3. Apple Loeal'l‘alk network cable

4. Apple Macintosh llct,

5. Apple 16" color “all!

6. Apple 14" color dlaplay

7. Vexta Z-pbaae stepper motor, 5.4 volta

8. Dymrnatlon SK-ohn potentlI-eter

9. Vexta 2-pbaae stepper motor, 5.4 volta

10. Dynamatlon SK-obln potentlolneaer

ll. Vesta Z-pbaaeatepper motor, 5.4 rob
12. Optboa 100 Watt Mlcrowave power aonrce
13. National Instruments NBIO-lo ham-hoe
14. Vexta PXZ Z-pbae stepper motor drivers

 

 

 

 

 

 

 

 

 

Figure 0.2 - Microwave Equipment and Configuration

The machine which is running the external control system software has a National
Instruments NBIO-lol“ I/O interface installed [National xx]. This interface provides a
number of analog and digital input and output channels which are used for monitoring
input power, reflected power and cavity and probe positions. The interface is also used to
control the sample loading, cavity tuning and probe tuning mechanisms and the micro-
wave power level.

A shared RAM disk is maintained on one of the machines. A RAM disk was chosen
because it was desirable to make shared access as fast as possible. The shared area is used
to pass messages between the two systems. The implementations on both machines post
and check for messages on the shared disk to facilitate communication between the Plan

Execution Monitor and the External Controller. This implementation has advantages of

181

simplicity, network independence and machine independence.

The power system wiring is shown in figure D3. The equipment on the left side of the
figure is used during fabrication runs. The output of the 100W, 2.45GHz fixed-frequency
power source is connected to a coaxial directional coupler which allows incident power
measurement (at ~20dB) while allowing the remaining power to flow directly through the
device unattenuated. Power then flows through a microwave circulator, a solenoid-actu-
ated coaxial switch and into the microwave coupling probe. Power reflected back from the
cavity (not coupled into the sample) flows from the probe, back through the switch, into
the circulator and finally into a 50 Ohm coaxial power resistor where it is dissipated as

heat. The power flowing into the resistor is monitored by a second power sensor.

 

 

)

 

 

 

 

 

 

)

 

 

 

-a an»

 

Key:

1.0thoelnetrumentsioowwGthowersoorce 9.8Ppowermeter(renected)
LNardacoaxhldirectiomlcoupler,-20dBMB 10.1‘ramcoRFSPD'l‘ransmhelonswltch
3. HP power sensor, 35W average power 11. Western Microwave directional coupler
4. HP power meter (incident) 12. Narda crystal detector .0l-lSGhZ

5. Ferrite circulator 13. Tektronix oscilloscope

GNarda coaxialdirectional coupler, -30dB/0rili 14. flPeweepoociilator

7. Termalbe coaxial resistor, 50 ohm, 500W 15. Marconi “217-47

tHPpoweraemorJoomWaveragepower

Figure 0.3 - Experimental Microwave Power Delivery Wiring

The right-most portion of figure D.2 contains equipment which is used for cavity tun-

ing operations. This equipment is not used during processing. The equipment includes a

182

sweep oscillator used to generate low-power swept-frequency EM signals in a narrow
band around a center frequency of 2.4SGHz. This signal is used in conjunction with a
reflected signal measurement from the probe to drive an oscilloscope in X-Y mode. When
configured correctly, the oscilloscope shows reflected power as a function of frequency
and can be used with either a loaded or unloaded cavity to locate modes. When a mode is
located, the reflected power is minimized and this is readily observable on the oscillo-

scope. Additional detail on the experimental setup can be found in [Adegbite 95].

IPCA External Controller Interface Protocol

This section describes the communications interchange protocol between the IPCA proto-
type and the External Controller. The overall software system can be hosted on a variety of
computing platforms on one or more machines connected by means of any AppleW-sup-
ported network including LocalTalkT“, EthernetTM and AppleTalk Remote Access”. The
information interchange activities between the two systems are facilitated through the use
of standard AppleShareT“ software and a set of shared files. This approach eliminates the
need for platform-specific interfaces to Visualworks\Smalltalk“‘ and National Instruments
LabVIEWT“. All that is needed to implement the protocol is the ability to open, read, write

and delete ordinary ASCII files.

Example Platforms

The overall system may run on a variety of computing platforms on one or more
machines. The only restrictions are that IPCA and the External Controller both have

access to the shared files used for information interchange.

183

Some example platforms are given below:

0 No Macintosh“ systems connected with a LocalTalkmcable.
0 TWO Macintoshmsystems connected with Ethernet 7”.

0 Two MacintoshTM systems connected with AppleTalk Remote Access T“
(using a modem).

0 One Macintoshmand one PC connected with Ethernet T”.

0 One 68040-based or higher Macintosh“ with AIUXTM running IPCA
and the External Controller concurrently.

Shared Files

The protocol utilizes a set of files located on a shared disk as a sort of “shared memory”
for information interchange between the processes (the Plan Execution Monitor and the
External Controller are each considered a process). The files are in common ASCII (plain
text) format. A shared RAM disk is used to host the files for increased performance. All
files must reside within the same folder on the same shared disk.

Any process attempting to read from or write to one of the shared files should observe
certain restrictions to ensure that no process makes it impossible for another process to
gain access to the data. In general, processes must not monopolize a shared file by holding
it open for reading or writing for more than 1 second. After a successful attempt to read/
write, the process should not immediately attempt to read or write the file again without
waiting for a courtesy interval of at least 1 second. Read and write file operations should
conclude within 1 second from the time the file is first opened. A process attempting to

check on the existence of a file (polling) should not attempt to access the file more fre-

184

quently than twice per second.

Files “SourceData” and “SourceDataAvailable”

The SourceData file contains sampled (sensed) process data which needs to be sent from
the External Controller to the Plan Execution Monitor. The External Controller writes to
this file to supply data to the Plan Execution Monitor. When valid SourceData is available
for the Plan Execution Monitor, the External Controller writes a zero-length file named
SourceDataAvailable as a flag to indicate that the SourceData file is available to the Plan
Execution Monitor. The Plan Execution Monitor politely polls (no more often than every
0.5 second) for the existence of the SourceData file, so the External Controller may be
unable to open it for writing at all times on the first attempt. Similarly, the Plan Execution
Monitor will not always be able to count on opening the SourceDataAvailable file for
reading (on the first attempt) because the External Controller may have it locked while try-
ing to create it. Since there is some (minimal) contention for these shared files, both the
External Controller and the Plan Execution Monitor must exercise care not to lock each
other out for extended periods of time. Thus, both processes should observe the timing
constraints given in the section above.

Note: The Plan Execution Monitor deletes the SourceDataAvailable file as soon as the

SourceData file is read. The External Controller is then free to write to both files
upon detecting that the SourceDataAvailable file no longer exists.

File Format

Each line in the SourceData file consists of a quadruple of values in the following format:

<channel #> <value> <reservedl> <reserved2>

185

The <channel #> field is an integer in the range [1..128]. If the channel # is outside this
range, the remaining data on the line will be ignored. Thus, the protocol allows for up to
128 input channels. The channels can be mapped in any way to the physical sensors, with
or without gaps in the sequence. One tuple appears in the file for each sensed channel. The
<value> field contains a floating point number which represents the sensed value for the
channel. The <reservedl> and <reserved2> fields should be treated as strings. All fields
(including the reserved fields) must contain no spaces and are restricted to no more than
31 bytes in length.

One or more spaces are used to separate fields and a standard CR character is used to
terminate each line. Each line is restricted to be no more than 128 characters in length. The
entire file can contain no more than 128 lines. The maximum size for this file is therefore
16K bytes. Typically, the file will be much smaller.

The SourceData file could be accessed with the following ‘C’ statements:

FILE *inFile;

int channelN umber;

float value;

char resl [32], re32[32];

inFile = fopen(“SharedVolumeName:PathToSharedFile:SourceData”, “r”);

if (inFile != NULL)
fscanf(inFile, “%i %f %s %s”, &channelNumber, &value, res], re52);

Files “SinkData” and “SinkDataAvailable”

The SinkData file contains control actions to be sent from the Plan Execution MonitOr to
the External Controller. The External Controller reads this file and the Plan Execution

Monitor writes to it. When the Plan Execution Monitor finishes writing the file, it creates

186

the SinkDataAvailable file to indicate that the SinkData file is available. The External
Controller pends on the existence of the SinkDataAvailable file. When the SinkDataAvail-
able file is found to exist, the External Controller opens the SinkData file, reads its con-
tents, closes the file and immediately deletes the SinkDataAvailable file.

Each line in the SinkData file consists of a 4-tuple as above. The only difference is that
the <channel #> field is restricted to be an integer in the range [129..256]. Any line with a
<channel #> field value outside this range should be discarded.

Note: The External Controller must delete the SinkDataAvailable file as soon as the
SinkData file is read. The Plan Execution Monitor is then free to write to both files
upon detecting that the SinkDataAvailable file no longer exists.

Protocol Example

The interchange protocol is adequate for information interchange, but is very simple.

A key feature of the protocol is that the processes are restricted to advance in lock-
step. Therefore, if no actions are indicated or no new sensed data is available, mm;

WWII by the External Controller (to SourceData, in step 5
below) or by the Plan Execution Monitor (SinkData, in step 8 below):

1. A shared disk with an empty folder is created to host the four communications files.

2. The Plan Execution Monitor is started up, initializes itself and it waits for the Source-
DataAvailable file to appear.

3. The External Controller is started up and it initializes itself.

4. The External Controller senses data from the process (temperature, power, position,
etc.).

5. The External Controller writes the data to the SourceData file, closes it and the creates
SourceDataAvailable. If no data is available to the External Controller, SourceData simply
ends up having a length of zero bytes. The External Controller then pends on the existence
of SinkDataAvailable, performing whatever control actions are appropriate between poll
attempts. Polling rate is restricted to no more than 2 polls per second (the External Con-
troller should have plenty of other things to do in the several seconds it may take to get
SinkDataAvailable).

6. The Plan Execution Monitor (which has been waiting for SourceDataAvailable) sees the
SourceDataAvailable file, opens the SourceData file, reads it, closes it and then deletes
SourceDataAvailable.

7. The Plan Execution Monitor determines appropriate actions to take given the process-

187

ing plan/goals and the current sensed information.

8. The Plan Execution Monitor writes appropriate control actions to the SinkData file,
closes it and creates SinkDataAvailable. If no actions are indicated, the SinkData file will
have a size of zero bytes.

9 The External Controller sees SinkDataAvailable, opens SinkData, reads it, closes it and
then deletes SinkDataAvailable. The External Controller takes the actions indicated.

10. Goto 4.

Sample Data Files

Example SourceData and SinkData files are given below. Some indication of the variety of
data which could appear in the reserved fields is shown. In practice, SourceData and Sink-
Data files are unlikely to display this much variety of appearance in the reserved fields.

SourceData:

1 25.0 resl resZ

2 35.24 ThisIsOKToo 0.0

3 45.2 0 0

4 52.8 A B

14 21 OKToHaveGapsInChannelNumbers !
15 34.2 0 0

SinkData:

129 10.5 resl resZ

130 15 0.0 0

150 367.21343 O 0

0 0.0 ThisIsIgnoredBecause TheChannelNumberIsOutOfRange

APPENDIX E
Domain Background

Composite Materials

Engineers and scientists have sought to enhance, perfect and extend the basic structural
materials for centuries. History has recorded entire eras named in recognition of the mate-
rials which predominated, including The Stone Age, The Iron Age and The Bronze Age. A
continuous flow of structural materials were discovered and produced, each offering
unique advantages over previous materials in terms of strength, stiffness, weight, corro-
sion resistance and other physical properties. Of course, the research and development of
new materials is presently more important than ever before. Many applications await the
development of high-strength, low-weight materials and there is tremendous promise in
the research, application and development of modern advanced composites to serve these
increasingly demanding needs.

The meaning of the term composite has evolved over many centuries. The initial com-
mon use of the term simply referred to any combination of two or more dissimilar materi-
als with properties superior to either of the starting materials. Bone, wood and reinforced
concrete are some examples of composites under this initial definition. Generally, com-

posite materials retain the macroscopic properties and individuality of their constituents.

188

189

Metallic alloys, for example, are not considered to be composites because the starting
materials co-mingle at the molecular level to produce an isotropic material. Physical
properties in an isotropic material are insensitive to the direction or orientation of the
material. Composites are generally anisotropic materials (their physical properties differ
with respect to their spatial orientation) and are also largely inhomogeneous [Strong 89].

As research and development efforts produced an increasing number and variety of
composites, the need for a refinement in terminology became evident and a distinction was
made between naturally occurring composite structures (such as bone) and synthetic com-
posite materials (such as reinforced plastics). Generally, when one speaks of composites
in a modern context, one refers to synthetic (man-made) composite materials which incor-
porate a reinforcement in a continuous matrix. The role of the reinforcements in these
materials is to accept loads transferred to them by the matrix. The reinforcements give a
composite material its strength. The matrix serves a dual role -- load transfer to the rein-
forcement and protection and isolation of the reinforcement from its surrounding environ-
ment. Some composites are arranged as a laminar structure with layers of reinforcements
held in a matrix binder.

Included among the advantages of composite materials are high strength, design flexi-
bility, corrosion resistance and excellent physical properties such as tensile modulus and
stiffness. The major disadvantages to composite material use are material, design and fab-
rication costs. The wide variety of applications for composites includes aerospace com-
ponents and structures, aircraft components, automotive parts, marine and sports

equipment, circuit boards and artificial human joints.

190

Reinforcements

One may characterize composite materials along a number of dimensions such as the
type of reinforcement or matrix used or the physical construction of the composite. Three
classes of reinforcements are common -- fibers, particulates, and whiskers (which are
sometimes alternately described as a subclass of particulates). A partial taxonomy of rein-

forcements is presented in figure E. 1.

Reinforcements

Fibers Particulates Whiskers

Glass Carbon Organic Metal Ceramic

ll“
salass EGIass

can” Kevlar

Figure E.1 - Partial Taxonomy of Reinforcement Types

Fibrous reinforcements are the most common of the three reinforcement types and
have a single comparatively long axis which is capable of supporting a significant tensile
load. Fibers are typically near-circular in cross section and dimensionally large when
compared to particulates and whiskers. Fibrous reinforcements are available in wide vari-
ety of forms including continuous fibers, various types of knitted fabrics, weaves and
braids. Woven or braided materials need not be homogeneous with respect to fiber com-
position or diameter. Some of the most common fibers are E—Glass, S-Glass, carbon/
graphite and aramids. Carbon/graphite fibers have very high strength and tensile medulus,
and are commonly used as reinforcements for polymeric composites. The interphase

region between fiber and matrix is critical to the properties of the composite. It is in this

191

region that loads are transferred from the matrix to the fibers, and so the bond strength
between the fiber and matrix is of critical importance.

Particulate reinforcements have no preferred orientation and their shape is less impor-
tant than with fibrous reinforcements. The particles may be of a wide variety of shapes,
sizes and compositions including ceramic, metal and glass. Examples of particulate com-
posite materials include concrete and particle board.

Whiskers are single crystalline particles with an elongated shape and a preferred orien-
tation. Their physical properties are generally superior to particles but they must be care-
fully oriented within the composite material matrix, presenting an additional challenge

during fabrication.

Matrix Materials

Three major types of matrix materials are common -- polymers, metals and ceramics.

A partial taxonomic classification of matrix materials is presented in figure B.2.

Matrices

———"‘—"’T--~“‘-.

Ceramic

 

DGEBPA

Figure E.2 - Partial Taxonomy of Matrix Types

Fillers such as chalk, clay or glass are often added to matrix materials in an effort to

reduce costs. Fillers are distinguished from reinforcements because their primary function

192

is to reduce cost and not to provide load—carrying capabilities within the material. Poly-
meric matrix materials (resins) are the most common of the matrix types and may be fur-
ther classified as either thermoset or thermoplastic. There are tens of thousands of
patented polymers and polymer systems. Polymer and copolymer structures may be lin-
ear, random, blocked, branched or cross-linked [Strong 89].

Thermoset materials are cured under heat and pressure. As a thermoset polymeric
composite undergoes the curing process, a non reversible chemical crosslinking reaction

occurs as illustrated in figure 3.3, which is adapted from [Strong 89].

193

Reactants

o o

r r
+c-or—or-c-o-o-5-or2-o); + so + @cu-cq

Polyester (master
(peroxide) Crosslinldng agent (styrene)

Initiation

o o

n n
+c-Q‘c-o-Oi-ar'0t;
Bridging o o

+g’@—FH-g-O'CH‘,- 01-0);-

on

\— Bridge between polymer and crosslinking agent
Crosslinking

O

-cH-6-o-cnz-cHz-o);

 

Figure E.3 - Chemical Crosslinking Behavior of Thermoset Composite

During the initial step in the reaction, an initiator breaks the carbon double bond on the
polymer. This creates a bonding site for the bridging step. During bridging, the crosslink-
ing agent bonds to this site. In the process, another bonding site is created on the
crosslinking agent, which is then free to break another carbon double bond on another
polymer chain. The polymers thus become crosslinked after many applications of this
procedure. Crosslinking reactions are typically exothermic (they release heat) and the

resulting materials cannot be reformed by re-heating. The greater the number of

194

crosslinks formed in the material, the greater the viscosity and, ultimately, the stiffer the
material becomes.

For demanding applications, epoxy/amine curing systems offer better performance at
higher costs. Amine groups are commonly used as hardeners which encourage greater
crosslinking density during the cure by reacting with epoxy groups. Among the common
epoxy products is diglycidyl esther of bisphenol-A (DGEBPA).

In contrast, thermoplastic polymeric materials such as polyoxymethylene (POM) and
polymethyl methacrylate (PMMA) are processed by heating the solid polymer until a liq-
uid melt phase is reached in which the viscosity of the polymer is relatively low. In this
liquid melt phase, the polymer has a low average molecular weight. A reactive cure cycle
is not required for thermoplastic materials, which helps to reduce manufacturing complex-
ity and cost. As the long-chained polymers subsequently gel, viscosity increases with
average molecular weight of the repeated polymer groups and the average degree of poly-
merization. The polymer chains generally become quite long as the polymer solidifies,
and very large interconnected structures are formed which offer excellent resistance to
creep failure. The polymer chains may be physically intertwined, but are not crosslinked
to one another chemically through a crosslinking agent (as is the case with thermoset
materials). Molecular weights range from 5,000 to 100,000 for typical polymers. Cata-
lysts and initiators may be used to encourage mass polymerization. Some conventional
thermoplastic polymers are nylon, polycarbonate, polyethylene, polystyrene, polyester
and acrylics. Thermoplastic materials may be re-heated and cooled repeatedly, offering
the potential for convenient recycling. Fiberglass reinforcements are commonly used with

thermoplastic materials in low-cost applications. The structure of a typical high-perfor-

195

mance thermoplastic (PEEK) is shown in figure E.4.

mascot),

PolyEtherEtherKetone (PEEK)

Figure E.4 - Chemical Structure of High-Performance Thermoplastic

Thermoplastic materials offer advantages in their shorter manufacturing times, lower
equipment investment costs and higher impact and creep resistance, but have some infe-
rior structural properties when compared with thermoset materials. However, advances
have been made with thermoplastics, and many high-temperature and high-strength
advanced therrnoplastics are becoming available to help narrow this traditional perfor-
mance gap. Thermoset materials offer generally excellent properties in terms of strength
and stiffness, and may have higher temperature capabilities. Polyamides are used in very
high-temperature polymeric applications up to temperatures of approximately 370° C.

Metal matrices such as aluminum, titanium, magnesium or copper are generally used
for very high-strength, medium-temperature applications. Metal matrix composites offer
greater thermal stability and superior physical properties over those observed in pure met-
als [Harris 86]. Particles and whiskers are commonly used as reinforcements due to diffi-
culties in wetting-out fibrous reinforcements during fabrication. Metal matrix composites
are often isotropic materials like their metal counterparts.

Ceramic matrices are used primarily for very high-temperature applications. Gener-
ally, these are crystalline, non-organic, ionic-bonded, ordered structures which are capable

of withstanding very high temperatures (1100-2200° C). Ceramic matrix composites offer

196

high dimensional stability and high resistance to environmental attack, but are very brittle
and are prone to catastrophic failure modes (especially cracking). They are not generally

used as structural parts for these reasons.

Layup and Processing Technologies

A number of methods exist for the fabrication of composite materials, including manual
lay-up, prepregging, filament winding, pultrusion, matched-die molding and resin-transfer
molding [Srivatsan 91].

Manual lay-up involves layering mats within a mold by hand. One simple method is
to layer the mat and cover with resin. This method obviously has the potential for poor
wetting of the reinforcement (and therefore poor bonding) but offers a simple, low-cost
method for non-critical applications. Mats are also available which are pre-impregnated
with resin. These materials are called prepregs and they generally achieve superior wet-
out of the reinforcing fiber, usually with the help of a technique called vacuum-bagging.
Manual lay-up methods have several problem areas including lack of uniformity, voids,
and shrinkage due to uneven resin application. Only one side of the material is finished
(smooth) if matched dies are not utilized. Accordingly, male molds are used to yield
smooth inner surfaces (such as a hot tub) while female molds result in finished outer sur-
faces (such as a boat hull).

Filament winding is a technique in which a continuous tape or fiber is wound on a
mandrel (a rotating mold mounted on a lathe) to form the resulting part. Dry filament
winding uses a prepreg tow as its winding medium, while in wet filament winding the con-

tinuous fiber is dipped (impregnated) in a resin bath just prior to winding. With wet wind-

197

ing, the viscosity of the resin is quite important - dripping may occur if the resin becomes
too fluid, while poor wetting may result if the resin is too viscous.

Pultrusion is a continuous process involving impregnation, shaping and curing steps.
Pultrusion offers high throughput and minimal material waste, but part cross-sectional
area must be largely constant. Automated equipment for pulling, winding and cutting
operations is typically employed.

Matched-die molding may utilize sheet molding compound (SMC) or bulk molding
compound (BMC) together with male/female mold pairs. Material is placed in one mold
half and pressed into the second mold half. Pressure and heat are then maintained to cure
the material, which displays two finished surfaces.

Resin-transfer molding (RTM) may be used for greater manufacturing productivity. In
this technique, a mold is pro-loaded with the reinforcement and closed up. Resin is then
injected under pressure (actually, a vacuum is usually drawn within the mold to assist in
creating a pressure differential) and the part is cured. RTM has an advantage over manual
lay-up in terms of throughput, and the process may be automated more easily as well.
There are several critical factors, however, including design of the mold, viscosity of the
resin and potential movement of the reinforcement during injection, which can make RTM

impractical in many applications.

Layup
A number of methods exist for the fabrication of composite materials, including manual
lay-up, prepregging, filament winding, pultrusion, matched-die molding and resin-transfer

molding [Srivatsan 91].

198

Manual lay-up involves layering mats within a mold by hand. One simple method is
to layer the mat and cover with resin. This method obviously has the potential for poor
wetting of the reinforcement (and therefore poor bonding) but offers a simple, low-cost
method for non-critical applications. Mats are also available which are pre-impregnated
with resin. These materials are called prepregs and they generally achieve superior wet-
out of the reinforcing fiber, usually with the help of a technique called vacuum-bagging.
Manual lay-up methods have several problem areas including lack of uniformity, voids,
and shrinkage due to uneven resin application. Only one side of the material is finished
(smooth) if matched dies are not utilized. Accordingly, male molds are used to yield
smooth inner surfaces (such as a hot tub) while female molds result in finished outer sur-
faces (such as a boat hull).

Filament winding is a technique in which a continuous tape or fiber is wound on a
mandrel (a rotating mold mounted on a lathe) to form the resulting part. Dry filament
winding uses a prepreg tow as its winding medium, while in wet filament winding the con-
tinuous fiber is dipped (impregnated) in a resin bath just prior to winding. With wet wind-
ing, the viscosity of the resin is quite important -- dripping may occur if the resin becomes
too fluid, while poor wetting may result if the resin is too viscous.

Pultrusion is a continuous process involving impregnation, shaping and curing steps.
Pultrusion offers high throughput and minimal material waste, but part cross-sectional
area must be largely constant. Automated equipment for pulling, winding and cutting
operations is typically employed.

Matched-die molding may utilize sheet molding compound (SMC) or bulk molding

compound (BMC) together with male/female mold pairs. Material is placed in one mold

199

half and pressed into the second mold half. Pressure and heat are then maintained to cure
the material, which displays two finished surfaces.

Resin-transfer molding (RTM) may be used for greater manufacturing productivity. In
this technique, a mold is pre-loaded with the reinforcement and closed up. Resin is then
injected under pressure (actually, a vacuum is usually drawn within the mold to assist in
creating a pressure differential) and the part is cured. RTM has an advantage over manual
lay-up in terms of throughput, and the process may be automated more easily as well.
There are several critical factors, however, including design of the mold, viscosity of the
resin and potential movement of the reinforcement during injection, which can make RTM

impractical in many applications.

APPENDIX F
Extending the Capabilities of IPCA: A Case Study

Extending the IPCA Framework

IPCA is designed to allow straightforward incorporation of new planning and control
method fragments into the existing problem-solving framework. It is possible to extend
the available set of IPCA control algorithms and planning approaches, to provide coverage
for additional fabrication technologies, new materials and so on. The IPCA architecture
serves as an extensible framework for representing knowledge about composite materials
processing.

Each extension potentially impacts IPCA in two areas:

Knowledge about the applicability of the extension is placed into an
appropriate agent within the DSPL-based Blueprint Generator

Procedural algorithms covering new control and planning methods
are added to the fragment library

Details regarding these extensions are reviewed in the sections below. Some common

types of extensions are shown in table F. 1.

200

201

 

 

 

 

 

 

 

 

 

. Impact on Impact on
Type Of Extensron Blueprint Generator Fragment Library Notes
Add new control Add knowledge Add procedural Method name must
method about method body of method to be unique in
applicability fragment library library hierarchy
Add new planning Add knowledge of Add procedural Method name must
method when method is body of method to be unique in
useful fragment library library hierarchy
Add new composite Add knowledge of None required Processing temps,
material material properties times, characteris-
tics of material
Add new fabrica— Add knowledge Add any new frag- Design DSPL
tion technology about when tech- ments required by structure and add
nology is indi- the technology to existing IPCA
cated; add DSPL framework
agents to design
blueprints

 

 

Table F.1 - Common IPCA Extensions

The final section in this appendix details the procedure for adding a new type of mode-

switching control method fragment to IPCA.

Blueprint Generator Extensions

Extensions to the DSPL framework of the Blueprint Generator are made by inserting
appropriate design agents into the existing IPCA DSPL hierarchy. This task is accom-
plished using the IPCA Blueprint Generator interface (see appendix B) to browse the
agent hierarchy and modify it. To accommodate a new fabrication technology, for exam-
ple, it is necessary to insert a new specialist for the technology in an appropriate position

under the fabrication technology specialist. Similarly, to add coverage for new matrix

202

materials, the design agents must be modified to provide for setting appropriate processing
temperatures for the materials involved.

Whenever new planing or control method fragments are added to the fragment library,
knowledge about the applicability of the fragments must be included in the constraints and
steps within the existing IPCA DSPL design hierarchy so the new fragments are specified

for use under the prOper conditions.

Fragment Library Extensions and Changes

Extensions to the IPCA fragment library are made by using the Fragment Librarian tool
(see appendix B). The Fragment Librarian is capable of adding, deleting and renaming
plan and control method fragments. Using the command key and selecting an item in the
Library, Fragment or Element lists causes a pop-up menu to appear which includes
options to add, delete and rename items in the lists. The lists are hierarchical, meaning
that each Library entry can have many Fragment entries, and each Fragment entry can
have many Element entries. Each Element, in turn, contains text which can be edited.
During plan assembly, items in the Library and Fragment lists are matched against
the key/value pairs in the blueprint being assembled. For example, if the pair (ModeCon-
trolPolicy ModeControlPolicyC) is included in the blueprint, this means that the Blue-
print Generator previously set the value of the DSPL variable Mode ControlPolicy to the
value ModeControlPolicyC. The Plan Assembler will attempt to match the key ModeCon-
trolPolicy against entries in the Library list. If a match is found, the assembler next
attempts to match the value ModeControlPolicyC against items in the Fragment list. If a

match is found, the Elements from the matched fragment are included in the process con-

203

trol plan.
Each fragment contains the following elements:

Parameters-Abstract: A list of the abstract parameters used by the fragment.
Example: (PartTemp PartGrad ElapsedProcessingTime).

Parameters-Input: A list of the input parameters used by the fragment.
Example: (T 1 Sensor T28ensor T3Sensor T4Sensor).

Parameters-Output: A list of the output parameters used by the fragment.
Example: (ModeActuatorl PowerActuatorl).

Parameters-Source: The SmallTalk source code for the executable portion of any abstract
parameters required by the fragment. This code is evaluated at each execution cycle for

all abstract parameters. See fragment library for examples of format.

Parameters- Variable: A list of the variable parameters used by the fragment.
Example: (ProcessingGoal InitializationComplete).

Plan-CoFragments: A list of fragments which need to be included with this fragment.
Usually this list is empty, but it can contain references to other dependant fragments.
Example: (ModeControlPolicy@ModeControlPolicyC)

Plan-Source: The SmallTalk source code for the executable portion of the fragment. See
fragment library for examples of format.

Case Study - Adding a new Fragment Type to IPCA

This section contains a case study detailing how a new control method fragment was
added to IPCA. The fragment supports a new mode-switching behavior which attempts to
equalize the average temperature of all probes. The new fragment was used for experi-
mental series F4.

Previous mode switching methods (used in experimental series Fl through F3)
attempted to minimize thermal gradients by selecting processing modes that would instan-

tantanously minimize the temperature differential between the probes. This approach

204

resulted in the observation that some probes were consistently warm while others were
consistently cool. This situation is undesirable during fabrication due to the effects of
incomplete chemical crosslinking in some of the cooler areas of the sample.

The new fragment overcomes these problems by switching between modes using a
different mode switching strategy. The new strategy attempts to achieve and maintain
equal average temperatures for all probes by basing mode switches on historical tempera-
ture data for each probe. For additional detail, see chapter 6.

The following steps were performed to design the new fragment and integrate it into
IPCA:

l. The algorithm for the new mode switching fragment was designed and coded. The
input, output, variable and abstract parameters used by the new fragment were included in
the element lists using the fragment librarian.

2. The SetModeControlPolicy step under the MDProcessStructure specialist in the Blue-
print Generator DSPL hierarchy was modified to recommend the new fragment as the
standard mode switching policy for all reactive mode switching applications.

3. A process control plan was generated and tested against the simulated control unit to

verify proper behavior and debug the new fragment.
4. The process control plan was tested in the laboratory in experimental series F4.

BIBLIOGRAPHY

F.L. Abrams, T.L. Lagnese, S.R. LeClair, CW. Lee, P.H. Garrett, and IE. Park, Qualita-
tive Process Automation for Autoclave Curing of Composites, Air Force Wright Aeronau-
tical Laboratories Technical Report 97-4083.

F.L. Abrams, An Expert System Process Controller for Advanced Composites, Proceed-
ings; Detroit Re Tec '87 (Society of Plastics Engineers, Detroit, MI).

L. Acar, and U. Ozguner, Design of Hierarchically Distributed Expert Controllers for
Large-Scale Systems, Proceedings of the 1989 IEEE International Symposium on Intelli-
gent Control, (Albany, New York, September 25-26, 1989), IEEE Computer Society Press,
pp. 12-17.

V.O. Adegbite, Dissertation for the degree of Doctor of Philosophy, Michigan State Uni-
versity, East Lansing, MI, 1995.

J.S. Albus, ATheory of Intelligent Systems, Fifth IEEE International Symposium on Intel-
ligent Control (Philadelphia, Sept. 5-7, 1990) pp. 866—875.

J.S. Albus, Lumia, R., and McCain, H, Hierarchical Control of Intelligent Machines
Applied to Space Station Telerobots, 1987 IEEE International Symposium on Intelligent
Control, (Philadelphia, Pennsylvania, January 19-20 1987), pp. 20-26.

R. Alterman, An Adaptive Planner, Proceedings of the Fifth National Conference on Arti-
ficial Intelligence (AAAI '86), pp. 65-69.

J .A. Ambros-Ingerson, and S. Steel, Integrating Planning, Execution and Monitoring, Pro—
ceedings of the Seventh National Conference on Artificial Intelligence (AAAI '88), pp. 83-
88.

PJ. Antsaklis, K.M. Passino, and 8.]. Wang, Autonomous Control Systems: Architecture

and Fundamental Issues, Proceedings of the 1988 American Control Conference, (Atlanta,
Georgia, June 15-17, 1988), American Automatic Control Council, pp. 602-607.

205

206

K.E. Arzen, Knowledge-Based Control Systems - Aspects on the Unification of Conven-
tional Control Systems and Knowledge-Based Systems, Proceedings of the 1989 Ameri-
can Control Conference, (Pittsburgh, Pennsylvania, June 21-23, 1989), American
Automatic Control Council, pp. 2233-2238.

J. Asmussen, et. al., Single-Mode or Controlled Multimode Microwave Cavity Applicators
for Precision Material Processing, Review of Scientific Instruments, Vol. 58, No. 8, 1987,
pp. 1477-1486.

K.J. Astrom, B. Wittenmark, Computer-Controlled Systems, Prentice-Hall, Englewood
Cliffs, New Jersey, Second edition, 1990.

M. Aynsley, D. Peel, and AJ. Morris, A Real-Time Knowledge-Based System for Fermen-
tation Control, Proceedings of the 1989 American Control Conference, (Pittsburgh, Penn—
sylvania, June 21-23, 1989), American Automatic Control Council, pp. 2239-2244.

V.V. Badami, P. Nielson, and J .B. Comly, An Intelligent Controller for Process Automa-
tion, Proceedings of the Fifth IEEE International Symposium on Intelligent Control, (Phil-
adelphia, Pennsylvania, September 5-7 1990), pp. 1076-1081.

R. Barletta, and D. Hennessy, Case Adaptation in Autoclave Layout Design, Lockheed AI
Center O/96-20 B/259.

M.R. Basila J r., and A. Cinar, A Model-Object Based Supervisory Expert System for Fault
Tolerant Chemical Reactor Control, Proceedings of the 1989 American Control Confer-
ence, (Pittsburgh, Pennsylvania, June 21-23, 1989), American Automatic Control Council,
pp. 1348-1353.

T.E. Bihari, T.M. Walliser, and MR. Patterson, Controlling the Adaptive Suspension Vehi-
cle, IEEE Computer, Vol. 22, No. 6 (June 1989), pp. 59-65.

M. Boddy, and T. Dean, Solving Time-Dependent Planning Problems, Proceedings of the
11th UCAI (1989), pp. 979-984

S.P. Boyd and CH. Barratt, Linear Controller Design - Limits of Performance, Prentice
Hall, Englewood Cliffs, New Jersey, 1991.

DC. Brillhart, and SF. Wible, An Expert System for Real-Time Process Characterization
and Control, Proceedings of the Seventh IEEE/CHMT International Electronic Manufac-
turing Technology Symposium (San Francisco, California, September 25-27, 1989), pp.
76-83.

W.L. Brogan, Modern Control Theory, Prentice Hall, Englewood Cliffs, New Jersey, Third
edition, 1991.

207

RA. Brooks, A Robust Layered Control System for a Mobile Robot, IEEE J oumal of
Robotics and Automation, Vol. 2, No. 1, March 1986, pp. 14-23.

C.A. Broverman, and W.B. Croft, Reasoning About Exceptions During Plan Execution
Monitoring, Proceedings of the Seventh National Conference on Artificial Intelligence
(AAAI- '88). Pp. 190-195.

D.C. Brown and B. Chandrasekaran, Design Problem Solving - Knowledge Structures and
Control Strategies, Morgan Kaufmann, San Mateo, California, 1989.

EC. Campbell, A.R. Mallow, F.R. Muncaster, E.L. Boman, and G.A. Blase, Computer-
Aided Curing of Composites, Air Force Wright Aeronautical Laboratories Technical
Report 86-4060.

S. Chand, On-line Self-Monitoring Tuner for Proportional Integral Derivative Controllers,
199] IEEE Decision and Control, Vol. 2, pp. 1905-1906.

B. Chandrasekaran, R. Bhatnagar, and DD. Sharma, Real-Time Disturbance Control,
Communications of the ACM, August 1991, pp.32-47.

B. Chandrasekaran, Design Problem Solving: A Task Analysis, AI Magazine, Winter,
1990, pp. 59-71.

B. Chandrasekaran, et. al., An Approach to Routine Planning, International Journal of
Man-Machine Studies, 1986.

B. Chandrasekaran, Generic Tasks in Knowledge-Based Reasoning: High-Level Building
Blocks for Expert System Design, IEEE Expert, Vol. 1, No. 3, Fall 1986, pp. 23-30.

B. Chandrasekaran, J. Josephson, A. Keuneke, and D. Herman, Building Routine-Plan-
ning Systems and Explaining their Behavior, International Journal of Man-Machine Stud-
ies, Vol. 30, 1986, pp. 377-398.

D. Chapman, Planning for Conjunctive Goals, Artificial Intelligence, Vol. 32, 1987, pp.
333-377.

D. Chapman, Penguins Can Make Cake, AI Magazine, Winter 1989, pp. 45-50.
J .C. Chautard, and C. Honnorat, PBX: A Reactive Procedure Based Decision Maker, 1991
IEEE Conference on Al Applications, pp. 368-371.

M. Cibrario, and J. Levine, Saddle-Node Bifurcation Control with Application to Thermal
Runaway of Continuous Stirred Tank Reactors, 1991 IEEE Decision and Control, Vol. 2,
pp. 1551-1552.

T.B. Co, J .G. Williams, and Z. Chen, Control Studies on the Advanced Composites Manu-
facturing Process, Composites Manufacturing, December 1990, pp. 253-259.

208

N. Coleman, An Emulation/Simulation for Intelligent Controls, Proceedings of the Work-
shop on Software Tools for Distributed Intelligent Control Systems, (July 17-19, 1990,
Pacifica, California), pp. 37-52.

R.D. Coyne, M.A. Roseman, A.D. Radford, M. Balachandran and J .8. Gem, Knowledge-
Based Design Systems, Addison-Wesley, Reading, Massachusetts, 1990.

J .R. Crosscope, and RD. Bonnell, An Integrated Intelligent Controller Employing both
Conceptual and Procedural Knowledge, 1987 IEEE International Symposium on Intelli-
gent Control, (January 19-20, 1987, Philadelphia, Pennsylvania), pp. 416-422.

J .B. Cruz, and AR. Stubberud, Knowledge-Based Approach to Multiple Control Coordi-
nation in Complex Systems, Proceedings of the 1987 IEEE International Symposium on
Intelligent Control, (Philadelphia, Pennsylvania, January 19-20, 1987), IEEE Computer
Society Press, pp. 50-53.

B. D'Ambrosio, M.R. Fehling, P. Raulefs, and B.M. Wilber, Real-Time Process Manage-
ment for Materials Composition in Chemical Manufacturing, IEEE Expert, Vol. 2, No. 2,
Summer 1987, pp. 80-93.

B. Dasarathy, Timing Constraints of Real-Time Systems: Constructs for Expressing them,
Methods for Validating Them, IEEE Transactions on Software Engineering, Vol. 11, No.
1, January 1985. PP. 80-86.

T. Dean, and M. Boddy, An Analysis of Time-Dependent Planning, Proceedings of the
Seventh National Conference on Artificial Intelligence (AAAI '88), pp. 49-54.

T. Dean, and M. Boddy, Incremental Causal Reasoning, Proceedings of the Seventh
National Conference on Artificial Intelligence (AAAI '88), pp. 196-201.

T. Dean, and G. Siegle, An Approach to Reasoning About Continuous Change for Appli-
cations in Planning, Proceedings of the Eighth National Conference on Artificial Intelli-
gence (AAAI '90), pp. 132-137.

D. Decker, J. Sticklen, et al., An Adaptive Intelligent Control Architecture for the Micro-
wave Fabrication of Polymeric Thermoset Composite Materials Utilizing an Empirical
Single-Mode Resonant Cavity Scanning Technique, Proceedings of the AIChE 1993 Sum-
mer National Meeting, (Seattle, Washington, August, 1993).

D. Decker, J. McDowell, M. Hawley, W. Punch, J. Sticklen, An Intelligent Control Archi-
tecture for Processing of Polymer Composites, Proceedings of the 4lst Annual Sagarnore
Army Materials Conference, (Plymouth, Massachusetts, August, 1993

209

D. Decker, W. Punch, J. McDowell, J. Sticklen, An Intelligent Control Architecture for the
Microwave Fabrication of Composite Materials Based on the Generic Task Approach to
Knowledge-Based Systems, Proceedings of the AAAI Fall Symposium on Intelligent Sys-
tems, (New Orleans, Louisiana, November 3-6 1994).

R. Doraiswami, and W. Liu, An Expert System for Monitoring the Status of Control Sys-
tems, Proceedings of the Fifth IEEE International Symposium on Intelligent Control,
(Philadelphia, Pennsylvania, September 5-7 1990), pp. 457-462.

M. Drummond, and J. Bresina, Planning for Control, Proceedings of the Fifth IEEE Inter-
national Symposium on Intelligent Control, (Philadelphia, Pennsylvania, September 5-7
1990), pp. 657-662.

M. Dudziak, M. Hall, and D. Zaret, IVC: An Intelligent Vehicle Controller with Real-Time
Strategic Replanning, 1987 IEEE International Symposium on Intelligent Control, (J anu-
ary 19-20, 1987, Philadelphia, Pennsylvania), pp. 145-152.

E.H. Durfee, and V.R. Lesser, Incremental Planning to Control a Blackboard-Based Prob-
lem Solver, Proceedings of the Fifth National Conference on Artificial Intelligence
(AAAI '86). Pp. 58-64.

E.H. Durfee, A Cooperative Approach to Planning for Real-Time Control, Proceedings of
DARPA Workshop on Innovative Approaches to Planning, Scheduling and Control
(1990), pp. 277-283. Morgan Kauffman.

C. Elkan, Incremental Approximate Planning, Proceedings of the Eighth National Confer-
ence on Artificial Intelligence, (AAAI '90), MIT Press, Cambridge, Mass, 1990, pp. 145-
150.

L.D. Erman, Ed. Intelligent Real-Time Problem Solving, Air Force Office of Scientific
Research, Workshop Rep. TI'R-ISE-90— 101, January 1990.

A. Farsaie, T.R. McKnight, B. Ferren, and CF. Harrison, Intelligent Controllers for an
Autonomous System, Proceedings of the 1987 IEEE International Symposium on Intelli-
gent Control, (Philadelphia, Pennsylvania, January 19-20, 1987), IEEE Computer Society
Press, pp. 154-157.

L.A. Fellows, Microwave Processing of Polyamides and Polyamide Composites, Thesis
(M.S.), Michigan State University, East Lansing, 1991.

J .A. Ferguson, Touring Machines: Autonomous Agents with Attitudes, IEEE Computer,
May 1992. pp. 51-55. '

210

RE. Fikes, and NJ. Nilsson, STRIPS: A New Approach to the Application of Theorem
Proving to Problem Solving, Artificial Intelligence, Vol. 2, 1971, pp. 189-208.

J. Fiksel, and F. Hayes-Roth, Knowledge Systems for Planning Support, IEEE Expert, Fall
1991, pp.l6-23.

R.J. Finocchio, A.J. Blanch, T.S. Lorincz, R. Meth, and B.P. Butz, Implementing a Real-
Tirne Knowledge-Based Controller In a Velocity Controlled, Servo Motor Control Sys-
tem, Proceedings of the Fifth IEEE International Symposium on Intelligent Control, (Phil-
adelphia, Pennsylvania, September 5-7 1990), pp. 811-815.

G. Fiorio, F. Lombardi, and A. Villa, Integration of Multi-Hierarchy Control Architectures
for Complex Systems, 1987 IEEE International Symposium on Intelligent Control, (J anu-
ary 19-20, 1987, Philadelphia, Pennsylvania), pp. 71-79.

R.J. Firby, An Investigation into Reactive Planning in Complex Domains, Proceedings of
the Seventh National Conference on Artificial Intelligence (AAAI -'88), pp. 202-206.

K.D. Forbus, Qualitative Process Theory, Artificial Intelligence, Vol. 24, pp. 85-168.

T.H. Fraichard, and C. Laugier, On-line Reactive Planning for a Non Holonomic Mobile
in a Dynamic World, Proceedings of the 1991 IEEE International Conference on Robotics
and Automation (Sacramento, California, April 1991), pp. 432-437.

P. Garrett, CW. Lee, and SR. LeClair, Qualitative Process Automation vs. Quantitative
Process Control, Proceedings of the American Control Conference, Minneapolis, MN,
American Automatic Control Council.

2. Geng, R. Carroll, M. Jamshidi, and R. Kisner, An Adaptive Learning Control Approach,
1991 IEEE Decision and Control, Vol. 2, pp. 1221-1222.

M.P. Georgeff, and FF. Ingrand, Decision Making in an Embedded Reasoning System,
Proceedings of the International Joint Conference on Artificial Intelligence, Morgan Kauf-
mann, San Mateo, California, 1989, pp. 972-978.

M.P. Georgeff, and AL. Lansky, Reactive Reasoning and Planning, Proceedings of the
Sixth National Conference on Artificial Intelligence (AAAI '87), Seattle, Washington, July
1987. pp. 677-682.

M.L. Ginsberg, Universal Planning: An (Almost) Universally Bad Idea, AI Magazine,
Winter 1989, pp. 40-44.

ML. Ginsberg, Possible Worlds and the Qualification Problem, Proceedings of the Sixth
National Conference on Artificial Intelligence (AAAI '87), pp. 212-217.

211

P. Godefroid, and F. Kabanza, An Efficient Reactive Planner for Synthesizing Reactive
Plans, Proceedings of the Ninth National Conference on Artificial Intelligence (AAAI
'91). pp. 640-645.

M. Gopal, Modern Control System Theory, John Wiley & Sons, New York, Second edi-
tion, 1993.

DA. Handelman, and RF. Stengel, Rule-Based Mechanisms of Learning for Intelligent
Adaptive Flight Control, Proceedings of the 1988 American Control Conference, (Atlanta,
Georgia, June 15-17, 1988), American Automatic Control Council, pp. 208-213.

E. Harris, Engineering Composite Materials, The Institute of Metals, 1986.

M.C. Hawley, and J. Wei, Microwave Processing of Polymer Composites, MRS Spring
Meeting, San Francisco, CA, April 1994.

C. Hayes, Using Goal Interactions to Guide Planning, Proceedings of the Seventh
National Conference on Artificial Intelligence (AAAI '88), pp. 224-228.

B. Hayes-Roth, Architectural Foundations for Real-Time Performance in Intelligent
Agents, Journal of Real-time Systems, Vol. 2 (1990) pp. 99-125.

R.F. Herrington, Time-Harmonic Electromagnetic Fields, McGraw-Hill, New York, 1961.

J PB. Hodgson, Structures for Intelligent Systems, 1987 IEEE International Symposium
on Intelligent Control, (January 19-20, 1987, Philadelphia, Pennsylvania), pp. 348-353.

L.E. Holloway, C.J. Paul, J .K. Strosnider, and EH. Krough, Integration of Behavioral
Fault-Detection Models and an Intelligent Reactive Scheduler, Proceedings of the Sixth
IEEE International Symposium on Intelligent Control (Arlington, Virginia, August 13-15
1991), pp. 134-139.

F. J ahanian, and A. Mok, Safety Analysis of Timing Properties in Real-Time Systems,
IEEE Transactions on Software Engineering, Vol. 12, No. 9, September 1986, pp. 890-
904.

M. Jamshidi, S. Baugh, D. Barak, and N. Vadiee, A Comparison of an Expert and an
Adaptive Fuzzy Control Approach, Proceedings of the 30th IEEE Conference on Deci-
sion and Control (Brighton, England, Dec. 1991), Vol. 2, pp. 1907-1908.

J. Jow, M.C. Hawley, and M. Finzel, Microwave Heating and Dielectric Diagnosis Tech-
nique in a Single-Mode Resonant Cavity, Review of Scientific Instruments, Vol. 60, No. 1,
January 1989. pp. 96-103.

212

J. Jow, M.C. Hawley, M. Finzel and T. Kern, Dielectric Analysis of Epoxy/Amine Resins
using Microwave Cavity Technique, Polymer Engineering and Science, Vol. 28, No. 22,
November 1988, pp. 1450-1454.

L.P. Kaelbling, An Architecture for Intelligent Reactive Systems, Proceedings of the 1986
Workshop on Reasoning About Actions and Plans, M.P. Georgeff and AL. Lansky, eds.,
1987, Morgan Kaufmann Publishers, Los Altos, California, pp. 395-410.

S. Karnbhampati, A Theory of Plan Modification, Proceedings of the Eighth National
Conference on Artificial Intelligence (AAAI '90), pp. 176-182.

A. Kamel, Dissertation for the degree of Doctor of Philosophy, Michigan State University,
East Lansing, MI, 1994.

G. Karsai, C. Biegl, S. Padalkar, J. Sztipanovits, and K. Kawamura, Knowledge-Based
Approach to Real-Time Supervisory Control, Proceedings of the 1988 American Control
Conference, (Atlanta, Georgia, June 15-17, 1988), American Automatic Control Council,
pp. 620-625.

H.A. Kautz, and LP. Allen, Generalized Plan Recognition, Proceedings of the Fifth
National Conference on Artificial Intelligence (AAAI '86), pp. 32-37.

H. Kern, and M. Fathi-Torbaghan, Concept for Combining features of Real-time and
Expert systems for On-Line Plant Diagnosis, Ninth Annual International Phoenix Confer-
ence on Computers and Communications (Scottsdale, Arizona, March 21-23, 1990), pp.
803-807.

C.A. Knoblock, J .D. Tenenberg, and Q. Yang, Characterizing Abstraction Hierarchies for
Planning, Proceedings of the Ninth National Conference on Artificial Intelligence (AAAI
'91). pp. 692-697.

W. Kohn, Declarative Control Architecture, Communications of the ACM, August 1991,
pp. 64-79.

W. Kohn, Declarative Hierarchical Controllers, Proceedings of the Workshop on Innova-
tive Approaches to Planning, Scheduling and Control, (N 0v. 1990).

W. Kohn, Advanced Architectures and Methods for Knowledge-Based Planning and
Declarative Control, IR and D BSC-021, To be presented as a paper in ISMIS'91, October
1991.

W. Kohn, and T. Skillmann, Hierarchical Control Systems for Autonomous Space Robots,

Proceedings of AIAA Conference on Guidance, Navigation and Control, Vol. 1, Minneap-
olis, MN, Aug. 15-18 1988. pp. 382-390.

213

T.J. Laffey, RA. Cox, J .L. Schmidt, S.A. Kao, and J .Y. Read, Real-Time Knowledge-
Based Systems, AI Magazine, Vol. 9, No. 1, Spring 1988, pp. 27-45.

T. Laffey, S. Weitzenkarnp, J. Read, S. Kao, and J. Schmidt, Intelligent Real-Time Moni-
toring, Proceedings of the Seventh National Conference on Artificial Intelligence (AAAI -
'88, August 1988, Lockheed Artificial Intelligence Center), pp. 72-76.

AL. Lansky, and BS. Fogelsong, Localized Representation and Planning Methods for
Parallel Domains, Proceedings of the Sixth National Conference on Artificial Intelligence
(AAAI '87), pp. 240-245.

A.A. Lazar, J .T. Amenyo, and S. Mazumdar, WIENER: A Distributed Expert System for
Dynamic Resource Allocation in Integrated Networks, 1987 IEEE International Sympo-
sium on Intelligent Control, (January 19-20, 1987, Philadelphia, Pennsylvania), pp. 159-
164.

SR. LeClair, FL. Abrams, and RF. Matejka, Qualitative Process Automation: Self-
Directed Manufacture of Composite Materials, AI EDAM, Vol. 3, No. 2, 1989, pp. 125-
136.

SR. LeClair, and FL. Abrams, Qualitative Process Automation, International Journal of
Computer Integrated Manufacturing, Vol. 2, No. 4, pp. 205-211.

S.R. LeClair, Sensor Fusion: The Application of Artificial Intelligence to Process Control,
1986 Rochester FORTH Conference Proceedings, Rochester, NY, pp. 15-22.

S.R. LeClair, and J .8. Park, Sensor Fusion, Statistical Process Control, edited by Keats
and Heuble (Marcel Dekker, New York), NY. pp. 235-267.

CW. Lee, Composite Cure Process Control by Expert Systems, Department of Chemical
Engineering, University of Dayton.

S. Lee, and Y.G. Shin, Multiple Agent Cooperative Problem Solving with Axiom-Based
Negotiation, 1987 IEEE International Symposium on Intelligent Control, (January 19-20,
1987, Philadelphia, Pennsylvania), pp. 469-474.

T.H. Lee, K.W. Lim, and WC. Lai, Real-Time Multivariable Self-Tuning Controller Using
a Feedforward Paradigm with Application to a Coupled Electric-Drive Pilot Plant, IEEE
Transactions on Industrial Electronics, August 1991, pp. 237-242.

W.I. Lee, and GS. Springer, Interaction of Electromagnetic Radiation with Organic
Matrix Composites, Journal of Composite Materials, Vol. 18, July 1984.

W.I. Lee, and GS. Springer, Microwave Curing of Composites, Journal of Composite
Materials, Vol. 18, July 1984.

214

V. Lesser, J. Pavlin, and E. Durfee, Approximate Processing in Real-Time Problem Solv-
ing, AI Magazine, Vol. 9, No. 1, Spring 1988, pp. 49-61.

T. Lenz, J .K McDowell, B. Moy, J. Sticklen and M. Hawley, Intelligent Decision Support
for Composite Material Design in an Integrated Design Environment, American Society of
Composites (ASC) 9th Technical Conference on Composite Materials, University of Dela-
ware, 1994.

TA. Linden, and J. Glicksman, Contingency Planning for an Autonomous Land Vehicle,
Proceedings of the 1987 International Joint Conference on A1, pp. 1047-1054.

A.C. Loos, and GS. Springer, Curing of Epoxy Matrix Composites, Journal of Composite
Materials, Vol. 17, March 1983, pp. 135-169.

D. McDermott, Robot Planning, AI Magazine, Summer 1992, pp. 55-79.

J .K. McDowell, A. Kamel, J. Sticklen and M.C. Hawley, Integrating Material/Part/Process
Design for Polymer Composites: A Knowledge-Based Problem Solving Approach, In D.
G. Newaz (Ed.), ASC 8th Technical Conference on Composite Materials, Cleveland, OH,
American Society of Composites, 1993.

A. Meystel, Planning/Control Architectures for Master Dependent Autonomous Systems
with Nonhomogeneous Knowledge Representation, 1987 IEEE International Symposium
on Intelligent Control, (January 19-20, 1987, Philadelphia, Pennsylvania), pp. 31-47.

D.J. Minnick, H. Kaufman, and G.W. Neat, Expert Hierarchical Adaptive Controller for
Robotic Systems, Proceedings of the 1989 IEEE International Symposium on Intelligent
Control, (Albany, New York, September 25-26, 1989), IEEE Computer Society Press, pp.
8- l l .

J .S.B. Mitchell, D.W. Payton, and D.M. Keirsey, Planning and Reasoning for Autonomous
Vehicle Control, International Journal of Intelligent Systems, Vol. 2 (1987), pp. 129-198.

D. Moldovan, W.L. Lee, L. Changhwa, and M. Chung, SNAP Parallel Processing Applied
to AI, IEEE Computer, May 1992, pp. 39-49.

F. Morant, P. Albertos, M. Martinez, and J .J . Serrano, Hierarchical Expert System as
Supervisory Level in an Adaptive Controller, Proceedings of the 1989 IEEE International
Symposium on Intelligent Control, (Albany, New York, September 25-26, 1989), IEEE
Computer Society Press, pp. 18-25.

B.J. Moy, J .K. McDowell, T. Lenz, J. Sticklen and M.C. Hawley, Expansion of an Intelli-
gent Decision Support System for Process Selection and Process Design in Polymer Com-
posites, Abstract submitted to the 26th International SAMPE Technical Conference, Feb.
1994.

215

DJ. Musliner, E.H. Durfee, and K.G. Shin, Execution Monitoring and Recovery Planning
with Time, 1991 IEEE Conference on Al Applications, pp. 385-388.

D.J. Musliner, J .A. Hendler, A.K. Agrawala, et. al, The Challenges of Real-Time AI, IEEE
Computer, January, 1995, pp. 58-66.

S.R. Nann, A. Ray, and S. Kumara, A Decision Support System for Real-Time Control and
Monitoring of Dynamical Processes, Proceedings of the 1989 American Control Confer-
ence, (Pittsburgh, Pennsylvania, June 21-23, 1989), American Automatic Control Council,
pp. 361-365.

National Instruments Corporation, LabVIEW 2 User Manual, September 1991 Edition,
Part number 320244-01, 6504 Bridge Point Parkway, Austin, TX.

A. Newell, The Knowledge Level, Artificial Intelligence 18 (1982), pp. 87-127.

A. Nitsche, H. Kern, and J. Janczak, Composites Design Based on Expert Knowledge,
Fifth Technical Conference of the American Society for Composites, Technomic Publish-
ing Company, Lancaster, PA, 1990, pp. 382-390.

D.A. Norman, Approaches to the Study of Intelligence, Artificial Intelligence 47 (1991),
pp. 327-346.

G.K.H. Pang, An Expert Adaptive Control Scheme in an Intelligent Process Control Sys-
tem, Proceedings of the 1991 IEEE International Symposium on Intelligent Control
(Arlington, VA, August 13-25, 1991), pp. 13-17.

W.J. Pardee, and B. Hayes-Roth, Intelligent Real-Time Control of Material Processing,
Technical report, Rockwell International Science Center, Palo Alto Laboratory, Palo Alto,
CA, 1987.

J. Park, Toward the Development of a Real-Time Expert System, 1986 Rochester FORTH
Conference Proceedings, New York, pp. 23-33.

C.J. Paul, A. Acharya, B. Black, and J .K. Strosnider, Reducing Problem-Solving Variance
to Improve Predictability, Communications of the ACM, August 1991, pp. 80-93.

C.J. Paul, L.E. Holloway, D. Yan, J .K. Strosnider, and EH. Krogh, An Intelligent Reactive
Monitoring and Scheduling System, IEEE Control Systems, June 1992, pp. 78-86.

D.W. Payton, Intemalized Plans: A Representation for Action Resources, Robotics and
Autonomous Systems, Vol. 6, No. 1, 1990, pp. 89-103. '

D.W. Payton, J.K. Rosenblatt, and D.M. Keirsey, Plan-Guided Reaction, IEEE Transac-
tions on Systems, Man and Cybernetics, Vol. 20, No. 6, 1990, pp. 1370-1382.

216

D.W. Payton, and TE. Bihari, Intelligent Real-Time Control of Robotic Vehicles, Commu-
nications of the ACM, August 1991, pp. 48-63.

W.F Punch, III, A.K. Goel and DC. Brown, A Knowledge-Based Selection Mechanism
for Control with Applications in Design, Assembly and Planning, accepted for publication
in the International Journal of Artificial Intelligence Tools, 1995.

W.F. Punch III, Dissertation for the degree of Doctor of Philosophy, The Ohio State Uni-
versity, 1989.

M. Rao, T.S. Jiang, and J .P. Tsai, From Knowledge Transfer to Knowledge Generation,
Proceedings of the 1989 IEEE International Symposium on Intelligent Control, (Albany,
New York, September 25-26, 1989), IEEE Computer Society Press, pp. 387-392.

M. Rao, T.S. Jiang, and J .P. Tsai, Software Environment and Architecture for Intelligent
Control, Proceedings of the 1989 IEEE International Symposium on Intelligent Control,
(Albany, New York, September 25-26, 1989), IEEE Computer Society Press, pp. 482-485.

A.A. Rizzi, LL. Whitcomb, and DE. Koditschek, Distributed Real-Time Control of a
Spatial Robot Juggler, IEEE Computer, May 1992, pp. 12-24.

J .L. Rose, Z.K. Li, C. Thomas, G. Chen, and R. Tupi], Sensor Monitoring and Control
Concepts for Intelligent Polymer Processing, Proceedings of the Fifth IEEE International
Symposium on Intelligent Control, (Philadelphia, Pennsylvania, September 5-7 1990), pp.
1193-1 198.

J .K. Rosenblatt, and D.W. Payton, A Fine-Grained Alternative to the Subsumption Archi-
tecture for Mobile Robot Control, Proceedings of International Joint Conference on Neu-
ral Networks (Washington, DC, June 1989), pp. 317-324.

E.D. Sacerdoti, Planning in a Hierarchy of Abstraction Spaces, Artificial Intelligence, Vol.
5, 1974, pp. 115-135.

J. Sanbom, and J. Hendler, A Model of Reaction for Planning in Dynamic Environments,
International Journal of Artificial Intelligence in Engineering, Vol. 3, No. 2, April 1988,
pp. 95- 102.

ON. Saradis, Knowledge Implementation: Structures of Intelligent Control Systems, Pro-
ceedings of the 1987 IEEE International Symposium on Intelligent Control, (Philadelphia,
Pennsylvania, January 19-20, 1987), IEEE Computer Society Press, pp. 9—17.

C.F. Schmidt, J .L. Goodson, S.C. Marsella, and J.L. Bresina, Reactive Planning Using a
“Situation Space”, Proceedings of the Annual AI Systems in Government Conference
(Washington, DC, March 27-31, 1989), pp. 50-55.

217

M.J. Schoppers, Automatic Synthesis of Perception Driven Discrete Event Control Laws,
Proceedings of the Fifth IEEE International Symposium on Intelligent Control (Philadel-
phia, Pennsylvania, January 19-20, 1987), IEEE Computer Society Press.

M.J. Schoppers, In Defense of Reaction Plans as Caches, AI Magazine, Winter 1989, pp.
51-60.

M.J. Schoppers, Universal Plans for Reactive Robots in Unpredictable Environments,
Proceedings of the 10th International Joint Conference on Al (August 1987), pp. 1039-
1046.

M.J. Schoppers, and T. Linden, The Dimensions of Knowledge-Based Control Systems
and the Significance of Meta-Levels, Proceedings of the AAAI Stanford Spring Sympo-
sium on Planning in Uncertain, Unpredictable or Changing Environments, J. Hendler, Ed.
Tech Rep. 45-90, Systems Research Center, University of Maryland, College Park, 1990.

RB. Schudy, and C.N. Duarte, Advanced Autonomous Underwater Vehicle Software
Architecture, Proceedings of the Symposium on Autonomous Underwater Vehicle Tech-
nology (Washington, DC, June 5-6, 1990), pp. 9-22.

L. Sha, and J .B. Goodenough, Real-Time Scheduling Theory and Ada, IEEE Computer,
Vol. 23, 1990. pp. 53-62.

D.D. Sharma, Plan Execution in Dynamic and Unanticipated Environments, 1991 IEEE
Conference on A1 Applications, pp. 361-367.

K.G. Shin, and X. Cui, Design of a Knowledge-Based Controller for Intelligent Control
Systems, IEEE Transactions on Systems, Man and Cybernetics, March/April 1991, pp.
368-375.

R. Shoureshi, and K. Rahmani, Intelligent Control of Building Systems, Intelligent Con-
trol Systems, Presented at the Winter Annual Meeting of ASME (December 10-15, 1989,

San Francisco, California), pp. 7-15.

T.S. Srivatsan, and S. Chandrashekhar (ed.), Processing and Manufacturing of Composite
Materials, ASME, Atlanta, Georgia, Dec. 1-6, 1991.

M. Stefik, Planning with Constraints (MOLGEN: Part 1), Artificial Intelligence, Vol. 16,
pp. 111-139.

M. Stefik, Planning and Meta-Planning (MOLGEN: Part 2), Artificial Intelligence, Vol.
16, pp. 141-170.

J. Sticklen, A. Kamel, M.C. Hawley, and V. Adegbite, Fabricating Composite Materials -
A Comprehensive Problem-Solving Architecture Based on Generic Tasks, IEEE Expert,
April 1992. Pp. 43-53.

218

J. Sticklen, and WE. Bond, Functional Reasoning and Functional Modeling, IEEE
Expert, April, 1991, pp. 20-21.

J. Sticklen, A. Kamel, M.C. Hawley, and J. Delong, An Artificial Intelligence-Based
Design Tool for Thin Film Composite Materials, MSU.

J. Sticklen, Problem Solving Architecture at the Knowledge Level, Journal of Experimen-
tal and Theoretical Artificial Intelligence, Summer 1989.

J. Sticklen, and B. Chandrasekaran, Integrating Classification-Based Compiled Level Rea-
soning with Function-Based Deep Level Reasoning, Applied Artificial Intelligence Vol. 3,
No. 2, 1989.

A. B. Strong, Fundamentals of Composites Manufacturing: Materials, Methods and
Applications, Society of Manufacturing Engineers, 1989.

M. Tambe, and PS. Rosenbloom, A Framework for Investigating Production System For-
mulations with Polynomially Bounded Match, Proceedings of the Eighth National Confer-
ence on Artificial Intelligence (AAAI -'90), August 1990, pp. 693-700.

J. Tenenberg, Planning with Abstraction, Proceedings of the Fifth National Conference on
Artificial Intelligence (AAAI '86), pp. 76-80.

E. Torgerson, and EW. Paul, A Hierarchical Semi-Autonomous Control Scheme for
Apparel Manufacturing, Intelligent Control Systems 1990, Presented at the Winter
Annual Meeting of ASME (November 25-30, 1990, Dallas, Texas), pp. 57-63.

E. Tse, and M. Fehling, An Intelligent Control System, Proceedings of the 1989 IEEE
International Symposium on Intelligent Control, (Albany, New York, September 25-26,
1989), IEEE Computer Society Press, pp. 194-199.

R.M. Turner, A Schema-Based Model of Adaptive Problem Solving, Doctoral Disserta-
tion, Georgia Institute of Technology, GIT-ICS-89/42, December, 1989, 229 pages.

L. Tychonievich, T.C. Smith, and R. Evans, Influence Networks: A Reactive Planning
Architecture, 1991 IEEE Conference on A1 Applications, pp. 354-360.

A. Valente, Knowledge-Level Analysis of Planning Systems, SIGART Bulletin, Vol. 6,
No. 1, January, 1995, pp. 33-41.

W. Van deVelde, Issues rn Knowledge Level Modelling, Second Generation Expert Sys-
tems, Jean-Marc David, et. al. (editors), Springer-Verlag, Berlin, 1993.

SM. Walsh, Artificial Intelligence: Its Application to Composite Processing, 35th Intema-
tional SAMPE Symposium, April 2-5, 1990. pp. 1280-1291.

219

J. Wei, J .D. DeLong, M. DeMeuse and M.C. Hawley, Comparison of Microwave and
Thermal Cure of Epoxy Resins, Polymer Engineering and Science, Vol 33, No. 17, 1993,
pp. 1132-1149.

M.P. Wellman, The STRIPS Assumption for Planning Under Uncertainty, Proceedings of
the Eighth National Conference on Artificial Intelligence (AAAI '90), pp. 198-203.

J .E. Whitlow, and K.A. Debelak, A Knowledge-Based Structure for Process Control, Pro-
ceedings of the 1989 American Control Conference, (Pittsburgh, Pennsylvania, June 21-
23, 1989), American Automatic Control Council, pp. 1354-1359.

D.E. Wilkens, Domain-independent Planning: Representation and Plan Generation, Artifi-
cial Intelligence, Vol. 22, 1984, pp. 269-301.

D.J. Williams, A.A. West, and OJ. Hinde, A Hierarchical Model of Intelligent Materials
Processing, 1991 IEEE International Symposium on Intelligent Control, pp. 330-333.

L. Winkelbauer, K. Fedra, ALEX: Automatic Learning in Expert Systems, Proceedings of
the 1991 IEEE Conference on Al Applications, pp. 59-62.

M. Winslett, Validating Generalized Plans in the Presence of Incomplete Information, Pro-
ceedings of the Sixth National Conference on Artificial Intelligence (AAAI '87), pp. 261-
266.

P.H. Winston, Artificial Intelligence, Addison-Wesley, Reading, Massachusetts, Third edi-
tion, 1992.

HT. Wu, and B. Joseph, Model-Based and Knowledge-Based Control of Pultrusion Pro-
cesses, SAMPE Journal, Vol. 26, No. 6, November/December 1990, pp. 59-70.

H.T. Wu, and B. Joseph, Knowledge-Based Control of Autoclave Curing of Composites,
SAMPE Journal, Vol. 26, No. 6, November/December 1990, pp. 39-54.

Q. Yang, and J .D. Tenenberg, ABTWEAK: Abstracting a Nonlinear Least Commitment
Planner, Proceedings of the Eighth National Conference on Artificial Intelligence (AAAI
'90). Pp. 204-209.

X. Zheng, and S. Srivastava, Events and Actions: An Object-Oriented Approach to Real-
Tirne Control Systems, Proceedings of the IEEE Pacific Rim Conference on Communica-
tions, Computers and Signal Processing (June 1-2, 1989), pp. 269-272.

S. Zilberstein, The Utility of Planning, SIGART Bulletin, Vol. 6, No. 1, January, 1995, pp.
42-47.

"‘titling