.. .i. . .
1 .-
“wir‘mnm...
I

1.5;.

[BALI-W
$33?
#3. 44..
. . ..
«iv

3 . , 9%

...

3...}. 1..

J5...
_

 

an.

 

“" “Hr. .

LIBRARY
Michigan State
University

This is to certify that the
thesis entitled

SEWER PIPELINE CONDITION PREDICTION USING

NEURAL NETWORK MODELS

presented by

GURUPRAKASH KULANDAIVEL

has been accepted towards fulfillment
of the requirements for the

Master of degree in Building Construction
Science Management

 

ML”! A/M'

Major Professor’s gigna‘ture

W27/ 2.004

Date

 

MSU is an Affirmative Action/Equal Opportunity Institution

 

 

 

_.—.--o-a-.-.-.-.-._. ..

 

PLACE IN RETURN Box to remove this checkout from your record.
To AVOID FINES return on or before date due.
MAY BE RECAILED with earlier due date if requested.

 

DATE DUE

DATE DUE

DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

2/05 c:/ClRC/DateDue.indd-p.15

 

SEWER PIPELINE CONDITION PREDICTION USING
NEURAL NETWORK MODELS

By

Guruprakash Kulandaivel

A THESIS

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

MASTER OF SCIENCE

Department of Building Construction Management

2004

ABSTRACT
SEWER PIPELINE CONDITION PREDICTION USING
NEURAL NETWORK MODELS
By

Guruprakash Kulandaivel

With an aging underground infrastructure, ever-encroaching population areas and
increasing economic pressures, the burden on the municipal agencies to efﬁciently
prioritize and maintain the rapidly deteriorating underground utilities is increasing.
Accurate forecasting of pipeline performance is essential for prioritizing and risk
management of the underground infrastructure. The essential function of a pipeline asset
management system is to consider the pipeline maintenance and improvement needs and
to arrive at the program of optimal rehabilitation, replacement, and maintenance. Hence,
the development of a pipeline condition prediction model will be indispensable to the
concerned authorities in prioritizing the care and rehabilitation of pipelines, and in
pipeline asset planning and management. This research developed an Artiﬁcial Neural
Network (ANN) model for predicting the condition of sewer pipes based on the historic
condition assessment data. The neural network model was trained and tested with
acquired ﬁeld data. The developed model is intended to aid in identifying the distressed
segments of the overall sewer pipeline network using a set of known input values. These
can then be directed toward assessing and prioritizing the maintenance measures needed

to prevent accelerated future distress and eventual failure of sewer pipes.

ACKNOWLEDGEMENT

I would like to express my sincere gratitude to my mentor and major advisor Dr.
Mohammad Najaﬁ for his continuous support, outstanding guidance, and for Offering
many learning opportunities. Most importantly I would like to thank him for his friendly
attitude, willingness to help and encouragement throughout my graduate study here at
Michigan State.

I would also like to thank the members of my advisory committee, Professors Dr.
Traiq Abdelhamid and Dr. Ghassan Abu-Lebdeh for offering their valuable time,
guidance, and inspiration. Their insightful comments and suggestions were signiﬁcant for
the ﬁnalization of this thesis.

My Special thanks to Dr. Tom Iseley, Interim Deputy Director, Buried Asset
Management Institute (BAMI), City of Atlanta, for his support and for directing me to the
most appropriate source of data. I would also like to acknowledge the help by Mr. John
Grifﬁn, Director of the BAMI, City of Atlanta, for providing necessary data and
insightful expertise without which this work would not have been possible.

I would like to thank the City of Detroit and City of Jackson, M1, for their interest
and willingness to participate in this research. I hope this thesis will be a guideline for
these municipalities to develop models with their own data in the future.

Many thanks go to Ms. Cathy Morrison, Ms. Vickie Lovejoy and Amanda
Simpson of the Michigan State University’s Construction Management Program for their

support and assistance throughout my graduate program.

iii

I would like to thank my fellow students Yu (Roy) Luo, KyounOk Kim, Thanveer
and Boo for all the fun and helping me in times of need.

I would like to express my indebtedness to my parents, V.V. Kulandaivel and
Poongodi Kulandaivel, and my sister Shuba Kulandaivel for their unconditional support
and love they have Offered my throughout my life. They are the reason I have succeeded

to where I am now and for them I dedicate this thesis.

iv

TABLE OF CONTENTS

LIST OF TABLES .....................................................................
LIST OF FIGURES .....................................................................
CHAPTER 1 — INTRODUCTION
1.1 Background and Overview ...................................................
1.2 Problem Statement ............................................................
1.2.1 State of the Sanitary Sewer ..........................................
1.3 Objectives and Methodology ...................................................
1.4 Scope of the Thesis ............................................................
1.5 Organization of the Thesis ...................................................

CHAPTER 2 —- LITERATURE REVIEW

2.1

2.2
2.3
2.4

2.5

Types of Pipes for Sewer Applications .....................................
2.1.1 Metallic Pipes .........................................................
2.1.2 Cement-Based Pipes ................................................
2.1.3 Clay Pipes ............................................................
2.1.4 Plastic Pipes ............................................................

Sewer Condition Assessment ................................................

Structural Condition Rating of Sewers
Pipeline Deterioration

2.4.1

2.4.1.1 Structural Defects ..........................................
2.4.1.2 Operational Defects ..........................................
Pipeline Deterioration Mechanisms ..........................................
2.5.1 Age of Sewer .......................................................
2.5.2 Sewer Size ............................................................
2.5.3 Sewer Section Length ..............................................
2.5.4 Sewer Gradient ...................................................
2.5.5 Sewer Joint Type ...................................................
2.5.6 Sewer Depth .........................................................
2.5.7 Surface Loading and Surface Type .................................
2.5.8 Frost Heave ............................................................
2.5.9 Frost Load ............................................................
2.5.10 Sewage Characteristics ............................................
2.5.11 Soil-Pipe Interaction ................................................
2.5.12 Pipe-Wall Temperature Gradients .................................
2.5.13 Corrosion ............................................................
2.5.14 Differential Pipe Temperature ....................................
2.5.15 Soil Type ............................................................
2.5.16 Soil pH ............................................................

ooooooooooooooooooooooooooooooooooooooooooooooooooooooo

Modes of Pipeline Deterioration ...................................

2.5.17 Groundwater Level ................................................... 54

2.5.18 Overburden Pressure ................................................ 54

2.5.19 Temperature ......................................................... 55

2.5.20 Precipitation (Snow/Rain) ........................................... 56

2.6 Pipeline Deterioration Models .............................................. 56

2.7 An Overview of Existing Deterioration Models ........................... 59

2.7.1 Statistical Models .................................................... 59

2.7.2 Physical Models ................................................... 71

2.8 Sewer Management System ................................................... 72
2.9 Application of Neural Networks in Pipeline Management and

Prediction Modeling ............................................................ 72

2.10 Summary and Conclusions ................................................... 74

CHAPTER 3 — NEURAL NETWORK METHODOLOGY AND APPLICATION

3.1 Artiﬁcial Intelligence and Neural Networks ................................. 75
3.2 Neural Networks and Statistical Modeling: A Comparison ............... 79
3.3 The Neural Network Algorithm .......................................... 82

3.3.1 Single Neuron ............................................................ 83
3.4 Backpropagation Neural Network (BPNN) ................................. 84

3.4.1 BPNN Modeling ............................................................ 86
3.5 Neural Networks and its Application in Pipeline Condition Prediction 89
3.6 Model Input Parameters .................................................... 91
3.7 Modeling Methodology ................................................... 92

3.7.1 Neural Network Design Process — An Overview ..................... 94
3.8 Summary and Conclusion ................................................... 95

CHAPTER 4 — DATABASE REVIEW AND PROCESSING

4.1 Data Acquisition ............................................................ 96
4.1.1 Background Information on Atlanta’s SSES Efforts ............... 97
4.1.2 Condition Assessment .......................................... 98
4.1.3 Condition Summary of Sewer Group 1 ........................ 99
4.2 Data Collection and Preprocessing .......................................... 103
4.2.1 Parameter Collection and Analysis ................................. 104
4.3 Software Selection for Data Preprocessing ................................. 105
4.4 Data Analysis ................................................................... 105
4.4.1 Interpretation of Results ............................................. 114
4.5 Database Transformation ................................................... 1 16
4.6 Summary and Conclusions ................................................... 118

CHAPTER 5 - MODEL DEVELOPMENT

5.1 Data Subdivision ............................................................ 119
5.2 Soﬁware Selection ............................................................ 120
5.3 Neural Network Design, Training and Testing ........................ 122

vi

5.4
5.5
5.6
5.7

5.3.1 Framework for Neural Network Development ............... 122

5.3.2 Neural Network Training and Testing ........................... 123
5.3.3 Detailed Overview of Model Development ......................... 128
5.3.4 Selection of Optimal Number of Hidden Neurons ............... 129
Model Performance Enhancement ........................................... 135
Weightage of Individual Parameters ............................................ 141
Results of the ANN Modeling Effort ....................................... 142
Summary and Conclusion 143

CHAPTER 6 — SUMMARY, CONCLUSIONS AND RECOMMENDATIONS

6.1 Summary ..................................................................... 144
6.2 Limitations of the Research ................................................... 145
6.3 Conclusions ..................................................................... 146
6.4 Recommendations for Future Work .......................................... 147
APPENDIX A ..................................................................... 151
APPENDIX B ..................................................................... 156
BIBLIOGRAPHY ..................................................................... 1 7 1

vii

LIST OF FIGURES

Figure 1.1 - Pipeline Asset Management Structure (McDonald et a1. 2001) ........ 5
Figure 1.2 - Conceptual Sewer Management Plan .................................. 8

Figure 1.3 - Optimal Renewal of Sewer Pipe with Low Cost of Failure
(Makar et a1. 2000) ..................................................................... 12

Figure 1.4 — Thesis Methodology ................................................... 14

Figure 2.1 - Single Corrosion Pit at the Outer Surface of Grey Cast Iron Pipe 20

Figure 2.2 - Combined Corrosion/Structural Failure of Grey Cast Iron

Pipe: (left) Blown Section; (right) Circumferential Fracture ....................... 20

Figure 2.3 - Combined Degradation/Structural Failure of Asbestos
Cement Pipe (Longitudinal Fracture) .......................................... 22

Figure 2.4 - Combined Degradation/Structural Failure of Asbestos

Cement Pipe (Complex Fracture) ................................................... 23
Figure 2.5 - Cracked Vitriﬁed Clay Pipe (NASSCO 1996) ..................... 23
Figure 2.6 - Brittle Fracture of a PVC Pipe .......................................... 25
Figure 2.7 - Rupture of a Polyethylene Pipe ....................................... 25

Figure 2.8 — Various Defects during the Life of Sewer Pipe (N ASSCO 1996) 31

Figure 2.9 — Typical Condition Deterioration Curve .............................. 36
Figure 2.10 - Pipeline Interactions Leading to Failure (O’Day et al. 1986) ...... 40
Figure 2.11 - Bath Tub Curve of Sewer Pipe Performance with Age ........... 44
Figure 3.1 - Anatomy of a Neural Network ....................................... 83
Figure 3.2 - Schematic Diagram of an Artiﬁcial Neuron ........................ 84

viii

Figure 3.3 - A Three-Layer Backpropagation Neural Network ............... 85
Figure 4.1 - Map Showing the Project Study Area (Sewer Group 1) .......... 98

Figure 4.2 - Proportion of Structural Deﬁciencies Observed from SSES
Inspections of SGl ..................................................................... 100

Figure 4.3 - Proportion of Service Condition Deﬁciencies Observed from

SSES Inspections of SGl ............................................................ 101
Figure 4.4 — Representation of the Sewer Material Distribution ............... 107
Figure 4.5 — Representation of the Sewer Age Group Distribution ............... 108
Figure 4.6 — Representation of Sewer Size Distribution ........................ 108
Figure 4.7 — Representation of the Sewer Depth Distribution .................. 109
Figure 4.8 - Representation of the Sewer Condition Distribution ............... 109

Figure 4.9 (a & b) — Representation of Sewer Age - Condition Relationships
for CO and VC Pipes ................................................................. 110

Figure 4.10 (a & b) — Representation of Sewer Depth — Condition
Relationships for CO and VC Pipes ................................................. 111

Figure 4.11 (a & b) — Representation of Sewer Size — Condition
Relationships for CO and VC Pipes ................................................ 112

Figure 4.12 (a & b) — Representation of Sewer Length — Condition
Relationships for CO and VC Pipes ................................................... 113

Figure 4.13 (a & b) — Representation of Sewer Gradient — Condition

Relationships for CO and VC Pipes ................................................... 114
Figure 5.1 - Database Subdivision ................................................... 120
Figure 5.2 — Neural Network Design .............................................. 121
Figure 5.3 - The Procedure for Neural Network Development .................. 123
Figure 5.4 - Schematic Architecture of the Neural Network ................... 125

ix

Figure 5.5 (a & b) — Connection Weights Histogram and Network Progress

in Training ..............................................................................
Figure 5.6— Neural Network Modeling Process .................................
Figure 5.7 - View of the NetMaker Data Processing ........................

Figure 5.8 — Graphical Representation of Training and Testing Errors for

different Architectures ............................................................
Figure 5.9 — Neural Network Training Progress .................................
Figure 5.10 — Training Results of Model #7 ......................................
Figure 5.11 — Testing Results of Model #7 .....................................

Figure 5.12 - Plot showing the Predicted and the Actual Condition of the
Testing Sample (Model #7) ..........................................................

Figure 5.13 —The Model after Training Maximum Facts ........................
Figure 5.14 —The Trained Model while Testing .................................
Figure 5.15 —Network Training Statistics for Model ........................

Figure 5.16 — Plot showing the Predicted and the Actual Condition of the
Testing Sample after 19945 Runs ...................................................

Figure 5.17 — Plot showing the Predicted and the Actual Condition of the
Testing Sample (40995 Runs) ...................................................

Figure 5.18 — Network Weight Matrices (40995 Runs) ........................

Figure 6.1 — Conceptual Integration of ANN Model ........................

LIST OF TABLES

Table 2.1 - Current Sewer Inspection Techniques — A Comparison 26
Table 2.2 - Sewer Pipe Structural Condition Evaluation ........................ 29

Table 2.3 - Structural Distress Conditions Included in the Evaluation of
Sewer Segments (WEF 1994) ................................................... 32

Table 2.4 - Condition for Various Levels of Distress in Sewer Pipes ........... 34

Table 2.5 - Five Degradation Sequences and their Severity Levels with
Abbreviations (Kathula 2001) ................................................... 35

Table 2.6 - Condition Coding System Incorporating with a Condition
Rating System (Mehle et al. 2001) ................................................................... 37

Table 2.7 - Typical Defects in Sewers Pipes (Davies et al. 2001) ............... 41

Table 2.8 - Summary of Statistical Prediction Models for Water and

Wastewater Pipelines (Kleiner and Rajani 2001) ................................. 56
Table 3.1 — NeuralWare Modeling Comparison ................................. 82
Table 3.2 — Ideal Input Parameters for Model Development ................... 91
Table 4.1 — Mainline Structural Defect Summary (City of Atlanta) ............. 101
Table 4.2 - Variables used for Neural Network Modeling ........................ 106

Table 4.3 - Format of the Transformed Database after the Selection of

Relevant Parameters ............................................................ 1 17
Table 5.1 - BPNN Architecture ................................................... 124
Table 5.2 — Training and Testing Errors of Different BPNN Architectures ...... 131
Table 5.3 — Network Architecture and Speciﬁcations of Model #7 ............. 133
Table 5.4 — Summary of Training and Testing Results (Model #7) 134
Table 5.5 — Testing and Training Network Statistics ............................ 137

xi

Table 5.5 — Testing and Training Network Statistics ............................ 139

Table 5.6 - Excluded Parameters and the Resulting Errors Generated
by the Model .............................................................................. 141

Table 6.1 - Recommended List of Parameters that needs to be incorporated in
Future ANN Models ............................................................ 148

xii

CHAPTER 1

INTRODUCTION

1.1 BACKGROUND AND OVERVIEW

The underground infrastructure systems span thousands of miles and form a signiﬁcant
part of the total US infrastructure. Sewer systems form one of the most capital intensive
infrastructure systems in the US and they are aging, overused, mismanaged and
neglected. Many of these systems are deteriorating and becoming more vulnerable to
catastrophic failures often resulting in costly and disruptive replacements. In spite of
recent increases in public infrastructure investments, municipal infrastructure is
deteriorating faster than it is being renewed. Study after study from the US.
Environmental Protection Agency to the American Society of Civil Engineers to the
American Water Works Association and the Water Infrastructure Network are estimating
from $150 billion to $2 trillion is needed during the next 20 years. The American Society
of Civil Engineers’ 2003 Report Card for America ’s Inﬁastructure gave wastewater
infrastructure a “D,” estimating an annual $12 billion Shortfall in funding needs
nationally. According to the American Water Works Association (AWWA), by the year
2020 the average utility will spend three times as much on infrastructure replacement as it
does today. The sewer infrastructure of the US must be assessed and upgraded to meet
the requirements of the EPA Sanitary Sewer Overﬂow Policy and the guidelines of the

Government Accounting Standards Board Statement 34 (GASB 34).

Factors such as population growth, tighter health and environmental requirements,
poor quality control leading to inferior installations, inadequate inspection and
maintenance, and lack of consistency and uniformity in design, construction and
operation practices have impacted adversely on municipal infrastructure. North
America’s water and wastewater pipelines (some of which are more than a century old)
are under daily assault from corrosion damage, moving soils, changing temperatures,
rainfall/snowfall, in-service stress and the continuous process of structural deterioration.
At the same time, an increased burden on infrastructure due to signiﬁcant growth in some
sectors tends to quicken the ageing process while increasing the social and monetary cost
of service disruptions due to pipeline failures. These environmental and operating stresses
inevitably lead to a number of pipeline failures throughout the year. Increasing concerns
over health, safety and the environment have contributed signiﬁcantly to raising the
visibility of pipeline risk management. The rapidly deteriorating old pipes and the
expansion of present network due to increasing demands require the municipalities to
prioritize the renewal, replacement and new installations of pipelines. However,
predicting and monitoring the condition of pipelines generally remains a difﬁcult task.

Maintaining and even enhancing wastewater collection systems is crucial in order
to have dependable transfer of wastewater to treatment facilities. When sewer systems
deteriorate, water from excessive inﬁltration and inﬂow (I/I) enters the system, resulting
in a decreased capacity of the sewer system as well as treatment facilities, increased
hydraulic loading at collection and treatment facilities, and consequently increased
capital and operation/maintenance costs. Therefore, it is necessary to maintain the sewer

system in a healthy condition.

Traditionally, municipalities have addressed the maintenance and operation of
sewer systems with a crisis-based approach. This practice results in the inefﬁcient use of
limited funds, causing more frequent sewer failures which end in difﬁcult and costly
rehabilitation or renewal (WEF-ASCE 1994). The cost of sewer failure, i.e., replacement
costs, disruptions, adverse publicity, and health and safety problems, could be
signiﬁcantly higher than the cost of rehabilitation and hydraulic upgrading. The major
reason for reactive approaches to sewer management is the sewer systems are most often
overlooked because they are underground infrastructure facilities whose existing
conditions are not readily visible to users. Thus, the actual problems caused by
deterioration are not evident until major failures occur. Another issue in sewer
management is the fact that the condition of these underground assets is generally not
fully documented. While condition assessment is very important in developing a
systematic procedure in effective sewer management, most cities do not have complete
documentation of sewer condition data in their management information systems. The
lack of data on the past and current condition of sewers hinders the system-wide
assessment of existing sewers, the development of prediction models, and the evaluation
of the effects of rehabilitation on sewer condition.

For underground sewer systems, without a predictive approach to
rehabilitation/renewal needs, condition assessment activities will be unfocussed and may
overlook high-risk assets. Some utilities try to avoid this by electing to conduct frequent
system-wide inspections at an unnecessarily high cost.

There have been efforts in the recent past years to develop a coordinated asset

management system to collect, analyze, and store massive quantities of pipeline related

data. Development of these new asset management systems open the door for many
advanced technologies and resources to be applied for state-of-the-art information
storage, retrieval, and management processes. The municipalities are looking beyond the
traditional reactive strategies to proactive maintenance of pipeline infrastructure, to
deliver the primary goals of a utility provider, which are reliable delivery of clean safe
water and wastewater services. Central to meeting these goals is the need for a robust
asset management plan that prioritizes the care, maintenance and improvement of
pipeline infrastructure, whilst taking into account the social and ﬁnancial risk
consequences of poor pipeline performance and failures. Rather than relying on a reactive
approach to pipeline repair and rehabilitation, it is important that municipalities develop
procedures that anticipate the need for repair. A systematic approach for the
determination of deterioration and obsolescence Of sewer systems is necessary to fully
gauge the status of these underground systems. This involves routine and systematic
sewer Structural and hydraulic condition assessments, establishment of a standard
condition rating system, and developing and updating prediction models for sewer
condition. Predictive modeling permits effective budgeting of inspection and

rehabilitation costs. Figure 1.1 depicts a typical asset management structure.

 

Inventory Database

 

 

 

 

 

 

Impact Assessment

——"l Prioritization |———

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Frequency of Next .
Inspection Inspection
7
Rehabilitation Condition Assessment

 

 

 

 

 

 

 

 

Decision-making on
Rehabilitation Actions

 

 

 

 

 

Figure 1.1 - Pipeline Asset Management Structure (McDonald et al. 2001)

In order to identify future improvement needs and perform technical-economic
analysis for each alternative, application of deterioration and/or condition prediction
models are required. Traditionally, age and material have been the only factors used in
prioritizing inspections. Infrastructure is widely distributed throughout a large
geographical area making it difﬁcult to track and maintain, and resulting in signiﬁcant
risks to the public and the environment in the event of failure. This traditional approach is
not sufﬁcient and does not take into account a deeper understanding of the variables
leading to failure, or of the impact of failures on the community and the environment. TO
be fully effective, the pipeline asset management system must have performance models

that combine the rate of deterioration and change in the actual pipeline condition

inﬂuenced by local factors such as in-service loading, pipe-soil interactions, corrosion
damage, pipe strength and resistance to stresses, depth of cover, soil corrosivity,
temperature, etc. Two aspects of information on pipeline performance are used in asset
management decision-making process: information on current condition, which is
Obtained through ﬁeld inspection, and information on future performance, which is
Obtained using deterioration models and forecasting tools. There are various types of
condition assessment tools available to evaluate the present state of pipelines. The
development of a model that can predict the condition of pipelines at any given time
could be beneﬁcially used to identify the distressed sections of the network. This indeed
can help in prioritizing the pipeline sections for further scrutiny and to implement
performance improvement measures. Up to now, physical deterioration models and
statistical models have been used to identify probabilistic condition and performance of

pipelines.

1.2 PROBLEM STATEMENT

1.2.1 STATE OF THE SANITARY SEWER

A sewer is an underground conduit or duct formed of pipes or other structures used for
the conveyance of wastewater. Sanitary sewer collection systems are an extensive and
vital part of the national infrastructure. In the United States, the average age of sewers
was reported to be 47 years, and the maximum age of greater than 100 years (Malik et al.
1997). Although major part of the deterioration is attributed to aging, there are other
factors like structural defects, hydraulic overloading, corrosion, etc. that accelerate the

rate of deterioration of pipes. Current sewer-condition information available to the asset

manager is often subjective, resulting in handicapped ﬁnancial justiﬁcation of
rehabilitation work, except for gross defects (Campbell et al. 1995). The knowledge of
how long a sewer pipe from an intact condition would degrade to cracking with
inﬁltration, and then to a more severe distress condition such as collapse, will allow
utility managers to make optimum decisions (Kathula 2001).

A major problem in assessing the condition of sewers is the lack of detailed
knowledge about pipeline degradation process. Being covered with soil, the condition of
buried pipelines cannot be directly and easily monitored. Moreover, their overall
condition changes so slowly that it appears as if they do not change at all. Conditions
assessment is the principle objective of any pipeline system inspection program. Optical
assessment of the physical attribute of the pipe must be made to establish the best
strategy for maintaining and rehabilitating the underground infrastructure. These physical
attributes include (1) inventory data deﬁning quantities, types, location of system
components, and (2) condition data describing the physical state of a facility or
component, e.g., cracking, deterioration, leakage, loss of strength, etc. (Iseley et al.
1997). Sewer system evaluation surveys (SSES) are the standard for gathering
information about the condition of sewers. These surveys include activities such as closed
circuit television (CCTV) ﬁlming, ﬂow monitoring, and manhole inspections. Performing
an SSES for the entire sewer network is an expensive and time-consuming process. The
budget constraints of most utilities allow only a portion of their sewer systems to be
investigated. Therefore it is important to prioritize these inspections to those sewers that
are likely to be the candidates for rehabilitation or repair so that the system is efﬁciently

managed as illustrated in Figure 1.2.

 

 

DEVELOP CONDITION PREDICTION MODEL

 

 

 

 

PRIORITIZE SEWERS FOR INSPECTION

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

INSPECTION
DAMAGE CLASSIFICATION
ASSESSMENT
CONSTRUCTION FUNCTION
- Condition of Wall - Discharge
- Water Coefﬁcients
Impermeability
FORECAST
Long-Term Maintenance Short-Term Periodical
Plan Maintenance Plan Measures
Repeat Inspections Project-Related - Cleaning
- Date Optimization Optimization of - Inspection
Rehabilitation Volume: Measures: - Repair
Budgets for - Repair
- Repair - Rehabilitation
- Rehabilitation - Replacement
- Replacement
FINANCIAL PLAN

 

Figure 1.2 - Conceptual Sewer Management Plan

 

It has been estimated that only 30-40% of local authorities have reasonably
satisfactory records, and 15-30% of all public sewers are not recorded at all (Read and
Vickridge 1997). According to a survey conducted by Malik et al. 1997, only 45% of the
cities use some kind of subjective criteria for repairing sewers in poor condition, 21% of
the cities base their decisions for the future upon the historical data, and with only 24% of
the cities making an attempt to predict the ﬁiture condition of the different sections of the
system for the repair and maintenance of their sewer systems. So the municipal agencies
tend to wait until there is a failure to take care of the network, after the damage is done.
Moreover, the distressed segments of the pipelines exert hydraulic stress on other parts of
the system, resulting in an expanding web of failures. This is not an ideal situation and
results in various ill effects and damages to society and the environment. Therefore, there
exists a tremendous need for developing prediction models that can give optimized
solutions for the decision-makers in order to provide uninterrupted service and to extend
sewer life.

In order to predict when the sections of the pipeline network need to be inspected
and maintained, it is necessary to predict the rate of those measurers for which criteria
have been established. A pipeline condition prediction model will provide base inputs for
pipeline inspection, maintenance and rehabilitation planning. Knowledge of pipeline
condition and performance characteristics will serve helpful in the following:

0 To inspect and rehabilitate the right pipe at the right time,
0 Determination Of the action year in which a pipe section deteriorates to the

minimum acceptable level,

- Forecasting of the future funding requirement to maintain the pipeline network at
an acceptable level,

0 Preventing major failures and risks associated with it,

o Prioritizing the segments of the network and better allocation of budget and
rehabilitation project phasing strategies.

Traditionally, statistical and physical models have been used to assess pipeline
condition, but they have been limited in their application. The use of neural network
based models to complement current models can be beneﬁcial to predict dynamically the
condition of sewers using historic data that is already available to the municipal agencies.
This type of information will enable the municipal authorities to make long-term strategic
decisions regarding pipeline maintenance, asset planning and operational management

using locally available data.

1.3 OBJECTIVES AND METHODOLOGY
Although there have been tremendous advancements in infrastructure management in the
past few years the impact of pipe degradation and failures on the ﬁnancial and service
level requirements of utilities remains signiﬁcant. To pre-empt these failures and reduce
their associated costs, planning models need to be developed to prioritize maintenance
and rehabilitation in pipeline networks (Burn et al. 2001).

The main objective of this research is to develop a pipeline condition prediction
model based on neural network algorithm, which can identify pipelines at risk of
degradation so that inspections can be prioritized. This model may aid the municipal

agencies in averting the inherent risks involved with pipeline failures by prioritizing the

10

parts Of the network that needs immediate action and optimize their limited inspection
and maintenance budget by applying resources where they are most effective. The
research is focused on developing a prediction model that will learn on historical
information to identify deterioration trends and predict future performance. The
developed model will provide adequate knowledge of condition of the assets to answer
the following questions that the municipal agencies often seek for:

I What will be the probable condition of a speciﬁc pipe — and the entire network?

I Which are the most vulnerable pipes in the network?

I How should the inspection projects be ranked?

I What is the future investments need?

I What is the optimal management of the underground sewer infrastructure asset?

An effective condition prediction model will allow the utility manager to optimize
the capital and maintenance budgets by identifying the parts of the network that are
critical and initiating further assessment of those segments that are potential candidates
for repair or renewal.

The expected cost of failure tends to increase with time due to the increase in the
probability of failure. On the other hand, the expected cost of intervention as well as
inspection and condition assessment tends to decrease over time due to discounting
(Kleiner 2001). The total cost thus typically forms a convex curve over time with t* being

the optimal rehabilitation time as illustrated in Figure 1.3.

11

Cost
(present value)

 
  
  
   
  

  

Total expected
cost

  

Expected cost
‘2’- of failure

min -
cost

Expected cost of
inspection + intervention

 

 

t" (optimal time) Time

Figure 1.3 - Optimal Renewal of Sewer Pipe with Low Cost of Failure
(Makar et al. 2000)

For the purpose of this thesis, Sewer System Evaluation Survey (SSES) data from
the city of Atlanta is considered for model development. The City of Atlanta is faced with
the typical problem of rapidly deteriorating sewer systems like most other cities in the
US. The city has developed a comprehensive plan to inspect, repair and where necessary,
replace its sanitary sewers. The city is currently able to inspect about 304 miles of a
possible 2,200 miles of local sewers for cracks, collapses and blockages as a part of their
SSET efforts. After extensive investigation and documentation of defects is completed, a
rehabilitation plan will be developed, identifying necessary sewer repairs and

replacement.

12

The speciﬁc objectives of this research are:

0 to review the existing models used to predict pipeline performance and failure
characteristics,

0 to review the City of Atlanta’s sewer pipeline condition assessment database to
identify useful pipeline performance data sources for deterioration model
development,

0 to develop a neural network model for condition prediction based on the
historical information, and

o to evaluate the performance of the neural network model with test data.

The research methodology for this thesis is represented in Figure 1.4.

13

hi I ~ﬂhﬂb .. - -nm A .m- —- an“ /.-um

1 Problem Deﬁnition

.
.
.

m... m-.- “-me _.

3 Literature Review

 

 

 

 

 

 

 

_. a... ....£.._n .._. .... -...__. - .__-EL...- ‘1‘”.-.

 

 

 

 

 

 

 

 

 

Review of the Current Review of the Modes
Pipeline Deterioration of Pipeline Failure
Models
Deﬁning the
Parameters for Pipeline
Failure

 

 

m-.. _...__.- l: _...._._._-_.

l Application of Artiﬁcial
Neural Networks in Pipeline
3 Condition Forecasting

__,_____ l... __

Data Collection for
Model Development

 

 

 

 

 

 

nar‘ _... .__.. Ml»- ‘ .
I
L

a Model Development

and Evaluation

 

 

 

Summary and
Conclusion

 

 

Figure 1.4 — Thesis Methodology

l4

1.4 SCOPE OF THE THESIS

The scope of this thesis is limited to the use of data from the City of Atlanta to develop
prediction models for their sewer systems using artiﬁcial neural networks. The
development of this model and its accuracy will rely heavily on the quality and quantity
of their historical and recorded data, condition assessment records and the extent of

detailed records of their underground sewer assets.

1.5 ORGANIZATION OF THE THESIS

Chapter 1 presents the background, nature of the problem and the objectives of this
thesis. Chapters 2 and 3 present an elaborate review of the current prediction models,
pipe failure modes and failure mechanisms, pipeline condition assessment techniques and
condition ratings, overview of artiﬁcial neural networks and their application in pipeline
condition prediction and the proposed methodology of the thesis. Chapter 4 presents the
data collection, assimilation and the modeling methodology. Chapter 5 presents the
neural network model development and summary Of results. Chapter 6 presents the thesis
summary, conclusions and recommendations for future work. The bibliography chapter

contains all the references and other related resources.

15

wr— 4-‘~:‘-~-A-"--' —- --'

 

CHAPTER 2

LITERATURE REVIEW

This chapter presents the background of sewer systems and builds upon it to describe the
motivation for this thesis. A detailed literature review is presented in this chapter that
covers pipeline management trends, review of the modes of pipeline failure, deterioration
models for pipes and a comprehensive analysis of the parameters that affect the
performance of pipelines. The mode and frequency of failure is dependent on the type of
pipe and the effect of environmental conditions. Each of these variables is discussed in

detail. The following ﬂow chart gives an outline for this chapter.

16

 

 

 

‘ CHAPTER 2
_; Literature Review
Review of Types Of Pipes Sewer Condition
for Sewer Applications 3 Assessment
2‘. .'

Structural Condition a Pipeline Deterioration

 

 

l

 

 

 

‘9-

 

 

 

Rating of Sewers Modes & Mechanisms

 

 

 

 

 

 

 

 

Sewer Management

Overview of
Deterioration Models
3 System

 

 

Application of ANN in
Pipeline Management &
Predication Modeling

 

 

 

 

 

 

2.1 TYPES OF PIPES FOR SEWER APPLICATIONS

There are several different pipe materials available for sewer systems, each with a unique
characteristic used in different conditions. Until 1850, sewers were generally constructed
using brickwork. Over time, because of aging, these sewers have suffered extensive
structural damage. Although some sewer systems still contain brick sewers, very few are
left. In the middle Of the nineteenth century, more and more clay pipes were used to build

the sewer systems. Concrete pipes were introduced during the early part of the twentieth

century. Modern sewers include polyvinyl chloride, ﬁberglass, high-density polyethylene,
ductile iron, steel and reinforced concrete. In general, pipe materials are grouped into

three categories:

- Metallics

. Cement-based
. Clay

. Plastics

The four different pipe materials that are most commonly used for sewer
applications are ductile iron, concrete, plastic, and vitriﬁed clay. Pipe material selection
considerations include trench conditions (geologic conditions), corrosion, temperature,
safety requirements, and cost. Key pipe characteristics are corrosion resistance (interior
and exterior), the scouring factor, leak tightness, and the hydraulic characteristics.

The stability of deteriorated sewers depends on the materials used for the
construction of the sewer pipe. Rigid pipe materials are usually designed to resist vertical
loading on their own, while brick sewers and ﬂexible pipe materials require side support
from the surrounding soil. Older sewers were typically constructed of vitriﬁed clay,
brick, or concrete. Presently, new materials are used such as plastic, ductile iron, steel,
reinforced concrete, and reinforced ﬁberglass. As shown in the ﬁgures below, different

pipe materials will fail by different mechanisms.

2.1.1 METALLIC PIPES
The most common types of metallic pipes are cast iron, ductile iron and steel pipes. The

ﬁrst ofﬁcial record of Cast Iron pipe installation was in 1455 in Siegerland, Germany.

18

Cast Iron pipe was introduced to the United States as early as 1817, when it was installed
in the Philadelphia water system. Today, more than 565 utilities (in the United States and
Canada) have had Cast Iron mains in continuous service for more than 100 years.
Additionally, at least 16 utilities have had Cast Iron mains in continuous service for more
than 150 years.

Ductile iron pipe (DIP) is an outgrowth of the cast iron pipe industry.
Improvements in the metallurgy of cast iron in the 1940's increased the strength of cast
iron pipe and added ductility, an ability to slightly deform without cracking. Ductile Iron
not only retains all of Cast Iron's attractive qualities, such as machinability and corrosion
resistance, but also provides additional strength, toughness, and ductility. Although its
chemical properties are similar to those of Cast Iron, Ductile Iron incorporates signiﬁcant
casting reﬁnements, additional metallurgical processes, and superior quality control
(DIPRA 2004). Corrosion control is achieved by using polyethylene encasement.

Steel pipes are versatile and has economic advantages Since it is stronger and thus
lighter for a given strength. In circumstances in which they are commonly used, they may
be susceptible to failure due to high external pressure, since their relatively thin walls
buckle easily. Steel pipes may also be more likely to be structurally damaged by
corrosion than iron due to their relatively thin walls. Under favorable conditions, the life
of steel pipes may exceed 50 years (McGhee 1991).

The main cause of deterioration in buried metallic pipelines is galvanic corrosion.
Soils of varying physical and chemical composition create galvanic potential differences
between different areas of the pipe. Under suitable soil electrolytic conditions, anodic and

cathodic areas are created, which leads to galvanic corrosion. Buried iron pipes are

19

vulnerable to anaerobic corrosion by sulfur-reducing bacteria under speciﬁc ground
conditions, whilst grey cast iron pipes are susceptible to a unique form of galvanic
corrosion, in which selective leaching of iron leaves a relatively weak graphitic network
in the pipe wall. This process is commonly referred to as graphitisation.

For example, in metallic pipes, failure can occur solely by corrosion (Figure 2.1),

or by corrosion combined with excessive loading (Figure 2.2).

 

Figure 2.1 - Single Corrosion Pit at the Outer Surface of Grey Cast Iron Pipe

 

Figure 2.2 - Combined Corrosion/Structural Failure of Grey Cast Iron Pipe: (left) Blown
Section; (right) Circumferential Fracture lhgtp://www.cmit.csiro.au/research>

20

2.1.2 CEMENT-BASED PIPES

Concrete pipes are manufactured in the form of reinforced concrete pipe (RCP), pre-
stressed concrete cylinder pipe (PCCP) or asbestos cement pipe (ACP). Generally, they
are manufactured by wrapping reinforced wire (high-tensile-strength wire in the case of
prestressed concrete) about a steel cylinder which has been lined with centrifugally
placed cement mortar. For prestressed concrete pipes, the wire is wound tightly
(prestressed) to prestress the core and is covered with an outer layer of concrete. In the
case of non-prestressed concrete, a similar pipe is manufactured without prestressing the
wire. In rare cases, where leakage is not important, plain concrete pipe may be used. A
reasonable estimate of concrete pipe service life is 75 years (McGhee 1991).

Amongst the advantages of concrete pipes, the following may be included: low
cost of maintenance, less corrosion if buried in ordinary soil or transporting non-reactive
wastes, expansion joints not normally required, and no specially skilled labor force is
required for its installation. However, it exhibits a certain tendency to leak due to porosity
and shrinkage cracks, has a low corrosion resistance in the presence of acids or alkalis,

and is generally difﬁcult to repair (Babbitt et al. 1962).

Asbestos cement pipe is a related product. It is composed of a mixture of Portland
cement and asbestos ﬁber which is built up on a rotating steel mandrel and then
compacted with steel pressure rollers into a dense homogenous structure in which a
strong bond is effected between the cement and the asbestos ﬁbers (Babbit et al. 1962,
McGhee 1991). Amongst its advantages may be mentioned the good corrosion resistance,

its light weight, and ease for making connections. Some disadvantages are the low

21

ﬂexural resistance of the pipe as a whole, low chemical resistance against petroleum

products, and may be easily damaged by excavating machinery.

Similar failure modes occur in cement-based pipes like other pipe types, but the

mechanism of degradation clearly differs (Figures 2.3 and 2.4).

 

Figure 2.3 - Combined Degradation/Structural Failure of Asbestos Cement Pipe
(Longitudinal Fracture)

Asbestos-cement and concrete pipes are subject to deterioration due to various
chemical processes that either leach out the cement material or penetrate the concrete to
form products that weaken the cement matrix. Presence Of inorganic or organic acids,
alkalis or sulphates in the soil is directly responsible for concrete corrosion. In reinforced
and pre-stressed concrete, low pH values in the soil may lower the pH of the cement
mortar to a point where corrosion of the prestressing or reinforcing wire will occur,

resulting in substantial weakening of the pipe (Dom et al. 1996).

22

 

Figure 2.4 - Combined Degradation/Structural Failure Of Asbestos Cement Pipe
(Complex Fracture) (http://www.cmit.csiro.au/researchi

2.1.3 CLAY PIPES

Vitriﬁed clay pipes are composed of crushed and blended clay that are formed into pipes,
then dried and ﬁred in a succession of temperatures. The ﬁnal ﬁring gives the pipes a
glassy ﬁnish. Vitriﬁed clay pipes have been used for hundreds of years and are strong,
resistant to chemical corrosion, internal abrasion, and external chemical attack. They are
also heat resistant. These pipes have an increased risk of failure when mortar is used in
joints because mortar is more susceptible to chemical attack than the clay. Other types of
joints are more chemically stable. It has been shown that the thermal expansion of

vitriﬁed clay pipes less than many other types (such as DIP and PVC).

 

Figure 2.5 - Cracked Vitriﬁed Clay Pipe (NASSCO 1996)

23

2.1.4 PLASTIC PIPES

Plastics are, in general, synthetic resins of high molecular weight, polymerized from
simple compounds by heat, pressure, and catalysis. Plastics used in the manufacture of
pipes belong principally to polyvinyl chloride (PVC) and cellulose acetate types.

Plastic pipes are cost effective and also have other advantages such as immunity
to corrosion due to chemicals commonly found in the vicinity of buried sewer systems,
freedom from damage due to freezing of ﬂuids inside them, ease of bending and joining,
adequate strength, resistance to shock, resilience and ﬂexibility. PVC pipes do not
deteriorate under attack from bacteria and do not serve as a nutrient to micro-organisms,
macro-organisms of fungi. Amongst their disadvantages are a low resistance to heat,
inability to conduct electrical current (which can constitute as an advantage too, by
making PVC pipes immune to electrolytic corrosion), high coefﬁcient of expansion, and
diminishing tensile resistance with an increase of temperature (Uni-Bell 1984).

Plastic pipes function effectively at temperatures between 32° to 90° F. with an
extreme temperature drop (below freezing, for example) PVC pipes loose impact strength
and become more brittle. Conversely, with an increase in temperature, PVC pipes loose
tensile strength and stiffness (Uni-Bell 1984).

Polyethylene (PE) is a thermoplastic material produced from the polymerization
of ethylene. PE plastic pipe is manufactured by extrusion in sizes ranging from 2" to 63".
PE is available in rolled coils of various lengths or in straight lengths up to 40 feet.

Generally small diameters are coiled and large diameters (>6" OD) are in straight lengths.

Whilst plastic pipes (such as PVC, polyethylene, etc.) are relatively 'young', their

failure mechanisms must also be understood to forecast future performance. As shown in

24

Figures 2.6 and 2.7, in the absence Of any obvious signs of degradation, fracture failure
can still occur in the ﬁeld. In general, failures in plastic pipes can be split into three
categories; plastic collapse, buckling and brittle fracture. Since the design pressures for
plastic pipes are based on the yield strength Of the pipe material, plastic collapse is rarely
seen in practice. Buckling failures, which result in local inversion Of the pipe
circumference, occur under high external loads when the ratio between pipe wall
thickness and diameter is below a critical value. As with plastic collapse, good design
practice limits the number of buckling failures observed in service. The majority of

failures reported in plastic pipes occur by brittle fracture.

 

Figure 2.6 - Brittle Fracture of a PVC Pipe Figure 2.7 - Rupture Of a Polyethylene Pipe

TO account for these various failure modes, pipeline deterioration models have to

be developed to estimate future deterioration rates.

2.2 SEWER CONDITION ASSESSMENT

A reliable condition assessment of a sewer system is essential for its maintenance and for

decisions regarding its rehabilitation. The City Of Atlanta has currently inspected

25

approximately about 304 miles of the total 2,200 miles of sewer system using various
condition assessment techniques. Considering the quantity and length of the sewer pipes,
inspection work is often intensive, requiring the collection of voluminous information. It
is therefore critical that the inspections are performed on the sections of sewer pipes that
are considered to be in the worst potential condition. The model developed in this
research will attempt to identify sewer groups in the network that are most vulnerable to
deterioration to facilitate prioritization for physical inspections. There are various
condition assessment techniques that are used for sewer inspection and can been
classiﬁed into three different groups (Makar 1999). The ﬁrst group, including
conventional CCTV and advanced SSETTM examinations, are techniques that determine
the condition of the inside surface of the sewer. The second group exarrrines the overall
condition of the sewer wall and, in some cases, the soil around the pipe. Finally, the third
group detects speciﬁc problems within or behind the sewer wall. Table 2.1 summarizes
the different condition assessment techniques and their utilization.

Table 2.1 - Current Sewer Inspection Techniques — A Comparison (Makar 1999)

 

 

 

Technique Where to use What will be found

 

Inspection of the Inner Surface

 

Surface cracks, visible
Empty pipes, partially ﬁlled deformation, missing
Conventional CCTV pipes above the water bricks, some erosion, visual
surface indications of

exﬁltration/inﬁltration

 

Pipes with less than 50 m.

 

 

Stationary CCTV distance between manholes As CCTV
. . Pipes where deformation is Better deformation
Light line CCTV measurements + CCTV

an issue
results

 

 

 

26

 

 

Technique

Where to use

What will be found

 

Computer Assisted CCTV

As CCTV, currently small
diameter pipes only

As CCTV, but with
quantitative measurements
of damage

 

SSETTM

Pipes of diameter ranging
from 8 - 24 inches

As CCTV, but with higher
sophistication and accuracy.
Can measure deformation
of pipes

 

Laser Scanning

Partially ﬁlled pipes, empty
pipes

Surface cracks,
deformations, missing
bricks, erosion losses

 

Ultrasound

 

Flooded pipes, partially
ﬁlled pipes, empty pipes

 

Deformation measurements;
erosion losses; brick
damage

 

Inspection of Pipe Structure and Bedding Condition

 

Microdeﬂections

Rigid sewer pipes

Overall mechanical strength

 

Natural Vibrations

Empty sewer pipes

Combined pipe and soil
condition, regions of
cracking, regions of

Exﬁltration

 

Impact Echo

 

Larger diameter, rigid
sewers

 

Combined pipe and soil
condition, regions of wall
cracking, regions of
exﬁltration

 

Inspection of Bedding

 

 

Ground Penetrating Radar

 

Inside empty or partially
ﬁlled pipes

 

Voids and Objects behind
pipe walls, wall
delaminations, changes in
water content in bedding
material

 

2.3 STRUCTURAL CONDITION RATING OF SEWERS

The condition rating which follows sewer evaluation is used to objectively determine the

current condition of sewers. A rating system that minimizes subjective evaluation and is

27

 

repeatable can be effectively used to predict future condition. It is acknowledged that it
does not make sense to develop a sophisticated condition rating system if the
deterioration process of a sewer structure is not fully understood, as is the case when new
methodologies or materials are involved. However, comprehensive and objective rating
systems can be developed for the most common sewer pipe materials and when adequate
historical performance records are available.

Most rating systems are based on assessment of structural conditions with little
consideration of hydraulics and Ill condition, because hydraulic and M conditions cannot
be easily evaluated. They require hydraulic modeling and Simulations (which include
comprehensive input data) and in-depth investigations of M, which can be expensive.

In the area of sewer management, there is no standard procedure to develop a
condition rating of sewer pipes. While a standard procedure for developing a
comprehensive sewer condition rating does not exist, several methods of sewer condition
rating (for brick and concrete/clay sewers) found in the literature have been reviewed in

order to gauge the status of condition assessment methodologies for sewer systems.

Water Research Center (WRc). The Sewer Rehabilitation Manual (WRc 1983) discusses
the development of the structural rating system for concrete and brick sewers in the UK.
The rating system involves three levels of structural condition. Each structural defect
found in a concrete pipe is numerically scored based on the severity of the defect and the
number of defects recorded in a pipe. These defects include: Open joint, displaced joint,
cracked, fractured, broken, deformed, and collapsed. For brick sewers, the defects are:

mortar loss, displaced bricks, missing bricks, surface damage, fractured, and dropped

28

invert. The inspector is provided with pictorial descriptions to determine the type and
severity of each defect.

For example, each “circumferential crack” found in a sewer pipe is assigned a
score of 1. “Longitudinal” and “Multiple” cracks are given scores of 2 and 5 (per crack),
respectively. A Single collapsed pipe is scored 165, etc.

The scores of all defects found in a pipe are then compiled to calculate “the peak
score” accumulated in any one-meter (3.18-ﬁ) length. Additionally, “the total score” and
“the mean score” for the entire length of sewer from upstream to downstream manholes
are calculated. Based on these three scores, sewers are rated as grade 1, 2, or 3, where
grade 3 represents the worst structural condition.

By considering the condition of the entire length of a sewer line from upstream to
downstream manholes, sewer lines Of different total lengths but similar scores are not
equally rated. Consequently, shorter lines with more serious defects will not be rated
below a longer line with less serious defects.

The pipeline assessment codes were developed in the United States by National
Association of Sewer Service Company (NASSCO) with the collaboration of Water
Research Center (WRc). Table 2.2 describes the various structural condition distress
terms proposed by NASSCO. Few of the types of defects encountered in sewer pipes are
shown in Figure 2.8 below.

Table 2.2 - Sewer Pipe Structural Condition Evaluation (NASSCO 1996)

 

Pipe Condition V Description

 

Complete loss of structural integrity of the pipe due to
Collapsed pipe fracturing and collapse of the pipe walls. Most of cross-
section area is lost to ﬂow.

 

 

Structural cracking with
Deﬂection

 

Pipe wall displacement plus cracks described by:

29

 

 

 

 

 

Pipe Condition Description

Longitudinal Defect runs approximately along axis of sewer.
Circumferential Defect runs approximately at right angles to the axis of sewer.
Multiple Combination of both longitudinal and circumferential defects.
Slab—out A large hole in the sewer wall with pieces missing.

Sag The pipeline invert drops below the downstream invert.
Structural cracking Sewer wall cracked longitudinally, circumferentially, or

without deﬂection

multiple, but not displaced.

 

 

 

Cracked joints The spigot and /or bell of a pipe is cracked or broken
Open Joints Adjacent pipes are longitudinally displaced at the joint
Holes A piece of a pipe wall or joint is missing.

 

Root intrusion

Tree or plant roots have entered the sewer through an
opening in the pipe wall or joint

 

Protruding joint material

Joint sealing material or gasket is displaced into the sewer
from its original location

 

Corrosion

Condition 1

Condition 2
Condition 3

When the cementitious pipe material shows evidence Of
deterioration illustrated by the following stages:

The pipe wall surface shows irregular smoothness, i.e. wall
aggregate is exposed

The reinforcing steel is exposed.

The reinforcing steel is gone and /or the pipe wall is no longer
intact revealing the surrounding soil.

 

Pulled joint

Adjacent pipe joints are deﬂected beyond allowable
tolerances so that the joint is Open.

 

Protruding lateral

A service outlet or pipe section that protrudes or extends into
the sewer varying in magnitude.

 

Vertical displacement

The spigot of the pipe has dropped below the normal joint
Closure

 

 

Depth of cover

 

The amount of soil covering the top of the pipe.

 

30

 

 

 

 

i\}l RIIQ I

Diliit7Illi - 'nV.ll’-ﬂl{

  

J

m S

Fracture - Circumferential (CF)

   

  

\

I W Collapse

   

 

Figure 2.8 — Various Defects during the Life of Sewer Pipe (NASSCO 1996)

31

 

Water Environment Federation - American Society of Civil Engineers. WEF-ASCE
(1994) suggests assigning an importance factor to each condition evaluation criteria for
the structural condition of brick and concrete/clay sewers. The structural condition of
brick sewers involves the following aspects: sags, vertical deﬂection and cracks, missing
bricks, lateral deﬂections, root intrusion, missing mortar, loose bricks, protruding lateral,
soﬁ mortar, and depth of cover. Concrete and clay sewer structural condition evaluation
criteria include: collapsed pipe, structural cracking with deﬂection (longitudinal,
circumferential, or both), slab-out sag, structural cracking without deﬂection, cracked
joints, Open joints, holes, root intrusion, protruding joint material, corrosion, pulled joint,
protruding lateral, vertical displacement, and depth of cover.

The sewer degradation is broadly classiﬁed into ﬁve degradation sequences. All
of the sequences started with an intact pipe and progressively degraded starting with the
distinct sequences of ( 1) cracks, (2) open joints, (3) displaced joints, (4) corrosion, and
(5) deformation, and ended in collapse. In each degradation sequence, there are various
severity levels of each distress before it reaches the collapse from the initial intact
condition. Table 2.3 contains a brief description of the degradation sequences.

Table 2.3 - Structural Distress Conditions Included in the Evaluation of Sewer Segments

 

 

 

 

 

 

(WEF 1994)
Structural Condition Description
Intact Best possible sewer condition
Separation of pipe materials that runs longitudinally
Crack . . .
or crrcumferentrally along of the sewer pipe
Open Joint Adjacent pipes are longitudinally displaced at the
jornts
Displaced Joint The pipe is not concentric with the adjacent pipe
. The cementious pipe material that shows evidence of
Corrosron

 

 

 

deterioration from chemical action. The pipe wall

 

32

 

Structural Condition Description

surface shows irregular smoothness and aggregate on
the cementitious material in the pipe is exposed

 

 

Deformation Original cross-section of the sewer is altered

 

There is complete loss of structural integrity of the

Collapse pipe. Most Of the cross-sectional area is lost.

 

 

 

 

Depending on the extent of the condition throughout a given sewer reach, a
“minor,” “moderate,” or “severe” multiplier factor, such as 1, 2, or 3, respectively, is
used. The overall numerical structural condition is then determined by calculating the
total score. Based on how likely it is the sewer will collapse, the internal condition rating
factor for overall structural condition can be determined. Sewers in rating 5 are in the
most serious condition. This rating can be adjusted based on external factors such as soil

types, surcharge, water table and ﬂuctuation, and trafﬁc condition.

Kathula (2000). Kathula (2000) proposed a structural condition rating system which
involves twenty levels of structural conditions. The various levels of sewer conditions are
based on the degree and the combination of structural defects commonly found in sewer
pipes. The structural defects considered in the evaluation include: intact (or no defects),
cracked, open joint, displaced joints, corrosion, holes, deformation, and collapse. Each
defect is then rated into three severity levels: low, medium, or high. The twenty

conditions of defects in rating structural conditions are listed in Table 2.4.

33

Table 2.4 - Condition for Various Levels of Distress in Sewer Pipes (Kathula 2000)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Condition Distress and Level
Number

1 Intact
2 Tight Crack (TC)
3 Open Crack + Inﬁltration Light (OC+IL)
4 Open Joint Light + Inﬁltration Light (OJL+IL)
5 Multiple Open Crack + Inﬁltration Light (MOC+IL)
6 Open Joint Medium + Inﬁltration Light (OJM+IL)
7 Corrosion Light (CL)
8 Multiple Open Crack + Small number of Holes (MOC+H1)
9 Open Joint Medium + Inﬁltration Medium (OJM+IM)
10 Displaced Joint Medium + Inﬁltration Medium (DJM+IM)
11 Corrosion Medium (CM)
12 Open Joint Severe + Inﬁltration Medium (OJ S+IM)
13 Displaced Joint Large + Inﬁltration Medium (DJL+IM)
14 Deformation Low (DL)
15 Corrosion Severe (CS)
16 Open Joint Severe + Inﬁltration Severe (OJ S+IS)
l7 Displaced Joint Large + Inﬁltration Severe (DJL+IS)
18 Corrosion Severe + Large number of Holes (CS+H2)
19 Deformation Severe (DS)
20 Collapse (X)

 

 

 

Each distress type has one, two, or three levels of severity, based upon the impact
that the defect has on the continued service of the sewer pipe. The three levels of severity

are I

34

 

1. Low severity level: Functionality is slightly impaired. The defect produces
little or no effect on the surrounding environment. Preemptive work in these
sewers would not be cost effective unless numerous failures occur in a short
pipe length.

2. Medium severity level: Functionality is signiﬁcantly impaired. Repair of these
failures has a signiﬁcant but not critically high cost.

3. High severity level: Functionality is seriously impaired. The cost of failure
under this condition would be high and affect the surrounding environment to
a great extent.

The following Table 2.5 shows the ﬁve degradation sequences and their severity
levels used by Kathula 2001 based on the degree and the combination of structural
defects found commonly in sewer pipe segments.

Table 2.5 - Five Degradation Sequences and their Severity Levels with Abbreviations

 

 

 

(Kathula 2001)
Degradation Sequence Severity Levels with their abbreviations
Cracks Tight Crack (TC)
Open Crack (OC)

 

Multiple Open Crack (MOC)

 

Multiple Open Crack + Small no. of Holes (MOC+H1)

 

 

 

 

Open Joints Small Open Joints (SOJ)
Medium Open Joints (MOJ)
Large Open Joints (LOJ)

Displaced Joints Small Displaced Joints (SDJ)

 

Medium Displaced Joints (MDJ)

 

Large Displaced Joints (LDJ)

 

 

Corrosion Light Corrosion (LC)

 

Medium Corrosion (MC)

 

Severe Corrosion (SC)

 

 

35

 

 

Degradation Sequence Severity Levels with their abbreviations

 

Severe Corrosion + Large no. of Holes (SC+H2)

 

 

Deformation Light Deformation (LD)

 

Medium Deformation (MD)

 

 

The defects can then be classiﬁed, for example as:
1 — Excellent condition, no defects present
2 — Good condition, only low risk defects present.
3 — Fair condition, pipe contains medium severity defects.

4 — Poor condition, pipe contains high severity defects and collapse is imminent.

A

I EXCELLENT

 

 

 

 

 

5 VERY POOR

 

2 GOOD §\\\\\\\
3 FAHr i ?\\\
4 POOR i ?\\

------ h----d

 

 

V

0% Percentage Effective Life Elapsed 100%

Figure 2.9 — Typical Condition Deterioration Curve

Mehle et al. (2001) proposed a modiﬁed Vani Kathula condition coding system

incorporated with a condition rating system that is represented in Table 2.6.

36

 

Table 2.6 - Condition Coding System Incorporating with a Condition Rating System
(Mehle et al. 2001)

 

 

 

 

 

 

 

 

Defects Excellent Good Fair Poor Failure
Cracks Intact TC, OC MOC MOC+H1 Collapse
Open Joints Intact SOJ MOJ LOJ Collapse
Displace Joints Intact SDJ MDJ LDJ Collapse
Corrosion Intact --------------- LC, MC SC, SC+H2 Collapse
Deformation Intact ------------------------------ LD, MD Collapse
Rating 1 2 3 4 5

 

 

 

 

 

 

City of Atlanta Defect Coding System. The city of Atlanta has developed their own
defect coding and condition making system based on NASSCO and We standards. A
detailed list of the defect coding is given in Appendix A.

Once the condition rating has been assigned to a particular pipeline, the worst
defect present is used as an indication of the overall sewer condition rating. Although the
pipeline may not be in poor condition throughout its length, the worst condition along the

length dictates its risk of collapse.

2.4 PIPELINE DETERIORATION

Pipeline systems require constant maintenance and can become impaired for a number of
reasons. A comprehensive study performed by We (Serpente 1993) concludes that the
concept of measuring the “rate of deterioration” of sewers is unrealistic, but deterioration
is more inﬂuenced by random events in a sewer life span (a storm or an excavation

nearby) and severe defects do not always lead immediately to collapse. Sewer pipes are

37

 

prone to certain types of failures based on the type of material, physical design, age,
functionality and external and internal environment. Distress and collapse Of a sewer are
the result of the complex interactions of various mechanisms that occur within and
around the pipeline (Kathula 2000). The impact of the deterioration of the sewer system
depends upon its size, complexity, topography and service. While it is almost impossible
to predict when a sewer will collapse, it is feasible to estimate whether a sewer has
deteriorated sufﬁciently for collapse to be likely. The mechanisms of pipeline
deterioration are:

I Structural — cracks, fractures, breaks, etc.

I Hydraulic — insufﬁcient capacity, ﬂooding, debris, encrustation and grease

I Operational problems -— roots, blockages debris, maintenance procedures, etc.

Pipelines can have defects classiﬁed as built-in or long term (Najaﬁ 1995). Built-
in defects are generated during pipeline construction and represent conditions that affect
the performance of pipes after installation. Long-term defects are caused as a result of the
deterioration process. Construction-related or built-in defects can be Offsets in alignment,
joints loosely ﬁtted or loosened by vibrations, ﬂattened or ovaled pipes, sags due to
settlement, stresses caused by dynamic loadings of backﬁll, removal of trench sheathing
and pilings, overburden compaction, etc. Joints can experience the construction defects,
such as pinching of rubber gaskets, misalignment of gaskets, and squeezing due to
“overshoving” of one pipe into another. A structural failure can be a crack, break, split,
cavitation of the pipe opening, or separation at a joint (N ajaﬁ 1995).
Examples of causes of long-term pipeline deterioration are sulfate corrosion due

to sewer gases, excessive hydraulic ﬂows, structural failures, leaks and inﬁltrations, and

38

 

erosions. Bacteria in the wastewater stream convert the sulfates to hydrogen sulﬁdes
which, when released into the sewer air space, become oxidized into sulfuric acid. The
sulphuric acid is reactive to some pipe materials making it to corrode. Severe corrosion
can jeopardize the structural integrity of a pipe or manhole and lead to collapse. Any
condition of pipeline deterioration which occurs over an extended time period and is not a
result of construction practice is considered a long-term deterioration. Proper
maintenance of pipelines is essential to keep the pipeline in good health.

The state of the surrounding soil is of fundamental importance in assessing the
structural condition of a sewer. The main factors that affect the rate of ground loss
include sewer defect size, hydraulic conditions (water table, and frequency and
magnitude of surcharge), and soil properties (cohesive or non-cohesive soil). Severe
defects (larger than 4 inches), high water table (above sewer level), frequent and high
magnitude of hydraulic surcharge, and soil types (silts, silty ﬁne sands, and ﬁne sands)
can have serious effects on ground loss. Loss of Side support will allow the side of the
pipe to move outward when loaded vertically, and collapse will likely once the pipe
deformation exceeds 10%. Uneven loading of pipes due to joint displacement also

accelerates the pipe deterioration process.

2.4.1 MODES OF PIPELINE DETERIORATION

Pipeline deterioration is a complex process; many factors are responsible for their
deterioration and failure — structural, hydraulic, environmental, functional, age of the
sewer, quality of initial construction, etc. The intensity of structural failures depends on

the Size of the defect, soil type, interior hydraulic regime, ground-water level and

39

ﬂuctuation, corrosion, method of construction, and loading on the sewer. Hydraulic
failures are caused by inﬁltration and inﬂow (l/I) problems. These I/I problems reduce
the planned hydraulic capacity of sewers, increasing the potential for collapse. Figure
2.10 illustrates various kinds of internal and external forces acting on a pipe. The modes

of failure depend on the type Of environment and pipeline material.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

     

 

 

 

 

 

External , Soil

Load Aeration Characteristics
, 3 ,/

son—T " ‘f Soil G d t

Temperature TNCK —* m‘ “"24”” Moistu roun wa er

onstruction—r @m‘"' I.

‘ g or ..—'- Leakage

Contraction —— __________
Load
/ gauging ZR
LMovement Con mo" --- Construction

 

 

 

 

 

 

Figure 2.10 - Pipeline Interactions Leading to Failure (O’Day et al. 1986)

Pipe breakage is likely to occur when the environmental and operational stresses
act upon pipes whose structural integrity has been compromised by corrosion,
degradation, inadequate installation or manufacturing defects. Pipe breakage types were
classiﬁed by O’Day et al. (1986) into three categories: (1) circumferential breaks, caused
by longitudinal stresses; (2) longitudinal breaks, caused by transverse stresses (hoop
stresses); and (3) split bell, caused by transverse stresses on the pipe joint. This
classiﬁcation may be complemented by an additional breakage type i.e., holes due to
corrosion. Circumferential breaks due to longitudinal stress are typically the result of one

or more of the following occurrences: (1) thermal contraction (due to low temperature of

40

the sewage in the pipe and the pipe surroundings) acting on a restrained pipe, (2) bending
stress (beam failure) due to soil differential movement (especially clayey soils) or large
voids in the bedding near the pipe (resulting from leaks, I/I, etc.), (3) inadequate trench
and bedding practices, and (4) third party interference (e.g., accidental breaks, etc.).
Table 2.7 lists the most typical type of defects found in sewers.

Table 2.7 - Typical Defects in Sewers Pipes (Davies et al. 2001)

 

Defect Description

 

Longitudinal cracks May occur at springing level as well as at the crown and invert. A
and fractures result of excessive ‘crushing’ or ‘ring’ stress.

Cracks are diagonal and spread from the point of overload which

Tensron “ad‘s is Often a hard spot beneath the pipe.

Relative vertical movement of successive lengths of pipe causing
cracks and/or fractures due to excessive shear or bending stresses.
Most likely to occur near joints.

Circumferential
cracks and Fractures

Occurs when pieces of a cracked or fractures pipe visibly move
from their original position. Normally represents a further stage in

Broken prpes deterioration of a cracked or fractured pipe and is a very serious
defect.
. Excessive pressure inside the joint due to the expansion of the
Socket bursting . . . . . .
jomtmg material may cause a bursting failure of the socket.
Deformed pipes Occurs when a longitudinally cracked or fractured pipe loses the

support of the surrounding ground.

 

2.4.1.1 STRUCTURAL DEFECTS

Structural defects failure mechanisms include cracks and fractures in the pipeline
material that are caused by a change in the forces around a pipeline or a change in the
ability of the pipe material to resist existing forces (ASCE 1994, Serpente 1993). The
inﬁltration of groundwater through existing structural defects creates or increases the size
of voids as the inﬁltrating water carries particles from the soil into the pipeline (Delleur

1989). The weakening of this soil makes the land above the pipe vulnerable to surface

41

 

collapse. The effects of inﬁltration on void formation are made worse by the process of
exﬁltration. Exﬁltration occurs when water leaves the sewer line through structural
defects during periods of hydraulic surcharge. Surcharge wastewater can scour or loosen
more ﬁnes at the perimeter of the voids (Delleur 1989, Stein et al. 1995).

Dynamic forces that cause structural defects are large one-time events or smaller
cyclic events that occur at a variety of frequencies (daily, seasonally, etc.). Large one-
time events include periods of heavy surface construction, in-ground utility construction,
or non-construction events such as earthquakes or landslides. These events are especially
signiﬁcant when coupled with a weakened material or voids in the soil. Many surface
collapse failures are associated with degraded but functioning sewers that fail due to a
large one-time event (Delleur 1989, WRc 1986). Smaller cyclical dynamic loads include
load transfer from above ground activities, such as routine truck, machinery, and bus or
train trafﬁc or in ground movements, such as those caused by expansive soils or frost

heave.

2.4.1.2 OPERATIONAL DEFECTS

Operational Defects failure mechanism originates from an increase in demand and a
decrease in capacity. Inﬁltration and inﬂow, often referred to as M, are the two types of
demand on a sewer system. Inﬁltration increases the demand as the number of structural
defects grows. Inﬂow is the demand on the system from service connections and storm
waters (ASCE 1994, EPA 1991). A decrease in capacity is the result of a decrease in the

effective diameter of the pipeline and an increase in the roughness coefﬁcient. The

42

effective diameter is reduced by structural defects such as Open joints, broken pipe

sections, root masses or collected debris.

2.5 PIPELINE DETERIORATION MECHANISMS
This section deals in detail about various mechanisms that would affect pipeline
deterioration. There are several theories that explain the deterioration mechanisms of
buried pipelines. Various pipe deterioration modes have been identiﬁed for different
types of pipelines, and the mechanisms thought to cause such defects have been studied.
Many sewer system deteriorations are attributable to the following predominant
mechanisms:
. Deterioration due to natural aging process - lack of maintenance exacerbates age-
related deterioration.
. Deterioration of pipes and joints due to soil-pipe interaction, operating conditions
and exposure to corrosive substances.
0 Freeze/thaw cycles, groundwater ﬂow, and subsurface seismic activity that can

result in pipe movement, warping, brittleness, misalignment, and breakage.

2.5.1 AGE OF SEWER

Aging is a part of life. From the minute the sewer pipe is installed, it begins to age. A
classical survival function relating the age of the pipeline to the failure rate is denoted by
a bath tub curve as Shown in Figure 2.11. The early part of the curve shows “infantile
failure” which for pipes is representative of failure due to human factors in the actual

laying of the pipe (manufacturing faults, tend to appear during that part). A period of time

43

follows in which failure rate is generally low. When failure does occur it may depend on
many factors, such as excessive loads not designed for, or settlement. As the pipes tend
towards the end of their useful life the failure rate increases exponentially. This classic
survival proﬁle is known as the “Bath Tub” curve. The “Bath Tub” curve can be applied
to an individual pipe, a group of pipes with similar characteristics or the whole

population of a pipe network.

Comparatively high Failure probability
failure probability due increases due to approach
to construction defects. of end of useful life

\

Steady, relatively low
failure probability

/

Failure Probability

 

 

 

V

Time

Figure 2.11 - Bath Tub Curve of Sewer Pipe Performance with Age
(http://www.pir.gov.on.ca/userﬁles_/HTML/nts 2 25528 l.html)

The factors thataccelerate the aging process of the sewers are discussed in detail

in the following sections.

2.5.2 SEWER SIZE
A number of authors have investigated the relationship between sewer size and structural

stability. Studies indicate that there is a decreasing trend in pipe failure rate with

44

increasing diameter and is directly attributed to the increasing wall thickness and joint
reliability with increase in pipe diameter. Larger wall thickness gives the pipe better
structural integrity and improved resistance to corrosion failures (Kettler and Goulter
1985). Many other studies have also shown that a larger proportion of failures have
occurred on the smaller diameter pipes (Rajani et al. 1996).

Pipe size also affects the mode of failure (O’Day, 1982). Smaller diameter mains
(6 - 8 inches) often experience beam (ﬂexural) failure because of poor bedding
conditions, however crushing failures (often longitudinal) are likely to occur due to the
relative length-to-diameter ratio. Conversely, larger mains (IO-inch or greater) are likely
to experience crushing failure, but are not likely to experience beam failure (O’Day

1982)

2.5.3 SEWER SECTION LENGTH
Generally, longer sewer runs are less likely to deteriorate at a faster rate than the Shorter
ones, which may be due to the fact that longer runs means less bends in the pipe to
accumulate debris, creating blockages or damage to the pipe from standing sewage.
Another possible reasoning is that the longer runs may be more of conveyance systems
rather than collection systems, thus having fewer laterals connected to the pipes which
can weaken a pipe system.

Other potential problem area is the length to diameter ratio. Although longitudinal
bending stresses increase with increasing pipe diameter, they do so at a slower rate than
the increase in the pipe’s section modulus, hence pipes which have high length to

diameter ratios may be more likely to suffer from excessive bending stresses (Young &

45

O’Reilly 1983). Despite the fact that this issue is well documented within the literature,
there is little evidence of any numerical or statistical investigation of the effect of high

pipe length to diameter ratios having taken place.

2.5.4 SEWER GRADIENT

The slope of the pipe is found to have an impact of the condition of the pipe. For all
condition states, the steeper the slope is, the higher the possibility that pipe segments
deteriorate. This may be due to the fact that steeper pipe segments induce faster ﬂow
rates, resulting in greater possibility for damage to the inside walls or joints of pipe

segments.

2.5.5 SEWER JOINT TYPE
The main ﬁinctions of a sewer joint are a follows (ASCE 1982):

a To be water tight;

0 To be durable;

c To be resistant to root intrusion
Joint failures are leak failures where pipe joints become separated. Joint type is an issue
since the type of joint will inﬂuence the susceptibility of the pipe to speciﬁc failures. A
large part of this may be owing to the amount of ﬂexibility and lateral constraint the joint

provides, as well as the pipe joint’s actual strength and its ability to resist corrosion.

46

2.5.6 SEWER DEPTH

In investigating the effect of depth on sewer structural condition, O’Reilly et al. (1989)
found a steady decreasing defect rate to a depth of 18 feet below which, the defect rate
began increasing with depth. It was suggested that the ﬁrst occurrence probably reﬂected
the decreasing inﬂuence of surface factors such as road trafﬁc and utility/surface
maintenance activity. The second occurrence or pattern was explained by the increasing
effect of overburden pressure. Jones (1984) suggested that, in shallow sewers, the effect
of seasonal moisture variations in the soil surround may be signiﬁcant. In an analysis of
over 4400 sewer failures, Anderson and Cullen (1982) reported that 65% of all incidents
occurred at a depth of 6.5 feet or less and 25% from 6.5 to 13 feet deep, although no
indication is given of overall sewer depth distribution. Changes in cover depth may also

be important in determining a sewer’s structural stability.

2.5.7 SURFACE LOADING AND SURFACE TYPE

The location of a sewer will obviously affect the magnitude Of surface loading to which it
is subject; for sewers beneath roads the main component of such loading is likely to be
that from trafﬁc. Pocock et al. (1980) monitored the bending strain developed in a
shallow buried pipeline due to static and rolling wheel loads. The measured bending
strains were found to increase linearly with axle load, the strains for any given load
tending to decrease with increasing vehicle speed. Maximum strains were always

associated with pipes that had been deliberately poorly bedded.

47

258 FROSTHEAVE
Frost heave is deﬁned as the vertical expansion of soils caused by freezing of the soil and
ice lens formation. All underground structures require the consideration of frost heave
effects as they are capable of displacing portions or the entire underground structure.
Differential heave causes sections of pipe to experience non—uniform
displacements, and this results in forceful ﬂexural stresses. Uniform heaving may also
prove to be a problem under certain circumstances where pipe joints are not subject to
movement. Under this scenario, the pipe experiences stresses similar to a simple beam
loading, in which case the pipe will experience bending stresses. Failure of pipe joints
may be the result of the frost heave process. This may be a function of the type Of
connection, and the type of ﬁll material used between joints.
The conditions for frost heave require the following:
1. The presence of a frost susceptible soil;
2. The presence of a sufﬁcient water source, whether it is capillary or a ground
water source (for lens formation) and;
3. A ground temperature below freezing point.
With all of the above factors present, there is the potential for damage due to frost
heave. The propensity for heave of a soil under freezing conditions is affected by
properties such as grain size, rate of freezing, the availability of water, and by applied

loads.

48

2.5.9 FROST LOAD

The failure of sewer pipes during winter could be attributed to increased earth loads on
the buried pipes, i.e., frost loads. In a trench, the frost load develops primarily as a
consequence of different frost susceptibilities of the backﬁll and the sidewalls of the
trench and the interaction at the trench backﬁll-sidewall interface. Trench width,
differences in frost susceptibilities of backﬁll and trench sidewall materials, stiffness of
the medium below the freezing front and shear stiffness at backﬁll-sidewall interface play
important roles in the generation of frost loads. Thus, it is preferable to use a backﬁll
material that has a matching or lower frost susceptibility than that of the sidewall in order

to mitigate against the development of excessive frost loads.

2.5.10 SEWAGE CHARACTERISTICS

Whilst domestic sewage is generally not aggressive to the fabric of sewer system, the
quality of the sewage varies from place to place and is dependent on several factors. It
can vary from relatively weak domestic sewage, perhaps diluted with large quantities of
stonnwater or inﬁltration, to strong and potentially aggressive sewage with a high

proportion of trade efﬂuent.

2.5.11 SOIL-PIPE INTERACTION

Soil-pipe interactions are also a possible cause of pipe deterioration. The resistance of the
soil-to-pipe union is important because the shear strength of the interaction can affect the
degree of mobility of the pipeline and hence its ability to displace. In cold temperatures,

the bond between the soil and pipe indicates the amount of restraint the pipe is allowed to

49

shrink axially. A high soil-pipeline interaction will not allow the pipe to contact, and
consequently the axial stress in the pipe will increase. It is also possible that a strong
bond between the iron pipe and soil will cause excessive soil-pipe interface shear that
may cause abrasion of the pipe coating. This abrasion may lead to premature corrosion of

the pipe exterior (Yen et al. 1981).

2.5.12 PIPE-WALL TEMPERATURE GRADIENTS

For longitudinal failures, a suspected failure mechanism is the high temperature gradient
occurring across the pipe wall. If the temperature difference of the transported efﬂuent
and surrounding soil is signiﬁcant, this temperature gradient can lead to unusually high
hoop stresses, subsequently leading to failure (Habibian 1994). Longitudinal failures may
also occur in combination with the weakening of the pipe wall due to corrosion, at the
weakest portion of the main wall. Another possible cause of longitudinal failure is due to

a crushing load. This usually occurs in the large diameter pipes (O’Day 1982).

2.5.13 CORROSION

Corrosion in metallic pipes essentially occurs by an electrochemical reaction between the
outer surface of an exposed pipe and its surrounding soil environment. For corrosion to
occur, there must be a potential difference between two points that are electrically
connected in the presence of an electrolyte, in this case, the surrounding soil. With these
conditions satisﬁed, a current will ﬂow from an anodic area, through the soil to a
cathodic area, and then back through the pipe wall to complete the circuit. The anodic

area becomes corroded by the loss of metal ions to the electrolyte (Romanoff 1964).

50

Upon initiation, the corrosion process is self-sustaining (Rossum 1969), resulting in the
formation of “pits” at the outer surface of the pipe, with a range of depths and widths.
Different pipe materials have different characteristics in their reaction to corrosion.
Factors like soil acidity, resistivity, pH content, oxidation-reduction, sulphides, moisture
and aeration level have all been reported to inﬂuence corrosion rate (Romanoff 1964) and
correlations have been proposed between corrosion rates and soil electrochemical
properties (Rossum 1969).

The risk of Interior Corrosion of a pipe interior depends upon the susceptibility
of the pipe material to corrosion and the amount of corrosive chemicals in the
wastewater. Interior pipe corrosion typically occurs from the formation and release of
hydrogen sulﬁde. Hydrogen sulﬁde is formed in anaerobic conditions, such as those
found in force mains, continually surcharged gravity pipes, debris piles, or pools caused
by sagging lines. It is assumed that anaerobic conditions exists in open channel ﬂow at
the wetted perimeter and therefore a small amount of hydrogen sulﬁde is generated and
some corrosion can occur if the conditions allow for release (ASCE 1994). Hydrogen
sulﬁde gas is released in turbulent conditions. Such conditions occur at siphon outlets,
drop structures greater than 2 feet, discharge of force mains, interceptor intersections, a
change in slope and during high wastewater velocities (Hahn et al. 2000).

Exterior Corrosion depends upon the susceptibility of the pipe material to acidic
ground substances and galvanic corrosion (Delleur 1989). Acidic soils or groundwater
attack unprotected cementious or metal pipe materials, whereas stray currents in the

ground cause a galvanic corrosion with metal or metal reinforced pipes.

51

2.5.14 DIFFERENTIAL PIPE TEMPERATURE

Some literature speculates that a high differential temperature between the internal and
external pipe wall can produce high temperature gradients. Under such conditions the
inner and outer ﬁbers will be subject to different temperature drops, resulting in

differential strains and circumferential stresses.

2.5.15 SOIL TYPE

The signiﬁcance of the type of soil cannot be overlooked, as it is one of the most
important factors, having effects on almost all of the above mechanisms. Its effects on
frost heave, strength of soil-pipeline interaction, and external corrosion can be important
for many failure mechanisms.

Frost susceptibility is deﬁned as the rate at which frost penetrates the ground. It is
generally regarded as one of the most important factors in characterizing frost heave
action. Frost susceptibility is ranked greatest to least for soil types in the following order:
silt, clay, sand, and then gravel. However, methods of further quantifying and thoroughly
characterizing soils in terms of ﬁost susceptibility are not consistent. Use of frost heave
rate (inch/day), total frost heave (inch), frost heave ratio (ratio of frost heave rate to total
frost heave) and segregation potential (to depict frost susceptibility) have been suggested
(Kujala 1993). However, these types of measures are Often difﬁcult to ﬁnd, or do not
translate accurately from laboratory to ﬁeld values (Konrad and Nixon 1994). Therefore
characterization of frost susceptibility, and hence frost heaving is difﬁcult using ﬁeld

measurements.

52

The type of soil the pipe is located in is also important for the aspect of
differential heaving and thaw settlement. If a pipe is located at the interface of two
different soil types, it has been shown that each soil will experience an uneven amount Of
frost heaving, and therefore have an inﬂuence on the amount of strain experienced by the
pipe (Nixon 1994). In the same manner, thaw settlement will lead to differential stress
distributions on the pipeline.

Soil corrosivity is a soil characteristic that must be considered for external
corrosion predictions. Physical characteristic (particle size, ﬁiability, uniformity, organic
content, color, etc.) have reﬂected corrosivity, based on observations and testing. Color
has also been linked to corrosivity. Soil uniformity is important because of the possible
development of localized corrosion cells. Corrosion cells may be caused by a difference
in potential between unlike soil types, with both soils being in contact with the pipe
(Smith 1968). If it can be assumed that for a particular soil classiﬁcatiOn the approximate
uniformity coefﬁcient can be estimated, then the possibility of corrosion can be

estimated.

2.5.16 SOIL pH

In order to characterize external corrosion, it is necessary to ﬁnd parameters which
indicate the corrosivity of the soil. Soil pH is a good indicator of external corrosion since
certain pH ranges allow for different corrosion mechanisms to occur. It has also been
found that resistivity is a function of pH [Morris Jr. 1967); (Jarvis and Hedges 1994)].

For that reason, only one of the two may be required for characterization.

53

2.5.17 GROUNDWATER LEVEL

Use of the soil water content parameter is important from several aspects. As mentioned
earlier, the rate of frost heave is controlled by the availability of free water (McGaw
1972). It is also important for external corrosion.

From the perspective of frost heave, it has been stated that the availability of a
water source is one of the necessary elements required for ice lens growth. In the absence
of a nearby ground water table, focus then shifts to the availability of water present in the
soil itself, i.e., soil water content. In reality, the water content may be a possible surrogate
measure for water table depth, as water may enter the soil above by capillary suction.

From the perspective of external corrosion, soil corrosion aggressiveness has been
related to moisture content. Soils with moisture content above 20 percent (wet basis) are
thought to be particularly corrosive (Jarvis and Hedges 1994). Studies substantiate that

moisture content is a factor contributing to soil aggressiveness (Booth et al. 1967).

2.5.18 OVERBURDEN PRESSURE
Overburden pressure is thought to be important due to its ability to help characterize ﬁ'ost
heaving and soil-pipeline resistance. It can be characterized by the depth of cover and soil
density. Literature indicates that the overburden pressure is important for the rate of
heaving [(Anderson et al. 1984); (Roy et al. 1992)].

Bury depth is an important factor. From the perspective of soil-pipeline
interaction, it has been demonstrated that the frictional soil resistance is affected by pipe
diameter and bury depth (Rajani et al. 1995). Also, from the perspective of mode of

failure, larger pipes are more susceptible than smaller pipes to crushing failure. This is

54

due to bury depth, or the external loadings the pipe is subjected to (i.e. roadways, large

structures, etc.) (O’Day 1982).

2.5.19 TEMPERATURE

The effects of temperature on pipe breakage rates have been observed and reported by
many. Walski and Pelliccia (1982) suggested that pipe breakage rates might be correlated
to the maximum frost penetration in a given year. To account for the lack of frost
penetration data, they correlated annual breakage rates with air temperature of the coldest
month, using a multiple regression analysis with age and air temperature as the covariates

N (t, r) = N (weAl eBT
Where 1 = pipe age; N (to) = breaks per mile at to; T = average air temperature in the
coldest month; A, B = constants.

Newport (1981) analyzed circumferential pipe breakage data and found that
increased breakage rates coincided with cumulative degrees-frost (usually referred to as
freezing index in North America and expressed as degree-days) in the winter as well as
with very d1}I weather in the summer. He attributed the increase in winter breakage rates
to the increase in earth loads due to frost penetration, i.e., frost loads, and the summer
breakage rates to the increase in shear stress exerted on the pipe by soil shrinkage in a dry
summer. He also observed that when two consecutive cold periods occurred, the breakage
rates (in terms of breaks per degree-frost) in the ﬁrst one exceeded those of the second
one. He rationalized that the early frost “purged” the system of its weakest pipes, causing

the later host to encounter a more robust system.

55

2.5.20 PRECIPITATION (SNOW/RAIN)

Snow is indicative of the insulating effect on ground temperature, as the snow will allow
for the entrapment of heat into the ground. Rain precipitation coupled with the soil type
may be indicative of moisture content or hydraulic conductivity if these parameters are
not measured regularly. Some literature indicates that corrosion resistance is enhanced
during dry periods of the year (Smith 1968). Therefore, inclusion of this parameter may
be necessary to help characterize climatic changes as well as to infer adjustments to soil

parameters.

2.6 PIPELINE DETERIORATION MODELS
Incorporating historical condition data to develop deterioration patterns for a city’s sewer
system is pivotal in obtaining a realistic assessment of the city’s infrastructure.
Deterioration models are necessary because the determination of cost-effective
maintenance actions requires information on the current condition as well as the
anticipated future condition. While current conditions are assessed based on sewer
inspections, future conditions may only be estimated from the deterioration models.
Although pipelines are designed for a particular lifespan under standard operating
conditions, their deterioration never follows a set pattern. The process of deterioration of
pipelines is rather complex simply because there are many factors which interactively
contribute to such deterioration. Environmental interactions (soil corrosivity, ground
movement, etc.) plus exposure to transported waste quality variations and operating
abnormalities ensure that pipeline deterioration is never uniform. Eventually, through a

combination of internal and external stresses, the pipe fails. Sometimes this process is

56

accelerated when defense measures such as protective coatings are damaged or not
repaired properly. The challenge for a condition prediction model is to analyze
information on the pipe and its environment to predict, as accurately as possible, its time
to failure or probability of reaching a certain condition state. Some of the factors that
contribute to deterioration are:
Construction features

Load transfer

Standard of workmanship

Sewer Size

Sewer depth

Sewer bedding

Sewer material

Sewer joint type and material

Sewer section length (Manhole to Manhole)

Sewer connections (Laterals)

Local external factors

Surface use

Surface loading and surface type

Ground disturbance

Groundwater level

Ground conditions

Soil/backﬁll type

Root interference

57

Other factors
Age of sewer
Sewer characteristics
Maintenance methods/Frequency

The basic idea of life assessment models is to try to estimate a function for each
individual pipe that will provide the probability for that pipe surviving beyond a future
time period. Life assessment models assume that pipe lifetimes can be treated as
independently and identically distributed random variables. The objective of these
models is to estimate the probability Of failure of a pipe within a time horizon.

There are two main categories of such predictive models: Statistical and Physical.
The statistical model can further be classiﬁed into: Aggregate Type models, Multiple
Regression Type models, Probabilistic Predictive models, Counting Process models,
Non—Homogeneous Poison Process models and Events Dependent Renewal Process
models. Among the non-purely-statistical ones, physical models of corrosion are the most
applicable (Melina and Kalles 2000).

A statistical approach based on historical maintenance data and pipelines
inventory is a technique that requires an undertaking of vast amounts of pipe sampling or
condition assessments and measurement of long lengths of pipes. The physical approach
is based on the knowledge of underlying process, using engineering-based equations in

developing simulation models that can be applied in making maintenance decisions.

58

2.7 AN OVERVIEW OF EXISTING DETERIORATION MODELS

2.7.1 STATISTICAL MODELS

Aggregate Type models group together pipes that have the same intrinsic properties and
then use linear regression to establish a relationship between the age of the pipe and the
number of failures. They describe the global evolution of failures on all the pipes in the
system (Walski 1986). Shamir and Howard (1979) proposed an exponential increase with

time of the form:
Mt) = i.(t0)e"““°’ .......... (2.1)

where Mt) is the number of failures/yr/ 1000 it at time t, to is the base year for analysis,

and A is the growth rate coefﬁcient.
The advantage of these models lies in their case of implementation. Their
drawback is that pipe characteristics, previous break history and environmental variables

are not taken into account.

Counting Process models are slightly different from Aggregate Type models because
they establish the cumulated number of failures of each group of “identical” pipes as a
function of time (Andreou and Marks 1986). In counting process models the pipe failures
are assumed to occur along the time, and no assumption is made regarding the status of
the pipe aﬁer the repair is completed. With counting process models one can see the
deteriorating (or improving) trend in time of a group of “identical” pipes and the rate of

occurrence of failure of a group of pipes.

59

Multiple Regression Type models give a regression equation between the number of
years from installation to the ﬁrst break (or the number of failures) and a set of
explanatory variables such as material, internal pressure, etc. to forecast the future
number of breaks. The advantage of this approach is that it enables explicit identiﬁcation
of the categories of mains. This modeling seems better suited to the school of thought that
favors the short-time selection of individual components for maintenance operations.

According to Lawless (1982) and Kalbﬂeisch & Prentice (1980), two classes of
regression models may be distinguished, namely, proportional hazards models (PHM)
and accelerated lifetime models.

The Weibull distribution is a very ﬂexible model for lifetime data. It has a hazard
rate, which either is monotone increasing, decreasing, or constant. It is the only
parametric regression model, which has both a proportional hazards representation and an
accelerated failure time representation.

Proportional Hazards Model (PHM) was originally developed for modeling
components, which can only fail once. In order to model a repairable system like pipeline

network, the lifetime is deﬁned as the inter-arrival time (i.e. time between failures).

A Non-Homogenous Poisson Process models the recurrence of events, assuming the
number of them occurring in a given time interval to be Poisson distributed. For an
ordinary Poisson process, the mean of the Poisson distribution is the product of the

interval length by the intensity function, which remains constant in time.

60

The Events Dependent Renewal Process is a generalization of the ordinary Renewal
Process, allowing the successive inter-arrival times to have different distribution
functions, which depend on the rank of the events. An ordinary renewal process models
the sequence of current events, occurring in a repairable system by assuming the delays

between events to be independent and identically distributed.

Probabilistic Predictive models estimate the probability that a break will occur at some
future time and/or the probability of a pipe to enter a particular state (e.g., severe
deterioration with multiple failures, Andreou 1986). This can then be used to calculate
the economic life of a pipe and therefore when it Should be replaced. Andreou et al.
(1987) proposed the use of a Cox proportional hazard model (Cox 1972) to relate the
hazard function to a set of explanatory variables. The basic form of this model is

represented in equation (2.2) below:
h(t : z) = home” .......... (2.2)

where, h(t : z) is the failure rate (termed hazard function), h0(t) is some unspeciﬁed
baseline hazard function, 2 is a vector of explanatory variables (diameter, soil, etc.), and b
is a vector of regression coefﬁcients.

Whilst various mathematical distributions can be used, the Weibull and Herz
distributions appear to be most suited to pipe failure statistics (Herz 1998). Provided it is
available, pipe replacement data can be used directly to Obtain ﬁtting parameters to these
distribution functions. For example, Herz, (1998) equated the ﬁaction of pipes that were
replaced in a given year to the ‘hazard’ (or failure rate) associated with pipes of that age.

The variation in hazard with pipe age was then used to obtain the lifetime probability

61

density function (Crowder et al. 1991). It should be noted that lifetime distributions
determined by this technique are not based on recorded failure data, and merely reﬂect
the replacement strategy used by the utility agency at the time of analysis. Alternatively,
in the absence of pipe replacement data, breakage rates can be extrapolated from recorded
failure data and used to determine the mean value of a lifetime probability density
function. For example, time-exponential equations can be used to forecast future
breakage rates, allowing the discounted costs for future pipe replacement to be
determined. The mean economic life of a homogenous group of pipes (i.e. the average
time for replacement) is assumed to end when the total repair costs attain a minimum
(Kleiner et al. 1998). Further research using a proportional hazard approach includes Gat
and Eisenbeis (2000) and Lei and Saegrov (1998). Both these authors used a Weibull
hazard model to model the useful life of a pipe.

Among various techniques for deriving probabilistic predictive models, the
survival analysis has been widely used. The objective of survival analysis is to develop
lifetime models based on survival data (or failure data, or lifetime data).

The main shortcomings of the survival analysis approach is that it groups similar
pipes and relies heavily on estimating the lifetime of the groups, which may itself be

highly variable and dependent on the individual pipe characteristics.

Probability-Based Markovian Models provide a reliable mechanism for developing

prediction models. Markov chains can be employed to model stochastic processes, which

have the distinct property that probabilities involving how the process will evolve in the

62

future depend only on the present state of the process and so are independent of events in
the past (Hillier and Liberman 1995).

The Markov process imposes a rational structure on the deterioration model
because it explains the rate of deterioration as uncertain, and it also ensures that the
projections beyond the limits of data will continue to have a worsening condition pattern
with time. This model has been successfully used in other types of infrastructure
deterioration modeling like bridges, pavements, etc.

To model the manner in which a sewer deteriorates with time, it iS necessary to
establish a Markov probability transition matrix. The transition matrix P is a square
matrix, m x m, where m is the number of possible states. Thus, if there are ﬁve categories
in sewer conditions, then ﬁve possible states will be involved in the matrix of size 5 x 5.
The components of P, namely pjj, are the probabilities of being in state i at time 0 and
transitioning to state j over a given period At. Kathula (2000) in her dissertation assumed
a time increment, At, of 5 yrs because sewer inspections Should generally be conducted
every 5 years. If the assumption is accepted that the sewer condition will not drop by
more than one state in any S-yr period, then the condition will either stay in its current
state or move to the next lower state in 5 years. Therefore, the one-step transition matrix

can be represented as follows (Kathula 2000):

p11 p12 0 0 0
0 p22 P23 0 0
P: 0 0 P33 p34 0 """"" (2-3)
0 0 0 p44 p45
L_0 0 0 0 l -

 

 

For each row of the transition matrix, 2; pi,- = l. The value of 1 in the last row

indicates an “absorbing” state corresponding to the fact that the sewer condition cannot

63

move from this state (the worst possible state) unless rehabilitation is performed. In this
particular transition matrix, the values of four unknown quantities (i.e., p11, p22, p33, and
p44) have to be determined. The application of the Markov process (Butt et al. 1994)
proposes a nonlinear programming approach to determine the probability values by
minimizing the sum of the absolute difference between actual data points and the
predicted condition for the corresponding time generated by the Markov chain.

The probability that the sewer is in state i at time t = t and will be in state j after it
periods is desired. Chapman-Kolmogorov equations provide a method for computing the
n-step transition probabilities, and the n-step transition probability matrix can be obtained
by computing the nth power of the one-step transition matrix (Hillier and Lieberman
1995). Thus, if the one-step transition matrix P corresponds to a 5-yr time period, then
the two-step (IO-yr time period) transition matrix Pm is represented by

Pm = P2 = P x P ---------- (2.4)

Besides the transition matrix P, the state matrix X representing the probability
distribution of being in m different states at time 0 (which is the fraction of sewer
network currently in each of the m possible states) is also required. X is a single-row
matrix (or state vector) where Z X; = 1 for i = 1,. . ., In. The state vector for any time cycle
t is obtained by multiplying the initial state vector by the transition matrix P raised to the
power of t. Thus, the prediction of sewer condition 10 yr from now is then represented by
x ‘2’.

[x ‘2’] = [X] x [P ‘2’] ---------- (2.5)
The use of the Markov chain prediction model is sufﬁcient to formulate the

problem as a dynamic programming problem because the knowledge of the current state

of the system conveys all the information about its previous behavior necessary for
determining the optimal policy henceforth. This property is required in dynamic
programming formulation (Hillier and Lieberman 1995). The application of the Markov
chain prediction model in conjunction with dynamic programming has several
advantages. It uses objective condition measures and has computational efﬁciency in
handling a large number of rehabilitation strategies for each sewer classiﬁcation/state
combination. However, the model development requires sufﬁcient statistical data for

establishing sound transition probability matrices.

65

 

.388 cognac 9 REESE
owSE> 2E 23 8165 Lo Dab

EzcoEmB 3 28:96 DQE mo “X. H WMN
JESS—gov

 

 

o6 mm. :03 8% mo 399 3592?. 13.53:: 3 Emto>o omamo X n N Emma);
.35 >255 was we 22:89:00 .35 a ESE Samoa 828%. n K m.>mo2o.q§§ 53.20. i
meiou 98 E3895 :8 .mocsmmoa 69a mo 88:86 M Q H mm;
message Sosa cosmgomﬁ mm 588 SE 8
:25 3 69¢ 2: Co 56:86 was cab 552ng 89a memo» Lo .5895 u x2 93? $48? mmmc? Amwm:
.boLmE owmxmoﬁ done—Ema: mo DEC. £56883 55.0.8ch H Sea 3? A? + 93? an n E .3 Ho ate—U
”mama
ammo ta c8 Snowmoa x85 =So>ov
was $8256 common momma ammo ta com
hocosvoﬂ Vick: c853 0.58 H ND
”not “mac AgnmwcmmEQV
8m Savage.“ x35 =So>3
“80:86 3% new magma 33 .«o wad 3x85 325E $88 no ogQOE Amwaa
@082: 2: no 833:0qu 33 83$ 5:» no: 58 AesmmcaemEav com «Boa—om
@3303 was E525 com me 83 08mm 55362.“ Meek: 5053 038 n G 93:32.6.6 n 82 one .5255
Eng
£38m BE owmxmoﬁ mo EEoEooo n V
.80 .mozmtoﬁﬁwno o8: E083 oﬁ an 0&9 2: we own .1. M
sound—€05 .03 :8 Gawain 3E 2:
.25 33 BE «:35 8 @6808 we gong—SEE mo 30> 05 “a 32 n A82
3388 3:on anowoﬁon ENDED—Eb can» 83:
mo gouache.“ $833 omexeos com Ewan: ES com 8:35 .02 n 32 9.23:
can 8% cone—:32: Ramse— omE meme.» 5 05on See towns—o 085 ”a 3:20.232 n 32 can :83

 

 

 

 

£252 Runeneaxmbﬁﬁ. cunt—=58:—

 

 

mEothsrom 8.5 ﬂ

neuauez

Eve—2

3:283—

 

 

Coom Esra use 5.55: 35—an 53383? new 53>? no.“ £382 SEEDS Eoumuﬁm mo ng .. ad 29¢.

66

 

 

 

 

 

 

.mmonm msocomﬁoﬁon 356508 563%»: Hams 2m 5a $33
mo cozwgom 3398 8% 088 8:85 0: FEB >3 ow< onaM
SSE: owmxmos can own .ﬁwaﬂ 35 a we bzﬁmnoa 2: mo 188302 n K Na + ﬁmcoq 5 + on u m 98 38mm

.Eowcomow
33533 2m Ho 88m 38%
3:82 Ho: 0%: 283628 533
95on 80:3 Eatong ﬂ $5.835
:8 mo watoquE mzosszcg EEBOQ x82 u 8
538 E Samoa cosoozoo 8% in :8 H ER
36an... Ea 325980 5530: 32:
638098 “on 55230 Saw 0680?. -828 bgumaou :8 @3833 u «h Eovod$§~wmm© Ammo:
USERS Ho: 332mb @859: Sam $320 x85 “m5 3 3E mo own u mmv. -memc.o+wn.mc H mMV 8:320:
£3:
83: $3888 coamokme n ox 5:30
c.8303 98 £528 “on ma 8% 98mm So» Sm $85 no #582“ n 2 N313 n 2 98 530M

 

2252 5354.053. 3238.58:

 

 

 

=8 038:8
>33: 5 3E mo «98 ooﬂSm n in
£8 bmzmoboo

32 5 game .88 conga u an
£8 ozmoboo 3880?":

93 32 a 532 2E “o .x. ”\EQ
£85 5E Boa 3E we own H»
.3355wa 05389 n 3%
£83“ Mo 8983 n max
A8888 voohomﬂe

n o asses u a 25 25 ”k
.28 ozmoboo

has a 2E «o 593 n 5
.EoEmoBZﬁ

 

 

 

 

67

 

 

 

 

 

 

 

 

mEHoot—w 8:282; mo :90? m u N 5.22
5.323: 3 :BaEsmu Amoo C
o: 9 8:23:08 :0 :08? n u :8qu :5
o2m> 0::qu n S 0:353:00
302833: Naoom H o
09$ :8 . £83.58 092% :5 2QO H m 6 3:0:
53:56 2:: . : mm: a: ﬁng— «Swazi
30:6:8 Stu: 98:35 . z:: B: 8:25 we 5:55: :88 U SE Emhrmme
8:38: 039m :aoE - um: 2:: n ~ a 83 u 3: 235
8:2:me
mﬁcmo— oEmb - :36? 2: mo $62838:
2:36:00 :::o:m . 28052:: :92? :5 Baum H a J
SEES: 8E . :ozoEd ESE n 3: $3:
8.85 2532: we 52:5: . 225 €35 9 2:: u: 3C6? n so: 20:65
3% n & u « n9mg.m 33 Q3:
._m 8 8—32
o>onm wontomo: SD3—
Ammo: .__w B 8:32 ago: .3
30:: 3 08mm Emma Ba— 2: 3 E8208 Emwa: H a 3 2:3 newsm bSm Ho =08::<
536850 :8 -
2:: E 3:35 5532: we 696:: ..
88 x85
Conwﬁ :8 288 an own 2:: ..
Eotmzﬁmi 82:8:— Eagxﬂ: 3 93258
no vote: :8 :owSE? o2: . on 9 £56508 :0 582, H :
E23205: mouwtgs go :90?» u N ml: xm Ammo:
6:2 >6. mo omﬁgﬁom . :osoce Emu“: 05—33 n 33 +k‘2 I 1o _ xm H 33 ._m 8 8:22
838:: mgﬁoao . 55:8 :5ch H AN .3: $238:
:ch2 2:: mo wo_ 123:: u #85 :8: 9 0:5 H s NE 302%: n AN .3: NgeEemci

 

 

a: 682239< :5“ «Bang: FEVER—Em I £352 3:_::>-_:=2 95:32:;

 

 

68

 

 

 

 

 

9:80— 050::

3:088 05: .

5:055: 2:: .

53:0— 0:: .5 w2 .
55:5 5:80:85:

8 35:50:: 050:? 50:8: n N
35:85 08::

5 3:058:00 0:80:88: 5 :800> u m.
8050:: 5555::

3: 3:08:00 0: 0: 53:88: n b

,0 :8 6:3 :8 03 .550: a?“ a :o 8:388:50 5:03
5:085: 09: :05: A..wo:m_:0:8:0:0 8500:: 5:30:50 5 3:558: “5% n 70:50:88535008 232535
:0 300.: 5:85 w:_::o:0 05:8: .5 3:558: 0:37:83»: HQ $55,250:: 50.2035

CLEN + B H ACNK

m 30:05 80> 0 V 08 8 :002:0: 05 :0: EB 2:30:
5:05 00:53: 9850830: 2:: :0: .3803 08: 00:90:00: n 0 Sam:
:0: 0E: 8:5:8 0.5 0:80:80 .5 C003: :88: 03:8: u a 2:30 + a :0 :0 500
::0 ”.580: 5 0:5: 8 03850:: . C003v :88: 568 n a a n Sm. A03: 503
38508 008.50 0:: 5 05:05 5.55 H:
E? 25:5 w:_::o:m ::0> - 5:25.: 533% n Em. ﬂ 0 + a”— 38:
.580: .5 0:5: 0:: . 5:05.: :88: H S: N 3;: n Sax 5935,:
008: 58:50:: 0:: . 5:05: 36:0: 3:558: u SK 2-50:: + 3 E509
m 20:32 98:0 85520.55 053—553:
08:? :28139850 .ﬂ. 3:...
0 8 :8 088200 a 8 :8: :05:

83 8—805 9850:: 5 55:5: . 33: :8
3555: :8 . A0555 :0 0:05:05:

35:8 :8 . :8 8:552: 08853 65:50 05:: C803 :04 8 050m

I

0:5
:08:0_000<

 

 

a 8:5 33 3:38. 2:: . :36»: a :30 S a .3? : :3 s :3
59:0— 0:: a 8 :05555: 05:55 80:8: n N

, 05v. 0:: . 0050.5? 3088—98 .5 :800> u a 0x5

95% 0mm 0:: a 05—5: 9x05 8 05: n .N N0 + n :x + 1 n ACE 8:50:08:

68: .3 3

05:85:00

 

08: owe—:05

 

 

”9.8:

 

 

69

 

.32 m: 3538:
:: 80:5 5008

 

 

 

 

 

 

 

2:: 0:88: 8 :82
:85: 8 88:88: 6:2:
0:3 :8 80:5 03500800 03: :0 58:2
:0>0: 0E8: 3:558: 05: 585080
15:08:: 2:: :0 :08: .08: 08080:: w:::0:::
8:285 08: 0:5 :0:: 0: 00:: v 08::
8:020 30:5 :0
a 5 0805 a: 0:: 50:: :0m:0_0 08:: H : :0 8:500
820 :83:
a 5 0805 a: 05: 88: 0088:: H : 5050:
5:80: 8:050 0:: 5 55000 x: 6:2: n E :8 85:00
005:8: 50500058 :0 85:5: n a .

080:5 580:0 :0800: :8 08:: 5:80: 8:050 05: 5 @5000 85:8: kl u 00: 95:08:50
:85 0:: 5:3 30:05 08080:: 0:: 5250058 :0 85:5: 80:: n 5 50:2: 88m
883
>085 :8
2A: _: =_m 8.: 8:330

>858: 80:50:: I 5:55:55:
88:50 0: 08:8 w::::0:w :0::0 8:505:00 . 25:0
0:3 0:5 0:: 85005 _: :0: 5:55:85 :35:
@805 08805 0:: e: 0:: :8 5:7: 05: 803:0: 05:: n _: :88» :0N:a:0:0w . A 5%»:
:0:“: :0: 8: _. u 0 cwo: .8
0:0 05000:: 8 50:0 :0:: 50800 8 :0 0088:8850 08m in: I c : r: + $5: :0 58:3:

2:05:50 @5825 .00:28:88.5 05: 5:200:50 :0 3:558: ”:50: “:2 In:

 

 

70

2.1.1 PHYSICAL MODELS

Physical models of the degradation process employ engineering-based equations to derive
structurally based estimates of pipe conditions. The physical mechanisms of pipe failure
involve three principal aspects: (a) pipe structural properties, material type, pipe-soil
interaction, and quality of installation, (b) internal loads due to operational pressure and
external loads due to soil overburden, traffic loads, frost loads and third party
interference, and (c) material deterioration due largely to the external and internal
chemical, bio-chemical and electro-chemical environment. The existing physical models
can broadly be classiﬁed into deterministic and probabilistic, and most cannot
simultaneously address all three principal aspects listed above. Based on actual failure
mechanisms, physical failure models can also be used to estimate changes in pipe
condition and future failure rate. To develop these models for the full range of pipe
materials in use, expertise is needed to quantify corrosion rates in metallic pipes, rates of
degradation in cement-based pipes and the fracture mechanics of plastic pipes. Additional
expertise is required to understand the interactions between the electrochemical
properties of surrounding soils and degradation rates. Physical models rely on input from
accurate condition assessment techniques, and can provide performance indicators. It
appears that the physical mechanisms that lead to pipe breakage are often very complex
and not completely understood, and little data are available to validate models based on

these mechanisms.

71

2.2 SEWER MANAGEMENT SYSTEM

For proper monitoring and maintenance of underground infrastructure, a thorough asset
management strategy is required to perform many ﬁmctions including inventory,
condition assessment, condition forecasting, inspection, scheduling, budget forecasting,
localized maintenance programs, and annual and long-range maintenance and
rehabilitation planning. More municipalities are beginning to realize of the obvious fact
that it is much more economical to repair or renew the sewers before they are fully
deteriorated. If maintenance and rehabilitation is performed during the early stages of
deterioration, substantial repair costs can be avoided in addition to avoiding service
disruption and other social costs. In today’s economic environment, as the sewer
infrastructure has aged, a more systematic approach to determine maintenance and

rehabilitation needs and priorities is necessary (Kathula 2001).

2.3 APPLICATION OF NEURAL NETWORKS IN PIPELINE
MANAGEMENT AND PREDICTION MODELING

In recent years, artiﬁcial neural networks have been advocated as an alternative to
traditional statistical models. Neural networks are application of an algorithm inspired by
research into human brain which can “learn” directly from the data. It can be deﬁned as
“highly simpliﬁed models of human nervous system, exhibiting abilities such as learning,
generalization, and abstraction.” One of the advantages of a neural network model is that
a well-deﬁned mathematical process is not required for algorithmically converting an
input into an output. A collection of representative examples of desired translation will

sufﬁce. Once trained, a neural network can perform classiﬁcation, clustering and

72

forecasting tasks. Thus, the pipeline industry can harness this technology to model
dynamic deterioration of pipes using the historically available data. Once the pipeline
deterioration pattern is modeled, it is then possible to predict future condition and
deﬁciencies of the pipelines. Feasible strategies can then be synthesized to further

examine the actual condition of those pipelines.

Past Research in ANN Application for Pipeline Application. Sacluti (1999) in his
Master’s thesis applied an artiﬁcial neural network (ANN) to predict the pipe breaks in
the water distribution system of a sub-division in Edmonton, Canada. The ANN model
was applied to the entire network as a single entity (rather than to individual pipes) and
was trained with data that included temperature (water and ambient), rainfall, operating
pressure and historical data on break numbers. Since the model considered an entire
network as a single entity, variants such as pipe age, type and diameter could not be
considered, as well as geographical varieties such as soil properties. The network
consisted of spun-cast 6-inch water mains. His work focused on the frequency modeling
of water distribution pipe failure mechanisms'in cold weather climates.

The ANN model was applied to a relatively small network with water mains that
were relatively homogeneous with respect to type of pipe and operational and
environmental conditions. A more heterogeneous set of water mains would likely require
more data. The model predicted the number of water main breaks based on a 7-day
weather forecast. This requirement limited its ability to short term response than its use

for long term planning purposes. In its present form, the model can only be applied to

73

homogenous groups of water mains, for short-term planning of the maintenance work

force required during an anticipated cold spell.

2.4 SUMMARY AND CONCLUSIONS

The literature review in this chapter indicates that research in pipeline deterioration and
forecasting model development has been in focus lately. Various modes and mechanisms
of pipeline deterioration were reviewed and the possibility of the application of neural
network in pipeline management and forecasting was discussed. Also, sewer condition
assessment techniques and structural condition rating of sewers were documented to
develop a broad-based understanding of the technology that are available in the market
for condition assessment and classiﬁcation of distressed sewer pipes. The literature
review indicates that there is a good possibility to develop a successful neural network
based model if the critical parameters that contribute to the deterioration of pipelines are
obtained. A neural network model for predicting pipeline performance trends based on

historical condition assessment data will be developed in this effort.

74

CHAPTER 3

NEURAL NETWORK METHODOLOGY AND APPLICATION

The previous chapter dealt with the literature review that gave the necessary background
for pursuing this thesis. This chapter presents the methodology used in this thesis for the
development of a sewer pipeline condition prediction model. The prediction model is
based on neural network modeling technique. A detailed description of neural networks is
presented in this chapter, along with the pragmatic method appropriate for modeling

sewer condition prediction.

3.1 ARTIFICIAL INTELLIGENCE AND NEURAL NETWORKS

The term artiﬁcial intelligence has be traditionally used to refer to the ﬁeld of computer
science dedicated to producing programs that attempt to be as smart as humans. Expert
systems and neural networks are two forms of artiﬁcial intelligence, each with distinct
strengths and weaknesses. Most implementations of artiﬁcial intelligence are programs
that simulate either the deductive or inductive intelligence of human being. Deduction
reasons in steps to a conclusion based on given premises. Deductive systems, which can
be simulated by expert systems, require rules or instructions executed one at a time to
arrive at the answer. By contrast, induction takes in a large amount of information all at
once and then draws a conclusion. Neural networks can be used to simulate the inductive

behavior of humans. Once trained, the neural network is able to look at input data and

75

produce an appreciate answer. In a comparison to expert systems, Garrett (1992) presents
the following advantages to the use of neural networks:

} Neural networks have the ability to present a model for a situation where only
examples are presented.

} Expert systems require “certain factors” or “levels of belief” as means of
accounting for uncertainty, whereas neural networks are trained to deal with
uncertainty since training data is obtained from situations very close to the
situations in which the network will operate.

i Expert systems are very brittle in that all data must be complete and correct in
order for a system to be analyzed. On the other hand, neural networks have the
ability to allow for minor errors or omissions in the input data and also for slight

deviations from existing training cases.

Neural Networks can be used to:

0 recognize patterns and images

- construct a decision tree to solve a problem

- classify data

- predict outcomes

0 study thematic evolution of a process and construct cost effective models

A neural network is a massively parallel distributed processor that has a natural

propensity for storing experiential knowledge and making it available for use. It
resembles the human brain in two aspects:

0 knowledge is acquired by the network through a learning process, and

76

0 connection strengths between neurons, which are known as synaptic weights, are
used to store the knowledge
In computing terms, neural networks have a unique set of characteristics derived
through its massively parallel-distributed structure and its ability to learn and generalize.
These two information-processing capabilities make it possible for neural networks to
solve complex problems in the real world. The key characteristics of neural networks can

be summarized as follows (Lou et al 1999):

Learning from experience: Neural networks are particularly suited to solve problems
whose solution is complex and difﬁcult to specify, but which provide an abundance of

observed data.

Generalizing from examples: An important attribute of neural networks is the ability to
learn from previous experiences and then give the correct response to the data that it has

not encountered.

Nonlinearity: Neural networks can be trained to generate nonlinear mappings, which
often give them an advantage for dealing with complex, real-world problems.
Nonlinearity is a particularly important property if the underlying physical mechanism is

inherently nonlinear.

Computational efﬁciency: Although the training of a neural network is computationally

intensive, the computational requirements of a fully trained neural network applied on

77

test data are modest. For large problems, speed can be gained through parallel processing,

as neural networks are intrinsically parallel structures.

Adaptivity: Neural networks have a built-in capability to adapt their synaptic weights to
changes in the surrounding environment. In particular, a neural network trained to operate
in a speciﬁc environment can be easily retained to deal with minor changes in the
operating environmental conditions.

A neural network is an excellent candidate for any application requiring pattern
recognition. Neural networks are able to recognize patterns, which may consist of visual,
numeric, or symbolic data, even when the data is noisy, ambiguous, or distorted. In

general, neural network tasks may be divided into ﬁve types of distinct applications:

Classiﬁcation: Deciding into which category an input pattern falls into.

Association: Acts as a content addressable memory that recalls an output with reduced

dimension. The opposite task, decoding, may also be of interest.

Simulation: The creation of a novel output for an input that acts as stimulus. The network

has been exposed to a sample of possible stimuli.

Modeling: The network mapping process involves nonlinear functions that can

consequently cover a greater range of problem complexity. Although other nonlinear

78

L5. /

techniques exist, the neural network is superior in its generality and practical ease in

implementation.

3.2 NEURAL NETWORKS AND STATISTICAL

MODELING: A COMPARISON
A neural network may be considered as a data processing technique that maps, or relates,
some type of input stream of information to an output stream of data. For example, the
input may be the pipeline condition data like pipe material, pipe age, diameter, slope,
environmental conditions, etc., and the output may produce an estimate of the probable
condition of the pipeline. On the other hand, the goal of statistical modeling is to ﬁnd an
equation that captures the general pattern of a relationship, which is usually derived from
observed examples. Therefore, the ﬁelds of statistical modeling and neural networks are
closely related in the context of input-output mapping. The principal difference between
these two ﬁelds is that traditional statistical models typically need an equation to be
speciﬁed, which could be difﬁcult in complicated nonlinear cases, while neural networks
have been mainly used to deal with nonlinear problems without requiring a pre-speciﬁed
function form. However, with the appearance of the backpropagation neural network
(BPNN), of which the learning paradigm is called supervised learning, these two ﬁelds
touch most closely in solving mathematical modeling problems. This technique solves
one of the central problems in neural networks, and it is a useful modeling tool as well.

Supervised learning involves the modiﬁcation of the synaptic weights of a neural
network by applying a set of training examples. Each example consists of a unique input

signal and the corresponding desired response. The network is presented an example

79

picked at random from the set, and the synaptic weights of the network are modiﬁed so as
to minimize the difference between the desired response and the actual response of the
network produced by the input signal in accordance with an appropriate statistical
criterion. The training of the network is repeated for many examples in the set until the
network reaches a steady state, where there are no further signiﬁcant changes in the
weights. Thus, the network learns from the examples by constructing an input-output
mapping for the problem at hand. The key characteristics of a neural network can be
' summarized as follows:
’ a large number of attributes can be considered in parallel
' neural networks learn by example, therefore, knowledge acquisition is not
difﬁcult
' quick response can be provided by a neural network model
D classiﬁcation based on given inputs can be attained and, input classiﬁcation
characteristics can be extracted
D an incomplete data set can be analyzed due to the neural network’s ability to
generalize
? a fault tolerant property allows for small errors in training data to have only a
slight effect on the processing elements
’ only a small amount of memory is required as only network weights need to be
stored for recall programs
Statistical modeling techniques are used to derive relationships between variables
from examples as well. In the case of pipeline condition prediction, the examples are the

performance history in the last few years. To derive the equation from the examples, the

80

values of the independent and dependent variables for each example need to be known. In
this case, the independent variables involve the performance history data and the other
pipeline descriptive information, and the dependent variable is the present condition of
the pipeline. During the running mode, a running ﬁle needs to be prepared, which
contains the independent variables of each new example for which an estimate of the
dependent variable is desired.

The prototypical example of a statistical modeling technique is linear regression.
The equation produced by a modeling method can be thought of as a mapping, because it
permits us to map any point in the space of the independent variables onto a point in the
space of dependent variables. The error of the mapping comes from two sources. The ﬁrst
source of error is noise, which includes inaccuracies in the data introduced by measuring
instruments, and inaccuracies due to the fact that the independent variables do not contain
all the information needed to determine the dependent variables. The second source of
error is the fact that the mapping function may not have the same form as the target
function. The so—called target function could be an idealized and unknowable function
that expresses the “true” relationship between the independent and dependent variables.

The fact that linear regression imposes a linear form on the mapping function can
severely limit its accuracy. In cases where the problem domain is not a linear space, it is
usually necessary to transform the variables so as to make the relationship linear. A better
approach is to automate the process of deciding what shape the mapping function should
have. What is needed, then, is a modeling technique based on a mapping function that is
complex enough to be ﬂexible. Although some simple curves, such as polynomial

regression and exponential equation, have been used to simulate the real world condition,

81

the optimum solution is a technique that can take on any form the data requires. One of
these advanced modeling techniques is the backpropagation neural network (BPNN).

NeuralWare (1993) compares the abilities of neural networks to other means of
artiﬁcial intelligence. Table 3.1 presents a comparison of neural networks to other means
of modeling problems.

Table 3.1 — NeuralWare Modeling Comparison

 

. AdvantageorNeum

Technique errtatron . Nemork

 

Neural networks are trained
and, therefore, can handle
unlimited numbers of
variations without additional
work.

The number of variations is
Traditional limited as each variation is

Programming required to be programmed
into the model.

 

Knowledge and explicit setting
of rules is not necessary for
System requires that an expert neural networks since

Expert Systems knowledgeable in the topic set historical data is used for

rule basis for processing. training (knowledgeable
checks and input, however, are
still advisable).

 

 

There are fewer limitations,
such as the need for a sufﬁcient
training data, to the number of
inputs that can be analyzed by
neural network.

Level of analysis is limited to a

Re ession Anal sis .
gr y certarn number of parameters.

 

 

 

 

3.3 THE NEURAL NETWORK ALGORITHM

There are many types of neural networks, but all have three things in common. A neural
network can be described in terms of its individual neurons, the connections between
them (topology), and its learning rule. Both biological and artiﬁcial neural networks
contain neurons, real or simulated. These neurons have many connections to each other

which transfer information. The knowledge of a network is distributed across the

82

interconnections between the neurons. A typical neuron receives input, either excitation
or inhibition, from many other neurons. A neuron calculates its own output by ﬁnding the
weighted sum of its inputs, generating an activation level and passing that through an
output on transfer function. The point where two neurons communicate is called a
connection. The strength of the connection between two neurons is called a weight. The
collection of weights arranged in rows and columns is called the weight matrix. A neural
network learns by changing its response as the inputs change. Because the weights in the
network can change, the relationship of the network’s output to its inputs can be altered
as well. In this sense, the learning rule is the very heart of a neural network, which

determines the behavior of the network and how that behavior can change over time.

Input Layer Hidden Layer Output Layer

 

Figure 3.1 - Anatomy of a Neural Network

3.3.1 SINGLE NEURON

Artiﬁcial neurons as information processing devices were ﬁrst proposed more than ﬁfty
years ago. As shown in Figure 3.2, a neuron computes a weighted summation of its n
inputs, the result of which is then thresholded to give a binary output y which is either +1
or -1. The bias weight, 9, is introduced whose input is ﬁxed at +1. This bias weight is

adaptive like the others and its use allows greater ﬂexibility of the learning process.

83

Input > y = f(Zwixi)

/

Figure 3.2 - Schematic Diagram of an Artiﬁcial Neuron

 

For a classiﬁcation problem, the neuron assigns input patterns, represented by the
vector of numbers x = (x1, xz, . ., xn), either to class A (for which y would be +1) or class

B (for which y would be -l). Thus:

1 when Zwixiw

y=f(:w.x.)= ------------- (3.1)

 

- 1 when ZZWixi 2 0

i=1

In the above equation (3.1), y is the neuron output and f is a hard-limiting of

threshold function, sometimes known as the neuron’s transfer function, which gives an

output of +1 whenever Zwixi is greater than zero (the threshold value) or -1 whenever
Zwixi is less than (or equal to) zero.

The learning process is to adjust all the weights and let the output y approach the

desired output so that the neuron performs the classiﬁcation task correctly. Multi-class

problems can also be solved by having a number of neurons operating in parallel.

3.4 BACKPROPAGATION NEURAL NETWORK (BPNN)
By far, the BPNN is the most popular one used for mathematical modeling.

Backpropagation is a supervised learning scheme by which a layered neural network with

84

continuously valued neurons is trained to become a pattem-matching machine. It
provides a way of using examples of a target ﬁmction to ﬁnd the weights that make a
certain mapping function hidden in the neural network approximate the target function as
closely possible. As shown in Figure 3.3, the neurons of the networks are structured in
multiple layers: input, hidden, and output. Each hidden-layer neuron receives input from
all neurons in the input layer through weighted connections (w). In addition, each neuron
is associated with a bias term, called the threshold, 9. This bias term works as a horizontal
shift for the origin of the transfer function to accommodate the magnitude of incoming
signals to the neuron. Speciﬁc values of both w and O for a given neural network are

determined during the training phase.

           
     
    

__,. A
’ \ll/ 0......
t N ——+
A Alli
Vl“< Output Layer

.
Input Layer H

”A.

Hidden Layers

Figure 3.3 - A Three-Layer Backpropagation Neural Network

85

3.4.1 BPNN MODELING

The BPNN network operates in two modes: mapping and training mode. In mapping
mode, information ﬂows forward through the network, ﬁ'om inputs to outputs. In the
training mode, the information ﬂow alternates between forward and backward. In the
mapping mode, the network processes one example at a time, producing an estimate of
values of the dependent variables based on the values of the independent variables for the
given example. First, a set of values for the independent variables is loaded onto the input
layer of the network. The input-layer neurons do no calculation — each neuron merely
sends a copy of its value to all the hidden-layer neurons. Each hidden neuron calculates
the weighted sum of the inputs using its unique connection strengths as weights. Next,
each hidden neuron computes a transfer function of its input sum and sends the result to
all the output-layer neurons. Then, each output-layer neuron performs a similar
calculation and outputs the resulting value as an estimate of the dependent variable it
represents.

The training mode refers to the process in which the network is exposed to
examples with correct output values known. The training algorithm consists of three
steps. In the ﬁrst step, the training patterns obtained from the database are fed into the
input layer of the network. These inputs are propagated through the network until
reaching the output layer. The output of each neuron is calculated by the following

transfer function (Lou et al. 1999):

a = Zn wixi """"" (3-2)
i=1

1

He83

 

o = f(a) = ---------- (3.3)

86

where:

values

0 = neuron output,
a = input to the transfer function,

‘1‘ input,

xi = i
w, = weight of connection i,

g = gain of sigmoid ﬁrnction, and

n = number of inputs to one neuron.

In the second step, the neural network outputs are subtracted from the desired

to obtain an error signal. This error signal is the basis for the coming

backpropagation step. The following equation (3.4) deﬁnes the error signal (Lou et al.

1999):

where:

 

 

N, N,
Z 2(Tjk'0jk)2
Ems= "' ---------- (3.4)
N.N.

Ems = root mean square error,

N0 = number of neurons in the output layer,

NC = total number of patterns in an epoch,

Tjk = target (desired) value of the jth neuron, and the kth pattern, and
Ojk = output of the j'h neuron, and the kth pattern.

In the third step, error is minimized by the backpropagation of the error signal

through the neural network. In this process, the respective contribution of each hidden

neuron is computed and corresponding weight adjustments needed to minimize the error

are derived.

87

For each output neuron k, compute the 5 value, deﬁned as follows:

5k = (H - 0k) f'(Xk) """"""" (35)
where:

6;, = adjusted error for output neuron k;

Tk = target value of output neuron k;

0k = output value of output neuron k; and

xk = input to output neuron k.

Backpropagate the 5 value through the network to the preceding hidden layer. For
each hidden layer neuron j connected to the output neurons k, compute the new 5 value

(Lou et al. 1999):

6]. = f‘ (xj)26kwj, .......... (3.6)
k

where:
Sj = adjusted error of hidden neuron j;
Xj = input to the hidden neuron j;
8k = adjusted error of output neuron k connected with hidden neuron j; and
wk,- = connection weight between neuron j and k.
The weight connecting any two neurons is updated by the following equation:
'1’ —‘) '9

qup = aqup ---------- (3.7)

Vw qp = adjustment of weight between preceding layer neuron p and proceeding

layer neuron q;

6., = adjusted error of proceeding layer neuron q;

88

Op = output of preceding layer neuron p; and

a = learning coefﬁcient (a positive constant).

The training process repeats steps 1 through 3 for all patterns in the training set
until the overall error is acceptably low based on a given criterion. If the network has not
converged then go back to the step 1, otherwise stop training. Once trained, the neural
network has the capability of adapting to changing input. If the trained network results in
good accuracy on the testing and validation data set, the development process is
completed.

Although theoretically complicated, the training process is typically implemented
by a computer program, within which the training algorithm has been incorporated.
Popular neural network development packages in the market include MATLABTM Neural
Network Toolbox, DataEngine, BrainMaker, NeuroSolutions, etc. These software
packages vary in terms of training speed, pre- and post- data processing utility, and
convenience of user interaction. Once trained, the BPNN can be incorporated into other
programs; the running of this model can be implemented through a user-ﬁiendly
interface. In this effort, BrainMaker is selected as the development tool, the detailed

training and testing process with BrainMaker is described in Chapter 4.

3.5 NEURAL NETWORKS AND ITS APPLICATION IN

PIPELINE CONDITION PREDICTION

The inherently nonlinear time series, such as that found in pipeline deterioration
process, are more suitable for analysis by the general nonlinear mapping provided by a

neural network, than by linear based statistical models. Neural networks are nonlinear

89

models that can be trained to map past and future data of a time-series, thereby
uncovering the hidden relationships governing the data.

Primarily, two distinct types of models can be developed for pipeline performance
forecasting similar to the pavement condition models proposed by Lou et al. (1999). The
ﬁrst type is a static model and can be conceptually described by the following equation:

PC, = f (S,, E,, M,, t, L, etc.) ---------- (3.8)

Where:

PC, = pipe condition at age t,

S, = pipe structural condition at age t,

E, = environmental conditions at age t,

M, = pipe material characteristics at age t,

L = external and internal load conditions.

The second type, a dynamic model, can be described by the following equation:

PC, = f (PC,-,,1, PC,-,,2. . .PC,-,,N) ---------- (3.9)
Pipeline condition at age t, PC,, is forecast using historical condition data at ages t-xl, t-
xz,. . ., t-xN. This type of model is based on historical performance of pipeline
characteristics, which is difﬁcult to obtain because of the lack of continuous monitoring
process by the municipal agencies.

The dynamic model requires historical data of continuous monitoring of the asset
which is not the common practice by the municipal agencies maintaining sewer pipelines.
A static model will be developed in this research effort due to the limited availability of

pipeline condition data. The association of different variables that contribute to the

90

deterioration of the sewer system will be analyzed and modeled with actual data from the

city of Atlanta.

3.6 ' MODEL INPUT PARAMETERS

To identify the condition of a sewer pipe, the type of defect needs to be classiﬁed and
severity level assigned. Each distress condition will be assigned severity levels based on
the degree and the combination of structural defects found commonly in sewer pipe
segments. Typically this information needs to be documented by the municipal agencies
while performing condition assessment. This information will serve as the input
parameters for the neural network.

As an outcome of the literature review in Chapter 2, various parameters that affect
the condition of sewer pipes during their lifetime were identiﬁed. The parameters that are
determined to have substantial impact on sewer pipe deterioration are summarized in the
Table 3.2 below.

Table 3.2 — Ideal Input Parameters for Model Development

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Pipe Material
Pipe Age
. Pipe Diameter

Pipe Data Length of Sections
Joint Type and Material
Wall Thickness

Zoning Residential/Commercial/Industrial
Typical depth — 8 to 15 feet

Depth of Cover Shallow — less than 8 feet + high live load
Deep - more than 15 ft

Gradient Slope of Pipe

Bedding Conditions '

Backﬁll Type Cohesrve SO11 .
Non-Cohesrve (Granular) Sorl

 

91

 

 

Soil Characteristics

Soil Type

 

Corrosivity

 

Resistivity

 

H Content

 

Sulﬁde Content

 

Moisture Level

 

High Groundwater table (pipe crown is below
GWT)

 

 

Groundwater Condition Low Groundwater table (pipe crown is above
GWT)
Ground Movement Due to expansive soils, etc.

 

Internal Service Loads

Operating pressures, surcharges, etc.

 

External Loads

Soil Load

 

Trafﬁc Loading

 

Other Surface Loads

 

 

Frost Load Factor

 

Wastewater Characteristics

 

Maintenance Frequency

 

 

Tree Root Problems

 

3.7 MODELING METHODOLOGY

The data once obtained will be analyzed for erroneous processing and reduction, after

which the original raw data will be transformed into a new format that is suitable for

ﬁrrther analysis.

To determine whether or not the database contains enough samples, two factors
will be considered important: (1) the form of the target function: to maintain a given
accuracy, sample size needs to increase as the target function becomes more complex,
and (2) the noise in the data: to maintain a given accuracy, sample size needs to increase
as noise increases. Given a target function of a certain complexity, and a certain amount
of noise in the data, there will be an absolute limit to the accuracy the model can achieve.

An inﬁnite sample size would be needed to achieve the limit of accuracy. For neural

92

 

network modeling, since the complexity of the target function is not a limit, noise alone
will determine the limit of accuracy. If the sample is large enough, the complexity of the
network’s mapping function could be increased to match the complexity of the target
function. Consequently, as sample size increases, neural network model’s accuracy will
be limited only by the noise in the data. Usually, neural network can beneﬁt more from
large samples than regression can. Because larger samples allow us to use more hidden
neurons or to continue training longer, the accuracy can be improved by ‘ increasing
sample size. On the other hand, neural network model does not require a larger sample
than a regression model. As the sample size gets smaller, we can use fewer hidden
neurons or halt training sooner to avoid overﬁtting. The basic rule, therefore, is to use the
largest sample available.

The model development in this study will include training, testing and validation.
Training a neural network involves repeatedly presenting a set of examples to the
network. The network takes in each example, makes a response as the output, checks this
response against the correct answer, and makes corrections to the internal connections.
Testing the network is the same as training it, except that the network is shown with the
examples it has never seen before, and no weight adjustments are made during testing.
Validation occurs after the neural network has been developed. Validating a network
consists of presenting it with new input data and gathering the network outputs. Unlike

testing, there is no known output, only known inputs in the validation.

93

3.7.1 NEURAL NETWORK DESIGN PROCESS - AN OVERVIEW

To design a neural network, the problem must be deﬁned clearly. The user should decide
what tasks the network is to perform. These tasks could be forecasting, recognizing, or
classiﬁcation. One cannot just throw all the spreadsheet data at the network and expect it
to ﬁgure out what to learn ﬁ'om the data.

The user also needs to choose the information on which the neural network will
base its forecasting, recognition or classiﬁcation. This should consist of whatever
information is available that is relevant in determining the desired output. Neural
networks learn by making associations between inputs and outputs. A network can
associate the inputs “red”, “medium”, “round”, and “fruit” with the output “apple”.

The user does not need to ﬁgure out procedures, rules, or formulas in the neural
network development. The user should think about what kinds of input data the neural
network can use to make an association with the desired output. Having a variety of data
types increases the chance that various signiﬁcant correlations can be found within the
data. A network would probably not be able to accurately predict stock prices based
solely on a collection of daily stock prices. It is better to have one or two extra items of
data than not enough. The neural network will learn to pay attention to the items that are
important and to ignore the few that don’t matter.

Another important part of design process is preparing to train the network by
gathering examples for which correct answers are known. For example, to recognize a
face, a network would need to have seen a picture of that face before. The training data
were organized as facts (patterns) in a spreadsheet format. A fact is a collection of inputs

coupled with the correct output(s). Each fact can be thought of as ﬂash card that is used

94

to train the neural network. One side of the card contains the input information, and the
other side contains the known answer which the neural network will learn to output
during training. The deck of training cards is called the training set.

i A random sampling of facts should be set aside from the training set of facts for
testing the network. Since the network generalization capability depends on its
performance on the testing data set, it is not as important for neural network to learn a
training set perfectly as it is for it to be able to provide correct answers for inputs it has
never seen before.

Once, trained, the neural network can be called from within some other program,
perhaps an integrated system. The network may also be downloaded onto a chip for fast
running. A trained neural network is considered intellectual property and may be

copyrighted in the United States.

3.8 SUMMARY AND CONCLUSIONS

The discussions in this chapter reinforced the suitability of using neural network
methodology for predicting the condition of pipelines. A comprehensive list of
parameters that affect the condition of the sewer pipes along with the modeling
methodology was presented. It was discussed previously that the availability of data from
the municipal agencies regarding the quantitative values for all the parameters is not a
reality, as such extensive documentation is not prevalent. The following chapter will
discuss the review of the data obtained from the City of Atlanta and preprocessing before

it is fed into the neural network software for modeling.

95

CHAPTER 4

DATABASE REVIEW AND PROCESSING

This chapter describes the detailed efforts of database review and data reduction. The
physical attributes of sewer systems owned by the City of Atlanta were extracted from
their existing database. The data items in Atlanta’s database include Pipe ID, Pipe
Material, Age, Diameter, Length, Depth of Cover, etc. that was surveyed during their
SSES efforts. It was found that not all the parameters that were identiﬁed in the literature
as contributing factors to sewer deterioration are readily available in the database. Hence,
the model was developed using the available factors. After erroneous data processing and
reduction, the original data was transformed into a new format which would be suitable

for further analysis and modeling.

4.] DATA ACQUISITION

The dataset used for the development of the prediction model using artiﬁcial neural
network for sewer pipes is described in this section. This information includes the
background information of Atlanta’s SSES surveys, attributes of the data, condition

assessment and rating standards and condition summary of sewer group one.

96

4.1.1 BACKGROUND INFORMATION ON

ATLANTA’S SSES EFFORTS
The dataset used in this thesis is acquired from the Department of Watershed
Management, City of Atlanta sewer asset database. The City of Atlanta’s Department of
Watershed Management manages approximately 2,200 miles of sanitary sewer lines.
Atlanta’s sewer system is comprised of 260 sewersheds, which are prioritized into six
separate ‘Sewer Groups’. The First Amended Consent Decree (F ACD) deﬁnes
“sewershed” as a subdivision of a sewerbasin that typically consists of 10,000 to 50,000
linear feet of hydraulically linked sewers that are tributary to a point in the sewer system.
The formal deﬁnition of a Sewer Group is a group of sewersheds within a common level
‘ of priority for evaluation, rehabilitation and relief requirement.

The SSES work on sewer group one has been completed at the time of this study
and will provide the base data for the model development that may aid in prioritizing the
inspection work for other sewer groups. The total length of pipes within the Sewer Group
1 database is 304 miles, whilst the length of the entire sewer network is 2,200 miles —
providing a sample size of approximately 13.8%. The Sewer Group 1 study area is
illustrated in Figure 4.1 below. Total linear feet of sewer in 8G], i.e., inventory is
1,655,117 LF out of which 1,340,943 LF has been inspected by CCTV. The database that
was obtained for this study consisted of condition assessment data ﬁ'om sewersheds

SRV10 and PTC19A of Sewer Group 1.

97

 

 

 

 

 

 

 

Figure 4.1 - Map Showing the Project Study Area (Sewer Group 1)

4.1.2 CONDITION ASSESSMENT
The condition rating system used for the inspection consists of 119 criteria. The entire
rating system is composed of four sub-groups: Structural, Service, Construction, and

Miscellaneous. Each sub-group contains rating criteria describing both the characteristics

98

and the severity of the defects (see Appendix A). The prioritization and rehabilitation
decisions are taken depending on the defect type and severity. The severity of defects is
classiﬁed as follows:
" l - Excellent condition, no defects present
2 — Good condition, only low risk defects present
3 — Fair condition, pipe contains medium severity defects
4 - Poor condition, pipe contains high severity defects and collapse is imminent

5 — Failure condition, pipe is no longer functioning and is not structurally intact

4.1.3 CONDITION SUMMARY OF SEWER GROUP 1
The results of the sewer group one inspections are summarized in this section. The
predominant sewer structural deﬁciencies observed from the SSES inspections of Sewer
GrOup 1 include (see Figure 4.2):

. Circumferential cracks

. Circumferential fractures

. Multiple ﬁactures

. Holes

. Displaced joints (medium)

. Defective junctions

. Open joints (medium)

99

Holes
M ultiple fractures 6%
6%
Circumferential fractures

6%

      
   

Other
33%

Open joints (medium)
7%

Circumferential C racks

11%

Displaced joints
(medium)
6%

Defective junctio ns

15%
Figure 4.2 - Proportion of Structural Deﬁciencies Observed from SSES
Inspections of SG1
The predominant sewer service condition deﬁciencies observed from the SSES
inspections of 861 include (see Figure 4.3):
o Debris (general)
- Grease
. Light Encrustation
. Fine roots
. Fine Roots at joints

. Root masses

100

Root masses

6%

  
   

Fine roots
8%

Fine roots at joints
23%

Debris (general)

14%

. . Other
Light Encrustatlo n 8%

6%

Grease
15%

Figure 4.3 - Proportion of Service Condition Deﬁciencies Observed from SSES
Inspections of SG1
This research will focus on modeling the structural condition of the sewers and
will not account for the service and other defects. Table 4.1, following, gives the
percentage distribution of mainline sewer structural defects encountered during the SSES
inspection of SG1. See Appendix A for full list of the city of Atlanta’s sewer defects
coding system and abbreviations.

Table 4.1 — Mainline Structural Defect Summary (City of Atlanta)

 

 

 

 

 

 

 

 

 

 

 

 

Main Line Defect Summary
Condition Defect Code Number Percentage
of Total
Structural Pipe Broken B 2,136 4.9%
MH Cover Cracked or
Structural Broken BC 2 0.0%
Structural Crack Circumferential CC 4,672 10.7%
Structural Crack Lonjitudinal CL 1,848 4.2%

 

101

 

Main Line Defect Summary

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Condition Defect Code Number Percentage
of Total
- Structural Cracks Multiple CM 1,599 3.7%
Structural Connection Intruding CNI 234 0.5%
Structural Connection Defective CX 1,518 3.5%
Connection Defective '
Structural Intruding CXI l ,278 2.9%
Structural Deformed D 122 0.3%
Structural Brick Displaced DB 3 0.0%
Structural Deformation Horizontal DH 7 0.0%
Structural Dropped Invert D1 5 0.0%
Structural Deformation Vertical DV 4 0.0%
Structural Exposed Pipe EXP 13 0.0%
Structural Fracture Circumferential FC 2,658 6.1%
Structural Fracture Longitudinal FL 1,363 3.1%
Structural Fractures Multiple PM 2,542 5.8%
Structural Hole H 2,395 5.5%
Structural Soil Fissures HOL 177 0.4%
Structural Hole in Storm Ditch HSD 6 0.0%
Structural Joint Displaced Large JDL 990 2.3%
Structural Joint Displaced Medium JDM 7,239 16.6%
Structural Junction Defective JX 6,468 14.8%
Structural Liner Defect LN 175 0.4%
Structural Brick Missing MB 2 0.0%
Structural Multiple Soil Fissures MLK 77 0.2%
Structural Missing Mortar Surface MS 8 0.0%
Structural Open Joint Large OJL 644 1.5%
Structural Open Joint Medium OJM 2,909 6.7%
Structural Open Joint Slight OJ S 9 0.0%
Surface Damage
Structural Corrosion Large SGL 23 0.1%
Surface Damage
Structural Corrosion Medium SGM 278 0.6%
Structural Storm Manhole SMH 5 0.0%
Surface Damage Spalling
Structural Large SSL 66 0.2%
Surface Damage Spalling
Structural Medium SSM 95 0.2%
Structural Surface Damage Wear SW 197 0.5%
Structural Surface Wear Large SWL 335 0.8%
Structural Surface Wear Medium SWM 871 2.0%

 

102

 

 

 

 

 

 

 

 

Main Line Defect Summary
Condition Defect Code Number Percentage
of Total
. Structural Surface Wear SLight SWS 629 1.4%
Structural Collapsed X 57 0.1%
Structural Collapsed Manhole XM 1 0.0%
Structural

Total 43,660 100%

 

 

 

 

 

 

4.2 DATA COLLECTION AND PREPROCESSING

The effectiveness of an ANN model depends on the availability of reliable input data.
Finding data that represents or corresponds to the possible factors reviewed was
important for representing the physical cause-effect relationships. The reliability of the
data is measured by the amount of “noise” inherent in the data (Sacluti 1999). Noise is
data patterns that contain inaccuracies and discrepancies, which does not allow the model
to make proper associations between input and output patterns. Use of data with little
apparent noise would result in a more accurate and precise model.

Data collection involves evaluating all available data based on accessibility,
relative ease of obtaining long-term relevant data, and the prospect of future availability
of the same type of data for future models. This data must have characteristics that are
signiﬁcant for model convergence. If all the proposed model input parameters are used
for the model, the run times for model training will be exceedingly long, and hence
would result in insufﬁcient use of time. Also, if insigniﬁcant (or inappropriate) data is not
eliminated initially, the redundant input parameters will be treated as “noise” by the ANN

model, and as such may decrease the likelihood of the model convergence.

103

4.2.1 PARAMETER COLLECTION AND ANALYSIS
Because of the nature of pipeline deterioration, literature indicates that to fully predict
their condition, it is necessary to have a wide range of representative data parameters.
Due to limitations in the collection of input data, it became necessary to restrict the scope
of the output being predicted to the overall condition of the pipes rather than the
probability of speciﬁc type and magnitude of defects. To determine the entire
deterioration pattern, it was assumed that no improvement activities were performed over
the life of the sewer pipes. Investigations into available data indicated that a large number
of the suggested input parameters that met the requirements (reliable and available in
reasonable abundance) were difﬁcult to collect. The database after initial screening and
preprocessing contains the following variables that are considered to have an impact on
pipe condition and for training the model:

Pipe Material

Pipe Age

Sewer Size (diameter)

Section Length (MH to MH)

Sewer Type (Sanitary, Storm. Combined)

Average Depth of Cover

Slope/Gradient

After the initial review of the information, it was observed that data was either
lacking or too general for application within the scope of this study: simplifying

assumptions had to be made. For example, it was assumed that the study area was a

104

uniform soil type, had similar operating conditions. bedding conditions. loading factors.
etc.

Sewer pipeline sections were classiﬁed in the study to group sewer sections with
similar properties. After carefully sifting and comparing, pipe type (concrete. clay or
other) and size were selected as grouping factors. It should be noted that the grouping of
the pipes were done only to perform initial statistical analysis. It was found in this study
that data preprocessing is necessary for BPNN model development. This preprocessed

database was then used to train, test, and validate the BPNN model.

4.3 SOFTWARE SELECTION FOR DATA PREPROCESSING

The original data was stored in Microsoft Access database. Microsoft Excel was selected
as a data processing and analysis tool because of its versatility for spreadsheet analysis.
The amount of data in this study required an integrated statistics software package which
could provide complete control over data access, management, analysis and presentation.
The MiniTab statistical software package was selected for data processing because of its

power. ﬂexibility, and ease of use.

4.4 DATA ANALYSIS

A total of seven variables were used in modeling process as shown in Table 4.2.
However, depending on the availability of data. other variables such as source of sewer
(industrial and residential), soils surrounding pipes. ground water level, trafﬁc volume

above pipe segments, and frequencies of overﬂow, etc. identiﬁed in the literature review

can be included in future analysis. This information. however. was not available for this

 

 

 

 

 

 

study.
Table 4.2 - Variables used for Neural Network Modeling
Name of Variable Description of Variable
Length Length of pipe segments between manholes
in feet
Size Diameter of pipe segments in inch
Type of material Concrete, Vitriﬁed Clay. PVC. etc.
Age of Sewer Age of pipe grouped on a ﬁve year period
Depth of Cover The average buried depth of the pipe

 

Slope of pipe segments between manholes
Slope Slope = (Elevation of upstream invert -—
Elevation of downstream invert)/Length

 

 

Sanitary/Stormwater/Combined
Type of Sewer (The sewer database obtained for this
research consisted of only sanitary sewers)

 

 

Data analysis involved the examination of all collected data as means of
determining preliminary factor inﬂuences on the sewer condition. Furthermore, the
analysis was used as means of exposing data inconsistencies and errors. Microsoft Excel
was used to perform statistical tests on the collected data. Minimum, maximum, mean.
mode, standard deviation and correlation values were developed for all factors. The
correlation of an input provides an indication of whether an input will properly, or
satisfactorily, train with a neural network. For instance. an input with a good correlation
(value close to either 1 or -1) will typically be more inﬂuential in a neural network than
an input with a poor correlation (value close to O). The correlation, however. can be

deceiving as it only accounts for the effect ofa single factor. The intent of this research is

106

 

to develop the combined effect of different factors. To this end, correlations were
examined only as means of preliminary input inﬂuence determination and not used to
eliminate factors deemed unimportant. A histogram and a scatter plot were also
developed for each input factor. The purpose of the histogram is to provide a
representation of the range and consistency of the collected data. The histograms exposed
a number of gaps in the collected data. The scatter plots were primarily used to expose
data inconsistencies. Furthermore, the scatter plots provided preliminary input inﬂuences.
The following Figures 4.4 — 4.13 represent the various statistical analysis performed with

the data.

 

Pipe Material Distribution in Database

 

 

  
  

 

" Insufﬁcient Data
Samples —
Excluded for
Future Analysis

 

 

 

 

No. of Samples
(I)
O

 

 

CO VC DI PVC
Sewer Pipe Material

 

 

 

Figure 4.4 — Representation of the Sewer Material Distribution
LEGEND
CO - Concrete Pipe
VC — Vitriﬁed Clay Pipe
DI — Ductile Iron Pipe

PVC — Poly Vinyl Chloride

107

 

 

Sewer Age Distribution in Database

 

01
O

   

&
C
1

No. of Samples
on
O

 

 

<=1011- 16- 21- 26- 31- 36- 41- 46- 51- 56- >60
15 20 25 30 35 40 45 50 55 60

Sewer Age Group (years)

 

Figure 4.5 — Representation of the Sewer Age Group Distribution

 

 

Sewer Size Distribution in Database

 

 

 

 

No. of Samples

 

 

8" 10" 1 5" 1 8" 30"

Sewer Size (inch)

 

Figure 4.6 — Representation of Sewer Size Distribution

108

 

 

 

 

 

Average Depth of Cover

Distribution of Sewer Depth

 

N
.0‘
o

N
.0
o

 

_A
.0‘
o

 

 

 

 

_|
.0
O

 

 

.U'
o

.0
o

1 14 27 40 53 66 79 92 105118131144157170183
Number Observed

 

 

Figure 4.7 — Representation of the Sewer Depth Distribution

 

 

No. of Samples

Distribution of Sewer Condition Rating

 

Condition 1 Condition 2 Condition 3 Condition 4 Condition 5
Sewer Condition (1-5)

 

 

Figure 4.8 — Representation of the Sewer Condition Distribution

109

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Age - Condition Relationship (Concrete Pipe)
30
25 - I dozens?"
E 20 .. f . I Condition 2
E 15 , . _. in Condition 3
3 10 - l— in Condition 4
5 . i ' f ‘2 7 3- ‘ . I Condition 5
0 - ' ' ‘Ln ’ . ; L [-L
31-35 36-40 41-45 46-50 51-55 56-60 >60
Sewer Age Group
Figure 4.9 (a)
Age - Condition Relationship (Vitriﬁed Clay Pipe)
14
12
10 in Condition 1
3 il Condition 2
In 8 i . .
E 6 in Condition 3
5 4 :0 Condition 4
2 il Condition 5
0 ,
31-35 36-40 41-45 46-50 51-55 56-60 >60
Sewer Age Group

 

 

 

Figure 4.9 (b)

Figure 4.9 (a & b) — Representation of Sewer Age — Condition Relationships for
CO and VC Pipes

llO

 

 

l Depth - Condition Relationship (Concrete Pipe)

 

 

 

 

 

 

 

 

 

 

5 _ - . . _
2’ 4 ~ e
a . ,
N i , r
0: 3 -———m 4
C . f
o ‘ . .
E 2 WWW—w
W e ‘4 e
a 1 1 . ‘ g .
0, . "i""“’4:’ A
0.0 5.0 10.0 15.0 20.0 25.0 1
E Depth of Sewer I
1
Figure 4.10 (a)

 

l Depth - Condition Relationship (Vitrified Clay Pipe)

9°
moi

 

N
or

 

Condition Ranking

 

O _L
OUI-‘OIN

0.0 5.0 10.0 15.0 20.0 25.0
Depth of Sewer

 

Figure 4.10 (b)

Figure 4.10 (a & b) - Representation of Sewer Depth — Condition Relationships for
CO and VC Pipes

111

 

Size - Condition Relationship (CO Pipe)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

l 5
l U!
l .E 4 e
l "e‘
I g 3 e e e e e
i 5
l E 2 e e e e
‘e’
o 1 e e 4v % e
o
0 a w I A i A rL A .
l O 5 10 15 20 25 30 35
: Sewer Size (inch)
1-- _ DBL,“ m...“ _ __ - L C
Figure 4.11 (a)
a Size - Condition Relationship (vc Pipe)
1
. a, 4
C
E
g 3 —~—- 4 4
n:
.2
f”: 1 r e
i S
l 0 0 . . 4
g 0 2 4 6 8 10 12 14
l Sewer Size (inch)
1
l

 

 

Figure 4.11 (b)

Figure 4.11 (a & b) — Representation of Sewer Size — Condition Relationships for

CO and VC Pipes

112

 

 

Length - Condition Relationship (CO Pipe)

 

 

 

 

 

 

 

 

 

 

5 , .
oi , - ,.
.E 4 ‘ 4
C , _ - .. .
g 3 -W e.
.3 2 W , f e
o 1 "
o p A . , f
0 ' r 1 ﬂ .
0 100 200 300 400 500 600 700 ‘
l Sewer Length (MH to MH)
l____ ___ _- _ _ _,_LL_HM
Figure 4.12 (a)
Length - Condition Relationship (VC Pipe)
a: 4 -
.E
E 3 at W»
n: .
= 2 +————40———e
.2
g 1 me e
C
O
0 0 A 4. .
50 100 150 200 250 300 350 400 ‘
Sewer Length (MHto MH)
Figure 4.12 (b)

Figure 4.12 (a & b) - Representation of Sewer Length — Condition Relationships for

 

 

 

 

 

 

 

 

CO and VC Pipes

113

i Slope - Condition Relationship (CO Pipe)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

i s
i or
‘ l I: A
l 3 v
| g A
' m V
C
l 3.
' ‘6
I i:
o
1 ° .
f -0.100 0.000 0.100 0.200 0.300 0.400 0.500 0.600 0.700 0.800 .
! Sewer Slope
Figure 4.13 (a)
} Slope - Condition Relationship (VC Pipe)
or 4
I C
l 3
£ 5 3 M <0
i:
l 5 2 -H
1 § 1 e————+
2 o
l U 0 t
; 0.000 0.020 0.040 0.060 0.080 0.100 0.120 0.140
i Sewer Slope
Figure 4.13 (b)

Figure 4.13 (a & b) — Representation of Sewer Gradient — Condition Relationships for
CO and VC Pipes

4.4.1 INTERPRETATION OF RESULTS
The signs of the parameter estimates are consistent over the estimation results.

Figure 4.4 represents the histogram representing the different Classes of materials in the

114

sewershed database in consideration. It can be seen that the majority of the sewers were
concrete pipes constituting about 75% of the sample with vitriﬁed clay pipes being the
second highest number followed by PVC and ductile iron pipes. The DI and PVC pipes
will be excluded from the model due to the insufﬁcient number of samples.

Figure 4.5 represents the distribution of the sewer age groups of the samples in
the database. It can be observed that the age of pipes in sewersheds in this study ranges
between 31 years and up. This might pose a setback as the model being developed will
not have learned the behavior of younger pipes.

Figure 4.6 represents the sewer size distribution obtained from the database. It can
be seen that the majority of sewer pipes are in the 8-inch category. Figure 4.7 shows the
distribution of the depth of cover within the group and Figure 4.8 shows the condition
ranking of sewers indicating that the average condition of sewers in these sewersheds are
in condition level 3.

Figures 4.9 to 4.13 show the plots of the present condition of the sewer with age,
depth, size, length and gradient of sewers. A few of the observations from the above
graphs are summarized in the following paragraph.

The relationship between the age and sewer condition ranking as observed in
Figure 4.9 indicates a slight correlation between the two. The condition of the sewers
tends to worsen as it ages. The relationship between the sewer depth and condition in
Figure 4.10 shows a slight pattern of increase in condition ranking with the depth. This
may be attributed to high overburden pressures acting on the pipe conforming to the
literature. The relation between size of the sewer and condition does not show any

correlation within this sample. The relationship between sewer length and condition

115

ranking is illustrated in Figure 4.12. It can be visually observed that the condition ranking
of the sewer has a higher value with the increase in length of the sewer. This scenario
may be related to the facts presented in the literature review. As per Figure 4.13, the
condition ranking increases as the slope or the gradient of the pipe increases.

Although there is slight relationships observed from the plots, there is no
substantial correlation seen with the condition of the sewer and the individual parameters.
This gives a strong case for adopting neural network modeling methodology as it is

capable of capturing such subtle global relationships with ease.

4.5 DATABASE TRANSFORMATION

After necessary parameters have been identiﬁed, the original database was
transformed to a new format that can be used for BPNN model development and analysis.
Table 4.3 represents the database that will be used for further analysis and modeling. This
database will then be fed ﬁrst to NetMaker, a preprocessing utility provided by
Brainmaker in order to assign the input-output parameters and then convert into ﬁles that

can be utilized by BrainMaker for training, testing and validation.

116

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

N :3. ea e e e e e e _ e e e e o> 2 SN ”:30
n 83. 2: e e e e e e _ e e e e o> m an ego
n wee... ed e e e e a e e e e e e o> e e: E0
N 83. 2: e e _ e e e e e e e e co 2 e2 ":50
N 82. me e e e e e a e e e e e 8 w ea "E5
n 83 a: e e _ e e e e e e e e 8 m. an EC
_ 8e... 92 e e _ e e e e e e e e co a 8e ":50
m 33 n: e e e e _ e e e e e o oo 2 ”2 CBC
_ 83. E e e _ e e e e e o e e co 2 E Ebo
_ Mae 2: e e _ e e e e e e e e co 2 can “.50
_ see a: e e _ e e e e e o e e on E SN ".50
82.55 5.2.6 an: 8A 8 m... e... we 3. mm 8 e" r.— e_uv .32 so 593 ~59
".5. er. .3 .3 .s. .8 .8 .3 a: .= a... 2.: e59
2:.5

95.5 ou< Stem

 

 

 

 

 

 

 

 

 

508588 «55.0% «o cocoa—om 2: Coca omen—zen 35.888... 2: mo ago...— - «.9 039—.

117

4.] SUMMARY AND CONCLUSIONS

This chapter presented a detailed discussion about the data acquisition, preprocessing and
analysis. The raw database was transformed into a standardized format is ready for neural
network model development. The available parameters for the model development were
identiﬁed and their relevance examined through statistical analysis. Although certain
variables have a low amount of correlation to the condition of sewers and deterioration, it
is important to include such parameters as neural network is capable of capturing even

subtle relationships.

118

CHAPTER 5

MODEL DEVELOPMENT

The previous chapter described the preprocessing and statistical analysis of the acquired
data. This chapter deals with the detailed account of the model development process. The
model development in this thesis includes training and testing. Training a neural network
involves repeatedly presenting a set of examples to the network. The network takes in
each example, makes a response as the output, checks this response against the correct
answer, and makes corrections to the internal connections. Testing the network is the
same as training it, except that the network is shown with the examples it has never seen
before, and no weight adjustments are made during testing. Validation occurs after the
neural network has been developed. Validating a network consists of presenting it with
new input data and gathering the network outputs. Unlike testing, there is no known
output, only known inputs in the validation. Due to time and resource constraints
validation was not performed for the model developed in this thesis. The following
sections provide a detailed discussion about the model development process. Figure 5.2

represents the neural network design structure used in this thesis.

'5.1 DATA SUBDIVISION
In this research, the purpose of preprocessing the data was to establish a database which
can be directly used for further model development. The processed data set was further

divided into three sub data sets. As shown in Figure 5.], approximately 85 percent of data

119

were used for network training and 15 percent of data were used to test the generalization

ability of the network when facing data unseen in the training period.
15%

i—.‘.—
I a Training

I Testing

 

85%

Figure 5.1 - Database Subdivision

5.2 SOFTWARE SELECTION
A developmental platform is usually a requirement to train neural network. Often referred
to as neural network simulators, these platforms are commercially available. Some
factors worth considering when choosing a suitable neural network simulator are required
level of expertise, complexity, and the pre- and post-processing facilities. In this study,
BrainMaker, a commercially available neural network simulator distributed by California
Scientiﬁc Software, was used for the development of the proposed neural network model.
The BrainMaker neural network simulator uses popular backpropagation training
algorithm for network training. BrainMaker reads three kinds of neural network ﬁles:
deﬁnition ﬁles, fact ﬁles, and network ﬁles. BrainMaker also creates different types of
statistics and output ﬁles. They are all human-readable and editable. A deﬁnition ﬁle
describes everything there is to know about the network to BrainMaker, such as the
number of neurons in each layer, the type of data, and what is going to be displayed on
the screen. BrainMaker uses the deﬁnition ﬁle to create the neural network. The default

extension for the deﬁnition ﬁle is “.def”. A fact ﬁle gets the data into BrainMaker. There

120

are fact ﬁles for training, testing, and running. The default extension for the training fact
ﬁle is “.fct”, for testing it is “.tst”, and for running it is “.in”. A network ﬁle is created by
BrainMaker during training using the data in the training fact ﬁle and the instructions in
the deﬁnition ﬁle. The network ﬁle contains the actual connection information as well as
training parameter information. The default extension for a network ﬁle is “.net”. The
network ﬁle plus the testing fact ﬁle are used for testing. When the answers are not

known, the training network ﬁle plus the running fact ﬁle are used for running.

' ngﬁatn}. ‘ D r- w w 3,}- w .

  
  

    

Acquire Sewer Data
Create Network Files

 

a r .1.

.n . ' a I
. m .33 by)!- 2.3.1.;

 

Train Network

 

 

l . a. .
Trained ‘ . N0

Successfully
9

YES

“at: 1." fan's-a“ rt" ....‘.' "73’

Test Network ‘

 

 

 

 

 

Figure 5.2 - Neural Network Design

121

5.3

NEURAL NETWORK DESIGN, TRAINING AND TESTING

To develop a neural network model, one must decide precisely what the neural network is

expected to forecast, generalize, or recognize. The original database needs some

preprOcessing before it can be used for training and testing. Then the neural network

needs to be trained and tested with some certain rules. Finally, a new data set is used to

validate the neural network.

5.3.1

FRAMEWORK FOR NEURAL NETWORK DEVELOPMENT

In this research, the BPNN model was developed through a procedure presented in Figure

5.3. Each developing stage shown in Figure 5.3 is discussed as follows:

>

Database Review — Database investigations was performed to ensure that the
database contained sufﬁcient information for neural network development and for
pipeline condition prediction.

Data Preprocessing - The database was processed and reorganized to form a new
database ready for model development. The new database was divided into
subsets to create a training data set and testing data set.

Network Design and Training - Several different architectures of the neural
network were designed and trained in order to obtain the best architecture
resulting in the best testing performance.

Network Testing and Error Analysis - Each of the architectures were tested
independently using the testing data set. Assessments were made for

generalization ability and accuracy.

122

2 Network Implementation - The best network architecture was then chosen and

proposed to be embedded in the working environment.

Network Develonment Procedure

 

Database review

 

 

 

 

Data analysis and
preprocessing

 

 

Network design
training

 

 

 

 

Network testing,
error analysis

 

 

 

 

 

Implementation

 

 

lZ-‘:(>

f—————:>

Documentation Produced

Data sources

 

Identifying
parameters

 

 

Architecture and
parameter settings

Performance
analysis results

 

Validation

 

ITIJIJM U

Figure 5.3 - The Procedure for Neural Network Development

It must be kept in mind that the model being developed will be able to predict the

probability of sewer pipe being in a certain severity level, but is not meant to predict the

probability for individual pipe defects. The main purpose of this model study is to

demonstrate the possibility of using ANN modeling for predicting the sewer condition so

that physical inspections can be prioritized. As such, further research and model

development will likely be necessary.

5.3.2 NEURAL NETWORK TRAINING AND TESTING

The speciﬁcation of input and output for the BPNN is presented in Table 5.1. The BPNN

is designed to predict the condition of sewer pipes given the variables that are identiﬁed

123

to affect its deterioration process. Training a neural network involves repeatedly
presenting a set of examples (facts) to the network. The network takes each input, makes
a guess as to the output, checks this guess against the output (correct answer), and makes
correCtions to the initial connections (weights) if its guess is incorrect. This process is
repeated for each fact in turn until the network learns the facts well enough to be useful.

Table 5.1 - BPNN Architecture

 

 

 

 

 

 

 

 

 

Neuron Type Neuron Number Description Range of Variables
Inputs 1 Pipe Material CO, VC, PVC, etc.
2 Pipe Age Age of Sewer in 5-yr
increments
3 Pipe Diameter 8- inches and up
4 Section Length Manhole to Manhole Length
in feet
5 Depth Average Depth
6 Slope Gradient of Pipe
7 Sewer Type Outfall/Lateral
Output 1 Condition 1-5

 

 

 

 

 

 

As described in Chapter 4, the above inputs are ﬁrst fed into NetMaker for
labeling the inputs and the outputs. The proportion of training and testing facts are
assigned before saving the ﬁle as BrainMaker readable and executable ﬁles. Now the
database is ready to be modeled using the BrainMaker software. Figure 5.4 depicts the

typical structure of the neural network with the given set of input parameters.

124

    

Pipe

o. it A1

‘ 1‘ ‘ v0].

1‘ s’ “ 3,2 . .
Section ‘Aﬂ'ﬂfﬂé v Condition
Len... Vilt’ii‘i‘".
v‘ I: '\ ’
'QJ‘QZ'

 

Gradient VI

6“)
/// ‘10

Type

 

Input Layer Hidden Layer Output Layer

Figure 5.4 — Schematic Architecture of the Neural Network

The training ﬁle created using NetMaker is accessed through BrainMaker for
modeling analysis. Histograms and the Network Progress Display, two useful tools

provided by BrainMaker, can help determine whether the network is making progress in

125

training and still has capacity to learn. Figure 5.5 (a) shows the training histogram of a
neural network (19 hidden neurons, 83 epochs) with the horizontal axis representing the
values of connection weights. This bell-shaped histogram indicates the network is healthy
and still has the capacity to learn. Another tool is the Network Progress Display, as
shown in Figure 5.5 (b). The top part of the screen shows a histogram of the errors over a
training run. It gives a quick snapshot of the distribution of errors, making it easy to see
how close the network is to achieving the pre-speciﬁed tolerance level. The bottom part
shows the progress of the Root Mean Square (RMS) error, which is deﬁned in Chapter 3,
during training. This graph shows how well training is progressing over runs. Figure 5.6
illustrates the step by step process of neural network model development for sewer

condition prediction.

 

P Siiiileiil UrdiiiMokrr MM‘X i‘ittt'lt'riilt'tl I 11")? 1111141,: 1 .11‘ 18 (it'll

mummhmcmm

Intuit: “:83. Facts: "IRWIN: "C lam: I.“ Tolerance: 0.2”
Fact: 5 Total: 11451 Dad: 19 Last: 63 Good: 34 Lost: 96 M: .3

Ian It fable Condition
H . - . . . . . W‘ :l .....
- rr-rs hn: -----

1.12; " zt-ao m

 

 

- Hidden Comedians

Slope 0

 

Hidden - Connections

     

 

4.0 0.0 0.0

 

 

Figure 5.5 (a)

 

 

 

 

Figure 5.5 (b)

Figure 5.5 (a & b) — Connection Weights Histogram and Network Progress in Training

126

 

 

Input Parameters (Independent Variables)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

v v v v i _ii
Type Length Dia. Matl. Age Depth Slope
l J J l l l
7

Import the Data into NetMaker
- Manipulate Data '
- Label Data
- Allocate Training & Testing
%
- Create BrainMaker Files

 

 

 

 

 

 

BrainMaker - Open Training File
- Model Conﬁguration
- Hidden Neurons, Layers,
Tolerances and Training
Criteria

 

 

 

 

 

 

Train Network Until it Converges to the
Maximum Extent

 

 

 

 

 

Test Network with Previously Allocated
Data

 

 

   

  

Tested No

 

 

Well??

    

 

 

Trained Model Ready to be Implemented

 

 

7

 

 

Update Model and Retrain when New
Data becomes Available

 

 

Figure 5.6— Neural Network Modeling Process

127

 

5.3.3 DETAILED OVERVIEW OF MODEL DEVELOPMENT

The ﬁrst step in creating a neural network model using BrainMaker is to
preprocess the data using NetMaker tool provided in the software bundle. NetMaker has a
number of utilities useful for manipulating the data to present the best possible data
structure for model development. It is here that Inputs and Outputs of the model are
assigned. The dataset is split randomly into training (85%) and testing (15%) facts using

the NetMaker preferences. Figure 5.7 shows the typical NetMaker window.

   

 

 

iii. hm Lllllllill'lt'l'did lliiIIIIM-illi” in
I. CM '4. l“ m 9” MCI
1'” W 10M m m in h“ m m h" w m m w m 2‘
Luge. met-1 co i It 1 (I101Ii-Isl"-20!II-ISIZGNJn-SSINIIIGIWJ 515g IIII] |'
#f i I I I I I I I I I I i I
__z_:iso u i I I I I I I I I I I i I
_3‘273 15 i I I I I I I I I I I 1 I
4362 30 1 I I I I I I I I I I i
L“ 30 i I I I I I I I I I I 1 I
__I__iII so i I I I I I I I I I I i I
_7_2n is i I I I I I I I I I I i I
_I_III so 1 I I I I I I I I I I i I
.3.” II 1 I I I I I I I I I i I
i_I324 is i I I I I I I I I I i I
L’" 15 i I I I I I I I I I I i I
_iz_357 15 i I I I I I I I I I I i I
L‘“ is i I I I I I I I I I I i I
_14_23I I i I I I I I I I I i I I
1s as I i I I I I I I I I 1 I I
Em I i I I I I I I I I i I I I
__i1_zoI I i I I I I I I I I i I I I
u 332 I i I I I I I I I i I I I
Ens I i I I I I I I I I i I I
__2_I_245 I i I I I I I I I I i I I I
_zi__isz I i I I I I I I I I i I I I
3.2“ I i I I I I I I I I I 1 I I
2_3254 I i I I I I I I I I I i I I
_szss I i I I I I I I I i I I I
L267 I i I I I I I I I I i I I I
Loco I i I I I I I I I I i I I I ,
S D

Figure 5.7 — View of the NetMaker Data Processing

After the completion of data preprocessing, the ﬁle is saved to a BrainMaker ﬁle.

NetMaker creates three ﬁles, namely, deﬁnition (.def), training (.fct) and testing (.tst).

The data is now ready to be operated with BrainMaker for modeling.

128

5.3.4 SELECTION OF OPTIMAL NUMBER OF HIDDEN NEURONS

Selection of optimal number of hidden neurons is an important issue in the neural
network training process. The goal of training is to obtain a neural network with best
generalization capability. Generalization is deﬁned as the ability of a neural network to
store in its weights general characteristics which are common to a group of examples.
Usually, a neural network with too few hidden neurons will not be able to learn
sufﬁciently from the training data set, whereas a neural network with too many hidden
neurons will allow the network to memorize the training set instead of generalizing the
acquired knowledge for future unseen examples. Unfortunately, there is no precise
formula for determining the ideal number of hidden neurons for a given application.
There are several ways to determine a good number of hidden neurons. One solution is to

train the neural networks with the number of hidden neurons calculated using the

 

 

formula:

Number of Hidden Neurons = of Data sets — Outputs ---------- (5.1)
C (# Input + # Output + 1)

Where, C = 2 — 5.

Therefore the # ofNeurons = (167 — 1) / 3 * (15 +1+1): 4 Neurons
The second equation suggested in BrainMaker manual is:
Number of Hidden Neurons = (# Inputs + # Outputs)/2 ---------- (5.2)
= (15 +1)/2 : 8 Neurons
The third solution is to begin with a small number of hidden neurons and add
more while training if the network is not learning. In this research, the ﬁrst method (4
Hidden Neurons) was used to train initially and gradually more neurons are added to train

several neural networks with varying number of hidden neurons. The neural network that

129

resulted in the least testing error was selected, resulting in the best generalization
capability. The “testing while training” method was used to trace the testing errors
(generalization ability) of the neural network during training process. After training, it
was convenient to ﬁnd the best network with the least testing errors.

In order to identify the best BPNN model for pipelinecondition prediction, a
variety of neural network architectures were experimented in this study. Table 5.2
presents the training and testing errors resulting from typical BPNN architectures that
were tested in this effort. Since the generalization capability of the neural network is
typically represented by the testing errors, the testing RMS error was selected as the
major criterion to evaluate the BPNN performance. It can be seen that Model #7
presented in Table 5.2 with 10 hidden neurons resulted in the best BPNN model with
lowest testing error.

Different network architectures were tried with the same facts to examine which
architecture best suites the problem in hand. In all the models experimented, 85% of the
total facts were used for training and 15% of the facts were set aside for testing. The
number of hidden layers used for all the models was one and the training and testing
tolerances were set at 0.3 and 0.3 respectively. It is assumed that these tolerances were
acceptable based on the fact that values for all the parameters that contributed to the
pipeline deterioration were not factored in the model due to the unavailability of data
attributed to the lack of monitoring of such data. After the model with optimal number of
hidden neurons was selected from this prescreening process, it will undergo further

processing in order to enhance the tolerance and accuracy levels.

130

Table 5.2 — Training and Testing Errors of Different BPNN Architectures

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Hidden Neurons

Model Architecture RMS-[mums RMSTEanG
1 15-4-1 0.2412 0.1881
2 15-5-1 0.2644 0.2115
3 15-6-1 0.2422 0.1882
4 15-7-1 0.2499 0.1769
5 15-8-1 0.2458 0.1795
6 15-9-1 0.2378 0.1760
7 15-10-1 0.2386 0.1478*
8 15-11-1 0.2411 0.1746
9 15-12-1 0.2713 0.1971
0.3 _ _ .
g M
W 0.15
in . .
E 01
' —o—RMSTRAINING
0.05 —I— RMSTESTING
0 . . . l
4 5 6 7 8 9 10 11 12

 

Architectures

131

Figure 5.8 — Graphical Representation of Training and Testing Errors for different

 

 

Figure 5.8 presents the RMS error and average error of the neural networks with
different number of hidden neurons. It can be seen that the network with 10 neurons in
the hidden layer resulted in the least RMS error and average error. Hence, Model #7 is
chosen as the best architecture and will be used in further analysis and model
development using BrainMaker.

From the BrainMaker interface, the number of hidden neurons is set to 10 and the
necessary values for the training and testing tolerances are made before the model is
ready for training. The training and testing statistics ﬁles were activated to capture the
network progress statistics that would be helpful in identifying the best run. Figure 5.9
shows the network during the beginning of the training. Figures 5.10 and 5.11 show the

network after training and testing processes with tolerances of 0.3 and 0.3 respectively.

 

 

 

1" \l 1111'! l HIHIIIMH‘H" MMK ﬁirt-lnrtil‘ 1]! 111,1. ilril

I" (a worm PING! (mm 0”
Intern “37:27 Pun: 71'6“..ch hon: I.‘ ‘leloeoneo: 0.3.
Feet: 2 total: 2815 led: I but: )9 (bed: I? Loos: 1.3 In: 146
out! IA! couom

1. Out 2.1?“
It 8122 he: 3.

248.” 8.”:

00 cc

1.- I.

31 35 36-40

I.“ I.

“-15 “-50

I.“ r.

Sg-SS 5:40

)6.“ DEPTN lﬂ Connection

I.“ 9.7.14

SLOPE

o.eeae

 

 

 

1.0000

 

 

 

Figure 5.9 - Neural Network Training Progress

132

Table 5.3 — Network Architecture and Speciﬁcations of Model #7

 

 

 

Inputs Hidden Layers No. of No. of Training Testing
Neurons Training Testing Tolerance Tolerance
Facts Facts
15 10 1 142 (85%) 25 (15%) 0.3 0.3

 

 

 

 

 

 

 

 

Figures 5.10 and 5.11 shows the training and testing results of the model #7 conﬁguration

(15-10-1).

 

f‘ \liuit'iii itmiiiM-ii-ri MHJI itri i-it-iiili'ii [i.‘llh'l 54 2111” i[
It no acorn Dee-nu W Out-v

 
 
 

Holt U:.:Il Pocu: PHIL!" been: 0.- toloI-oneor I.”
Feet: 5 Ion]: 2476483 I“: 28 Loot: 11 Good: 122 Loos: 121 he: rm

comm
Out: 2.3]!3
he: I.“

 

 

 

 

ILIIII

 

 

 

Figure 5.10 - Training Results of Model #7

The training is stopped when the neural network settles at the lowest possible

training and testing errors and there is no visible convergence in the training statistics.

The summary of training and testing statistics is given in Table 5.4 below.

133

 

'{ Student Hldluﬂah'l Mnix in micron-1| Ii lNAl iir'll

III (a (borne Pam Cm Meir
0.1th I:u:. Poets: 'lmL-‘IE
25 led:

feet: 35 total: 3
W" tar M!!!

0. 1." Out: 2.3303
1.93111 SIZE he: 2.“

3..” 0.”.
CO C
0.“ I.“
31-35 36-40
0." .
“-45 06-50
0.“ .
Sl-SS 56-60
I.“ I.“

 

Leon 0

been: 0.- Tolenneo: 0.3.
Good: 22 Lou: 0

in: 17440

 

 

 

- Hidden Connectieno

 

Hidden -

 

Connections

  

 

 

 

 

 

Figure 5.11 — Testing Results of Model #7

Table 5.4 - Summary of Training and Testing Results (Model #7)

 

 

 

 

Training (156 Facts) Testing (27 Facts)
Bad Good Bad Good
20 (14%) 122 (86%) 3 (12%) 22 (88%)

 

 

 

 

Although the model learned 86% of the facts presented to it and predicted 88% of

the testing facts right, the tolerances are on the higher side. Figure 5.12 shows the plot of

the predicted condition versus the actual condition. It can be observed that the model

almost always under predicted the sewers in condition rating level 3. This indicates that

the model will have to be further ﬁne tuned to pick up the condition rating 3 scenario.

134

 

 

9"
0301

1"
or

  
 
 
 

_I
401k)

+77PrediCted-Conditionl
+ﬁgtual 9°ndi£i°ii - l

Sewer Condition Rating

.0
our

13 5 7 91113151719212325
Testing Facts

 

 

 

Figure 5.12 -— Plot showing the Predicted and the Actual Condition of the
Testing Sample (Model #7)

5.4 MODEL PERFORMANCE ENHANCEMENT
There are some tools available in BrainMaker that help in enhancing the performance of
the models. The application of these utilities to get better results with a lower tolerance is
explained in this section. The main objective of this effort is to converge the neural
network training as much as possible with the lowest tolerances so that the network learns
most number of facts presented. This model will then be tested with a lower testing
tolerance of 0.25 as compared to the tolerance of the initial model which was set at 0.3.
The network was set to test and save every 20 runs automatically to enable the best
network conﬁguration to be selected after training.

Model #7 had previously learned 86% of the facts presented with a testing
tolerance of 0.3 and it had tested successfully 88% of the facts with testing tolerance of
0.3. First, the model 7 is opened in BrainMaker to be retrained and tested with a tolerance

of 0.25. One of the techniques available in the software is to randomize the network

135

connections and to add some noise to the network. This helps the network to learn some
hard to learn facts which it has difﬁculty learning. The connections were randomized
with a constant of 5 (default) and a noise value of 0.15 is set. The network was then
retrained until it learns the most number of facts and also tests successfully. The
reconﬁgured network was able to learn maximum of 92% of the training facts. The
training is stopped at this point and the facts are ready to be tested. Figure 5.13 shows the

trained model with 131 of the total 142 facts learned by the network.

 

f \ttiiit‘iii Hi iiinMnIkm' MM): in t vii-Mimi li-‘llNi ‘i'mtl Hi I I

RICOOIIIOIPoo-Iosmodn
.200

             
  

this“ It's. Pun: PIMLJec lam: 0.- l'oloI-onoo: 0 , - - * _~~ -~—-- ~ ‘ ~ -
Feet: 5 loco]: 1032193 led: 11 Loos: 10 Good: 131 Loot: 122 Mi: 19945 *~“3~§315‘~7‘4?‘§."5‘g
wt, L" his: 1.55;; 00 oHidden Connedtone

mm 012! Pen: 2.“

ll .
3.." 0.”.
0.- 1.-

31-35 36-4.
0." .

41-45 461-50

-0.0 0.0

Hidden - Connections

   

0.” 0.7030 10
DOPE

 

 

 

1.0000

 

 

 

 

Figure 5.13 —The Model after Training Maximum Facts
The neural network ﬁle was initially programmed to test the allocated data
automatically for every 5 runs while training progresses. The testing statistics ﬁle that
tracks the network conﬁguration was pulled to select the best network that has the least
testing as well as training RMS errors. Figure 5.14 shows the snapshot of the testing

phase. This model tested 21 out of the total 25 facts set aside for testing with a tolerance

of0.25.

136

 

Newman-norm”

   

      

3““, “:"z" he": P]|.|,,ggg tau-n: 0.- I'd-nee” .35. W‘"‘“' '" "'“" - " ”-1” ~
’“g: 35 Ind: 25 0nd: 1 Lou: 0 Good: 31 k": . h": ‘”‘s _WE;EM€;.-v—;£Fe§él- ‘,“.&
our! IA! 0“ will; 50 - Hidden Conncuiene

MI "2! Pen: 2.“

3'.” '0 on

I

31 35 3:40

“-15 40-50

I.“ I.

sa—ss 5:40 4.0 M

>60 Imu

Connections

 

 

 

1.0000

 

0.000

 

 

 

Figure 5.14 —The Trained Model while Testing

The network information for this model can then be accessed from the training

statistics ﬁle as shown in Figure 5.15. The testing conﬁguration of this network is given

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

in Table 5.5.
Table 5.5 — Testing and Training Network Statistics
it‘ll? i323: Good Bad “Willi?" A1153? Ell-“oi
Training Conﬁguration
19945 142 131 11 0.25 0.1770 0.1966
Testing Conﬁguration
19945 25 21 4 0.25 0.1649 0.1979

 

 

 

 

 

 

 

 

 

137

E ilNAl lllAlNINGSllllS Nah-(Mil

 

TDKFICE‘

3900‘ 2027015
09007 202305

19000 202409!
19009 2024241
1903 2024303
19051 2024525
1909 2024667
1909) 101400.
1009 207435 1

19905 2025511
19906 202 $555
19907 2016797
1090 1025919

10011 2027303
XSSIZ 2817507
1991) 2027649
00014 2027791
1991 203793]

19941 2011625
10942 2031‘67
10943 2031909
19044 20320! I

a

VODOVn

 

omuwurs
,,

ovoonmo

 

Toni-ed earn olcrence voError lszrror murmur! A
.0000 .2500 .1051 .1961 00:01‘
.0000 .2000
.0000 .2500
.0000 .2500
.0000 . 00
.0000 .2500
.0000 .2500
.0000 .2800
. 00 .2500
.0000 .2500
.0 00 .2500
.0000 .2500
. 00 . 00
.0000 .2900
.0000 .2500
.0000 .2500
.0000 .2500
. 00 .2 00
.0000 .2500
.0000 .1500
.0000 .2500
. 00 .2 00
.0000 .2000
0000 .2500
.0000 .2500
.0000 .2500
. DC .2 00
.0000 .2500
.0000 .2500
.0000 .2500
.0000 .2500
.0000 .1900
.0000 .2500
.0000 .2500
.0000 .2300
.0000 .2500
.0000 .2500
.0000 .2500
.0000 .1500
.0000 .2500
.0 00 . 00
.0000 .2500
.0000 .2500
.0000 .2500
.0000 .1500
. 00 .1500
.0000 .2500
.0 00 .1500
.0000 .2500
.0000 .1500
. 00 .2 00
.0000 .2500
.0000 .2500
.0000 .2500
.0000 .2500
. 00 .2500
.0000 .2500
.0000 .2500
.0000 2500 _

 

 

 

 

 

 

 

 

Figure 5.15 —Network Training Statistics for Model

 

 

3.5

2.5

1.5

Sewer Condition Rating

0.5

  
  

+ Predicted Condition
——n— Actual Condition

13 5 7 91113151719212325
Testing Facts

 

Figure 5.16 - Plot showing the Predicted and the Actual Condition of the

Testing Sample after 19945 Runs

138

 

The model automatically tests 15% or 25 facts that were set aside initially with a
testing tolerance of 0.25. The model tested 21 out of 25 facts as good or 84% successfully
within the 0.25 testing tolerance. Figure 5.16 shows a comparative plot of the predicted
condition versus the actual condition of sewers. This graph indicates a better trend than
the earlier results represented in Figure 5.12.

Finally, the network is again calibrated to a lower testing tolerance level of 0.2
and then trained longer. The network in this mode could test only 15 out of 25 facts, but
with higher accuracy level. Figure 5.17 represents the plot of the predicted and the actual
condition. This plot has more accurate facts than the previous models because of the
stringent tolerance levels. It is assumed that a testing tolerance of 0.2 is acceptable and
will be able to give a fair judgment as to what the condition ranking of the pipe will likely
be, given the combination of input parameters. The model’s accuracy can be increased as
more input parameters identiﬁed to affect the sewer condition are available for modeling.

Table 5.5 — Testing and Training Network Statistics

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

1111.1? 32?: Good Bad “3111521216“ A33?“ SMS
Training Conﬁguration

40995 142 126 16 0.2 0.1568 0.1868
Testing Conﬁguration

40995 25 15 10 0.2 0.1367 0.1792

 

 

 

 

 

 

 

139

 

 

3.5

2.5

   
 
 

1.5

+ Predicted Condition
+ Actual Condition

Sewer Condition Rating

0.5

13 5 7 91113151719212325
Testing Facts

 

 

 

Figure 5.17 — Plot showing the Predicted and the Actual Condition of the
Testing Sample (40995 Runs)

The neural network automatically conﬁgures the weights of the hidden neurons
based on the training. Figure 5.18 represents the weight matrices for the hidden neurons.
There are two blocks (input to hidden and hidden to output) in the Figure 5.18. The ﬁrst
block has 16 numbers per row (15 inputs plus the threshold) and there are 10 rows (for
the 10 hidden neurons). The second block has 11 numbers per row (10 hidden plus the
threshold) and there is only 1 row (for 1 output). Weights from threshold neurons always

go at the end of the row.

140

 

F Untitled Notepad

Fl: Ed Format M Heb
[77315—31:577'-3.ffs'6”57..370775.333-"2.3390 6.053? 1.5676 4.95“ -6.7076 7.302 4.“): 7.9994 2.9154 0.23“

 

1.8444 o7.9942 -7.9972 6.3362 6.3512 -3.3596 -7.9472 -6.I616 0.9140 1.7350 -7.9912 '2.0114 -7.9994 -4.0554 7.9994
9663226 -7.I710 -2.6436 7.9994 5.9916 5.0076 7.7606 2.0234 3.7532 -2.6934 4.2506 7.7026 -7.9994 7.9946 7.0052
:::§g;g -7.3360 7.9904 -7.0396 -7.9904 -7.7404 -7.0630 -4.6536 -7-2972 1.3790 6.2400 -7.9060 -7.9994 -7.9994 7.9994
:6::25: ~6.5206 7.9002 7.9994 -7.0642 7.9994 7.9994 -6.4144 -6.6236 -7.9556 5.5924 -7.9994 7.9994 7.6512 '7.9994
-;.:37i -7.6944 3.5726 -3.2510 -5.4204 -5.5066 3.1232 5.0212 5.9252 7.4010 7.0020 1.3434 -7.9994 4.3602 -0.6764
:6.;::: -1.1070 4.0990 '1.9056 7.0134 7.3092 7.1116 1.2374 1.9104 5.9500 '2.3922 *1.9196 7.9994 -3.0744 3.6246
:;:2:26 -1.9524 -7.9972 -5.3796 6.3400 4.7752 6.7322 2.5992 2.4502 '0.1556 7.6540 2.7520 '7.9994 2.5994 -7.9044
:;:;z:0 -7.6100 7.9736 -7.0320 3.9760 7.1664 7.9994 7.2646 2.5400 7.9994 6.0556 6.5160 '7.9994 -7.9916 7.9994
-E.E§§g -7.1016 '7.9044 7.9724 6.0676 6.9376 -4.7162 6.6074 -3.5026 4.5716 '0.2064 5.1110 '7.9994 -7.9960 2.0432

3.6404 ‘3.0214 -3.4246 7.9994 1.5390 '3.6392 -4.6164 -4.5762 -1.1206 '1.6092 5.4094

 

 

 

Figure 5.18 — Network Weight Matrices (40995 Runs)

5.5 WEIGHTAGE OF INDIVIDUAL PARAMETERS

In order to determine the importance of each of the parameters, different models were
trained with the targeted parameter excluded during training. The corresponding
parameter that is excluded in the model with the highest RMS Error will be the most
important parameter followed by the other parameters ranked descending based on the
RMS error value. The following table lists the parameters in descending order of their
importance based on the Average and RMS errors resulting from the elimination of the
particular parameter.

Table 5.6 - Excluded Parameters and the Resulting Errors Generated by the Model

 

Excluded No. of Runs Training Average

. RMS Error
Parameter Tolerance Error

 

Size 19945 0.25 0.2481 0.2733

 

Type 19945 0.25 0.2311 0.2646

 

Length 19945 0.25 0.2093 0.2376

 

Age 19945 0.25 0.1961 0.2188

 

 

 

 

 

 

 

141

 

 

 

Depth 19945 0.25 0.1884 0.2045
Material 19945 0.25 0.1706 0.1968
Slope 19945 0.25 0.1612 0.1844

 

 

 

 

 

 

 

5.6 RESULTS OF THE ANN MODELING EFFORT

Although the data obtained was noisy and inadequate for a thorough statistical analysis, it
was demonstrated that the neural network was adept in capturing the subtle relationships
that gives an indication of the pipeline condition state. This experiment indicates that the
neural network is capable of learning the deterioration trends and relates it to the
condition ranking of sewers during training. But the fact that it could learn only 70% with
ease indicates that additional parameters identiﬁed in the literature review are needed to
account for the full deterioration pattern to predict the actual condition of sewer pipes. As
access to more detailed information identiﬁed in the literature is available, the network
can perform at a better rate and the tolerances can be set at a lower value to result in a
more accurate model.

This model is essentially designed to be able to predict the condition probability
of the sewer pipes. If the pipe attributes are known, such as pipe age, average depth of
cover, manhole to manhole length, pipe diameter, pipe material, the deﬁciency
probability can be predicted from the trained neural network model. The output of the
model ranges from 1 to 5, 1 being the best and 5 being the worst possible condition. High
priority should be placed on the pipe with high deﬁciency probability. If the resulting

condition predicted by the model is higher than a set threshold (usually determined by the

142

municipal agency), the recommended action is to perform a physical inspection to those
sewer sections to determine the condition state of that pipe.

The model developed in this thesis effort was intended to generate the deﬁciency
probability that will aid the decision-maker along with other factors like expert judgment.
location importance factors, etc. to prioritize “at risk” sewers. Physical inspections can
then be scheduled to these prioritized sewers to optimize the inspection resources and to

carry out any appropriate performance improvement measures in a proactive manner.

5.7 SUMMARY AND CONCLUSIONS

This chapter presented the detailed overview of the development of the neural network
model. Various conﬁgurations were experimented and the best architecture among them
was chosen for further description and development. It was observed that the model
exhibited a good learning tendency towards the facts presented, but there were problems
due to noise in the data because of the fewer number of available facts, outliers and the
missing parameters that account for rest 0f the deterioration process. It is concluded that
the application of neural network to solve the problem of condition prediction of sewers
is feasible and the accuracy of the model depends on acquiring a larger and more

inclusive sample size.

143

CHAPTER 6

SUMMARY, CONCLUSIONS AND RECOMMENDATIONS

6.1 SUMMARY

According to the literature review results, many researchers have developed regression
models for pipeline deterioration and condition forecasting. Although widely used and
easy to understand, these models are less accurate due to the complexity of deterioration
mechanism involved in pipeline deterioration models. On the other hand, development of
the traditional models needs a function form to be pre-speciﬁed. This could be difficult in
the presence of a huge pipeline condition data because of the multitude of variables
associated with pipeline deterioration. As an alternative, this thesis attempted to develop
a sewer pipeline condition prediction model using artiﬁcial neural networks, which does
not require a pre-speciﬁed function form. The outcome of this thesis demonstrates the
ability to develop neural network based condition prediction models for sewer pipes
using the available condition assessment data. Although complicated in the neural
network training algorithm, the development of neural network is usually implemented
by commercially available software packages. In this research, the BPNN model was
developed with the City of Atlanta’s sewer condition assessment database by using
BrainMaker, a popular neural network training platform. After the training process is
completed, the BPNN model can remember all the necessary information in its weight

matrix and can be drawn to validate new datasets.

144

There are several factors involved in determining the threshold of the pipe
condition. If the predicted pipe condition exceeds the threshold, inspection to that pipe
section must be performed to prevent worsening of the condition. If the predicted pipe
condition is lower than the threshold, low priority is placed in the inspection of that pipe
section. Budget availability, historical record and experience will play important role in
determining the threshold.

The predicted pipe condition using this model‘is essential in sewer rehabilitation
scheduling as it provides the decision-maker with a priority based inspection ranking
system. At project level, the model can help identify the maintenance needs and strategies
for sewers. A higher priority should be placed on the higher deﬁciency prediction of the
sewer system. Physical inspections can be scheduled to optimize inspection so that
necessary‘repairs can be performed in a proactive manner prior to sewer collapse or other
adverse effects and to rehabilitate the pipe at the most optimal time. At the network level,’
the decision-makers can propose the annual budget and maintenance plans by using the

above information.

6.2 LIMITATIONS OF THE RESEARCH

As indicated previously, this research is undertaken mainly to demonstrate the possibility
of using neural network models as a screening tool to prioritize inspections. The
availability of fewer numbers of deterioration parameters and limited data availability
posed the primary drawback to effective neural network training and caused the main

limitation to this thesis.
3.

145

Environmental parameters affecting the pipe, such as overburden pressure, soil
type and properties, soil pH, soil water content and other factors identiﬁed in the
literature were omitted due to lack of monitoring of the data. Review of literature showed
these parameters to be appropriate measures of corrosion and soil-pipe interaction. Hence
the largest limiting factor for modeling ease and accuracy was the unavailability of all

encompassing and comprehensive data.

6.3 CONCLUSIONS

Due to their low visibility, rehabilitation of underground sewer system is often
neglected until a catastrophic failure occurs. This, more often than not, results in costly
and difﬁcult rehabilitation due to the urgent nature of ensuring that the sewer system is
operational. A majority of sewer repair projects are executed on a “reactionary” basis,
rather than adopting a “proactive” approach. There are two main reasons for this: the ﬁrst
is the unavailability of adequate information regarding the condition of the sewer system;
and the second is the ineffectiveness of predicting sewer deﬁciency prior to failure or an
adverse condition so that inspection and repairs could be performed prior to failure of the
system that might lead to a costly ﬁx and other risks.

The main contribution of this thesis is the development of a neural network model
to assess the probable condition of sewers to prioritize inspection requirements. This
prediction model is developed to improve the objectivity of proactive management of
sewer systems.

The neural network model was developed utilizing the City of Atlanta’s condition

assessment survey database. Since all the parameters that were identiﬁed to affect the

146

 

sewer deterioration, as identiﬁed in the literature were not readily available to be
incorporated in this model, it is recommended that the model be expanded to encompass
those parameters and retrained. Through this process the neural network will keep
learning the updated information and adjust its hidden weights to ensure the forecasting
accuracy.

To accurately quantify the effect of certain input parameters for sewer
deterioration, it will be useful to develop a neural network model, as demonstrated in this
thesis as an initial starting base. However, subsequent models with more descriptive
parameters will enhance the understanding of the effects of inﬂuencing input parameters

on sewer systems.

6.4 RECOMMENDATIONS FOR FUTURE WORK

Since the developed model does not include a number of parameters thought to be
important to sewer deterioration, the model developed in this exercise is not complete.
While it demonstrates the utility of using Artiﬁcial Neural Networks for predicting sewer
condition, further work for data collection and model development is required to ensure
that the model is more accurate and reliable for future applications.

Having made the above conclusions, it is clear more work is required to facilitate
future use of the model. This thesis illustrates the need for the following actions, to
facilitate ease, and more comprehensive development of Artiﬁcial Neural Network
models for sewer condition prediction:

7 Inclusion of more descriptive data.

147

 

7 Collection technique improvements of present data. Advances in embedded
sensor technology can be used to get more information on the deterioration trends
in the sewer pipes, and information obtained from this can be used to enhance the

' prediction capabilities of the model.

7 Exploration of input parameter importance, and other factors affecting sewer
deterioration.

To further the development of neural network models that are accurate and
ﬂexible, inclusion of more descriptive data is needed. The model developed in this
research required making assumptions that were scope limiting since it required values
from some factors that may affect the sewer deterioration process. The availability of
detailed soils parameters, physical pipe characteristics, and in-situ pipe conditions would
be assets to fully understand and model the deterioration of sewers and accurately predict
their condition. A list of possible parameters that can be factored into the model is listed
below.

Table 6.1 — Recommended List of Parameters that needs to be incorporated in

 

 

 

 

 

 

 

 

Future ANN Models
Parameter Range of Variables
Surface Loads High (1)/Low (0)
Groundwater Level High (1)/Low (0)
Frost Heave Factor High (1)/Low (0)
Bedding Condition Good (1)/Poor (O)
Backﬁll Soil Type Cohesive (1 )/Non-Cohesive (0)
Soil Aggressivity High (1)/Low (0)
Soil Stability Factor High (1)/Low (0)

 

 

 

 

148

 

 

 

 

 

 

 

Parameter Range of Variables
Sewerage Characteristics Corossivity, pH, etc.
Number of Laterals Number
1 Tree Root Problem High (1)/Low (0)
Sewer Location Residential/Commercial/Industrial
Construction Quality Expert Factor (O-l)
Ground Movement High (1)/Low (O)

 

 

 

 

The neural network based sewer condition prediction model can then be
integrated with a comprehensive infrastructure asset management system to aid the
municipal agencies in better planning and spending of their limited available budget.
Figure 6.1 illustrates a ﬂow chart of the proposed integrated model to make decisions for

inspection prioritization using the built neural network model.

149

 

    

Sewer Database Integrated
with Neural Network

  
    
    
  
 

 

 

 

Pipe Attributes
- Age
- Material
- Size
- Length
- Depth of Cover
- Gradient
Sewer Type

 

 
 

 

ANN Model Predicts
Sewer Condition

 

 

Condition State

1 - “good”
2 — “fair” 1;
3 — “moderate” -‘
4 — “poor”

5 — “severe”

 

 

Select Acceptable

Threshold Condition Level 7.;

 

 

Assess Consequence of
Failure and Importance
Factor

 

    

 

Prioritize and Inspect

Figure 6.1 — Conceptual Integration of ANN Model

150

APPENDIX A

151

INTERNAL SEWER CONDITION GRADES OF DEFECTS - CITY OF
ATLANTA

     

 

   

 

CN 7 Connection 1 7 77 onstruction
_CNA Connection Abandoned _ ,1_. _ _, , __an§_trp£ti9r,l_-_ . - -
DC Dimension Change 17 7 7 7 7 Construction
JN Junction 7 7 7 7 7 1 Construction
JNA7 Junction Abandoned 1 7 Construction
LC Liner Changes 7 7 17 77 7 Construction
LD7 Line Deviates Down 7 1 Construction 7
LL Line Deviates Left 1 7 7 7 Construction
LR7 Line Deviates Right 7 7 7 1 _ 77 Construction _
LU 7 Line Deviates Up 1 77 Construction
MC Material Change 71 7 Construction
MH Manhole-_ -. , . _ - - L ,_ ._iC_0.n8,t_r.u<.=ti_on - ,,
SC Sewer Shape Changes 1 Construction
CU Camera Underwater 1 7 Miscellaneous
FH Finish Survey 1 77 Miscellaneous
GO General Observation 1 Miscellaneous
GOA General Observation Abandon 1 Miscellaneous
SA Survey Abandoned 1 Miscellaneous
ST Start Survey 1 Miscellaneous
WL Water Level 1 Miscellaneous
ABS Abandon Service 1 Service

DE Deposit 2 7 Service

DEG Deposit Grease 2 Service

DEJ Debris at Joint 2 Service

DEP Defective Plumbing in Bldg 1 Service

DES Deposit Silt 2 Service

EH Encrustation Heavy 4 Service

EHJ Encrustation Heavy at Joint 4 Service

EL Encrustation Light 2 Service

ELJ Encrustation Light at Joint 2 Service

EM Encrustation Medium 3 Service

EMJ Encrustation Medium at Joint 3 Service

ESH Scale Heavy 4 Service

ESL Scale Light 1 Service

ESM Scale Medium 3 Service

HI MH Below Grade 1 Service

ID Inﬁltration Dripper 2 Service

 

152

 

 

RTJ

EXP

7 Missing Cleanout Cover—

.Fine Roots- - -

 

Inﬁltration Dripper at Joint _
Inﬁltration Gusher

Inﬁltration Runner

Inﬁltration Runner 7at7J70int 7 77 7
Inﬁltration Seeper

Inﬁltration Seeper at Joint
MH Above Grade 7 7
Obstruction 7

Roots Fine at Joint 7
Mass Roots 7 7
Roots Medium at7J7oin7t _ 77

Tap Root

Roots Tap at Joint

Surface Spalling 7 77 7 _ 7
Surface Damage Spalling Slight
Vermin — Rats

Area Drain

Pipe Broken

MH Cover Cracked or Broken
Broken Cleanout

MH Frame Cracked or Broken
Sewer Broken at Joint

Catch Basin

Crack Circumferential 7

Crack Circumferential _at Joint
Crack Longitudinal

Crack Longitudinal at Joint
Cracks Multiple

Crack Multiple at Joint
Connection Intruding
Connection Defective
Connection Defective Intruding
Deformed

Brick Displaced

Deformation Horizontal
Dropped Invert

Deformation Vertical
Driveway Drain

Exposed Pipe

153

2.
74
3
, 3_
72
E.
-1
2 ,
.2.
,1_
,3
.3...
3
..3,.
2
1
4
4
5
3
5.
7 4
5
2
2
2
2
3
3
4
4
4
4
3
4
4
4
4
3

7 7 Service
7 7 .Service
.,S§rv.ic_s- _
, Service,
. -. Service .-
..S._e.rxi£¢_ _
Service ,
Service. ,.
Service

_, Service

Service.
. Service

l

Scryice.
7 Service

l
i

- Service.
Service
Structural
Structural
Structural
Structural
7 Structural
Structural
Structural

.. -.Stru¢tural.
‘ Structural
Structural
Structural
Structural
Structural
Structural
Structural
Structural
Structural
Structural
Structural
Structural
Structural
Structural
Structural

_ _‘ $833931- ._ _

 
   

 

 

 
 

OJM
OJ S

SGL
SGM

SMH
SSL
SSM

SW
SWD
SWL
SWM
SWS

Fracture Circumferential
Fracture Circumferential at Joint
Frame/Cover Leaks

Foundation Drain

Fracture Longitudinal

Fracture Longitudinal at Joint
Fractures Multiple

Fracture Multiple at Joint

Hole

Soil Fissures _7 7

.Hole in Storm Ditch M

Joint Displaced Large
Joint Displaced Medium 7
Joint Displaced Slight
Junction Defective
Liner Defect77 7

Brick Missing

Manhole Frame/Cover
Manhole Structure
Multiple Soil F issures
Missing Mortar Medium
Missing Mortar Surface
Missing Mortar Total
Open Joint Large

Open Joint Medium
Open Joint Slight

Roof Leader Connected
Surface Damage Corrosion Large
Surface Damage Corrosion
Medium

Storm Manhole

Surface Damage Spalling Large
Surface Damage Spalling
Medium

Surface Damage Wear

Stairwell Drain

Surface Wear Large

Surface Wear Medium

Surface Wear Slight
Windowwell Drain

Collapsed

154

'. .. ~.-
‘J; L» (a) b) b.) 03"...)7 "‘ <7
- I u ;“

_Strsctural-.-
,-Sm_19tura! ,
7 Structural
..St_mqtur_al_ . ..
Structural 7
Structural
_. SEECQPQI.

IDJAA

J
I

I
l

.—-4>.‘N 3+th

1
l

w-hA—‘v-‘NWHNUwa-bi

Lab-I:-

UINNUJANA

 

V Structura W
' Structural

Structural

,.Strucm,ra1

Structural
Structural
Strycmral -

. Structural

Structural

Structural 7
Structural
Structural
Structural
Structural
Structural
Structural
Structural
Structural

7 Structural7 _ 77

Structural
Structural
Structural

Structural
Structural
Structural

Structural
Structural
Structural
Structural
Structural
Structural
Structural

 

 

Code

. Deﬁnition _

 

 

7 Conditiolig. .
; _ Gracia-.- ?

1

 

L‘"

. .. ; Application) 3

~33”

 

 

XM

Z

I CollapsedManhol‘e7777

Multiple

5

3

Structural 7
Structural

 

155

 

APPENDIX B

156

DATA SAMPLES USED FOR MODELING

TRAINING FACTS

OUTF LAT LENGTH SIZE CO VC 31-35 36—40 41-45 46-50 51-55 56-60 >60

DEPTH SLOPE OUTPUT
F: ........ 1

5.4

l O 215 18 l 0 0 0 0 0 l 0 0 15.9 0.001

10350181000001001850013

fl ________ 3

l 0 273 15 l 0 0 0 0 0 1 0 0 9.7 0.005
3

Li -------- 4

l 0 416 30 l 0 O 0 0 0 1 0 0 11.4 0.011
3

U -------- 5

l 0 166 30 l 0 0 O 0 0 1 0 0 9.2 0

1 0 241 15 1 0 0 O 0 0 l 0 0 12.6 0.023

l; -------- 7

l 0 494 30 1 0 0 0 0 0 1 0 0 13.5 0.027
3

L1 -------- 8

l 0 489 30 l 0 0 0 0 0 l O 0 9 0.001

3

U -------- 9

1 0 324 15 l 0 0 0 0 0 1 0 0 14.3 0.007
3

Li -------- 10

l O 367 15 l 0 0 0 0 0 1 0 0 10.1 0.01

3

L1 -------- 11

l 0 404 15 1 0 0 0 0 0 l 0 0 9.3 0.004
Ll -------- 12

0 1 238 8 l 0 0 0 0 l 0 0 0 9.6 0.015

3

U -------- 13

O l 293 8 1 0 0 0 0 1 0 0 0 10.3 0.008

157

-------- 15
1 332

-------- 16
1 205

1 245

1 240
1 254

1 267

-------- 23
1 112

-------- 24
1 170

-------- 25
1 249

l 148

158

8.6 0.01

8.9 0.009

9.8 0.009

8.3 0.003

8 0.016

8.4 0.006

10 0.019

12.5 0.008

11.3 0.012

10.4 0.004

7.9 0.012

9.7 0.002

9.3 0.008

12 0

8.7 0.005

0 1 114
2

D -------- 30

0 1 122
3

L -------- 31

0 1 256
1

13 -------- 32

0 1 202
3

Li -------- 33

0 1 401
3

U -------- 34

0 1 126
3

L: -------- 35

0 1 487
3

L3 -------- 36

0 1 144
2

L1 -------- 37

0 1 290
3

u -------- 38

0 1 130
1

u -------- 39

0 1 84
2

U -------- 40

0 1 258
3

L: -------- 41

0 1 139
3

Li -------- 42

0 1 226
3

L1 -------- 43

0 1 316
3

1.1 -------- 44

159

0 0 8

0 0 9 3
0 0 9.7
0 0 9.3
0 0 8.7
0 0 14
0 0 8.9
0 0 8.3

0.001

0.003

0.018

0.001

0.002

0.031

0.006

0.004

0 0 10.9 0.015

0 0 9

0 0 8.5

0.002

0.014

O 0 11.4 0.005

0 0 11.8 0.001

0 0 9.2

0 0 9.3

0.004

0.002

0 l 177
2

Li -------- 45

0 1 183
2

L1 -------- 46

0 1 440
3

U -------- 47

0 l 375
3

U -------- 48

0 1 394
3

L1 -------- 49

0 1 158
3

D -------- 50

0 1 372
3

L1 -------- 51

0 1 397
3

U -------- 52

0 l 345
2

LI -------- 53

1 0 259
2

L1 -------- 54

l 0 232
L1 -------- 55

l 0 437
1

L1 -------- 56

1 0 142
3

U -------- 57

l 0 247
U -------- 58

l 0 212
2

U -------- 59

1 0 128

10

10

10

10

160

0

0

9.4 0.005

12.2 0.002

11 0.004

9.9 0.001

12 0.002

11.3 0.005

9.9 0.004

9.5 0.004

10 0.002

12.2 0.017

19.9 0.057

12.1 0.013

8.1 0.013

11.1 0.01

11.8 0.006

11.6 0.005

0 230

0 255

-------- 64
0 500

‘- -------- 65

1 139

‘ ........ 66

0 263
-------- 67
0 256

0 312

0 331

0 190

10

10

10

10

15

18

18

18

18

18

18

0 0 1 0 0 0 0 11.3 0.004

0 0 1 0 0 0 0 9.3 0.001

0 0 1 0 0 0 0 9.2 0.004

0 0 1 0 0 0 0 9.9 0.007

0 0 0 0 1 0 0 13.9 0.011

0 0 1 0 0 0 0 10.5 0.022

0 0 0 0 1 0 0 14 0.053

0 l 0 0 0 0 0 9.1 0.001

0 0 0 0 1 0 0 13.5 0.023

0 0 0 0 1 0 0 17.4 0.019

0 0 0 0 1 0 0 12 0.009

0 0 0 0 1 0 0 11.9 0.009

0 0 0 0 1 0 0 10.7 0.002

0 0 0 l 0 0 0 10.5 0.001

0 0 0 l 0 0 0 8.4 0.002

161

0 1 202
3

L1 -------- 76

0 1 232
3

u -------- 77

0 1 109
1

1;: -------- 78

0 1 131
2

U -------- 79

0 1 219
2

n -------- 80

0 1 384
3

U -------- 81

0 1 170
3

n -------- 82

0 1 176
3

[J -------- 83

0 1 345
3

Li -------- 84

0 1 151
3

L1 -------- 85

0 1 352
2

L1 -------- 86

0 1 214
1

LJ -------- 87

0 1 180
1

u -------- 88

0 1 500
3

u -------- 89

0 1 286
4

u -------- 90

162

9.5

8.2

12.5

6.8

7.2

6.8

13.2

10.4

7.8

8.7

6.7

5.8

10

6

0.011

0.045

0.001

0.014

0.006

0.018

0.012

0.009

0.003

0.003

0.013

0.003

0.014

0.003

1 311

-------- 95
1 291

-------- 96
1 121

-------- 97
1 267

-------- 98
1 281

1 129

-------- 103
1 321

' -------- 104

l 197

-------- 105
1 216

163

8.4

9.6

6.5

9.7

11

6.5

6.5

9

0.007

0.001

0.007

0.006

0.003

0.002

0.03

0.026

10.3 0.002

10.4 0.002

10.2

9.6

10.9

8.8

9.7

8.2

0.013

0.012

0.006

0.007

0.016

0.002

3

U -------- 106
0 1 200
3

U -------- 107
0 l 290
l

U -------- 108
0 1 428
1

U -------- 109
0 1 419
3

U -------- 110
0 l 217
3

B -------- 111
0 1 378
3

U -------- 112
0 1 278
3

U -------- 113
0 l 135
2

U -------- 114
0 l 77
2

U -------- 115
0 1 311
1

U -------- 116
0 l 200
2

U -------- 117
0 1 133
3

U -------- 118
0 1 331
3

U -------- 119
0 l 169
3

U -------- 120
0 1 282
3

15

10

15

15

164

12 0.029

11.9 0.026

18.7 0.038

7.3 0.006

10 0.008

9.2 0.013

8.6 0.002

10.9 0.06

4.9 0.037

6 0.007

5.65 0

9 0.011

10.2 0.001

12 0.008

9.8 0.004

_1N""[:

CWHI

12

165

0 8.8
0 9.5
0 9.4
0 7.6
0 6.6
0 9.8
0 10.4
0 12.1
0 3.8
0 14.2
0 9.4
0 7.4
0 7.4
0 7.5
0 19.7

0.003

0.012

0.01

0.006

0.002

0.007

0.016

0.004

0.079

0.005

0.014

0.002

0.12

1 0 139 8 0 l l 0 0 0 0 0 0 8.1 0.016
3

U -------- 137

1 0 149 8 0 l 1 0 0 0 0 O 0 6.9 0.002
3

U -------- 138

1 0 146 8 0 1 l 0 0 0 0 0 0 6.7 0.003
3

U -------- 139

1 0 298 8 0 l 1 0 0 0 0 0 0 10.7 0.007
U -------- 140

0 l 60 8 0 1 0 0 0 0 0 l 0 6.8 0.004
2

U -------- 141

0 l 65 8 0 l 0 0 0 0 0 1 0 5 0.004
2

U -------- 142

O l 109 8 0 1 0 0 0 0 0 1 0 5 0.007
2

TESTING FACTS

OUTF LAT LENGTH SIZE CO VC 31-35 36-40 41-45 46-50 51-55 56-60 >60
DEPTH SLOPE OUTPUT(DESIRED)

10362301000001001250003
3

U -------- 2
10259151000001001030
3

L -------- 3
0120881000100008.80.02
3

U‘ -------- 4
01255810001000010.70.013
3

U -------- 5
0120481000010001020002
2

U -------- 6
0113481000010008.70.002
2

166

0 1 184 8 l 0 0 0 0 l 0 0 0 11.9 0.003

U
0 1 117 8 1 0 0 0 l 0 0 0 0 9.2 0.006
2

L -------- 9
0122781000100009.50.004
2

L -------- 10
102928100010000120019
3 .

U -------- 11

0 253 8 1 0 0 0 1 0 0 0 0 8.3 0.009

1
3
U
0 l 200 8 1 0 0 0 l 0 0 0 0 9.9 0.009
3
U

01968100100000930
2

L -------- 14
0118681000010008.60.044
2

L -------- 15
011538100000001930002
3

L -------- 16
0120181000010006.30.002
1

L -------- 17
0124781000000106.10.004
2

L -------- 18
01125810000001010.10.005
3

L -------- 19
01175810000001080.011
2

L -------- 20
0134481000000011050012
3

L -------- 21
011978010010000110.008
3

167

0 1 154 8 0 1 0 0 l 0 0 0 0 9.6 0.021
0 l 235 8 0 1 0 0 1 0 O 0 0 10.6 0.001
3

'0 253 8 0 1 1 0 0 0 0 0 0 6.9 0.019

0 l 300 8 0 1 0 0 0 1 0 0 0 8.7 0.004
2

TESTING OUTPUT (@ 40995 RUNS)

OUTF LAT LENGTH SIZE CO VC 31-35 36-40 41-45 46-50 51-55 56-60 >60
DEPTH SLOPE OUTPUT(PREDICTED)

U -------- 1

1.0000 0.0000 362.00 30.002 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000
0.0000 0.0000 12.500 0.0029

2.8144

L] -------- 2

1.0000 0.0000 259.09 15.000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000
0.0000 0.0000 10.302 0.0000

3.2861

U -------- 3

0.0000 1.0000 208.06 8.0020 1.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000
0.0000 0.0000 8.8012 0.0200

2.7632

L] -------- 4

0.0000 1.0000 255.01 8.0020 1.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000
0.0000 0.0000 10.699 0.0130

0.0000 1.0000 204.09 8.0020 1.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000
0.0000 0.0000 10.200 0.0020

2.5039

U -------- 6

0.0000 1.0000 134.05 8.0020 1.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000
0.0000 0.0000 8.7030 0.0020

168

0.0000 1.0000 184.00 8.0020
0.0000 0.0000 11.902 0.0029
2.7097

U -------- 8

0.0000 1.0000 117.08 8.0020
0.0000 0.0000 9.2022 0.0060
2.1201

U -------- 9

0.0000 1.0000 227.08 8.0020
0.0000 0.0000 9.5009 0.0040
2.8489

U -------- 10

1.0000 0.0000 292.07 8.0020
0.0000 0.0000 12.000 0.0190
1.7839

U -------- 11

1.0000 0.0000 253.07 8.0020
0.0000 0.0000 8.3020 0.0090
2.3120

U -------- 12

0.0000 1.0000 200.01 8.0020
0.0000 0.0000 9.9018 0.0090
2.7588

U -------- 13

0.0000 1.0000 96.026 8.0020
0.0000 0.0000 9.3004 0.0000
1.3906

U -------- 14

0.0000 1.0000 186.04 8.0020
0.0000 0.0000 8.6008 0.0440
1.3657

U -------- 15

0.0000 1.0000 153.06 8.0020
0.0000 1.0000 9.3004 0.0020
2.9873

U -------- 16

0.0000 1.0000 201.08 8.0020
0.0000 0.0000 6.3013 0.0020
1.9158

U -------- 17

0.0000 1.0000 247.06 8.0020
1.0000 0.0000 6.1009 0.0040
3.0840

U -------- 18

0.0000 1.0000 125.03 8.0020
1.0000 0.0000 10.102 0.0050

1 .0000

1.0000

1.0000

1 .0000

1.0000

1 .0000

l .0000

1 .0000

1.0000

1.0000

1.0000

1 .0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

169

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

1 .0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

1.0000

1 .0000

l .0000

1 .0000

1 .0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

1.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

1.0000

0.0000

1.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

0.0000

 

2.4826

U -------- 19

0.0000 1.0000 175.08 8.0020 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
1.0000 0.0000 8.0033 0.0109 '

2.8503

U -------- 20

0.0000 1.0000 344.06 8.0020 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
0.0000 1.0000 10.503 0.0120

2.9880

U -------- 21

0.0000 1.0000 197.00 8.0020 0.0000 1.0000 0.0000 0.0000 1.0000 0.0000 0.0000
0.0000 0.0000 11.002 0.0080

2.8533

U -------- 22

0.0000 1.0000 154.03 8.0020 0.0000 1.0000 0.0000 0.0000 1.0000 0.0000 0.0000
0.0000 0.0000 9.6031 0.0210

2.7178

U -------- 23

0.0000 1.0000 235.03 8.0020 0.0000 1.0000 0.0000 0.0000 1.0000 0.0000 0.0000
0.0000 0.0000 10.601 0.0010

2.9038

U -------- 24

1.0000 0.0000 253.07 8.0020 0.0000 1.0000 1.0000 0.0000 0.0000 0.0000 0.0000
0.0000 0.0000 6.9027 0.0190

2.9997

U -------- 25

0.0000 1.0000 300.02 8.0020 0.0000 1.0000 0.0000 0.0000 0.0000 1.0000 0.0000
0.0000 0.0000 8.7030 0.0040

2.1377

170

 

 

BIBLIOGRAPHY

Abraham, D. M., Wirahadikusumah, R., Short, T. J. and Shahbahrami, S. (1998).
“Optimization Modeling for Sewer Network Management,” J. of Construction
' Engineering and Management, pp. 402-410, September/October 1998.

American Society of Civil Engineers (1994). “Existing Sewer Evaluation and
Rehabilitation,” ASCE, New York, NY.

Anderson, D., and Cullen, N. (1982). “Sewer Failures 1981, the Full Year.” WRc
External Report No. 73E.

Andreou, S. and Marks, D. H. (1986). "A New Methodology for Modeling Water Pipe
Breaks", Water Forum ‘86', ASCE, 2, 1726-1733.

Andreou, D., Marks D. and R. Clark (1987). “A New Methodology for Modeling Breaks
Failure Patterns in Deteriorating Water Systems: Theory.” A Journal of advanced
Water Resources, March 1987.

Andreou, S. A., Marks, D. H., and Clark, R. M. (1987b). “A New Methodology for
Modeling Break Failure Patterns in Deteriorating Water Distribution Systems:
Applications.” Advance in Water Resources, 10, 11-20.

ASCE (1982). “Gravity Sanitary Sewer Design and Construction,” ASCE.

Babbitt, H. E., Doland, J. J., Cleasby, J. L. (1962). “Water Supply Engineering,” 6m
Edition, McGraw-Hill, New York, NY.

Benson, A., (1995). “Applications of Ground Penetrating Radar in Assessing Some
Geological Hazards,” Journal of Applied Geophysics, vol. 33, no. 1-3.

Birks, A. and Green, R. (1991). “Nondestructive Testing Handbook Second Edition,”
Volume 7: Ultrasonic testing, American Society for Nondestructive Testing,
Columbus, Ohio.

Booth, G. H., Cooper, A. W., Cooper, RM. and Wakerley, D.S. (1967). “Criteria of Soil
Aggressiveness Towards Buried Metals.” British Corrosion Journal, no. 2, pp.
104-118, May, 1967.

Brémond, B., (1997). “Statistical Modeling as Help in Network Renewal Decision.”

European Commission Co-operation on Science and Technology (COST),
Committee C3 — Diagnostics of Urban Infrastructure, Paris, France.

171

 

Bungey, J. (1995). “Testing Concrete by Radar,” Concrete, November/December,
Concrete Society, London, England.

Burn, L. S., Davis, R, DeSilva, D., Marksjo, 8., Tucker S. N., Geehman, C. (2001). Proc.
Plastics Pipes XI, Mﬁnchen, Germany.

Butt, A. A., Shahin, M.Y., Carpenter, S. H., and Camahan, J. V. (1994). “application of
Markov Process to Pavement Management Systems at Network Level.” Proc., 3rd
International Conference on Managing Pavements, TranSportation Research
Board, National Reaearch Council, Washington, DC.

Gilbert, J., Campbell, G., and Rogers, K. (1995). “Pirat — A System for Quantitative
Sewer Assessment,” International No-Dig, Hamburg, Germany, September, pp.
455-462.

CERF (2001). “Evaluation of SSET: The Sewer Scanner & Evaluation Technology,”
Civil Engineering Research foundation (CERF) Technical Evaluation Report,
CERF Report #40551, March 2001.

Chae, M. J ., Iseley, T. and Abraham, D. M. (2003). “Computerized Sewer Pipe Condition
Assessment,” New Pipeline Technologies, Security, and Safety, Proc. of the
ASCE International Conference on Pipeline Engineering and Construction,
Baltimore, MD, July 13-16, 2003, pp. 477-493.

Clark, R. M., Stafford, C. L., and Goodrich, J. A. (1982). “Water Distribution Systems: A
Spatial and Cost Evaluation.” .1. Water Resources Planning and Management
Division, ASCE, 108(3), 243-256.

Clean Safe Water for the 21St Century (2000). “A Renewed National Commitment to
Water and Wastewater Infrastructure,” Water Infrastructure Network
(http://www.amsa-cleanwater.org/advocacv/winreport/winreport2000.pdt), April,

2000.

Clemena, G. and Sprinkel, M. (1987). “Use of Ground Penetrating Radar for Detecting
Voids under a Jointed Concrete Pavement,” Transportation Research Record, no.
1109.

Constantine, A. G., and Darroch, J. N. (1993). “Pipeline Reliability: Stochastic Models in

Engineering Technology and Management.” S. Osaki, D.N.P. Murthy, eds.,
World Scientific Publishing Co.

Constantine, A. G., Darroch, J. N., and Miller, R. (1996). “Predicting Underground Pipe
Failure.” Australian Water Works Association.

Cox, D. R. (1972). “Regression Models and Life Tables.” Journal of Royal Statistical
Society, 1972.

172

 

 

 

Crowder, M. J., Kimber, A. C., Smith, R. L., and Sweeting, T. J. (1998). “Statistical
Analysis of Reliability Data,” Chapman and Hall, London, 1998.

CSIRO http://www.cmit.csiro.ﬁa_1_1/_

Daniels, D. and Schmidt, D. (1995). “The Use of Ground Probing Radar Technologies for
Non-Destructive Testing of Tubes,” International Symposium of Nondestructive
Testing in Civil Engineering (NDT-CE) pp. 429-436, German Nondestructive
Testing Institute (DGZfP), Berlin, Germany.

Davies, J. P., Clarke, B. A., Whiter, J. T. and Cunningham, R. J. (2001). “Factors
Inﬂuencing the Structural Deterioration and Collapse of Rigid Sewer Pipes,”
Urban Water 3, March 2001, pp. 73-89.

Davis, P. and Burn, L. S. (2002). “Incorporating Failure Models into Asset Management
Strategies for Plastic Pipes,” CSIRO Building Construction and Engineering,
Highett, Australia, 2002.

Dawn of the Replacement Era — Reinvesting in Drinking Water Infrastructure (2001).
“An Analysis of Twenty Utilities’ Needs for Repair and Replacement of Drinking
Water Infrastructure,” American Water Works Association (AWWA),
http://www.win-water.org/win reports/infrastructurepdf, May 2001.

Deb, A. K., Hasit, Y., Grablutz, J. F. M., and Herz, R. K. (1998). “Quantifying Future
Rehabilitation and Replacement Needs of Water Mains.” AWWA Research
Foundation, Denver, CO.

Delleur, J. W. (1994). “Sewerage Failure, Diagnosis and Rehabilitation,” Urban Drainage
Rehabilitation Programs and Techniques, William A. Macaitis (ed.), 11-28.

Digest (2001). “Deteriorating Infrastructure — Congress Prepares to Act — Billions of
Dollars Are at Stake,” (http://www.acppaorg/pdf/spring2001.pdf) Volume 31,
Number 1, Spring 2001.

DIPRA (2004). Ductile Iron Pipe Research Association (www.dipra.ﬁorg).

 

Doleac, M. L., Lackey, S. L., and Bratton, G. N. (1980). “Prediction of Time-to Failure
for Buried Cast Iron Pipe.” Proceedings of the Annual Conference of the
American Water Works Association, Denver, Colorado, pp. 31—38.

Dom, R. A. (1989). “Conducting a Waterline Corrosion Evaluation,” National

Conference on Environmental Engineering, American Society of Civil Engineers,
Austin, Texas, pp. 168-175.

173

 

Eisenbeis, P., Rostum, J., and Le Gat, Y. (1999). “Statistical Models for Assessing the
Technical State of Water Networks - Some European experiences.” Proc. A WWA
Annual Conference, Chicago.

Fenner, R. A. and Sweeting, L. (1998). “A Bayesian Statistical Model of Sewer System
Performance Using Historical Event Data,” Proc. Of HydraStorm ’98, Adelaide,
' Australia, 27-30 September, 1998.

Garrett, J. H. (1992). “Neural Networks and their Applicability within Civil
Engineering,” Computing in Civil Engineering, Proc. of the Eight Conference,
Dallas, Texas, 1992, pp. 1155-1162.

Gat, Y. and Eiesenbeis, P. (2000). “Using Maintenance Records to Forecast Failures in
Water Networks.” Urban Water, 2000: p. 187-220.

Gokhale, S. and Hastak, M. (2003). “Automated Assessment Technologies for Sanitary
Sewer Evaluation,” Proc. of ASCE New Pipeline Technologies, Security, and
Safety Conference, Baltimore, MD, July 2003.

Goodman, D. (1994). “Ground Penetrating Radar Simulation in Engineering and
Archeology,” Geophysics, vol. 59, no.2.

Goulter, I. C., and Kazemi, A. (1988). “Spatial and Temporal Groupings of Water Main
Pipe Breakage in Winnipeg.” Canadian J. Civil Engineering, 15(1), 91-97.

Goulter, 1., Davidson, J. and Jacobs, P. (1990). “Predicting Watermain Breakage Rates,”
Water Resources Infrastructure: Needs, Economics, and Financing. American
Society of Civil Engineers, Fortworth, Texas, pp. 66-69.

Goulter, I. C., Davidson, J., and Jacobs, P. (1993). “Predicting Water-Main Breakage
Rate.”J. Water Resources Planning and Management, ASCE, 119(4), 419-436.

Gustafson, J. M., and Clancy, D. V. (1999). “Modeling the Occurrence of Breaks in Cast
Iron Water Mains using Methods of Survival Analysis” Proc. AWWA Annual
Conference, Chicago.

Habibian, A. (1994). “Effect of Temperature Changes on Water-Main Breaks,” Journal of
Transportation Engineering, 120 (2), pp. 312-321, March/April, 1994.

Hahn, M.A., Palmer, R.N., Merrill, MS. and Lukas, A. B. (2000). “Knowledge
Acquisition and Validation of an Expert System for Prioritizing the Inspection of
Sewers,” ASCE Conference Proceedings, 2000.

“Handbook: Sewer System Infrastructure Analysis and Rehabilitation.” (1991). 625/6-
91/030, US. Environmental Protection Agency, Cincinnati, Ohio.

174

 

 

Herz, R. K. (1996). “Ageing Process and Rehabilitation Needs of Drinking Water.”
Journal Water SRT, 1996. 45(5): p. 221-231

Herz, R. K. (1998). “Exploring Rehabilitation Needs and Strategies for Water
Distribution Networks.” Journal of Water SRT-Aqua, 1998. 47(6): p. 275-283.

Hibino, Y., Nomura, T., Ohta, S. and Yoshida, N. (1994). “Laser Scanner for Tunnel
Inspections,” International Water Power and Dam Construction, June, IPC
Electrical-Electronic Press, London, England.

Hillier, S. F. and Liberman, G. J. (1995). Introduction to Operations Research, 6th
Edition, McGraw-Hill Book Co, Inc., New York.

Iseley, T., Abraham, D. M., and Gokhale, S. (1997). “Condition Assessment of Sewer
Systems,” Trenchless Pipeline Projects, Proceedings of the Conference Sponsored
by Pipeline Division, ASCE, Massachusetts, January, pp. 43-50.

Iseley, T., Abraham, D. M., and Gokhale, S. (1997). “Intelligent Sewer Condition
Evaluation Technologies,” Proceedings of International No-Dig Conference, pp.
254-265.

Jacobs, P., and Kamey, B. (1994). “G18 Development with Application to Cast Iron
Water Main Breakage Rate.” 2nd Int. Conf on Water Pipeline Systems, BHR
Group Ltd., Edinburgh, Scotland.

Jarvis, M. G. and Hedges, M. R. (1994). “Use of Soil Maps to Predict the Incidence of
Corrosion and the Need for Iron Mains Renewal,” Water and Environmental
Management, 8(1), pp. 68-75, February, 1994.

Jones, G. M. A. (1984). “The Structural Deterioration of Sewers,” Proc. International
Conference on the Planning, Construction, Maintenance and Operation of Sewer
Systems, Reading, UK, September 1984.

Kalbﬂeisch, J. D., and Prentice, R. L. (1980). “The Statistical Analysis of Failure Time
Data,” John Wiley and Sons Inc., New York.

Kartam, N., Flood, 1. and Garrett, Jr., J. H. (1997). “Artiﬁcial Neural Networks for Civil
Engineers: Fundamentals and Applications,” American Society of Civil
Engineers, New York, NY.

Kathula, V. S. (2000). “Distress Modeling for Sanitary Sewer Management Systems,”
Masters Thesis, Louisiana Tech University, May 2000.

Kathula, V. S., McKim, R. A., and Nassar, R. (2000). “Presiction of Sanitary Sewer Pipe

Performance Using Markovian Methods,” North American No-Dig 2000, April 9-
12, 2000, Anaheim, California.

175

Kettler, A. J. and Goulter, I. C. (1985). “An Analysis of Pipe Breakage in Urban Water
Distribution Networks,” Canadian Journal of Civil Engineering, 12, pp. 286-293.

Kleiner, Y., and Rajani, B. B. (1999). “Comprehensive Review of Structural
Deterioration of Water Mains: Physical Models.” J. Infrastructure Systems,
' ASCE.

Kleiner, Y. and Rajani, B. B. (1999). “Using Limited Data to Assess Future Needs,”
Journal of American Water Works Association, v.91, no.7, July 1999, pp. 47-62.

Kleiner, Y. (2001). “Optimal Scheduling of Rehabilitation and Inspection/Condition
Assessment in Large Buried Pipes,” 4th International Conference on Water
Pipeline Systems - Managing Pipeline Assets in an Evolving Market, 2001, pp.
181-197.

Kleiner, Y. and Rajani, B. B. (2001). “Comprehensive Review of Structural Deterioration
of Water Mains: Statistical Models.” Urban Water, v. 3, no. 3, Oct. 2001, pp. 131-
150

Konard, J. M. and Nixon, J. K. (1994). “Frost Heave Characteristics of a Clayey Silt
Subjected to Small Temperature Gradients,” Cold Regions Science and
Technology, 22, pp. 299-310.

Krstulovic, N., Woods, R.D., Al Shayea, N. (1996). “Nondestructive Testing of Concrete
Structures using the Rayleigh Wave Dispersion Method,” American Concrete
Institute Materials Journal, Vol. 93, no. 1, pp. 75-86.

Kujala, K. (1993). “Evaluation of Factors Affecting Frost Susceptibility in Soils,” Frost
in Geotechnical Engineering, Anchorage, Alaska, pp. 83-87.

Kulkarni, R. B., Golabi, K., and Chuang, J. (1986). “Analytical Techniques for Selection
of Repair-or-Replace Options for Cast Iron Gas Piping Systems — Phase 1.” Gas
Research Institute, PB87-1141 12, Chicago, IL.

Lawless, J. F. (1982). “Statistical Models and Methods for Lifetime Data,” John Wiley &
Sons, Inc., New York.

Lawrence, J. (1993). “Introduction to Neural Networks.” 5th edition, California Scientiﬁc
Software, Nevada City, CA.

Lei, J. and Saegrov, S. (1998). “Statistical Approach for Describing Failures and Lifetime
of Water Mains.” Water Science and Technology, 1998. 36(6): p. 209-217.

Levin, B. R., Epstein, R. R, Ford, E. T., Harrington, W., Olson, E., and Reichard, GE.
(2002). “US. Drinking Water Challenges in the Twenty-First Century,”

176

Environmental Health Perspectives Supplements Volume 1 10
(http://ehpnetl .niehs.nih.gov/docs/2002/suppl-1/43-521evin/abstract.htrn1),
Number 1, February 2002.

Lou, Z., Lu, J. J. and Gunaratne, M. (1999). “Road Surface Crack Condition Forecasting
Using Neural Network Models.” Florida Department of Transportation, Report
WPI# 0510816, October, 1999.

Makar, J. M. (1999). “Diagnostic Techniques for Sewer Systems,” Journal of
Infrastructure Systems, v.5, no.2, June 1999, pp.69-78.

Makar, J. M., Kleiner, Y. (2000). “Maintaining Water Pipeline Integrity,” AWWA
Infrastructure Conference and Exhibition, Baltimore, Maryland, March 12-15,
2000.

Malik, O., Pumphrey, N. D. Jr., and Roberts, F. L. (1997). “Sanitary Sewer: State-of-the-
Practice,” Proceedings of the Conference on Infrastructure Condition Assessment:
Art, Science, and Practice, ASCE, New York, pp. 297-306.

Marks, H. D., et al., (1985). “Predicting Urban Water Distribution Maintenance
Strategies: A Case Study of New Haven Connecticut.” US Environmental
Protection Agency (Co-operative Agreement R810558-01-0).

Marks, H. D., Andreou, S.,'Jeffrey L., Park, C., and Zaslavski, A. (1987). “Statistical
Models for Water Main Failures” US Environmental Protection Agency (Co-
operative Agreement CR810558) M.I.T Oﬂice of Sponsored Projects No. 94211.
Boston, Mass.

McDonald, S. E., and Zhao, J. Q. (2001). “Condition Assessment and Rehabilitation of
Large Sewers,” Institute for Research in Construction, National Research Council
Canada, Ottawa, Canada.

McGaw, R. (1972). “Frost Heaving Versus Depth to Water Table,” Highway Research
Record, no. 393, pp. 45-55.

McGhee, T. J. (1991). “Water Supply and Sewerage,” 6‘h Edition, McGraw-Hill, New
York, NY.

McMuIlen, L. D., (1982). “Advanced Concepts in Soil Evaluation for Exterior Pipeline
Corrosion”, Proceeding A WWA Annual Conference, Miami, FL.

Melina, G. and Kalles, D. (2000). “Water Network Maintenance Models,” PLAN project,
Cluster Task D3 Technical Report, Computer Technology Institute, Greece,
March, 2000.

177

Morris Jr., R. E. (1967). “Principal Causes and Remedies of Water Main Breaks,” Journal
of American Water Works Association, no. 59, pp. 782-798, July, 1967.

Moser, A. P. (1990). “Buried Pipe Design.” McGraw-Hill, NY.

Najaﬁ, M. and Iseley, T. (1995). “Trenchless Pipeline Rehabilitation,” National Utility
Contractors Association, 1995, Arlington, VA.

NASSCO (1996). “Recommended Speciﬁcation for Sewer Collection System
Rehabilitation.” Florida: NASSCO publications.

Newport, R. (1981). “Factors Inﬂuencing the Occurrence of Bursts in Iron Water Mains.”
Water Supply and Management, 3, 274-278.

Nixon, J. F. (1994). “Role of Heave Pressure Dependency and Soil Creep in Stress
Analysis for Pipeline Frost Heave,” Cold Regions Engineering, Canadian Society
for Civil Engineers, Edmonton, Alberta, pp. 397-412.

O’Day, D. K., Weiss, R., Chiavari, S., and Blair, D. (1986). “Water Main Evaluation for
Rehabilitation/Replacement.” American Water Works Association Research
Foundation (90509), Denver, Colorado.

O’Reilly, M. P., Rosbrook, R. B., Cox, G. C., and McCloskey, A. (1989). “Analysis of
Defects in 180 km of Pipe Sewers in Southern Water Authority.” TRRL Research
Report 172.

Peabody, A. W. (1977). “Control of Pipeline Corrosion.” National Association of
Corrosion Engineers, Houston. Texas.

Peters, L., Daniels, J. and Young, J. (1994). “Ground Penetrating Radar as a Subsurface
Environmental Sensing Tool,” Proceedings of the IEEE, vol. 82, no. 12,
December, pp. 1802-1822.

Price, T. (1995). “Inspecting Buried Plastic Pipe using a Rotating Sonic Caliper,”
Proceedings of the Second International Conference on Advances in Underground
Pipeline Engineering, Bellevue, Washington, June 25-28, 1995, American Society
of Civil Engineers, New York.

Pocock, R. G., Lawrence, G. J. L. and Taylor, M. F. (1980). “Behavior of a Shallow
Buried Pipeline Under Static and Rolling Wheel Loads.” TRRL Laboratory
Report 954. '

http://rebaiecn.purdue.edu/JTRP_Completed Proiecthocuments/SPR_2453/Fina1Repor
t/spr_2453_Form1700.pdf

178

 

 

Rajani, B. B. and Kleiner, Y. (2001). “Comprehensive Review of Structural Deterioration
of Water Mains: Physically Based Models,” Urban Water, v. 3, No. 3, Oct. 2001,
pp. 151-164.

Rajani, B. B, and McDonald, S. (1995). “Water Mains Break Data on Different Pipe
Materials for 1992 and 1993,” Report No. A-70l9.1, National Research Council
' of Canada, Ottawa, Ontario.

Rajani, B. B., Robertson, P. K. and Morgenstern, N. R. (1995). “Simpliﬁed Design
Methods for Pipelines Subject to Transverse and Longitudinal Soil Movements,”
Canadian Geotechnical Journal, no. 32, pp. 309-323.

Rajani, B. B., Zhan, C. and Kuraoka, S. (1996). “Pipe-soil Interaction Analysis of Jointed
Water Mains,” Canadian Geotechnical Journal, no. 33, pp. 393-404.

Read, G. F ., and Vickridge, I. G. (1997). “Sewers-Rehabilitation and New Construction.”
Arnold Publication, New York.

Romanoff, M. (1964). “Exterior Corrosion of Cast-Iron Pipe.” Journal of the American
Water Works Association, 56: 1129-1143.

Rossum, J. R. (1969). “Prediction of Pitting Rates in Ferrous Metals from Soil
Parameters.” Journal of the American Water Works Association, 61: 305—3 10.

Sacluti, F. R. (1999). “Modeling Water Distribution Failures Using Artiﬁcial Neural
Networks,” Master of Science thesis, Department of Civil and Environmental
Engineering, Edmonton, Alberta, Spring 1999.

Serpente, RF. (1993). “Understanding the Models of Failure for Sewers.” Proceeding of
the International Conference on Pipelines, ASCE, New York, 86-100.

Shamir, U. and Howard, C. D. D. (1979). “An Analytical Approach to Scheduling Pipe
Replacement.” Journal of the American Water Works Association, 71(5), 248-
258.

Smith, W.H. (1968). “Soil Evaluation in Relation to Cast-Iron Pipe,” American Water
Works Association Journal, no. 60(2), pp. 221-227, February, 1968.

Stein, D., Kentgens, S. and Bommann, A. (1995). “Diagnosis and Assessment of
Damaged Sewers Concerning Their Structural Capacity,” Proceedings of
International Conference on Advances in Underground Pipeline Engineering,
June 25-28, 1995, Bellevue, WA, pp. 1-13.

Uni-Bell (1984). “PVC Pipe — Technology Serving the Water Industry,” Uni-Bell PVC
Pipe Association, Dallas, TX.

179

 

 

Walski, T. M. (1986). "Making Water System Rehabilitation Decisions," Water Forum
‘86', ASCE, 467-474, 1986.

Walski, T. M., and Pelliccia, A. (1982). “Economic Analysis of Water Main Breaks.” J.
AWWA, 74(3), 140-147.

Water Infrastructure Now (2000). “Recommendations for clean and safe water in the 21st
Century,” Water Infrastructure Network (http://www.win-
water.org/win_reponLh)ub2/winow.pdﬂ, February, 2000.

Water Research Center (1986). “Sewerage Rehabilitation Manual,” 2nd Edition. Water
Research Center/Water Authorities Association.

WEF-ASCE (1994). Existing Sewer Evaluation and Rehabilitation. ASCE Manuals and
Reports on Engineering Practice, No. 62, Alexandra, VA.

WRc (1983). Sewerage Rehabilitation Manual. 1” edition, Water Research Center, U.K.

WSA/FWR (UK Water Industry Sewers and Water Mains Committee) (1993). “Materials
Selection Manual for Sewers, Pumping Mains and Manholes”

Yen, B. C., Tsao, C. and Hinkle, RD. (1981). “Soil-Pipe Interaction of Heated
Pipelines.” Transportation Engineering Journal, January, 1981, pp. 1-14.

Young, 0. C. and O’Reilly, M. P. (1983). “A Guide to Design Loadings for Buried Rigid
Pipes.” TRRL, Department of Transportation.

180

 

".

IIIIIIIIIIIIIIIIIIIIIIIIIIIII

1!1ill/IllWilli/W11Ill/11111111111 WWW/U!

3 1293 02736 3450