TOWARDS DISCRETE - PULSE - BASED NETWORKING AND EVENT DETECTION ARCHITECTURES FOR RESOURCE - CONSTRAINED APPLICATIONS By Saptarshi Das A DISSERTATION Submitted to Michigan State University in partial fulfillment of the requirements for the degree of Electrical Engineering Doctor of Philosophy 2019 ABSTRAC T TOWARDS DISCRETE - PULSE - BASED NETWORKING AND EVENT DETECTION ARCHITECTURES FOR RESOURCE - CONSTRAINED APPLICATIONS By Saptarshi Das In this dissertation thesis, we develop a scalable and energy - efficient discrete - pulse - based networking architecture along with a Spiking - Neuron - based low - power detection framework for use in resource - constrained settings. Applications such as Structural Health Monitoring (SHM) using wir eless sensor networks powered by ambient energy harvesting are particularly suited for such a framework. The key idea in pulse - based networking is to eschew unnecessary overhead as incurred in traditional packet - based networking and encode only the essenti al information using small number of discrete pulses and their positions with respect to a synchronized time frame structure. The baseline pulse networking does not scale well with increase in network size. In order to ameliorate this, we develop a scalabl e time frame structure for use in applications with large network size while preserving the energy advantages of pulse networking. In addition, we stress the importance of judicious use of erratic energy availability in ambient energy harvesting powered sy stems. To that effect, we build energy - awareness syntaxes within the pulse networking framework for better utilization of energy resources in such systems. We also demonstrate the feasibility of pulse networking over a through - substrate ultrasonic link lay er and the advantages thereof in terms of utilizing existing infrastructure and removing the need for radio retrofits. We explore how the protocol performance varies for an airplane stabilizer monitoring application powered by ambient vibration energy harv esting in different energy availability scenarios. Beyond this, we also develop a Spiking - Neuron - based low - power event pattern detection architecture and illustrate how this can be incorporated within a pulse - networked SHM system. The Spiking Neuron based architecture is evidenced to be simpler in terms of implementation but more efficient in terms of computation and energy usage , thus enabling in - situ detection even at intermediate nodes in the network and robust low - power event pattern detection immune to pulse drifts and errors. Copyright by SAPTARSHI DAS 2019 v Dedicated to my Maa and Baba . Thanks for all your patience, love and inspiration. vi A CKNOWLEDGEMENTS I would like to express my heartfelt gratitude to all family and friends who stood by me during my years as a Ph.D. student. Every little gesture of patience and kindness from you all helped me get farther along on my journey and kept me motivated to work till I could reach my goal . A lot of thanks go to my advisor Dr. Subir Biswas for supporting and nurturing me throughout my Ph.D. years. I am grateful for his guidance in shaping my research direction as well as preparing me to be a good researcher even be yond my Ph.D. tenure. I would like to thank my committee members Dr. Mahapatra, Dr. Ren and Dr. Kulkarni as well for their patient support throughout my Ph.D. years, always being available for productive discussions, and helping me perfect my final thesis with their valuable inputs on my research work. I also want to thank the National Science Foundation (NSF) for partly funding the work done as part of this thesis through the grant CNS 1405273. Working at the NeEWS lab at MSU has been truly fulfilling during my Ph.D. years, not just because of the varied project experience but also because of the wonderful lab members that I have had the joy of working with over the years , including Debasmit, Dong, Faezeh, Yan, Feng, Rui, Henry, Brandon and Shahrukh. Even when the work was hard, it was a relief to have friends in the lab like you that I could count on and share the experience with . My friends at MSU outside work have been a constant source of support throughout my years here and I consider them as close as my own family. I believe my Ph.D. experience would not have been half as fulfilling if not for their company through my years at MSU . Special mention vii in this regard should be made of my roommates throughout the years Chetan, Aritra and Preetam , with whom I have shared many wonderful moments during my time at MSU and who ha ve always been there for me in moments happy or sad. I would also like to thank so many other members of my Spartan family from seniors in Debasmit Da , Arko Da, Shreya Di, Faezeh and Yan to peers like Oishi, Chetan, Tridip, Piku, Tias, Soumen, Saptarshi, Portia and juniors in Aritra, Preetam, Sayli, Sabya sachi , Kanchan, Tarang and Yashesh. I have learnt a lot from you all and lived a lot as well thanks to all your wonderful company as I worked toward my Ph.D. Thanks for always being there for me . I would like to thank my family back in India for being a constant source of motivation even from the other side of the world. Special mention goes to my cousins Rivu, Shaon, Jhini, Sreeshty, Mamon and Mishtu and all my aunts and uncles who have been the re for me and my parents as I spent years away from home in pursuit of my Ph.D. I would also like to thank my elder brother Rajarshi , my sister - in - law Debolina Di and my lovely little niece Raya for their continued inspiration through my Ph.D. studies. Las t, but not the least, I would like to thank my Maa and Baba (parents) for their patience and constant motivation as I worked through my Ph.D. far away from them all these years. You have lived through many hard times but smiled through it all so that I cou ld complete my studies well. I wanted to let you know that this Ph.D. was as much yours as mine and I hope I could make some of your dreams come true as well by completing this . viii TABLE OF CONTENTS LIST OF TABLES ................................ ................................ ................................ .................. xi LI ST OF FIGURES ................................ ................................ ................................ ............... xii CH APTER 1: INTRODUCTION ................................ ................................ ............................. 1 1.1. Motivation ................................ ................................ ................................ ........................ 1 1.2. Application Domain Structural Health Monitoring ................................ ......................... 3 1.3 . Scalability in Discrete - Pulse - Based Networking ................................ ............................... 4 1.4. Energy - Aware Discrete - Pulse - Based Networking ................................ ............................. 5 1.5. Ultrasonic Through - Substrate Communication ................................ ................................ .. 6 1.6. Spiking Neuron Based Event Pattern Detection ................................ ................................ . 7 1.7. Dissertation Objectives ................................ ................................ ................................ ..... 8 1.8. Scope of Dissertation Thesis ................................ ................................ ............................. 8 CH APTER 2: RELATED WORK ................................ ................................ ......................... 10 2.1. Packet - Based Networking Alternatives ................................ ................................ ........... 10 2.2. Energy Aware Networking ................................ ................................ .............................. 11 2.3. Structural Health Monitoring ................................ ................................ .......................... 13 2.4. Through - Substrate Ultrasonic Communication ................................ ................................ 14 2.5. Spiking Neuron Based Event Pattern Detection ................................ ............................... 15 2.6. Summary ................................ ................................ ................................ ........................ 17 CH APTER 3: SYNCHRONOUS PULSE NETWORKING ................................ .................. 18 3.1. Pulse Abstraction ................................ ................................ ................................ ............ 18 3.2. Network Model Cellular Abstraction ................................ ................................ ............ 19 3.3. Cellular Pulse Protocol Frame Structure ................................ ................................ .......... 20 3.3.1. Pulse as a Protocol Data Unit ................................ ................................ .................... 20 3.3.2. Joint MAC - Routing Frames ................................ ................................ ...................... 21 3.3.3. Protocol Features ................................ ................................ ................................ ...... 23 3.4. Need for Scalability and Energy Awareness Improvements ................................ ............. 24 3.5. Summary ................................ ................................ ................................ ........................ 25 CH APTER 4: DEVELOPMENT OF A SCALABLE PULSE FRAME STRUCTURE ....... 26 4.1. Scalable Cellular Pulse Networking (SCPN) Frame Structure ................................ ......... 26 4.2. Simulation Setting and Performance Results ................................ ................................ ... 30 4.2.1. Network, simulation and event model ................................ ................................ ....... 30 4.2.2. Performance Results ................................ ................................ ................................ . 33 4.3. Summary ................................ ................................ ................................ ........................ 38 CH APTER 5: ENERGY - AWARE PULSE NETWORKING ................................ ............... 39 5.1. Energy - aware Pulse Switching Protocol ................................ ................................ .......... 39 5.1.2. Frame Adaptation ................................ ................................ ................................ ..... 40 ix 5.1.3. Binary Event Buffer ................................ ................................ ................................ . 40 5.1.4. Energy - aware Forwarding Syntaxes ................................ ................................ ......... 41 5.2. Performance Results ................................ ................................ ................................ ....... 43 5.2.1 Results with Synchronous Energy Generation ................................ ........................... 44 5.2.2. Results with Synchronous Energy Generation ................................ .......................... 52 5.3. Summary ................................ ................................ ................................ ........................ 66 CH APTER 6: THROUGH - SUBSTRATE ULTRASONIC PULSE NETWORKING ......... 67 6.1. System Architecture ................................ ................................ ................................ ........ 67 6.2. Prototype Ultrasonic Transceiver and Link Characterization ................................ ........... 69 6.3. Application and Network Model ................................ ................................ ..................... 70 6.3.1. Application Model ................................ ................................ ................................ .... 70 6.3.2. Network Model ................................ ................................ ................................ ........ 71 6.3.3. Performance Needs ................................ ................................ ................................ ... 71 6.4. Structural Vibration Model ................................ ................................ ............................. 72 6.5. Energy Harvesting Model ................................ ................................ ............................... 75 6.6. Integrated Evaluation Architecture ................................ ................................ .................. 77 6.7. Performance Results ................................ ................................ ................................ ....... 79 6.7.1. Network, Energy and Event Generation Model ................................ ......................... 79 6.7.2. Network Node Energy Traces ................................ ................................ ................... 80 6.7.3. Event Reporting Performance ................................ ................................ ................... 81 6.7.4. Impacts of Adaptive Route Diversity ................................ ................................ ........ 88 6.7.5. Impacts of Error ................................ ................................ ................................ ....... 90 6.8. Summary ................................ ................................ ................................ ........................ 92 CH APTER 7: DISTRIBUTED COGNITION USING NETWORKED PULSES AND SPIKING NEURONS ................................ ................................ ................................ ............. 93 7.1. Introduction ................................ ................................ ................................ .................... 94 7.2. System Architecture ................................ ................................ ................................ ........ 98 7.2.1. Network Model ................................ ................................ ................................ ...... 100 7.3. Spiking Neuron Based Learning ................................ ................................ .................... 102 7.3.1. Key Concepts ................................ ................................ ................................ ......... 102 7.3.2. Neuron Description and Tempotron Learning Rule ................................ ................. 102 7.4. Baseline Pulse - based Networking Approaches ................................ .............................. 106 7.4.1. Pulse Position Coding Protocol ................................ ................................ ............... 106 7.4.2. Pulse Time Encoded Networking ................................ ................................ ............ 109 7.5. Adaptations of Spiking Neuron Learning for Pulse Networki ng ................................ .... 112 7.5.1. Networking Adaptations ................................ ................................ ......................... 112 7.5.2. Membrane Time Constant Selection ................................ ................................ ....... 112 7.5.3. Training Methodology ................................ ................................ ............................ 113 7.5.4. Test Methodology ................................ ................................ ................................ .. 116 7.6. Simulation Process and Performance Results ................................ ................................ 116 7.6.1. Simulation Environment and Process ................................ ................................ ...... 116 7.6.2. Effect of varying the event interval range (beta) ................................ ..................... 119 x 7.6.3. Effect of varying positive pattern length ................................ ................................ . 121 7.6.4. Effect of Pattern Type ................................ ................................ ............................ 124 7.6.5. Effect of Spike Jitter ................................ ................................ ............................... 125 7.6.6. Effects of Learning Rate Selection for Training ................................ ...................... 127 7.6.7. Effects of Membrane Time Constant Sel ection for Training ................................ ... 128 7.6.8. Training Error Analysis ................................ ................................ .......................... 129 7.6.9. Synaptic Weights Evolution Analysis ................................ ................................ ..... 130 7.6.10. Effects of Pulse Loss and False Positive Errors ................................ ..................... 130 7.7. Summary ................................ ................................ ................................ ...................... 133 CH APTER 8: S UMMARY AND FUTURE WORK ................................ ........................... 135 8.1. Summary ................................ ................................ ................................ ...................... 135 8.2. Application Architecture ................................ ................................ ............................... 136 8.3. Extending Single - Layer Spiking Neuron - based Event Pattern Detection ....................... 138 8.4. Development of Energy - Harvesting Awareness in Pulse Networking ............................ 139 BI BLIOGRAPHY ................................ ................................ ................................ ................. 141 xi L IST OF TABLES Table 5.1. Asynchronous Harvesting Model Parameters ................................ ............................ 53 Table 6.1. PLR/FPPR for ultrasonic communication over Al 2024 alloy plate using prototype TUPN module ................................ ................................ ................................ ........................... 70 Table 7.1 . Experimental Parameters ................................ ................................ ........................ 117 xii L IST OF FIGURES Figure 1.1 . Thesis Outline ................................ ................................ ................................ ........... 9 Figure 3.1. Cellular Network Model on a Rectangular Plate Structure ................................ ....... 19 Figure 3.2. MAC - Routing Frame for Pulse Switching [85] ................................ ........................ 21 Figure 3.3. Pulse Forwarding in the Localization Area [8] ................................ ......................... 22 Figure 4.1. Joint MAC - Routing frame for Scalable Cellular Pulse Networking [85] .................. 27 Figure 4.2. Demonstration of multi - hop event forwarding in SCPN ................................ ........... 28 Figure 4.3. Simulated Network Topolo gies ................................ ................................ ............... 31 Figure 4.4. Spatio - temporal variation in harvested energy ................................ ......................... 31 Figure 4.5. Network Cell Count vs Average Per Hop Delivery Delay / Event Throughput ......... 33 Figure 4.6. Heat maps for source - sink event delivery latency ................................ .................... 34 Figure 4.7. Effect of Energy Constraint on Per Hop Delivery Latency ................................ ....... 35 Figure 4.8. Effect of Information Content Size on Delivery Latency of Pulse Protocols ............. 37 Figure 5.1. Summary of Energy - Aware Pulse Switching ................................ ........................... 41 Figure 5.2. Routing Decision based on Energy Levels ................................ ............................... 43 Figure 5.3. Capacitor charging for different duty cycles ................................ ............................ 44 Figure 5.4. Energy trace at a node on even t propagation route ................................ ................... 46 Figure 5.5. Event Reporting Delay for Different Duty Cycles ................................ .................... 48 Figure 5.6. Event Delivery Ratio at Different Duty Cycles ................................ ........................ 49 Figure 5.7. Route diversity for different energy availability ................................ ....................... 50 Figure 5.8. Event buffering/storage delay (in frames) distributions ................................ ............ 51 Figure 5.9. Rectangular plate designs with various anchor configurations ................................ .. 53 Figure 5.10. Spatial variation of node charging profiles at a fixed duty cycle ............................. 55 xiii Figure 5.11. Temporal variation in harvested energy availability using a duty cycled approach . 56 Figure 5.12. Cellular network model on a r ectangular plate structure ................................ ......... 57 Figure 5.13. Energy traces for selected nodes on the rectangular plate ................................ ....... 59 Figure 5.14. Delivery Delay characteristics for selected source nodes on the rectangular plate .. 60 Figure 5.15. Average Harvested Energy and Event Reporting Delay distributions across rectangular plate with different edge anchor configurations ................................ ....................... 61 Figure 5.16. Distribution of Average Harvested Energy and Event Reporting Delay across rectangular plate with central anchor configuration ................................ ................................ ... 63 Figure 5.17. Per hop transmission route diversity distributions for selected source nodes .......... 64 Figure 5.1 8. Buffer storage time distributions vs harvesting efficiency (Duty Cycle) ................. 65 Figure 6.1. Event Monitoring using a Through - Substrate Sensor Network ................................ . 68 Figure 6.2. Network Model on Airplane Stabilizer Structure ................................ ..................... 71 Figure 6.3. (a) 3D model of stabilizer, (b) Accln. of stabilizer based on Fi nite Element analysis 73 Figure 6.4. Total pressure profile and simplified triangular pressure profile vs. chord length (aircraft speed 800 km/h) ................................ ................................ ................................ .......... 74 Figure 6.5. Average Acceleration based on node coordinates (at aircraft speed of 800 Km/h) .... 75 Figure 6.6. Piezoelectric Harvester Circuit Model ................................ ................................ ..... 76 Figure 6.7. Architecture of integrated evaluation software ................................ ......................... 78 Figure 6.8. Pulse network mapping on a target aircraft stabilizer structure ................................ . 79 Figure 6.9. Acceleration and harvested energy profile at chosen TUPNs ................................ ... 80 Figure 6.10. Event reporting delay with different vibration intensities ................................ ....... 83 Figure 6.11. Event reporting delay with different energy storage capacity ................................ . 84 Figure 6.12. Reporting delay for different electromechanical coupling ................................ ...... 85 Figure 6.13. Spatial distribut ion of harvested energy ................................ ................................ . 86 Figure 6.14. Spatial distribution of event reporting delay ................................ ........................... 86 xiv Figure = 1.96*10 - 4 ) .......... 87 Figure 6.16. Transmission route diversity with varying coupling constant ................................ . 88 Figure 6.17. Route diversity with varying storage capacitance ................................ ................... 89 Figure 6.18. Route diversity for different average acceleration levels ................................ ........ 89 Figure 6.19. Impacts of pulse loss on event reporting delay ................................ ....................... 91 Figure 7.1. Application Overview ................................ ................................ ............................. 96 Figure 7.2. Spiking Neuron Learning in Pulse Communication Networks ................................ .. 99 Figure 7.3. Network Topology ................................ ................................ ................................ 101 Figure 7.4. Synaptic Voltage Evolution across Training Epochs Effect of Positive and Negative Training Patterns ................................ ................................ ................................ ..................... 104 Figure 7.5. PPCP PDU Spike Representation ................................ ................................ .......... 107 Figure 7.6. Event Spikes vs PPCP Spikes ................................ ................................ ................ 108 Figure 7.7. Pulse Time Encoding Frame Structure ................................ ................................ 110 Figure 7.8. Event Spikes vs Pulse Time Encoding Spikes ................................ ........................ 111 Figure ................................ ................................ ................................ ................................ ............... 120 Figure 7.10. Effect of event interval communication ................................ ................................ ................................ ........................ 120 Figure 7.11. Effect of positive pattern length (pl) on detection accuracy across different negative training set sizes when using PPCP communication ................................ ................................ 121 Figure 7.12. Effect of positive pattern length on detection accuracy when using PPCP communication ................................ ................................ ................................ ........................ 122 Figure 7.13. Effect of positive pattern length on detection accuracy when using TDMA communication ................................ ................................ ................................ ........................ 122 Figure 7.14. Similarity of unknown pattern to positive pattern trajectory and its effect on detection accuracy for different positive pattern trajectory lengths - PPCP ................................ ............. 123 Figure 7.15. Detection Accuracy for Different Positive Pattern Types - PPCP ......................... 124 xv Figure 7.16. Detection Accuracy for Different Positive Pattern Types TDMA ...................... 124 Figure 7.17. Detection Accuracy for Different Pattern Jitter Levels - PPCP ............................. 126 Figure 7.18. Detection Accuracy for Different Pattern Jitter Levels TDMA .......................... 126 Figure 7.19. Detection Accuracy for Different Learning Rates ................................ ................ 127 Figure 7.20. Detection Accuracy for Different Membrane Time Constants .............................. 128 Figure 7.21. Positive and Negative Pattern Training Error Evolution across Training Mini - Batches ................................ ................................ ................................ ................................ ............... 129 Figure 7.22. Synaptic Weights Evolution across Training Epochs for different sizes of negative train set ................................ ................................ ................................ ................................ ... 130 Figure 7.23. Performance in t he presence of Single Pulse Loss Errors - Positive and Unknown Pattern Detection Accuracy across different number of negative trajectories used for training . 132 Figure 7.24. Performance in the presence of False Positive Pulse Errors - Positive and Unknown Pattern Detection Accuracy across different number of negative trajectories used for training . 133 Figure 8.1. Structural Health Monitoring Platform based on Pulse Communication and Spiking Neuron Based Detection ................................ ................................ ................................ .......... 1 37 1 CHAPTER 1: INTRODUCTION 1.1. Motivation Applications in diverse domains such as environment monitoring [1] [3] , habitat monitoring [4] [6] , structural monitoring [7] [10] , target tracking [11] [14] , and industrial process control [15] [17] rely on the collection of granular information across areas of concern using distributed sensing devices. The data from such devic es can be collected at a central location for processing and drawing higher level inferences in the context of the application involved. Energy - efficient w ireless sensor networks [18] [21] are well suited for such applications because they allow low - cost deployment and data communication across the sensor nodes as well as to a central Base Station. Packet - based networking [22] [24] is the dominant mo de of networking for such applications. Various routing strategies and energy - efficiency improvements [25] [34] have been proposed in the literature to suit such packet - based wireless sensor networking in low - power environments. The latter is the case when the sensing devices are equipped with small batteries or are powered by erratic energy harvesting sources such as wind or ambient vibrations. However, depending on the application involved and data size as well as latency requirements, packet - based networking is not always the most energy - efficient approach. It is to be noted in this regard that most of the energy usage in a typical sensing device for a wireless sensor network is expended in the wireless (radio, ultrasound etc.) transmission/reception/idling costs and less so in actual sensing. Therefore, reducing the former is very important for resource - constraine d applications. Also, packet - based networking generally involves a considerable amount of overhead information beyond the actual data payload being transported, which is constituted of components such as the 2 packet preamble and header. The preamble uses a chosen bit pattern to synchronize transmission timing between senders and receivers in a wireless networking context. This is especially important when long data sequences need to be transported and in an environment with many other wireless devices using the same transmission channel. The packet header contains other information needed to ensure reliable transmission (error detection / correction), addressability, quality of service etc. Thus, the preamble and header are needed to ensure data reliability, transmission synchronization, low latency etc., especially when the Protocol Data Unit (PDU) being transported is long. However, in application instances such as Structural Health Monitoring, the essential data is often just binary event information from d istributed sensors and thus very small. In such scenarios, Packet - based networking approaches can often be overkill as overhead costs easily overwhelm actual data networking costs. It has been shown in prior work [35] that discrete - pulse - based networking approaches c an be much more energy - efficient compared to packet - based networking approaches when the information content is small, latency requirements are relaxed, and network event rate is not exceedingly high. This can be achieved by using a on in time within a synchronized time frame structure to encode next - hop information and origin location. In this thesis, we will focus on applications which have the above - mentioned properties (i.e. low information content, less - stringent latency bounds) and explore several improvements that need to be incorporated to adapt the baseline discrete - pulse - based networking to various relevant application scenarios. The latter includes systems powered by scarce energy harvesting sources as well as large distrib uted networks where the baseline pulse networking might not easily scale. Specifically, we will develop scalable and energy - aware discrete - pulse - based networking architectures that can be applied in Structural Health Monitoring (SHM) applications. We will also 3 explore through - substrate networking using ultrasonic communication as a feasible and energy - efficient mode of communication for distributed sensors in an SHM context. It will be shown that discrete - pulse - based networking can be well - adapted to such a communication scenario in a variety of energy availability situations. In addition, we will also consider how to process the communicated information from discrete - pulse - based networking using a low - energy detection architecture to create a holistic energ y - efficient communication and detection platform for resource - constrained wireless sensor networks. For the detection purposes, we will demonstrate the use of a Spiking Neuron based architecture for its simplicity in implementation and inherent energy effi ciency, yet robust generalizability in detecting spatiotemporal spike patterns. We outline the various research challenges involved in developing such an architecture and how these are being addressed in ongoing work. We also include performance results sh owing the efficacy of the Spiking Neuron based implementation in realistic application scenarios. 1.2. Application Domain Structural Health Monitoring The core objectives in structural health monitoring (SHM) [36] [38] are to collate information such as unusual stress, faults and cracks from many strategically positioned sensors over a target structure and infer the health of the structure using such information [39] . Ene rgy - efficient wireless sensor networks [7], [8], [40], [41] are often deployed for multi - hop data collection from such sensors to an access - point or a sink, where the collective data processing functions can be placed. A notable observation is that after a certain amount of local processing at a sensor, often the transportable information from the sensor to an access - point is merely an event. This can be a threshold crossing of local stress, detection of a pre - defined temporal stress pattern, or even a crack in the extreme case. Since the event information is binary (yes or no), a single pulse can be used 4 to communica te this, thus eliminating the need for packets and their associated overheads such as synchronization preambles. This would lead to significant benefits in energy - starved environments. The low energy usage can also improve reliability of operation in syste ms powered by energy harvesting, by keeping the consumption rate lower than the energy generation rate. The primary design questions here are how to: 1) transport event localization information using a single pulse, 2) route a pulse multi - hop without expli cit node addressing, and 3) provide reliable event delivery even under energy constraints. These problems can be architecturally addressed in a discrete - pulse - based networking framework by integrating a pulse's (i.e., event's) area of origin within a MAC - r outing protocol framework and incorporating several energy - awareness syntaxes within the same framework as will be discussed in later sections in this thesis. Thus, Structural Health Monitoring can be a good application instance for showcasing the advant ages of discrete - pulse - based networking and discussing the changes needed for such an architecture to adapt it to specific application needs such as extreme resource constraint or scalability issues. A low - energy detection architecture would also be very u seful for interpreting the communicated event information in the form of discrete pulses, especially if it can easily be interfaced with the networked pulses. Spiking Neuron based architectures can be useful in this regard because of their inherent low - com plexity design yet robust generalizability and will be developed as an event pattern detection solution in this thesis. 1.3 . Scalability in Discrete - Pulse - Based Networking The central foundation of the discrete - pulse - based architecture is a synchronized Frame Structure employed to control pulse transmission schedules such that the time of transmission can encode various information aspects that are useful to the application context. The size of such a frame is known to be related to the number of network cells [35], [42] . More precisely, the frame 5 size scales quadratically with increase in the number of network cells. It is also known that the frame size restricts the event delivery latency and event throughput of the network. Since, the core premise of discrete - pulse - based networking is to trade off latency (i.e. event delivery delay) and throughput for energy benefits, this works well for small to medium - sized networks if the limits for allowable delay and throughput are relaxed. In applications where such delay / throughput limits are more restrictive, but network cell count scaling needs to be supported in addition to energy efficiency, the frame structure does not adapt well. Network cell count might need to be incr eased to enhance maximum sensing localization resolution for the network or when the network area is increased. Hence, there is need for some redesign to adapt the baseline discrete - pulse - based networking approach for a more generic and scalable scenario a s will be developed in this thesis. 1.4. Energy - Aware Discrete - Pulse - Based Networking Various modern approaches employ sensing devices powered by energy harvested from the ambience e.g. solar power, vibration harvesting etc. Though harvesting sources hav e theoretically infinite potential, they can be exceedingly erratic in the short term. This can be shown to affect networking performance significantly in a baseline discrete - pulse - based networking approach. This is because, if networking is unaware of ene rgy availability, transmission can often be wasteful when economy is warranted, leading to intermittent network failure due to power loss. Often, some amount of energy - awareness incorporated in the networking protocol can significantly improve network upti me while not sacrificing network throughput. We try to demonstrate this with our own improvements to the baseline discrete - pulse - based networking architecture. We also evaluate how such performance is affected for various energy harvesting 6 situations and s how the utility of our approach across a wide variety of such scenarios in later sections of this thesis. 1.5. Ultrasonic Through - Substrate Communication Many structures that need to be monitored e.g. airplane wings / stabilizers, bridge beams etc. are constituted of substrate materials such as metal or composites, which themselves can be used as communication media for signals such as ultrasound. This has motivated the idea of developing a through - substrate sensor network. Tiny sensors, when embedded or affixed on a bridge or an aircraft wing, can communicate with each other using ultrasonic pulses propagating through the structure's solid substrate. This c an eliminate the need for out - of - substrate radio or wired links. A prototype modem as described in [8] , which deals with the various challenges in such a design, had been developed in our laboratory to demonstrate the feasibility of this approach and we discuss v arious physical layer implications of such an approach in the current thesis. Moreover, opportunities exist in harvesting energy from ambient vibrations in several structures such as airplane wings, stabilizers, and bridge beams. Such harvesting can provi de energy for sensing and communication of collected sensor information. Self - powered sensors have already been demonstrated [43], [44] which can use the energy from the signal being sensed to power the sensing, computation and non - volatile storage operation. Work is under way on a collaborative project to design a piezoelectric - based transducer that can use vibrations insid e a structure to power communications in addition to sensing and buffering needs. The notable fact here is that the same transducer (i.e., a substrate - embedded piezoelectric module) can be used for sensing (ultrasonic fault signatures), communication (ultr asonic link), as well as energy harvesting (ultrasonic vibration harvesting) to power all operations. Such convergence of functionality in a through - substrate approach leads to a cleaner design by removing the need for separate retro - fitted 7 components for sensing, communication and, energy generation. We will demonstrate the feasibility of such an approach through simulation results in an integrated harvesting and communication framework in the current thesis. 1.6. Spiking Neuron Based Event Pattern Detecti on Structural Health Monitoring applications and others of the same family often require identification / classification of spatiotemporal event occurrence patterns [14], [15], [39], [45] [49] using distributed sensor measurements in order to make higher level inferences based on the same. A key requirement in such applications is that the sensing architecture be flexible enough to be able to identify a variety of event occurrence patterns i.e. if new patterns need to be detected, the architecture should be able to adapt to the new scenarios. Another aspect is the need for generalizability over a range of similar inputs i.e. detection robustness to minor changes in the same event pattern. Such a pplications also assume some amount of energy - efficient / energy - aware operation because many modern distributed sensing architectures for such aim to create cheap and maintenance - free operation by relying on small sensing devices with limited energy stora ge but theoretically infinite energy generation capacity (harvesting from environmental sources) albeit at low / erratic rates (harvesting source unpredictability) as discussed in the previous sections. Spiking Neuron based approaches can be shown to have these properties and work with much less energy compared to comparable approaches and are well amenable to discrete - pulse - based networking approaches. In this thesis, we propose such an architecture melding discrete - pulse - based networking and spiking - neuro n - based event pattern detection and discuss the research challenges associated with designing such a system. We also evaluate the developed architecture in the context of both synchronous and asynchronous discrete - pulse - based networking protocols 8 to show h ow the system can be a feasible and energy - efficient means of spatiotemporal detection especially combined with pulse networking. 1.7. Dissertation Objectives The core objectives of this dissertation would be to design the following 1) A scalable discr ete - pulse - based communication protocol 2) An energy - aware discrete pulse - based networking protocol, and analyze how performance varies in different energy harvesting / availability scenarios 3) An ultrasonic through - substrate computing architecture, based on discrete - pulse - based communication and powered by vibration energy harvesting, for structural health monitoring applications, for example in airplane wing / stabilizer structures 4) A spiking - neuron - based architecture for energy - efficient spatiotemporal event pattern detection which can easily interface with a discrete - pulse - based networking architecture 1.8. Scope of Dissertation Thesis In the current thesis, we cover all the objectives of the dissertation as listed in the last sub - section, including published results on the same. In a prior thesis proposal, we had presented then - completed work on the first 3 Dissertation Objectives and laid the foundations for the work on Spiking Neuron based pattern detection architecture (Dissertation Objective 4) a nd how future work would be used to evaluate this architecture. 9 Figure 1. 1 . Thesis Outline In the present thesis, we have extended the work from the proposal and completed the Spiking Neuron based detection architecture development and evaluation. In all, we have tried to develop a holistic solution for energy - efficient detection of structural a nomalies using a through - substrate network of inexpensive sensors powered by ambient vibration harvesting. In Figure 1 .1 , we provide an outline of the main thesis objectives and clearly indicate what has been achieved grouping the contributions under vario us pertinent application areas. In the following chapters, we will start with a survey of the existing state - of - the - art in these areas and then follow up details on our specific contributions. 10 CHAPTER 2: RELATED WORK 2.1. Packet - Based Networking Alternatives There are few reported approaches in the literature that address the energy and capacity overheads of packet - based network communication. This is mainly because networking for application niches such as Structural Health Monitorin g, that require low volume information transport with relaxed latency requirements have not been explored very far. Research on networked systems with extreme energy constraints such as those powered by small piezo - electric energy harvesting devices is not very mature either. We aim to bridge the gap in research in this domain through the present thesis work. Strategies like aggregation of short payloads [50] and binary sensing models [51] have been proposed for reducing the networking energy burden, but the inherent packet overhead limitations remain. In [52] , it is shown that in single - hop networks, the wors t - case performance of pulse - based communication is better than packet - based ones, albeit with a worse delay performance. Inspired by this result, the authors in [35], [53], [54] developed the concept of multi - hop pulse switching as an alternative to energy - inefficient packet - based communication. This concept is very well - applicable when the data to be transmitted is binary in nature and the tolerable delay is high. However, the m ulti - hop MAC - Routing protocol framework for event routing, as mentioned in [35], [42] , has some limitations in the context of applications using energy - harvesting or extreme energy constraint. The latter is generally envis ioned for various maintenance - free Structural Health Monitoring applications powered by ambient energy harvesting. The first shortcoming of the baseline pulse switching is that it implicitly assumes constant energy availability, which is not valid when the available energy is variable depending on the harvesting 11 conditions. This assumption can be detrimental in terms of energy management and can lead to unwanted power losses when the energy availability changes drastically. This is because the network energ y usage is not matched with changing energy availability. Such erratic changes in energy generation are very probable particularly in harvesting - powered systems. In the current thesis, we will discuss energy - aware pulse routing syntaxes that have been deve loped to address this limitation. The second limitation is that the synchronized time frame structure proposed for the protocol in [35], [42] does not scale well for larger networks. We develop a new time - frame structure t o ameliorate this. Yet another factor is that the architecture in [35], [42] , is designed using Ultra Wide Band Impulse Radio as the physical layer. In the present thesis, we will show how this architecture can be adapted specifically for an ultrasound - based through - substrate physical layer which presents various challenges of its own. 2.2. Energy Aware Networking Various other approaches have been proposed in the literature [55] [57] with regards to energy conservation, including data - driven approaches such as reducing data through in - network aggregation and compression [50] and interest driven data acquisition [58] . Another approach is to keep local energy costs low at the sensing nodes by adopting efficient sleep schedules and topology control (only a limited number of nodes are active, that is not in sleep), based on data and energy conservation requi rements [55] . There have also been many approaches on energy - aware MAC and routing protocols in packet - based sensor networks operating under various f orms of energy constraints including from harvested energy. A general direction is to adapt the sensing duty cycle based on a harvesting model [59], [60] or by tracking the battery energy level and its depletion pattern [57] . Other approaches [58] include using an energy - efficient networking approach such as directed diffusion while building energy awareness features on top to prevent 12 energy leakage when the network is energy - constrained. Energy - harvesting aware protocols like [61] have also been discussed in the lit erature, which consider the harvesting rate and current consumption characteristics combined into a node's current energy state and utilize that to design the optimal routing approach such that the system can survive on the harvested energy. While these ap proaches provide generalized energy - saving concepts, they are designed for traditional packet - based operation and thus cannot directly be applied to the pulse - based approach considered here. In the current thesis, we discuss the development of energy - awa re mechanisms using similar principles to the literature, but specifically for pulse - switching protocols. Like the literature, these would attempt to achieve energy neutrality, but for a pulse switching network in a vibration based harvesting environment, and thus deal with the unique challenges presented there. In the thesis, we also present a fresh look at the pulse networking architecture presented in [35], [42], [54] toward solving the inherent disadvantage of very high latency when the network cell count is scaled. A scalable pulse architecture, as will be discussed in later sections of this thesis, also provides better applicability in terms of larger information content routing, lower latency, fair access latency to the network sink from across the network without increasing the energy cost. Deployment strategies and their effect on energy - efficiency and reliable data acquisition in Wireless Sensor Networks (WSNs) has also been covered extensively in the literature [62] [64] . For example, non - uniform deployment of sensors has been proposed to deal with disproportionate energy consumption at the sensor nodes close to the sink thus increasing lifetime of network operation. Also, strategies like separati on of relaying (forwarding) and sensing functionalities has been suggested to increase energy efficiency. It is to be noted that such strategies can be used in the Pulse Networking architecture as well because they mainly deal with network organization and set - up which is independent of the actual networking scheme in this case. Use of mobile nodes for 13 data acquisition from WSN sensor nodes has also been put forward in the literature [65] . Since, the autonomous energy - rich sinks can travel to the source nodes and collect information, there is no need for multi - hop transmission and the sensing node ene rgy is conserved (due to absence of forwarding function). Authors in [66], [67] discuss various challenges in a mobile sink scenario which includes various strategies for synchronizing sink acquisition schedules with node sensing schedules. It is to be noted that though mobile sink - based approaches have been shown to be advantageous in terms of energy conservation of the sensing nodes, they involve network maintenance in terms of using a source - coordinated sink mobility schedule. In the SHM monitoring applications, like the one we have mentioned before, the intent is to keep the data acquisition maintenance - free and the multi - hop approach works well in such scenarios due to absence of any extraneous control (for sinks etc.). 2.3. S tructural Health Monitoring Structural Health Monitoring has been an area of growing interest in recent years. This is because of the vast array of infrastructure in need of maintenance and the fact that maintenance and repairs represent a staggering fra ction of infrastructure costs. For instance, maintenance and repairs represent a quarter of airplane operating costs and the U.S. spends more than $200 billion dollars on maintenance of plant, equipment and facilities. The motivation in SHM is in reducing such costs by replacing scheduled maintenance with as - needed maintenance [68] [70] . Other goals include anticipating structural lifetime and rate of degradation. Wireless sensor network - based approaches [7], [40], [71] provide a seamless way to deploy such monitoring solutions. Energy - harvesting powered wireless networks can be even more useful because the network of sensors can function almost maintenance - free just based on ambient energy harvesting [72] [74] . Various works have been proposed in the literature in this regard [59], [60], [72], [75] [78] . However, 14 almost all these approaches rely on traditional pulse - based networking approaches which have been shown to be energy - expensive in such low information content scenarios. The differentiation proposed in this thesis is mainly in terms of adapting the pulse n etworking architecture to this domain. We will show that by incorporating pulse networking along with energy - awareness syntaxes as well as a scalable framing architecture, we can get better and more resilient performance in the SHM application context. We will also propose further low - energy detection solutions which can be embedded within the network nodes themselves and how a holistic low - power networking and detection architecture can be developed based on the same. 2.4. Through - Substrate Ultrasonic Co mmunication There have been other contemporary approaches towards the use of ultrasonic communication for sensor networking. The authors in [79] have outlined the advantages of using ultrasonic communication for human intra - body applications as opposed to radio communication. The latter is inefficient around human bodies because of water being a dominant constituent. The authors have discussed var ious adaptations to the MAC - layer protocol such as rate adaptation and stochastic channel access techniques. However, the advantages of these adaptations hold true mostly in the intra - body context (where radio propagation has disadvantages compared to ultr asonic), whereas the advantages due to the Pulse - based networking approach can be utilized even when the underlying technology is based on radio such as UWB [54] . Our choice of ultrasound as the medium for communication is mainly motivated by the advantages in integrating the sensing, communication and energy harvesting aspects in structural health monitoring scenarios which we detail further in the upcoming chapters of this thesis. 15 2.5. Spiking Neuron Based Event Pattern Detection The work in [80] was one of the first among many [81] [84] to describe a biologically plausible supervised synaptic learning rule that Linear Integrate and Fire (LIF) neurons can use to efficiently learn and read out spike - timing based neural codes. This is useful because neural codes that embed information in the spatiotemporal structure of spike patterns have been known to be computatio nally a very efficient means of encoding information. Thus, spiking neurons once equipped with the Tempotron learning rule should be able to leverage this. The Tempotron work [80] mainly attempts to show that a gradient descent - inspired learning rule, based on an error which is represented as the voltage distance between expected membrane potential for spike / non - spike and the maximum actual sub threshold membrane voltage, can be used with a very high capacity to decode information encoded in spatiotemporal spike patterns. This work [80] is primarily concerned with demonstrating that Tempotron can successfully distinguish between patterns from two different classes within some bounds on the number of total patte rns given the number of synaptic inputs to the neuron. In the current thesis, we develop a framework for applying the Tempotron learning rule to a spiking neuron for detecting selected event occurrence patterns vs others. We also lay out the scenario in te rms of a specific application (event pattern monitoring) and the spike patterns involved therein due to use of spike - based networking protocols. The - specific infe rences like structural anomaly detection, tracking progression of a structural crack along the structural substrate and so on. We provide the architectural details on how to create a Spiking Neuron - based architecture for such applications and why it would be useful in terms of energy - efficiency and deployment ease. We also articulate the research challenges and then provide a detailed characterization of the effect of different system and learning parameters on the 16 system performance. These include choice o f training patterns, choice of pulse - based protocols range of inter - event (spike) intervals, a presentation which has not been made in prior work. Other wor ks [81] [83] have also covered the topic of spatiotemporal pattern recognition using Tempotron - like learning rules. For example, in [81] the authors aim to create a Spiking neuron - based learning framework that is able to distinguish among multiple classes of input patterns wher e information is embedded in the precise timing of spikes relative to each other to generate precisely timed output spikes. The authors suggest two specific learning rules and the main improvement over the Tempotron rule is the fact that the output spike j itter can be lower than the input spike jitter thus making the system more robust to noise. It is to be noted here that our current proposal of Spiking Neuron - based detection can also be reproduced in a similar setting replacing the Tempotron - based learnin g with the Chronotron rule as mentioned in [81] . However, we want to show the baseline operation of the application system in the context of our f unctional specifications and so consider only the Tempotron learning scenario. Future work will cover the implications of using different learning mechanisms. Yet another approach to the spatiotemporal pattern classification problem using Spiking neurons is provided in [82] . This techniqu e, called ReSuMe, applies a similar learning approach to Tempotron [80] , but using a Widrow - Hoff rule to on the distance of actual output spike trains produced from the desired spike trains. It is to be noted here that approaches like ReSuMe can produce different spike trains for different classes of input spike train patterns such that the precise timing of the output spikes can be read out to make conclusions about the pattern class. This is more generalized than the Tempotron approach where the output pattern can be either a spike or no spike indicating only tw o classes. However, the latter 17 approach is suitable for the purposes of our application i.e. binary event classification and hence we chose to leverage the simpler Tempotron approach. In future work, we have plans to integrate other works like ReSuMe and r eport the efficacy of the system. Our main contribution in relation to the other established approaches [80] [84] to precise spike time based pattern classification lies in the fact that we adapt it with energy - efficient discrete - pulse - based networking protocols and consider the detection scenarios for different classes of unknown negative patterns and noise scenario s that would be present in a practical application scenario. 2.6. Summary In the following chapters of this thesis, we will develop scalability and energy - awareness within the inherent energy - efficient discrete - pulse - based networking paradigm. In addition, we will present an architecture for through - substrate ultrasonic communic ation for SHM applications using pulse networking and evaluate the system performance using realistic simulations for an airplane stabilizer structure. We will also develop a low - power Spiking Neuron based architecture for event pattern detection which we aim to incorporate in our pulse communication - enabled SHM architecture. We will mention the research challenges in this pursuit and how we tackle these. We will also include results from performance evaluation experiments for this low - cost detection archit ecture establishing its feasibility. 18 CHAPTER 3: SYNCHRONOUS PULSE NETWORKING In this chapter, we will cover briefly the fundamentals of the Synchronous Pulse Networking approach as has been publis hed in [35], [42], [54] . This is necessary because in subsequent chapters we will refer to various augmentations to the baseline protocol architectu re discussed here in response to application needs like network scalability, energy - harvested operation, and through - substrate communication. The basic premise is that a discrete - pulse - based networking approach is better suited in terms of energy efficienc y for applications with low information and latency demands. We will present here the protocol syntaxes and features which lend the discrete - pulse - based architecture its energy economy in target applications like Structural Health Monitoring. 3.1. Pulse Ab straction In the pulse networking domain, the key mechanism is to use individual pulses or the absence of such to indicate binary event information, that is presence or absence of an event. In low - information - density applications requiring ultra - low energ y operation, use of the pulse abstraction can significantly lower the energy overhead of transmissions compared to traditional packet - based mechanisms. Often, just the binary event information from distributed sensors when collated at a central Base Statio n is valuable for making high - level inferences about the system being monitored. However, beyond the binary event information, at least two more pieces of information are required for such systems, namely the event location information and the next - hop inf ormation to facilitate multi - hop routing. Event location is needed to give proper context to the event occurrence info gathered while multi - hop routing is generally preferred as it enables the use of small, inexpensive individual sensing devices. In order to preserve the energy advantages of pulse abstraction over packet - based networking, at least these two essential pieces of information 19 need to be encoded without incurring further transmission costs. In further sections, we will cover how this is achieved . Figure 3. 1 . Cellular Network Model on a Rectangular Plate Structure 3.2. Network Model Cellular Abstraction Before delving into the pulse networking semantics, it is important to understand the general network model envisioned for such applications. The network of sensors for collecting, for instance structural event information, will consist of sensor nodes dis tributed uniformly across the structure being monitored and a strategically placed Base Station (sink) where information from sensors across the structure are collated. For simplicity, in Figure 3.1 , we have shown an illustrative rectangular plate structur e with sensor nodes embedded throughout the structural area. Data collection nodes are referred to as non - sink nodes while the data integration node is called the sink node (placed on the left edge of the plate). The sensor nodes are equipped necessary eve nt sensing as well as forwarding functions (in terms of a pulse networking interface). Each node also has some pre - programmed localization information. Localization is accomplished with the 20 resolution of pre - defined sensor - cells like the hexagonal cells as shown in Figure 3.1 . Each cell represents an event area, with a unique Cell - ID. Since spatial localization resolution is at the cell level, shrinking the sensor cell size can increase the resolution. This feature gives us the flexibility to tune our local ization resolution based on application needs. Multiple sensor devices in each cell also enable data reporting redundancy which is valuable when the individual sensors might not be robust enough. Each sensor node belongs to one of these event areas (cells) and is pre - programmed with the Cell - ID of its own cell and those of its geographical neighbors. Although the cells in Figure 3.1 are shown to be hexagonal, there are no specific architectural requirements in terms of their symmetry, shape, and size. Generally, a hexagonal cell structure is chosen for maximal cell packing with minimum number of cells. Due to the cellular abstraction , the sensors are not individually addressed, and therefore no per - sensor addressing is necessary at the MAC or routing layers. 3.3. Cellular Pulse Protocol Frame Structure A summary of the Baseline Pulse Switching Protocol is provided in this section. T he objective is to highlight its major features, which make it an interesting choice for the application domain discussed earlier. The details of the protocol were originally presented in [35], [42], [54] . 3.3.1. Pulse as a Protocol Data Unit Upon detecting an event at a network node, a discrete pulse is sent to the sink using a multi - hop pulse routing process. Localization information about the received event is inferred by the sink from the time - of - arrival of the pulse with respect to a specified MAC - Routing frame structure as presented below. 21 Figure 3. 2 . MAC - Routing Frame for Pulse Switching [85] 3.3.2. Joint MAC - Routing Frames As reported in [35], [42], [54] , a network - wide synchronized frame structure, as shown in Figure 3.2 , is followed by each network node. The frame is controlled and enforced by the sink node. The frame is divided into many slots as shown in gray in Figure 3.2 . Each slot in a frame is used for sending a single pulse. The width of a slot is chosen based on the minimum pulse separation delay for the physical link under consideration. For example, this can be on the order of nanoseconds (ns) [53] for UWB pulse radio systems while in the order of milliseconds (ms) [8] for ultrasonic pulse communication systems. This slot duration should also be large enough to accommodate any cumulative clock - drift during a frame, as well as the propagation delay for the physical layer communication medium used. It is to be noted that the propagation delay effects might lead to sync pulses reaching different nodes at slightly shifted time instants, but if the pulses arrive within the designated slot, the system operates correctly. As shown in Figure 3.2 , each frame includes various down link and uplink areas. Operational details for these are presented below. 22 Figure 3. 3 . Pulse Forwarding in the Localization Area [8] A. Frame Synchronization Every frame starts with a Sync Area, which is allocated to the sink for network synchronization. During the Sync Area, the sink sends a pre - defined pulse pattern, which other nodes can detect and use to identify the start of a frame. The sink, having unlimited energy supply, can transmit the sync pulses with high enou gh power such that they reach all nodes in the network. Generally, this is a single pulse transmitted in the sync slot. B. Pulse Forwarding The uplink Localization Area of the frame is the key to pulse forwarding. In a network with M sensor - cells (excludi ng the sink cell), this area contains M slot - clusters, each cluster containing (M+1) individual slots. Each slot - cluster corresponds to a specific Cell - cell of origin. In each slot - cluster, individual slots correspond to specific Cell - IDs, which represent the next - hop cell for a transmitted pulse within the slot cluster. Figure 3.3 shows an example scenario to demonstrate pulse forwarding using the Localization Area. It is important to note that the slot cluster for a pulse remain s unchanged during the entire forwarding from the origin sensor 23 to the sink node, with the specific transmission slot dependent on the next hop as per the routing decisions. Hence, when the pulse arrives at the sink, its slot cluster indicates the cell of origin. maintains a sorted list of next - hop cells based on the hop - counts of the resulting routes based on a discovery process discussed hereafter. There is a lso a Control Area prior to the Localization Area of the frame, which is used for announcing impending transmissions within the Localization Area. All non - transmitting nodes can use this information to decide sleep and wake - up schedules for additional ener gy savings. C. Route Discovery and Response Mechanism These are optional components of the frame structure. The first, that is Route Discovery, is a continuous background process that creates and maintains the routing table in each sensor in terms of the next - hop Cell - IDs. The Response mechanism, on the other hand, enables a single - pulse acknowledgement for enhancing one - hop transmission reliability in presence of pulse errors. It is to be noted that the protocol works even without a Route Discovery scheme if there is a static routing table configured in each node before networking operation begins, but dynamic route discovery can help the network better adapt to changing conditions of routes e.g. in an energy dynamic environment. Response mechanism provide s communication reliability but might be skipped in low Pulse Loss Rate environments to save energy on redundant pulse transmissions. Details of these two components are covered in [42], [54] . 3.3.3. Protocol Features The baseline Synchronous Pulse Switching protocol also provides other useful features such as Route Diversity (RD), Spatial Compression a nd Pulse Merging. RD is a protocol 24 parameter, which indicates the number of separate paths along which the pulse information should be propagated towards the sink providing redundancy and better reliability in case of cellular fault scenarios. Spatial Comp ression is a feature used to prevent identical transmissions from multiple nodes in the same cell. Yet another feature is Pulse Merging, which provides inherent in - network aggregation for events originated from same cells. It is to be noted that only nodes in the same cell can transmit simultaneously and, in such cases, there is no interference, rather the pulses are reinforced on superposition. Details of these are discussed in [42], [54] . 3.4. Need for Scalability and Energy Awareness Improvements It is to be noted that with increase in number of cells, the delivery delay for the Baseline Synchronous Cellular Pulse Networking scheme described above increases quadratically with the number of cells due to the Localization area size of the frame. Hence, the protocol is mostl y targeted for networks with small number of cells for a reasonable localization resolution. When covering large areas, the protocol still works, though either with a reduction in spatial resolution (larger cells but lesser in number) or further relaxed de livery - delay specifications. However, when spatial resolution demand is high and further latency cannot be afforded, scalability suffers. This necessitates a fresh look at the fundamental time frame structure enabling pulse networking for possible changes to improve scalability. The baseline Synchronous Pulse Protocol architecture also implicitly assumes that the energy availability in the network is constant. In erratic energy - harvesting - powered systems, such an assumption is invalid and will lead to misma nagement of energy resources resulting in nodes running out of power prematurely. Hence, there is a need to incorporate energy - aware mechanisms within the baseline architecture to better equip the systems for energy - harvesting - powered operation. 25 3.5. Summ ary In this chapter, we covered the foundations of Cellular Synchronous Pulse Networking and event information. We detailed the networking protocol componen ts to show how medium access is managed and multi - hop routing achieved. We also consider the limitations of using this framework as - is in a dynamic and constrained energy environment or in networks with large cell counts. In the next chapter, we will devel op a revised Synchronized Pulse Networking Frame Structure which can scale for larger networks without sacrificing the energy - efficiency benefits of Baseline Pulse Networking or realizing further latency concessions. In subsequent chapters of this thesis, we will also develop various energy - awareness syntaxes within the Baseline Pulse Networking framework which can provide more judicious management of dynamic energy resources and thus better performance in energy - harvesting - powered systems. 26 CHAPTER 4: DEV ELOPMENT OF A SCALABLE PULSE FRAME STRUCTURE Building on the baseline synchronous pulse networking detailed in Chapter 3, in this chapter, we will develop a Scalable Cellular Pulse Networking (SCPN) architecture. We will design a novel time Frame Structur e which can enable graceful scaling of Pulse Networking latency with an increase in the number of network cells. As mentioned earlier, increase in number of network cells can be mandated by need for a higher resolution of detection/monitoring or increase i n the network area being covered. The basic premise is still to use discrete pulses to structure which is smaller compared to the baseline version, allowing bet ter latency performance without energy - efficiency concessions. 4.1. Scalable Cellular Pulse Networking (SCPN) Frame Structure The scalable time frame structure as shown in Figure 4.1 is designed to be comprised of the following components: A. Discovery Su b Frame (DSF): This is a downlink portion of the frame which has similar functions to the corresponding one in the baseline synchronous CPN frame structure as discussed in Chapter 3 of this document. This sub - frame is involved in discovering neighbor nodes during the in itial stages of network set - up and thus aids in formation of routing tables to be used during multi - hop event transmission from source to sink. The Discovery sub - frame uses up to (M+1) slots where M is the number of non - sink cells in the network. B. Next H op Sub Frame (NHSF): This is a part of the frame where a pulse in a designated slot indicates to the receiver that it should be awake during the Event Origin Sub Frame (discussed hereafter) part of the frame. The NHSF is used to notify the appropriate next hops about 27 information coming their way. In the NHSF, each cell i is allocated the corresponding slot i . Nodes with cell - ID i are only awake to receive in their own cell - slot i.e. slot i in the NHSF. Nodes are asleep in other slots unless they have pendin g information to forward. If a node has information to forward, it checks its neighbor table and selects the appropriate next hop cell - ID based on energy - aware routing strategies (detailed further in Chapter 5 ). If a node i decides to transmit to its neigh boring cell j , it sends this transmission in the slot where j is receiving i.e. slot j . A pulse at slot s in the NHSF tells a receiver with cell - id s that it should be awake during the whole period of the Event - Origin Sub - Frame (EOSF) to receive event data coming its way. The NHSF uses up to (M+1) slots. Figure 4. 1 . Joint MAC - Routing frame for Scalable Cellular Pulse Networking [85] C. Event Origin Sub Frame (EOSF): The Event Origin Sub - Frame is used to indicate the event origin location for the information being forwarded. A pulse received at slot s in EOSF indicates an event originated at a node with cell - id s . It is to be noted here that the only nodes receiving i n the EOSF will be the ones that have been notified about events coming their way during NHSF. So, all nodes indicated as next hops in NHSF will be awake and receiving in all slots of EOSF while all other non - transmitting nodes will be sleeping. All transm itting nodes will 28 be awake and transmit a pulse in the respective event origin cell - slots. The EOSF uses up to M slots where M is the number of non - sink cells in the network. Response Sub Frame (RSF): The Response Sub - Frame is identical to the Response Ar ea in the baseline Pulse Networking Scheme and an optional component to guarantee reliable detection of pulse errors / false positives. The RSF uses up to (2*M + 1) slots. Figure 4. 2 . Demonstration of multi - hop event forwarding in SCPN Figure 4.2 demonstrates a multi - hop forwarding operation using the SCPN frame structure. An event originating in a node with cell - ID 6 is forwarded through nodes in cells 5, 3, 1 up to the sink (cell - ID 0). In Frame 1, the node is cell - I D 6 transmits a pulse in the NHSF slot 5. This indicates to the node in cell - ID 5 that an event is to be sent to them. As a result, nodes in cell - ID 5 keep awake and listen during all slots in the EOSF of the same frame. Since the event is originated at ce ll - ID 6, the node in cell - ID 6 transmits a pulse in slot 6 of the EOSF which the node in cell - ID 5 receives. In the next frame, the node in cell - ID 5 forwards the event it received in the last frame by first sending a pulse in slot 3 of NHSF to notify its neighbors in cell - ID 3. In the EOSF of this frame, the node in cell - ID 5, transmits in slot 6 (indicating the origin cell of the event) which 29 is received by the node in cell - ID 3 which is awake in EOSF because of the NHSF transmission of node in cell - ID 5. Hence, the event originated in node with cell - ID 6 has now been propagated up to the node in cell - ID 3. Using a similar procedure, the node in cell - ID forwards the event up to 1 and then on to the sink (cell - ID 0) as shown in Figure 4.2 . It is to be noted that the NHSF and EOSF together combine the functionality of the Localization Area in the baseline Pulse Networking protocol. The important distinction is that while the Localization Area uses up to O(M 2 ) pulse slots, the NHSF and EOSF in the new frame st ructure uses only ( M + 1 + M) = (2*M + 1) i.e. O(M) pulse slots and is thus much better scalable with increase in the number of network cells. This saving in terms of the frame size can be utilized in multiple ways e.g. reducing the overall delivery delay or toward increasing the amount of information encoded in each frame while keeping the delivery delay at par with the baseline approach. The trade - off in the current approach is the lack of distinction between event origin data heard from different nodes i n the EOSF. During the EOSF, there is a possibility of nodes, which are awake, overhearing pulses which are not meant for them and thereafter forwarding these pulses to their neighbors. As a result, there might be extra energy used or leakage for communica ting redundant information (because the same information will also be propagated through the node where it was originally intended). In the current study, we have not proposed any solution to this because as the simulations demonstrate the scalable frame s tructure gives better energy performance compared to the baseline Pulse protocol despite the overhearing leakage if any. Moreover, overhearing is not always a problem. It can only become a source of leakage when data originated from different sources are b eing propagated in the same frame across different routes and the nodes on these routes are neighbors such that one can overhear pulses meant for the other. 30 Such confluence of factors is very rarely the case. For example, in many cases the events being pr opagated across different routes are originated at the same node. In such cases, there will effectively be a merging of the pulse in the same slot and overhearing is not an issue. Also, if the routes in consideration are spread apart, there would be no ove rhearing. Another important thing to note is that the root cause of overhearing is the fact that event pulses meant for multiple destinations are being transmitted in the same sub frame. This problem can be ameliorated by staggering the transmissions from neighboring nodes if there are multiple nodes forwarding the same data. 4.2. Simulation Setting and Performance Results 4.2.1. Network, simulation and event model An event - driven C++ simulator implementing MAC - Routing framing for both CPN and SCPN was developed. The baseline simulated network consists of TUPN sensor nodes evenly distributed on a rectangular plate structure with the sink node placed at the center lef t corner as indicated in Figure 4.3 . The nodes are grouped into regular hexagonal cells with an average of 3 nodes per cell. The spacing between individual nodes is 0.25 m and the transmission range is kept at 0.75 m. Height of each baseline hexagonal cell is 0.5 m. We used different dimensions of the plate to vary the network size and hence the number of cells in the network to compare performance of SCPN vs CPN as the network cell count is scaled. The effect would be similar if the cell count were increas ed by reducing the cell size. The latter is also often the case when higher localization granularity is required. As shown in Figure 4.3 , we used 6 different plate sizes of dimensions 4m X 3m , 6m X 3m , 8m X 3m , 10m X 3m , 12m X 3m with the respective cell c ounts as 32, 41, 59, 68, 86 and 113 respectively. The plate width was kept similar, to have comparable 31 routing across the network despite the scaling when a common cell is chosen as source among all the networks. Figure 4. 3 . Simulated Network Topologies Figure 4. 4 . Spatiotemporal variation in harvested energy the idle listening and s Energy availability 32 at the nodes was designed to have both spatial and temporally variable components. Temporal variation was incorporated with the use of a duty cycle (DC) parameter which controls the proportion of time when harvesting power is available. Harvesting power av ailability is modeled as a stochastic process with energy availability (charging ON time) and non - availability (charging OFF time) intervals being exponentially distributed. The means of these distributions are chosen based on the DC selected. With this co nfiguration, the harvesting power availability for different DCs is demonstrated in Figure 4.4 (a) with higher DC involving denser power availability leading to more harvested energy. Spatial distribution of energy is incorporated using a spatial distribut ion function which is linear in the current paper as shown in Figure 4.4 (b). Thus, the nodes near the left edge have less energy harvested compared to the nodes near the right edge assuming the anchor of the plate is at the left edge and tapering away tow ard the other edge as shown in the inset in Figure 4.4 (b). This is representative of the energy available using vibration harvesting on the plate which is more intense away from the anchor than near. It is to be noted that all nodes are charging synchrono usly in the simulated model because energy availability events are synchronous across the plate area. Events were generated at different source nodes across the rectangular plate and the corresponding average performance characteristics were noted. Events were assumed to be far apart from each other so that they can be considered as independent transmissions. For event transmission from any source node, 50 similar experiments were performed, and their average noted to take care of the stochasticity in the c harging model introduced as part of the exponentially distributed Duty Cycle formulation. 33 4.2.2. Performance Results A. Delivery Delay and Event throughput To demonstrate the advantages of the SCPN design as compared to baseline CPN, we evaluated the network performance across different network sizes in terms of event delivery latency and throughput. In the initial experiments, we considered single event generation from cells across the plate to the sink and the highest level of energy availability i.e . no energy constraint. Further effects of energy constraint have been discussed in subsection C hereafter. Figure 4. 5 . Network Cell Count vs Average Per Hop Delivery Delay / Event Throughput As shown in Figure 4.5 , network cell count has a more marked bearing on the delay / throughput parameters in the case of CPN as opposed to SCPN. It is to be noted that the y - axis of the delay / throughput plots is exponentially scaled which is intentional to demonstrate how qui ckly the average delay per hop and average event throughput scale in CPN when the cell count is increased. This is opposed to SCPN where the event delay increases much more gracefully i.e. slope of latency increase or throughput decrease is lower when the cell count is increased from 32 to 113. It is to be noted that due to the use of a linearized frame structure, SCPN offers per hop latencies in the order of 1 - 5 seconds while CPN for comparison has per hop delays which are on 34 the order of 10 - 100 seconds. S imilarly, for the average throughout, SCPN offers much better values of 0.1 - 1 events/second for the used network topologies as opposed to CPN which offers only 0.1 or lower. Both results are a consequence of the fact that latency and throughput are functio ns of the frame size and CPN aims to gracefully scale the same. The improved performance ranges open up the feasibility of more time - sensitive applications as well as ones where a high throughput is required for pulse - based networking. Figure 4. 6 . Heat maps for source - sink event delivery latency B. Source to Sink Event Delivery Latency For a better visualization of the latency advantages provided by SCPN, we have charted the source to sink delay (complete routing delay) for single events generated at different nodes across the plate in Figure 4.6 . As seen from the heat maps provided in Figure 4.6 , events that are located farther away from the sink, have much higher event delivery latency from source to sink compared to t hose located nearer to the sink. Such disparity becomes even higher when the 35 network cell count is increased as the network size increases. Thus, the distribution of event delivery delays is very non - uniform in the case of CPN and nodes farther away from t he sink are implicitly disadvantaged. In the case of SCPN though, the source to sink delivery delays scale much more reasonably as the distance from the sink changes as well as due to network cell count changes. This can lead to a fairer event reporting sc hedule across the network. C. Effect of energy constraint on event delivery latency To evaluate the effect of energy constraint on the network performance, we considered two different network sizes - 4x3 with cell count 32 and 6x3 with cell count 41. As s hown in Figure 4.7 , when the energy constraint in the network increases i.e. the duty cycle decreases, the average delay per hop increases for SCPN and much faster than that of CPN. Figure 4. 7 . Effect of Energy Constraint on Per Hop Delivery Latency At the highest energy constraint i.e. duty cycle 10%, the delay for SCPN and CPN are approximately similar. This nature of the latency is a consequence of the fact that with energy 36 constraint involved, the per hop lat ency not only a function of the frame structure, but also related to the delay brought about due to energy - aware event buffering. When the energy is low, irrespective of the protocol, the latency is large because it is primarily dictated by the time it tak es for the nodes to acquire enough energy to transmit the pulses needed for event information transport. Hence, for either case of SCPN vs CPN, the latencies would be similar which masks the disadvantages of the longer frame structure of CPN compared to SC PN. In some low energy cases, as in DCs < 40 % in Figure 4.7 , the delivery delay from SCPN can be higher than CPN. This is because, since more frames are completed for SCPN which involves more sync pulses as well as more attempts at node discovery, SCPN us es up more energy compared to CPN in the same time. This affects further transmission along the route and thus effective delivery delay. The observation from this graph is important because it indicates two things one, that SCPN is roughly on par with CP N at high energy constraint, but is far more advantageous in network situations with low energy constraint; secondly, SCPN has lower increase in delivery delay with increase in network cell size irrespective of energy constraint. D. Energy Consumption It is to be noted that the energy consumption for both CPN and SCPN are roughly similar throughout the frame structure which is by design. In CPN, for a single hop transmission, the energy consumed by the network includes the transmission energy for the so urce / forwarding node and the reception energy for the receiving nodes i.e. idle listening. The number of pulses that need to be transmitted per event from a source / forwarding node for both CPN and SCPN are equal. For example, assuming a route diversity of 2, SCPN would have to transmit 4 pulses - 1 pulse for sync, 2 pulses for next hop and 1 pulse for event origin. Similarly, CPN would also involve 4 pulses 1 pulse for sync, 1 for control, 2 for localization. Also, the number of slots that the 37 receivi ng nodes need to be idle listening for events is the same for a network with M non - sink cells, CPN receivers must be idle listening in all the slots in control frame i.e. M slots and 2 slots in Localization Frame, i.e. a total of M + 2. Similarly, in SCP N, receivers would be idle listening for 2 slots in NHSF and M slots in EOSF i.e. M + 2. Hence, the total energy consumption due to transmission and idling is identical for both protocols. SCPN stands to make some extra energy consumption when overhearing is involved as explained earlier. However, with proper strategies in place such as merging and staggering of events, such extra consumption is essentially removed as seen in our simulations. Thus, using SCPN, we can obtain better network performance with s imilar levels of energy consumption when the energy constraint is not very high. Figure 4. 8 . Effect of Information Content Size on Delivery Latency of Pulse Protocols E. Higher granularity information transport As mentioned pre viously, using the better scaling provided by SCPN we can include more information within the frame structure without sacrificing on the energy efficiency or latency. This 38 is demonstrated in Figure 4.8 above. In the baseline case, we use only 1 bit of event data i.e. the origin event area id. We can include further bits of information e.g. whether the available node energy is over / under a threshold etc. If such information were to be included in the C PN frame structure, the delay would increase significantly as demonstrated in Figure 4.8 for 1 - bit vs 2 - bit and 3 - bit information. In SCPN though, such increase of information density does not need to come at a significantly higher latency price. In fact, even 2 - 3 bits of event information can be used in SCPN with a latency far less than the latency for 1 - bit info using CPN. Such advantages are more pronounced when the network cell count is larger. 4.3 . Summary In this chapter of the thesis, we developed a novel time frame structure for discrete - pulse - based networking that scales well with increase in the number of network cells. We also evaluated the performance of the new scalable pulse networking architecture vs the baseline synchronous cellular pulse ne tworking approach introduced in Chapter 3. We have shown that the SCPN protocol can provide similar energy - efficiency to the CPN while lowering the delivery delay. The protocol can also be reliably used in a wide range of harvesting situations. In the next chapter, we will discuss energy awareness features that should be included in the Baseline Synchronous Cellular Pulse Networking approach for more judicious energy management in extreme resource - constrained or dynamic energy availability scenarios. We wil l detail the energy aware protocol design and show the advantages of these new syntaxes through simulation experiments for a vibration - harvesting - powered network. 39 CHAPTER 5: ENERGY - AWARE PULSE NETWORKING In earlier chapters of this thesis, we have menti oned that we envision the use of networks of sensors powered by ambient energy harvesting. For instance, in Structural Health Monitoring applications, such harvesting can be done from ambient structural vibrations using piezoelectric transducers [85] [88] . The advantage in a harvesting - powered network is that the individual sensors can be less expensive and bulky (because large batteries need not be included) while being practically maintenance - free (as energy availability is theoretically limitless). Howe ver, such ambient harvesting sources are by nature very erratic and energy availability cycles might not be well - synchronized with energy use cycles. As a result, in absence of appropriate energy management, there might be leakage and eventual sensor power loss affecting effective network lifetime. In this chapter, we will develop an energy - aware pulse networking protocol which can deal with these shortcomings of the baseline synchronous pulse networking architecture. We will show that in systems powered by various levels of energy harvesting availability, the energy - aware version of the protocol performs significantly better compared to the baseline version. We will also show how the performance varies for different harvesting scenarios based on vibration h arvesting on a simple rectangular plate structure. 5.1. Energy - aware Pulse Switching Protocol This section introduces specific protocol syntaxes developed to deal with stochastic energy harvested , for instance, from the ambience (e.g., vibration) of a target structure. As mentioned in Chapter 3, the synchronous pulse networking architecture consists of the use of discrete pulses to indicate event information. The baseline architecture does not change transmission schedules based on energy availability. When an event is available, transmissions are attempted. In case of energy - constrained scenarios, for example, slow harvesting, the baseline protocol would attempt 40 transmissions even when enough energy is not available to transmit all necessary component p ulses. Since the transmission is incomplete, this energy is essentially wasted. In the proposed energy - aware version of the protocol, the main idea is to manage the stochastically available energy in a prudent manner such that the wastage on incomplete pul se routing is minimized while the number of fully routed events is maximized . 5.1.2. Frame Adaptation The same frame structure , as discussed in Chapter 3 without the Route Discovery and Response Module , is used for this energy - aware version. The route discovery part is omitted due to its energy intensiveness. A static routing operation with pre - defined routing table entries is chosen instead. The response area which is responsible for dealing with pulse losses is also removed . Given the low pulse loss rates (PLR) of the physical layer proposed i.e. ultrasound communication [8] , it was decided that the benefits of the response mechanism may outweigh its energy overheads. Thus in the revised frame format, every frame starts with a Sync Area followed only by the Control , Localization and Protection areas. 5.1.3. Binary Event Buffer Due t o insufficient energy, a pulse sensor network can become partitioned i.e. a pulse may arrive at a node but cannot be forwarded due to the lack of energy required for transmission. To accommodate such scenarios, a binary event buffer is provided which can b e used for storing such a pulse until enough transmission energy is available for a forwarding operation. As shown in Figure 5.1 , the length of the buffer at each node is M , where M is the total number of sensor cells in the network. At node - j , the i - th bu ffer bit is set to 0 or 1, depending on the absence or presence of a pulse originated at sensor cell i , buffered at node - j . At the start of each frame, every node first 41 checks whether there are any buffered events, and tries to transmit them in the respective slots based on its energy availability. When a pulse is generated or received (for forwarding) at a node, it is stored in the binary event buffer in the bit correspo nding to the cell of its origination. If the i - th bit of the buffer is already 1 in a node, and a new pulse originated at Cell - Id - i is received by the node, it simply ignores the new pulse without any changes in the buffer. This causes the old event to be merged with the new one at this point in the route. Such merging results in a loss of temporal resolution of event detection. For many structural health - monitoring applications, however, such a loss can be acceptable due to their low requirements on tempor al granularity of event detection. Figure 5. 1 . Summary of Energy - Aware Pulse Switching 5.1.4. Energy - aware Forwarding Syntaxes Each networked sensing unit contains an energy storage super capacitor that is charged at a rate based on the harvesting source. For structural health monitoring applications using Through - Substrate Ultrasonic Pulse Networking (TUPN) units as discussed in Chapter 1, this harves ting source can be the ambient vibration s in the structure where the TUPN is embedded . All baseline forwarding decisions, as explained in Chapter 3 , are modulated by the amount of energy available in the super capacitor and the estimated energy expenditure for sending the pending events that are 42 buffered in the binary event buffer. For example, the event buffer in Figure 5.1 indicates buffered/pending events from cell IDs 3 and M - 1 that need to be forwarded. The protocol estimates the energy needed to send those event pulses with different route diversities (as defined in chapter 3 ) and makes selective forwarding decisions based on that estimation and the available energy in the super capacitor. While checking for available energy, the protocol first asce rtains whether there is enough energy to transmit all pending events with the highest allowed RD , and if not then with the next lower RD and so on. If energy is not sufficient to transmit even with lowest RD (i.e., 1), then the node tries to transmit one e vent less than the number of pending events, with RD s starting from the highest and so on. If not even one event can be transmitted with the lowest RD , the node suspends transmissions in the current frame. Total energy consumption E c is estimated as: E c = (n ctrl + n loc ) * e pulse + p idling * F ( 5. 1) where n ctrl and n loc are the number of Control and Localization area pulses respectively, e pulse is the energy consumption for a single pulse generation, p idling is the idling power, and F the frame durat ion. The idling energy consumption for a full frame is considered as a conservative estimation to remove the chance of nodes running out of energy in the middle of a frame. It is to be noted that n loc in the above equation is a function of the Route Divers ity ( RD ), and hence estimates the energy expenses for different route diversity values. Depending on the amount of energy in the super - capacitor and the estimated amount needed, a node may be able to forward only a subset of all the buffered events. To ens ure fairness, the events from the least recently used bits in the event buffer are selected for transmission first. 43 Figure 5. 2 . Routing Decision based on Energy Levels Figure 5.2 shows different energy availability situations in the super - capacitor and how they influence the routing decisions as per the protocol described above. A level E - Ti - Rj indicates the estimated energy needed for transmission of i events with a route diversity j . Each column in the figure shows an energy availability scenario (i.e. the y - axis levels) and a corresponding forwarding decision , marked as a {Allowed Events (AE), Route Diversity (RD)} pair. After the allowed events are s ubsequently transmitted, the y are removed from the event buffer. 5.2. Performance Results An event - driven C++ simulator implement ing the MAC - Routing framing was developed. Both the non - energy aware (NEA) and energy aware (EA) versions of the protocol were simulated. Note that the baseline energy advantages of pulse switching over traditional packet - based event reporting have already be en established in [ 4 , 5 ]. In this chapter, we mainly validate and characterize the effectiveness of the energy - awareness of the architecture in a harvesting scenario . We first consider a scenario where the energy generation is synchronous across the networ k i.e. all nodes receive similar energy harvesting. In the next scenario, we cover the efficacy of energy - aware pulse networking in scenarios where the energy generation is non - uniform across the network. 44 5.2.1 Results with Synchronous Energy Generation A. M odel for Vibrational Harvested Energy Synchronous The objective is to develop pulse network protocols that can operate using sparse energy, harvested from system ambience such as aircraft wing vibration. A piezoelectric harvesting mode using synchronous vibration is assumed. This implies that all the TUPN units mounted on a structure experience similar levels of ambient vibrations and resulting charging of their respective super capacitors. Discharging of the capacitors, however, will depend on the gener ated events and their resulting spatial - temporal patterns in different parts of the structure. An on - off charging model is used for evaluation of the protocol performance. In this model, corresponds to the availability of structure v ibration sufficiently intense for harvesting. Figure 5. 3 . Capacitor charging for different duty cycles To simulate memory - less vibration patterns, exponentially distributed random processes harvesting is parameterized by the average duty cycle of the on - off pattern . A higher duty cycle 45 piezoelectric harvesting mechanism and its efficiency [89] . For the results in the next section 4.2., charging patterns for two different vibration duty cycles are shown in Figure 5.3 . The presented protocols were evaluated with different harvesting/charging rates representing different harvesting technologies and their efficiency. B. Network Structure and Event Generation The simulated network consists of 210 evenly distributed TUPN sensor no des with a sink node placed at the lower left corner of a rectangular terrain that measures 6m x 3m . The nodes are grouped into 84 regular hexagonal cells with an average of 3 nodes per cell. The spacing between individual nodes is around 0.3 m and the tra nsmission range is kept at around 0.75 m. Height of each hexagonal cell is 0.5 m with cell side length of around 0.288 m. Events are generated at different network locations. For each event, a pulse is sent to the sink repeatedly with a pre - specified int er - pulse interval , such that even if a specific pulse is lost on its way due to insufficient energy availability or unreliable ultrasonic links, successive transmissions will increase the probability of successfully informing the sink about the event. Expe riments were conducted with a variety of such pulse repetition intervals in order to understand their implications on event delivery latency under differen t energy constraints. C. Network Node Energy Traces Simulation experiments were performed using three d uty cycles (DC) i.e. 20%, 30%, 40%, and three different pulse repetition intervals (PRI) i.e. 20, 10, and 5 frames for both NEA and EA pulse switching. The energy budget for a single pulse transmission was chosen as 10 mJ, and constant idling power consump 46 synchronous structure vibration as discussed in Section 4.1.5 . All the graphs in Figure 5.4 demonstra te the remaining energy in the super capacitor of a node on the event delivery route. The first notable observation is that for the same vibration DCs (i.e. degree of harvesting energy constraint) and PRIs, the NEA protocol often leads the super - capacitor to zero - energy condition, whereas the EA version manages to avoid it. Figure 5. 4 . Energy trace at a node on event propagation route It should also be noted that irrespective of the duty cycle , the consumption cycles in the NEA protocol operation are more sensitive to the pulse repetition intervals . For example, when the event generation interval is 5 frames, there is visible energy consumption after every 5 frames in the NEA protocol at all DCs, indicating pulse transmissions . However, at the lower DCs , the observed drop in energy is not enough for transmission of all the pulses, which requires at least 20 47 mJ for the lowest RD . This indicates that only a subset of all the needed pulses for an event are being sent. As a result, certain amount of energy is wasted even though the event is not successfully transmitted. This can affect future event transmission s in an adverse manner. This is especially true in the low DC scenario s where there is never enough energy to transmit all the pulses for an event at the lower PRI s (i.e., 5 and 10 frames). For the higher PRI s, since the node has more time to accumulate energy between successive event transmissions, incomplete transmissions are less frequent. In contrast, for the EA protocol, irrespective of the PRI , consumption happens only when sufficient amount of energy (i.e., 20 mJ) has been accumulated for sensing all the pulses n eeded for an event transmission. This avoids wastage caused by partial transmissions and therefore improves the overall energy economy. In all scenarios, as the duty cycle increases, the operation of the EA protocol becomes similar to the NEA case because of higher energy availability. D. Event Reporting Performance Figure 5.5 depicts the event reporting delay, which is the primary performance index of the system. Such delay is defined as the latency between the transmission of first pulse for an event fro m the source node and the first time that particular pulse or a subsequently repeated pulse for that event arrives at the sink node. This indicates the effective delay in reporting a structural failure event. Such delay is caused due to: 1) end - to - end rout ing delay, 2) pulse losses because of channel errors, 3) deferred/stalled pulse routing due to event buffering caused by energy unavailability in the energy - aware routing. Due to the pulse repetition feature of the event generation model, and the event buf fering in the energy - aware routing protocol, an event is guaranteed to be reported with infinite delay. 48 Figure 5. 5 . Event Reporting Delay for Different Duty Cycles To accommodate practical settings, however, a notion of A llowed D elay (AD) has been developed while defining the concept of successful reporting. An event is regarded as successfully reported if it reaches from source to sink within a certain Allowed Delay . Figure 5.5. reports results for four different ADs, namely, 60, 120, 180 and 240 frames. For the required frame duration of 76 sec corresponding to our network , in Figure 5.5 those delays correspond to approximately 1, 2, 3, and 4 hours respectively. Tho se A D numbers are deemed reasonable for typical structural health monitoring applications. Observe in Figure 5.5 , for all vibration DC s and AD s , the EA version of the protocol reports events quicker than the NEA version. Also , as expected, with higher A D limits, the average reporting delay for the NEA protocol increases , as more events can be considered as successfully delivered in the longer intervals. The EA protocol, however , appears to be insensitive to the A D . This is because of its ability to deliver all events with reporting delays that are well below the lowest A D even at the lowest vibration DC . We do not consider DC s below 20% because in such ca ses events are never delivered within the highest A D for the NEA version. Note that in 49 all the scenarios, for high DC s (i.e., 30% and above ) both EA and NEA protocols behave similarly. The reporting delay stabilizes for both protocols since they are able t o deliver all events successfully within the lowest A D due to high energy - availability. It is notable that improvements in delay demonstrated by the EA protocol are more pronounced under severe energy constraints, i.e. the low DC scenarios. Improvements ar e also higher with higher A Ds . This can be explained using the next set of results showing the event delivery ratio in Figure 5.6 . Figure 5. 6 . Event Delivery Ratio at Different Duty Cycles Event delivery ratio is defined as the fraction of generated events that are successfully reported to the sink for a given A D . As can be seen in Figure 5.6 , for the NEA protocol, the delivery ratio improves with more structure vibration up to around DC 30%. Delivery ratio also improves with higher AD . Beyond that point, the system enjoys enough harvested energy, which result s in near - 100% event delivery. The EA protocol, however, is significant ly less sensitive to both vibration DC and A D because of its judicious energy usage and event buffering as presented in Section V. In summary, the proposed EA protocol syntaxes show significant improvements over 50 its NEA counterpart in low structur al vibrat ion situations. Such improvements are consistent for both event reporting delay and delivery success. Figure 5. 7 . Route diversity for different energy availability E. Impacts of Adaptive Route Diversity A key component of the EA protocol is its ability to choose appropriate route diversity (RD) based on the instantaneous energy availability. Figure 5.7 reports the occurrence distribution of RD s under different vibration DC s and PRI s. It can be observed th at for a given PRI , higher diversity routes occur more frequently as energy becomes plentiful (i.e. higher vibration DC s). For a given energy availability (i.e. DC ), routes are more diverse for larger PRI s and less for smaller ones. This is because with mo re frequent pulse repetitions (small PRI) at the source, there is a bigger drain on the available energy, which prevents the protocol from choosing energy - expensive higher route diversities. These results validate the protocol s ability to choose pulse RD in a manner that is adaptive to the instantaneous available energy. For the NEA version, such diversity 51 is statically chosen and can cause energy wastage leading to higher reporting delays and lower delivery success , depicted in Fig s . 5.5 and 5.6 . F. Ev ent Buffering Characteristics In Section 4.1.3, we presented how event buffering can be leveraged in the absence of sufficient pulse forwarding energy in the proposed protocol for reducing event losses and the resulting reporting delay reductions. Figu re 5. 8 . Event buffering/storage delay (in frames) distributions Figure 5.8 reports an example distribution of the per - hop buffering delay (i.e., averaged over all the hops on an end - to - end route) experienced by an event on its way to the sink node. As s the lower end of the buffering delay as the vibration DC increases. This implies that when energy is scarce (i.e. low DC), the protocol employs longer storage times to allow sufficient accumulation of harvested energy before transmitting all the needed p ulses for complete event forwarding. As the energy 52 becomes more abundant, the need for storage becomes less important and storage times approach the lower limit of one frame per hop. These observations validate the event - buffering process and its intended impacts. 5.2.2. Results with Synchronous Energy Generation A. Model for Vibrational Harvested Energy Asynchronous The main objective of following evaluation was to evaluate the performance of Energy - Aware Pulse Switching protocol in dynamic energy harv esting scenarios without any limitations like synchronous energy harvesting in all nodes of the network. To that effect, a few representative vibration harvesting models were constructed for the evaluation experiments. In order to simulate an asynchronous vibration energy harvesting model, we used an exponentially distributed ON OFF charging (energy accumulation) model modulated by a spatial distribution function. The exponentially distributed ON OFF model helps to simulate the temporal variation in vibrat ion (and hence, charging) while the spatial distribution function incorporates the spatial variation in harvested energy profiles. In order to parameterize the temporal variation in harvesting efficiency, we define a parameter called Duty Cycle as follows Duty Cycle = Average of ON time distribution / (Average of ON time distribution + Average of OFF time distribution) , where the ON time and OFF times of the model are both exponentially distributed to emulate memory less processes mirroring the fact that structural vibrations available are essentially stochastic in nature and not correlated. Duty Cycle, therefore, gives us a measure of the harvesting 53 efficiency of the system on a temporal level. The spatial distribution function is multiplied with the duty cycled energy to obtain a spatiotemporally variable vibration profile. We have used four different spatial distribution functions (as detailed in Equations 5.2 5.5 ) to evaluate the efficacy of Energy - Aware Pulse Protocol operation when powered by a variety of vibration profiles. These vibration / harvested energy profiles correspond to the four anchor configurations for the rectangular plate as shown in Figure 5.9 (a) - (d). Figure 5. 9 . Rectangular plate designs with various anchor configurations In each equation below, the energy harvested during a single ON time of the vibration profile is evaluated. This energy is accumulated over the course of time as more ON times are encountered, leading to energy harvesting in a particular node. In relation to the equations below, the definition of component parameters has been provided in Table 5.1 . Table 5. 1 . Asynchronous Harvesting Model Parameters Parameter Name Parameter Representation Parameter Definition Charge Interval t charge ON time interval (modulated by the duty cycle) Charging Power P charge Harvesting power during the ON time (assumed to be constant) - m, c, k parameters determining the spatial variation 54 - x X coordinate of the node whose energy is being calculated - y Y coordinate of the node whose energy is being calculated - x i X coordinate of the node i - y i Y coordinate of the node i Energy Charged e charge Energy accumulated in one Charge Interval = ( 5.2 ) = 5.3 ) = 5.4 ) = 5.5 ) The first configuration (Fig 5.9 (a) ) assumes that the anchor is continuously attached along the whole length and breadth of the rectangular plate. As a result, there is no spatial variation involved as indicated by the co nstant value of the Harvesting Power Multiplier in Fig 5.10 (a). This is because all the area of the plate experiences a consistent level of vibration due to similar anchoring configuration throughout. This effect is modeled in the Equation ( 5.2 ). In Figure 5.9 (b), an anchor configuration is demonstrated which tapers down from the left edge of the wing to the right edge. The anchor width is unchanged along the y - axis for all points which are located at a fixed distance from the left edge. Such a varia tion is modeled using the Equation ( 5.3 ) with the corresponding harvesting multiplier variation shown in Figure 5.10 (b). 55 Figure 5. 10 . Spatial variation of node charging profiles at a fixed duty cycle The third scenario illustra tes a case where the anchor is present only near the left edge of the plate as shown in Figure 5.9 (c). As a result, rate of increase of vibration levels is higher when we move farther from the left edge. The corresponding harvesting power multiplier is demonstrated in Figure 5.10 (c) where the slope increases in quadratic manner in accordance with Equa tion ( 5.4 ). Figure 5.9 (d) demonstrates a slightly different anchoring scheme where the plate is affixed at the center instead of along the edge. This is represented by the charging equation in ( 5.5 ) where the energy accumulated depends on the position awa y from the center (in this case calculated using the mean of all the node positions since they are uniformly distributed along the two axes). The corresponding harvesting multiplier variation is shown in Figure 5.10 (d). As mentioned above, 56 the spatial var iation in the equations ( 5.2 ) - ( 5.5 ) is demonstrated in the following temperature maps where the harvesting power multiplier based on the equations is plotted versus node location. This represents the available harvesting power distribution across the rec tangular plate at a constant duty cycle or alternatively at a certain instant in time. Figure 5. 11 . Temporal variation in harvested energy availability using a duty cycled approach The temporal variation (change in duty cycle) for a specific spatial variation is demonstrated in Figure 5.11 . As we move from Figure 5.11 (a) - (e), the duty cycle increases and as a result we can see that the temporal availability of harvesting power avail able is higher i.e. harvesting occurs at a higher rate leading to faster accumulation of energy. As we will show in the following sections, the spatial distribution function and the temporal harvesting efficiency combine to create a spatiotemporal variatio n in the harvesting profile which is then used to evaluate the operation of the Energy - Aware Pulse Switching Protocol when powered by the same. 57 B. Network structure, simulation parameters and event generation model The simulated network consists of 1057 T UPN sensor nodes evenly distributed on a rectangular plate structure with a sink node placed at the center left corner as shown in Figure 5.12. The nodes are grouped into 84 regular hexagonal cells ( like Figure 5.12 ) with an average of 3 nodes per cell. The spacing between individual nodes is around 0.3 m and the transmission range is kept at 0.75 m. Height of each hexagonal cell is 0.5 m with cell side length approximately 0.288 m. An energy model is used where the energy for transmitting a single pulse is 10 mJ while the reception energy budget is 10 W (10uJ/sec) including the energy required for pulse detection and idle listening. Figure 5. 12 . Cellular network model on a rectangular plate structure Events were generated at different source nodes across the rectangular plate and the corresponding performance characteristics were noted. Events were assumed to be far apart from each other to be considered as independent trans missions. For event transmission from a particular source node, 100 similar experiments were performed, and their average noted to take care of the 58 stochasticity in the charging model introduced as part of the exponentially distributed components Duty Cycl e formulation. C. Network Node Harvested Energy Traces In order to demonstrate the spatiotemporal variation of harvested energy across the simulated rectangular plate, the energy traces for three representative nodes on the structure were obtained at varying levels of energy availability (signified by the Duty Cycle of ON - OFF charging). For example, the energy traces shown in Figure 5.13 , were obtained from simulations on the rectangular plate anchored at the center. It is evident from the graphs that the energy harvested increases monotonically with increase i n the temporal frequency of vibrations represented by the Duty Cycle (DC). The spatial variation in energy harvesting is also easily discernible with nodes near the anchor (e.g. Node 111) enjoying much less in terms of harvested energy compared to nodes aw ay from the anchor (e.g. Node 183). It is to be noted that the former node always has lower levels of vibration being closer to the anchor location and thus consistently lower levels of harvested energy compared to the latter node irrespective of the duty cycle of vibrations. The positions of the selected nodes on the rectangular plate have also been indicated to demonstrate the dependence of node location on the harvesting rate. The figures thus demonstrate the spatial variation in energy profile (among t he selected nodes at each duty cycle) as well as the temporal variation (for each node across the range of duty cycles observed). The range of duty cycles observed is divided into exponential intervals (0.025, 0.063, 0.158 etc.) such that the temporal vari ation for a large range of duty cycles can be demonstrated. 59 Figure 5. 13 . Energy traces for selected nodes on the rectangular plate D. Networking Performance Characteristics In this section, we evaluate the networking characteristics on using Energy - Aware Pulse Switching on the spatiotemporally variable harvesting profiles discussed in Section 3. 60 Figure 5. 14 . Delivery Delay characteristics for selected source nodes on the rectangular plate In Fig 5.14 , the average delivery delay goes down exponentially as the energy availability in the system is enhanced. Fig 5.14 shows the average delivery delay for 5 representative nodes on the surface of the plate with anchor configurati on at the center. It can also be seen that the variability in the delivery delay increases with decrease in duty cycle. This is because with decrease in duty cycle, energy availability is more erratic , and the temporal variation becomes more pronounced in this case. As a result, even for the same source node for various node deliveries, event delivery delay can differ substantially. In Figs. 5.15 and 5.16 , we demonstrate the average harvested energy (over 100 seconds) and corresponding network delivery dela y characteristics for events originated at source nodes across the rectangular plate using a temperature map. 61 Figure 5. 15 . Average Harvested Energy and Event Reporting Delay distributions across rectangular plate with different edge anchor configurations As the color moves from blue to red (i.e. height of the surface increases), the value of the appropriate parameter being showcased increases. Figure 5.15 demonstrates the energy generation and delay characteristics for the anchor configurations along edge of the rectangular plate ( Figure 5.9 (a) - (c)) while Figure 5.16 demonstrates the same for the center anchor configuration (Fig 5.9 (d)). We also show the variation for separate duty cycles. As seen in Figure 5.15 (a), t he harvested energy generation is consistent over the whole area of the rectangular plate (based on spatial distribution function shown in Figure 5.10 (a)). As a result, the delivery delay characteristics are mainly modulated by the distance of the source nodes 62 from the sink. This is seen in Figure 5.15 (b), where the nodes near the right edge of the rectangular plate have much higher delays compared to those near the left edge because the latter are situated closer to the sink. In Figure 5.15 (c), we demonstrate a spatial variation in harvested energy which varies approximately in a linear fashion along the x coordinate of the wing (corresponding to the spatial distribution described in Figure 5.10 (b)). As a result, the delay characteristics a re also changed from the previous scenario. As can be seen in Figure 5.15 (d), at a lower duty cycle, when the energy is at a premium, the location of the nodes near the right edge of the plate allows for an energy benefit which even trumps the disadvantag e of being further away from the sink in distance. As a result, the delivery delay for events sourced at nodes near the right edge of the plate turns out to be lower than even those on the left edge which are closer to the sink. The latter nodes despite be ing closely situated to the sink, receive very low levels of energy being near the anchor of the plate which affects their event reporting delay. This effect is less pronounced when the energy availability is less constrained, and the delay characteristics resemble the characteristics in the constant spatial variation scenario. A similar trend is seen in Figure 5.15 (f), because the corresponding energy availability curve in Figure 5.15 (e), is even steeper in the x - coordinate direction which ensures that n odes on the right edge of the plate have delays even less than the previous case for the higher duty cycle scenario. In Figure 5.16 (a), we show the average harvested energy for the anchor configuration of Figure 5.9 (d), corresponding to spatial distribut ion function as in Figure 5.10 (d). We consider the event delivery delay characteristics across the node for three different Duty Cycles 6.3%, 25.1% and 100% as shown in Figure 5.16 (b), (c) and (d) respectively. 63 Figure 5. 16 . Distribution of Average Harvested Energy and Event Reporting Delay across rectangular plate with central anchor configuration It is noted that nodes near the edges of the plate have lower delays consistently compared to those near the central regions. With the increase in duty cycle, the overall energy availability increases and as a result, even the central regions of the plate deliver events with less delay. The dark red regions correspond to higher delays which are notably s een to get thinner with increasing energy availability. The dark red regions extend from the center towards the right edge of the wing in the lower duty cycle scenarios ( Figure 5.16 (b), (c)) because at these harvesting levels the distance of the nodes on the right edge of the plate from the sink also plays a role is modulating the delivery delay values. Overall, it is noted from the delivery delay characteristics that the network performance is a function of the energy availability and the Energy - Aware Pu lse Protocol 64 works based on the energy availability to modulate the transmission and lead to this behavior as we will further explore in the following sections. The increased delay allows for enough energy generation for event propagation instead of event losses and that is the benefit of the energy awareness in the Energy - Aware Pulse Protocol. Figure 5. 17 . Per hop transmission route diversity distributions for selected source nodes E. Route Diversity Distributions The Energy - aware Pulse Switching protocol helps to modulate the energy usage of individual nodes depending on their available energy levels based on energy harvesting. When energy is in abundance, the protocol assures that the nodes utilize maximum number of routes to ensure higher reliability which is represented by use of a higher Route Diversity. When the energy is scarce, nodes conserve their power and stagger the communication by storing generated events in the binary buffer until energy becomes availa ble. Even when energy is available, if it is just enough for transmission, the nodes select a lower route diversity sacrificing more reliable transmissions (with higher route diversity) keeping in mind the energy limitations. This is 65 demonstrated in the gr aphs of Figure 5.17 where the selected route diversity distributions have been shown for various levels of energy harvesting. When the duty cycle is high representing a faster rate of energy harvesting, the nodes have a higher percentage of transmissions w ith greater route diversity compared to lower duty cycle cases where the energy availability is limited. Figure 5. 18 . Buffer storage time distributions vs harvesting efficiency (Duty Cycle) F. Buffer Storage Time Distributions In Section 4.1.3, we have discussed how a Binary Event Buffer has been incorporated per node within the Energy - Aware Pulse Routing scheme in order to reduce event losses due to energy scarcity. The latter can effectively translate into reduc tion in overall event reporting delay because the reporting of the event does not need to rely on subsequent repeat transmissions. In Figure 5.18 , the distributions of the per - hop event buffering delay (averaged over all hops on an end - to - end route) have b een demonstrated. As is evident from the distribution graphs, the peak moves towards the lower end of the storage times when the vibration DC increases. Thus, when the energy availability is scarce (lower DC), the percentage of larger buffer storage times is 66 higher than when energy is gratuitous (higher DC). This is effectively a consequence of the energy - awareness in the pulse protocol coming into play. With energy at a premium, the protocol employs longer storage times to accrue enough energy for transmis sion of pulses needed for event forwarding. As energy becomes more abundant, the need for buffering is reduced and the per hop storage times gradually approach the lower limit of 1 frame delay which in this case corresponds to a time of 1.5 minutes as show n in Figure 5.18 . Thus, the observations validate the event buffering process and its proposed impacts. 5.3. Summary In this chapter, we have developed an energy - aware pulse networking protocol and show the advantages of using the energy - aware version ove r the baseline version of Pulse Networking in a vibration energy harvesting setting. We also provide evaluation of the energy - aware pulse protocol performance in a variety of harvesting scenarios which consider spatiotemporal variation of the energy harves ting availability. We conclude that the energy - aware protocol provides better performance over the baseline version as demonstrated in a wide variety of harvesting scenarios. In the next chapter, we will develop a Structural Health Monitoring framework uti lizing Energy - Aware Pulse Networking in a piezo - electric transducer - based through substrate communication scenario. We will also evaluate the performance of the developed protocol framework using realistic vibration traces from an airplane wing structure. 67 CHAPTER 6: THROUGH - SUBSTRATE ULTRASONIC PULSE NETWORKING In this chapter, we develop a through - substrate ultrasonic communication framework utilizing the energy - efficient discrete - pulse - based networking architecture detailed in the previous chapters. We also evaluate the efficacy of this networking architecture in a realistic Structural Health Monitoring (SHM) setting, namely event monitoring on an aircraft stabilizer structure. For the evaluation process, first, a realistic acceleration prof ile across an airplane stabilizer is developed using dynamic response simulation based on Finite Element Modeling. Thereafter, a simulated model of harvested energy is obtained from the spatiotemporal vibration profiles on the stabilizer. Finally, Energy - A ware Pulse Networking (as detailed in Chapter 5) simulation is performed on an array of nodes distributed over the stabilizer. In this process, we also consider pulse networking time frame adaptations relevant for ultrasonic through - substrate operation. U sing this evaluation framework, we demonstrate the performance advantages of the Energy - Aware Pulse Protocol in such an SHM scenario. 6.1. System Architecture The high - level system architecture envisioned for the representative SHM application of airplane stabilizer monitoring is shown in Figure 6.1 . The broad idea is to use a through - substrate sensor network as shown in the figure, where a collection of Through - Substrate Ultrasonic Pulse Networking (TUPN) units are mounted / embedded on the structure bein g monitored. The example in Figure 6.1 (a) and 6.1 (b) demonstrates how an event transportation network can be formed on an airplane stabilizer through the stabilizer substrate itself. Each TUPN is comprised of a piezo - electric transducer and a low - power c omputing module involved in sensing and networking control as shown in Figure 6.1 (e). In Figure 1(e), the computing module is a prototype, while in Figure 68 6.1 (f), we show a more recent work - in - progress integrated circuit (IC) chip with a smaller footprin t both in size and in energy usage. Figure 6. 1 . Event Monitoring using a Through - Substrate Sensor Network The advantage of using a piezo - electric transducer in this application is three - fold. First, the transducer can be used for sensing purposes because it reacts to vibrations and can be used to identify vibration patterns that indicate structural anomalies. Secondly, the piezo transducer can be used to harvest energy from the ambient structural vibrations which can be used to power the TUPN modules. Finally, the piezoelectric modules can be used for communication through the very substrate that they are embed ded on, thereby removing the need for any retrofitted components for communication such as radio transceivers. As shown in Figure 6.1 (c), neighboring TUPNs can form short ultrasonic communication links through the substrate (e.g., aluminum alloy or compos ite). A TUPN detected event (i.e., strain, fatigue etc.) results in an ultrasonic pulse, which is transported multi - hop to a data logger or sink node in such a manner that the latter can 69 indicate: 1) the very occurrence of the corresponding event, and 2) i ts location of origin with a pre - defined resolution. Resolution is based on a cellular abstraction ( Figure 6.1 (b)) in which the TUPNs are addressed not individually but based on the IDs of the cells in which they reside. Even with such limited information , several application level conclusions can be derived at the sink by correlating multiple event pulses [39] . 6.2. Prototype Ultrasonic Transceiver and Link Characterization A prototype TUPN, as shown in Figure 6.1 (e), was developed and used for characterizing pulse - based ultrasonic data links ( Figure 6.1 (d)) through metal substrates. Each TUPN can both transmit and receive using an ultrasonic piezoelectric attached. The Pul se Loss Rate (PLR) and False Positive Pulse Rate (FPPR) for communication using the above - mentioned transceiver through a 2024 Aluminum alloy plate (substrate) is listed in Table 6.1 . We consider the ultrasonic link performance over a variety of distances and two different source voltages. It is to be noted that the choice of substrate material (Al 2024) was motivated by the fact that this is the most prevalent material used in aircraft stabilizer construction [90] . The two source voltage levels (i.e. 3V, 6V) were chosen keeping in mind the different voltage requirements for transmission in two network roles source (forwarding) and sink (synchronization) as will be discussed in Chapter 5 for the pulse networking protocol. As shown in Table 6.1 , for the distances and voltage levels considered in the experiments, the ultrasonic signal propagation was found to be reliable with PLR/FPPR range s in the order of 10 - 6 . It is to be noted that the architecture is specifically designed for short links and an overall small network size. Hence, the choice of distance less than 1m for 3V source voltage experiments (pulse forwarding links) and up to 4m f or 6V source voltage experiments (frame synchronization 70 pulse links). At such distances, the ultrasonic SNR levels and signal shape are good enough for efficient and robust signal reconstruction at the receiver. Pulse Link Length (meters) Pulse Voltage (80 pulses) Pulse Loss Rate (PLR) False Positive Pulse Rate (FPPR) 0.5 3 2.16X10 - 6 4.02X10 - 6 0.75 3 2.89X10 - 6 5.77X10 - 6 1.00 3 3.02X10 - 6 7.21X10 - 6 4.00 6 1.62X10 - 6 3.4X10 - 6 Table 6. 1 . PLR/FPPR for ultrasonic communication over Al 2024 alloy plate using prototype TUPN module 6.3. Application and Network Model 6.3.1. Application Model TUPN units, as discussed above, are distributed across the structure to be monitored and form an event transportation network . We use an airplane horizontal stabilizer as the target structure in this work. Each TUPN is equipped with piezoelectric sensors for structural event sensing, and piezoelectric transceivers for generating ultrasonic pulses when events are detected in the locality. The TUPNs are powered by energy harvested from structural vibrations using a piezoelectric transducer. The TUPN network is used for transporting local event information to a centralized Access Point / Base Station ( Figure 6.1 (b)) where pulses re ceived from multiple such units across the structure are collated to make inferences about the overall health of the structure, which is the stabilizer in our case. 71 6.3.2. Network Model A cellular network abstraction is used for organizing the TUPNs dis tributed across the stabilizer. Localization is accomplished with the resolution of pre - defined regular hexagonal sensor - cells as shown in Figure 6.2 . Each cell represents an event area with a unique Cell - ID. Since spatial localization resolution is at the cell level, shrinking the sensor cell size can increase the resolution. Each sensor node (i.e., a TUPN) belongs to one of these event areas (cells) and is pre - programmed with the Cell - ID of its own cell and those of its geographical neighbors. Although the cells in Figure 6.2 are shown to be hexagonal, there are no specific architectural requirements in terms of their symmetry, shape, and size. Due to the cellular abstraction, the sensors are not individually addressed, and therefore no per - sensor addressing is necessary at the MAC or routing layers. Figure 6. 2 . Network Model on Airplane Stabilizer Structure 6.3.3. Performance Needs Event reporting delay from sensor to base station is considered the primary performance index for this architecture. Delay in this architecture depends on the route lengths and energy 72 availability in different parts of the target structure. As explained la ter (Section IX), pulse losses on the structure also translate into additional delay. In traditional networks, the delay typically ranges from few milliseconds to 100s of milliseconds. However, in a through - substrate network that runs on vibration - harvest ed energy, the expected delay can be much larger - in the range of few minutes to 100s of minutes. Such large delay can be acceptable for niche applications such as aircraft structure monitoring on a per - flight basis. For example, non - critical sensed event s (e.g., anomalous stress pattern in certain parts of the structure such as an aircraft wing or stabilizer) during a flight can be reported to an access point while the aircraft is in flight. After landing, the collected data can be downloaded from the acc ess point, thus avoiding the need for a ground - based structure inspection after each flight. For such an application, the reporting delay needs only to be smaller than typical flight duration, which can be up to tens to hundreds of minutes. 6.4. Structural Vibration Model This section summarizes dynamic response simulation of an aircraft stabilizer for generating spatiotemporal vibration (i.e., acceleration) data. This data is then used for modeling energy harvesting for the target sensor network. The simu lations were performed using a finite element method and the model was based on the geometry of the Boeing 737 horizontal stabilizer. Extracted results consisted of acceleration time histories at the TUPN node locations along the stabilizer surface. Struc turally, a typical aircraft stabilizer consists of an internal framework of stringers (for bending resistance), and diaphragms (for shear resistance and load distribution) enclosed by a continuous surface, or skin (for torsional resistance and aerodynamic lift). The lift pressures 73 generated from the wind flow are primarily controlled by the airfoil shape and the skin; while the structural behavior (i.e., dynamic response) is primarily controlled by the internal framework. The finite element simulations wer e conducted with the program Abaqus [91], [92] . The geometry for the model was simplified to represent the mai n structural components for the stabilizer element and obtain a realistic dynamic response from the simulation. The airfoil geometry of the stabilizer was ignored. Figure 6. 3 . (a) 3D model of stabilizer, (b) Accl n. of stabilizer based on Finite Element analysis A view of the model is shown in Figure 6.3 (a). All parts were modeled with continuum - type shell elements with uniform thickness. Additional stiffness from the stringers was modeled by placing spars (beam ele ments) on the top edges of the leading and back stiffeners. Thickness of the shell elements including the box structure and the stiffeners was 5 mm. This thickness value was determined to obtain realistic dynamic properties due to additional mass from non - structural components in the stabilizer. The material properties assigned to the model were those of Aluminum 2024 [93] with an elastic modulus of 73.1 GPa, a Poisson ratio of 0.33 and density of 74 2780 kg/m 3 [93] . Figure 6.3 (b) shows the i nstantaneous acceleration available at different parts of the stabilizer structure at a particular time, in this case the 7600 th second after start of simulation . Figure 6. 4 . Total pressure profile and simplified triangular pressure profile vs. chord length (aircraft speed 800 km/h) Frequency analyses were conducted to obtain vibration properties and a dynamic analysis was performed to determine spatiotemporal acceleration profiles under simulated cruising conditions for an aircraft. The aero - dynamic loading pressure on the stabilizer was estimated using FoilSim III [94] for a speed of 400 km/s and extrapolated to 800 km/s. The lift pressure distribution across the stabilizer chord length is illustrated in Figure 6.4 . The loading pressure was then applied to the model as an incremental ramp with a noise perturbation of 10% over a time domain of 7600 seconds. It should be noted that the simulated demand neglected the rigid - body flight dynamics of the plane and thus the s imulation captures only the relative response of the stabilizer element. A surface contour plot of the average acceleration at the bottom surface nodes is shown in Figure 75 6.5. As expected, nodes at the tip experience greater acceleration on average than th e nodes at the mid span of the stabilizer. Figure 6. 5 . Average Acceleration based on node coordinates (at aircraft speed of 800 Km/h) 6.5. Energy Harvesting Model An energy harvesting simulation model was developed for translating the above spatiotemporal acceleration data into energy generation across the target stabilizer structure. A bridge rectifier based piezoelectric energy harvester circuit model as shown in Figure 6.6 is used. The piezoelectric module on the harvester is used to transform the ambient vibration in the structure to electric power, which is subsequently rectified and stored in an attached super - capacitor for driving the TUPN module operation. I n Figure 6.6 , the left part represents the mechanical equivalent circuit for a single degree of freedom piezoelectric energy generator. The resistor R m represents the damping constant, K p represents the spring constant, and M represents the equivalent mass all for the piezoelectric device. The transformer with gain ratio is used for modeling the transfer of mechanical energy 76 to electrical energy. The parameter represents the electromechanical coupling constant, which is an inherent physical property of the piezoelectric device material. The parameter x represents the displacement of the piezo generator device (i.e., from its mean position) because of the vibration applied on the structure to which the device is attached. For a device mass of M , the g enerated force can then be written as , where is the generated acceleration. Figure 6. 6 . Piezoelectric Harvester Circuit Model The right part of Figure 6.6 represents the equivalent circuit for generated electrical energy regulation and storage. A bridge rectifier is used to rectify and store the generated energy in a storage super - capacitor with capacitance C storage . Another capacitor with capacitance C piezo represents the internal capacitance of the piezoelectric device itself. The TUPN sensor nodes are to be connected across the storage capacitor as the load. 6.1 ) 6.2 ) 6.3 ) 77 (6.4 ) The equations 6.1 6.4 as shown above m odel the circuit in Figure 6.6 . The differential equations have been formulated in terms of the four variables v 0 , v 1 , v 2 , v 3 . These variables capture the displacement (i.e., x ), its first derivative ( ), the voltage across the piezo capacitor (i.e. ), and the voltage across the storage super - capacitor (i.e. ) respectively. The parameters and represent the accelerations on the piezo module at consecutive discrete sim ulation time instants and respectively. The quantity expresses the interpolated acceleration at any time instant t. The equations have been obtained using the equations of motion [43] of the piezoelectric element and the closed loop circuit equations. By solving these differential equations simultaneously for a specific input acceleration profile (i.e., and ), it is possible to generate the corresponding harvested energy profile . The efficiency of harvesting is sensitive to many parameters including the electromechanical coupling co - oelectric material, as well as its dimensions. The amount of energy that can be stored in the super capacitor and the rate of energy accumulation depend on the capacitance of the storage capacitor. These parameters and their effects on the network performa nce during event reporting across the stabilizer structure will be presented in S ection 6.7. 6.6. Integrated Evaluation Architecture An end - to - end software simulation architecture was developed for evaluating network performance in the presence of models for structural vibration and the resulting harvested energy. As shown in the right column of Figure 6.7 , a set of sensor nodes (i.e., the TUPNs) are distributed 78 over a target structure of airplane stabilizer to form a cellular pulse network. Each node ex ecutes the energy - aware pulse protocol as outlined in Chapter 5 . The left column shows how spatiotemporal acceleration profiles (Section 6.4) are used for generating harvested energy using the model presented in Section 6.5. Based on the specific placement of a TUPN node, each sensor node can make use of a certain time - varying energy profile that is applicable for its specific location on the stabilizer structure. Using their specific time - varying energy input, as they execute the energy - aware pulse protoco l, the reserve energy in the storage capacitor (see circuit in Figure 6.6 ) cycles with time. An example of such cycling for a specific TUPN sensor node is shown in the bottom middle part of Figure 6.7 . Figure 6. 7 . Architecture of integrated evaluation software 79 End - to - end event transportation performance is determined by the structure vibration profile, event generation profile, and the network performance. The integrated evaluation platf orm integrates all these components within a single software module that was written in C++. 6.7. Performance Results 6.7.1. Network, Energy and Event Generation Model The simulated network consists of 1057 TUPN sensor nodes evenly distributed on the described airplane stabilizer structure with a sink node placed at the center left corner as shown in Figure 6.8 . The nodes are grouped into 56 regular hexagonal cells (as i n Figure 6. 2 ) with an average of 3 nodes per cell. The spacing between individual nodes is around 0.3m and the transmission range is kept at 0.75 m. Height of each hexagonal cell is 0.5m with cell side length approximately 0.288 m. The entire simulated str ucture is approximately 6m x 3m x 0.005m. Figure 6. 8 . Pulse network mapping on a target aircraft stabilizer structure Energy needed for transmitting a single pulse is chosen to be 1 mJ. This transmission budget has been chosen in line with the estimates for ultrasonic message transmission budgets as reported in [8] . The power consumption for reception as well as idling operation is chosen to be 1 80 see datasheets for ATTiny25/45/85 at [95] [97] ). H igher values for these budgets (as compared to the references) are chosen to demonstrate the worst - case event reporting delays delivered by the proposed architecture. Event generation locations are uniformly scattered across the stabilizer. Single events a re generated at different source locations across the stabilizer and their reporting characteristics are studied. Spatial and temporal energy generation is controlled using the harvested model described in the previous sections. 6.7.2. Network Node Energy Traces Figure 6. 9 . Acceleration and harvested energy profile at chosen TUPNs Figure 6.9 shows the applied acceleration, generated from structure vibration, at four different nodes whose locations on the stabilizer are shown in Figure 6.8 . A title {NID i , EA j } represents the corresponding node - id and event area (cell ID) that the node belong s to. In addition, 81 absence of any expenditure due to network activities. In particular, for one sample source node NID 51, the figure also demonstrates the variatio n in energy generation rate as a function of the It should be observed the node situated near the outer tip of the stabilizer structure (i.e., NID 58 EA 55) experiences the highest level of vibration, and hence the fastest harvesting out of all other selected nodes. The node closest to the anchor of the stabilizer (i.e., NID 35 EA 11) experiences the least amount of vibration, and hence the lowest rate of energy build - up. The energy harvesting rates of the other two selected nodes are somewhere in between. The energy build - up for node {NID 51 EA 33} is shown for three different electromechanical coupling constants (i.e., its resulting storage capacitor charging rate. The harvesting rate is important since the performance of the event transportation network protocol does depend on such rates across the structure as will be discussed in the following subsections. 6.7.3. Event Reporting Performance The performance of multi - hop event reporting from a source sensor node (i.e., TUPN) to the sink node is characterized in terms of the reporting delay. Such delay is characterized with varying: 1) structure vibration inte nsity, 2) per - node storage capacitor size, and 3) electromechanical coupling efficiency of the energy harvesting piezoelectric material. Vibration intensity is controlled by a scalar acceleration multiplier, which is used for scaling the raw acceleration values generated by the finite element model. The maximum node accelerations are varied in the range 10 m/sec 2 to 20 m/sec 2 (i.e., approximately from g to 2g, where g is the acceleration due to gravity). Such a range for the acceleration values has been ch osen 82 based on data from prior studies with representative unmanned aerial vehicles [98] which indicate that the accelerations range between g and g [98] . Three acceleration multiplier values have been used to cr eate such variation viz. 10, 15 and 20. These correspond to maximum node acceleration values of 10.4652 m/sec 2 , 15.6978 m/sec 2 and 20.9304 m/sec 2 respectively. The corresponding average node acceleration values across the stabilizer are 1.0124 m/sec 2 , 1.51 86 m/sec 2 and 2.0248 m/sec 2 respectively. to - efficient s. I IV), a physical property of the piezoelectric generator material and its dimensions, is changed in the approximate range of 1.96*10 - 5 Coulomb/meter to 1.96*10 - 4 Coulomb/meter. This is used for varying the available energy - harvesting rate. This maxim um value for the co - - 4 ) was calculated using Equation 6.5 [99] (6.5 ) Here we assumed a PZT ceramic type piezo material with the following characteristic parameter values [99] = 320x10 - 12 Coulomb/Newton = 5.0x10 10 Newton/meter 2 = 6.2x10 10 Newton/meter 2 It is to be noted that , and are inherent properties of the piezo material) [99] . The dimensions of transducer were chosen as follows 83 Length (l) - 31.7 mm Breadth (b) - 16 mm Thickness (t) - 0.0028 mm Figure 6.10 shows the delivery delay experienced by events (e.g., stress anomaly in the structure) detected at four different event areas of the aircraft stabilizer structure. The exact locations of origin of those events and the sink node are shown in Figure 6.8 . Th e four event areas are subject to different vibration levels (at the same instant and across time) and correspondingly different levels of energy harvesting. Such variations in energy availability in these areas, and along the route to the sink, cause diff erent event reporting delay for the pulse protocol. Figure 6. 10 . Event reporting delay with different vibration intensities Delays for those four areas are evaluated with three different vibration intensities, which are indicate d by the corresponding average acceleration values in Figure 6.10 . As expected with higher vibration intensities, the events from all four areas are reported to the sink with lower delays, mainly due to higher energy availability and lower event buffering at the intermediate nodes on 84 the route. It should also be observe d that the relative patterns for the delay experienced by the events from all four areas are similar across all vibration levels. Figure 6.10 shows that with intense structure vibration, the reporting delay on the target stabilizer structure can be as low as a few minutes. With very low vibration, however, the delay can be as large as hundreds of minutes. High event reporting delays (i.e., minutes) can be acceptable for non - critical monitoring applications as outlined in the Performance Needs subsection in Section 6.3.3. Figure 6. 11 . Event reporting delay with different energy storage capacity Figure 6.11 depicts the dependence of event delivery delay on the available storage for harvested energy at individual nodes. As the super capacitor capacitance ( C storage ) (as shown in Figure 6.6 ) is increased, the net energy availability per node increases in a give n time because for the same electrical voltage available, the higher capacitance can store a larger amount of electrical charge and thus energy. Thus, the event - buffering phenomenon (as introduced in the protocol in Chapter 5) occurs less frequently. This helps in reducing the overall end - to - end event reporting delay. This effect is evident in the delay for all four sample sensor nodes shown in Figure 6.11 . 85 The rate of energy harvesting can be also increased by using a piezoelectric material with higher Ele 6.1 - 6.4 ). Figure 6.12 depicts the delay improvement as a function of that parameter. As observed in Figs. 6.10 and 6.11 , the relative delay patterns for the sensor nodes 35, 51, 256, and 58 average. Figure 6. 12 . Reporting delay for different electromechanical coupling It is interesting to note that in all the three Figs. 6.10 , 6.11 and 6.12 , the nodes situated near the tip of the stabilizer and the edge e.g. 58, 55 enjoy less delivery delay compared to nodes situated near the heart of the stabilizer farther from the tip e.g. 256, 35. Such spatial variation in performance is explored in detail in the following results. Spatial distribution of energy availability and event reporting delay are shown in Figs. 6.13 and 6.14 respectively . The average harvested energy over a 10 0 second period at different parts of the stabilizer structure is shown in Figure 6.13 . The energy values are shown for different electromechanical coupling coefficient (i.e., ). It can be observed that for all , the available 86 energy is higher at locatio ns farther from the anchor of the stabilizer. As expected, with higher , more energy is harvested, causing a more visible gradient from the anchor to the tip. Figure 6. 13 . Spatial distribution of harvested energy Figure 6. 14 . Spatial distribution of event reporting delay Figure 6.14 depicts event reporting delay in the presence of the energy profile shown in Figure 6.13 . The delay value in the temperature map at a given location represents the delay experienced by an event that is generated in that specific location on the stabilizer. The dark brown color near the anchor represents high delivery delay due to very low ene rgy availability in that region. The blue color near the stabilizer tip indicates lower delay, owing to higher energy 87 availability in those regions as shown in Figure 6.13 . It can be observed that with higher electro - mechanical coupling constant , the ove rall delay reduces across the whole structure due to higher harvesting efficiency. A maximum allowable event reporting delay of 300 minutes was set for all the experiments corresponding to the results in Figure 6.14 . This maximum allowable delay explains the sharp transition in the delay values near the middle part of the structure. Depending on the specific from the anchor of the stabilizer. Beyond that distance, the delay starts decreasing, thus causing the multicolor bands somewhere in the middle of the structure. Beyond that band area towards the tip, where the harvested energy is plentiful, the delay becomes very small as indicated by the blue color. It is to be noticed that the band moved towards the anchor with increasing energy availability Figure 6. 15 - 4 ) To better understand the delay bands observed in the top view in Figure 6.14 , we present a lateral view of the spatial delay profile in Figure 6.15 , when is set to 1.96*10 - 4 . It can be seen in Figure 6.15 that the delivery delay falls quickly as the event source is moved toward the tip of the 88 stabilizer. This sudden fall in the delay causes the multi - color delay band to be quite narrow in the top - view temperature maps in Figure 6.14 . 6.7.4. Impacts of Adaptive Route Diversity As an energy - aware routing syntax, adaptive route diversity is explored for delay reduction, especially when extra energy is available. Figure 6.16 depicts the effects of available energy on the effective route diversity for node (NID 58 EA 55), which is near the tip of the stabilizer. The figure shows that for the largest coupling constant (i.e., 1.96*10 - 4 ), when the harvested energy along the stabilizer is the highest, a large fraction of the transmissions uses a route diversity of 2. This is an attempt to reduce the event delivery delay by exploiting alternate routes, even though they consume higher energy which is abundant in this scenario. Figure 6. 16 . Transmission route diversity with varying coupling constant is 1.96*10 - 5 , almost all the transmissions refrain from using multiple diversity (i.e., they use route 89 diversity of one). Note that for all experime nts corresponding to Figure 6.16 , a maximum allowable route diversity of 2 was chosen. Figure 6. 17 . Route diversity with varying storage capacitance Figure 6. 18 . Route diversity for different average acceleration levels The impacts of other energy - availability factors, namely, super capacitor storage capacitance and structure vibration intensity (i.e., the maximum acceleration) on route diversity 90 are shown in Figs. 6.17 and 6.18 respectively. The same trends (i.e., as observed in Figure 6.16 ) of high effective route diversities for high - energy situations can be observed in both Figs. 6.17 and 6.18 . High storage capacitance and large acceleration values create such high - energy situ ations. 6.7.5. Impacts of Error In the absence of pulse loss, the energy - aware routing architecture ensures guaranteed event reporting even when the delay is increased due to energy unavailability along the route. However, when pulses are lost due to nois e and detection errors, such guarantees cannot be made. To address that, a repetitive pulse generation model is adopted in which upon event detection, the source node starts generating pulses with a fixed periodicity (e.g., two frames) for indefinite perio d. This way, even when a pulse is lost, one of the subsequent pulses from the same event eventually gets delivered to the sink, causing an effective event delivery. The tradeoff of this pulse generation model is that the temporal resolution of event report ing is lost. Event reporting delay under this model depends on the pulse generation periodicity and the pulse loss rate (PLR). Event reporting delays from two different sensors to the sink on the aircraft stabilizer structure is shown in Figure 6.19. In both scenarios, the acceleration multiplier was set to 20 (i.e. an average acceleration of 2.0248 m/sec 2 ), and the energy storage capacity at the super - capacitor - efficie nt The first observation is that with increasing rates of pulse loss, the delay increases or stays constant. This is because after a pulse is lost, the event can only be reported when a later pulse corresponding to the s ame event makes its way to the sink node. In some cases, such as for node 256 (where vibration intensities are low), even with an increase in PLR, the delivery delay does 91 not increase because the event buffering effect due to energy constraint overshadows the effect of delay due to pulse losses. Figure 6. 19 . Impacts of pulse loss on event reporting delay The second observation is that for a given , the delay for events from the sensor NID 58 i.e., on tip of the stabilizer is lower than that for sensor NID 256, which is near the center of the stabilizer. This difference is mainly due to the highest energy availability near the tip, as demonstrated i n Figure 6.13 available harvested energy. Overall, the simulation results demonstrate that pulse losses affect the event reporting delay, which increases with an increase in PLR, but such increase can be minimal when there is already a significant amount of buffering delay in the system due to energy constraints. 92 6.8. Summary In this chapter of the thesis, we developed and evaluated a discrete - pulse - communication - based sensor net work architecture that uses through - substrate ultrasonic links and is powered by energy harvested from ambient structural vibrations. An airplane stabilizer is used as the target structure for both monitoring and harvesting purposes. Using an integrated si mulator, it was first shown that in the presence of vibration energy harvesting, reliable network performance in terms of event reporting delay could be accomplished by employing the proposed energy - aware protocol syntaxes. Simulation results are then used for analyzing the network performance sensitivity to key system parameters, namely, structural vibration intensity, energy harvesting efficiency of the used piezoelectric material, the energy storage capacity at the pulse switching sensor nodes, and pulse losses caused by ambient vibration noise present in the structure. In the following chapter, we develop a Spiking - Neuron - based learning architecture for event sequence pattern detection purposes particularly suited to pulse communication systems. We will also demonstrate how this can be combined with the low - power event sensing and networking architecture developed in Chapters 4, 5, 6 to create a holistic low - power framework for event monitoring and detection for Structural Health Monitoring applications. We will support this design using evaluation results showing how the architecture can be effective for low - power detection. 93 CHAPTER 7: DISTRIBUTED COGNITION USING NETWORKED PULSES AND SPIKING NEURONS The need for identification of spatiotemporal pattern s in distributed event occurrence sequences turns out to be a recurring theme in various popular application scenarios such as object tracking [14] , intruder detection [45] , structural health monitoring [39] , and environment monitoring [46] . Wireless sensor networks are well - suited to s uch applications for seamlessly capturing the event occurrence information and sending to a centralized base station to facilitate pattern detection and inference. In - situ detection at local cluster - heads in the network, when possible, would be more energy - efficient and facilitate faster response, instead of transporting the data all the way to the Base Station. However, sensing nodes are generally energy - constrained and thus in - situ detection is less practical. In this chapter, we demonstrate that using a single - layer Spiking Neuron architecture, we can efficiently detect the occurrence of such pre - defined event occurrence sequence patterns and with much less energy compared to traditional neural network - based approaches. This opens the possibility of in - place detection of such pattern sequences leading to less networking costs and quicker response. The proposed architecture can also easily be interfaced with various energy - efficient pulse networking architectures like those describe d in C hapters 4, 5, and 6 of this dissertation . It can thus be part of a complete discrete pulse - based sensing, networking and detection architecture aimed at optimal resource utilization. The latter can be very useful when employed in applications with sm all and inexpensive sensing devices or those powered by scarce energy harvesting and/or equipped with small batteries. Despite the simplicity of the architecture and inherent energy - efficiency, we demonstrate that high detection accuracy can be achieved, a nd the approach is well - generalizable over a reasonable range of event interval variation. The detection results are robust enough to handle practical amounts of pulse drift errors, and detection performance can be extended to cover pulse loss and false po sitive error scenarios 94 as well using appropriate modifications to the training process. In the current chapter, we provide details of the architecture, implementation and learning parameters as well as training methodology and the corresponding rationale f or their choice to make the case for use of Spiking Neurons in energy - efficient distributed cognition applications. 7.1. Introduction Applications in diverse domains such as habitat monitoring [46] , target tracking [14] , industrial process control [15] , structural health monitoring [39] etc. are known to require the identification / classification of spatiotemporal event occurrence patterns using distributed sensor measurements in order to make higher level inferences based on the same . Wireless sensor networks (WSNs) [47] often provide a seamless way to implement such solutions by facilitating low - cost distributed sensing and communication among sensing nodes. Application instances could range from moving object iden tification [14] , intrusion detection [45] to environmental change [6, 7] , structural anomaly detection [39] etc. A key requirement in such applications is that the sensing architecture be flexible enough to be able to identify a variety of event occurrence patterns i.e. if new patterns need to be detected, the architecture should be able to adapt to the new scenarios. Another aspect is the need for generalizability over a range of similar inputs i.e. detection robustness to minor changes in the same event pattern. Applications i n the field usually also assume some amount of energy - efficient / energy - aware operation because many modern distributed sensing architectures for such, aim to create cheap and maintenance - free operation by relying on small sensing devices with limited ene rgy storage but theoretically infinite energy generation capacity (harvesting from environmental sources) albeit at low / erratic rates (harvesting source unpredictability). 95 Neural network - based learning approaches have been shown to offer a good degree of accuracy when identifying / classifying spatiotemporal patterns in general and with a good degree of generalization robustness which can guarantee suitable operation in the field where conditions are highly dynamic even though underlying patterns might re main the same. However, perceptron - based neural networks [100] (also referred to as second - generation neural networks) which are the workhorse of such applications, generally consume a considerable amount of energy in terms of computation because good recognition is associated with the use of extensive number of neurons. Learning systems like th e actual human brain, however, are known to be much more efficient in terms of energy consumption for such computational tasks. This has motivated the development of the third generation of neural networks also referred to as Spiking Neural Networks which try to recreate the mammalian brain neuron functionality more faithfully than perceptron - based approaches. Spiking Neurons have been demonstrated to offer significantly higher computational capacity per neuron [80] and can thus, handle larger computational tasks at a lower energy budget due to lesser number of neurons being employed. This is mainly because Sp iking Neurons can recognize inputs in the form of spike trains instead of mere values as in the perceptron - based networks. Encoding inputs in spike trains opens a much larger set of possibilities in terms of temporal coding, rank coding etc. which the spik ing neurons can identify if suitably trained, all of which would be much more energy expensive to build / support using a perceptron - based architecture. In terms of hardware implementation as well, Spiking Neuron approaches can be more energy - optimized bec ause they can be implemented using simple highly energy - optimized analog circuitry, eschewing the need for complex digital circuitry and the associated energy burdens for perceptron - based approaches. 96 Figure 7. 1 . Application Overview In the current work, we demonstrate the use of a Spiking Neuron based architecture and associated learning / training mechanism, based on a learning model called the Tempotron rule [80] , to design an energy - efficient spatiotemporal pattern classification architecture for distributed networked sensing applicati ons. We envision a scenario as demonstrated in Figure 1, where a distributed set of event occurrences across a sensor network are available to a central node such as a Base station or cluster - head. The latter has a pre - trained Spiking Neuron which can look at the sequence of event occurrences and detect a known pattern to indicate cognition of an environmental situation that needs to be acted upon or logged for future analysis. As discussed already, Spiking Neurons support extreme energy - efficient operation as well as good detection robustness. Further, because Spiking neurons accept spike trains as inputs, we can adapt them for use with various discrete pulse - based networking approaches as described in [101] and [102] , which further helps energy - efficient operation. It has been shown in prior work [35] that discrete pulse - based communication approaches can be significantly more energy - efficient compared to traditional packet - based data communication approach es when the data content is small and 97 delivery delay constraints are relaxed. This is quite the case for the class of applications we mentioned earlier and, in particular, structural health monitoring, because the only information to be transferred are the time of event occurrence and its location (i.e. data content is small) and the event occurrence intervals are in most cases far larger than the data communication delays e.g. a person walking between sensors placed about 10 meters from each other can trig ger events in intervals of about 4 8 seconds (assuming normal walking speeds) while communication of such event information from source to sink can generally be achieved in much less time than that, often in the order of milliseconds, thus allowing event information delivery latencies much larger than usual in communication networks. The primary objectives of the work presented in the current chapter are as follows - 1) To develop an ultra - low - energy architectural solution for detection of spatial - temporal progression of events in WSNs, of spatiotemporal patterns of event occurrence across a chosen terrain (as in target trajectory detection, enviro nmental change detection, structural anomaly detection applications etc.) using event occurrence data transmitted from sensing nodes dispersed across the terrain. Such event data will be available only as discrete low - energy asynchronous spikes with event source id information encoded, 3) Develop reasonable learning mechanisms and training approaches such that the architecture can recognize a large variety of patterns as well as ensure robustness to reasonable pattern jitter. The key contributions are as fo llows 98 a) Mapping spiking neuron - based pattern detection (Tempotron learning rule) to a pulse - based wireless networking context, b) Use of a Spiking Neuron to learn unique temporal characteristics of expected pattern(s) c) Use of pulse networking schemes (Pulse Position Coded Pulses / Pulse Time Encoded Networking) for low - energy transmission of event occurrence information from source nodes to a central sink (single - hop / multi - hop) within constraints necessary for proper spiking neuron - based detection. I n the following section, we first discuss the general application and network models considered in this work before going into the details of the learning and networking architecture and demonstrating performance / evaluation results. 7.2. System Architect ure Before we delve into the specifics of the Spiking Neuron Based learning approach and the discrete - pulse based coding architecture, it is reasonable to discuss in some detail about the nature of applications where this architecture would be best - suited . We envision applications where the need is to infer some higher - level conclusions based on a distributed set of sensor measurements across a terrain. For example, we can imagine a geographical area of chosen size (referred to as Sensing Area ) with sensor s deployed uniformly throughout at regular intervals as shown in the green shaded area to the left of Figure 7. 2. Such sensors might be of diverse nature e.g. shadow detection [103] , sound detection [104] , temperature detection [105] , structural anomaly detection [106] etc. The important notion here is that the nature of sensing is not as important as the fact that an event can be detected e.g. threshold crossing of the sensed value. The more important task here is to understand the order in which the even ts were detected at different sensors across the Sensing 99 Area and how far apart. This can give us clues to understanding several higher - level events e.g. if a person has been moving along a specific trajectory multiple times, if an intruder has moved into a restricted zone within the terrain, if a forest fire is spreading too far, too fast, if a crack is developing over a particular area of a structure etc. Other applications can include detection of toxic gas diffusion in a controlled environment for appro priate safety measures, trajectory detection of a customer in a retail store to understand retail shopping preferences, stealthy motion detection across a terrain indicating an intruder etc. Figure 7. 2 . Spiking Neuron Learning in Pulse Communication Networks In the architecture discussed here, each occurrence of a notable event will trigger generation of a spike train at the corresponding time. The spike train can be composed of one or multiple spikes depending on the t ransmission protocol choice and should incorporate the most important information about the event occurrence such as event location as well as other information such as granularity of detection etc. as relevant. Figure 7.2 shows such spike trains 100 being gen erated across multiple sensors (see (A) in Figure 7.2 (shaded in blue)) in response to an event occurrence trajectory (see (X) in Figure 7.2 shaded in red). Such spike trains from multiple sensing devices will travel across the communicating media (see (B) in Figure 7.2 shaded in orange) before they are collated at central locations such as Base Stations or cluster - heads (see (E) in Figure 7.2 shaded in yellow). Here, an appropriate address decoder is used to extract the event source id information and the relevant spike trains are fed to the correct inputs of a prior - trained Spiking Neuron based on where they were generated. The Spiking Neuron would be able to analyze the input spike trains and indicate detection of one or more event sequence patterns with suitable robustness. The neuronal synaptic inputs (see (C) in Figure 7.2 ) receive spikes from different synaptic inputs (shown in different colors) and based on these, the Spiking Neuron either generates an output spike or not (see (D) in Figure 7.2 ) to in dicate detection of positive or negative patterns respectively. The panel on the lower right of Figure 7.2 outputs for different positive and negative patterns. It is to be noted that there is a baseline positive pattern and a ll patterns, which are slightly jittered versions of the baseline, would be regarded by the Spiking Neuron as positive patterns. Negative patterns can have totally different sequence of synaptic inputs or vastly large jitter levels even if like the baselin e pattern in terms of component inputs and this is indicated by the lack of an output spike generation by the Spiking Neuron. 7.2.1. Network Model It is prudent to discuss some of the assumed characteristics of the network model that would be best suited to the event pattern detection architecture that we discuss in this chapter. The sensing nodes will be equipped with low energy networking functions which can be used to transmit appropriate spike trains when sensing events are detected. 101 Figure 7. 3 . Network Topology In Figure 7.3 , we show a representative network of sensors in a square topology as is used for our simulation results discussed in Section 7.6 . The nodes have Manhattan connectivity i.e. each node can transmit only to neighboring nodes located within a grid distance of 10 meters. The nodes can encode event location information in spike train form and, at the receiving sink such information can be decoded correctly and efficiently to recreate the approximate event occurrence pattern. This can then be fed into the Spiking Neuron for detection of known sequences. The need for encoding individual node event information in spike trains arises because the nodes need to function in various multi - access channel scenarios where simultaneous transmissions (due to simultaneous even t occurrences) would cause collisions and garbled transmissions. The other motivation is to reduce energy cost of transmissions. As will be described in the pulse - based communication protocols in Section 7.4 , certain time - domain techniques can be used to e nsure communication with minimal collision while incorporating the essential event information in low - energy pulse trains. It is to be noted that such encoding and decoding is not necessarily error - free and can introduce drift errors into the recreated ev ent pattern sequence that is input to the Spiking 102 Neuron at the Base station. The choice of Spiking Neuron as the learning architecture is particularly in order to protect against such jitter scenarios and still produce robust detection results. In the cur rent work, we also consider only one - hop connectivity between source and sink as shown in Figure 7.3 . However, the detection architecture and associated networking approach proposed in this chapter can also be extended to multi - hop networks. 7.3. Spiking N euron Based Learning 7.3.1. Key Concepts The primary motivation for the use of a Spiking Neuron - based learning scheme for the kind of applications discussed in Section 7 [80] and easy amenability to synchronous / asynchronous spike - based input patterns (which is how the energy - efficiently transmitted data from sensi ng nodes looks like using different pulse communication protocols as discussed in Section 7.4 ). Spiking Neuron - based learning schemes also provide robustness in terms of reasonable jitter tolerance and the ability to learn a larger diversity of patterns (compared to ordinary perceptron - style learning schemes) with similar number of neurons [80] . Finally, there is no need for manual coding as the pattern detection requirement changes i.e. the spiking neuron just needs to be re - trained with the new expected scenarios instead of requiring to be logically programmed for each unique pattern, thus saving on some circuit implementation costs as the neuron itself can be designed as a simple analog circuit ins tead of flexible (less energy - efficient) digital design. 7.3.2. Neuron Description and Tempotron Learning Rule Since the spiking neural architecture is primarily inspired by the human brain, we use the characteristics of neurons in the brain in our archit ectural neurons as well. The Leaky Integrate 103 and Fire (LIF) Model [80] is one of the predominant neuron m odels in the literature which closely maps the operation of neurons in the mammalian brain. The LIF neuron essentially models a neuron as driven by exponentially decaying synaptic currents contributed by several input synapses. The synaptic currents drive the evolution of the neuronal voltage (integration) and when essentially a spike will be emitted which can be transmitted to other neurons connected to the ou tput. The sub - threshold membrane voltage of an exponentially decaying LIF neuron at any time t is given by 7. 1) where and, spike times for synaptic input , K is a causal filter vanishing for and are the decay constants for membrane and synaptic integration respectively. The neuron initially starts with a voltage and is then activated by several of its input synapses which try to raise the output voltage based on their relative importance indicated by the respective synaptic weights . When the neuron voltage crosses a threshold, say , the neuron emits a spike and the voltage returns to th e original state i.e. . The voltage is held at for a short period called the refractory period irrespective of any input contribution during this period. After the refractory period, the neuron is ready to again accept synaptic input c ontributions and fire accordingly. 104 Figure 7. 4 . Synaptic Voltage Evolution across Training Epochs Effect of Positive and Negative Training Patterns of the neuron to different input patterns. So, the idea is that by properly adjusting (learning) the appropriate weights of the synaptic inputs, the neuron can be t rained to fire only when certain spatiotemporal input spike patterns are detected. Figure 7.4 shows the response of an LIF neuron to different input patterns in terms of membrane voltage and output spike creation at different stages of training as will be discussed in more detail in Sections 7.5 , 7.6 . In the figure, the shaded portions show the change in voltage as different spikes (part of a pattern positive or negative) are input at the neuron synapses. As can be seen, the voltage pattern corresponding to similar input patterns seems to change over time and for the positive patterns, near the end of training, the voltage pattern is such that the threshold is crossed right after the last spike in the pattern and this results in an output spike. In case of the negative patterns, even at the end of training, the input spike patterns should not be able to elicit an output spike i.e. the membrane voltage needs to below 105 threshold as is the case shown in Figure 7.4 . Our intention is to come up with a training mechanism which can create such outcomes. Various methods have been cited in the literature to accomplish this kind of training, out of which the Tempotron rule stands out because it was one of the first and r obust enough to handle most of the application needs as mentioned in Section 7. 1 . The assumption of the Tempotron rule is that the neuron needs to be trained to distinguish between two sets of patterns i.e. a Tempotron neuron will be able to accomplish bin ary classification by firing a spike at the output when stimulated by a positive pattern spike train and, remaining quiescent when given a negative pattern spike train at the input. In order to achieve this, the Tempotron learning rule is applied over mult iple epochs of training when a defined error in the output is used to motivate changes in the contributing input synaptic weights, thereby reducing the errors over time in a gradient descent - style training approach. The Tempotron error is defined as follows For positive patterns, error = For negative patterns, error = , where is the maximum voltage achieved during an epoch of training and is the membrane threshold voltage . This works because for posit ive patterns, error is generated when the max voltage in an epoch is lower than the threshold i.e. spikes are not fired. On the other hand, for negative patterns, the error case occurs when the max voltage is higher than the threshold voltage i.e. a spike is fired, though it is not expected. Based on this error function, the Tempotron learning rule is defined as follows 106 For the positive error scenario i.e. if no output spike is elicited in response to a positive pattern, weights at synaptic input are i ncreased by , where = Time at which post - synaptic potential ( ) reaches its maximal value and, > 0 specifies the maximum size of the synaptic update per input spike (or can also be looked upon as the learning rate for training purposes). For the negative error scenario, i.e. if an output spike is elicited in response to a negative pattern, weights at synaptic input are decreased by the amount specified in Eq. ( 7.2 ). Eq. 7.2 shows how the time of occurrence of the maximum output voltage helps to determine the amount of training updates and how the kernel function helps to distribute the updates across the different synaptic inputs based on their contribution to the error. In particular, spikes which are far away from the max voltage time have little bearing on its value and thus weights for synaptic inputs corresponding to these are modified to a lesser extent than those which have input spikes very close to the max voltage time. The kernel function helps to modulate the rate at which the effect of a particular spike impacts voltage increments and then decay. 7.4. Baseline Pulse - based Networking A pproaches 7.4.1. Pulse Position Coding Protocol The Pulse Position Coding Protocol (PPCP) architecture, first explored in [102] aims to produce savings in terms of energy consumption by encoding any information to be conveyed in terms of the interval between consecutive spikes (discrete pulses). Depending on the value being encoded and the base of representation chosen, such inter vals can be quite large. Hence, as shown 107 in [102] , based on the energy consumption budget (transmission, reception, idling etc.) and the value to be encoded, we can choose an optimal base for representation of the value using multiple inte rvals (for each digit in the corresponding base representation) separated by discrete pulses or spikes. The trade - off is mainly in terms of energy vs delivery delay because the PPCP delivery latency depends on the interval between spikes and thus the value encoded and base of representation. If the number of spikes is increased by encoding the value in multiple small component intervals (digits), then delivery latency can be improved but at the cost of transmitting more spikes. On the other hand, larger int ervals with less spikes can provide better energy utilization but at the cost of delay. However, this architecture can be very useful when the range of values to be encoded is limited. In either case, this approach has been shown to use significantly less energy compared to traditional packet - based networking approaches [102] . Figure 7. 5 . PPCP PDU Spike Representation Figure 7. indicate a node id = 6. The protocol data unit (PDU) consists of a header portion constituted by a 108 pre - selected number of spikes at short intervals (called slots) to indicate start of data, a tail portion using one less pulse compared to header to indicate end of data and the actual value encoded within with different digits separated by single spikes. The number of spikes needed for the different delimiters e.g. header, tail, digit separator etc. depends on the choice of base and the max value to be supported. For example, if we need to support only values between 0 - 9 and the base is chosen to be 10, then we will not require multiple digits, thus no digit delimiter spikes. We c an choose two spikes for the header and 1 spike for the tail. This is the case shown in Figure 7. 5. As requirements get more involved though, the number of spikes comprising header and tail will need to be increased to still distinguish these from each oth er and the digit delimiter. Figure 7. 6 . Event Spikes vs PPCP Spikes The main idea here is that PPCP PDUs can be used to encode the essential information e.g. origin location of an event indicated by the corresponding sensing node id. This can be done using only three spikes (two for header, 1 for tail) if the value is limi ted in range and depending on the 109 base selection. Figure 7. 6 shows how an event pattern in time would be translated into the corresponding PPCP spike patterns that will be sent over the channel with good energy utilization. The PPCP PDUs can be decoded at pattern and the event spikes corresponding to the different nodes can be fed to the proper synaptic inputs of the pattern - detecting neuron in order. The Spiking Neuron can then fire on its output if the pattern closely matches one of the positive patterns. 7.4.2. Pulse Time Encoded Networking The PPCP protocol discussed in the previous section does not require any synchronization among the nodes for operation. However, in many applications, such synchronization can be achieved with the use of a central Base Station which can periodically broadcast sync messages and keep the network in sync. This is especially true for small network scenarios. In [35] as well as in Chapters 3, 4, 5, 6 in the current thesis, the authors discuss a pulse networking protocol that is well - suited to such scenarios and can use the synchronization to achieve data transmission with very small number of pulses per event. For exa mple, in the Pulse Networking protocol mentioned in [35] origin information can be encoded in the timing of the afore - mentioned pulse with respect to a synchronized time frame. Thus, using such a protocol, the number of pulses required is one - third of those required for PPCP transmissions in the best case. However, the trade - off here is the latency of such transmissions, because the synchronized time frame structure that is requ ired here imposes inherent latency bounds on transmission. Moreover, synchronization is assumed which might not always be a given. However, assuming latency restrictions are relaxed, and synchronization is guaranteed, synchronous pulse networking can be a very efficient networking and medium access control policy. In the current work, we use a simplified version of the Pulse Networking 110 architecture as shown in Figure 7.7 , with specific time slots assigned to every node for its event pulse transmission. This is also referred in the literature often as a Time Division Multiple Access scheme. Figure 7. 7 . Pulse Time Encoding Frame Structure Each node has a specific slot assigned for its transmission and depending on the number of slots for each node in the local vicinity. For example, Figure 7.7 sh ows the case for a network with a medium access scheme comprising a frame with M slots. When an event is detected, nodes will transmit their information usin g a pulse only in their respective slot as shown for nodes 4 and 2 in Figure 7.7 . Only one node can transmit in a specific frame in order to maintain event order resolution. 111 Figure 7. 8 . Event Spikes vs Pulse Time Encoding Spikes In Figure 7.8 , we show how this translates to our application scenario. Once the events are generated, the sensing nodes trigger spikes in their respective time slots and the resultant spike train is transmitted through the channel. The lat ter is like the original event generation spike train though slightly delayed because of the need to transport spikes from different nodes in their genera ted at the same node. This would make the transmitted pulse sequence somewhat different from the original event sequence pattern. However, when the number of nodes is limited and the slot size is small, the frame structure can be much shorter than the even t intervals involved. Thus, small discrepancies in the spike timing due to framing can be smoothed out by the Spiking Neuron detection. Essentially, such small changes can be modeled like pulse drift and as we will show in Section 7.6 , these can be handled very robustly using the Spiking Neuron architecture. 112 7.5. Adaptations of Spiking Neuron Learning for Pulse Networking 7.5.1. Networking Adaptations As mentioned in Section 7.2 , due to the need for a decoder at the receiver to read the event source id from the corresponding protocol specific PDU, the number of spikes that will be input to the Spiking Neuron at any synaptic input for a particular event will always be 1. What is m ore important is that the event intervals be sufficiently large compared to the protocol spike intervals or frame lengths. As mentioned in Section 7.2 , we envision applications where the event intervals are on the order of seconds e.g. person walking acros s a network area with sensors spaced at a resolution of 10 meters or so. Such event intervals are indeed much larger compared to the actual transmission time for individual spikes in a PPCP PDU or the length of a TDMA frame. The latter is a function of the slot size used for transmission which is the minimal spacing between spikes such that they can be unambiguously resolved at the receiver and is generally on the order of milliseconds or less. We define the event intervals as the parameter beta and use val ues of beta ~ (4.5, 7.5) seconds as predicated by our sample application (discussed in Section 7.2 ), while using slot sizes in the order of milliseconds which leaves our event sequences unaltered by the choice of networking architecture. 7.5.2. Membrane Ti me Constant Selection The membrane time constant and the synaptic integration time constants are two important parameters of the neuron model that decide which time scales the neuron will be working at and the temporal resolutions of patterns that the neuron can identify. If we have a general notion of the range for the event intervals that we are expecting for an application, then we can adjust the membrane and synaptic integration time constants accordingly such that the neuron is well - 113 equipped to han dle event patterns for that application. Synaptic and membrane time constants determine the level of contribution for spikes at different times to the membrane voltage of the neuron. Thus, with a larger synaptic integration time constant, spikes would take longer to influence the membrane voltage. On the other hand, larger membrane time constants make sure that the spike contributions to membrane voltage do not die out fast. The membrane time constant selection is important based on the application, because it determines the resolution of the spike patterns i.e. if the precise timing of each individual spike will decide the membrane voltage or a certain subset of this. In the case of our application, we need the neuron to adapt to the intervals in the positi ve pattern specified but also not be too tied to the precise spike timing because we want to ensure sufficient generalization of the detection mechanism to effects like pulse drift. In order to achieve this, we chose a membrane time constant which is on th e order of the event intervals chosen but slightly larger e.g. for event intervals in the range 4.5 7.5 seconds, we choose a membrane time constant around 16 seconds with a synaptic integration time constant much smaller than that (precisely 16 / 64 = 0. 25 seconds) such that spikes have nearly immediate effect on the membrane voltage but fall off slowly (larger membrane time constant) such that effect of multiple consecutive spikes can be accumulated to result in the final voltage and thus output spiking. 7.5.3. Training Methodology In our application scenario, event patterns can arise due to two scenarios the order of the nodes triggering events and the timing between such event occurrence. In order to represent the first scenario, we define something called a trajectory. A traject ory is a sequence of network nodes which comprise an event order. For example, if a man walks around the network and triggers events at nodes 1, 2, and 3 in order, then the trajectory he traversed would be 1 - 2 - 3. Given the network topology and connectivity , we can only have certain number of valid trajectories. In the 114 application scenario, we will be interested in certain fraction of these valid trajectories as positive trajectories while all others will be designated as negative trajectories. In order to t rain the neuron to identify the positive and negative patterns, the neuron needs to be exposed to representative positive and negative patterns over multiple epochs. In our application scenario, we envision having p valid positive trajectories and n valid negative trajectories. Each such trajectory can be used to generate a pattern by selecting the event intervals in the range corresponding to the application (controlled using the beta parameter). In order to keep the training uniform for positive and negat ive patterns, we expose the neuron to equal number of positive and negative pattern instances in every epoch. A collection of such epochs where all the positive and negative train patterns have been shown to the neuron is referred to as a batch. Every batc h training is then repeated until the positive and negative training error values stop improving based on a convergence criterion as will be mentioned in more detail in Section 7.6 . The training and testing process es work as follows 1) We design a set of positive patterns, one for each positive trajectory called the baseline positive patterns. During creation of the baseline patterns, we use various random event intervals in the specified range to provide generality. Similarly, a set of baseline negative patterns are also designed. For the negative baseline set, it is important to make an informed choice such that we can cover a large variety of negative patterns corresponding to all valid negative trajectories with a small number of negative baseline patt erns. We have tried various strategies for selection of the negative patterns, but it appears that if we choose only patterns corresponding to trajectories which are most like the positive trajectory, we can get best results with least number of negative p atterns. In order to calculate similarity, we use the relative percentage of similar spikes based on the gestalt pattern matching metric as available in the Python difflib helper called SequenceMatcher [107] . 115 We modified this slightly to make sure that when the second pattern contains the first pattern, then the similarity value is capped at 100%. 2) Based on the positive and negative baseline patterns, we create some positive t rain and negative train patterns by introducing jitter in a reasonable amount to the baseline patterns as expected for the application. The motivation here is to expose the neuron to various generalized versions of the positive and negative patterns during training such that it can be robust during actual operation. The jitter is mainly used to model the effect of pulse drift errors. The number of both positive and negative training patterns is chosen as the maximum of the number of positive and negative ba seline patterns. For both positive and negative cases, if the number of baseline patterns are less than the number of training patterns, some of the baseline patterns are repeated during training. 3) During training, the patterns in the training set are in put to the neuron. Every positive train pattern is followed by a negative input pattern. For each pattern as input, the neuron integrates the input spikes and generates a spike or not based on the membrane voltage evolution. If the neuron response is not a s expected (i.e. spike for positive pattern or no spike for negative pattern), an error is calculated based on the maximum voltage achieved by the neuron as stated in Section 7. 3 . Based on the error values, the input synaptic weights are updated to reflect the contribution to the error from multiple synaptic inputs. After the weights are updated, the process is repeated for the next training pattern and so on, until all training patterns are exhausted. This completes a batch of training. Training batches ar e repeated until the average error values over batch windows stop improving. 116 7.5.4. Test Methodology We generate test positive patterns based on the positive baseline patterns and by introducing various amounts of jitter but unrelated to the train test patterns are generalized versions of the baseline patterns, but different from the train patterns. For the negative patterns test, we decided to choose patterns based on all valid negative trajectories and we choos e multiple instances of patterns with each trajectory but with different beta values in the range specified. We refer to these as unknown patterns as the Spiking Neuron has not necessarily been shown any of these (though some negative trajectories are cove red in negative training) and we would like to train the neuron such that all or most of these unknown patterns can be detected. By varying both the trajectories as well as jitter, we aim to evaluate the generalization efficacy of the Spiking Neuron detect ion architecture. 7.6. Simulation Process and Performance Results 7.6.1. Simulation Environment and Process In order to simulate the Spiking Neuron - based learning architecture using a Tempotron - style rule, we created a network simulation and learning fra mework in Python. In the framework, we can create arbitrary network topologies and simulate spike - based pulse communication using PPCP and Pulse Time Encoding - based protocols (as detailed in earlier section 7.4 ). In order to simulate the Linear Integrate a nd Fire neuron characteristics (as discussed in Section 7.3 ) and the corresponding output behavior in response to input spike patterns, we utilize the python interface to the neural simulation tool NEST also known as PyNEST [108] . This program can efficiently simulate the behavior of various neuron models, but we chose the LIF neuron because it is simple to understand yet sophisticated enough for our learning demonstration. The neuron parameter 117 settings such as membrane time constant, synaptic integration time constant, refractory period, threshold v oltage etc. have been noted in Table 7.1 . Most of these parameters have been chosen to use standard values [80] , though we have adapted the membrane and synaptic integration time constants and spiking threshold v oltage based on our application timing requirements as explained in the Section 7.5 . We have also included various parameters related to the learning process in Table 7.1 . These include the epoch time which was chosen to incorporate the longest possible pa ttern trajectory (i.e. trajectory of length = no. of nodes) and the largest possible event intervals (i.e. 7500 ms between each pair of events). The number of batches that the simulation learning process can run for is also kept high (500) and a convergenc e criterion is defined as follows which allows for training termination. We realized that the training should be regarded as converged when the training error can no longer be appreciably improved upon over subsequent batches of training. Thus, we defined the training as converged when the relative average error change over a window of the last 5 batches falls below 1%. Table 7. 1 . Experimental Parameters Symbol Parameter Name Parameter Value N No. of synaptic inputs 9 n %. of negative trajectories chosen for negative pattern generation Variable 1, 2, 4, 8, 16. Event Interval Uniform (4.5, 7.5) seconds s Slot Size 50 ms Learning Rate Variable 0.01 (default) Threshold Voltage - 40 mV Reset Voltage - 70 mV E Epoch Length W Set of Synaptic Weights All initialized to 0.01 (small, identical, non - zero value) 118 Table 7.1 . Membrane Time Constant Variable - 16 secs (default) Synaptic Integration Time Constant Refractory Period 2 ms j Jitter (%) - % of beta used as jitter Variable 0, 4, 8, 16. Using the above - mentioned settings, we were able to train the Spiking Neuron with a variety of positive and negative patterns as discussed i n Section 7.5 and then evaluate the learning efficacy using a range of selected test patterns. We finalized on using small, non - zero and identical weights for all synaptic inputs to start the training in order to remove any initialization bias. We also not iced that this performs better compared to a random initialization of synaptic weights because negative patterns have less chance of spiking in the absence of initial bias. For our negative training patterns, we chose only the ones which were very similar to the positive pattern trajectory as mentioned in Section 7.5 . However, we made sure that the negative patterns which have trajectories with the positive pattern trajectory as a prefix are not part of the negative training set. This makes sense because, f or example, if we have a positive trajectory 1 - >2 - >5 - >6, we would not expect our neuron to properly classify 1 - >2 - >5 - >6 - >9 as a negative because the neuron has already seen the first part as a positive trajectory. Hence, we do not include such trajectories as part of negative training and the failure to identify these is the limit of the Spiking Neuron performance. We report various statistics mainly regarding the detection accuracy performance of the neuron in a variety of pattern scenarios. We cover some internal probes into the learning architecture as well and discuss how these demonstrate the learning mechanism working efficiently. We also aim to demonstrate through our findings that the Spiking Neuron architecture discussed here generalizes well for a variety of application and error scenarios underscoring the importance of this research work. 119 7.6.2. Effect of varying the event interval range (beta) In order to understand whether the current neuron architecture is robust enough to generalize its event sequence detection to a suitably large range of individual event intervals, simulations were performed with various event interval ranges, starting from a narrow one (4.5 to 4.75 seconds) up to a reasonably large one (4.5 to 7.5 seconds). The latter is lar ge enough to mimic event intervals in a target application e.g. detection of person walking (speeds ~ [3 5 Kmph]) o n a given trajectory inside the network area. As can be seen in Figs. 7.9 and 7.10 on the next page (for PPCP and TDMA cases respectively), the true positive rates and the unknown positives rates for all the interval scenarios are very similar when a moderately high value of n is used (e.g. n > 8% of all valid negative trajectories), which tells us that the platform is suitable for application across a wide variety of event interval ranges and should thus be applicable to a number of similar application scenarios. We also notice that there is minimal variability as a result of the networking protocol used i.e. PPCP or TDMA and th us the architecture should work well for either communication protocol scenarios. 120 Figure 7. 9 Figure 7. 10 121 7.6.3. Effect of varying positive pattern length In Figure 7.11 , we show the effect of positive pattern choice of different lengths (indicated by parameter pl ) on the detection accuracy performance. As is evident from the figures, the unknown accuracy results are best for the case of the positive patterns of largest l ength. This is intuitively expected of a well - working detection architecture, since the number of negative (unknown) patterns which can be similar to the positive patterns decrease as the positive pattern length increases. Figure 7. 11 . Effect of positive pattern length (pl) on detection accuracy across different negative training set sizes when using PPCP communication 122 Figure 7. 12 . Effect of positive pattern length on detection accuracy when using PPCP communication Figure 7. 13 . Effect of positive pattern length on detection accuracy when using TDMA communication As seen in Figs. 7.12 and 7.13 , the true positive results seem to decrease with increase in the trajectory length, but this is a trade - off with the unknown positive (which goes down). We can adjust the threshold voltage of the output neuron to ameliorate this situation as our applicati on prioritizes true positives vs unknown positives. The unknown positives rate is exactly the result we expected from our close negatives training process because the training should be able to adjust the weights to distinguish the positive pattern from al l but the ones very similar to it i.e. ones which 123 have the positive pattern as some subset. In particular, as evidenced by the data logs collected, in the case for highest n (% of negative trajectories used for training) for each positive trajectory length , the Spiking Neuron failed to identify only unknown patterns which were based on trajectories that had the positive pattern trajectory as a prefix. All other patterns were correctly flagged as negative. This is also evidenced from the pattern similarity v s unknown positives rate performance as shown in Figure 7.14 . Figure 7. 14 . Similarity of unknown pattern to positive pattern trajectory and its effect on detection accuracy for different positive pattern trajectory lengths - PPCP At low n values, some unknown patterns which have less similarity to the positive patterns are also mis - identified. However, with higher n, the percentage of such detections goes down and only a small percentage of unknown positives which ar e very similar (100%) to positive pattern are identified wrongly. Another observation is that the performance trajectories are similar for both PPCP and TDMA networking paradigms. 124 7.6.4. Effect of Pattern Type Figure 7. 15 . Detection Accuracy for Different Positive Pattern Types - PPCP Figure 7. 16 . Detection Accuracy for Different Positive Pattern Types TDMA In order to understand whether our detection accuracy results are generalizable over a large variety of patterns, we decided to investigate the accuracy results for different positive patterns of 125 the same length. The results shown in Figure 7.15 and 7.16 a re for different pattern types indicated by the value of pt = 1, 2, 3, 4, 5, corresponding to the positive pattern trajectories {1 - >2 - >5}, {6 - >9 - >8}, {7, - >4 - >5}, {3, - >6 - >9} and {2 - >5 - >8} respectively, each of length 3. As can be seen, the performance is la rgely similar across the pattern types in terms of both true positives and unknown positives and so is the case irrespective of the protocol used (PPCP / TDMA). 7.6.5. Effect of Spike Jitter In Figs. 7.17 and 7.18 ( see the next page) , we show the effect o f varying amounts of spike jitter on the detection performance of the Spiking Neuron. Spike Jitter as defined before is an amount of drift introduced in the train as well as test patterns in order to incorporate the practical implications of real - life puls e drift errors. Such pulse drifts can occur due to transmitter, receiver issues as well as energy constraints in the system which delay transmission / reception to conserve energy. Spike Jitter is incorporated as a relative proportion of the beta intervals used i.e. the event interval range. As can be seen in the results graphs, spike jitter up to a range of 16% seems to have little if any effect on the true positives as well as unknown positives rates for the Spiking Neuron. Only at the highest ji tter in this range is the True Positives Rate slightly affected while the Unknown Positives Rate remains largely unchanged. This indicates that the Spiking Neuron architecture described here should work well for varying amounts of realistic spike jitter en suring 126 Figure 7. 17 . Detection Accuracy for Different Pattern Jitter Levels - PPCP Figure 7. 18 . Detection Accuracy for Different Pattern Jitter Levels TDMA 127 7.6.6. Effects of Learning Rate Selection for Training In order to understand the effect of learning rate on the event detection performance, we consider the Spiking Neuron learning performance at three different learning rates from 0.01 to 1.0 as shown in Figure 7.19 . As is evident from the performance plots , a lower learning rate seems to provide better performance, though we see that below 0.05, the performance improvement is not appreciable, so using 0.05 can be good enough especially because convergence would faster in this case. In our current applicatio n domain, convergence time is not of prime importance, so we can afford to use an arbitrarily low learning rate. This is turn ensures that the training process is smooth and can go down the error gradient without much unnecessary oscillation and thus reach the proper optima (and not get stuck at locally optimal solutions). Figure 7. 19 . Detection Accuracy for Different Learning Rates 128 7.6.7. Effects of Membrane Time Constant Selection for Training As shown in Figure 7.20 , we have also investigated the effect of the neuron membrane constant choice on the detection performance. As can be seen, the performance is best when the membrane time constant is chosen at 16 seconds and falls off on either side for membrane time constants a n order of magnitude higher and lower , especially for the lower scenario. It is to be noted here that the membrane time constant in our experiments is tied to the synaptic integration time constant as well (which is a fraction of the membrane time constan t) and thus this parameter performance by noting that this membrane time constant is of the order of the event intervals chosen which are in the range 4.5 7.5 secon ds. Figure 7. 20 . Detection Accuracy for Different Membrane Time Constants 129 7.6.8. Training Error Analysis If we analyze the error plots, shown in Figure 7.21 , we can see that the error values initially fall sharply due to positive pattern training until positive error falls to a very low value. This is when close negative patterns start contributing to the negative error until some of the extraneous synaptic weights are reduced with the final positive and error values settl ing. Convergence criteria based on relative change of windowed error seems to identify convergence well. Absolute threshold - based convergence criteria might miss the negative pattern training completely giving sub - par results, while our final convergence c riteria based on relative change of the windowed average error stops the training much more gracefully and at a point where we would intuitively like to stop the training process. The appropriateness is also indicated by the superior detection accuracy res ults mainly in terms of unknown positives. Figure 7. 21 . Positive and Negative Pattern Training Error Evolution across Training Mini - Batches 130 7.6.9. Synaptic Weights Evolution Analysis In Figure 7.22 , we show the synaptic weights evolution for different synaptic inputs as the training proceeds from start to convergence. In the case illustrated in Figure 7.22 , the pattern trajectory is 1 - >2 - >5 and we would expect that the weights for these synaptic inp uts should go up while others should go down. This is exactly the case as we go toward convergence. Moreover, we see that as n is increased i.e. more negative patterns are used during training, the weights of synaptic inputs which are not part of the posit ive pattern are also changed, in this case on the negative side such that the patterns which are close to the positive pattern but not the same cannot trigger unwanted spikes. Figure 7. 22 . Synaptic Weights Evolution across Training Epochs for different sizes of negative train set 7.6.10. Effects of Pulse Loss and False Positive Errors In order to demonstrate good adaptability of the current learning approach to practical application scenarios, we also evaluated the performa nce of the network in various spike insertion and deletion error scenarios. The former scenario is often referred to as a Pulse Loss scenario 131 which can result due to certain communicated pulses of the event pattern not reaching the final sink / destination due to channel collisions, reception problems etc. The latter scenario i.e. spike insertion in the original pattern can occur due to channel noise, receiver abnormality etc. and is generally referred to as a False Positive Error scenario. These two kinds of errors are important to consider in the context of the Spiking Neuron performance because abrupt deletion or insertion of spikes has the possibility to create valid patterns which might be mis - identified by the Spiking Neuron. It is to be noted here tha t, errors in the system will not all necessarily reach all the way up to the Spiking Neuron and affect performance. Since there is a decoder at the sink module (see Section 7.2 ) right before the Spiking Neuron, certain more obvious abnormalities in the pul se data can be recognized and removed / remedied even before reaching the Spiking Neuron and thus the neuron performance would not be affected. However, in certain cases as allowed by the network topology, communication protocol semantics etc., pulse loss or false positive errors might not create invalid patterns that can be detected at the decoder and thus creep into the Spiking Neuron input. We want to demonstrate here that in such situations, we can adapt the Spiking Neuron to the error scenarios by appr opriately anticipating such errors and training the neuron to properly identify and classify these. This would be achieved with the trade - off of losing some granularity in negative pattern detection. It is to be noted here that the Pulse Loss Error as well as False Positive Pulse Error probabilities are generally very low, on the order of 10 - 5 [8] for the network architectures we are considering. Hence, we consider only single pulse insertion or deletion errors, assuming that two or more such errors within a singl e pattern would be much less probable (because the error rates get multiplied as the number of errors go up, each error being independent of others) and can be neglected for all practical purposes. For example, in the case of the positive pattern trajector y 1 - 2 - 132 5 - 6, we can assume pulse loss errors to create patterns such as 1 - 2 - 5, 2 - 5 - 6 (one pulse loss) etc. and not 1 - 2, 2 - 5 (two pulse losses) etc. False positive errors, on a similar token, would create patterns such as 4 - 1 - 2 - 5 - 6, 1 - 2 - 5 - 6 - 3 etc. In order to not mis - classify such error patterns as negative, we train with pattern instances of these trajectories as positive patterns while pattern instances of these trajectories are removed from the negative training scope. For each positive and negative traject ory chosen, we use multiple instances as before during training, such that the Spiking Neuron can generalize well for both event intervals and event occurrence sequence trajectories. Figure 7. 23 . Performance in the presence of Single Pulse Loss Errors - Positive and Unknown Pattern Detection Accuracy across different number of negative trajectories used for training As shown in Figs. 7.23 and 7.24 , the true positive pattern detection (output spike) as well as unknown negative pattern detection (no output spike) accuracy approaches close to ideal value (100%) for high n i.e. negative training with a reasonably large percentage (12% - 16%) of the total number of valid negative trajectories. The True Positives detection ac curacy is particularly 133 consistent for Pulse Loss errors across all n (see Figure 7.23 ), while in the False Positives case, there is more variability (see Figure 7.24 ), albeit in a small range (90 100%), but the highest n performance is ideal i.e. 100%. Unknown Positives Detection Accuracy gets progressively lower with increasing n for both Pulse Loss and False Positive Error scenarios with the final values at n = 16% close to the ideal scenario for the positive trajectory used. Thus, we can reasona bly conclude that with proper adjustments to the training procedure, we can adapt the Spiking Neuron architecture discussed here to effectively handle Pulse Loss and False Positive Error scenarios. Figure 7. 24 . Performance in the presence of False Positive Pulse Errors - Positive and Unknown Pattern Detection Accuracy across different number of negative trajectories used for training 7.7. Summary In this chapter, we demonstrate that using a single - layer Spiking Neuron architecture, we can efficiently and effectively detect the occurrence of pre - defined event occurrence sequence patterns which can be valuable in applications like Structural Health Monitoring. The prop osed 134 architecture can easily be interfaced with various pulse networking architectures (for energy - efficient transport) and operate with high detection accuracy generalized over a reasonable range of event interval variation. The detection results are robu st to decent amount of pulse drift errors and detection performance can be extended to cover pulse loss and false positive error scenarios using modified training pattern sets. We lay down details of the architecture implementation and learning parameters, training methodology and the corresponding rationale. Because of the simple architectural design but robust performance, this approach can be a good choice for resource - constrained applications. In the next chapter which will conclude this thesis, we will talk about future research work in this direction. Especially, we will discuss how we can intend to extend the application and network size and scope to demonstrate that the Spiking Neuron detection performance scales well irrespective of the size of the network. We will also outline plans to use layering of the spiking neurons to achieve even better performance when the problem complexity is increased. 135 CHAPTER 8: SUMMARY AND FUTURE WORK 8.1. Summary In this thesis, we have laid the foundations for the development of a holistic framework for applications like Structural Health Monitoring using maintenance - free wireless sensor networks. The proposed solution would be powered by ambient vibration energy harvesting to provide maintenance - free operation. Communication across the network will be based on energy - efficient through - substrate ultrasonic pulse - based communication. This would enable reliable networking performance and consistent network uptime despite the unpredictability of harvesting - powered o peration. Through - substrate links based on piezo - electric transducers would enable wire - free communication without the need for retro - fitted wireless radio infrastructure deployment. We also outlined a Spiking Neuron based low - complexity event pattern dete ction architecture. The latter would enable easy identification of structural anomaly patterns based on the spatiotemporal binary event information available from across the structure. The low - energy Spiking Neuron architecture can also enable some amount of in - network processing, even on intermediate energy - constrained network sensing modules, instead of delegating all processing to the Central Base Station. This can provide savings in terms of networking cost and faster detection and response. In the var ious chapters of this thesis, we have developed the different components of the final envisioned architecture. This includes scalable and energy - aware pulse - based networking, through - substrate pulse networking in energy - harvesting - powered systems as well a s design and evaluation of a single - layer Spiking Neuron based spatiotemporal event pattern detection architecture. 136 Future work on this broad topic can go along various routes. To start with, further research will need be carried out on the Spiking Neuron based detection architecture including an extensive evaluation of the same when used in more complicated (non - binary) pattern detection scenarios. Specifically, the weaknesses of the single - layer spiking neuron architecture will be scrutinized in such sce narios and improvements such as multi - layer designs will be considered and evaluated. The Spiking Neuron detection architecture can also be incorporated into a realistic Structural Health Monitoring application based on pulse networking, for instance, on a n airplane wing and performance of the complete system can be evaluated in terms of detection accuracy to establish the advantages of this architecture. Beyond this, further research can also look into developing energy - harvesting - awareness mechanisms with in the pulse networking framework (in addition to the energy - awareness syntaxes as discussed in prior chapters of this thesis) to make it even better suited to ambient harvesting - powered operation. In the following sections, we first discuss the applicatio n architecture envisioned in this thesis and then consider the future work proposals in some more detail. 8.2. Application Architecture In Figure 8.1 , we depict a high - level vision of the application architecture being proposed. As an outcome of this thes is, we envision a Structural Health Monitoring application based on through - substrate pulse communication and spiking neuron based low - power detection of anomalous spatiotemporal event sequence patterns. In an example application scenario of an airplane wi ng structure monitoring, the network of sensors would be deployed over the wing substrate as shown in Figure 8.1 (a). The distributed sensors will be grouped into a cellular abstraction (hexagonal cells shown here) with multiple sensors per cell for redund ancy. All event addressing will be on a cellular resolution. 137 Figure 8. 1 . Structural Health Monitoring Platform based on Pulse Communication and Spiking Neuron Based Detection It is to be noted that all the sensors will be equipped with ultrasonic through - substrate communication - enabled pulse communication transceivers and such modules would be referred to as Through - Substrate Ultrasonic Pulse Networking (TUPN) units. When an ev ent is detected based on local sensing, the corresponding TUPN unit would transmit a pulse to indicate occurrence of the event. This pulse would then be routed multi - hop along the network toward the Base Station over ultrasonic links based on the pulse net working semantics. The pulse networking protocol enables preservation of the source id information as well as next - hop routing information which enables the event information to reach from source to sink, that is the Base Station for further processing. As shown in the shaded red area on the bottom of the wing network, four cells (shown by the colored sensors) have event occurrences in a specific order as indicated by the arrows. This is 138 a sequence of events laterally across the wing and in a sequence fro m left to right, which might be indicative of a stress pattern of note. In order to detect such an event sequence pattern at the Base Station, the sequence of event pulses will be fed into a Spiking Neuron after appropriate address decoding to differentiat e the pulse sources. The Spiking Neuron will look at the event sequence pattern and based on its pre - trained synaptic input weights, process the input pattern and create an output spike when an event pattern of note is detected. We have already evaluated v arious components on this architecture in Chapters 4 7. In future work, we intend to incorporate all these elements into a real application such as airplane wing monitoring and evaluate the performance in various ambient energy availability and network a rchitecture scenarios. Using the performance analysis of such a study, we aim to establish the proposed architecture as a prime candidate for use in SHM applications. 8.3. Extending Single - Layer Spiking Neuron - based Event Pattern Detection In Chapter 7, w e have shown preliminary results on the detection performance of a single - layer Spiking Neuron based detection architecture for a simple application scenario and a limited network topology. Notably, we have considered detection of a single positive pattern vs multiple negative patterns. We have done quite some analysis on the effect of spike jitter to ensure robustness of the detection mechanism as well as how to adapt the learning when channel errors are appreciable. We have also shown that the detection a rchitecture is well adapted to work in consort with various energy - efficient discrete pulse - based networking mechanisms. In a more typical application though, we might have many such positive patterns which need to be detected by the same application. We h ave also noticed in some preliminary experiments that when the number of positive patterns increases, a single - layer Spiking Neuron architecture might be limited in terms of detection performance. Hence, future work can involve exploring multi - neuron and 139 m ulti - layer Spiking Neuron approaches and when they would be better suited compared to a single - layer single neuron architecture to ensure detection performance. The premise is that multiple neurons can share the detection load by handling different parts ( sub - patterns) of the chosen pattern. A progression of layers can be trained to look at higher level features with low granularity in the starting layers, requiring individual neurons in the starting layers to be less accurate and less computationally power ful. In such scenarios, important considerations would be how to design the multi - layer neuronal connectivity to achieve best detection performance with the least number of additional neurons. Work can also be done in applying other Spiking Neuron learning rules apart from the Tempotron learning mechanism as used here. Use of more sophisticated learning mechanisms might enable better performance with minimal number of neurons added (and thus energy expended). 8.4. Development of Energy - Harvesting Awareness in Pulse Networking In Chapter 5 - 6, we have developed energy - aware syntaxes within the discrete pulse - based networking framework. These enable the protocols to operate well in slow harvesting scenarios such as ambient energy harvesting. It is to be note d that such performance does not assume any knowledge of the energy harvesting availability. It only considers the energy availability in each availability is availab le, theoretically an improved utilization of the energy input can be achieved. Upcoming work can be aimed at tackling this aspect. Specifically, research attempts can be made to predict energy harvesting profiles based on machine learning approaches to ena ble a true energy - harvesting - aware pulse networking platform for optimal network energy utilization. If historical data on the harvesting availability profiles are available, various time series prediction mechanisms can be utilized for forward prediction and planning including traditional mechanisms like 140 exponential moving averages and Auto Regressive Integrated Moving Averages (ARIMA) as well more sophisticated mechanisms such as the one mentioned in [109] as well as Deep Neural Network - based mechanisms that utilize recurrent layers such as Long Short Term (LSTM) memory networks [110] . The idea is that being able to predict the harvesting availability will enable better planning of system resources and enable more efficient utilization with less wastage. 141 BIBLIOGRAPHY 142 BIBLIOGRAPHY [1] in Procedia Engineering , vol. 41, pp. 1204 1210, Jan. 2012. [2] in Proceedings of the IEEE , vol. 98, no. 11, pp. 1903 1917, Nov. 2010. [3] J. Zhao, R. Govindan, a Proceedings of the First IEEE International Workshop on Sensor Network Protocols and Applications, 2003. , 2003, pp. 139 148. [4] A. Mainwaring, D. Culler, J. Polastre, R. Szew Proceedings of the 1st ACM International Workshop on Wireless Sensor Networks and Applications , New York, NY, USA, 2002, pp. 88 97. [5] R. Szewczyk, A. Mainwaring, J. Polastre, J. Proceedings of the 2 n d International Conference on Embedded Networked Sensor Systems , New York, NY, USA, 2004, pp. 214 226. [6] Hanbiao Wang, J. Elson, L. Girod, D. 2003 IEEE International Conference on Acoustics, , 2003, vol. 4, pp. IV 844 - IV - 847 . [7] N. Xu et al. , Proceedings of the 2Nd International Conference on Embedded Networked Sensor Systems , New York, NY, USA, 2004, pp. 13 24. [8] - Substrate Event Reporting using Harveste d Energy in Proceedings of IEEE Globecom 2015 , San Diego, CA, 2015. [9] Proceedings of the 6th International Conference on Information Processing in Sensor Networks , New York, NY, USA, 2007, pp. 69 78. [10] S. Kim et al. in Proceedings of the 6th International Conference on Information Processing in Sensor Networks , New York, NY, USA, 2007, pp. 254 263. [11] A. Arora et al. twork for target detection, in Computer Networks , vol. 46, no. 5, pp. 605 634, Dec. 2004. 143 [12] Wei - in IEEE Transactions on Mobile Computing , vol. 3, no. 3, pp. 258 271, Jul. 2004. [13] - Quality Tradeoffs for Target Tracking Information Processing in Sensor Networks , 2003, pp. 3 2 46. [14] - in EURASIP J. Adv. Signal Process , vol. 2003, pp. 378 391, Jan. 2003. [15] vent Proceedings of the 31st International Conference on Very Large Data Bases , Trondheim, Norway, 2005, pp. 769 780. [16] Principles, in IEEE Transactions on Industrial Electronics , vol. 56, no. 10, pp. 4258 4265, Oct. 2009. [17] F. Salvadori et al. in IEEE Transac tions on Instrumentation and Measurement , vol. 58, no. 9, pp. 3104 3111, Sep. 2009. [18] in Computer Networks , vol. 52, no. 12, pp. 2292 2330, Aug. 2008. [19] V. Rajendran, K. Obraczka, and J. J. Garcia - Luna - - efficient, Collision - free in Wirel ess Netw orks , vol. 12, no. 1, pp. 63 78, Feb. 2006. [20] - resi lient, Energy - efficient in SIGMOBILE Mob. Comput. Commun. Rev. , vol. 5, no. 4, pp. 11 25, Oct. 2001. [21] in wireless s PCCC 2005. 24th IEEE International Performance, Computing, and Communications Conference, 2005. , pp. 535 540 , 2005 . [22] in IEEE Transactions on Communications , vo l. 22, no. 5, pp. 637 648, May 1974. [23] - efficient and low - 18th International Parallel and Distributed Processing Symposiu m, 2004. Proceedings. , 2004, pp. 224 - 235 . [24] in Second IEEE Workshop on Dependability and Security in Sensor Networks and Systems , 2006, pp. 54 64. 144 [25] R. C. Shah an in 2002 IEEE Wireless Communications and Networking Conference, 2002. WCNC2002 , 2002, vol. 1, pp. 350 355 vol.1. [26] - aware localized routing in in IEEE Transactions on Parallel and Distributed Systems , vol. 12, no. 11, pp. 1122 1133, Nov. 2001. [27] - efficient target coverage in Proceedings IEEE 24th Annual Joint Conference of the IEEE Computer and Communications Societies. , 2005, vol. 3, pp. 1976 1984 vol. 3. [28] S IEEE INFOCOM 2003. Twenty - second Annual Joint Conference of the IEEE Computer and Communications Societies (IEEE Cat. No.03CH37428) , 2003, vol. 3, pp. 1713 1723 vol. 3. [29] - efficient, Delay - aware, and Lifetime - in IEEE/ACM Trans. Netw. , vol. 23, no. 3, pp. 810 823, Jun. 2015. [30] - Efficient Data Collection for Large - in IEEE Internet of Things Journal , vol. 3, no. 4, pp. 511 519, Aug. 2016. [31] L. Lin, N. B. Shroff, - aware Routing for in IEEE/ACM Trans. Netw. , vol. 15, no. 5, pp. 1021 1034, Oct. 2007. [32] ributed energy harvesting aware routing algorithm for ad - hoc multi - World of Wireless Mobile and Multimedia Networks (WoWMoM), 2010 IEEE International Symposium on a , 2010, pp. 1 9. [33] M. Younis, M. Youssef, and K. Arisha - aware routing in cluster - based sensor Proceedings. 10th IEEE International Symposium on Modeling, Analysis and Simulation of Computer and Telecommunications Systems , 2002, pp. 129 136. [34] J. C. C. Restrepo, C. G. Gruber, and C. M. 2009 IEEE International Conference on Communications Workshops , 2009, pp. 1 5. [35] Sensing and Localization in Sensor Network Distributed Computing and Networking , D. Frey, M. Raynal, S. Sarkar, R. K. Shyamasundar, and P. Sinha, Eds. Springer Berlin Heidelberg, 2013, pp. 208 224. 145 [36] Technolo Proceedings of the 4th International Workshop on SHM , Stanford, CA, 2003, pp. 42 52. [37] F. - K. Chang, Structural Health Monitoring 2003: From Diagnostics & Prognostics to Structural Health Management , in Proceedings of the 4th Internati onal Workshop on Structural Health Monitoring, Stanford University, Stanford, CA, September 15 - 17, 2003. DEStech Publications, Inc, 2003. [38] in Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences , vol. 365, no. 1851, pp. 303 315, Feb. 2007. [39] H. Salehi, S. Das, S. Chakrabartty, S. Biswas, and R. Burgueño, Damage Identific ation Algorithms Using Binary Data , Proceedings of the ASME 2015 Conference on Smart Materials, Adaptive Structures and Intelligent Systems , vol. 2 , Integrated System Design and Implementation; Structural Health Monitoring; Bioinspired Smart Materials and Systems , pp. 1 - 10, Sep . 2015 . [40] in IEEE Transactions on Dependable and Secure Computing , vol. 14, no. 4, pp. 363 376, Jul. 2017. [41] 2011 IEEE/ACM International Conference on Cyber - Physical Systems (ICC PS) , 2011, pp. 206 214. [42] Proceedings of the IEEE Global Communications Conference (GLOBECOM) , 2012, pp. 379 384. [43] N. G. Elvin, N. Lajnef, and A. in Smart Mater ials and Struct ures , vol. 15, no. 4, p p . 977, 2006. [44] - Powered CMOS Sensor - Data - Logger With a 13.56 in IEEE Journal of Solid - State Circuits , vol. 47, no. 2, pp. 476 489, Feb. 2012. [45] in International Journal of Dist ributed Sensor Networks , vol. 9, no. 5, p. 167575, May 2013. [46] in Methods Inf Med , vol. 47, no. 1, pp. 70 75, 2008. [47] G. Werner - Allen et al. in IEEE Internet Computing , vol. 10, no. 2, pp. 18 25, Mar. 2006. 146 [48] - Based Event Detection for 3D Environment in IEEE Transactions on Knowledge and Data Engineering , vol. 20, no. 12, pp. 1699 1711, Dec. 2008. [49] Wireless Sensor Networks , 2008, pp. 90 105. [50] - hoc in Computer Science Technical Reports (CU - CS - 960 - 03) , University of Colorado, Boulder, 2003. [51] Z. Wang, E. Bulut, and B IEEE International Conference on Communications , 2008, pp. 306 310. [52] C. Fragouli, A. Orlitsky, É. Polytechnique, and F. Lausa nne, Silence is Golden and Time is Money : Power - Aware Communication for Sensor Networks in Proceedings of the 43rd Annual Allerton Conference on Communication, Control, and Computing , 2005, pp. 1026 - 1035 . [53] 2011 8th Annual IEEE Communications Society Conference on Sensor, Mesh and Ad Hoc Communications and Networks , 2011, pp. 197 205 . [54] - Less Protocol Paradigm in IEEE Transactions on Mobile Computing , vol. 12, no. 1, pp. 35 50, Jan. 2013. [55] G. Anastasi, M. Conti, M. Di Francesco, and A. Passarell in Ad Hoc Networks , vol. 7, no. 3, pp. 537 568, May 2009. [56] - efficient MAC Protocol for Wireless Proceedings of the 1st International Conference on Embedded Networked Sensor Systems , New York, NY, USA, 2003, pp. 171 180. [57] - 4th Annual IEEE Com munications Society , 2007, pp. 21 30. [58] Communication Paradigm for Sensor Networ Proceedings of the 6th Annual International Conference on Mobile Computing and Networking , New York, NY, USA, 2000, pp. 56 67. 147 [59] in ACM Trans. Embed. Comput. Syst. , vol. 6, no. 4, Sep. 2007. [60] Proceedings of the 4th International Symposium on Information Processing in Sensor Networks , Piscataway, NJ, USA, 2005. [61] - harvesting 7th International Conferenc e on Communications and Networking in China , 2012, pp. 742 747. [62] in Computer Communications , vol. 91 92, pp. 1 16, Oct. 2016. [63] S. in ICC 2001. IEEE International Conference on Communications. Conference Record (Cat. No.01CH37240) , 2001, vol. 2, pp. 472 476 vol.2. [64] M. Younis and K. Akkaya, Strategies and techniques for node placement in wireless sensor networks: A surve y, in Ad Hoc Networks , Jun. 2008, vol. 6, no. 4, pp. 621 - 655 . [65] for in IEEE Transactions on Mobile Computing , vol. 13, no. 9, pp. 1942 1954, Sep. 2014. [66] 2015 IEEE 16th International Symposium on A World of Wireless, Mobile and Multimedia Networks (WoWMoM) , 2015, pp. 1 9. [67] R. Figura et al. 2016 12th Annual Conferenc e on Wireless On - demand Network Systems and Services (WONS) , 2016, pp. 1 8. [68] in Structural Safety , vol. 33, no. 1, pp. 26 41, Jan. 2011. [69] Health Monitoring and Intelligent Transportation Systems for Bridge Condition Assessment: in IEEE Transacti ons on Intelligent Transportation Systems , vol. 17, no. 8, pp. 2107 2122, Aug. 2016. [70] - in Proceedings of the 4th International Worksho p on Civil Structural Health Monitoring , Berlin, Germany, 6 - 8 Nov . 2012 . 148 [71] J. P. Lynch, K. J. Loh, J. P. Lynch, and K. J. Loh, A Summary Review of Wireless Sensors and Sensor Networks for Structural Health Monitoring , in The Shock and Vibration Digest , vol. 38, no. 2, Mar. 2006, pp. 91 - 128 . [72] Park Gyuhae, Rosing Tajana, Todd Michael D., Farrar Charles R., and Hodgkiss William, in Journal of Infrastructure Systems , vol. 14, no. 1, pp. 64 79, Mar. 2008. [73] harvesting (WSN - HEAP) - 2009 1st International Conference on Wireless Communication, Vehicular Technology, Information Theory and Aerospace Electronic Systems Technology , 2009, pp. 1 5. [74] - in Smart Mater. Struct. , vol. 10, no. 2, pp. 293 299, Apr. 2001. [75 ] Optimization of a Solar Energy Harvester System for Self - Powered Wireless Sensor in IEEE Transactions on Industrial Electronics , vol. 55, no. 7, pp. 2759 2766, Jul. 2008. [76] energy harvesting technologies: a game - in IEEE Wireless Communications , vol. 14, no. 4, pp. 90 96, Aug. 2007. [77] - in IEEE/ACM Transactions on Networking , vol. 21, no. 4, pp. 1117 1130, Aug. 2013. [78] Z. A. Eu, H. Tan, and W. K. G. Seah, in wireless sensor networks powered by ambient energy harvesting , in Computer Networks , vol. 54, no. 17, pp. 2943 - 2966, Dec. 2010 . [79] - body networking: Interference modeling, stochastic chann 2015 IEEE Conference on Computer Communications (INFOCOM) , 2015, pp. 2425 2433. [80] - based in Nat ure Neurosci ence , vol. 9, no. 3, pp. 420 428, Mar. 2006. [81] R. Florian, in Nature Precedings , Nov. 2010 , https://doi.org/10.1038/npre.2010.5190.1 . [82] Spiking Neural Networks with ReSuMe : Sequence Learning , in Neural Computation , vol. 22, no. 2, pp. 467 510, Oct. 2009. [83] - Based Spike Pattern in Neural Computation , vol. 21, no. 5, pp. 1259 1276, Dec. 2008. 149 [84] in Neural Computation , vol. 21, no. 2, pp. 340 352, Sep. 2008. [85] N. Elvin and A. Erturk, Advances in Energy Harvesting Methods in Springer Science & Business Media , 2013. [86] (2003 in Smart Mater. Struct. , vol. 16, no. 3, pp. R1 R21, May 2007. [87] P. Shas hank, , in Journal of Electroceramics , vol. 18, issue 1, pp. 167 - 184 , Sep. 2007 . [88] H. S. Kim, J. - vibrati in Int. J. Precis. Eng. Manuf. , vol. 12, no. 6, pp. 1129 1141, Dec. 2011. [89] in Smart Materials and Structures , vol. 13, no. 5, pp. 1131 1142, Oct. 200 4. [90] in Progress in Organic Coatings , vol. 41, no. 4, pp. 201 216, May 2001. [91] Manual , Providence, RI. [92] http://dsk.ippt.pan.pl/docs/abaqus/v6.13/index.html. [Accessed: 14 - May - 2019]. [93] Michael Bauccio, , 3 rd Edition, ASM International, 1993 . [94] , National Aeronautics and Space Administration (NASA), 2014. [95] - 8 - https://www.microchip.com/wwwproducts/en/ATtiny25. [Ac cessed: 14 - May - 2019]. [96] - 8 - https://www.microchip.com/wwwproducts/en/ATtiny45. [Accessed: 14 - May - 2019]. [97] - 8 - https://www.microchip.com/wwwproducts/en/ATtiny85. [Accessed: 14 - May - 2019]. [98] Active and Passive Smart Structures and Integrated Systems 2008 , 2008, vol. 6928, p . 692824. [99] https://piezo.com/collections/piezo - sheets - plates?_=pf&pf_t_quantity=Quantity__1. [Accessed: 14 - May - 2019]. 150 [100] Encyclopedia of Computer Science , Chichester, UK , John Wiley and Sons Ltd., pp. 1383 1385. [101] S. Das and S. Biswas, , in Proceedings of the 10 th International Conference on Communication Systems and Networks (COMSNETS) , Jan. 2018 , pp. 206 - 213 . [102] 2017 14th IEEE Annual Consumer Communications Networking Conference (CCNC) , 2017, pp. 498 503. [10 3] - space Approach to Tracking and Sensor Proceedings of the 1st ACM International Workshop on Wireless Sensor Networks and Applications , New York, NY, USA, 2002, pp. 131 139. [104] G. Simon et al. - Proceedings of the 2Nd International Conference on Embedded Networked Sensor Systems , New York, NY, USA, 2004, pp. 1 12. [105] L. Yu, N. Wang, and X. Meng - time fore st fire detection with wireless sensor Proceedings. 2005 International Conference on Wireless Communications, Networking and Mobile Computing, 2005. , 2005, vol. 2, pp. 1214 1217. [106] ibuted Piezoelectric Sensor in Sensors (Basel) , vol. 17, no. 6, May 2017. [107] Helpers for computing deltas Available: ht tps://docs.python.org/3/library/difflib.html. [Accessed: 13 - Nov - 2019]. [108] J. M. Eppler, M. Helias, E. Muller, M. Diesmann, and M. - in Front. Neuroinform. , vol. 2, 2009. [109] A. Cammara - Energy: A novel energy prediction model for solar and wind energy - 2012 IEEE 9th International Conference on Mobile Ad - Hoc and Sensor Systems (MASS 2012) , 2012, pp. 75 83. [110] S Learning Approach - 2018 Fifth International Conference on Emerging Applications of Information Technology (EAIT) , 2018, pp. 1 4.