Electrical and Electronic Engineering - Theses
Permanent URI for this collection
Search Results
1 - 10 of 57
-
ItemThermal Multispectral Imaging and Spectroscopy with Optical Metasurfaces and Deep Learning( 2022-12)Spectral imaging captures information in one or more selective bands across the electromagnetic spectrum, permitting the objects in the world to be identified by their absorption or reflection characteristics. Advancements in spectrally selective imaging have primarily been in colour imaging in the visible domain; however, infrared detectors have also enjoyed technological advances that position them ideally for thermal spectral imaging. Advanced spectrally selective imaging systems in longwave infrared (LWIR) thermal wavelengths of 8-14 microns can produce unique thermal fingerprints of objects by recording the heat radiation emitted from objects, thereby creating additional knowledge of the world otherwise difficult to acquire with colour cameras. Therefore, advanced spectral imaging finds important applications in precision agriculture (e.g., early detection of plant diseases), non-invasive medical diagnosis (e.g., vein and dental analysis, skin screening), mining (e.g., non-destructive testing), environmental monitoring (e.g., greenhouse gas detection) and recycling (e.g., plastic classification). However, existing LWIR multi- and hyperspectral imaging systems are expensive and bulky (with cryogenic cooling) and demand time and resources to process several images. Further, LWIR spectral imaging is hindered by the lack of materials responding to thermal wavelengths to design wavelength filters and the low resolution of thermal sensors to design a multi-band filter mosaic compared to their counterpart in the visible wavelengths. Recently, miniaturized infrared spectrometers were reported in the thermal wavelengths. However, they work only with a single isolated object using an active blackbody in the background and fail to detect multiple objects in real scenes. They collect an average emission from multiple objects using single or multiple detectors, which cannot be further resolved due to missing spatial information. There has been an ever-increasing demand for miniaturized and CMOS-compatible LWIR sensors performing imaging spectroscopy to realize their full potential with increased on-chip integration and new compact applications. In this thesis, I design and demonstrate lightweight and high-performance computational infrared imaging technology to enable joint spatial and spectral data acquisition in LWIR wavelengths. I propose and discuss promising solutions for handheld, mass-producible and affordable LWIR multi- and hyperspectral sensing systems using existing monochrome thermal sensors with a focus on plasmonic filters, sensor engineering and artificial intelligence. The first part of this thesis is focused on designing narrowband filter technology towards LWIR multi- and hyperspectral imagers. I begin by presenting optical metasurfaces and designing nano-optical filters with hexagonal lattices of hole/disk geometries to create surface plasmonic resonances in the LWIR regime. I perform comprehensive detector studies and detailed analyses of nano optical filters to accurately tailor the spectral responsivities of the LWIR plasmonic filters for imaging applications. I propose CMOS standard infrared plasmonic filters offering horizontal scalability, narrow spectral width, micron size thickness, and high transmission features. In the second part of this thesis, I explore time-resolved and spatially-resolved multispectral imaging systems for acquiring spatial image information in selective spectral bands. I substantiate the findings from the plasmonic filter simulations by experimentally realizing the novel LWIR plasmonic filters. Their instrumentation is explored by stacking into thermal image sensors through a filter wheel, and by integrating the filter mosaic into the camera to make a compact single-sensor imaging system. I experimentally demonstrate their time- or spatial-multiplexing performance in real-time and recover high-resolution multispectral images with deep imaging. In the third part of this thesis, I develop a deep learning-based LWIR imaging spectroscopy system prototype for acquiring more spectral information with selective spatial images in real time. This is a computational LWIR spectral imaging system acquired by the joint design of a snapshot multispectral imager at the hardware front, and a novel deep learning-based algorithmic spectroscopy concept for rapid spectral reconstruction at the software front. Snapshot images are acquired in selective spectral bands using LWIR plasmonic filters stacked to multiple detectors, which are further processed with deep neural network architecture to rapidly predict the spectra. The power of our deep learning-based imaging spectrometer is experimentally demonstrated by identifying four minerals: amethyst, calcite, pyrite, and quartz. The proposed technique is a simple and approximate 'uncooled LWIR thermal hyperspectral imaging system', which can be used to identify multiple objects by retrieving the spectral fingerprint in a real scene without recording a large number of images and without needing an active blackbody source. I thus demonstrate next-generation thermal sensing systems by merging nanoplasmonic sensors and artificial intelligence. Our results will form the basis for a snapshot, lightweight, compact, and low-cost hyperspectral LWIR imagers enabling diverse applications in chemical detection, precision agriculture, disease diagnosis, environmental sensing and industry vision.
-
ItemPrinted Radio Transceivers( 2022-10)The application of additive manufacturing technologies in the field of radiofrequency electronics is explored. Low-cost manufacturing technology is applied at the component level to fabricate P-N junctions and combined with other components to form a simple printed radio transceiver. An analysis is performed for large-scale antenna arrays - amenable to low-resolution additive manufacturing technologies - with a view to further improve the capabilities of printed radiofrequency systems. A low-resolution PCB printer with a minimum feature size in the order of 200 micrometres and thermal curing temperature up to 515 kelvin is demonstrated capable of fabricating P-N junctions suitable for application as a varactor at 100MHz to 2GHz. P-N junctions were formed by printing N type inks with inorganic ionic compounds (zinc oxide and diindium trioxide) and P-type inks with conjugated polymers (PEDOT:PSS and pentacene) as the active ingredients. For two P-N junction chemistries (the first combining PEDOT:PSS with diindium trioxide, the second combining PEDOT:PSS with zinc oxide), device impedance characteristics showed promise for application in a 50 ohm system operating at UHF frequencies. These inks were characterised with a source meter to demonstrate a current-to-voltage characteristic according to the Shockley equation and a network analyser to quantify the depletion region capacitance and equivalent series resistance as a function of frequency and bias voltage. A simple radio transceiver was demonstrated by printing a P-N junction as the terminating impedance to a printed antenna. Fabrication was completed in less than four hours and did not require temperatures exceeding 500 kelvin. As a transmitter, load modulation was applied to the P-N junction (a varactor) to phase modulate an incident radiofrequency signal before reradiating. Over-the-air measurements quantifying the power received from the transmitter as a function of distance, carrier frequency, signal bandwidth and BER for a 2FSK signal; demonstrated reasonable agreement with a Friis-equation-based path loss model. As a receiver, an AM signal received by the antenna experienced self-mixing due to the nonlinear current versus voltage characteristic of the P-N junction. The mathematics describing the harmonic content of the demodulated signal showed reasonable agreement to SPICE simulation data and empirical measurements. To overcome inverse quartic losses that commonly afflict such backscatter-based systems, antenna array techniques were demonstrated as a viable strategy. The reflected signal from an infinite size, infinitely dense antenna array was shown to converge to an optical mirror (i.e. inverse square losses) under certain circumstances. Further reductions in radiative power loss are obtained under the condition of (retrodirective) phase conjugate matching. Under ideal conditions, a planar array of isotropic antennas can return up to 25% of the energy from an isotropic source to a collocated receiver. Accounting for incomplete coverage area, improperly specified apertures, cable loss and antenna efficiency, the model was compared against FEM simulation data and empirical measurements for a linear array, showing reasonable agreement. Thus, low-resolution additive manufacturing was demonstrated as a viable technology for fabricating radiofrequency systems. Low-cost printers in conjunction with semiconducting and conducting inks may be used to fabricate simple radio transceivers, which may be combined in retrodirective arrays for improved performance characteristics.
-
ItemInformation Theory and Machine Learning: A Coding Approach( 2022-11)This thesis investigates the principles of using information theory to analyze and design machine learning algorithms. Despite recent successes, deep (machine) learning algorithms are still heuristic, vulnerable, and black-box. For example, it is still not clear why and how deep learning works so well, and it is observed that neural networks are very vulnerable to adversarial attacks. On the other hand, information theory is a well-established scientific study with a strong foundation in mathematical tools and theorems. Both machine learning and information theory are data orientated, and their inextricable connections motivate this thesis. Focusing on data compression and representation, we first present a novel, lightweight supervised dictionary learning framework for text classification. Our two-stage algorithm emphasizes the conceptual meaning of dictionary elements in addition to classification performance. A novel metric, information plane area rank (IPAR), is defined to quantify the information-theoretic performance. The classification accuracy of our algorithm is promising following extensive experiments conducted on six benchmark text datasets, where its classification performance is compared to multiple other state-of-the-art algorithms. The resulting dictionary elements (atoms) with conceptual meanings are displayed to provide insights into the decision processes of the learning system. Our algorithm achieves competitive results on certain datasets and with up to ten times fewer parameters. Motivated by the similarity between communication systems and adversarial learning, we secondly investigate a coding-theoretic approach to increase adversarial robustness. Specifically, we develop two novel defense methods (eECOC and NNEC) based on error-correcting code. The first method uses efficient error-correcting output codes (ECOCs), which encode the labels in a structured way to increase adversarial robustness. The second method is an encoding structure that increases the adversarial robustness of neural networks by encoding the latent features. Codes based on Fibonacci lattices and variational autoencoders are used in the encoding process. Both methods are validated on three benchmark datasets, MNIST, FashionMNIST, and CIFAR-10. An ablation study is conducted to compare the effectiveness of different encoding components. Several distance metrics and t-SNE visualization are used to give further insights into how these coding-theoretic methods increase adversarial robustness. Our work indicates the effectiveness of using information theory to analyze and design machine learning algorithms. The strong foundation of information theory provides opportunities for future research in data compression and adversarial robustness areas.
-
ItemA Blockchain-based Solution for Sharing IoT Devices( 2022)The Internet of Things (IoT) includes billions of sensors and actuators (which we refer to as IoT devices) that harvest data from the physical world and send it via the internet to IoT applications to provide smart IoT solutions. These IoT devices are often owned by different organizations or individuals who deploy them and utilize their data for their own purposes. Procuring, deploying, and maintaining IoT devices for exclusive use of an individual IoT application is often inefficient, and involves significant cost and effort that often outweigh the benefits. On the other hand, sharing IoT devices that are procured, deployed, and maintained by different entities (IoT device providers or simply providers) is efficient, cost-effective and enables rapid development and adoption of IoT applications. Currently, most IoT applications themselves procure, deploy, and maintain the sensors they need to collect the IoT data they require as there is limited support for sharing IoT devices and their costs. Therefore, there is a need for developing an IoT device sharing solution that allow IoT applications to 1) discover already deployed IoT devices, 2) use discovered IoT device data (IoT data) for their own purposes, and 3) share-cost of IoT device deployment via a “pay-as-you-go” model similar to cloud computing. To address the aforementioned problems, in this thesis we propose, develop, implement, evaluate, and validate a solution namely IoT Devices Sharing (IoTDS). IoTDS enables scalable and cost-efficient discovery and use of IoT devices by IoT applications. IoTDS incorporates services for IoT device registration, IoT device query, IoT device payment and IoT device integration. To support these services, we propose 1) a novel IoTDS ontology, an extension of Semantic Sensor Network (SSN) ontology to describe IoT devices and their data to enable IoT device registration and query services. The IoTDS ontology also provides for describing the payment and integration information that is used by IoT device payment and integration service; 2) a special-purpose blockchain namely IoTDS Blockchain that has been developed specifically to support the needs of the IoTDS services i.e., supporting decentralised and scalable query, integration and payments services for IoT devices and applications. Specifically, IoTDS Blockchain incorporates a distributed semantic triple store and functions to register IoT devices, and specialised transactions for supporting IoT device payments (we propose a new cryptocurrency namely SensorCoin); 3) a novel IoT marketplace (IoTDS marketplace) that offers an interface and a protocol (IoTDS protocol) to support the interactions between IoT devices, IoT applications, IoTDS Blockchain, and IoTDS services. IoTDS solution 1) facilitates IoT devices deployed across the globe by different providers to be queried by any IoT application; we term this global, 2) enables via the IoTDS Blockchain a non-ownership model; we term this IoT-owned i.e., no individual/organisation owns it or controls it, 3) able to handle the vast and ever-increasing number of IoT devices and IoT applications; we term this scalable, and 4) able to support and integrate heterogenous of IoT devices and their data; we term this interoperable. In this thesis, we provided implementation details of IoTDS services that includes IoTDS ontology, marketplace, and blockchain. the IoTDS ontology has been modelled using Protegee and Owl and implemented using RDF. The IoTDS Blockchain and corresponding functions are implemented using NodeJS and Web Socket. The IoTDS marketplace and corresponding IoTDS protocol has been implemented using NodeJS and MQTT. We conducted large-scale experimental evaluation of IoTDS solution by deploying it on Nectar cloud (20 instances) using both real and simulated (5,000,000) IoT devices and IoT applications (5000) to assess and validate the scalability and performance of IoTDS. We also developed and validated a mathematical model that can be used to estimate the performance of the IoTDS with the increasing number of IoT devices. Experimental outcomes show that the proposed IoTDS solution performs great (linear scalability) in supporting global discovery, use, and cost-share of large numbers of IoT devices and applications. The main contributions of this thesis are 1) an IoTDS solution for sharing IoT devices, 2) a survey of techniques for supporting IoT device sharing; 3) a special purpose Blockchain to support sharing of IoT devices, 4) a novel Marketplace to support registration, querying, payment, and integration of IoT devices, 5) a novel protocol for autonomic control of integrating IoT devices and fetching their data, and 6) an implementation and experimental evaluation of the IoTDS solution.
-
ItemAdvanced Control of Wind Energy Conversion Systems for Grid Frequency Support( 2022)Renewable energy sources have been increasing rapidly recently. These generation units which are converter based technologies are replacing the conventional generator systems in the power grid. As a result fewer generators are participating in frequency regulation services and the frequency deviation from its nominal value has increased lately. An example of these converter based technologies are wind turbines which normally operate in maximum power point tracking meaning that they generate as much energy that they can possibly harvest from the wind. The problem of frequency deviation has increased the attention of many researchers to tackle the problem and investigate the possibility of wind turbines to participate in frequency regulation services. Frequency regulation is done by controlling the appropriate active power supplied to the transmission lines. There are various responses to an event that happens in the grid. While the inertial response and primary frequency response are the first two controls of the system to bring the frequency back to its operating point, secondary frequency regulation aims to eliminate steady state error of the frequency from its nominal value. Secondary frequency regulation is managed by the market operator. This means that the generation units have to track a power command signal generated by the market operator. This problem could be modeled as a tracking problem, since the wind generation unit has to track a power command signal sent by the market operator and reject the disturbances such as wind variation or fluctuations of the terminal voltage of the converter based generators. The problem is inherently difficult due to the time varying power commands as references and the stochastic disturbances such as wind variations. Technologies such as Light Detection And Ranging (LIDAR) has absorbed the attention of many researchers. This technology provides preview information for the coming wind disturbances up to several seconds ahead. Methods to use this information for better tracking the power command signal or better performance in maximum power point tracking has been the topic of many research articles so far. In this thesis, we will investigate the capability of a classical control methodology to provide wind turbines with the capability to participate in frequency regulation services. This control methodology is known as exact output regulation. It considers a time invariant plant model and has the capability to track a known reference signal and reject disturbance signals. The wind information could be modeled using an exo-system and produce the disturbance signal and the market operator will produce the reference signals to be tracked. This thesis will have two different scenarios considering the problem of secondary frequency regulation. In the first scenario the wind turbine is modeled by a high fidelity aero-elastic simulator known as Fatigue, Aerodynamics, Structures, and Turbulence (FAST) in conjunction with a simple generator. In this section the control is basically on the wind turbine and LIDAR wind preview information is also used to obtain the disturbance signal. In the second scenario we will investigate adding a Doubly Fed Induction Generator (DFIG) instead of a simple generator and design the control for both the generator and wind turbine. We have investigated two different types of output regulation which are well suited for each problem. To have a realistic results we have employed FAST 5 MW reference for the turbine model. The DFIG is implemented in MATLAB Simulink and to simulate stochastic wind signals we have used Turbsim which is able to generate different practical classes of wind signals. Then the two different problems has been addressed and compared against the performance of baseline controllers. The baseline controllers are the most widely used methods for wind turbine active power control. Different control objectives are defined in each chapter for the purpose of comparison between the proposed controller and the baseline controllers. These include the root mean squared error between the generated power and the power command signals, fatigue loads and actuator usage. Results show that the proposed output regulation methods in both scenarios are able to track the command signals better than the baseline controllers. In terms of the fatigue loads, in the first scenario the controller is able to reduce the fatigue loads in most of the considered fatigue cases such as blade and tower bending moments as well as the low speed shaft torque, however in the second scenario the fatigue loads for tower and blade bending moments were similar to baseline controller and only the low speed shaft torque was improved. The input command signals was smoother in both of the scenarios when using the proposed output regulation control techniques.
-
ItemAutomated Assessment of Motor Functions in Stroke using Wearable Sensors( 2022)Driven by the aging population and an increase in chronic diseases worldwide, continuous monitoring of human activities and vital signs have become a major focus of research. This has been facilitated by the advent of wearable devices equipped with miniaturized sensors. Compared to bench-top devices in hospitals and laboratories, wearable devices are popular in improving health outcomes, because of their compact form factors and unobtrusive nature. Stroke, a neurological disorder, is a major concern among all chronic diseases because it causes high rates of death and disability globally every year. Motor deterioration is the most common effect of stroke, leading to one-sided weakness (i.e., hemiparesis), and limiting movements and coordination. Stroke survivors require regular assessments of motor functionality during the acute, sub-acute and chronic phases of recovery, leading to dependence on human intervention and massive expenditures on patient monitoring. Therefore, an automated system for detecting and scoring hemiparesis, independent of continuous specialized medical attention, is necessary. This thesis develops various methods to objectively quantify motor deterioration related to stroke using wearable motion sensors, for automated assessment of hemiparesis. In the first part of the thesis, we use accelerometer data acquired from wrist-worn devices to analyze upper limb movements and identify the presence and severity of hemiparesis in acute stroke, during a set of spontaneous and instructed tasks. We propose measures of time (and frequency) domain coherence between accelerometry-based activity measures from two arms, that correlate with the clinical gold standard National Institutes of Health Stroke Scale (NIHSS). This approach can accurately distinguish between healthy controls, mild-to-moderate and severe hemiparesis through supervised pattern recognition, using a hierarchical classification architecture. We propose additional descriptors of bimanual activity asymmetry, that characterize the distribution of acceleration-derived activity surrogates based on gross and temporal variability, through a novel bivariate Poincare analysis method. This leads to achieving further granularity and sensitivity in hemiparesis classification into four classes, i.e., control, mild, moderate and severe hemiparesis. The second part of the thesis analyzes the quality of spontaneous upper limb motion captured using wearable accelerometry. Here, velocity time series estimated from the acquired data is decomposed into movement elements, which are smoother and sparser in the normal hand than the paretic hand, and the amount of smoothness correlates with hemiparetic severity. Using statistical features characterizing their bimanual disparity, this method can classify mild-to-moderate and severe hemiparesis with high accuracy. Compared to the activity-based features, this method is more interpretable in terms of joint biomechanics and movement planning, and is robust to the presence of noise in the acquired data. In the third part of the thesis, we propose unsupervised methods for bimanual asymmetry visualization in hemiparesis assessment, using motion templates representative of well-defined instructed tasks. These methods are aimed at creating models for assessing the qualitative progression of motor deterioration over time instead of single-point measurements, or when class labels representing clinical severity are not available. We propose variants of the Visual Assessment of (cluster) Tendency (VAT) algorithm, to study cluster evolution through heat maps, by representing instructed task patterns through local timeseries characteristics, known as shapelets. These shapelets transform high dimensional sensor data into low-dimensional feature vectors for VAT evaluation. We show the significance of these methods for efficient and interpretable cluster tendency assessment for anomaly detection and continuous motion monitoring, applicable not only to hemiparesis assessment, but also in identifying motor functionality in other neurological disorders or activity recognition problems. Finally, in the fourth part of the thesis, we show applications of the above methods to objectively measure gait asymmetry in stroke survivors, using lower limb position data from wearable infrared markers and camera-based motion capture devices. These methods can efficiently quantify the severity of lower limb hemiparesis, thereby being suitable for automated gait monitoring during extended training and rehabilitation in the chronic phase of recovery.
-
ItemThermo-mechanical energy storage applications for energy system decarbonisation( 2022)This research explores the prospective application of thermo-mechanical energy storage technologies for energy system decarbonisation. It characterises, first, the techno-economic performance of one such technologies, Liquid Air Energy Storage (LAES), when operated within the power system to supply energy and reserve services. Then, Liquid Air Energy Storage operation as a multi-energy asset is studied. To conclude, the potential of six between established and novel thermo-mechanical energy storage concepts is cross-compared and benchmarked with incumbent storage technologies for long-duration energy storage applications.
-
ItemScenario Based Optimization over Uncertain System Identification Models( 2022)A model describes the relationship between inputs and outputs of a system and offers an explanation of the system behaviour. Mathematical models of dynamical systems are widely used in many field of science and engineering. There are uncertainties associated with the models, and it is important to quantify these uncertainties. System identification can be used to obtain models of dynamical system. A typical system identification approach is to select a parameterized model class and estimate the unknown parameters from observed data. This includes method such as least squares, prediction error methods and instrumental variable methods that give a point estimate for the unknown parameters. There is always an error between the estimate and the true model parameter when the observed data are noisy. This error represents the model uncertainty. Another way to describe the model uncertainty is to use confidence regions. A confidence region contains a set of plausible values for the unknown true parameter and contains the true parameter with a certain probability. There are many ways to construct confidence regions. By using asymptotic system identification theory and assuming the number of data points goes to infinity, confidence regions that contains the true model parameter with a certainty probability can be found. In recent years, new methods such as Leave-out Sign-dominant Correlation Regions (LSCR) and Sign-Perturbed Sums (SPS) have also been developed which provide confidence regions when only a finite number of data points are available. In optimization based design problem, a cost function reflecting the design target is minimized. The found decision variable and minimized cost will depend on the model parameter, which is unknown but can be described by the confidence region. One approach is to apply a robust design, which minimizes the worst case value of the cost function over a set of unknown parameters. This problem can be difficult to solve in some cases. Therefore a computationally tractable method called scenario approach is used in this thesis. In the scenario approach, samples are drawn from the model uncertainty set and an optimization problem is then solved based on the drawn samples. In this thesis approaches that combine system identification methods and the scenario approach for different data generating systems are investigated. They are considered in three different settings. In the Bayesian framework, a Bayesian approach to system identification is considered and the system parameters are viewed as a realization of a random vector. Using the observed data, the posterior density of the system parameters can be computed and used as the model uncertainty set from which samples are drawn from to be used in scenario approach. Algorithms are provided to obtain approximately i.i.d samples from the posterior distribution. In the non-Bayesian framework, a model is obtained via system identification and the uncertainty associated with the model is characterized by the distribution of estimation error. By knowing this estimation error, samples can be built based on this estimation error. Algorithms are provided for obtaining the samples and theoretical results are also given. In the Sign-Perturbed Sums framework, we use SPS method to find the confidence region and this confidence region is considered as the model uncertainty. By knowing whether a given point belongs to the SPS confidence region, algorithms for combining SPS and scenario approach are designed.
-
ItemUniformly Bounded State Estimation over Multiple Access Channels( 2022)In this doctoral thesis, a characterization of the zero-error capacity region for three different classes of multiple access channels (MACs) is derived. The first type of channels considered in this work is a two-user MAC with a common message that captures the correlation between transmitters. Next, this model is extended by considering an arbitrary number of users M >= 2. The last class of MACs represents a further extension to a more general case where inter-user correlation is modeled by a common message seen by all users as well as pairwise shared messages. In this research, we look at the zero-error capacity, which differs from the more commonly studied small-error capacity, from a nonprobabilistic angle. In fact, the obtained characterization is based on the so-called nonstochastic information, and is valid not only for asymptotically large coding block-lengths but also for finite lengths. Understanding how to coordinate unambiguous communication through MACs, such that several unrelated senders can simultaneously send as much information as possible is of great interest, especially with the emergence of new paradigms such as the Internet of Things (IoT) and Machine-to-Machine (M2M) communication. Next, using the characterization of the zero-error capacity region for the two-user MAC, we investigate the problem of distributed state estimation under the criterion of uniformly bounded estimation errors. It is shown that if there exists a coder-estimator tuple that achieves the desired criterion, namely uniformly bounded estimation error, the vector of topological entropies of the linear systems, whose state is being estimated, must lie within the zero-error capacity region of the communication channel. Additionally, we prove that if the to-be-observed plants have a topological entropy vector inside the interior of the zero-error capacity region, the existence of a coder-estimator tuple achieving uniformly bounded state estimation errors is guaranteed. This result relates the channel properties to the plant dynamics and paves the way toward understanding information flows in networked control systems with multiple transmitters. Finally, we seek to characterize the fundamental tradeoff between the communication data rate, code-length, system dynamics and state estimation performance. To this end, a universal lower bound on the time-asymptotic estimation error is obtained using volume-based analysis. Additionally, to provide a guarantee on the estimation performance, an upper bound on the error is derived when the measurements are quantized. When the code-length is large, we show that these lower and upper bounds converge to the same limit.
-
ItemArchitecture and Policy Design for Next-generation Access Networks( 2022)With more than twenty years down the twenty-first century, communication networks are undergoing a paradigm shift. Due to the increase in available computing power, computing-intensive applications in form of augmented/virtual reality, internet of vehicles, remote automation, etc., have emerged in addition to the traditional voice, video, and data. The increasing role of computing in executing applications over networks led to the emergence of cloud, and subsequently edge and fog computing. Next-generation networks are envisioned to be application-driven and designed to satisfy end-to-end (E2E) quality of service (QoS) and quality of experience (QoE) requirements of the applications, considering both network and computing paradigms. This thesis focuses on achieving the aforementioned goals at the access part of the network that connects end-users to the service providers. The access segment of the network experiences significant dynamic behavior as compared to its core counterpart due to the independent and random nature of users and customers. The thesis investigates two popular access network technologies: Passive Optical Networks (PON) and Radio Access Networks (RAN), in the context of fog and edge computing. While PON is a wired access network, RAN connects to users wirelessly with a wired counterpart from radio stations to the edge/core servers (also known as backhaul). It is desired that fog/edge nodes be integrated with PON in a cost-effective and seamless manner, without altering the protocols in place. In addition, it is important to investigate and design dynamic bandwidth allocation (DBA) policies that can satisfy the strict QoS requirements. Chapter 3 of this thesis, demonstrates how to design a cost-effective fog-integrated architecture for PON. It also delineates a dynamic bandwidth allocation protocol that enables the communication between fog node and users without significantly changing the existing DBA of PON. In Chapter 4, the problem of satisfying strict QoS requirements is solved using the Model Predictive Control (MPC) technique. For this, an innovative delay tracking mechanism using virtual queues is developed, allowing one to take far-sighted decisions in contrast to short-sighted ones that is commonly employed in the literature. In RAN, it is envisioned that future networks be zero-touch, implying that the network is able to intelligently automate its policies according to demands, thus significantly reducing human intervention. Orchestrators such as software defined network (SDN) controller (for networks) and Kubernetes (for servers) play an important role as key enablers of a zero-touch implementation. To meet the different QoE requirements of new applications, networking and computing resources are virtualized and sliced-up for each application type, also known as network slicing. To achieve E2E QoE, the two orchestrators should work jointly and create slices of networking and computing resources to satisfy E2E QoE. In addition, they must establish the relationships between E2E QoE and resources so that resource requirements are decided for both deterministic and dynamically changing environments in an automated manner. Chapter 5 of this thesis develops sequential distributed learning and optimization models to learn the relationships under static and dynamic conditions and take robust slicing decisions to achieve E2E QoE at the backhaul of RAN. The learning process requires the incorporation of artificial intelligence (AI) in the slicing process which is a crucial step towards zero-touch network design. To summarize, this thesis demonstrates how next-generation access network architectures involving fog/edge computing are designed, operated and maintained in an automated and seamless manner.