Electrical and Electronic Engineering - Theses

Permanent URI for this collection

Search Results

Now showing 1 - 7 of 7
  • Item
    Thumbnail Image
    Model Predictive Controller Tuning by Machine Learning and Ordinal Optimisation
    Chin, Robert Alfred ( 2021)
    While for the past several decades model predictive control (MPC) has been an established control strategy in chemical process industries, more recently there has been increased collaboration in MPC research between academia and automotive companies. Despite the promising work thus far, one particular challenge facing the widespread adoption of MPC in the automotive industry is the increased calibration requirement. The focus of the research in this thesis is to develop methods towards reducing the calibration effort in designing and implementing MPC in practice. The research is tailored by application to offline tuning of quadratic-cost MPC for an automotive diesel air-path, to address the limited time-availability to perform online tuning experiments. Human preferences can be influential in automotive engine controller tuning. Some earlier work has proposed a machine learning controller tuning framework (MLCTF), which learns preferences from numeric data labelled by human experts, and as such, these learned preferences can be replicated in automated offline tuning. Work done in this thesis extends this capability by allowing for preferences to be learned from pairwise comparison data, with monotonicity constraints in the features. Two methods are proposed to address this: 1) an algorithm based around Gaussian process regression; and 2) a Bayesian estimation procedure using a Dirichlet prior. These methods are successfully demonstrated in learning monotonicity-constrained utility functions in time-domain features from data consisting of pairwise rankings for diesel air-path trajectories. The MLCTF also constitutes a plant model, yet there will typically be some uncertainty in an engine model, especially if it has been identified from data collected with a limited amount of experimentation time. To address this, an active learning framework is proposed for selection of the next operating points in the design of experiments, for identifying linear parameter-varying systems. The approach is based on exploiting the probabilistic features of Gaussian process regression to quantify the overall model uncertainty across locally identified models, resulting in a flexible methodology which accommodates for various techniques to be applied for estimation of local linear models and their corresponding uncertainty. The framework is applied to the identification of a diesel engine air-path model, and it is demonstrated that measures of model uncertainty can be quantified and subsequently reduced. To make the most of the limited availability for online tuning experiments, an ordinal optimisation (OO) approach is proposed, which seeks to ensure that offline tuned controllers can perform acceptably well, once tested online with the physical system. Via the use of copula models, an OO problem is formulated to be compatible with the tuning of controllers over an uncountable search space, such as quadratic-cost MPC. In particular, results are obtained which formally characterise the copula dependence conditions required for the OO success probability to be non-decreasing in the number of offline controllers sampled during OO. A gain-scheduled MPC architecture was designed for the diesel air-path, and implemented on an engine control unit (ECU). The aforementioned non-decreasing properties of the OO success probability are then specialised to tuning gain-scheduled controller architectures. Informed by these developments, the MPC architecture was firstly tuned offline via OO, and then tested online with an experimental diesel engine test rig, over various engine drive-cycles. In the experimental results, it was found that some offline tuned controllers outperformed a manually tuned baseline MPC, the latter which has comparable performance to proprietary production controllers. Upon additional manual tuning online, the performance of the offline tuned controllers could also be further refined, which illustrates how offline tuning via OO may complement online tuning approaches. Lastly, using an analytic lower bound developed for OO under a Gaussian copula model, a sequential learning algorithm is developed to address a probabilistically robust offline controller tuning problem. The algorithm is formally proven to yield a controller which meets a specified probabilistic performance specification, assuming that the underlying copula is not too unfavourably far from a Gaussian copula. It is demonstrated in a simulation study that the algorithm is able to successfully tune a single controller to meet a desired performance threshold, even in the presence of probabilistic uncertainty in the diesel engine model. This is applied to two case studies: 1) `hot-starting' an online tuning procedure; and 2) tuning for uncertainty inherent across a fleet of vehicles.
  • Item
    Thumbnail Image
    Adversarial Robustness in High-Dimensional Deep Learning
    Karanikas, Gregory Jeremiah ( 2021)
    As applications of deep learning continue to be discovered and implemented, the problem of robustness becomes increasingly important. It is well established that deep learning models have a serious vulnerability against adversarial attacks. Malicious attackers targeting learning models can generate so-called "adversarial examples'' that are able to deceive the models. These adversarial examples can be generated from real data by adding small perturbations in specific directions. This thesis focuses on the problem of explaining vulnerability (of neural networks) to adversarial examples, an open problem which has been addressed from various angles in the literature. The problem is approached geometrically, by considering adversarial examples as points which lie close to the decision boundary in a high-dimensional feature space. By invoking results from high-dimensional geometry, it is argued that adversarial robustness is impacted by high data dimensionality. Specifically, an upper bound on robustness which decreases with dimension is derived, subject to a few mathematical assumptions. To test this idea that adversarial robustness is affected by dimensionality, we perform experiments where robustness metrics are compared after training neural network classifiers on various dimension-reduced datasets. We use MNIST and two cognitive radio datasets for our experiments, and we compute the attack-based empirical robustness and attack-agnostic CLEVER score, both of which are approximations of true robustness. These experiments show correlations between adversarial robustness and dimension in certain cases.
  • Item
    Thumbnail Image
    Novel Defenses Against Data Poisoning in Adversarial Machine Learning
    Weerasinghe, Prameesha Sandamal Liyanage ( 2019)
    Machine learning models are increasingly being used for automated decision making in a wide range of domains such as security, finance, and communications. Machine learning algorithms are built upon the assumption that the training data and test data have the same underlying distribution. This assumption fails when (i) data naturally evolves, causing the test data distribution to diverge from the training data distribution, and (ii) malicious adversaries distort the training data (i.e., poisoning attacks), which is the focus of this thesis. Even though machine learning algorithms are used widely, there is a growing body of literature suggesting that their prediction performance degrades significantly in the presence of maliciously poisoned training data. The performance degradation can mainly be attributed to the fact that most machine learning algorithms are designed to withstand stochastic noise in data, but not malicious distortions. Through malicious distortions, adversaries aim to force the learner to learn a model that differs from the model it would have learned had the training data been pristine. With the models being compromised, any systems that rely on the models for automated decision making would be compromised as well. This thesis presents novel defences for machine learning algorithms to avert the effects of poisoning attacks. We investigate the impact of sophisticated poisoning attacks on machine learning algorithms such as Support Vector Machines (SVMs), one-class Support Vector Machines (OCSVMs) and regression models, and introduce new defences that can be incorporated into these models to achieve more secure decision making. Specifically, two novel approaches are presented to address the problem of learning under adversarial conditions as follows. The first approach is based on data projections, which compress the data, and we examine the effect of the projections on adversarial perturbations. By projecting the training data to lower-dimensional spaces in selective directions, we aim to minimize the impact of adversarial feature perturbations on the training model. The second approach uses Local Intrinsic Dimensionality (LID), a metric that characterizes the dimension of the local subspace in which data samples lie, to distinguish data samples that may have been perturbed (feature perturbation or label flips). This knowledge is then incorporated into existing learning algorithms in the form of sample weights to reduce the impact of poisoned samples. In summary, this thesis makes a major contribution to research on adversarial machine learning by (i) investigating the effects of sophisticated attacks on existing machine learning models and (ii) developing novel defences that increase the attack resistance of existing models. All presented work is supported by theoretical analysis, empirical results, and is based on publications.
  • Item
    Thumbnail Image
    Impact of Rooftop Solar on Distribution Network Planning
    Gupta, Ashish Bert ( 2019)
    Electricity networks have been undergoing significant transformation recently, especially in terms of embedded generation. There has been a lot of focus on demand fluctuations from solar and wind farms that are being connected onto high voltage (HV) grids in energy markets. But the distribution low voltage (LV) grid may prove the most challenging for the network owners and market operators. This is because rooftop solar, whether installed in commercial or residential areas, is leading to high demand fluctuations within the last mile. Customer-installed solar is also causing voltages to rise, but it is the Distribution Network Operator (DNO) on which the responsibility of voltage regulation falls. There is hence greater importance for the DNO to have full visibility of the LV feeder voltages at all times, accurately analysing proposed connections, and meeting the regulators’ and government expectations of enabling solar penetration. Voltage monitoring and regulating infrastructure at the LV level, though, is expensive to implement and hence scarce due to its huge scale. Utilities hence employ empirical or statistical techniques to calculate voltage drop and voltage rise. Conservative allowances for demand diversity and unbalance can lead to erroneous results and can form the basis of considerable utility capital expenditure programs. Utility expenditure in turn usually leads to an increase in customer bills over time. A small number of utilities in the world have access to voltage data from smart metering infrastructures, such as in Victoria, Australia, but ownership of data is becoming an open question. Data availability also presents a different problem to them, as these meters are leading to an extraordinary amount of near real-time data, which they are failing to fully embrace. They see smart-technology driven initiatives as a form of disruption and are slow or unwilling to adapt to the changing nature of the grid. This dissertation details the use of data analytics for forecasting future voltages on the network. Standard machine learning techniques are used to create a non-linear regression model fit to train parameters that reflect the operational status of the feeder. These parameters reflect load diversity and unbalance as well as generator diversity and unbalance. The trained model consequently accurately predicts voltages on the feeder with additional connections. A load-flow simulation of a real-world network is carried out. Training and testing are performed on data from different halves of the year. Predicted voltages are compared to simulation results to confirm the high accuracy, even though consumption patterns and solar irradiation patterns change due to different seasons in the test data. Hence, by leveraging interval metering data, it is shown how standard machine learning methods can be used to develop forecasting capabilities. The methodology developed in this thesis can used as a planning tool to quickly and accurately evaluate future rate of recurrence of voltage violations; and predict the voltage headroom available on the LV feeder. This is a significant outcome as predictability of LV feeder voltages is a concern for the utilities, consumers as well as regulating bodies. The presented method will enable more loads and PVs onto the network without the need of new assets such as distribution transformers or LV feeders, that may be left underutilised. It will also help resolve certain quality of supply issues such as voltage drop complaints; and help better prioritise and technically analyse constrained areas of the network. It is clear that high-quality, high-volume data analysis will play a key role in resolving the needs of the electricity industry. This thesis serves as an interface between network planning engineers and data scientists who will solve the emerging energy constraints, play a part in minimising customer energy prices and assist in the transition to decentralised clean energy sources.
  • Item
    Thumbnail Image
    Privacy-preserving machine learning and data aggregation for Internet of Things
    Lyu, Lingjuan ( 2018)
    The proliferation of Internet of Things (IoT) devices has contributed to the emergence of participatory sensing (PS) and collaborative learning (CL), where multiple participants collect and report their data to a cloud service to analyse the union of the collected data in the server-based framework. While in the decentralized framework, multiple participants collaboratively train a more accurate global model or multiple local models. However, the possibility of the cloud service or any participant being semi-honest or malicious pose a serious challenge of preserving the participants' privacy. Privacy-preserving machine learning and data aggregation aim to discover or derive useful statistics without compromising privacy. This thesis systematically investigates state-of-the-art techniques for privacy-preserving machine learning and data aggregation in a range of IoT applications. Extensive theoretical and experimental results are provided to support the following primary contributions. First, we explore three privacy-preserving machine learning applications. Examples include collaborative anomaly detection, human activity recognition and decentralized collaboration in a biomedical domain. We tackle security challenges in collaborative anomaly detection with a two-stage scheme called RG+RT: in the first stage, participants individually perturb their data by passing through a nonlinear function called repeated Gompertz (RG); in the second stage, the perturbed data are projected to a lower dimension using a participant-specific uniform random transformation (RT) matrix. The nonlinear RG function is designed to mitigate maximum a posteriori (MAP) estimation attacks, while random transformation resists independent component analysis (ICA) attacks. For human activity recognition, a similar two-stage scheme called RG+RP is proposed, the difference lies in the second stage, where participants project their perturbed data to a lower dimension in an (almost) distance-preserving manner, using a random projection (RP) matrix. The random projection can both resist ICA attacks and maintain model accuracy. These proposed two-stage randomisation schemes are assessed in terms of their recovery resistance to MAP estimation attacks. Preliminary theoretical analysis as well as experimental results on synthetic and real-world datasets indicate that both RG+RT and RG+RP exhibit better recovery resistance to MAP estimation attacks than most state-of-the-art techniques, meanwhile high utility is guaranteed. To mitigate the inherent limitations in the centralized framework, and investigate the applicability of the decentralized framework, we study the decentralized collaboration in a biomedical domain. In particular, we develop an efficient Decentralized Privacy-Preserving Centroid Classifier (DPPCC) considering three practical scenarios, where distributed differential privacy (DDP) is combined with distributed exponential ElGamal cryptosystem to preserve privacy and maintain utility. We realize DDP using discrete Gaussian mechanism without any restriction on ε as in the traditional Gaussian mechanism, and only the encrypted noisy model parameters or test results are shared among all parties. It ensures each party learns nothing but the noisy sum of local statistics. Second, we examine privacy-preserving data aggregation in smart grid application. To this end, we propose a multi-level aggregation framework based on fog architecture, which
  • Item
    Thumbnail Image
    Optimal design of low carbon residential microgrids
    Percy, Steven ( 2018)
    Internationally, the residential sector makes up 31% of global energy use [1]. In 2015, Australian households were responsible for 23.5% of Australia's total net electricity demand [2] at 456 PJ annually [3]. Without considering the potential of the Emissions Reduction Fund, coal-fired generation is projected to continue to supply the bulk of Australia’s electricity requirements to 2035 [4]. However, the residential sector has the potential to move to a low carbon future through the increased application of distributed generation and distributed storage, microgrid systems, new demand response methods, innovative passive building designs and improved energy efficiency [5]. The size of many early distributed generation, distributed storage and microgrid trials is often limited by funding, where the financial risk of deployment is minimised by keeping the trial small [6]–[8]. Through improved modelling methods, the financial risk can be reduced. Improved modelling and optimisation has the opportunity to simultaneously reduce financial risk (through higher fidelity estimates of operation and performance) and improve the overall business case (by providing guidance on the best configuration, design and structure of deployments). In this thesis, we present an integrated modelling and optimisation framework that has new capabilities to design, build and test the business case for low Carbon microgrid precincts with storage and distributed generation technologies at their core. This modelling framework analyses the impacts of low carbon infrastructure on carbon emissions, precinct demand and household demand. The modelling framework consists of four components: 1) An electricity demand model for the estimation of household and precinct load behaviour in urban precincts. 2) A standalone solar and battery model for analysing impacts on demand, costs and selecting optimal capacities. 3) A microgrid model for comparing distributed and utility storage options, and analysing impacts of installed capacity on grid demand and cost. 4) A PowerFactory microgrid network model for verifying the network voltage levels and losses of the microgrid designs. To develop the electricity demand model we draw on an extensive residential energy consumption dataset of more than 6,000 homes, we develop a demand model for the estimation of half-hourly electricity demand for individual households based on a small set of household characteristics. The Adaptive Boost Regression Tree algorithm was presented as the best approach for our application. We contribute a new modelling method that substitutes household demographic survey data for the Mosaic demographic estimate dataset and evaluate the model on three case study precincts in New South Wales, achieving an R2 value of between 0.72 and 0.86, and the Lochiel Park Precinct, achieving an R2 value of 0.81. This is an improvement in fidelity over prior pioneering works in this space. To understand the impact of standalone residential solar and battery systems on demand and cost we developed a linear programming residential hybrid solar PV/battery/grid-connected power system model. We contributed an extension to modelling techniques by considering NPV energy costs, solar lifetime, commercially available modular battery sizes and DOD limits. The model estimates an upper limit (break-even) cost for homeowners for the installation of residential distributed energy storage. Retail energy price forecasts from the Australian Energy Market Operator have been applied to determine how the economics of residential solar and battery systems are impacted due to future energy price growth. The model has been applied to the measured and simulated demands of a case study in Sydney NSW, to show that the demand model can be applied to design solar and battery systems accurately and provide confidence to the future applications of the demand model. A new comprehensive microgrid design model using mixed-integer linear programming was developed to evaluate the design and impact of microgrid precincts. The model formulation allowed individual household demands to be considered within a single optimisation problem. The formulation facilitates the modelling of load diversity, network impacts, utility storage and distributed (household) storage in a microgrid. We applied this model to identify an optimal microgrid configuration for the Lochiel Park precinct in Adelaide, South Australia. We presented how the costs of off-grid microgrids are poor, although can be improved using a higher installed solar capacity. The microgrid model is then extended through integration with the demand model. In the model integration, we first develop individual household, half-hourly, load demand simulation, using information about the precinct. Next, we provide these demand profiles to the microgrid model for microgrid design, outputting the microgrid capacity levels, resource dispatch and energy costs. This modelling process delivers a new method to design and evaluate microgrid precincts where load demand data is unknown. Australian Energy price growth forecasts have been used to present how the business case for microgrids is impacted under different price growth scenarios. To explore voltage and power flow performance of microgrid solutions, we develop an AC and DC Digsilent PowerFactory microgrid model to simulate the Lochiel Park case study solution designed by the microgrid modelling tool chain. The network model is grounded in design data provided by the South Australian electricity distribution utility, SA Power, ensuring high-fidelity estimates of power line type and length, load locations, transformer ratings and transformer locations; providing a new level of detail not seen in other comparable research. The integration of demand, microgrid and PowerFactory microgrid models provides a pathway forward for the evaluation of the technical business case for AC and DC microgrid systems for residential precincts. By way of illustration, for the case study precinct, results showed that the network losses for the DC systems are lower than the AC systems. Moreover, we show that using distributed batteries result in lower network losses than utility battery storage due to less network power flow. Despite the lower losses of the distributed storage microgrid, our results showed that the lowest cost microgrid configuration for Lochiel Park, under the current pricing mechanism and available real-world battery (energy storage) capacity levels, was an AC microgrid using utility storage. The models developed in this thesis have integrated data science, economics and electrical engineering to provide a detailed scope of the total costs, emissions reduction potential, impacts on demand and impacts on the distribution network for future low carbon precincts. The methods presented in this thesis provide a new level of utility with the potential to be applied by developers, precinct planners, and local governments for new developments, helping identify the costs of and opportunities for a low carbon energy supply future enabled by contemporary distributed energy and microgrid solutions. Our focus is predominantly on Australian energy demand and microgrid opportunities, though the techniques and methodologies are applicable more broadly, and the findings present a basis for equivalent studies in other national contexts (where tariff, regulatory, environmental and other conditions may lead to different conclusions regarding optimal design and operation of microgrid solutions).
  • Item
    Thumbnail Image
    Medical image processing with application to psoriasis
    George, Yasmeen ( 2017)
    Psoriasis is a chronic, auto-immune and long-lasting skin condition, with no clear cause or cure. Psoriasis affects people of all ages, and in all countries. According to the International Federation of Psoriasis Associations (IFPA), 125 million people worldwide have psoriasis. The severity of psoriasis is determined by clinical assessment of affected areas and how much it affects a person's quality of life. The most common form is plaque psoriasis (at least 80% of cases), which appears as red patches covered with a silvery white build-up of dead skin cells. The current practice of assessing the severity of psoriasis is called "Psoriasis Area Severity Index" (PASI), which is considered the most widely accepted severity index. PASI has four parameters: percentage of body surface area covered, erythema, plaque thickness, and scaliness. Each measure is scored for four different body regions: head, trunk, upper-limbs, and lower-limbs. Although, PASI scores guide the dermatologists to prescribe a treatment, significant inter- and intra- observer variability in PASI scores exist, and are a fact of life. This variability along with the subjectivity and time required to manually determine the final score make the current practice inefficient and unattractive for use in daily clinics. Therefore, developing a computer-aided diagnosis system for psoriasis severity assessment is highly beneficial and long over due. Although, research in the area of medical image analysis has advanced rapidly during the last decade, notable advances in psoriasis image analysis and PASI scoring have been limited and only recently have started to attract the attention. In this thesis, we present the framework of a computer-aided system for PASI scoring using 2D digital skin images by exploring advanced image processing and machine learning techniques. From one side, this will greatly help improve access to early diagnosis and appropriate treatment for psoriasis, by obtaining consistent, precise and reliable severity scoring as well as reducing the inter- and intra- observer variations in clinical practice. From the other side, this can improve the quality of life for psoriasis patients. The framework consists of (i) a novel preprocessing algorithm for removing skin hair and side clinical markers in 2D psoriasis skin images, (ii) psoriasis skin segmentation method, (iii) a fully automated nipple detection approach for psoriasis images, (iv) a semi-supervised approach for erythema severity scoring, (v) a robust, reliable and fully automated superpixel-based method for psoriasis lesion segmentation, and (vi) a new automated scale scoring method using bag of visual words model with different colour and texture descriptors.