Doctoral Degrees (Electrical and Electronic Engineering)

Permanent URI for this collection

https://scholar.sun.ac.za/handle/10019.1/327

Browse

Recent Submissions

Now showing 1 - 5 of 226
  • Thumbnail Image
    Item
    On Eigendecomposition-based algorithms as feature extraction techniques used with hidden Markov model for the detection of whale vocalisations
    (Stellenbosch : Stellenbosch University, 2024-03) Usman, Ayinde Mohammed; Versfeld, Daniel Jaco J.; Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering.
    ENGLISH ABSTRACT: Whales emit a variety of distinctive sound signals for communication, echolocation, and other social functions, which are gathered through passive acoustic monitoring (PAM). Different automated methods have been proposed in the literature for analysing PAM datasets to detect and classify whale species, including the use of the hidden Markov model (HMM). This thesis proposes eigendecomposition-based (ED) algorithms as feature extraction (FE) techniques used with HMM for the detection of whale vocalisations. Specifically, the principal components analysis (PCA) and the dynamic mode decomposition (DMD) are deployed to extract the latent underlying characteristics of whale signals in PAM datasets. In addition, enhanced FE techniques are proposed through the kernelisation of PCA and DMD. The emerging ED-based hidden Markov models (ED-HMMs): PCA-HMM, kPCAHMM, DMD-HMM, and kDMD-HMM are grouped according to the underlying algorithm deployed for the FE: the PC-based hidden Markov models (PC-HMMs) and the DMD-based hidden Markov model (DM-HMMs). Each of the models is tested on PAM datasets containing southern right whale (SRW) and humpback whale (HW) vocalisations. Their performances are evaluated using metrics such as the true positive rate (TPR), precision (PREC), error rate (ERR), and F1 scores. Performance outcomes vary subject to different experimental conditions like the dimension of the feature vectors, the size of the training data, and the species vocalisations. The models demonstrated good performance across different evaluation metrics. For the PC-HMMs, the kPCA-HMM did not only outperform the PCA-HMM in terms of TPR and PREC, but it also exhibited a lower ERR. However, the kCPA-HMM exhibits a higher computational cost when compared to the PCA-HMM. Similarly, for DM-HMMs, the kDMD-HMM outperformed the DMD-HMM in terms of TPR and PREC, and it also exhibited lower ERR as well as a lower computational cost. The comparison showed that PC-HMMs stabilised faster than DM-HMMs in terms of performance. Thus, the PC-HMMs are less complex than the DM-HMMs in terms of dimension. However, the DM-HMMs outperformed the PC-HMMs, albeit at higher https://scholar.sun.ac.za Abstract iii dimensions. The reliability of the developed models was confirmed with F1 scores, as all the models achieved F1 scores > 0.9 at their respective optimal dimensions. Lastly, the results of the proposed ED-HMMs are compared with the existing FE techniques used with HMM in the literature for the detection of whale vocalisations. The ED-HMMs do outperform the existing HMM methods. A general observation is that every model displays better performance with an increase in the number of samples deployed for training. Hence, large window sizes are recommended for model training. The different experimental results showed that a model’s performance must be evaluated on a species-to-species basis. It is also important that the training data be a subset of the datasets for testing, or at least using recordings from the same region. This is to avoid bias that may arise from the variation that does exist between the vocalisations of the same species. The ED-HMMs proposed in this study can be further tested on other whale vocalisations to confirm their robustness. Besides, they can be explored by researchers working on the automatic detection of other vocalising animal species.
  • Thumbnail Image
    Item
    How to automatically detect calls from cetaceans and fish without large-scale training data
    (Stellenbosch : Stellenbosch University, 2024-03) Van Wyk, Jacques; Versfeld, Jaco; Du Preez, Johan; Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering.
    ENGLISH ABSTRACT: Passive acoustic monitoring has become pivotal in the acquisition of vocalisations produced by marine wildlife, however, there is a need for suitable detection methods that can segment these vocalisations from the many hours of noise present in the retrieved audio recordings. Automatic detection methods can aid in this, labelling potential animal calls in bioacoustic recordings faster than manual annotation. One problem modern automatic detection methods suffer from, particularly deep learning methods, is overfitting due to small training sets. This calls for detection methods that can quickly and accurately detect bioacoustic vocalisations without the need for copious amounts of training data. Such detection methods are proposed here for the automatic detection of cetacean and fish vocalisations in underwater recordings. The first method is a silence detector based on a temporal power calculation that can be used as a pre-processing step to quickly find potential target calls. The harmonic structure of vocalisations is then exploited to aid in their detection. The second method is based on the summation of sound harmonics, with a normalisation scheme that provides robust performance in the presence of wide-band noise. The third method extends this normalisation scheme to spectrogram masking, a faster alternative to spectrogram cross-correlation. After this the use of unsupervised deep learning methods are investigated, in an attempt to determine their ability to detect cetacean vocalisations in the presence of imbalanced data.
  • Thumbnail Image
    Item
    Enhancing customer engagement in e-commerce: improving e-marketing open rates through model-free reinforcement learning.
    (Stellenbosch : Stellenbosch University, 2024-03) Grobler, Abraham; Engelbrecht, Herman; Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering.
    ENGLISH ABSTRACT: This thesis aims to optimise the delivery time of e-marketing methods such as emails and push notifications, with the intention of increasing customer engagement with an e-commerce platform. This optimisation can be performed using model-free reinforcement learning (RL) methods. First, we aim to develop a statistical, non-stationary model of a customer’s probability to interact with e-marketing at different hours of the day. The model is built using a small sample of anonymous, real customer data. From this sample, we train a Gaussian Mixture Model, which allows us to generate a large synthetic customer base. This customer model acts as the environment of the RL experiments. We then develop several different RL agents, employing algorithms such as Q-learning and DQN, to try and find the best time to deliver e-marketing messages to each customer. We then compare the different agents in terms of learning rate, adaptability and stability. A novel method for epsilon-greedy exploration, tailored to each customer through a parameter-specific approach, is also proposed and tested. Our experiments demonstrate that this method outperforms traditional exploration techniques in the context of our experiments. Our findings demonstrate that RL-based optimisation of delivery time provides a promising method of potentially increasing the open rate and customer engagement, providing valuable insights for e-commerce platforms.
  • Thumbnail Image
    Item
    Autonomous guidance and conflict avoidance for multiple Unmanned Aerial Vehicles (UAVs) in urban environments
    (Stellenbosch : Stellenbosch University, 2024-03) Hughes, Merrick; Engelbrecht, Japie; Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering.
    ENGLISH ABSTRACT: This thesis presents an autonomous guidance and conflict avoidance system for multiple unmanned aerial vehicles (UAVs) in urban environments. The guidance system comprises a long-term route planner and a short-term cooperative conflict avoidance function. The route planner generates long-term routes for the UAVs to follow to satisfy their missions, whereas the conflict avoidance function continuously predicts and resolves short-term impending collisions/conflicts with terrain and other UAVs. These conflicts are resolved cooperatively between all UAVs involved while attempting to minimise each UAV’s deviation from its long-term route during conflict avoidance. The guidance system is conceptualised to support multiple independent UAVs flying within an urban environment that contains terrain as well as wind. A 3D model of real urban terrain is implemented using established computational geometry techniques and the motion constraints of the UAVs are set to replicate those of real-life UAVs. Monte Carlo simulations are designed to evaluate and compare the performance of the guidance system in different scenarios by analysing both illustrative and statistical results. The illustrative results indicate that the route planner and conflict avoidance function behave as expected in given scenarios. The statistical results confirm that the performance of the route planner decreases at lower altitudes and also suggest that it may struggle with grid-like terrain, while the failure rate of the conflict avoidance function does not appear to be strongly correlated with any given scenario. A combined set of avoidance manoeuvres is the most effective for conflict avoidance but has the longest execution times, whereas airspeed manoeuvres used in isolation are the least effective and horizontal manoeuvres are the least optimal with the largest average deviations from a UAV’s long-term route. A single avoidance manoeuvre type used in isolation exhibits promising execution times for a real-time application.
  • Thumbnail Image
    Item
    Open-set learning with augmented category by exploiting unlabelled data (Open-LACU)
    (Stellenbosch : Stellenbosch University, 2024-03) Engelbrecht, Emile; Du Preez, Johan ; Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering.
    ENGLISH ABSTRACT: Neural network classifiers provide scalable means to analyse categorical patterns within datasets. However, current machine learning policies fail to consider certain nuances developed in real-world applications. The vast number of patterns represented in certain datasets and the continual collection of new data means classifiers must be aware of the observed-novel category and the unobserved novel category. To address these challenges, this dissertation combines semi-supervised learning and novelty detection into a single learning framework called open-set learning with augmented category by exploiting unlabelled data or Open-LACU. Although Open-LACU requires further development, we show and argue that Open-LACU classifiers will have reduced annotation cost, improved practicality and enhanced safety. Semi-supervised learning trains models using partially labelled datasets to reduce annotation costs. Novelty detection ensures classifiers are able to separate all data samples outside the domain of interest for enhanced safety. When working with partially labelled datasets in a domain where novel patterns exist, several inconsistencies appear in existing literature. More specifically, there is no distinction between those novel patterns which are unrepresented during training but appear during testing, and those novel patterns that are represented in unlabelled training data. Considering the unique properties of these different novel category types, we argue that classifiers must generalise these separately. In Open-LACU, classifiers must generalise 1) those K > 1 number of source categories for which labels are provided, 2) an additional K + 1’th observed-novel category for those novel patterns in the unlabelled training data, and 3) an additional K + 2’nd unobserved-novel category that encapsulates all those novel patterns unobserved during training but seen during testing. To introduce Open-LACU, we pursue several objectives that integrate different learning frameworks. For each of these integrating steps, we experiment on small-scale vision datasets to simulate different categorical scenarios. Our results both confirm the feasibility of Open-LACU and reveal several insights into the challenges that future research must address.