Masters Degrees (Electrical and Electronic Engineering)

Permanent URI for this collection

https://scholar.sun.ac.za/handle/10019.1/766

Browse

Browsing Masters Degrees (Electrical and Electronic Engineering) by browse.metadata.advisor "Botha, Matthys"

Now showing 1 - 2 of 2
  • Thumbnail Image
    Item
    The application of machine learning for computational electromagnetic solver selection
    (Stellenbosch : Stellenbosch University, 2023-03) de la Bat, Willem; Botha, Matthys; Ludick, Danie; Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering.; Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering.
    ENGLISH ABSTRACT: The field of Computational Electromagnetic (CEM) encompasses a number of methods for resolving the electromagnetic (EM) response of conducting objects with arbitrary shapes and sizes. The specific geometry and configuration of the problem, along with substantial user expertise, are usually required in order to choose the most effective and accurate CEM solver techniques. The process of selecting the optimal techniques can therefore become highly time-consuming. By utilizing Machine Learning (ML) models to predict, prior to simulation, whether a specific CEM technique will be sufficiently accurate or not, we are able to streamline the antenna design process. In this work, this approach was applied to the hybrid Method of Moments (MoM) and Single Reflection Physical Optics (SRPO) technique, the hybrid MoM and Multiple Reflection Physical Optics (MRPO) technique as well as the Domain Green’s Function Method (DGFM) technique. The first two focusing on feed-reflector type antenna problems and the latter on regular and irregular antenna arrays. Using ML algorithms that include the Decision Tree Classifier (DTC), Logistic Regression Classifier (LRC), Artificial Neural Network Classifier (ANNC) and the Random Forest Regressor (RFR), amongst others, it is shown that it is, in fact, possible to make highly accurate predictions of the accuracy of these CEM techniques, prior to simulation.
  • Thumbnail Image
    Item
    Classification of synthesised ISAR images of small complex targets
    (Stellenbosch : Stellenbosch University, 2023-03) Stewart-Burger, Cullen; Ludick, Danie; Botha, Matthys; Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering.
    ENGLISH ABSTRACT: This study examines the application and comparison of several machine learning algorithms to the problem of classifying inverse synthetic aperture radar (ISAR) images of electrically small, geometrically complex targets. These algorithms include k-nearest neighbours, logistic regression, a fully connected neural network and a capsule network. A novel classifier is proposed, utilizing a capsule network with a reconstruction sub-network, to perform open-set classification. A dataset of synthetic ISAR images was created from simulated electromagnetic (EM) target returns and used to train and test the models. The EM simulation process was performed using a method of moments solver to compute the backscattering from models of the targets, which are represented as triangular meshes. Specifically considering geometrically small targets with low radar cross-sections, particular attention is paid to the performance of the classifiers when signals are received in the low signal-to-noise ratio regime. The use of a capsule network is found to be highly effective for both closed-set and open-set classification tasks, out-performing the other traditional machine learning and deep learning based classifiers investigated in this study (logistic regression, support vector machines, k-nearest neighbours and fully connected neural networks). The proposed method of comparing the images formed by the capsule network’s reconstruction subnetwork to the input image is demonstrated to be an effective technique for identifying observations of ISAR images of targets that do not belong to any known classes, i.e. targets which are “unknown” to the classifier. Additionally, it is demonstrated that the use of the zero-mean normalised cross-correlation coefficient to compare the input and reconstructed images makes the proposed open-set recognition method more resilient to noisy inputs when compared to the use of the mean-squared error between the images. This addresses a commonly overlooked problem that an operational radar’s automatic target recognition algorithm is not guaranteed to have been trained for all the possible target types that it will sense in the surveillance volume. The proposed classifier achieves an F1-score of greater than 0.9 for a test set containing two known and two unknown classes with signal-to-noise ratios of 6 dB and above.