AI-driven enhancement and risk detection in microwave systems

Abstract

This thesis investigates advanced deep learning techniques to improve the accuracy, reliability, and interpretability of microwave sensing systems for subsurface object detection and classification. The work is structured around three main contributions, each demonstrating the integration of neural network-based models to address different challenges in microwave imaging and detection. The first contribution focuses on the application of convolutional neural networks (CNNs) for breast tumor detection. The proposed CNN framework is capable of simultaneously detecting and localizing tumors by accurately estimating their center coordinates and radii. In addition, the network predicts the dielectric permittivity (𝜖r) of each identified object, enabling classification into distinct material categories. By leveraging the rich spatial and spectral information contained in microwave data, this approach achieves robust and precise object detection and classification. Simulation results confirm the efficacy of the CNN in extracting both geometric parameters and material properties, highlighting its potential for experimental breast cancer screening and non-invasive medical diagnostics. The second contribution addresses the enhancement of microwave inverse imaging using deep learning. Inverse scattering problems often suffer from noise and non-linearities, which degrade the quality of reconstructions. To overcome these challenges, a novel deep neural network model is introduced that refines multi-frequency qualitative indicators derived from the Linear Sampling Method (LSM). Scattered field data are simulated for various dielectric object configurations using a two-dimensional method of moments and an antenna array, and then processed via the LSM to generate input indicators. The proposed network, comprising eleven layers of convolutional and deconvolutional operations without dropout or pooling layers, effectively reduces noise while preserving essential structural information. Quantitative evaluation using the Jaccard index demonstrates that this approach improves reconstruction accuracy and quality. The combination of a simple yet effective neural network with LSM results highlights the method's ability to enhance inverse imaging performance in complex environments. The third contribution presents a lightweight neural network integrated with a microwave-based detection system for identifying buried objects in real-world scenarios. The model is trained and validated exclusively on experimental measurements, ensuring practical relevance and robustness. S-parameter measurements from a 16×16 antenna array are transformed into a 256-dimensional feature vector that captures the microwave response of subsurface materials. This compact representation allows for a computationally efficient network architecture while maintaining high detection accuracy. Experimental results demonstrate exceptional performance, achieving an accuracy of 99%, an F1 score of 0.99, and a recall of 0.98, outperforming standard CNN, DRN, and EfficientNet architectures. These findings illustrate the model's suitability for defense and security applications, where rapid and accurate threat identification is critical. Overall, this thesis demonstrates the substantial benefits of integrating deep learning techniques into microwave sensing workflows. By combining CNN-based tumor detection, neural network-based enhancement of LSM indicators, and efficient classification of real-world S-parameter data, the work provides novel methodologies for both object detection and material characterization. The results underscore the potential of machine learning to advance electromagnetic sensing systems, improving detection performance, robustness to noise, and interpretability of microwave imaging data. Collectively, these contributions provide a foundation for future developments in both biomedical and security focused subsurface sensing applications.