Enhancing Wi-Fi Sensing Performance Using Advanced Deep Learning Approaches

Abstract

Smart sensing technologies are increasingly integrated into our daily lives, creating demand for complementary sensing approaches that are unobtrusive, noninvasive, and capable of operating in complex indoor environments. Wi-Fi sensing has emerged as a promising solution as it leverages existing wireless infrastructure and relies on metrics such as Received Signal Strength (RSS) and Channel State Information (CSI) to capture spatial information and variations in signal propagation patterns. Existing methods can be broadly categorized as model-based or learning-based. Model-based approaches are interpretable but often limited by simplified assumptions, while learning-based approaches can capture complex features yet are hindered by noise, limited data, and environmental dynamics. To overcome these challenges and develop a more robust Wi-Fi sensing framework, it is necessary to both enhance the quality of CSI data and extract meaningful representations from it. In this thesis, we propose a simple yet effective Wi-Fi sensing method that combines Conditional Generative Adversarial Networks (CGANs) for CSI data augmentation with denoising and feature-extraction techniques based on Convolutional Neural Networks (CNNs). The proposed system increases the diversity of the training dataset, reduces the impact of noisy measurements, and improves the discriminative capability of learned features. Experimental evaluations demonstrate that our method achieves high recognition accuracy under limited-data conditions, surpasses baseline deep learning models, and maintains competitive performance on unseen data. These results indicate that integrating generative models with feature-enhanced learning-based techniques offers a promising direction for robust and practical Wi-Fi sensing applications.