With the rapid advancement of digital technologies, wireless communication and sensing systems have become increasingly integral to our daily lives. These systems utilize wireless signals not only as data carriers but also as a medium for radio sensing. Model-based approaches have traditionally been a popular choice for addressing existing challenges in communication and sensing. However, model-based approaches struggle to accurately characterize signal propagation, especially at higher frequencies, and optimizing them for communication is even more difficult. Moreover, extracting human motion-related information from these complex signals is often challenging with conventional methods. Recent progress in artificial intelligence (AI) has opened new avenues for addressing these challenges. This thesis explores learning-based approaches to uncover the hidden information embedded within wireless signals. By doing so, it aims to enhance the efficiency of wireless communication systems and enable fine-grained human motion sensing, thereby pushing the boundaries of wireless systems.The first part of this thesis explores the capability of various RF signals to sense different levels of human motion using learning-based approaches. We begin by proposing AuthIoT, a gesture-based wireless authentication scheme designed for IoT devices. AuthIoT leverages a convolutional neural network (CNN) to learn human gesture features from Wi-Fi channel state information (CSI) and maps them to specific letters for device authentication. To enhance robustness and enable gesture recognition across diverse environments, the system employs a feature fusion approach that integrates location-independent features, ensuring strong transferability. Next, we shift our focus to tiny motions and propose RadSee, a system capable of recognizing fine-grained handwriting. We develop a 6 GHz FMCW radar system along with a tailored deep neural network to identify handwritten letters through walls. The model combines a bidirectional long short-term memory (BiLSTM) network with an attention mechanism to leverage temporal dependencies and capture critical features—such as turning points—in radar phase sequences for accurate recognition. We push the limits of this system further with a novel learning framework and introduce RadEye, a system designed to recognize eye movements. Given the subtle nature of eye motion and the challenge of detecting it in RF signals, we adopt a transformer encoder as the feature extractor to more effectively exploit temporal dependencies in the phase sequences. To further enhance performance, we incorporate a state-of-the-art vision-based method to provide guidance and prior knowledge during the learning process. The second part of this thesis focuses on leveraging learning-based solutions to improve the efficiency of wireless communication systems, with particular emphasis on enhancing the throughput of mmWave communication systems. We begin by proposing an uplink multi-user MIMO (MU-MIMO) mmWave communication (UMMC) scheme for WLANs. MU-MIMO techniques are well-known for increasing network efficiency and throughput. A key innovation in this work is a learning-based Bayesian optimization (BayOpt) framework for joint beam search across multiple antennas. This approach eliminates the need for complex channel modeling and identifies optimal beamforming directions with only a few search iterations, significantly reducing beamforming overhead. We then further explore the beamforming problem in mmWave communications, shifting our focus to mobile mmWave networks. In such dynamic environments, beamforming overhead becomes more pronounced. To address this challenge, we leverage the temporal correlation of wireless channels to aid in beam selection. Specifically, we propose a Temporal Beam Prediction (TBP) scheme that enables a mobile mmWave device to predict its future beam direction based on its historical beam selection profile. At the core of this scheme is a modified LSTM architecture, complemented by an adversarial learning model to improve the robustness and generalizability of the beam steering process. This thesis presents efficient communication schemes and novel sensing applications based on learning-driven approaches, paving the way for the design of AI-enabled next-generation wireless communication and sensing systems. It provides detailed descriptions of system implementations, experimental setups, and performance evaluations of the proposed schemes in real-world environments. Furthermore, it offers an in-depth analysis of the limitations of these systems and discusses open challenges in developing future wireless communication and sensing systems using learning-based techniques.
Read
