Machine learning enabled integrated sensing and communication for next generation wireless networks

Integrated sensing and communication (ISAC) has been identified as a key technology for the sixth-generation (6G) wireless networks. It enables wireless networks to simultaneously transmit information and receive sensing echoes through a unified infrastructure and shared resources, thus improving both spectrum and energy efficiencies. This thesis investigates the interplay of sensing and communication functionalities in ISAC systems and explores machine learning approaches to facilitate end-to-end system design. First, we study sensing-enabled communications in vehicular networks. We investigate the problem of ISAC-assisted beamforming prediction to maintain communication links between the roadside unit (RSU) and vehicles. Based on the collected sensing signals reflected from the vehicles, we leverage machine learning techniques for an end-to-end predictive beamforming design, which directly maps the sensing signals to the beamformers without estimating the state parameters of the vehicles or the channel state information (CSI). The proposed ISAC-enabled predictive beamforming design can reduce the signaling overhead and enhance the communication performance. Then, we study communication-assisted sensing and develop a cooperative ISAC framework for target sensing. By using cell-free multiple-input multiple-output (MIMO) architecture, we consider multiple distributed access points (APs) collaboratively performing multistatic sensing under the control of a central processing unit (CPU). To tackle the fronthaul overhead issue between the CPU and each AP, we propose a collaborative learning approach which enables APs to locally preprocess the sensing signals, thus reducing the volume of data transmitted over the fronthaul links while ensuring that useful sensing information is obtained by the CPU. Simulation results show a significant sensing performance improvement over existing approaches with a lower fronthaul signaling overhead. Finally, we investigate the joint optimization of the communication and sensing performance from the transmitter perspective, and explore a new architecture by using movable antennas to fully exploit the spatial degrees of freedom (DoFs). We propose a heterogeneous graph neural network (GNN)-based approach to capture the spatial relationships of distributed AP antennas and extract features from sensing and communication channels. Simulation results demonstrate the advantages of the proposed GNN-based approach over existing deep neural network (DNN) approaches, and the performance improvement brought by movable antennas.