Novel Over-the-Air Federated Learning via Reconfigurable Intelligent Surface and SWIPT

To provide sustainable energy support while meeting the demands for serving a rapidly increasing number of devices, in this paper, we propose a new Reconfigurable intelligent surface (RIS) Assisted simultaneous wireless information and Power transfer (SWIPT) and over-the-aIr computation (AirComp) feDerated learning system which is termed RAPID. Specifically, at each training iteration, an access point (AP) first simultaneously broadcasts global model and transfers wireless energy to the selected devices via the RIS-assisted SWIPT. These devices then use the harvested power to compute and upload their local gradients to the AP for aggregation via the RIS-assisted AirComp. To identify the performance improvement to federated learning, we first analyze and derive the expected convergence rate for the proposed RAPID system taking into account factors of device selection and wireless communication. We then formulate a joint learning-communication optimization problem in terms of device selection, transmit beamforming, power splitting, receive beamforming, and RIS coefficients design. To solve the formulated non-convex problem, we propose a new two-stage algorithm by successively solving the downlink SWIPT and the uplink AirComp sub-problems based on alternating optimization and successive convex approximation techniques. Simulation results are presented to demonstrate that our proposed RAIPD system can significantly improve the convergence and accuracy of federated learning compared with benchmark schemes.