Feature Purification Using Extreme Learning Machine for RIS-ISAC Channel Estimation

Abstract

Integrated sensing and communication (ISAC) technology enables the joint integration of communication and sensing functions through the efficient utilization of spectrum/energy resources. By further incorporating with the reconfigurable intelligent surfaces (RISs), the wireless propagation environment of ISAC systems can be dynamically controlled, thereby enhancing the overall sensing and communication (SAC) performance. In this context, accurate channel estimation is a fundamental prerequisite for the efficient operation of the RIS-assisted ISAC systems. However, the strong coupling between SAC signals from direct and reflected channel links, as well as severe noise interference, limits the SAC channel estimation accuracy. This paper proposes a novel two-stage channel estimation scheme for RIS-assisted ISAC systems, where the direct and reflected SAC channels are respectively estimated in the first and second stages. To mitigate the negative effects of noise interference in each estimation stage, an extreme learning machine-based feature purification module is precisely designed, improving the quality of received SAC signals and generating purified channel features. Then, the dedicated deep neural network adopts the purified channel features to estimate SAC channels. Simulation results demonstrate that, under different signal-to-noise ratio conditions and channel dimensions, the proposed scheme achieves superior estimation accuracy and strong robustness compared to the benchmark methods.