Abstract
In high-mobility communication scenarios, leveraging reconfigurable intelligent surfaces (RISs) to assist orthogonal time frequency space (OTFS) systems proves advantageous. Nevertheless, the integration of RIS into OTFS systems increases the complexity of channel estimation (CE). Utilizing the benefits of machine learning (ML) to address such intricate issues holds the potential to reduce CE complexity. Despite this potential, there is a lack of investigations of ML-based CE in RIS-assisted OTFS systems, leaving significant gaps and posing challenges for intelligent applications. Moreover, ML-based CE methods encounter numerous difficulties, including intricate parameter tuning and long training time. Motivated by the inherent advantages of the single-hidden layer feed-forward network structure, we introduce extreme learning machine (ELM) into RIS-assisted OTFS systems to improve CE accuracy. In this method, we incorporate a threshold-based approach to extract initial features, aiming to remedy the inherent limitations of the ELM network, such as inadequate network parameters compared to the deep learning network. This initial feature extraction contributes to an enhanced ELM learning ability, leading to improved CE accuracy. Applying the classic message passing algorithm for data symbol detection, simulation results demonstrate the effectiveness of the proposed method in improving the symbol detection (SD) performance of RIS-assisted OTFS systems. Furthermore, the SD performance exhibits its robustness against variations in modulation order, maximum velocity, and the number of sub-surfaces.