Deep Learning for Reliable RIS-Enabled SISO Localization Under Practical Model Mismatch

Abstract

Reconfigurable Intelligent Surfaces (RISs) have shown strong potential for improving user localization in single-input single-output (SISO) systems, where both the base station (BS) and the user equipment (UE) are equipped with a single antenna. However, in practical wireless environments, localization accuracy can be affected by signal-to-noise ratio (SNR) variations, multipath propagation, and hardware-related impairments. These factors may create a mismatch between the assumed system model and the received signals, thereby reducing the performance of conventional model-based localization estimators. To address this issue, this work proposes a mixed-domain deep learning approach for three-dimensional downlink localization in RIS-enabled SISO systems. The proposed method constructs a learning observation tensor from OFDM pilot measurements by transforming the received signals into the delay domain using the inverse fast Fourier transform (IFFT), followed by appropriate normalization. The network is trained under different SNR levels and impairment conditions to improve its robustness against practical model mismatches. Simulation results show that the proposed method improves localization accuracy, as measured by root mean square error (RMSE), and enhances reliability, as measured by the 90th percentile error (p90). Compared with conventional model-based estimators, the proposed approach provides more stable performance across different SNR values and multipath scenarios, particularly in reducing the number of high-error cases.