Channel Estimation in RIS-Assisted LoRa Systems via Lightweight Residual CNNs

Abstract

Accurate channel state information (CSI) is critical for reconfigurable intelligent surface (RIS)-assisted LoRa systems. Due to the severe channel fading in long-range transmissions, the received LoRa signals have small strength, thus resulting in unsatisfactory channel estimation (CE) performance during convolutional neural network (CNN) training. To properly optimize the network parameters during training, a data pre-processing (DP) method is proposed to pre-process the datasets consisting of small values. Moreover, considering the energy normalization of LoRa pilots, a lightweight residual CNN (LRCNN) is proposed for CE in RIS-assisted LoRa systems. Simulation results show that DP method can significantly improve the performance of CNN-based estimators. Furthermore, the proposed LRCNN can obtain more accurate CSI than the traditional least squares (LS) and linear minimum mean square error (LMMSE) estimators under different spreading factors (SFs) of LoRa modulation. Compared to the existing CNN-based estimators, the proposed LRCNN performs better with lower network complexity.