CNN-RAGNet Architecture for CFO Estimation in RIS-Assisted MIMO-OFDM Systems

Abstract

This letter presents a deep learning (DL) supervised model of estimating carrier frequency offset (CFO) for reconfigurable intelligent surfaces (RIS)-assisted multiple input multiple output-orthogonal frequency division multiplexing (MIMO-OFDM) systems without the channel state information. The proposed architecture consists of the convolution neural network (CNN) with enhanced residual (Res), attention and dense gated linear unit (GLU) blocks, collectively referred to as CNN-RAGNet architecture. The integration of enhanced Res block facilitates feature extraction from various received antenna samples and mitigates the vanishing gradient problem. The attention and D-GLU blocks are incorporated into the model to prioritize relevant features and enhance the CFO estimation accuracy. Furthermore, the proposed architecture is adaptable to various modulation schemes and RIS elements, and works on the realistic 3GPP TR38.901 tapped delay line channel model. The simulation results indicate its outperformance over existing statistical based methods and DL based approaches. The proposed architecture has lower computational complexity than the existing methods.