Hybrid Signal Processing and Machine Learning Techniques for Robust Wireless Channel Estimation

Abstract

Channel state information (CSI) is the key to high-quality data gaining and spectral efficiency in contemporary wireless systems based on a MIMO-OFDM, especially in the conditions of a noisy and time-varying channel propagation. The standard channel estimators like the Least Squares (LS) are inexpensive to compute but very sensitive to noise compared to minimum mean square error (MMSE) estimation which is more accurate but much more costly in terms of computation and requires prior statistical information. All-data based deep learning methods are potentially good, but lack physical explanation and highly need generalization to different signal-to-noise ratio (SNR) situations. The paper is a proposal of a hybrid signal processing and machine learning system of robust wireless channel estimation. The approach is a combination of model-driven pre-estimation using LS with the light-weight residual-learning neural network which further details the structured estimation errors. The hybrid model maintains a physical basis of the classical estimators but adds resilience by correcting predictions using data. It is under Rayleigh and Richardson fading channels whereby extensive Monte Carlo simulations are done within a SNR of 0-30 dB. The findings prove that the recommended method limits the normalised mean square error (NMSE) 28 times smaller than LS and performance as good as MMSE at significantly reduced computational complexity. Further tests with pilot density reduction, Doppler mobility and SNR mismatch are found to have better generalization and stability. The system proposed provides an efficient and scalable system of next-generation wireless communication systems.