BEM-ANN Receiver With Linear Complexity for Severe Doubly-Selective Fading Channels

In orthogonal frequency division multiplexing (OFDM) communication over fast time-varying frequency-selective fading channels, also known as doubly-selective channels (DSC), channel estimation (CE) and OFDM demodulation are challenging tasks. This is because high Doppler spread in DSC introduces severe inter-carrier interference (ICI) that distorts the OFDM data, and spreads the pilot tones away from their designated frequency positions to make channel estimation inaccurate. In this paper, we present a robust OFDM communication receiver for severe DSC, which incorporates a novel deep learning (DL) based channel estimator (CE) and a two-stage equalization architecture named 2s-BEM-ANN. The proposed 2s-BEM-ANN architecture operates in two stages. In the first stage, a pilot-aided CE is performed using an attention neural network (ANN) known as BEM-ANN, followed by equalization using a banded-minimum mean square error (BMMSE) algorithm. In the second stage, data-aided CE is carried out using another BEM-ANN, followed by maximum-likelihood (ML) equalization. Both channel estimators model the DSC using a complex-exponential basis expansion model (CX-BEM), and two training strategies are proposed for the pilot-aided and data-aided BEM-ANNs. Simulation results show that the proposed 2-stage basis expansion model attention neural network (2s-BEM-ANN) receiver significantly improves the bit-error rate (BER) performance over the conventional CE and equalization techniques. Furthermore, the BER performance of the proposed 2s-BEM-ANN approaches very close to the BER lower bound, and without requiring any turbo iteration between the CE and equalization modules. The proposed 2s-BEM-ANN is also robust against signal-to-noise ratio (SNR) and Doppler uncertainties, as well as variations in the channel power-delay profile. Finally, the computational complexity of the proposed 2s-BEM-ANN grows only linearly with the OFDM block length $N$.