Spatial-Temporal Attention Enhanced Deep Complex-Valued Convolutional Networks for Signal Detection in OFDM Communications

Accurate signal detection is crucial and also challenging in reliable orthogonal frequency division multiplexing (OFDM) communications. In this paper, we propose a spatialtemporal attention (STA) enhanced deep complex-valued convolutional network (DCCN), termed as the STA-DCCN, for signal detection, which performs well across various channel environments. In the STA-DCCN, we first employ the DCCN to extract local frequency-domain features. Secondly, the STADCCN enhances its feature extraction capability by incorporating the STA, which establishes interactive channels between temporal and frequency-domain features to adaptively and dynamically weight the feature space, thereby effectively suppressing interference and enhancing signal features from the received signals. The simulation results demonstrate that, under flat fading channel with the signal-to-noise ratios (SNRs) ranging from 0 dB to 20 dB, the STA-DCCN reduces the bit error rate (BER) by 35%$\mathbf{8 1. 8 2 \%}$ and 18.75%-76.92% compared with the LMMSE and the DCCN, respectively.