Attention-Based Deep Learning for Pilot-Assisted Channel Estimation and Signal Detection in 6G Massive MIMO Networks

Accurate channel estimation and reliable symbol detection are essential for exploiting the spatial multiplexing gains of massive multiple-input multiple-output (MIMO) wireless systems. This paper investigates an attention-based deep-learning approach for refining pilot-assisted least-squares (LS) channel estimates in an 8 × 8 MIMO simulation. The simulated link employs a Rayleigh fading channel, additive white Gaussian noise (AWGN), and 16-quadrature amplitude modulation (16-QAM). The proposed approach refines the LS estimate using an attention-based neural network, after which zero-forcing (ZF) detection is applied to recover the transmitted symbols. Its performance is evaluated against LS, minimum mean-square-error (MMSE), and feed-forward neural network (FFNN) baselines over signal-to-noise ratios (SNRs) from 0 to 20 dB. The reported experiment uses 3,000 training samples, 500 testing samples, and 15 training epochs. At 20 dB SNR, the attention-based model achieves a bit error rate (BER) of 0.22525 and a symbol error rate (SER) of 0.39663, compared with 0.43219 and 0.68150, respectively, for the FFNN baseline. However, the attention-based model does not outperform MMSE across the evaluated SNR range, and its BER improvement becomes limited at higher SNRs. It also performs worse than LS at the higher evaluated SNR points. These findings indicate that attention-based refinement can improve upon the evaluated FFNN baseline, but the current implementation does not establish superiority over conventional estimators. Further validation using repeated trials, a fully documented network architecture, and controlled comparisons is required to assess its robustness and generalisation.