WeFTNet: DNN-Aided Weighted Fourier Transform Approximate MMSE Detection for Massive MIMO

Abstract

The minimum mean square error (MMSE) detector is widely adopted for massive MIMO systems, but its reliance on matrix inversion results in prohibitive computational complexity. To solve this problem, linear iterative detectors, e.g., weighted Neumann series approximation (wNSA), can circumvent matrix inversion and substantially reduce complexity. However, their performance degradation in high-dimensional high-order MIMO scenarios remains unacceptable. In this paper, we propose WeFTNet, a deep neural network (DNN)-aided weighted Fourier transform (FT)-based approximate MMSE detector for massive MIMO systems. Specifically, the proposed WeFTNet integrates two unexplored features: weighted FT-based MMSE approximation and DNN-driven weighting factor optimization. Furthermore, a geometrically decreasing weighting factor model is theoretically derived and introduced to enhance the training stability and performance robustness. Numerical results demonstrate that the WeFTNet consistently achieves near-MMSE bit error rate (BER) performance in various MIMO scenarios. Compared with the other linear iterative detectors, the proposed WeFTNet succeeds in delivering up to 3.5 dB performance gains while achieving 57% complexity reduction in $256\times128 \ 256$ -QAM MIMO systems.