Based on Deep Neural Networks: A Machine Learning-Assisted Channel Estimation Method for MIMO Systems

This paper presents a machine learning-assisted approach for channel estimation in massive multiple-input multipleoutput (MIMO) systems, focusing on deep neural networks (DNNs) to enhance performance over traditional methods like least squares (LS) and minimum mean square error (MMSE). In 5G and beyond networks, accurate channel estimation is crucial for mitigating challenges such as pilot contamination and high mobility, which degrade system reliability. Our proposed DNN architecture, incorporating multi-layer perceptrons with ReLU activation, specifically consisting of 3 hidden layers (256,128, and 64 neurons per layer respectively), using Adam optimizer (learning rate 1e-4) and mean square error (MSE) loss function, learns from pilot signals to predict channel matrices, achieving lower normalized mean square error (NMSE) and bit error rate (BER) across various signal-to-noise ratio (SNR) levels. Simulations using COST 2100 public standard dataset (a widely recognized MIMO channel dataset for 5 G, instead of synthetic datasets), comprising 10,000 samples of $4 \times 4$ MIMO channels under urban macro scenarios, demonstrate that the DNN outperforms LS and MMSE by 3-5 dB in NMSE at medium SNR, with robust performance in high-mobility scenarios. The study evaluates metrics including NMSE vs. SNR, BER vs. SNR, and sensitivity to pilot length, antenna configurations, and computational complexity. Specifically, the DNN achieves 2.3 GFIOPs computational complexity, 15.6 k parameters, and 1.8 ms inference time on edge devices (Raspberry Pi 4), verifying deployment feasibility. Results indicate the DNN’s superiority in reducing estimation errors, particularly under noisy conditions, while maintaining feasible computational overhead. This work contributes to advancing ML integration in wireless communications, paving the way for efficient resource allocation and improved spectral efficiency in next-generation networks. Future extensions could incorporate more real-world datasets and hybrid architectures for even better generalization.