CNN-Based Iterative Decoding for Polar Codes in MIMO Systems: Performance Analysis and Computational Complexity Evaluation

Abstract

With the continuous development of 5G technology, wireless communication systems demand higher spectral efficiency and robust error correction. In this study, we propose an iterative decoding algorithm for polar codes in MIMO systems, based on a convolutional neural network (CNN). The CNN architecture comprises three 1D convolutional layers with ReLU activation, a depthwise separable convolution layer, and two fully connected layers. Training was conducted using a synthetic dataset of 100,000 samples generated under AWGN, Rayleigh, and Rician channel models, with 80% used for training and 20% for validation. Evaluation metrics include bit error rate (BER), throughput (Mbps), and computational complexity (ms/symbol). Compared to BP, SC, and SCL algorithms, our model achieves a 30% reduction in BER at 6 dB SNR, throughput improvements of up to 25%, and reduced processing latency across 2×2, 4×4, and 8×8 antenna configurations. The experimental results show that the proposed algorithm exhibits better performance than traditional decoding methods under varying SNR levels, channel models, antenna configurations, and data rates. While not entirely eliminating complexity, the model leverages depthwise separable convolution to reduce parameter size and training overhead, making it more efficient than conventional iterative decoders. This research provides a promising step toward resolving the open challenge of designing decoders that balance adaptability and computational feasibility in MIMO systems.