Channel Statistics-Aware Unsupervised Learning-Based Massive MIMO Channel Estimation

Channel estimation (CE) in massive MIMO systems, essential for wireless communications, faces growing challenges as antenna numbers rise and environments become more complex. Conventional methods like MMSE incur high computational complexity, while supervised deep learning (DL) struggles with labeled data scarcity in real-time scenarios. To tackle these issues, we propose an unsupervised learning (UL)-based method, which exploits channel correlations through a symmetric autoencoder (AE) design. Incorporating statistical feature extraction, it overcomes labeled data dependency for efficient, real-time CE. By incorporating statistical channel properties, our method reduces data dependency, enabling efficient and real-time CE. To maximize its potential, we introduce a two-phase training approach: offline joint training captures statistical features from unlabeled data, and online deep image prior (DIP)-inspired denoising refines performance. Therefore, our method eliminates labeled data requirements, offering a practical opinion for real-time CE. Experimental results confirm the superior performance with reducing computational cost, and excellent low-SNR performance for real-time applications of our proposal.