Abstract
In this paper, an unsupervised deep learning-based framework based on dual-path model-driven auto-encoders (AE) is proposed for angle-of-arrivals (AoAs) estimation in massive MIMO systems. Specifically designed for AoA estimation, the proposed framework differs from the conventional AE in two aspects. Firstly, unlike conventional auto-encoders, our framework employs a dual-path neural network for the encoder, decoupling the estimated parameters and enabling independent updates of each paths. Secondly, the decoder has fixed weights that implement the signal propagation model, instead of learnable parameters. This knowledge-aware decoder ensures the output of meaningful physical parameters (i.e., AoAs) which is unattainable by conventional AEs. We also conduct a thorough analysis to characterize the multiple global optima and local optima of the estimation problem. This analysis inspires the design of a low-complexity two-phase training scheme and confirms the convergence of our proposed framework. Consequently, our framework addresses two key challenges in unsupervised learning: the lack of interpretability and the convergence to local optima. Extensive simulations validate our theoretical analysis and demonstrate the performance improvements of our proposed framework.
Funding Information
  • National Key Research and Development Program of China (2022YFB2902004)
  • National Natural Science Foundation of China (62271513)
  • Open Research Fund from Guangdong Provincial Key Laboratory of Big Data Computing
  • Chinese University of Hong Kong, Shenzhen (B10120210117-OF09)
  • National Science Foundation (NRT-AI 2244574)

This publication has 46 references indexed in Scilit: