Knowledge-Guided Auto-Encoder for Unsupervised Angle-of-Arrival Estimation

In this paper, we propose a highly accurate unsu-pervised deep learning framework based on auto-encoder (AE) for angle-of-arrival (AoA) estimation in massive MIMO systems. Our method builds on an improvement of the vanilla AE by incorporating the knowledge of signal propagation models into the decoder. In our proposed knowledge-guided AE (KG-AE), instead of having learnable parameters, the decoder has fixed weights that implement the signal propagation model. Such modification forces the encoder to output meaningful physical parameters of interest (i.e., AoA), which cannot be achieved by standard AE. Furthermore, we rigorously analyze the multiplicity of local optima in unsupervised channel estimation problems. Our analysis informs the design of the cost function and the training scheme for the proposed KG-AE. Specifically, we design a two-stage training scheme, different loss functions in the two stages to achieve good initial points and boost the performance of the proposed KG-AE, respectively. Finally, extensive simulations are performed, and the results corroborate the analysis and demonstrate the performance improvements of the proposed KG-AE over the subspace-based algorithms and the state-of-the-art unsupervised learning-based algorithm.