Regularized Deep Generative Model Learning for Real-time Massive MIMO Channel Tracking

Conventional deep neural network (DNN)-based channel estimator is trained offline and deployed online with fixed weights, lacking the adaptivity and generality to fast varying channels. In this paper, we propose a real-time compressive channel tracking (CT) algorithm based on regularized deep generative model (DGM) to recover time-varying channels recursively. Specifically, we propose an untrained DGM for online CT, which does not require training over large datasets and can learn the channel variations on the fly. Moreover, DGM can capture the sparse structure of massive MIMO channels automatically. To further improve the tracking performance, we propose to regularize the DGM adaptively using a temporal prior to exploit the dynamic sparsity of channels, which is learned by an online trained sliding long short term with memory (LSTM) network. Simulations demonstrate the superior performance of the proposed algorithm compared to the model-based and supervised DNN-based CT algorithms.