The low-rank property of the channel covariances can be adopted to reduce the overhead of the channel training in massive MIMO system. In this paper, we exploit such low-rank property through virtual channel representation (VCR) under the time-varying channel scenario. Firstly, we reformulate the dynamic massive MIMO channel as one sparse signal model through VCR. Then, an expectation maximization (EM) based sparse Bayesian learning (SBL) framework is developed to estimate the statistical parameters of the sparse virtual channel. Specifically, the Kalman filter (KF) and the Rauch-Tung-Striebel smoother (RTSS) are applied to track the posterior statistics of the angle domain sparse channel in the expectation step, while a fixed-point theorem based algorithm and a low-complexity searching algorithm are separately developed to recover the temporal varying characteristics and the spatial signatures in the maximization step. Finally, we demonstrate the efficacy of the proposed schemes through simulations.