To take full advantage of massive multiple-input multiple-output (MIMO) technology, it is indispensable to obtain accurate channel state information (CSI) through channel estimation. In this paper, a novel beam domain channel estimation algorithm based on sparse Bayesian learning (SBL) is proposed. First of all, we provide a channel model considering massive MIMO non-stationarity in the beam domain. Then the compressive sampling matching pursuit assisted block sparse Bayesian learning (CoSaMP-BSBL) algorithm is proposed. While leveraging the potential block sparsity of the channel, the proposed algorithm incorporates compressive sampling matching pursuit (CoSaMP) to pre-screen the small number of nonzero channel elements. Simulation results illustrate that the proposed CoSaMP-BSBL has higher estimation accuracy compared with other schemes in different scenarios. In addition, this algorithm shows the advantage of no need for prior channel information while maintaining reasonable complexity.