Unitary matrix transformation is a commonly used real-valued method， which can effectively reduce computational complexity. However， the dimension of the observation matrix is doubled in the existing downlink channel estimation method for massive Multiple input multiple output （MIMO） systems based on unitary matrix transformation. Without dimensional compression， the goal of reducing computational complexity is difficult to achieve. Although the orthogonality of signal space and noise space can compress the dimension， the signal space can only be approximately calculated， leading to performance loss. To improve channel estimation performance， the signal space matrix is regarded as a variable and adaptively adjusted in the process. Since the signal space matrix and sparse signal matrix are highly coupled， the traditional Bayesian inference is not applicable. Therefore， the column-independent variational Bayesian inference (VBI） factorization is adopted to decouple the signal space matrix and sparse signal matrix successfully. Simulation results show that this method can significantly improve the channel estimation performance.