Aiming at the non-stationary and non-linear characteristics of rolling bearing degradation, a deep-network-driven state-space modeling method for rolling bearing degradation prediction is proposed. First, the bearing degradation-sensitive features are merged and downscaled using principal component analysis (PCA) to construct health indicators (HI). Then, a deep network-driven state space model is built and trained to predict the bearing degradation. The proposed model is based on the state-space modeling framework, which uses ECA-BiLSTM network as the state update equation to capture the long-term dependency information in the data; and the variational self-encoder as the degradation measurement equation to achieve data compression and feature learning to capture the degradation-sensitive features. Experimental results show that the proposed model can effectively predict the bearing degradation.