Abstract: For the wind turbine in variable working conditions and complex environment,the fault monitoring and early warning of its bearing faces the problems of high latent characteristics of and the difficulty in fault threshold setting. Aiming to the problems,an early warning method based on multi-layer deep mutual information variational network（MDMIVN） is proposed. The coding layer of variational auto-coder is extended to multi-layer encoding,and the decoded signal is encoded again to improve the network robustness to the noise in the original fault signal. To improve the ability of modeling the bearing normal state space,the maximum mutual information between the latent variables and the input signal,between the latent variables and the secondary coding features are introduced as the loss functions. The health index is established based on the reconstruction errors between the secondary coding features and the latent variables. And then,the health baseline is adaptively set,by combining the triple exponential weighted moving average model and the updating iteration of the health indexes. The experimental results on two wind turbine bearing vibration data sets show that compared with the traditional early fault detection method based on model reconstruction,the proposed method has high fault warning accuracy,anti-interference ability.