Aiming at the problem of fault diagnosis of generator bearings of wind turbines, a deep learning network fault diagnosis model based on sliding window-KL divergence and improved stacked autoencoder was proposed. The improved variable learning rate stacked autoencoder was used to reconstruct the temperature state of the generator bearing. The sliding window-KL divergence algorithm was used to diagnose the generator bearing, and the fault diagnosis results were compared with the Euclidean distance and the 3σ criterion. The results shows that the sliding window-KL divergence algorithm has high accuracy and low false alarm rate in fault diagnosis.