Wind turbine generator has the characteristics of complicated structure and difficult maintenance. In order to evaluate its health, this paper combines the characteristics of denoising auto-encoder and sparse auto-encoder to improve the traditional stacked auto-encoder model, and uses the reconstruction error of the model to monitor the running state of the wind turbine generator. The reconstruction error obtained by off-line testing is compared with the reconstruction error obtained by online monitoring, and the health of the wind turbine generator is obtained by combining three different indicators. The health assessment model is trained and tested by using the actual data of a wind farm in Hebei Pro- vince. The example analysis shows that the method can effectively track the state change of the wind turbine generator and has the function of early identification of faults.