Abstract: Atrous Convolutional Neural Networks（ACNN）and Bidirectional long-term and short-term memory network（Bi-LSTM）are combined to realize the early fault identification of wind turbines. First,the input variables are determined by Pearson correlation coefficient. Then,based on ACNN and Bi-LSTM,the spatial and temporal features of SCADA data are extracted,with active and reactive power output as the predictive variables. Finally,the root mean square error（RMSE）of the output predicted value is calculated,and the adaptive threshold is set by EWMA to identify the wind turbine state. After extracting the spatial features with ACNN,the Bi-LSTM is used to perceive the changes of spatial features in time series,which improves the training efficiency of the model and the sensitivity of the early fault of the wind turbine. Through the analysis of the actual wind turbine SCADA data,it is proved that this method can effectively identify the early faults of wind turbines.