In the study of wind turbine fault early warning based on deep learning, aiming at the prediction accuracy of the model and the accuracy of fault early warning, a combined model early warning method combining CNN, BiLSTM and attention mechanism Attention is proposed. Firstly, aiming at the problem of low quality of SCADA raw data, the parameter-optimized DBSCAN algorithm is combined with the control principle of wind turbine to complete data cleaning, and the GRA analysis method is used to screen the original features to reduce the redundancy between features. Aiming at the problem of model prediction accuracy, in order to improve the feature extraction ability of BiLSTM network and the focusing ability of key features, CNN and attention mechanism are introduced respectively to build a combined network model. Finally, the exponential weighting method is used to smooth the power residual, so as to determine the early warning threshold and realize the fault early warning of wind turbines. The effectiveness of the method is verified by the SCADA data of a wind farm. The experimental results show that compared with the BiLSTM model, the error indexes RMSE and MAE of the proposed model are reduced by 29.8 % and 30.7 % respectively, and the fitting degree R2 is increased by 4.8 %. The warning time is 2~6 hours earlier than the SCADA alarm log.