Abstract: An integrated approach,which is based on SCADA data analysis,sparse self-encoder and deep neural network algorithms,is proposed for wind turbines online condition monitoring. Firstly,the complex intrinsic features of SCADA high-dimensional data are learned by sparse auto-encoder,and the reduced dimension data is obtained. Secondly,deep neural network is used to predict the output power of wind turbine based on the reduced dimension data,wind turbine’s condition is judged by analyzing the residuals between the predicted active power and the actual active power. Finally,SCADA data of a wind turbine for nearly one and a half years are used to verify the proposed method. Results show that the proposed approach can detect anomalies of wind turbine generator 5 days before it is shut down for maintenance which can avoid the shutdown caused by catastrophic failures,reduce the maintenance cost,and improve the competitiveness of the wind energy.