Research On Fault Prediction Of Wind Turbine Driving Chain Based On Deep Learning Method

With energy crisis and environmental pollution becoming increasingly prominent,wind energy as a clean,renewable green energy has been rapidly developed in recent years.As the installed capacity of wind power is increasing rapidly,the problem of wind turbine faults has become the focus of attention due to wind turbine in a complicated and changeable environment prone to failure.So,this problem needs to be solved urgently.As key components of wind turbines,gearboxes,generators and main bearings,their faults are the main cause of long downtime of wind turbines.Therefore,it is of great significance to study and analyze the fault characteristics of gearboxes,main bearings and generators,and realize fault prediction to improve the reliability and availability of wind turbine driving chains and provide a new method for wind turbine health management in wind farm data centers.In order to realize the fault prediction analysis of wind turbine driving chain components,a deep learning method of deep autoencoders(DAE)network based on supervisory control and data acquisition(SCADA)is proposed in this paper.In this method,several restricted boltzmann machines are connected to form a deep autoencoders network.The SCADA state parameters data is selected under normal conditions as training data to train this network,so that the network intelligently extracts the intrinsic relationship from the main bearing,gearbox and generator SCADA parameters layer by layer to construct their deep encoding learning model of components.According to the relationship between DAE input and output and the dynamic parameters become non-stationary change of the components under fault condition,the reconstruction error(R_e)is calculated by using the initial input and reconstruction value of the deep learning network of wind turbine components,taken as the detection index to reflect the whole condition of components.The DAE network model is established based on the SCADA parameters of main bearing,gearbox and generator respectively.Through deep analysis of the characteristics of R_e changes,different thresholds were set to detect its change trend,and finally we found that the reconstruction error with non-stationary and extreme characteristics.Therefore,the extreme value theory is used to set the adaptive threshold to monitor the trend of reconstruction error,which reduced the influence of false alarm and R_e extremum,and finally acts as the judgment criterion of anomaly alarm.Through the verification and analysis of the recorded data before and after the main bearing,gearbox and generator faults in the actual wind farms,the analysis results verified that the method can detect the fault characteristics in advance and send anomaly alarms to achieve the fault prediction.So,the deep autoencoder network method for wind turbine driving chain components fault prediction is achieved.After alarming the anomaly of wind turbine components,the possible causes of failure were further analyzed according to the residual trend of the component SCADA parameter and physical laws.The analysis results are consistent with the actual fault information.