Anomaly Detection Of Wind Turbine Drivetrain Based On Generative Adversarial Networks

In recent years,the installed capacity of wind turbines has been increasing year by year,and the capacity of individual wind turbines has become larger and larger.Compared with the ground equipment,the failure rate of wind turbines is higher due to the severe working environment such as high altitude,random wind load excitation and extreme temperature difference.The failure of the drivetrain of wind turbines will lead to a long period of downtime and a large economic loss.Therefore,it is of great practical significance to carry out the state monitoring and fault warning of the drivetrain of wind turbines.At present,the Supervisory Control And Data Acquisition system of large wind turbines,namely the SCADA system,has the function of threshold alarm,but it can hardly be directly applied to the fault warning of wind turbines.The main reason is that the state information in the massive SCADA data is complex and irregular.In the actual fault warning,the fault deterioration state of the equipment is ignored only by the threshold alarm,and it is difficult to reflect the mechanical failure that has little correlation with the monitoring parameters.Under these circumstances,this paper takes the drivetrain of wind turbines as the research object,and based on the large amount of data collected by the SCADA system,it studies the anomaly detection method of the drivetrain of wind turbines based on Generative Adversarial Networks,aiming at providing guidance for the operation and maintenance of wind turbines.The main research contents are as follows:(1)According to the operation characteristics of wind turbines under variable conditions,the SCADA data preprocessing method to eliminate abnormal points was studied.Based on the operation principle of wind turbines and the regular pattern of data statistics,the processing methods of SCADA data,such as segmentation,denoising and screening outliers,are designed to screen out data points irrelevant to the wind turbines’ health status in SCADA data,so as to provide accurate data sources for the establishment of subsequent data models.(2)Anomaly detection model design of wind turbine based on Generative Adversarial Networks.The technical characteristics of Generative Adversarial Networks were studied and a CGAN,which is convolution GAN,model was built based on the regular pattern of SCADA data of wind turbines,and the health index reflecting wind turbines’ state was designed.Aiming at the potential training instability of the CGAN model,the Wasserstein GAN with Gradient Penalty model was established to optimize the training process through the gradient penalty function and the improved loss functions.Considering the randomness of the noise information sampled by the traditional Generative Adversarial Networks,the Autoencoder and GANs were further fused into the GANomaly model.In this model the reconstructed samples were regarded as generated samples,which avoids the instability of the results.(3)Quantification and visualization of anomaly detection results.The test results of Generative Adversarial Networks are just the quantized results of the degree of abnormality of the wind turbine drivetrain at all times.Then the numerical results are smoothed to make the change trend in time domain more obvious.The control chart method is introduced into the threshold setting of abnormal state to realize the accurate prediction of abnormal state and final fault alarm.