Research On Fault Diagnosis Methods Of Wind Turbine Based On Generative Adversarial Networks

Wind power is one of the most powerful ways to promote the development of renewable energy.The sudden-stop of wind turbines will not only result in interruption of power supply,but also causes great damage to machine components,and even shorten the service life of the whole wind turbine.Therefore,it is of great significance to monitor wind turbine operating states and detect fault or warning events in time based on various collected signals.Therefore,aimed at imbalanced problems and varied operating conditions in wind farms,this thesis studies wind turbine fault diagnosis based on improved generative adversarial network(GNA).Aimed at class-imbalanced fault diagnosis problems in wind farms,an intelligent anomaly detection approach based on GAN is proposed.In particular,an encoder-decoderencoder three-sub-network generator is trained on normal samples alone and diagnoses faults by generating much higher anomaly scores when a fault sample is fed into the trained model.The proposed approach can distinguish abnormal samples from normal samples with100% accuracy on a benchmark rolling bearing dataset acquired by Case Western Reserve University and another rolling bearing dataset which is acquired by our laboratory.Vibration signal of rolling bearings in wind turbines is not complex,so using GAN is enough to capture their feature distribution patterns and the simple design of anomaly score is enough to detect abnormal samples.However,in most cases,the distribution characteristics under normal and abnormal conditions is hard to distinguish just through training GANs.In most scenarios of fault diagnosis for wind turbines,only using the anomaly score cannot accurately identify the fault samples.In addition,when there are fault samples,ignoring them completely will definitely cause the loss of important information.To solve the above problem,this thesis puts forward an intelligent fault diagnosis framework,which is composed of two parallel GANs followed by a CNN classifier.After training GANs for each classes,the CNN classifier is used to further extract the differences between these features.The proposed two-stage training strategy can help to improve the performance.The effectiveness of the proposed method is validated by SCADA data of three wind turbines from the first Chinese Industrial Big Data Innovation Competition.The experimental results are state-of-the-art compared to SVC and traditional CNN-based networks.Operating conditions are always different from turbine to turbine,so a fault diagnosis framework trained on dataset from one wind turbine is always not suited to diagnose fault events of another wind turbine.To improve the diagnosis accuracy on a new wind turbine,a detection model based on GAN and transfer learning is proposed.Firstly,the data distribution characteristics of each class are obtained by using GANs,and then transfer learning is used to capture domain-invariant features between source domain and target domain.The experimental results on the data mentioned above show that the proposed framework effectively improves the model transfer performance.