Research On Condition Monitoring Of Generator Subsytem In Doubly Fed Induction Generator Wind Turbine Based On Deep Learning

With the development of wind industry,wind turbines have been widely deployed to convert wind energy into electricity.But wind turbines work in a harsh environment,they need high operation and maintenance(O&M)costs.The cost is about 20 percent of the wind farm revenue,and it limits the development of wind industry severely.In order to increase the reliability of wind turbines,reduce the fault caused by mechanical and electrical components,and optimize the operation and maintenance strategy,it is necessary to carry out online condition monitoring of components or subsystems in wind turbines,and abnormal should be found in the early stage of fault,so as to reduce the loss caused by wind turbine fault.In order to increase the reliability of generator subsystem in wind turbines,deep learning is used to online condition monitoring of generator subsystem in doubly fed induction generator(DFIG)in this thesis.Firstly,online condition monitoring technology is reviewed,and as one of its extension technologies,the remaining useful life prediction technology is introduced.Then supervisory control and data acquisition system(SCADA)data from a real DFIG and condition monitoring system(CMS)data from experiments are got,they are cleaned and the features are extracted.Mahalanobis distance is used to calculate the distance between features of SCADA data,mahalanobis space is built using Johnson transformation to judge the condition of generator subsystem in DFIG.Sparse auto-encoder and neural network are combined to predict the active power of DFIG,it's more accurate,and through the analysis of residual between the predicted power and the real power,the state of generator subsystem in DFIG is monitored.The generator in generative adversarial network(GAN)is used to obtain generation data,which has the same dimension with real SCADA data,the generated data and real SCADA data are used to train GAN,a discriminator model which can distinguish the condition of generator subsystem in DFIG is got after training.In view of abnormal component found by condition monitoring method,long short-term memory are used to predict the remaining useful life of it,and lebesgue sampling is introduced to increase the accuracy of prediction.Finally,experimental data of CMS is used to verify the proposed method.In this paper,the condition monitoring methods of DFIG wind turbines based on mahalanobis distance,sparse auto-encoder network,deep neural network and generative adversarial network are discussed.The results show that all of these three methods can detect the early abnormal behaviors of the DFIG.The method of mahalanobis distance is simple and does not need to build a model,but it need to inverse the features' matrix,which limits the sample size of features' matrix,so it's more appropriate in short-term monitoring,the method of sparse auto-encoder network requires the labels of data which need to be sorted out,it is sensitive and easy to be affected by noise signal,so it's more appropriate in special time,and the method of GAN is not as sensitive as the method of sparse auto-encoder network in finding abnormal data,but it is more robust,so this method can be used in daily time.