Research On Application Of Deep Learning Method In Wind Turbine Fault Diagnosis

Wind energy has gained widespread attention in recent years as a sustainable and clean energy source.Wind turbines with complex structure are usually installed in harsh operating environments,resulting in the increasing probability of wind turbine faults,affecting the economy and safety of wind farm operations.In recent years,the wind turbine fault diagnosis method based on the deep learning algorithm has attracted many researchers.One of the most representative algorithms in deep learning,the convolutional neural network(CNN),is able to downscale high-dimensional input data to distil deep features.However,most of the current research has used CNNs for image recognition,while there is relatively little analysis on the application of supervisory control and data acquisition(SCADA)data.The proper operation and maintenance of wind turbines is facilitated if the advantages of convolutional neural networks can be exploited in the faults diagnosis of wind turbine.The following research is conducted in this paper based on convolutional neural networks in deep learning methods for large component fault diagnosis in wind turbines:Firstly,for the application of convolutional neural network in fan fault diagnosis,a icing diagnosis method of 1D-CNN network wind turbine blades with feature extraction is proposed.Feature extraction of the data is achieved by feature selection and reconstruction.The XGBoost model is used for feature selection of SCADA data,and features with high correlation with the icing state of wind turbine blades are screened out.Then,new features are constructed based on the residual principle for the important features.The features selected by XGBoost and the three new features reconstructed are combined into a new feature set,which is used as the input of 1D-CNN network for blade icing diagnosis.Secondly,for the application of convolutional neural network and long short-term memory(LSTM)network in wind turbine fault diagnosis,a new diagnosis model of wind turbine power reduction at high temperature based on the combination of vine-Copula network model and CNN-BiLSTM-attention algorithm is proposed.The vine-Copula model was first used for correlation analysis to obtain a high-dimensional vine-Copula structure consisting of highly correlated features together with the high temperature power reduction state.The selected features were input into the CNN-BiLSTM-attention model,using the advantages of CNN and temporal features of BiLSTM,and improving the BiLSTM network,to obtain the evaluation results of the high temperature power reduction state of the wind turbine.Finally,for the application of improved convolutional neural network for probabilistic prediction in wind turbine fault diagnosis,a yaw angle zero drift diagnosis method using temporal convolutional neural network quantile regression(TCN-Quantile)is proposed.Firstly,DBSCAN is used to process the wind speed bins.Secondly,a power probability prediction curve based on TCN-Quantile neural network is established to estimate the yaw zero drift angle.Finally,the yaw zero drift diagnostic effect of the model proposed in this chapter is verified by actual wind farm operation data.