Fault Diagnosis Of Direct-drive Wind Turbine Based On Spatio-temporal Feature Mining Of SCADA Data

Wind energy,as a clean,low-cost renewable energy source,has attracted more and more attention from wind power companies around the world.The installed capacity and installed capacity of wind turbines are increasing day by day,The structure of wind turbine(WT)is more complex,but the problems that follow have emerged: wind turbines fail frequently.Therefore,it is urgent to study fault diagnosis methods for WTs.SCADA data has become an important data support for the study of wind turbine fault diagnosis methods due to its large amount of data,easy access,and no need for additional sensors.The SCADA data of WTs has many variables and complex coupling relationships between variables,and the data has time correlation.Based on the characteristics of the multivariable time series of SCADA data and considering the data imbalance with few fault samples and many normal samples,this paper proposes a multi-kernel fusion convolutional neural network model(MKFCNN),a spatio-temporal fusion neural network model(STFNN),and a spatio-temporal multi-scale neural network model(STMNN)to solve the problem that the fault features contained in the SCADA monitoring data of wind turbines are difficult to extract,and provides a new technology for wind turbine fault diagnosis.All the proposed network models have verified the effectiveness and reliability through experiments.The data used are the Benchmark simulation data set and the actual wind farm SCADA data set.The main work of the paper is summarized as follows:(1)Discuss and analyze the structure,operation principle,main fault types and inducing factors of direct-drive wind turbines;Taking economic,practical and convenient factors into consideration,SCADA data is determined to be used as the data basis for model development,and the characteristics of SCADA data are deeply discussed and analyzed to lay a foundation for the subsequent development of fault diagnosis model.(2)Research on the spatial multi-scale feature extraction method based on MKFCNN model.Inspired by Goog Le Net,MKFCNN model is designed based on one-dimensional convolutional neural network to solve the problem that fault features of wind turbines are difficult to be extracted,and the concept of making full use of computer resources is adopted.The model can increase the depth of the network without increasing the consumption of resources and extract the implied fault features from the data at multiple scales,and the validity of the model is verified on Benchmark simulation data and SCADA data sets of wind farms.(3)Research on the features extraction method based on STFNN network model of spatio-temporal two dimensions and spatial multi-scale.In view of the large number of SCADA data variables of wind turbines and the complex coupling relationship between the variables,and the characteristics of correlation in time,the STFNN network model is proposed,which uses the MKFCNN network module and the long short-term memory network module(LSTM)to extract separately spatial multi-scale features and temporal correlation features from the SCADA data,and finally the spatiotemporal features are fused together and sent to the classifier for the final fault diagnosis.The method was verified on the Benchmark simulation data and the actual wind farm SCADA data set,which proved the effectiveness and reliability of the method.(4)Research on spatio-temporal two-dimensional multi-scale feature extraction method based on STMNN network model.In view of the large amount of SCADA data,the complexity of variables,the characteristics of time series and the difficulty of extracting fault features,the STMNN network model is proposed.The multi-scale deep echo state network module(MSDeep ESN)and the multi-scale residual neural network module(MSRes Net)are designed to extract the temporal multi-scale features and spatial multi-scale features hidden in the SCADA data,and considering the serious data imbalance in SCADA data,the cost function focal-loss is used as the loss function to overcome the impact of data imbalance to the model training.The model is validated on the wind farm SCADA dataset.