Fault Diagnosis Of Wind Turbines Based On Data-mining Method

Wind energy is particularly rich in the earth,it is not expensive and the process of generation does not have any influence on our surroundings,so the wind power industry will develop very well in the future.However,the resources are mostly distributed in harsh circumstances,and the season and weather intensified the uncertainty and uncontrollability of wind energy.Considering the safety,stability and economy of the generation process,it is very necessary to research and develop an efficient condition monitoring and fault detection system.Technology is changing with each passing day,so the storage and processing of data is becoming easier.The large-scale information generated in the operation of wind power plants provides a rich soil for real-time online fault diagnosis using data-driven algorithms.Statistics and neural network are two common techniques,which have developed rapidly in the domain of detecting malfunctions in wind turbines in the past decade.The former can accurately and quickly solve problems such as the high dimensional,dynamic,interrelated variables and the uncertainty of the system itself;the latter can imitate the human learning behavior,but it is not suitable for high dimensional variables.Based on the above two techniques,this paper comes up with improved CVA and two uses of LSTM.The former creates a new statistic in view of CR,mapping CR to the feature space of that statistic by PLSR.The latter uses LSTM to build the cross-LSTM forecasting model and sequence-LSTM classification model respectively,combined with feature extraction method of sliding window,simultaneously solving the time series correlation problem and its high dimension properties in space.The conclusions from benchmark show that the accuracy and sensitivity of the two methods are better than that of traditional CVA and machine learning methods.Also,actual operation data are selected to judge whether the blades freeze or not.The simulations verify that improved CVA and LSTM prediction and classification have the same validity for benchmark data and actual operation data,indicating that they can indeed improve the performance of fault diagnosis,but they apply to different scenarios.