Research On Data Unbalance Processing And Fault Prediction Method Of Wind Turbine

At present,the comprehensive development of green and low-carbon energy is of great significance to the survival and development of mankind,and wind energy is an indispensable part of this type of energy.The number of wind turbines in China is increasing,and wind power generation has gradually changed from supplementary energy to alternative energy.However,the internal structure of the fan is complex,the working environment is harsh,and it is in the complex tuyere area all the year round,which causes serious failure of the fan equipment and higher operation and maintenance costs.If the abnormality can be detected in time and dealt with in the early stage of the failure,it can avoid aggravating the failure and causing a huge disaster.Therefore,it is of great significance to predict fan failures,which can effectively reduce the related costs of equipment maintenance.Through the analysis of the faults encountered in actual production and the research on the current wind turbine fault prediction methods,it is found that the following problems still exist:(1)The wind turbine data set categories in actual production are not balanced,which leads to the poor failure prediction effect of the model.(2)The existing fault prediction model randomly disrupts the data set during the training phase,causing the previous and subsequent running data to be independent of each other,destroying the relevance and changing trend of the data.In order to solve the above-mentioned problems,this article takes the fan gearbox as the research object.The main research contents and results are as follows:(1)Aiming at the problem of unbalanced wind turbine data sets,an improved BSMOTE-Sequence sampling algorithm is proposed,which takes the spatial and temporal characteristics into consideration when synthesizing new samples,and uses Tomek Links algorithm to clean the new samples.Can effectively filter out noise sample points.First,find the neighbor samples of each minority sample,and divide the minority samples into safety samples,boundary samples,and noise samples according to the proportion of these neighbor samples.Then,for each boundary class sample,the minority class sample set with the closest spatial distance and time span is selected,new samples are synthesized by linear interpolation,and noisy samples and overlapping samples between classes are filtered out.Finally,SVM,CNN,and LSTM are used as wind turbine gearbox fault prediction models,and F1-Score,AUC(Area Under Curve),and G-mean are used as model performance evaluation indicators.The actual wind turbine data sets are compared with a variety of commonly used sampling algorithms.The experimental results show: Compared with existing algorithms,the classification effect of samples generated by the BSMOTESequence algorithm is better.F1-Score,AUC,and G-mean are increased by an average of3%,which can be effectively applied to the wind turbine faults with unbalanced data with sequential rules Forecast field.(2)Aiming at the problem that the current fault prediction method randomly disrupts the data set during the model training stage,destroys the correlation and change trend of the data before and after,a wind turbine fault prediction method based on the operating data graph is designed.First,perform operations such as abnormal point processing,missing value processing,feature selection,standardization,and dimensional compression on the original data set to obtain a clean and condensed data set.Then,according to the set data graph size and failure ratio and other parameters,the one-dimensional data records are assembled into two-dimensional data graphs.Then,use the convolutional neural network to complete the feature extraction of the data graph,so as to obtain the relationship features of the internal structure of the data,and build the gearbox fault prediction model.Finally,the test set is used to verify the effect of the method,and the data graph size,failure ratio and other parameters are discussed,and their influence on the model prediction effect is analyzed.The experimental results verify the effectiveness of the wind turbine failure prediction method based on the data graph.