Research On Time Series Data Imbalance Processing And Fault Detection Methods Of Wind Turbine Based On Neural Network

As an important part of the intelligent management in wind farm,the fault detection for key components of wind turbine can improve the production efficiency and reduce the operation cost in wind farm,which is a hot issue in the current scientific research field.It has been widely concerned by researchers all over the world,and various wind turbine fault detection methods are proposed.The study found that the current fault detection methods of wind turbine still have the following problems:(1)Due to the category imbalance of wind turbine data sets,the wind turbine fault detection model tends to large category samples in training,which makes the effect of the model unsatisfactory;(2)Due to the single feature learning ability of the existing wind turbine fault detection model,the model can not make full use of multi-level feature information,resulting in poor performance of the model.To address these issues,the fault detection problem of wind turbine gearbox is studied in this paper,the main research contents and contributions are as follows:(1)Aiming at the problem of imbalance category of wind turbine data sets,a Stacked Generative Adversarial Networks(Stack-GANs)method is proposed,which considers comprehensively the correlation and time sequence of wind turbine data features when generating new samples.The Stack-GANs method generates data in a progressive way to maintain the strong correlation and weak correlation between features,and uses the Recurrent Neural Network to construct the generator and discriminator in Generative Adversarial Networks(GANs)of each stage to capture the time-series characteristics between data.At the same time,the features with high importance are selected by using random forest algorithm,so as to reduce the difficulty of constructing the Stack-GANs model and improve the training speed of the model.Stack-GANs is divided into two stages:In stage 1,according to the analysis results of Pearson correlation coefficient and Maximal Information Coefficient,GANs are trained in groups.Each GANs learns the strong correlation of features in the group,and generates the feature subset of corresponding groups.In stage 2,the feature data of each group generated in stage 1 are spliced as input,and the weak correlation of features between each group is corrected to generate minority class of sample data with high simulation.Compared with the existing algorithms for processing data imbalance,the experimental results show that the Stack-GANs method can generate effectively more real wind turbine fault data,which makes the wind turbine fault detection model learn the feature distribution of fault data better,so as to improve the overall detection performance of the model.(2)Aiming at the problem of single feature learning ability of existing wind turbine fault detection models,a wind turbine fault detection model LCL integrating Linear neural network,Compressed Interactive Network and Long-Short Term Memory neural network is constructed.LCL model has the learning ability of low-order features,high-order cross features and time-series features.Compared with the neural network model with single structure,it can explore the data features of wind turbine from multi-level and multi-angle,so that the model mines fully effective information to improve the detection performance.Compared with other advanced wind turbine fault detection methods,the experimental results show that LCL model achieves the optimal values in each evaluation indexes and can meet better the application requirements in the field of wind turbine fault detection.