Research On Fault Detection Technology Of Wind Turbine Bearing Based On Deep Learning

The generator and gearbox of the wind turbine are important components of the wind turbine.At the same time,these two components are also the most serious causes of wind turbine shutdown components.The cause of the failure is mostly caused by the damage of the rolling element bearing sub-components of the generator and gearbox.Therefore,the research on the fault detection technology of the rolling element bearing sub-components of the wind turbine can reduce the occurrence of the wind turbine failure and ensure the stable operation of the wind turbine.The main research contents include :(1)The specific contents of Supervisory Control and Data Acquisition(SCADA)data of wind farm and vibration signal data of wind turbine bearing are studied.At the same time,the principle of deep learning algorithm involved in this thesis is studied.(2)Aiming at the abnormal state detection technology of wind turbine bearings,a Time Convolutional Neural Networks and Long-Short-Term Memory Networks(Time-CNN-LSTM)fusion model considering the time series characteristics of data is adopted.Firstly,considering the influence of the seasonal distribution difference of SCADA data on the prediction accuracy of the model,the normal state prediction model of wind turbine bearing in each season is constructed,and the prediction error of the normal state and abnormal state of the bearing on the normal state prediction model is analyzed.The error perception interval of the normal state prediction model in each season is determined,and the determination rules of the abnormal state of the bearing are stipulated.The simulation results show that this method has faster bearing abnormal state recognition ability than SCADA early warning report and single simple network model.(3)Aiming at the fault diagnosis technology of wind turbine bearings,a new scene wind turbine bearing fault diagnosis method based on Multiple Wide Kernel Convolutional Neural Networks(MWKCNN)and transfer learning fusion is adopted.This method uses the bearing vibration signal data in the source domain to train the MWKCNN model,and then based on the model transfer learning method,according to the similarity between the source domain and the target domain,the source domain MWKCNN model structure is fine-tuned by layer freezing and layer replacement.The simulation results show that the fault diagnosis accuracy of the MWKCNN model in the source domain reaches 99.48 %,and the fault diagnosis accuracy in the target domain reaches more than 94 %.Compared with other simple network models,the MWKCNN model has better diagnostic effect after migration.