Research On Fault Diagnosis Method Of Wind Turbine Transmission System Based On Deep Learning

The adjustment of energy structure is crucial to the realization of the medium and long-term goals of "peak carbon dioxide emissions" and "carbon neutrality".Wind power generation is developing rapidly as a green and clean energy.However,with the wind turbine running for a long time,various equipment components will inevitably occur various faults.If these faults cannot be found and dealt with in time,they will cause huge losses to wind power enterprises.Therefore,fault diagnosis of wind turbine has gradually become a hot research direction.In this thesis,vibration signals of rolling bearings and gearboxes in the transmission system of wind turbine are studied.Starting with signal processing based on Variational Mode Decomposition(VMD),vibration signal features are analyzed and taken as input of Convolutional Neural Network(CNN)to intelligently extract and diagnose fault features.In view of the adverse effects of noise,compound faults and variable working conditions on the diagnosis results of wind turbine during actual operation,three different diagnosis methods were proposed,and each module in the CNN was optimized and verified.The main contents are as follows:1.Aiming at the end effect and mode mixing problems of Local Mean Decomposition(LMD)and Empirical Mode Decomposition(EMD),VMD was introduced,and a method to determine the mode number based on the correlation coefficient was proposed.The effectiveness of the method was verified by simulation signals,data under constant operating conditions and actual operation data of wind turbine.The results show that VMD exhibits the characteristics of band-pass filtering,which can detect the bearing unbalance fault and rolling bearing fault.2.Aiming at the problem that traditional fault diagnosis methods need a lot of professional background knowledge and experience,a fault diagnosis method based on.CNN is proposed.Several network models based on Inception module are designed by analyzing the influence of network depth and various optimization techniques on the diagnosis results in the classical Lenet-5 network structure,and each mode obtained by VMD is input into the network with different strategies.The validity of the model is verified by experimental comparison.The results show that the accuracy of the proposed network model for fault diagnosis can reach about 99%,and the whole process does not need human intervention,which realizes the intelligent fault diagnosis.3.In order to solve the problem that noise and compound faults have great influence on the diagnosis results of wind turbine during actual operation,a fault diagnosis method based on deep residual network(ResNet)and attention model was proposed.Firstly,the experimental comparison shows that the diagnostic accuracy of the existing network models decreases when different degrees of noise are added and compound faults exist.Secondly,the ResNet module and the attention module are improved.Finally,based on the original network model,the module is expanded and the network model based on deep residuals is proposed.The results show that the ResNet can effectively deepen the network and avoid gradient explosion and gradient disappearance.The attention model uses the error back propagation algorithm to optimize the weight,which can effectively reduce the influence of noise on the diagnosis results.The processing method based on VMD can effectively decouple the compound faults,and the diagnosis accuracy can be improved by using the decomposed modes as network input.4.A fault diagnosis method based on deep transfer learning(TL)was proposed to solve the influence of variable working conditions on the diagnosis results of wind turbine.By analyzing the prediction accuracy of the existing network models for the same working conditions with different loads,different working conditions with the same loads and different working conditions with different loads,it can be found that the network model has a certain generalization ability and the data used has the prerequisite of TL.On this basis,a fault diagnosis method based on model parameter transference and feature mapping and several transfer learning programs are proposed.A transfer learning program is determined through comparative experiments,and the effectiveness of the method is verified.At the end of the thesis,the main works and conclusions of the research are given,and the future research direction is prospected.