Research On Data-driven Fault Diagnosis Method Of Three-level Inverter

In recent years,with the increasing scale of offshore oil platforms,how to improve the stability and reliability of crude oil production system has become the focus of research.Three-level inverter is an important part of the frequency converter of oil production electric submersible pump.After the fault occurs,it is often checked by technician.There are some problems such as low efficiency,which will seriously affect the stable operation of the crude oil production system.Therefore,the research on fault diagnosis of three-level inverter has important engineering value and practical significance.In this thesis,the neutral point clamped three-level inverter in the frequency converter of oil production electric submersible pump is taken as the research object.Without additional sensing equipment,the open circuit fault of power switch and soft fault of capacitor are identified and located by data-driven method.The details are as follows:In view of the interference of the noise in the actual operating condition,the open circuit fault feature extraction method of power switch based on variational mode decomposition was studied,and the time-frequency domain parameters of each order intrinsic mode function were further calculated as the characteristic index to characterize the operation state of the inverter.In order to solve the problem of the high dimension of the fault features in time-frequency domain,a supervised local linear embedded manifold learning method was proposed,which used Euclidean distance with class information to improve neighborhood point search,and the problem of unobtrusiveness of feature discrimination caused by ignoring class information in LLE algorithm was solved.The simulation results show that the method based on VMD and SLLE can effectively extract open-circuit fault features from nonlinear and non-stationary fault signals,and it also has good noise robustness.Aiming at the problem of low recognition accuracy of traditional machine learning method in open circuit fault diagnosis of power switch,a model of extreme learning machine based on BSO was studied.In the introduced BSO algorithm,the update rules of each particle are based on the particle swarm optimization algorithm,and their own judgment on the environment is added in the iteration to improve the global search efficiency.BSO algorithm is used to optimize the input weight and hidden layer bias of ELM,which solves the problems of poor accuracy and long learning time in the fault diagnosis of ELM.The simulation results show that the BSO-ELM method has higher diagnostic accuracy than BP,SVM,ELM and PSO-ELM.Considering the characteristics of poor differentiation of different capacitive soft faults,a capacitive soft fault diagnosis model based on VMD and Res Net-LSTM was proposed.First,the IMF fault feature map was constructed by VMD algorithm,and then the feature map samples were extracted by Res Net in deep space.In order to further extract the timeseries features of the samples,LSTM was used to learn the time-series of the extracted spatial features.This model overcomes the problems such as gradient disappearance and accuracy decline with the increase of the number of layers of CNN,and avoids the problem of insufficient feature extraction caused by only learning spatial features from samples.The simulation results show that the fault diagnosis model based on VMD and Res Net-LSTM has higher fault identification accuracy compared with Res Net and CNN-LSTM,thus verifying the applicability effectiveness of the proposed model.