Permanent Magnet Synchronous Motor Based On Long Short Term Memory Network Fault Research On Health Status Prediction Method

Permanent magnet synchronous motor is a kind of high performance drive equipment,which is widely used in control and automation fields.The interturn short circuit fault is a frequent and highly destructive fault.If the fault is not found and eliminated in time at the early stage,the work efficiency will be affected,and the production will be suspended.In order to minimize the economic loss caused by interturn short circuit fault,it is particularly important to study the interturn short circuit fault of permanent magnet synchronous motor in the early stage.This paper is based on the deep research on the interturn short circuit fault of PMSM.On the one hand,the information maximizing generative adversarial nets for data expansion is introduced to solve the problem that the original negative sequence current data is not sufficient and no noise is introduced.Finally,the stable part of the generated data is selected to splicing with the original data.This method not only expands the data set,but also introduces noise to some extent to enhance the robustness of the prediction network.On the other hand,the recurrent neural network,the autoregressive model network and the long and short term memory network are used to predict the interturn short circuit fault of PMSM respectively.The experimental results show that prediction accuracy of the long and short term memory network is the highest.In this paper,the health status prediction of permanent magnet synchronous motor based on long and short term memory network are described from the three aspects:(1)Analysis of fault characteristics of permanent magnet synchronous motor.Firstly,the mathematical models of permanent magnet synchronous motor under normal ABC phase quadrant,under?-?quadrant and under d-qquadrant are introduced.Then the mathematical model of PMSM in the case of interturn short circuit is introduced and the negative sequence current is selected as fault feature.The experiment proves that when the interturn short circuit fault of permanent magnet synchronous motor occurs,the rise of short circuit current will cause the rise of negative sequence component at the same time.Therefore,it is reasonable to use the negative sequence current to study the interturn short circuit fault.(2)Research on information maximizing generative adversarial nets data expansion.Firstly,the generative antagonistic network and the conditional generative antagonistic network are selected for data expansion.The accuracy of the expanded data respectively was 64.5% and 85.6%.Because the accuracy of network expansion is too low,information maximizing generative adversarial nets is further selected to data expansion.The experimental results confirmed that when the number of hidden layers of the network was 3,the learning rate was 0.01,and optimization algorithm was Adam,the network state was optimal and the accuracy of data expansion was the highest,up to 96.8%.Finally,200 points in the stable part of the generated data are selected and splice with the original data to form a new continuous data set,which not only expands the data set,but also introduces noise to a certain extent to enhance the robustness of the network.(3)Research on prediction of interturn short circuit fault of permanent magnet synchronous motor based on long and short memory network.Firstly,the recurrent neural network is selected as the fault prediction network.The results show that the recurrent neural network can not be used to process long time and non-periodic data.The network has the problem of the gradient disappearing.Then,the autoregressive model network is selected for prediction research.The results show that although the autoregressive model network can converge quickly,it has obvious oscillation points after convergence.Finally,the permanent magnet synchronous motor fault prediction is studied by using long and short term memory network.The experimental results show that when the size of the network window is 200,the learning rate is 0.01,and the optimization algorithm is Adam,the network state is optimal.The network can overcome the disadvantages of the recurrent neural network,and not only does it avoid the problem of gradient disappearance or gradient explosion,but also has the ability of long term memory.Long and short term memory network is accurate in fitting the negative sequence current curve of permanent magnet synchronous motor,and its prediction accuracy is also the highest,reaching 98.3%.The network only oscillated slightly after 13,000 training sessions without gradient explosion.Therefore,the l long and short term memory network is selected as the permanent magnet synchronous motor health state prediction research network.