Fault Diagnosis And Prediction Of Permanent Magnet Synchronous Motor Based On Deep Learning

Permanent magnet synchronous motor as a high-performance driving device,plays a key role in the field of high-precision control and industrial automation.However,the occurrence of interturn short circuit and permanent magnet loss of excitation affects the efficiency of the motor,and burns the motor seriously,resulting in irreparable losses.Therefore,in order to improve efficiency and reduce losses,it is of great significance to study the interturn short circuit and loss-of-excitation faults of permanent magnet synchronous motor.Based on the in-depth study of the above-mentioned permanent magnet synchronous motor faults,a deep learning fault diagnosis method based on generation model is proposed in this paper.On the one hand,aiming at the problems of cumbersome fault data processing,single feature and sparse sample,data expansion model is introduced to conduct unsupervised learning of ready-made samples,and simulate sample features to synthesize pseudo-data.On the other hand,aiming at the problems of high fault diagnosis complexity,low recognition efficiency and poor robustness,an Auto-Encoder Net based on sparsity principle is proposed as a fault diagnosis model of permanent magnet synchronous motor.In this paper,the fault diagnosis method of deep learning permanent magnet synchronous motor is described from the above four parts.(1)Fault characteristic analysis of permanent magnet synchronous motor.Firstly,a permanent magnet synchronous motor model is built to collect real-time parameters such as phase current,frequency domain current,flux density and electromagnetic torque.Then,the characteristics are obtained by calculating,and the correlation between the characteristics and motor fault is analyzed from the mathematical point of view.The simulation results show that it is reasonable to use each feature combination as the fault sample set of permanent magnet synchronous motor.(2)Research on data expansion method of Generative Adversarial Nets.Firstly,the advantages of this method are highlighted by comparing the Generative Adversarial Nets with the traditional generation model.Then,the network structure and connotation algorithm are deeply studied,and the generation mechanism of Generative Adversarial Nets is explained from two aspects(generation stage and discrimination stage).Next,the measurement criteria of the results are derived,and the network tends to the best standard by iterating the optimization algorithm repeatedly.The experimental results show that the Generative Adversarial Nets can be effectively applied to sample expansion,but due to its principle constraints,the application scope is limited to small-dimensional data generation(3)Research on data expansion method of Vriational Auto-Encoder Net.In order to remedy the shortcomings of Generative Adversarial Nets,a Vriational Auto-Encoder Net is proposed.The Vriational Auto-Encoder Net is divided into encoder and decoder.According to these two internal models,the working mechanism of the Vriational Auto-Encoder Net is studied.The reconstruction error criterion of the method is described from the mathematical point of view,and the visual comparison of the results is made based on the two error functions.The experimental results show that the Vriational Auto-Encoder Net has better accuracy than the Generative Adversarial Nets in terms of performance,and the Generative Adversarial Nets are simpler in terms of network complexity.(4)Research on fault diagnosis method of Sparse Auto-Encoder Net.Firstly,the method of introducing sparsity is studied.Then,the process of building Sparse Auto-Encoder Net model is described from the mathematical point of view.The advantages of this method in fault diagnosis are analyzed.The experimental results show that this method is more accurate and feasible than the traditional fault diagnosis method.