关键词: 永磁同步电机, 三相逆变器, 堆叠降噪自编码器, 前馈神经网络, 故障诊断

Abstract: As the power source of electric vehicles, the safety of the motor drive system is the key factor affecting the reliability and safety of the vehicle, and it is also one of the key technologies restricting the development of electric vehicles. Motor drive system is an electromechanical coupling system, and its fault occurrence mechanism involves many factors, so the fault diagnosis of motor is more complicated than that of other equipment. Stacked Denoising Autoencoder-Feedforward Neural Network (SDAE-FFNN) model was proposed to solve the problems of the motor drive system, such as difficulty in capturing fault signals, difficulty in precise feature characterization and poor fault diagnosis accuracy. Based on the simple structure and high efficiency of feedforward neural network, the last layer network of stack noise reduction autoencoder is replaced by feedforward neural network, which reduces the scale of neural network and improves the efficiency of the model. Additionally, this method makes full use of the relationship between the deep features of nonlinear empirical features and improves the accuracy of model fault diagnosis and recognition. Specifically, firstly, this paper simulates different types of open-circuit faults of three-phase inverter, and extracts three-phase stator current output of permanent magnet synchronous motor as the fault research object. Secondly, due to the high reliability design principle of the inverter, the fault data is distributed in a long-tail and the number of fault features is insufficient. Thus, multi-feature fusion methods, such as fast Fourier transform, short-time Fourier transform, Gabor transform, wavelet transform, empirical mode decomposition and linear regular transform, are used to integrate and form high-dimensional fault feature data sets. Then, the input signal with noise is reconstructed through the SDAE-FFNN model, which makes the model have good robustness and makes the deep features have the ability to prevent overfitting. Finally, the proposed model method is used to identify the open circuit fault of the three-phase inverter of the permanent magnet synchronous motor drive system. In order to verify the effectiveness of the proposed method, the paper compares the fault diagnosis with the single feature extraction method and the traditional deep network algorithm, and verifies the superiority of the proposed method. The experimental results show that the fault recognition rate of single feature extraction method is much lower than that of multi-feature extraction fusion method. At the same time, the method presented in this paper can effectively complete the fault classification identification, and the average recognition accuracy is as high as 98.8021%, which is better than the traditional deep learning method.