Fault Diagnosis Method of Motor Bearing Based on Deep Transfer Learning

Abstract

Aiming at the problem that the fault diagnosis effect of motor bearing fault is poor when the effective data samples are insufficient under variable working conditions, a motor bearing fault diagnosis method based on deep migration learning is proposed. Firstly, the fault mechanism of motor bearing is analyzed, and the collected original vibration signal is transformed by SVD denoising wavelet packet transform to obtain a color two-dimensional time-frequency map conducive to the training of convolutional neural network; Secondly, the network is constructed, the structure and parameters are determined through training, and the over fitting is suppressed by data enhancement and dropout mechanism; Finally, transfer learning is introduced to freeze the trained network bottom structure, and fine tune the network top structure with small sample data under different working conditions. The example analysis shows that the introduction of transfer learning can realize the accurate classification of small samples under other working conditions, and solve the problem of poor fault diagnosis effect when there are insufficient samples in practical engineering application.