Fault Diagnosis Method of Motor Bearing Based on Improved Residual Networks and Transfer Learning

To overcome the limitations of conventional deep learning models in fault diagnosis under diverse operating conditions, a new bearing fault detection technique combining SE-ResNet and transfer learning is introduced. Raw vibration data is processed using wavelet transformation to create a color 2D time-frequency plot, which facilitates the training of convolutional neural networks. The Squeeze and Excitation Networks, which utilize an attention mechanism, are integrated into the residual blocks of the network. The resulting 2D images are then used to train the SE-ResNet, with the network's architecture and parameters being established during this phase. Transfer learning is applied to lock the core structure of the trained network while adjusting the parameters of the fully connected layer to adapt to different operating conditions and classify various fault types. The findings indicate that SE-ResNet achieves higher accuracy and better generalization across varying conditions and for fault diagnosis across different equipment.