A Fault Diagnosis Method Based on Convolutional Decomposition and Domain Adversarial Transfer Learning for Rolling Bearing under Variable Operating Conditions

This paper proposes a rolling bearing fault diagnosis method for variable operating conditions based on convolutional decomposition and domain-adversarial migration learning. Through resampling method to balance the data samples of different categories in the dataset. The convolutional layer of the neural network is used to replace the wavelet packet transform, and the time-frequency decomposition parameters are optimized with the improved inter-domain loss to improve the time-frequency decomposition capability and reduce the time-frequency decomposition difference of migration learning for signals from different operating conditions. Multi-core feature convolution method is utilized to obtain features in the time-frequency matrix with higher correlation coefficients with migration learning fault diagnosis. Utilizing the Wasserstein distance to measure the disparity in joint distribution between domains and employing multi-kernel maximum mean discrepancy (MK_MMD) to improve domain adversarial neural network (DANN) loss function structure. This approach improves the model’s ability to fit domain distributions, reduces oscillations during optimization, and consequently enhances the model’s generalization capability and fault diagnosis accuracy in the target domain. The experimental results show that the migration learning fault diagnosis accuracy of the method proposed in this paper on the CWRU dataset is improved by 9.03% on average compared with the accuracy of the comparison models such as convolutional neural network and domain adversarial neural network, which shows an obvious comprehensive advantage compared with the comparison methods.