An Improved Transfer Learning Method for Bearing Diagnosis under Variable Working Conditions Based on Dilated Convolution

摘要

Although intelligent fault diagnosis methods of rolling bearings have been extensively developed, the diagnosis results are usually disturbed by the change of working conditions and environmental noise, which will lead to significant decrease in diagnostic accuracy. Most methods assume that training data and testing data are under the same distribution but it is always impractical in real-world production. To improve diagnostic performance, a new transfer diagnosis method is proposed. On the basis of simultaneously aligning the marginal distribution and conditional distribution of datasets, the entropy penalty is added to the objective function to improve the separability of inter-class features. The dilated convolutional layer with a jagged expansion rate is added to the diagnostic model, and it will better extract the local and global features of the raw signal. The experimental results show that the proposed method can achieve a higher diagnostic accuracy under variable working conditions comparing with related methods and can acquire good robustness under the interference of different intensity gaussian noise.