Adversarial Domain-adapted Bearing Fault Diagnosis of Variable Working Conditions Based on ResNet and LIMMD

Rolling bearings are used to support rotating shafts, reduce friction and ensure smooth operation of equipment. In industrial applications, bearing fault diagnosis often faces severe challenges due to the complexity of actual working conditions, load and rotational speed fluctuations, etc. Deep features are difficult to be extracted, poor sample correlation between domains, and possible negative transfer, greatly affecting the model’s diagnostic accuracy. The domain adaptation method narrows the gap between working conditions by aligning the feature distributions of two domains to realize fault diagnosis, but the existing methods have the problems of difficulty in extracting the deep features of the data and distribution bias. To address the existing problems, this paper proposes an adversarial domain adaptation framework based on Residual Network (ResNet) and Local Inter-class Maximum Mean Discrepancy (LIMMD). The improved ResNet can effectively extract domain-invariant features, and the Adversarial Domain Adaptation Module utilizes adversarial learning and LIMMD to optimize the feature distribution, thus improving the diagnostic accuracy and robustness. Finally, validated on the public bearing dataset as well as on the selfbuilt bearing dataset, respectively.