Collaborative and Conditional Deep Adversarial Network for Intelligent Bearing Fault Diagnosis

Abstract

In real applications, the data distribution may be inconsistent due to the different working conditions of the equipment, resulting in significant performance degradation of a well-trained faults diagnosis model. To this end, unsupervised domain adaptation (UDA) technique is proposed to bridge the distribution gap between the source and the target domain. Most studies on UDA-based faults diagnosis put much attention on extracting domain-invariant features without considering how to improve the discriminant capability of these features in the target domain. In this study, the Collaborative and Conditional Deep Adversarial Network (CCDAN) is proposed for cross-domain bearing faults diagnosis. In CCDAN, characteristic fault features at different scales are learned and utilized for domain adaptation. The domain classifier at each scale is fed with discriminative information conveyed in the label predictions, resulting in better alignment of feature distributions per class between different domains. By collaborative and adversarial learning across different scales, the learned features are expected to be domain-invariant and also discriminant for faults diagnosis in the target domain. In addition, the multiple kernel variant of maximum mean discrepancies (MK-MMD) is introduced to further reduce the feature distribution discrepancy. Experiments performed on two different datasets proved the effectiveness and superiority of our model.