Duplex adversarial domain discriminative network for cross-domain partial transfer fault diagnosis

Domain-adaptation technologies have been widely developed for mechanical fault diagnosis. Most related methods assume the same label space between the source and target domains, whereas the partial domain adaptation problem is more widespread and challenging. Moreover, most previous studies focused on data distribution alignment but failed to address fine-grained distributional differences of subdomains. Hence, this study proposes a novel duplex adversarial deep discriminative network (DADDN) for fault diagnosis in cross-domain partial transfer cases. First, a dual-domain adversarial attention mechanism is adopted in the DADDN to discriminate the label space and evaluate the transferability of the data samples. Second, a new metric function of the distribution difference is employed to learn both the fine-grained and discriminative features of the data samples and a weight-joint adaptive distribution mechanism is proposed to improve the domain confusion capability. Third, a center-of-balance weighting strategy is utilized to expand the categories of the target domain and reduce label space asymmetry. In addition, to avoid negative transfer, a weighting mechanism is designed to filter outlier categories, and a dynamic adaptive factor is introduced to weigh the importance of the marginal and conditional distributions. Finally, the effectiveness, robustness, and progressiveness of the proposed method are validated using several partial-transfer diagnostic tasks based on three rotating machinery datasets. (c) 2023 Elsevier B.V. All rights reserved.

Automated summary

This study proposes a novel duplex adversarial deep discriminative network (DADDN) for fault diagnosis in cross-domain partial transfer cases. By adopting a dual-domain adversarial attention mechanism, a new metric function of the distribution difference, and a center-of-balance weighting strategy, the accuracy and robustness of fault diagnosis are improved.