A Feature Separation Prototype Network for Open Set Cross-Domain Fault Diagnostics
Shuai Tan, Xiaojian Lin, Qingchao Jiang, Xin Peng, Shumei Zhang
Abstract
In recent years, domain adaptation (DA) techniques have demonstrated significant advancements and promising potential in cross-domain fault diagnosis. However, most approaches assume a shared label space between the source and target domains. In real-world industrial scenarios, unknown faults may arise in the target domain, leading to limited knowledge available for transfer and the possibility of misalignment. To address these challenges, this study introduces a feature separation prototype network (FSPN), which can perform cross-domain fault diagnosis with few fault classes available. First, a feature separation module is constructed to separate domain-shared and domain-specific features. Then a weighted adversarial module is designed for finer domain alignment at the sample level. Finally, a sample separation module is used to construct representative fault prototypes (FPs) and separate unknown faults. Experimental results from three industrial cases validate the exceptional diagnostic performance of the proposed method in cross-domain fault diagnosis with inconsistent classes.