Abstract: The complex and variable operating conditions of mechanical equipment not only lead to significant domain shift in the distribution of state data, but also pose great challenges for diagnostic models to identify unknown faults under unseen conditions. Therefore, enhancing the generalization ability of intelligent diagnostic models in open-set scenarios is of great importance for effective real-time diagnosis in practical engineering applications. Inspired by how humans learn new things, this paper proposes a perception proxy-based approach for open-set domain generalization in mechanical intelligent fault diagnosis, aiming to improve the effectiveness of diagnosing mechanical faults under unknown conditions. Specifically, the method employs known class perception adversarial learning strategies to eliminate the assumption that training samples are pre-set as unknown categories, thereby establishing decision boundaries between known and unknown fault classes. Additionally, the approach constructs unknown class proxies and designs an open-set proxy contrastive learning strategy to enhance sample feature compactness within classes and discriminability between classes. Experimental results on two mechanical fault datasets demonstrate that the proposed method outperforms existing advanced contrastive techniques in open-set domain generalization fault diagnosis tasks.

Key words: adversarial learning, contrastive learning, mechanical fault diagnosis, open set domain generalization