Abstract
ABSTRACT
Reliable bearing fault diagnosis is critical for the safety of rotating machinery. Traditional closed‐set diagnostic methods assume that the training set encompasses all possible fault types, leading to frequent misclassifications of unknown faults in real‐world industrial scenarios. Furthermore, existing Open Set Fault Diagnosis (OSFD) approaches largely rely on static boundaries or passive probability estimation, lacking adaptive mechanisms for active unknown fault discrimination. To address these limitations, this paper proposes the HabitOSR‐Net framework, inspired by human cognitive habits regarding unknown class discrimination. By jointly optimizing known class classification, virtual classifiers, and mixed feature loss, the framework enhances unknown fault identification capabilities while maintaining classification accuracy for known faults. First, drawing on the cognitive process of “rejecting the known to accept the unknown,” we propose a strategy to cultivate an “unknown‐priority” discrimination habit. This involves masking known class labels and forcing their classification into virtual unknown categories. Second, a habit activation mechanism is constructed by introducing virtual classifiers at the output layer to act as “receivers” for unknown classes. This combines mask training with random virtual target assignment to solidify the model's responsiveness to unknown faults. Finally, we implement feature mixing enhancement by generating mixed features of known classes via Beta distribution linear interpolation and adding Gaussian noise. This simulates the distribution of unknown faults to optimize decision boundaries. Validated across three datasets, the proposed method achieves higher accuracy than existing state‐of‐the‐art methods and demonstrates adaptive decision boundaries.