Joint Open-Set and Closed-Set Classifier-Guided Domain-Invariant Learning for Universal Compressor Component Fault Diagnosis

Advancing the practical application of intelligent fault diagnosis algorithms in engineering is a key objective of their rapid development. Considering the limited exploration of generalized fault diagnosis in existing intelligent diagnostic methods, this study proposes a universal domain adaptation (UDA) Method guided by joint open-set and closed-set classifiers for domain-invariant learning (JCCGDIL). Initially, invariant learning is introduced to enhance the clustering of features belonging to identical fault categories, combined with a self-paced neighborhood search strategy and weighting mechanisms to extract highly discriminative target features. Subsequently, to address the issue of class misalignment in target features, an interpolation-based feature augmentation approach is employed, generating auxiliary features by blending source and target domain features to improve the classifier’s recognition capability for unknown categories. Finally, to prevent degradation in the model’s recognition of known categories, a mutual learning mechanism between the open-set and closed-set classifiers is constructed, achieving a superior balance in identifying both known and unknown categories. Extensive experiments conducted on two benchmark datasets demonstrate the superiority of JCCGDIL. The proposed method achieves an average diagnostic accuracy of 88.77% and 93.00% under two different domain adaptation (DA) scenarios, outperforming state-of-the-art baselines such as EDLAN (87.39% and 91.70%).