Multistrategy Progressive Adaptation for Generalized Open-Set Cross-Working Condition Fault Diagnosis in Rotating Machinery

Abstract

In real-world industrial environments, frequent condition changes, mechanical degradation, and incomplete fault labeling often lead to data distribution shifts and label space asymmetry between the source and target domains. Moreover, compounded by the emergence of previously unseen fault types, severely undermine the generalization capability of conventional intelligent diagnostic methods. Although existing open-set domain adaptation (OSDA) methods attempt to address unknown classes, most still rely on the assumption of full label space consistency, limiting their applicability under complex and uncertain industrial conditions. To overcome these limitations, this article proposes a novel model for generalized open-set cross-working condition fault diagnosis, named multistrategy progressive adaptation network (MSPAN). The model allows for partial class overlap between source and target domains, each containing private classes, thereby reducing reliance on strict label alignment, complete class coverage, and prior knowledge of the target domain. This significantly enhances the model’s adaptability and flexibility under realistic operating conditions. MSPAN integrates three core strategies: centroid-guided expansion (CGE), progressive consensus filtering (PCF), and localized knowledge integration (LKI). CGE expands the label space with pseudo-target domain samples, alleviating interference from private classes in the source domain; PCF combines dual measure-driven ranking and consensus-aware estimation to adaptively filter unknown classes in the target domain; LKI focuses on learning unknown-class representations while enhancing the transferability of general diagnostic knowledge across domains. Extensive experiments on the PU and NLN-ESP public datasets demonstrate that MSPAN achieves both accurate shared-class diagnosis and effective unknown-class separation under varying degrees of class overlap and openness. These results validate its robustness, generalization ability, and practical potential for deployment in complex industrial fault diagnosis scenarios.