A Dual-Component Elastic Adaptation Network for Rotating Machinery Incremental Fault Diagnosis under Variable Operating Conditions

In practical industrial environments, the data distribution of rotating machinery drifts as operating conditions vary, causing a marked deterioration in the performance of traditional fault diagnosis methods that rely on the assumption of identical distributions. Incremental learning provides a promising pathway to address dynamic operating conditions. However, existing approaches typically depend on replaying historical data and still struggle to strike a balance between stability and plasticity. To overcome these limitations, this paper proposes a dual-component elastic adaptive network (DCEAN) designed for incremental fault diagnosis of rotating machinery under varying working conditions. The proposed framework operates without access to previous data and simultaneously achieves knowledge retention and feature correction. Specifically, a sensitive parameter constraint (SPC) mechanism is introduced to curb excessive updates to parameters identified as critical, thereby stabilizing previously learned knowledge. In parallel, a feature drift self-calibration (FDSC) mechanism is employed to estimate and compensate for distribution shifts induced by condition variations, promoting consistency of feature representations across domains. Through the coordinated action of these two mechanisms, DCEAN establishes an incremental learning paradigm that harmonizes stability with adaptability. Two case studies demonstrate that the proposed method delivers superior diagnostic performance in variable operating environments, underscoring its robustness and effectiveness.