Feature Disentanglement Augmented Network for Vibration-Based Fault Diagnosis of Rotating Machinery Under Unknown Operating Conditions
Accurate fault diagnosis of rotating machinery based on vibration measurements is essential for reliable condition monitoring and predictive maintenance in industrial systems. However, conventional data-driven diagnostic models often experience significant performance degradation under varying operating conditions, primarily due to distribution shifts in the acquired vibration signals. This article proposes a feature disentanglement augmented network (FDAN) to improve the robustness and generalization capability of vibration-based fault diagnosis for rotating machinery to unseen target domains. In FDAN, a disentanglement network is employed to explicitly separate the input features into domain-invariant and domain-specific components. An independently trained attention module is further introduced to identify the most classification-relevant components, which are subsequently used for augmentation to enrich the diversity of the learned representations. The proposed method requires no test-condition data during training, making it well-suited for industrial scenarios. Extensive experiments conducted on the widely used DIRG and SDUST fault datasets demonstrate that FDAN consistently outperforms several domain generalization (DG) approaches, achieving significant improvements in diagnostic accuracy under varying working conditions. Additional validations, including feature analysis and ablation studies, confirm the effectiveness and necessity of each component. These results demonstrate the potential of FDAN as an effective and deployable solution for robust intelligent fault diagnosis in vibration measurement-based monitoring of rotating machinery in industrial applications.