17 January 2023 in IEEE Transactions on Industrial Informatics
End-to-end intelligent diagnosis of rotating machinery under speed fluctuation and limited samples is challenging in industrial practice. The existing limited samples methods usually focus on the data distribution or learning strategy with particularity. Generative adversarial network (GAN) provides a data generation solution with portability in fault diagnosis with limited samples. However, GAN has problems with gradient vanishing, weak extraction of global features, and redundant training. This article proposes a dual-threshold attention-guided GAN (DTAGAN) to generate high-quality infrared thermal (IRT) images to assist fault diagnosis. First, Wasserstein distance and gradient penalty are combined to design loss function to avoid gradient vanishing. Second, attention-guided GAN is constructed to extract global thermal-correlation features of IRT images. Finally, dual-threshold training mechanism is developed to improve the generation quality and training efficiency. The comparative experiments show that DTAGAN is superior to comparison methods in fault diagnosis of rotor-bearing system under speed fluctuation and limited samples.
20 November 2024 in IEEE/ASME Transactions on Mechatronics
As a complex industrial system, wind turbine (WT) will inevitably experience new faults during long-term operation. Incremental fault diagnosis can continuously accumulate new fault knowledge from data streams, thereby expanding the model's diagnostic capabilities and overcoming catastrophic forgetting. However, when dealing with complex multicomponent faults of WT and limited incremental fault samples, the incremental diagnostic model still suffers from the “stability-plasticity” dilemma. Therefore, a few-shot class-incremental learning (FSCIL) method called forward-back compatible representation is proposed to realize WT system-level fault diagnosis. First, a forward virtual prototype is designed to reserve sufficient embedding space for limited new classes by clustering the intraclass feature distribution in the base task. Second, a backward memory bank is constructed to avoid catastrophic forgetting during the multicomponent incremental process by refining representative old diagnostic knowledge. In the WT system-level fault diagnosis experiment, the superiority and efficiency of the proposed method in the FSCIL scenario are compared with mainstream incremental learning methods and validation through the ablation experiment.
23 April 2025 in IEEE Transactions on Transportation Electrification
Transmission systems are prone to developing new faults over long-term operations. Incremental learning offers a solution by continuously acquiring new knowledge from online data streams while preserving existing knowledge, making it ideal for ongoing diagnosis in industrial settings. However, the scarcity of incremental fault data and the time-intensive updating of incremental diagnosis methods reduce the effectiveness of existing models in real-time incremental diagnosis. To address this issue, a Cosine-represent Real-time Evolved Network (CREN) is proposed to cope with the challenge of incremental few-shot fault diagnosis (IFFDs) for transmission systems. First, a cosine-represented learning strategy is employed to achieve bias imputation in few-shot scenarios by redesigning the mapping between features and weights. Second, a continuous evolution classifier is constructed to achieve real-time incremental updating by embedding decoupled nonparametric class prototypes into the classification weights. The efficiency of the proposed method is verified through IFFD experiments on the two transmission systems.