Deep learning-assisted compound fault diagnosis method of rotating machinery under limited sample and noisy environments

Rotating machinery often operates under complex working conditions where compound faults frequently occur. However, diagnosing such compound faults under limited samples and noise scenarios remains a significant challenge in industrial applications. To address these issues, this paper proposes a novel model—Siamese comb filter network. Specifically, comb filter network is employed to enhance the feature extraction capability, which capture the fault characteristics frequency components and mitigate the unrelated information simultaneously. Besides, Siamese neural network (Siamese Network) is introduced to build a feature space under limited sample conditions through metric learning, improving the ability of the model to discern sample similarities. By imposing loss functions on each branch, a multi-loss optimization strategy is deployed to guide feature extraction, facilitates more focused and discriminative feature learning through meaningful representations. Comprehensive experimental results on the laboratory-collected dataset and publicly available bearing dataset reveal that our proposed method outperforms comparative deep learning models under limited samples and noise environment.