Adapted Oversampling-based Multi-layer Support Vector Machines for interpretable multi-class imbalance fault...

In engineering practice, intelligent fault diagnosis for high-end equipment often involves small-sample, multi-class imbalanced data with noise and complex intra- and inter-class distributions. Existing sampling methods may generate low-quality samples, amplify noise, and depend heavily on hyperparameters. To address these issues, this paper proposes an interpretable fault diagnosis framework, termed Adapted Oversampling-based Multi-layer Support Vector Machines (AM-SVMs). The framework embeds a Multi-mechanism Adaptive Oversampling Technique (MAOTE) into a multi-layer LSSVM architecture. MAOTE adaptively determines sampling strategies according to data characteristics and employs a Newton-Raphson-inspired evolutionary mechanism to search for high-quality candidate solutions guided by multi-class classification performance. The optimized solutions are reorganized into diverse and representative fault samples, and a classifier-feedback-based evaluation mechanism improves distributional consistency and interpretability between generated and real samples. Finally, balanced feature samples are used to train a multi-class LSSVM classifier, yielding a robust and generalizable diagnostic model. Experiments on public bearing datasets and self-collected data show that AM-SVMs outperforms ten data augmentation methods and eight multi-class classifiers in diagnostic accuracy and robustness, demonstrating its effectiveness for imbalanced fault diagnosis in intelligent manufacturing.