ArcFace-enhanced few-shot meta-learning for industrial fault diagnosis using multi-modal representations with lightweight 1D CNNs

Machine Health Monitoring Systems (MHMS) often struggle in industrial environments due to scarce labelled data, limited fault classes, and the overfitting tendency of large deep models. This study introduces a lightweight multi-Skip 1D CNN (MS-1D-CNN) built with depth-wise separable convolutions and skip connections, coupled with an ArcFace-enhanced Few-Shot Learning (FSL) framework and multi-modal feature extraction to ensure discriminative and robust representations under limited data. ArcFace enforces margin-based class separation, while episodic FSL training enhances generalization with only a few samples per class. The framework employs a two-stage protocol: meta-training on the clean CWRU dataset to learn stable fault-related patterns, followed by meta-testing on noisy industrial vibration data to evaluate real-world generalization without overfitting. Both single-sensor signals and image-based representations are utilized during evaluation to verify the framework’s adaptability across various data modalities. The proposed framework achieves near-perfect accuracy with only 1–5 samples per class, confirming their robustness, strong generalization capability, and suitability for deployment in data-constrained industrial MHMS applications.