Detecting Low-Intensity Cyber Attacks in Electric Drives via Contrastive Few-Shot Learning

This paper proposes a novel anomaly detection framework for intelligent motor drives. The framework integrates contrastive learning (CL) and transfer learning (TL) into a CNN-based architecture. The CL module enhances the discriminative capability of feature embeddings. It groups samples from identical operational states and enforces separation between samples collected under different conditions. This allows the model to effectively distinguish between normal and anomalous states. Meanwhile, TL enables the effective transfer of knowledge from abundant simulation data to limited real-world experimental data, improving generalization. Experiments on a motor drive testbed demonstrate that our method achieves 99.8% accuracy using the full training set and 99.69% accuracy with only 40% of the data—surpassing the TL-based model (99.0% and 89.0%,). In a challenging few-shot setting (12.5% of training data), it still maintains 91.6% accuracy, compared to 86.5% for TL. Moreover, our method yields a lower false alarm rate (2.1% vs. 4.5% for TL), showing superior robustness and data efficiency in real-world deployment scenarios.