This paper proposes a physics-informed few-shot anomaly detection framework for intelligent motor drives by inte- grating contrastive learning into a CNN-based architecture. The framework leverages physics-informed parameters—specifically, second-harmonic components of stator current spectra—to guide the embedding space learned through CL. This enhances the model’s ability to discriminate between normal and anomalous states, particularly under low-data conditions. All training cases assume a few-shot label setting, where only a limited number of labeled attack samples are available. TL is first applied to transfer generalizable knowledge from simulation data to real-world domains, and CL is then fine-tuned with a physics-weighted loss to emphasize subtle anomalies. The proposed method provides a practical and interpretable solution for cyber-attack detection in intelligent motor drives with minimal labeled data.
