Transformer partial discharge identification via supervised contrastive learning and Multi-scale perceptual networks

Accurate partial discharge (PD) identification is essential for the operational integrity of traction transformers in electrified railways, yet intense electromagnetic interference and similar discharge patterns often diminish diagnostic robustness. This study proposes a framework integrating supervised contrastive learning with a Multi-scale Perceptual Network (MSDP-Net) to enhance diagnostic reliability. A dual-view adaptive augmentation strategy, involving Gaussian noise injection and temporal shuffling, is first applied to improve invariant representation learning. The physics-informed MSDP-Net utilizes a morphological branch with dilated convolutions for long-range envelopes and an instantaneous branch with singularity-aware operators for energy mutation detection. Supervised contrastive learning is then employed to optimize the latent feature space, strengthening intra-class compactness. Additionally, a new electromagnetic pulse amplitude sequence dataset is established and publicly released to address the scarcity of PD benchmarks. Experimental results demonstrate that the proposed method achieves 98.42% identification accuracy across four typical defects, surpassing both general-purpose deep architectures and re-implemented PD-specific strategies. To exclude session-level information leakage, run-wise group-wise and leave-one-run-out cross-validation, in which all pulses of a given experimental run are confined to either the training or the test partition, are additionally conducted and retain 97.29% and 96.88% accuracy respectively. The framework exhibits exceptional robustness in high-noise environments, significantly outperforming current state-of-the-art models and offering a more reliable solution for railway power supply monitoring. Finally, a complexity analysis shows that the framework requires only 0.0882 GFLOPs and 0.94 ms per-pulse inference latency, and that the contrastive projection head is discarded at inference and therefore adds no deployment cost.