Robust open-set partial discharge diagnosis based on hybrid supervised contrastive learning and SVM framework

Abstract

Automated partial discharge (PD) diagnosis using machine learning models is useful for high-voltage equipment (HVE) insulation condition monitoring. However, without a mechanism to identify unknown PD classes (defined as new classes not present in the training data), models will misclassify unknown classes as one of the known classes. To address this, a novel hybrid open-set recognition (OSR) framework based on Supervised Contrastive Learning (SupCon) is proposed to address a previously unexplored direction in the PD diagnosis domain. The framework integrates discriminative representation learning with both unified and per-class rejection strategies using one-class classification, enabling effective separation of known and unknown PD classes. The main contribution is the synergistic integration of SupCon for constructing structured latent spaces, SVM for precise closed-set classification, and dual OCSVMs for adaptive unknown rejection, together forming a unified pipeline that achieves both fine-grained discrimination and robust unknown detection. To evaluate the effectiveness of the proposed framework, comprehensive experiments are conducted across 30 OSR tasks, covering 12 PD classes from three types of high-voltage equipment under varying openness levels. The proposed framework is benchmarked against five state-of-the-art approaches, including ArcFace, GAN-Flow, a convolutional neural network (CNN), Autoencoder, and Vision Transformer. Experimental results demonstrate that the proposed framework achieved the best performance, with a mean normalized accuracy of 97.66 % and a Youden’s index of 0.953, confirming its robustness, generalization capability, and potential to advance open-set PD diagnostic methodologies.