Partial Discharge Recognition Algorithm Based on Ultraviolet Image Features and Deep Learning Techniques

Introduction: Traditional partial discharge detection methods often rely on manual judgment, which suffers from high subjectivity and low efficiency. To improve the automatic recognition capability of ultraviolet partial discharge images during converter station inspection, this paper proposes an improved ResNet18 recognition method deployed on a quadruped robot platform. Methods: The ultraviolet imaging principle of partial discharge is first analyzed, and the collected ultraviolet images are preprocessed, and then model training is carried out. ResNet18 is adopted as the backbone network, with ImageNet pre-trained weights used for transfer learning to improve feature extraction under limited data conditions. Mixup data augmentation is applied to expand the effective sample distribution, and a Region of Interest (ROI)-guided spatial attention mechanism is integrated into the network to strengthen its focus on small discharge-related regions. Results: The proposed model was trained and evaluated on a self-built dataset containing 1200 ultraviolet images. The experimental results show that the improved model achieved an accuracy of 96.67% on the test set. To further verify its practical applicability, the trained model was deployed on a quadruped robot inspection platform and tested for two months at a converter station in Ningxia, China. The field recognition accuracy reached 94.23% in the converter transformer yard and 90.61% in the Direct Current (DC) area. The average inference time was approximately 5.3 ms per image, meeting the real-time requirements of on-site inspection. Discussion: Although the proposed model achieved good recognition performance, strong reflection and complex background interference may still affect field results. Further dataset expansion and robustness improvement are needed. conclusion: The framework demonstrates high recognition performance and robustness, and has been integrated into a quadruped-robot inspection system to support real-time partial discharge identification in practical field environments. The results confirm its potential for intelligent, non-contact condition monitoring of high-voltage power equipment. Conclusion: This paper proposes an improved ResNet18 model. Compared with traditional methods, the proposed model shows advantages in recognition accuracy, environmental adaptability, and processing speed. Combined with field tests, it is suitable for intelligent partial discharge inspection tasks in converter station scenarios.