Article Details

Abstract

To address engineering challenges such as insufficient single-modality information representation, scarcity of labeled samples, and poor generalization capability across operating conditions in the fault diagnosis of electrical power equipment, this study proposes a multimodal fault diagnosis method integrating contrastive learning and transfer learning. An improved contrastive learning framework is pretrained on large-scale unlabeled multimodal data to extract operating-condition-invariant deep feature representations. On this basis, a transfer learning module based on the joint optimization of Maximum Mean Discrepancy (MMD) and Correlation Alignment (CORAL) is introduced to achieve model adaptation and cross-domain generalization across different types of distribution equipment and operating conditions. Experimental results show that the method achieves an average diagnostic accuracy of 95.3% when only 3～5 labeled samples per class are available in the target domain, representing an improvement of approximately 8.6% over traditional fine-tuning methods. The proposed method provides a feasible technical pathway for intelligent fault diagnosis of power equipment such as distribution terminals and primary-secondary fusion equipment.

Keywords

contrastive learning; transfer learning; electrical power equipment; multimodal fault diagnosis; cross-domain generalization

References

1. [1] Lu Feiyu, Tong Qingbin, Jiang Xuedong, etc An interpretable method for mechanical fault diagnosis based on conditional metric transfer learning [J]. Journal of Instrumentation, 2024, 45 (7): 250-262
2. [2] Zuo Jianhui, Shi Fu, Song Xuemeng, etc A method for evaluating transient voltage stability in power grids combining active learning and self training [J]. Power System Automation, 2026, 50 (2): 177-187
3. [3] Ye Jianbo, Yuan Peihan Design and Implementation of Real time Communication System for Electrical Construction Sites Based on WebRTC Hybrid Architecture [J]. Modern Information Technology, 2025, 9 (15): 14-18
4. [4] Xiang Ling, Wang Ning, Bing Hankun, etc A rotor fault diagnosis method based on multi-scale residual dynamic domain adaptive network under different operating conditions [J]. Journal of Vibration Engineering, 2026, 39 (2): 595-604
5. [5] Ding Wenhua, Du Junwei, Hou Lei, etc Fashion content and style transfer based on generative adversarial networks [J]. Computer Engineering and Applications, 2024, 60 (9): 261-271
6. [6] Qiao Qi, Wang Hongjun, Ma Kang, etc Gas turbine rotor fault diagnosis based on improved DenseNet ViT joint network and transfer learning [J]. Journal of Electronic Measurement and Instrumentation, 2024, 38 (11): 40-47
7. [7] Zheng Hanbo, Hu Junhao, Li Jinheng, etc Research on Thermal Fault Diagnosis of Power Equipment Based on Improved Instance Segmentation and Deep Neural Network Temperature Fitting [J]. Journal of Motor and Control, 2025, 29 (2): 116-125
8. [8] Xu H Z，Xu C，Xu J T. Multi-sensor signals augmented multimodal MAML-1DCNN-RBEAM deep transfer learning algorithm for wind turbine bearing fault diagnosis [J].Nondestructive Testing and Evaluation，2026，41（4）：1897-1925.
9. [9] Xie Fengyun, Dong Jiankun, Fu Yu, etc Transfer learning and CNN based motor fault diagnosis methods [J]. Mechanical Science and Technology, 2024, 43 (3): 513-519
10. [10] Cai Qinghuai, Luo Qingquan, Yu Tao, etc Generalization method for load identification based on active transfer learning [J]. Electric Power Automation Equipment, 2025, 45 (3): 209-216

About the authors

Guo Xiaogang (1990-), male, Han ethnicity, from Pingdingshan, Henan Province, China, intermediate engineer, bachelor's degree, research direction: research on medium voltage gas insulated switchgear.