A Novel Capsule Network for Multi-source Partial Discharge Diagnosis in Gas-insulated Switchgear
Wang Yanxin,Yan Jing,Wang Jianhua,Geng Yingsan
DOI: https://doi.org/10.19595/j.cnki.1000-6753.tces.220104
2024-01-01
Abstract:Monitoring and diagnosing partial discharge in gas-insulated switchgear（GIS） is an effective means to evaluate its insulation state. Deep learning methods have shown excellent performance in GIS partial discharge diagnosis in recent years because they can automatically learn the nonlinear relationship between GIS partial discharge signals and fault types. However, the existing deep learning methods are all trained on the premise of massive data, and the model’s performance decreases severely when the amount of data decreases.Therefore, this paper proposes a novel Siamese network for diagnosing the few-shot partial discharge of GIS. The meta-training approach can achieve on-site small-sample GIS partial discharge（PD） diagnosis accurately and robustly.Firstly, a multi-level second-order attention convolutional network is constructed to mine comprehensive feature information in partial discharge data and extract discriminative and representative features. Secondly, the metric learner is designed to learn the similarity of metric features. Then, the similarity metric compares or matches samples under the target task to achieve classification. Finally, a multi-level second-order attention Siamese network is designed for partial discharge diagnosis in GIS. Through the A-way K-shot training, the ability of model learning is endowed, and the learned meta-knowledge can be transferred to new tasks through fine-tuning, which solves the problem of small samples.The experimental results show that the GIS partial discharge diagnosis accuracy of the method proposed in this paper reaches 93.76 % when each type of support sample is 5, which is 57.34 %, 45.53 %, 25.97 %, 65.91 %,and 33.97 % higher than direct training, FT, DATL, TCNN, and SVM, respectively. Simultaneously, box plots and scatter plots show that the method proposed in this paper has the most negligible dispersion, indicating that it is robust. Moreover, through the confusion matrix and t-distribution stochastic neighbor embedding, compared with traditional methods, this method proposed in this paper effectively solves the problem of small sample diagnosis and improves diagnostic performance under small samples. In addition, the proposed method has strong fault tolerance for unbalanced samples.The following conclusions can be drawn:（1） The constructed Siamese network can realize high-precision and high-robust diagnosis of partial discharge in small-sample GIS on-site, especially for scenarios where UHV GIS samples are scarce.（2） The introduction of multi-scale convolution and second-order attention modules effectively improves the diagnostic accuracy of GIS partial discharge.（3） The training and testing of the method proposed in this paper are carried out under small samples, and the training is strictly in accordance with the A-way K-shot. K samples are randomly selected for each task from many training sets as the support set for model training, which has a strong tolerance for unbalanced samples.