摘要详情

ID / 提交时间

257 / 2024-07-12 15:26:34

标题

Breakdown Mechanism Classification in Conditioning Based on Two Deep Learning Models

关键字

Breakdown mechanism; Deep learning; VGG16; ResNet; Convolutional block attention module (CBAM); Conditioning

主题及专题

7)Fundamental physics in Switchgear

状态

终稿

作者

Xu Xunchen / Nanjing University of Aeronautics and Astronautics

Jie Fang / Nanjing University of Aeronautics and Astronautics

Shimin LI / Nanjing University of Aeronautics and Astronautics

Chaohai Zhang / Nanjing University of Aeronautics and Astronautics

Yingsan Geng / Xi’an Jiaotong University；State Key Laboratory of Electric Power Equipment

zhiyuan liu / Xi’an Jiaotong University；State Key Laboratory of Electric Power Equipment

Jianhua Wang / Xi'an Jiaotong University

xiao zhang / AVIC

摘要

Breakdown mechanism classification is important for vacuum conditioning physics, which could improve the vacuum circuit breaker insulation ability. Deep learning is an effective technology to classify the breakdown mechanism. This paper aims to compare the breakdown mechanism classification performance by the VGG16 and ResNet-CBAM. The VGG16 and ResNet-CBAM are tested by three groups of voltage and current waveforms for breakdown mechanism classification. The accuracies of VGG16 are 79.38%, 89.84% and 89.05%, while the accuracies of ResNet-CBAM are 83.69%, 91.71% and 91.90%. The precision, recall and F1-score of ResNet-CBAM, which reach 82.99%, are also higher than the VGG16’s (≥ 79.01%). The VGG16 takes 15 hours to complete the training and classifies one breakdown with 5 seconds. The ResNet-CBAM takes 3 hours to complete the training and classifies one breakdown with 1 second, which are one-fifth of the VGG16’s. The parameters for operation and storage of VGG16 are 138 M and 524 M respectively. The parameters for operation and storage of ResNet-CBAM are 22.7 M and 83.3 M respectively, which are about one-sixth of the VGG16’s. The results show that the ResNet-CBAM has a better performance for breakdown mechanism classification compared to VGG16.