电工技术学报  2024, Vol. 39 Issue (13): 4153-4163    DOI: 10.19595/j.cnki.1000-6753.tces.230617
电器装备及智能化 |
基于深度学习的冲击电压老炼过程中真空击穿机制甄别优化方法
李世民, 徐勋晨, 张潮海
南京航空航天大学多电飞机电气系统工信部重点实验室 南京 210016
Optimization Method for Classifying Breakdown Mechanism During Impulse Voltage Conditioning Process Based on Deep Learning
Li Shimin, Xu Xunchen, Zhang Chaohai
Center for More-Electric-Aircraft Power System Nanjing University of Aeronautics and Astronautics Nanjing 210016 China
全文: PDF (6998 KB)   HTML
输出: BibTeX | EndNote (RIS)      
摘要 冲击电压老炼技术是提高真空电极间隙绝缘能力的有效手段,快速准确地甄别真空击穿机制对揭示冲击电压老炼过程的物理演化机理具有重要意义。该文提出了一种基于深度学习的通过突显击穿前过程提高真空击穿机制甄别准确度的优化方法。对五对相同的无氧铜球形电极开展同样的冲击电压老炼试验,分别获得时间提取范围为0~400 μs的完整击穿电压电流波形样本和0~200 μs的突显击穿前过程的击穿电压电流波形样本,并通过深度学习模型对两种击穿电压电流波形样本开展脉冲电流诱发、场致发射诱发和微粒诱发三种真空击穿机制的甄别训练与测试,并将测试结果与真实结果进行对比分析与评估。结果显示:时间提取范围为0~200 μs的突显击穿前过程的击穿电压电流波形样本的击穿机制甄别准确率均在87.99%以上,平均提高了2.55%,其对应的精确率、召回率和F1分数均更优。研究结果表明,突显击穿前过程的击穿电压电流波形处理能够有效地优化真空击穿机制甄别的效果,具有良好的工程应用前景。
服务
把本文推荐给朋友
加入我的书架
加入引用管理器
E-mail Alert
RSS
作者相关文章
李世民
徐勋晨
张潮海
关键词 冲击电压老炼击穿机制深度学习突显击穿前过程击穿波形    
Abstract:Impulse voltage conditioning technology is an effective means to improve the insulation ability of vacuum circuit breaker (VCB). Classifying the breakdown mechanism quickly and accurately has a great significance to reveal the physical evolution of impulse voltage conditioning and improve the VCB withstanding voltage level. The traditional method to classify the breakdown mechanism needs to eliminate the displacement current through mathematical compensation algorithm and fit the Fowler-Nordheim formula, which is complicated to obtain the breakdown mechanism. Deep learning has an obvious advantage in image recognition and feature extraction. In this paper, an optimized method to classify the breakdown mechanism was proposed through enlarging the pre-breakdown period in breakdown waveform based on deep learning.
Five identical sphere oxygen-free copper electrode pairs A, B, C, D and E were applied the same impulse conditioning. All the breakdown waveforms were processed into two kinds: 0~400 μs, containing the whole breakdown waveform, and 0~200 μs, pre-breakdown period enlarged breakdown waveform. The corresponding breakdown mechanisms of A and B were labeled as pulsed current induced vacuum breakdown (PB), field emission induced breakdown (FEBD) and particle induced vacuum breakdown (PBD) through the traditional method. Then, breakdown waveforms of A and B (1 530) in 0~400 μs and 0~200 μs were for the breakdown mechanism classification training, and breakdown waveforms of C, D and E (1 398) in 0~400 μs and 0~200 μs were for breakdown mechanism classification test, respectively. The corresponding breakdown mechanisms of C, D and E were classified into PB, FEBD and PBD with deep learning. In addition, the breakdown mechanisms of C, D and E were also obtained through the traditional method. The deep learning outputs were compared with that through the traditional method. The test results were evaluated and analyzed by the evaluation parameters such as precision, recall, F1-score and so on.
The results showed that the breakdown mechanism classification accuracies of C, D and E (0~200 μs) were 88.92%, 87.99% and 92.78%, respectively, and all the accuracies of 0~200 μs were higher than 87.99%. The breakdown mechanism classification accuracies of C, D and E (0~400 μs) were 85.23%, 84.90% and 91.90%, respectively. Compared with 0~400 μs, the breakdown mechanism classification accuracies of 0~200 μs were improved by 3.69%, 3.09% and 0.88%, respectively. The accuracy of 0~200 μs had an average improvement by 2.55% than that of 0~400 μs. Precision, recall and F1-score of 0~200 μs were also higher than those of 0~400 μs. The results showed that 0~200 μs, pre-breakdown period enlarged breakdown waveform had a better performance in breakdown mechanism classification.
Conclusions were drawn as following: (1) The classification accuracy for breakdown mechanism through deep learning could be improved by enlarging the pre-breakdown period in the breakdown waveform. (2) The breakdown mechanism classification can be completed quickly and accurately, whose accuracy could be higher than 87.99% with the effectiveness verified by precision, recall and F1-score. It has a theoretical guidance for a promising conditioning technology to improve the VCB voltage level in industry application.
Key wordsImpulse voltage conditioning    breakdown mechanism    deep learning    pre-breakdown period enlarged    breakdown waveform   
收稿日期: 2023-05-04     
PACS: TM561.2  
基金资助:国家自然科学基金(52207162)、江苏省自然科学基金(BK20210307)和中央高校基本科研业务费专项资金(NJ2023012, NJ2023014)资助项目
通讯作者: 李世民 男,1988年生,讲师,硕士生导师,研究方向为高压真空断路器、真空放电与绝缘、深度学习、等离子体模拟等。E-mail:dianqilishimin@163.com   
作者简介: 徐勋晨 男,2000年生,硕士研究生,研究方向为真空击穿与深度学习。E-mail: xuxunchen@nuaa.edu.cn
引用本文:   
李世民, 徐勋晨, 张潮海. 基于深度学习的冲击电压老炼过程中真空击穿机制甄别优化方法[J]. 电工技术学报, 2024, 39(13): 4153-4163. Li Shimin, Xu Xunchen, Zhang Chaohai. Optimization Method for Classifying Breakdown Mechanism During Impulse Voltage Conditioning Process Based on Deep Learning. Transactions of China Electrotechnical Society, 2024, 39(13): 4153-4163.
链接本文:  
https://dgjsxb.ces-transaction.com/CN/10.19595/j.cnki.1000-6753.tces.230617          https://dgjsxb.ces-transaction.com/CN/Y2024/V39/I13/4153