摘要
考虑到万能式断路器触头系统机械故障是一个从轻微到重度的演变过程,准确识别其运行状态可以大大提高断路器的可靠性。提出一种单信号输入和多任务输出的MTL-SEResNet网络模型以兼顾故障诊断和程度评估。首先采用连续小波变换对触头系统振动信号进行时频分析,得到相应的二维时频图像;其次将SENet结构引入到改进的ResNet18中,利用多任务学习共享机制构建MTL-SEResNet网络模型;并通过调整故障分类和程度评估两个任务损失函数的权重比例,对模型进行优化;最后,通过模拟的触头系统的故障数据对所提方法进行实验验证。结果表明,模型的性能更佳,类型及程度准确率分别为99.78%和99.36%,可以有效地实现万能式断路器故障程度评估。
The mechanical fault of the contact system for a conventional circuit breaker is a process from slight to severe. The accurate identification of its operating state can greatly improve the reliability of the circuit breaker. In this article, a single signal input and multi-task output MTL-SEResNet model is proposed for fault diagnosis and degree evaluation. Firstly, the raw vibration signals of the contact system are analyzed using a continuous wavelet transform. And the corresponding two-dimensional time-frequency images are obtained. Secondly, the improved ResNet18 network is combined with the SENet structure, and the multi-task learning sharing mechanism is used to formulate the MTL-SEResNet model. The model is optimized by adjusting the weight ratio of the two task loss functions for fault classification and degree evaluation. Finally, the proposed method is verified by experiments with simulated fault data of the contact system. The results show that the proposed model has better performance with 99.78% and 99.36% accuracy in type and degree, respectively, which can effectively evaluate the fault degree of the conventional circuit breaker.
作者
孙曙光
张婷婷
王景芹
魏硕
邵旭
Sun Shuguang;Zhang Tingting;Wang Jingqin;Wei Shuo;Shao Xu(State Key Lab Reliability and Itelligence of Electrical Equipment,Hebei Unirersity of Technology,Tianjin 300130,China;School of Arificial Intelligence,Hebei Universily of Technology,Tianjin 300130,China)
出处
《仪器仪表学报》
EI
CAS
CSCD
北大核心
2022年第6期162-173,共12页
Chinese Journal of Scientific Instrument
基金
河北省自然科学基金(E2021202136)
河北省自然科学基金创新群体(E2020202142)项目资助。
关键词
万能式断路器
触头系统
故障程度评估
连续小波变换
多任务学习
conventional circuit breaker
contact system
fault degree evaluation
continuous wavelet transform
multi-task learning
