基于卷积神经网络的核电多相无刷励磁系统旋转整流器故障诊断  

Fault Diagnosis of Rotating Rectifier in Nuclear Multi-Phase Brushless Excitation System Based on Convolutional Neural Network

作　　者：梁郑秋 郝亮亮[1] 周艳真 段贤稳 王光[4] Liang Zhengqiu;Hao Liangliang;Zhou Yanzhen;Duan Xianwen;Wang Guang(School of Electrical Engineering Beijing Jiaotong University,Beijing 100044 China;Department of Electrical Engineering Tsinghua University,Beijing 100084 China;China Nuclear Power Operations Co.Ltd,Shenzhen 518172 China;Nanjing NR Electric Co.Ltd,Nanjing 211102 China)

机构地区：[1]北京交通大学电气工程学院,北京100044 [2]清华大学电机工程与应用电子技术系,北京100084 [3]中广核核电运营有限公司,深圳518172 [4]南京南瑞继保电气有限公司,南京211102

出　　处：《电工技术学报》2023年第20期5458-5472,共15页Transactions of China Electrotechnical Society

基　　金：中央高校基本科研业务费专项资金项目(2020JBM070);中广核集团公司科技项目(3100077013)资助。

摘　　要：多相无刷励磁系统广泛应用于大容量核电机组,对其旋转整流器故障的准确判断是保障机组安全稳定运行的重要前提,但是现有方法难以准确区分所有类型故障。为此,该文提出基于卷积神经网络的多相无刷励磁系统旋转整流器故障诊断方案。在给出故障诊断信号选取的理论基础上,将励磁电流作为输入特征,采用一维空洞卷积神经网络实现励磁电流的特征提取和故障模式分类,并结合置信度指标对诊断结果的可靠程度进行评价。进一步使用Score-CAM类激活映射算法分析模型的诊断机制,兼顾了准确性和可解释性。通过11相无刷励磁系统的动模实验进行验证,结果表明,所提方法能够准确区分旋转整流器的不同二极管开路故障模式,在含噪声情况下仍然有较高的准确率,具有良好的应用前景。The multi-phase brushless excitation system is widely used in large-capacity nuclear power plants.The safe and stable operation of nuclear power plants depends on accurately diagnosing rotating rectifier faults.However,the existing methods cannot accurately distinguish all types of faults,which makes it challenging to meet the demands in practice.Recently,as a representative artificial intelligence method,deep learning has provided new ideas for solving this problem.Therefore,this paper develops a fault diagnosis scheme for the rotating rectifier based on a convolutional neural network.First,taking an 11-phase brushless excitation system as an example,the excitation current characteristics after the rotating rectifier fault are analyzed.The results show that the excitation current waveforms contain amplitude and phase information of all fault harmonics,which can be varied under different fault modes and used for fault diagnosis of the rotating rectifier.Then,a one-dimensional dilated convolutional neural network(1D-DCNN)model is introduced to realize fault diagnosis.The 1D-DCNN establishes a mapping relationship from the input(excitation current waveforms)to the output(fault mode).Meanwhile,the difference between the largest and second-largest probability of the output is defined as the confidence based on the conditional probability of samples belonging to different fault modes given by the model,which is used to measure the reliability of the diagnosis results.The process of fault diagnosis based on the 1D-DCNN model includes three stages.At stage 1,raw data at several voltage levels are acquired to ensure the model has applicability and generalizability under different voltage levels.Then,the data are preprocessed regarding sliding window segmentation,filtering,and mean normalization.At stage 2,the 1D-DCNN model is trained to diagnose rotating rectifier faults online.At stage 3,the Score-CAM method is introduced to clarify the interpretability of the 1D-DCNN model.Finally,the prototype experiment verifies the

关 键 词：无刷励磁系统 旋转整流器 深度学习 卷积神经网络 故障诊断 

分 类 号：TM341[电气工程—电机]