航空学报 > 2022, Vol. 43 Issue (3): 325121-325121   doi: 10.7527/S1000-6893.2021.25121

基于OCKELM与增量学习的在线故障检测方法

戴金玲1, 许爱强1, 申江江2, 王树友1   

  1. 1. 海军航空大学 航空作战勤务学院, 烟台 266001;
    2. 91206部队, 青岛 266108
  • 收稿日期:2020-12-18 修回日期:2021-01-26 出版日期:2022-03-15 发布日期:2021-02-08
  • 通讯作者: 许爱强 E-mail:1159416532@qq.com

Online fault detection method based on incremental learning and OCKELM

DAI Jinling1, XU Aiqiang1, SHEN Jiangjiang2, WANG Shuyou1   

  1. 1. Naval Aviation University, Yantai 266001, China;
    2. No. 91206 Troops of PLA, Qingdao 266108, China
  • Received:2020-12-18 Revised:2021-01-26 Online:2022-03-15 Published:2021-02-08

摘要: 针对航空电子设备故障检测样本少、以及缺乏在线实时检测理论研究的问题,将一类核极限学习机(OCKELM)和数据增量学习(IL)相结合,实现对贯序来临的样本数据在线故障检测。基于正常状态下的样本数据,给出了OCKELM的核化形式,并推导了核函数和核权重向量的表达式;根据增量学习方法,在吸收新样本时更新核权重向量并估计样本输出值;最终基于2种阈值准则给出模型的检验阈值,对测试样本进行在线故障检测。将所提方法应用于UCI数据集和某航空电子设备的测试数据,实验结果表明:该方法的时间消耗在毫秒级别,实现了在线检测;且相比于现有的SVDD、PCA、OC-SVM方法,该方法在F1、AUC、G-mean和故障检测率等性能指标方面均表现优异。

关键词: 一类分类, 核极限学习机, 增量学习, 在线检测, 故障检测

Abstract: To solve the problems of less samples for avionics equipment fault detection and lack of online real-time detection theory, the One Class Kernel Extreme Learning Machine (OCKELM) is combined with the data driven method of Incremental Learning (IL) to realize online sequential fault detection. Based on the sample data in the normal state, the kernel form of OCKELM is given, and the expressions for the kernel function and kernel weight vector are deduced. According to the incremental learning method, the kernel weight vector is updated and the sample output value is estimated when new samples are absorbed. Finally, based on two kinds of threshold criteria, the inspection threshold of the model is given, and online fault detection of the test samples is carried out. The proposed method is applied to the test data of UCI data set and an avionics device. The experimental results show that the method can realize online detection. Compared with the existing SVDD, PCA and OC-SVM methods, the proposed method has better performance in terms of F1, AUC, G-mean, fault detection rate, and other performance indicators.

Key words: one class classification, kernel extreme learning, incremental learning, online detection, fault detection

中图分类号: