计算机工程与应用 ›› 2018, Vol. 54 ›› Issue (9): 5-12.DOI: 10.3778/j.issn.1002-8331.1712-0425

• 热点与综述 • 上一篇    下一篇

集成散度的MKL模型在模拟电路诊断中的应用

张  伟,许爱强   

  1. 海军航空大学 航空作战勤务学院,山东 烟台 264001
  • 出版日期:2018-05-01 发布日期:2018-05-15

Application of MKL model incorporated within-class scatter in analog circuit diagnosis

ZHANG Wei, XU Aiqiang   

  1. School of Aeronautical Operations and Service, Naval Aeronautical University, Yantai, Shandong 264001, China
  • Online:2018-05-01 Published:2018-05-15

摘要: 为提升模拟电路故障诊断精度,结合基于故障特征间一维模糊度的特征选择算法,提出一种新的多核超限学习机诊断模型。该模型通过设置虚拟的基核,将正则化参数融入基核权重求解过程中;同时,通过将特征空间类内散度集成到多核优化目标函数中,在最小化训练误差的同时,使得同一模式的故障样本更加集中,有效提升了故障模式间的辨识力。通过两个模拟电路诊断实例表明:相比于单核学习算法,所提方法可以显著提升诊断精度,并且可以将难以辨识的故障样本更加准确地隔离到相应模糊组中;相比于一般的多核学习算法,所提方法在取得相似诊断精度的同时,时间花费更少。

关键词: 故障诊断, 多核学习, 散度矩阵, 超限学习机, 自适应正则化

Abstract: In order to improve the fault diagnosis accuracy of analog circuit, a novel multi-kernel Extreme Learning Machine (ELM) diagnostic model is presented by combining with feature selection algorithm used one-dimensional ambiguity among fault features. In this model, the optimization of regularization factor is incorporated into the solving process of basis kernel weight coefficients by setting a fictitious kernel function. Moreover, the within-class scatter of training data in feature space is also incorporated into optimized objective function of multi-kernel ELM, which makes the samples from same fault pattern more concentrated when the training error is minimized so that the identifiability is effectively enhanced. Experimental results on two analog circuits show that the diagnostic accuracy is significantly improved compared with single kernel learning algorithms, and those faults which are difficult to be identified can be more accurately isolated into relevant ambiguity groups. In addition, compared with common multi-kernel learning algorithms, the similar diagnostic results can be obtained, but the proposed model costs less time.

Key words: fault diagnosis, multiple kernel learning, scatter matrix, extreme learning machine, adaptive regularization