摘要
字典学习是一种有效的统计机器学习方法,也被应用于工业过程的故障诊断。初始字典直接影响对数据特征的表达,从而影响故障诊断效果。本文采用PCA/KPCA来对初始字典进行优化,以获得原子间冗余度小、特征表达准确的初始字典。提出了用PCA/KPCA优化初始字典的故障诊断算法框架,讨论了离线学习和在线诊断的核心问题。以机车变压器冷却油泵为对象进行了试验,取6对差动感测线圈输出为样本,对比了5种算法的故障诊断、故障定位及故障类型判别性能。结果表明:基于字典学习的故障诊断优于传统PCA、KPCA算法;用PCA/KPCA优化初始字典的确能提升故障诊断性能;在初始字典优化试验对比中KPCA优于PCA,表明KPCA更加适合油泵的非线性特征。
Dictionary learning,as an effective statistical machine-learning algorithm,is applied to fault diagnosis of industrial processes as well.The initial dictionary has an effect on the expression of structural features hidden behind the original data,thereby affects the performance of fault diagnosis.In order to obtain an initial dictionary with accurate feature expression and low redundancies among atoms,principal components analysis(PCA)and kernel principal components analysis(KPCA)are employed for dictionary optimization.The framework of the fault diagnosis algorithms with PCA/KPCA optimized initial dictionaries is proposed.Then,the key issues of offline dictionary learning and online fault diagnosis are discussed in details.Finally,experimental studies are carried out with a cooling oil pump of locomotive transformers.The outputs of six differential coils are formed as samples in the following fault diagnosis.Five fault diagnosis algorithms are compared in terms of fault diagnosis,fault location and fault type identification.The results indicate that the dictionary learning based fault diagnosis algorithms behave much better than the traditional algorithms as PCA and KPCA.In addition,the PCA/KPCA optimized initial dictionaries can improve the performance of fault diagnosis indeed.Furthermore,initial dictionaries optimized by KPCA show better performance than PCA optimized dictionaries because KPCA suits the nonlinearity of the oil pump more than PCA.
作者
周清雅
张世荣
黎贤钛
汪佳宝
Zhou Qingya;Zhang Shirong;Li Xiantai;Wang Jiabao(School of Electrical Engineering and Automation,Wuhan University,Wuhan Hubei 430072,China;Zhejiang Erg Science Technology Co.,Ltd.,Taizhou Zhejiang 317113,China)
出处
《传感技术学报》
CAS
CSCD
北大核心
2020年第11期1599-1607,共9页
Chinese Journal of Sensors and Actuators
基金
国家自然科学基金项目(51475337)。
关键词
故障诊断
字典学习
特征表达
初始字典
主成分分析
变压器油泵
fault diagnosis
dictionary learning
feature expression
initial dictionary
principal components analysis
oil pump
