一种带有模型更新机制的JITL-MPLS故障监测方法  

A new MPLS fault monitoring method based on just in time learning

作　　者：张亚堃 高学金 曹彩霞[1,2,3,4] 李亚芬 王普 ZHANG Yakun;GAO Xuejin;CAO Caixia;LI Yafen;WANG Pu(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Engineering Research Center of Digital Community,Ministry of Education,Beijing 100124,China;Beijing Laboratory For Urban Mass Transit,Beijing 100124,China;Beijing Key Laboratory of Computational Intelligence and Intelligent System,Beijing,24,China)

机构地区：[1]北京工业大学信息学部,北京100124 [2]数字社区教育部工程研究中心,北京100124 [3]城市轨道交通北京实验室,北京100124 [4]计算智能与智能系统北京市重点实验室,北京100124

出　　处：《计算机与应用化学》2018年第6期457-468,共12页Computers and Applied Chemistry

基　　金：国家自然科学基金(61640312,61763037);北京市自然科学基金(4172007);北京市教育委员会资助

摘　　要：发酵过程具有时变性、动态性和多阶段性的特点,对其进行故障监测主要采用离线建模方式,但这种方法并不能很好地反映当前生产过程的数据特征。近年来有学者使用即时学习（Just in Time Learning,JITL）在线建模策略来建立精确的在线模型并进行故障监测,但是即时学习在线建模策略存在着模型更新频繁、计算量大的问题？本文提出一种带有模型更新机制的即时学习多向偏最小二乘（JITL-MPLS）的故障监测方法：依据马氏距离相似度,选择相似历史样本建立多向偏最小二乘监测模型;而后通过对比上一时刻的质量测量值和当前时刻的质量预测值的差值是否超限来判断模型是否需要更新,当其差值没有超限,即上一时刻监测模型能够表征当前时刻的数据特征,不更新模型,而是继续沿用,否则更新模型。最后将此方法应用于青霉素发酵仿真系统的在线监测,验证了该方法的有效性。The fermentation process has the characteristics of time-varying, dynamic and multi-stage. The fault monitoring mainly adopts the offline modeling method, but this method does not reflect the data characteristics of the current production process well. In recent years, some scholars use the online learning model of Just in Time Learning （JITL） to establish an accurate online model and conduct fault monitoring, but there is a problem of frequent model updating and large amount of computation in online learning modeling. A fault monitoring method based on JITL-MPLS is proposed, which is based on the model updating decision mechanism. Based on the similarity of Mahalanobis distance, similar historical samples are selected to establish the multi- directional partial least squares （PLS） monitoring model. By comparing whether the residuals of the MPLS model exceed the limits, it is judged whether the model needs to be updated. When the monitoring model can characterize the data features of the current moment, the model is not updated but continues to be used; otherwise, the model is updated. Finally, this method is applied to online monitoring of penicillin fermentation simulation system, which verifies the effectiveness of the method.

关 键 词：即时学习(JITL) 多向偏最小二乘(MPLS) 模型更新 信息熵 马氏距离 故障监测 

分 类 号：TP206.3[自动化与计算机技术—检测技术与自动化装置]