基于AP-MDFE-MDPCA的间歇过程故障监测  

AP-MDFE-MDPCA-based fault monitoring for batch processes

作　　者：沈亚慧 马红玉 王亚君[1] SHEN Yahui;MA Hongyu;WANG Yajun(School of Electronic and Information Engineering,Liaoning University of Technology,Jinzhou 121001,Liaoning Province,China)

机构地区：[1]辽宁工业大学电子与信息工程学院,辽宁锦州121001

出　　处：《化学工程》2024年第11期83-88,共6页Chemical Engineering(China)

基　　金：辽宁省教育厅面上项目(LJKZ0624);辽宁省自然科学基金资助项目(2020-MS-291)。

摘　　要：为解决多阶段间歇生产过程操作随机性、数据不等长、呈动态分布等问题以及提高故障监测准确性,提出基于AP(近邻传播聚类)和MDFE(混合数据特征提取)的MDPCA(多动态主成分分析)故障监测方法AP-MDFE-MDPCA。首先,针对多阶段间歇过程中操作随机性使得数据相关性发生改变,不能用整个批次数据统一建模问题,先对阶段进行精确划分,在同一阶段采用AP算法,将相似特征数据进行聚类,并采用BFO(菌群优化)算法对其偏向参数p优化,获得最优权重和最佳初始聚类中心,构建批数据的分类模型。其次,为解决聚类后相同类中的批次数据不等长问题,利用MDFE算法进行批次数据等长化处理。同时考虑到过程变量内的动态特性,按照不同阶段聚类数目的不同,构建MDPCA模型。最后,将所提方法应用于精炼炉炼钢过程,并与基于LNS-MDPCA(局部邻域标准化的多动态主成分分析)方法进行比较。实验结果表明提出的方法监测准确性更高,监测速度更快。In order to solve the problems of operation randomness,unequal data length and dynamic distribution of multi-stage batch production process and to improve the accuracy of fault monitoring,a MDPCA(multi-dynamic principal component analysis)fault monitoring method AP-MDFE-MDPCA based on AP(affinity propagation clustering)and MDFE(mixed-data-feature extraction)was proposed.Firstly,to address the problem that the randomness of operation in the multi-stage intermittent process makes the data correlation change and the whole batch data cannot be used to model the problem uniformly,the stages were firstly divided accurately,and the AP algorithm was used in the same stage to cluster the data of similar features and the bias parameter p was optimized by using BFO(bacterial colony)algorithm to obtain the optimal weights and the best initial clustering centre to construct a classification model for the batch data.Secondly,in order to solve the problem of unequal length of batch data in the same class after clustering,the MDFE algorithm was used to equalise the batch data.Meanwhile,considering the dynamic characteristics within the process variables,the MDPCA model was constructed according to the difference in the number of clusters at different stages.Finally,the proposed method was applied to the steelmaking process in the refining furnace and compared with the LNS-MDPCA(local neighbourhood standardized-multi-dynamic principal component analysis)method.The experimental results show that the proposed method has higher monitoring accuracy and faster monitoring speed.

关 键 词：近邻传播聚类 混合数据特征提取 多动态主成分分析 过程监测 精炼炉炼钢过程 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置]