基于多动态核聚类的间歇过程在线监控

Multiple dynamic kernel clustering based online monitoring for batch processes

针对传统的多元统计监测方法不能有效检测工业过程中由于初始条件波动较大所引发的弱故障问题,提出一种基于多动态核聚类的核主元分析(DKCPCA)监控策略,实现多阶段间歇过程的弱故障在线监控。该方法首先针对过程中各阶段每一批次数据结合自回归移动平均时间序列模型(ARMAX)和核主成分分析(KPCA)方法分别建立动态核PCA模型,然后根据各批次模型间载荷的相似性采用分层次聚类方法进行聚类,最后将聚在一起的批次数据进行展开重新再建立动态核PCA模型,随着聚类数目的不同从而建立多个类模型。当在线应用时给出了多模型选择策略,以提高监测精度。将此方法应用于青霉素发酵过程的监控中,监测结果表明此方法取得了比DKPCA和MKPCA更好的监测性能。

Since weak faults induced by large fluctuations under poor initial conditions could not be effectively detected by traditional multivariate statistical monitoring methods, a novel kernel principal component analysis monitoring strategy based on multiple dynamic kernel clustering （DKCPCA） was proposed to improve weak faults detection performance for multi-stage batch processes. The proposed method firstly combined auto-regressive moving average exogenous time series model and kernel principal component analysis （KPCA）. The dynamic kemel PCA model was built for each batch in each stage. Then hierarchical clustering was implemented through load matrix similarity among batch models. Finally, the batch data belonging to the same cluster were unfolded to build dynamic kernel PCA model again. The multiple models were established along with different cluster numbers. When online monitoring, multiple model selection strategy was given to improve monitoring precision. The monitoring method was applied to fault detection for benchmark of fed-batch penicillin production. The monitoring results showed that the proposed method had better performance than DKPCA and MKPCA.