基于MICA-OCSVM的间歇过程故障监测

Fault Detection of Batch Process Based on MICA-OCSVM

针对多向独立成分分析(multi-way independent component analysis,MICA)需要假设过程变量服从非高斯分布的要求,以及MICA基于马氏距离构造的监控统计量会导致故障检测率降低的问题,研究了一种将多向独立成分分析与单类支持向量机(one-class support vector machines,OCSVM)相结合的MICA-OCSVM监测方法.首先采用MICA提取间歇过程所有批次的独立成分;然后分别对每个时刻的所有批次的独立成分进行OCSVM建模,利用确定的决策超平面构造非线性的监控统计量;最后计算所有建模数据的监控统计量,并利用核密度估计确定相应的控制限.将该方法应用到青霉素发酵过程仿真平台,实验结果表明:该方法相比于传统的MICA故障监测方法,无需考虑过程变量服从何种分布,能够有效利用独立成分的结构信息,故障的误报率、漏报率明显降低.

To solve the problems that multi-way independent component analysis （MICA） need to assume process variables conforming non-Gaussian distribution and the monitoring statistics based on the Mahalanobis distance of MICA will cause reduction in fault detection rate, a new monitoring method based on Multi-way Independent Component Analysis and One-Class Support Vector Machines （MICA- 0CSVM） was researched. Firstly, the independent components （ICs） from all batches of batch process were extracted by MICA. Secondly, OCSVM was used to model for all batches＇ ICs at each time, separately. Meanwhile, decision hyper-plane of the OCSVM model was chosen to construct monitoring statistics. Finally, the confidence limits were determined using kernel density estimation by the monitoring statistics calculated from all modeling data. The method was applied to fed-batch penicillin fermentation process. The experiment results show that in contrast to the fault detection methods based on traditional MICA, the proposed method can make full use of ICs＇ structure information regardless of the distribution of process variables and can reduce the rate of misinformation and omission effectively.