(383d) Feed Characteristics Dependent Control Parameter Estimation By Multivariate Analysis in Batch Processes

In this study, we propose a method to estimate the parameters of a control model for a batch process by using previous batch data. We focus on the case of multistep batch processing, where appropriate control input values often strongly depend on the prior processing history of the feed (called “feed characteristics” hereafter), e.g., the equipment or operating conditions used in the previous processing steps. In such cases, it is a common practice to predict control inputs of new batch by using the data from those previous batches with identical feed characteristics. As batch operations become more complicated, the variety of feed characteristics is increased and consequently the chance of finding recent batch data with identical feed characteristics is reduced. To combat the shortage of usable data in same feed characteristics, it is important to enable the utilization of not only data from batches of identical feed characteristics but also those from batches of “similar” feed characteristics. This paper attempts to address this need in a practical manner. By using multivariate clustering analysis, statistical similarities among the estimated parameter values for different feed characteristics can be evaluated and substitutable sets of the feed characteristics can be identified. MANOVA (multivariate analysis of variance), K-means clustering and hierarchical clustering are representative statistical clustering analysis. Based on the results from the different clustering algorithms, appropriate clustering method is selected for identifying similar feed characteristics. This approach can increase the amount and/or recency of the data used in the batch control input calculation. We suggest some specific rules for selecting among available previous batch data by considering both the feed characteristic similarity and time-immediacy. The proposed method firstly tested on artificially generated data for pre-analysis. Then, the method was applied to real manufacturing industrial data and the results showed practical viability and significant potentials of the method.