Batch Process Monitoring By Dynamic Time Warping and k-Means Clustering

In the production of membranes for gas separation, the raw material goes through a series of batch processes before it is converted into a finished product. Each of these batch processes is governed by a recipe, which often involves a combination of process control logic and operator interventions. This work presents a method for monitoring the quality of a batch process by detecting deviations from the recipe, through a combination of Dynamic Time Warping (DTW) and k-Means clustering.

Historical data (e.g., flows, temperatures, pressures) are gathered from a membrane manufacturing process, which consists of several steps of variable duration. Through a combination of interactions between researchers and subject matter experts, a set of “ideal” and “bad” recipes are identified which is used as a training set for the model. Historical operation is then compared with the ideal and bad recipes, and a unified distance metric is used to then cluster the operational data using k-Means clustering. Encouraging results are reported from practice.