Phase identification for product quality prediction in batch processes: Application to industrial resin production

This paper addresses the phase identification problem in the development of a soft-sensor for quality variables prediction in batch processes. Batch processes are characterized by time-varying dynamic behavior. This means that input-output variable correlation is not constant during the whole process and so using a single statistical/deterministic model is not usually accurate enough to describe the plant behavior. The multi-phase approach overcomes this problem by grouping process data in different clusters according to their statistical structures. Each phase is then modeled independently using, e.g., the Partial Least Square (PLS) method. In this paper we propose a method to identify process phases based on the k-means clustering algorithm. The main advantages are that this method is generic, automatic and requires no a-priori knowledge of the process. The proposed methodology has been validated on data coming from an industrial plant for production of a resin.