Data-Driven Modeling for Multiphase Processes: Application to a Rotomolding Process

This paper addresses the problem of capturing the multiphase nature of a rotational molding process using subspace identification (SSID) to enable improved control. Existing SSID techniques are not designed to utilize any known, multiphase nature of a process in the model identification stage. This work adapts existing SSID methods to account for multiple phases by splitting the data into phases during the identification step and building a distinct SSID model for each phase while carefully connecting the individual models through the means of subspace states. This is achieved via a partial least-squares (PLS) model that relates the final states of the preceding phase to the initial states of the proceeding phase. This multiphase subspace identification (MPSSID) approach exploits the ability of SSID techniques for dynamic modeling of batch processes, which allows for model construction using batches of nonuniform length. In this work, the proposed approach is applied to the rotational molding process. For rotational molding, the final product quality is dependent on the temperature trajectory of the polymer inside the mold, and the process goes through visibly distinct phases that can be recognized when a specific temperature (not time) is reached. Data from past experiments are used to build the model and validate it, comparing the predictive ability of multiphase models to conventional one-phase models. Results demonstrate the ability of the multiphase models to better predict both the temperature trajectories and final product quality of validation batches.