Regression modeling and quality prediction for multiphase batch processes with inner-phase analysis

• The inner-phase evolution is tracked and modeled for quality analysis.
• After dividing a batch into phases, a phase is further divided into parts.
• Transition-steady part-transition is identified as typical inner-phase structure.
• Different modeling strategies are proposed for different inner-phase parts.
• The application to injection molding process illustrates the feasibility.

Many batch processes have multiple phases which may exhibit significantly different underlying behaviors. Besides, within each phase, processes in general evolve following certain underlying rules, called inner-phase evolution here. In this paper, a new statistical process analysis and quality prediction method is proposed for multiphase batch processes. A two-level phase division algorithm is proposed to capture the changes of relationship between process variables and quality variables within each phase. It reveals that the quality-related inner-phase evolutions in general goes through three statuses sequentially, i.e., transition, steady-phase and transition. Partial least squares (PLS), canonical correlation analysis (CCA) and qualitative trend analysis (QTA) are effectively combined to distinguish different inner-phase process statuses. Their different characteristics are then analyzed respectively for regression modeling and quality analysis. Meanwhile, the uneven-length problem of batch processes caused by operation conditions is handled properly according to their different characteristics in each inner-phase parts. Cumulative effect is considered and modeled both within inner-phase parts and between inner-phase parts for quality perdition. During online application, different quality-related process behaviors within each phase are tracked, revealing the inner-phase evolution. Online quality prediction is performed at each time by adopting different regression models. The application to a typical multiphase batch process, injection molding, illustrates the feasibility and performance of the proposed algorithm for uneven-length batch group quality prediction.