A novel reinforced incomplete cyber-physics ensemble with error compensation learning for within-batch quality prediction

Abstract

This study addresses the challenge of real-time quality monitoring in batch operation by emphasizing the significance of within-batch quality estimation. While data-driven machine learning models are easy to construct, they often lack reliability and interpretability when dealing with sparse quality data. Conversely, first-principles models (FPMs) are interpretable but struggle with accuracy and adaptability to changing conditions. To overcome these issues, a three-phase reinforced incomplete cyber-physical ensemble plus error compensation learning (RICPE-P-ECL) method is proposed. This method enhances the adaptability of the incomplete cyber-physical model (IncompCPM), which relies on partially-available FPMs, for online quality prediction under varying conditions. The innovation in RICPE-P-ECL lies in its ensemble design and error compensation strategy. Phase 1 constructs IncompCPMs to predict quality for each operating condition, creating base models for ensemble learning. Phase 2 combines these IncompCPMs, with real-time information assigning weights to each model. Phase 3 involves an error compensation agent that adjusts the real-time ensemble prediction, addressing the limitations of FPMs and sparse data. The method is evaluated using a fed-batch bioreactor as the process model, and the results demonstrate that RICPE-P-ECL outperforms traditional data-driven models such as semi-supervised latent dynamic variational autoencoder and semi supervised dual attentioned latent dynamic complementary state space model, achieving R² values close to 1 for real-time within-batch quality prediction across five new testing conditions.