Improved yellowness index (YI) control in ABS compounding process through virtual control using an RNN-based neural network soft-sensor model

Soft sensors have gained widespread acceptance due to their ability to handle highly complex systems. However, their application in process control remains limited, primarily due to the lack of interpretability associated with most soft sensors, which are often black-box models without a clear physical basis. In an effort to address this issue, this study proposes a novel approach to soft-sensor construction by integrating the eXtreme Gradient Boosting (XGBoost) technique with the Gated Recurrent Unit (GRU) in chemical engineering. The approach is applied to a real-world industrial ABS (Acrylonitrile-Butadiene-Styrene) resin compounding process. After the predictive performance and physical correctness of the soft sensor are validated, a virtual fuzzy control system is constructed using data from the soft-sensor prediction. The results show that the proposed method with superior predictive performance and physical correctness, and its use in the control system improves the stability of the product quality and decreases grade transition times.