Hybrid factors latent Gaussian process modeling with wasserstein distance for soft sensing of extruder processes

Abstract

In the domain of polymer production, twin-screw extruders are crucial, necessitating precise monitoring of key quality variables. However, challenges arise due to time-variant and intricate production processes, as well as high-dimensional and hybrid screw configuration data, hindering efficient quality inference. To overcome these hurdles, a latent variable-based online soft sensing approach, termed just-in-time hybrid factors Gaussian process latent variable regression (GPLVR) via the Wasserstein metric, is proposed for twin-screw extruders. Specifically, composite variables are firstly constructed that combine both quantitative and qualitative factors, ensuring a comprehensive representation of the process. Subsequently, the just-in-time method is adopted to ensure continuous updating of the model. To describe the mixed-type dataset of extrusion process data including both continuous and discrete elements. The Wasserstein distance is utilized to select historical samples that closely match the distribution of online query samples. Furthermore, the GPLVR is used to address the curse of dimensionality associated with high-dimensional screw configuration elements. The effectiveness of our proposed method has been verified through its application in the production process of polypropylene, demonstrating its potential to improve the accuracy and reliability of quality inference in twin-screw extruder operations.