(224c) Data-Driven and Hybrid Modeling of Integrated Paper Production Systems

The paper industry is instrumented with sensing systems that generate large streams of data. These rich data streams provide distinctive opportunities for improved optimization and control of these energy and water intensive processes. Recently, researchers have been trying to utilize these data streams to achieve better system performance [1], predict system variables [2–4], or to find the root cause of the problem [5]. At the same time, paper production is comprised of a set of process steps for mixing and drying streams containing fibers, for which partial first-principle knowledge is available. However, due to the complex nature of data and models, previous approaches are limited to single process-level analysis, which leads to suboptimal operation.

To overcome these challenges, we propose a system-level analytical framework which combines first-principle knowledge and low-dimensional learning from high-dimensional streaming data. We take a hybrid approach that combines the domain knowledge with data-driven modeling, through model calibration techniques utilizing Bayesian analysis [6]. To achieve more accurate predictions and improve interpretability, we present a scheme for the integration of physics-based equations for the prediction of pulp fiber orientation and final paper properties with Gaussian process model training. Due to the high-dimensionality of the data, dimensionality reduction techniques are required to compress the input space into the important reduced-space features. The system studied includes main downstream process steps for paper manufacturing, including the headbox, wire forming, wet pressing, and drying sections. Through this approach, we are able to develop a tractable integrated process model and quantitatively link process-design-material variables to final paper quality attributes for the first time. We will also discuss how the techniques presented for the paper production system can be generally applied to a variety of different manufacturing systems.

[1] P. Zhou, H. Wang, M. Li, Z. Zhao, and T. Chai, “Data-driven ALS-SVR-ARMA2K modelling with AMPSO parameter optimisation for a high consistency refining system in papermaking,” IET Control Theory Appl., vol. 10, no. 14, pp. 1620–1629, 2016, doi: 10.1049/iet-cta.2015.0850.

[2] K. Ciesielski and K. Olejnik, “Application of neural networks for estimation of paper properties based on refined pulp properties,” Fibres Text. East. Eur., vol. 22, no. 5, pp. 126–132, 2014.

[3] A. Alonso, C. Negro, A. Blanco, and I. San Pío, “Application of advanced data treatment to predict paper properties,” Math. Comput. Model. Dyn. Syst., vol. 15, no. 5, pp. 453–462, 2009, doi: 10.1080/13873950903375445.

[4] O. T. Okan, I. Deniz, and S. Tiryaki, “Application of artificial neural networks for predicting tensile index and brightness in bleaching pulp,” Maderas Cienc. y Tecnol., vol. 17, no. 3, pp. 571–584, 2015, doi: 10.4067/S0718-221X2015005000051.

[5] J. Fu and P. W. Hart, “Leveraging mill-wide big data sets for process and quality improvement in paperboard production,” Tappi J., vol. 15, no. 5, pp. 309–319, 2016, doi: 10.32964/tj15.5.309.

[6] M. C. Kennedy and A. O’Hagan, “Bayesian calibration of computer models,” J. R. Stat. Soc. Ser. B (Statistical Methodol., vol. 63, no. 3, pp. 425–464, 2001, doi: 10.1111/1467-9868.00294.