Data-Driven Soft Sensors in Refining Processes – Pulp Property Estimation Using ARX-Models

Artikel i vetenskaplig tidskrift, 2023

This paper focuses on estimation of shives(wide) and fiber length in RGP82CD-refiners using an AutoRegressive eXogenous (ARX) structure in a data-driven soft sensor concept. Both external and internal variables are considered as model inputs. The pulp properties were sampled every 15 min from an on-line device positioned after the latency chest, whereas other process data were sampled every 6 seconds. Notably, despite the high data sampling rate, the development of robust models necessitated a dataset spanning over two months of process information. The external variables studied in this paper were specific energy, the sawmill chip content, plate gaps, and dilution water feed rates to each refining zone. Additional internal variables, such as the inlet flat zone temperature, the maximum temperature, and the periphery temperature in the conical zone, were also used as model inputs. It was concluded that both shives(wide) and fiber length can be estimated with relatively good accuracy although large uncertainties exist in the measured properties. Finally, it was shown that fast pulp property dynamics in the blow-line can be followed, which outperforms current practices of using pulp measurement devices positioned after the latency chest. This offers implementation of more advanced future pulp property control concepts.