Semi-supervised learning for data-driven soft-sensing of biological and chemical processes

Abstract

Continuously operated (bio-) chemical processes increasingly suffer from external disturbances, such as feed fluctuations or changes in market conditions. Product quality often hinges on control of rarely measured concentrations, which are expensive to measure. Semi-supervised regression is a possible building block and method from machine learning to construct soft-sensors for such infrequently measured states. Using two case studies, i.e., the Williams-Otto process and a bioethanol production process, semi-supervised regression is compared against standard regression to evaluate its merits and its possible scope of application for process control in the (bio-) chemical industry. The case studies show that semi-supervised regression can serve as a valuable building block in construction of a soft sensor to reliably predict steady state data even in case of very few quality measurements (hourly or daily measurements). However, no reliable prediction of process dynamics can be ensured in case of common measurement frequencies for offline quality measurements.