Adaptive soft-sensor update by Latest Sample Targeting Frustratingly Easy Domain Adaptation

Soft-sensors are widely used in manufacturing processes to estimate key process variables; however, their performance may deteriorate when process characteristics change. Although Just-In-Time (JIT) modeling techniques have been proposed for adaptive soft-sensor design, they do not always adapt to abrupt changes. Transfer learning (TL) has been suggested as a means to address this issue, with Frustratingly Easy Domain Adaptation (FEDA) being used for soft-sensor design. This study proposes a new TL method called Latest Sample Targeting-FEDA (LST-FEDA) for JIT-based soft-sensor, which can handle both sudden and gradual changes in process characteristics. LST-FEDA updates soft-sensors using a fixed number of latest samples whenever a new sample is obtained. The effectiveness of the proposed method was demonstrated using simulation data from a vinyl acetate monomer (VAM) process and actual operation data from a fluorine-based monomer (FM) process. LST-FEDA accurately estimated objective variables during sudden malfunctions and scheduled maintenance, contributing to efficient and safe process operation. • A novel soft-sensor update algorithm utilizing LST-FEDA is proposed. • LST-FEDA is a new transfer learning method for Just-In-Time modeling. • The usefulness of LST-FEDA is validated through simulation data of a VAM process. • The application results of LST-FEDA to real data of a FM process are reported.