Soft-Sensor Construction Method Based on Adaptive Modeling and Transfer Learning for Manufacturing Process Including Maintenance Periods

Various soft-sensor construction methods based on transfer learning have been proposed to quickly adapt soft-sensors to changes in process characteristics caused by process maintenance. However, most of the conventional methods adopt transfer learning immediately after the maintenance. Since process characteristics may also change due to other than process maintenance, it is necessary for the conventional methods to switch to a soft-sensor that is applicable to the current process characteristics at the appropriate time; however, it is difficult to properly determine the switching timings of soft-sensors. This study proposes a new adaptive soft-sensor construction method called Latest Sample Targeting Frustratingly Easy Domain Adaptation (LST-FEDA) based on adaptive modeling and FEDA which is a kind of transfer learning method. In the proposed LST-FEDA, the application range of FEDA is updated and rebuilds a soft-sensor when a new query sample is measured. This approach enables soft-sensors to automatically adapt to changes in process characteristics without intentional soft-sensor switching. In order to demonstrate the effectiveness of the proposed LST-FEDA, it was applied to actual operation data of a distillation process including maintenance periods. The results demonstrated that LST-FEDA improved the RMSE by 19% on average in comparison with the conventional methods.