Employing Adaptive Just-In-Time-Learning in a Transfer Learning Frame for Soft-Sensor Design

Abstract

Abstract Online learning methods, such as moving window (MW) and just in time learning (JITL), have been proposed in the literature to remedy concept drift problem deteriorating soft-sensor performance. While these methods are effective against different types of drifts, a single method may not be sufficient in combating against heterogeneous concept drifts. In the current study, we propose combining MW and JITL methods within a transfer learning frame coupled with a relevant instance selection method to improve the prediction accuracy offered by either method. The proposed method involves i) forming a relevant sample of historical observations via backward elimination of the clusters composed of the extended nearest neighbors of the query point, ii) constructing a task transferred JITL model via kernel ridge regression, and iii) using a transductive MW learner. Employing the proposed method on two publicly available real benchmark datasets yields highly accurate predictions, showing convenience for industrial applications.