Online tuning of predictor weights for relevant data selection in just-in-time-learning

Abstract

Just-in-Time-Learning (JITL) is one of the most frequently used adaptive methods in data-based soft sensor design for chemical processes. While JITL is an effective method to combat concept drifts, samples selected via similarity metric in a Euclidean space consisting of a large number of equally-weighted predictors may diminish the performance of the JITL, due to curse of dimensionality and the varying degrees of nonlinearity between the predictor and response variables. Algorithms involving offline tuning of predictor weights were developed to tackle this issue, but changes in process conditions may depreciate a currently used set of weights for measuring similarity. In the current study, an adaptive method is developed for adjusting predictor weights in the Euclidean space used in measuring similarity between samples, named JITL via Online Weighted Euclidean Distance (JITL-OWED). JITL-OWED mainly consists of four steps: i) Relevant data is selected from an online weighted Euclidean space, and multiple models using different weights in their similarity measures are simultaneously constructed. ii) Control charts are used to detect changes in the prediction accuracy of the multiple models constructed online, hence triggering new subset searches. iii) A small number of search steps is used in subset selection, akin to early stopping in gradient search, with the “modest” aim of moving towards a local minimum for feasible online implementation. Exponentially Weighted Moving Average (EWMA) filtering is employed to stabilize the results from feature searches. iv) Final predictions are obtained via stacking multiple models. Employing JITL-OWED on one synthetic and four publicly available real datasets shows that online predictor weighting may indeed improve the prediction accuracy of the traditional JITL up to 80%, and JITL-OWED is both easy to implement and tune.