Locally weighted slow feature regression for nonlinear dynamic soft sensor modeling and its application to an industrial hydrocracking process

Abstract

Latent variable (LV) models have been extensively constructed to obtain informative low-dimensional features for process soft sensors. However, static LV models are more often adopted, which cannot describe the process dynamics. Recently, slow feature analysis and its regression model (SFR) have been introduced for dynamic LV modeling in industrial processes. However, linear SFR is limited in its modeling capacity because most industrial processes have time-varying and nonlinear characteristics. To alleviate this problem, a novel locally weighted slow feature regression (LWSFR) is proposed in this paper for nonlinear dynamic modeling. Unlike other static locally weighted learning, two weighting techniques are designed in LWSFR. First, sample weighting is used to deal with static nonlinear relationships based on Euclidean distance. Temporal weighting for the first-order sample time difference is then designed to locally linearize the nonlinear dynamics. The effectiveness of the proposed method is validated on an industrial hydrocracking process.