A hybrid kernel-based nonparametric system identification approach for multiphase batch processes

Shuyu Wang

doi:10.1016/j.compchemeng.2025.109472

A hybrid kernel-based nonparametric system identification approach for multiphase batch processes

2025-10-27

0

PRE

AI

DOI：10.1016/j.compchemeng.2025.109472

Original

Original request for help

Save

Abstract

En 中文

In this study, a hybrid kernel-based nonparametric identification approach for multiphase batch processes is proposed. In accordance with the kernel-based identification framework, the time-varying impulse response is modeled as realizations of a zero-mean Gaussian process, whose characteristics are described by the kernel function. However, existing kernel functions for identification are single kernels that cannot accurately describe the abrupt changes in dynamic behaviors caused by multiphase operations. To overcome this limitation, the proposed approach uses a hybrid kernel defined as a set of piecewise kernel functions over partitioned regions to describe different impulse response characteristics. Then, on the basis of the repetitive nature of batch processes, a transition point-based partitioned region representation is developed; it can automatically form a complete/nonoverlapping partition of 2D time–response plane. Building on the partition, we transform the nonparametric identification task into a joint estimation problem for the transition points and kernel hyperparameters, thus overcoming the suboptimality resulting from sequential estimation and improving identification accuracy. Given the introduction of unobservable transition points, we solve the estimation by maximizing the marginal likelihood with the expectation–maximization algorithm. Two cases are studied to demonstrate the effectiveness of the proposed method.

A hybrid kernel-based nonparametric system identification approach for multiphase batch processes

Abstract

Journal

Organization