Dynamical process monitoring using dynamical hierarchical kernel partial least squares

Abstract

In the paper, dynamical hierarchical kernel partial least squares (DHKPLS) algorithm is proposed and a new monitoring approach is proposed based on DHKPLS. The original multiway partial least squares (MPLS) process monitoring method has the following disadvantages: 1) MPLS is a linear projection method, which can't effectively capture the nonlinear features existing in most batch processes. 2) It is limited that complete batch process data is indispensable. Hierarchical kernel partial least squares (HKPLS) can solve these problems. However, HKPLS is a fixed-model monitoring technique, which gives false alarms when it is used to monitor real processes whose normal operation involves dynamical changes. In this paper, a DHKPLS algorithm is proposed to solve dynamical problems. DHKPLS diagnoses faults without need of estimating the uncompleted portion of the testing batch. It reduces false alarms and provides reliability when applied it to monitor complex processes. Then the online monitoring system is set up, which can track different types of variations closely.