A reduced order soft sensor approach and its application to a continuous digester

In many industrial processes, the primary product variable(s) are not measured online but are required for feedback control. To address this challenge, there has been increased interest toward developing data-driven soft sensors using secondary measurements based on multivariate regression techniques. Among different data-driven approaches, the dynamic partial least squares (DPLS) soft sensor approach has been applied to several industrial processes. However, despite its successful applications, there is a lack of theoretical understanding on the properties of the DPLS soft sensor. Specifically, whether it can adequately capture process dynamics and whether it can provide unbiased estimate under closed-loop operation have not been examined rigorously. In this work, we provide a theoretical analysis to answer these questions. In addition, we propose a reduced-order DPLS (RO-DPLS) soft sensor approach to address the limitation of the traditional DPLS soft sensor when applied to model processes with large transport delay, i.e., large number of lagged variables are required to be include in the regressor matrix in order to capture process dynamics adequately. Compared to the traditional DPLS soft sensor, the proposed RO-DPLS approach not only reduces model size and improves prediction but also provides multiple-step-ahead prediction. The performance of the proposed RO-DPLS is demonstrated using both a simulated single-vessel digester and an industrial Kamyr digester. (C) 2011 Elsevier Ltd. All rights reserved.