Advanced Monitoring and Soft Sensor Development with Application to Industrial Processes

Abstract

This dissertation presents the research performed to develop innovative multivariate process monitoring and soft sensor approaches for industrial processes. In the operation of modern industrial processes, it is attempted to improve product quality and yield in shorter periods of time while observing tighter environmental regulations and safety guidelines. Furthermore, these goals are tried to be accomplished facing a severe global competition and shorter product life cycles. The most reliable approach to achieve these goals is the development of first principle models (e.g. based on dynamic mass and energy balances) to describe the physics, chemistry, and the different phenomena occurring in the process. However, because industrial processes are generally quite complex to model and are characterized by significant inherent nonlinearities, a rigorous theoretical modeling approach is often impractical, requiring a great amount of effort. An alternative approach to address these aspects is using process data. In most modern industries, a large amount of variables are measured and stored automatically by the governing control system (e.g. DCS). The stored data corresponds to different measurement instruments that are used for closedloop control or merely as indicators of different conditions in process units. The process data can be utilized to address several aspects. These include monitoring the process over time in order to detect special events (e.g. faults) and assign causes for them. Furthermore, the data can be used to predict key process variables that are difficult to measure online (e.g. quality) based on easy-to-measure process variables. In the first part of this work (PART I), a new multivariate method to monitor continuous and batch processes is developed based on the statistics pattern analysis (SPA) framework. The SPA framework is developed to address challenges present in industrial data such as nonlinearities and multi-modal distributions, which are characteristics of process data that ii normally affect traditional multivariate monitoring methods. In addition, different unique characteristics of the SPA framework are discussed in detail. Also presented are comparisons of the monitoring performance based on SPA with the monitoring performance of different linear and nonlinear multivariate monitoring methods. The comparison results clearly demonstrate the superiority of the SPA-based monitoring in key aspects such as detecting and diagnosing different faults. The research findings indicate that the SPA-based monitoring is a promising alternative for monitoring industrial processes, especially where nonlinearities and multi-modal distributions are present. In the second part of this work (PART II), the idea of obtaining quality estimates of key process variables using multivariate projection methods is explored. In many industrial processes, the primary product variable(s) are not measured online or not measured frequently, but are required for feedback control. To address these challenges, there has been increased interest toward developing data-driven software sensors (or soft sensors) using secondary measurements (e.g. variables easy-to-measure) based on multivariate regression techniques. The prediction of these key process variables not only provides a way to have real time measurements, but also aids in improving quality and preventing undesirable operating conditions. In this work, the properties and characteristics of dynamic partial least squares (DPLS) based soft sensors are studied in detail. In addition, a reduced-order DPLS (RO-DPLS) soft sensor approach is proposed to address the limitation of the traditional DPLS soft sensor when applied to model processes with large transport delay. The RO-DPLS soft sensor is also extended to its adaptive version to cope with processes where frequent process changes occur. Also presented are alternatives to address practical issues such as data scaling and the presence of outliers off-line and online. The application of the proposed soft sensor is illustrated with various simulated and industrial case studies of a continuous Kamyr digester. In both simulation and industrial case studies, the proposed RO-DPLS soft sensor shows iii that it can provide quality estimates of key process variables in the presence of different disturbances, data outliers, and time-varying conditions.