Design of an event-based early warning system for process operations

Abstract

This thesis proposes a new methodology to design an event-based warning system as an alternative to the conventional variable-based alarm system. This study initially explores the options for grouping process variables for alarm allocation. Several grouping methods are discussed and an event-based grouping procedure is detailed. Selection of the key variables for a group is performed considering the information that the variables contain to distinguish between an abnormal and a normal condition. The information theory is used to quantify the information content of a variable about an event to select the key variables. The cross-correlation analysis between pairs of key variables is used to identify the redundant variables. Simulation study using the model of a continuous stirred tank reactor (CSTR) is used to demonstrate the methodology. The proposed event-based early warning system utilizing online measurements is detailed in the thesis. In this approach, warnings are assigned to plant abnormal events instead of individual variables. To assess the likelihoods of undesirable events, the Bayesian Network is used; the event likelihoods are estimated in real time utilizing online measurements. Diagnostic analysis is conducted to identify root-causes of events. By assigning warning to events, the methodology results in significantly lower number of warnings compared to traditional variable-based warning (alarms) system. It also enables early warning of a possible event along with an efficient diagnosis of the root-causes of the event. Experimental testing using a level control system is presented to demonstrate the efficacy of the proposed method. Simulation study using the model of a CSTR is also presented to demonstrate the performance of the algorithm. Both, experimental and simulation studies, have shown promising results.