Pergamon Expert Systems WithApplications, Vol. 12, No. 4, pp. 455-464, 1997 © 1997 Elsevier Science Ltd Printed in Great Britain. All rights reserved 0957-4174/97 $17.00+0.00 PII: S0957-4174(97)00005-5 Development of Operation-Aided System for Chemical Processes KVUNG J o t Mo, YOUNG SEOK OH AND EN SuP YOON Department of Chemical Engineering, Seoul National University, Shillim-dong, Kwanak-ku, Seoul 151-742, Korea CHANG WOOK JEONG Process EngineeringDepartment,SamsungEngineeringCo, Ltd, GlassTower,946-1, Daechi-Dong, Kangnam-ku,Seoul 135-280,Korea Abstract--This paper presents the development of a knowledge-based operation-aided system for polypropylene process. The important part of this system is the search for the root cause of faults by detecting and analyzing the symptoms which occurred in the process in the case of abnormal situations. In this system, an artificial neural network which is able to handle pattern recognition is used for qualitative interpretation of sensor data and generating symptoms. For effective fault diagnosis, two causal effect models which are based on SDG (Signed Directed Graph) are developed. One model, RCED (Reduced Cause Effect Digraph) uses only the measurable sensor data of the process and is constructed off-line and stored in the knowledge base of the system. The other model, PGTT (Pattern Graph Through Time) is generated in the real-time mode during the diagnosis period. It is generated from symptoms-- status and~or tendency change--and can handle dynamic state effectively. By implementing the developed qualitative interpretation method and two causal effect graph models, the operation-aided system for the polypropylene process, FINDS/PP (Fault Isolation and Detection System/FolyPropylene) was devel- oped. This system was developed with the expert system tool G2 and showed good results. © 1997 Elsevier Science Ltd 1. INTRODUCTION As CONCERNABOUTsaving materials and energy, safety and environment has increased, chemical processes have been integrated into complex networks. As a result, when any equipment failure or human error occurs in these complex processes, it is very difficult to find the causes of the failures, therefore economical loss may be very large. So, concern regarding an automatic fault diagnosis system that can help to maintain product quality and decrease the operator's load is increasing (Dvorak & Kuipers, 1991). A variety of approaches to automated fault diagnosis have been explored by various researchers, and no single method has proven to be universally superior. Method- ologies for fault diagnosis can be broadly classified as model-based or experience-based. A model-based approach will incorporate a structured model of plant behavior based on fundamental engineering principles. Models can be qualitative, expressed in the form of numerical equations, or qualitative, expressed in the form of logical relationships. A number of competing technologies for qualitative methods, such as the experi- ence-based expert system, digraph-based system, qualitative simulation and neural network, were reviewed by Becraft et al. (1991). In this paper, the development of a fault diagnosis system for the polypropylene process is presented. An artificial neural network is used for qualitative inter- pretation of process data and two causal effect models are developed for fault diagnosis. 2. QUALITATIVE INTERPRETATION 2.1. Conventional Methods Real-time application of a knowledge-based system as a fault diagnosis system requires conversion of inherently numeric sensor data into a useful symbolic inter- pretation. Whiteley and Davis (1992) defined this numeric sensor data converting into symbolic informa- tion as qualitative interpretation (QI). The simplest and most commonly used method is limit checking. The basic form of this is: rm~n<r(t)<rm~x 455