Copyright 0 IFAC On-Line Fault Detection and Supervision in the Chemical Process Industries, Lyon, France, 1998 Semi-quantitative Model Based Chemical Process Fault Diagnosis via Episodic FUzzy Rules ibrahim Burak Ozyurt O , Aydin K. Sunol·· and Lawrence O. Hall"· ·University of South Florida, Department of Chemical Engineering and Department of Computer Science f1 Engineering, Tampa, FL, USA. Email:ozyurtOkosh.csee.usf.edu ··University of South Florida, Chemical Engineering Department, Tampa, FL, USA. Email:sunoIOeng.usf·edu ···University of South Florida, Computer Science f1 Engineering Department, Tampa, FL, USA. Email:hallCcsee. usf . edu Abstract. A method for chemical process fault diagnosis using numerical behavior envelopes is described in this paper . A sequence of fuzzy rules is generated from numerical behavior envelopes for each fault class where any rule in a sequence is only valid within the bounds of its time interval . each observed process variable in each fault class two sequences of episodic fuzzy rules , which prOVide a parsimonious, qualitative description of the trend of the variable, are automatically generated one for the lower and one for the upper numerical behavior envelope. The observed process are matched against the fuzzy rules for the current time in the rule base. In case of an region defined by behavior envelopes, the introduced distance and time based behef scalmg allows ranking of fault candidates. A novel abnormal situation will not pass the mtroduced systeI? undetected due to a novel class detection mechanism. The diagnostic performance of the system IS shown in two case studies. Copyright <0 1998IFAC Key Words. Dynamic behavior envelopes; fault diagnosis; qualitative trend analysis; fuzzy clus- tering; semi-quantitative model based fault diagnosis 1. INTRODUCTION Knowledge about how a system works facilitates the determination of why it is not working cor- rectly, which is the idea behind model based di- agnosis. In model based diagnosis, a model of the physical process is operated in parallel with the process. The outputs of the model indicate how the system should operate, while the outputs of the system indicate how it is actually operating. The discrepancy between actual output and model predictions are used as an indicator of abnormal behavior. The main objective of model based di- agnostic systems (MBDS) is then determination of the fault(s) causing the discrepancy. Semi-quantitative methods, such as fuzzy quali- tative simulation (Shen and Leitch, 1993) and Q2 (Kuipers and Berleant, 1988) produce overly con- servative bounds since they use a simulation time step determined by qualitative distinctions. How- ever, the numerical behavior envelopes method (Kay and K ui pers, 1993) makes better use of avail- able quantitative information by generating nu- merical envelopes on all model variables which bound them from above and below. For most of the chemical processes at least some understand- ing of the underlying chemical and/or physical phenomena is available, which will enable con- 41 struction of a process model. However, the deter- mination of the precise parameters for the model is usually the limiting step in the model construc- tion process. The numerical behavior envelopes method allows use of an interval instead of a pre- cise value for model parameters. This method, however, requires the numerical integration of two systems, to generate the upper and lower behav- ior envelopes for each process variable. In real time model based fault diagnosis, for a moderate to large scale process, evaluation of the numeri- cal envelopes may be computationally intractable. The problem becomes worse, in the case that more than one fault model candidate has to be simul- taneously computed in order to monitor an ab- normal process behavior. In this paper, a method to convert off-line generated numerical behavior envelopes into a parsimonious sequence of fuzzy rules is described for automatic rule base genera- tion. The generated rule base is used for real time process monitoring and fault diagnosis. Equipped with distance and time based fault belief scal- ing mechanisms and novel class detection capa- bilities, the developed MBDS minimizes the com- putationalload problem by eliminating the evalu- ation of numerical behavior envelopes in real time, it allows more refined diagnosis than numerical be- havior envelopes alone and provides a novel class