Chemical process fault diagnosis using pattern recognition and semi-quantitative model based methods

No industrial process operates without any deviation from the scheduled operation mode, during its economic life. Deviations from normal operation, if not detected and prevented, may lead to decreased plant life and even to disastrous events. Hence for process safety and preventive maintainance fault diagnosis is of vital importance. To minimize the risk of human error and as a decision support aid, computerized decision support systems for fault detection and diagnosis are valuable tools. The vast size of the design space for diagnostic systems makes a single general approach practically impossible.; In this dissertation, two pattern recognition based and two semi-quantitative model based approaches to chemical process fault diagnosis are developed. First, a symbolic fuzzy genetic algorithm based inductive learning system FGAL is developed. Equipped with a fuzzy c-means based partitioner, FGAL generates symbolic, general fuzzy rules from numerical plant data. This is not possible from a neural network. Unlike discriminative neural network and decision tree based approaches, FGAL is a hybrid generative-discriminative system avoiding novel class problems associated with discriminative learning systems. FGAL is tested on a hydrocarbon dichlorination fault diagnosis problem, showing 98% and 95.7% diagnostic performance for 1.0% and 1.5% Gaussian noise, respectively.; A second system developed is an extension of FGAL obtained by retrofitting it with kernel densities and using hidden Markov models to stochastically model the time component of the underlying dynamic process. The resulting hybrid generative-discriminative system is capable of extracting symbolic knowledge, uses only relevant dimensions for classification and can learn the HMM parameters via the Baum-Welch method. The system was tested successfully on two case studies: a gravity tank and a cascade controlled continuously stirred tank reactor (CSTR), correctly diagnosing all faults in each test case.; The third approach is a model based diagnostic system (MBDS), based on semi-quantitative fuzzy qualitative simulation (Shen and Leitch, 1993). The system makes use of partial process knowledge to generate a semi-quantitative model of the diagnosed process and qualitatively mimics the behavior of the diagnosed system under abnormal conditions. The introduced MBDS uses model selection rules generated from the off-line, one step ahead fuzzy qualitative simulation of fault models unlike QSIM (Kuipers, 1986) based MIMIC's (Dvorak and Kuipers, 1991) dependency tracking, which is not suitable for systems with recycles and control loops. The system was tested on a gravity tank and a constant holdup CSTR.; The last approach converts semi-quantitative dynamic bounding behavior envelopes that result from numerical interval simulation into compact, episodic fuzzy rule sets which are further used to monitor the process under consideration. Equipped with distance and time based belief scaling and novel class detection mechanisms, the system was successfully tested in a gravity tank and a closed-loop CSTR case study. Despite the conservative and highly overlapping envelopes for CSTR, the system was able to detect the correct fault in each test case. The automatic rule base generation capability facilitates rule base maintainance and fault library extension.