Use of genetic algorithms in training diagnostic rules for process fault diagnosis

The self learning of diagnostic rules can ease knowledge-acquisition effort, and it is more desirable in cases where experience about certain faults is not available. Applications of genetic algorithms to the self learning of diagnostic rules for a pilot-scale mixing process and a continuous stirred-tank reactor system are described in the paper. In this method, a set of training data, which could be obtained from simulations and/or from the recorded data of the previous operations of the real process, is required. The training data is divided into various groups corresponding to various faults and the normal operating condition. Corresponding to each fault, there is a group of initial rules which are coded into binary strings. These rules are evaluated by a genetic algorithm which contains the three basic operators, reproduction, crossover and mutation, and an added operator which preserves the best rule ever discovered. Through this biological-type evaluation, new fitted rules are discovered. The results demonstrate that diagnostic rules fitted with a given set of training data can be efficiently discovered through genetic learning, and, hence, that genetic algorithms provide a means for the automatic creation of rules from a set of training data. It is also demonstrated that bad training data and the inappropriate formulation of rules could degrade the performance of the learning system.