A modular approach to generating fuzzy rules with reduced attributes for the monitoring of complex systems

The building of intelligent monitoring and diagnostic systems for complex industrial domains tends to be hindered by the knowledge-acquisition bottleneck. Creating good knowledge bases for such tasks is notoriously difficult, especially where human experts are not readily available. High dimensionality of the domain attributes presents a further obstacle for a number of rule-induction algorithms which would, otherwise, have the potential for automating knowledge acquisition: This paper attempts to tackle both problems, by proposing a highly modular framework for data-driven fuzzy ruleset induction incorporating a dimensionality-reduction step based on rough set theory. This removes redundant and information-poor attributes from the data, thereby significantly increasing the speed of the induction algorithm, which is employed to generalise historic data into fuzzy association rules. The aid of dimensionality reduction extends past the training stage of the system into its runtime. By removing information-poor attributes, the implemented system is kept simple by requiring fewer connections to physical instrumentation, while the system's response times are increased. The paper introduces the techniques jointly forming the proposed framework, and demonstrates the applicability of the approach by building a monitoring system for an urban water treatment plant. The results of this application are presented and discussed, and comparisons to alternative approaches are given. (C) 2000 Elsevier Science Ltd. All rights reserved.