Computational intelligence methods for process discovery

The challenge in commissioning and maintaining industrial processes is one of managing complexity. When faced with a highly complex system, an analyst may make simplifying assumptions, ignoring features that are assumed to ha ve minimal impact, assuming that data will take values in specified regions or that system dynamics will have specific forms. When the specification o f the real process is incomplete there may be occasions when the perceived situation does not fit within the constrained frame of reference or the res ponses do not have the expected effect. Computational intelligence methods can enable the analyst or process engineer greater ability to cope with the natural complexity of industrial processes. Unsupervised learning methods can be used to classify modes of operation. Neural network models can be tr ained from process data and used on-line to simulate or replace inefficient tests. Evolutionary algorithms can be used to effect optimal closed-loop s upervisory-level control of processes. Any or all of these technologies can be applied to process monitoring, fault anticipation and aversion, Fault d iagnosis and resolution, or process optimization. An example of the use of these methods is presented in the domain of metalcasting. (C) 1998 Publishe d by Elsevier Science Ltd. All rights reserved.