Diagnostic applications of artificial neural networks

This paper examines some of the issues which must be resolved in order to develop efficacious and accurate diagnostic neural networks. Two case studies are presented which investigate different aspects of the use of artificial neural networks for diagnosis. Both case studies involve the diagnosis of faults in chemical process systems. In first case study, an industrial furnace is examined concerning the use of binary/trinary input data representations, the effect of primary and secondary symptoms on fault diagnoses, and the relative importance of diagnostic threshold selection on network performance. In second case study, a multi-column distillation plant is examines concerning the use of continuous-valued input data representations, hierarchically-structured neural networks for diagnosis of large systems, the effect of input data noise on diagnostic performance, and the resolution of novel fault situations.