A neural network methodology for process fault diagnosis

TLDR

Knowledge‑based expert systems promise to automate complex process engineering tasks that require symbolic reasoning, but their deployment is hindered by tedious knowledge acquisition, limited learning ability, and unpredictable performance outside their domain. This study proposes a neural‑network methodology to address these limitations in process fault diagnosis. The approach is illustrated on an oil refinery fluidized catalytic cracking case study and benchmarked against a knowledge‑based system. The neural network accurately diagnoses trained faults, generalizes to unseen fault combinations, and tolerates incomplete or uncertain data.

Abstract

Abstract The ability of knowledge‐based expert systems to facilitate the automation of difficult problems in process engineering that require symbolic reasoning and an efficient manipulation of diverse knowledge has generated considerable interest recently. Rapid deployment of these systems, however, has been difficult because of the tedious nature of knowledge acquisition, the inability of the system to learn or dynamically improve its performance, and the unpredictability of the system outside its domain of expertise. This paper proposes a neural‐network‐based methodology for providing a potential solution to the preceding problems in the area of process fault diagnosis. The potential of this approach is demonstrated with the aid of an oil refinery case study of the fluidized catalytic cracking process. The neural‐network‐based system successfully diagnoses the faults it is trained upon. It is able to generalize its knowledge to successfully diagnose novel fault combinations it is not explicitly trained upon. Furthermore, the network can also handle incomplete and uncertain data. In addition, this approach is compared with the knowledge‐based approach.