Qualitative modeling for fault diagnosis based on physical knowledge and historical operation data under normal operating conditions

Abstract

Fault diagnosis is a critical task in the daily operation of chemical processes. In this paper, a hybrid fault diagnosis method is proposed that combines a process-knowledge-based qualitative reasoning technique with fault detection based on a data-driven process-monitoring technique, without using any faulty datasets. Extended attributes, which are additional process feature variables generated from normal-operating-condition knowledge, are utilized to integrate the two techniques. The process qualitative reasoning model is simplified for combining these techniques and easing the modeling. This fault diagnosis method provides multiple reasoning routes for several potential fault root candidates. Each candidate and variable in its reasoning routes are weighted according to the results of the data-driven fault-detection method. Therefore, a priority list is presented to chemical engineers for further field examinations. The effectiveness of this method is validated using the Tennessee Eastman process, and novel diagnosis results are subsequently achieved.