Fault Detection and Isolation for Multimode Processes with Recursive Principal Component Analysis

Abstract

Contribution plots of the monitored statistics, Q and T2, are investigated to locate faulty variables when the statistics are out of their control limits. It is a popular method for fault isolation; however, it is well known that the smearing out of contributions leads to misdiagnose the faulty variables. Alternatively, the reconstruction-based contribution approach is claimed to guarantee correct diagnosis. It has been examined in this paper that the approach fails to precisely locate faulty variables when encountering multiple sensor faults. A fault isolation chart on principal component (PC) subspace is provided to locate faulty variables for a process with multiple operating regions. The results of the quadruple-tank process simulation show the proposed approach successfully locate faulty variables in a case of multiple sensor faults, as long as the process behavior can be depicted by the scores on the PC subspace.