Multivariate statistical diagnosis using triangular representation of fault patterns in principal component space

Abstract

A pattern-based multivariate statistical diagnosis method is proposed to diagnose a process fault on-line. A triangular representation of process trends in the principal component space is employed to extract the on-line fault pattern. The extracted fault pattern is compared with the existing fault patterns stored in the fault library. A diagnostic decision is made based on the similarity between the extracted and the existing fault patterns, called a similarity index. The diagnosis performance of the proposed method is demonstrated using simulated data from Tennessee Eastman process. The diagnosis success rate and robustness to noise of the proposed method are also discussed via computational experiments.