A Fault Magnitude-Based Strategy for Effective Fault Diagnosis and Isolation

Abstract

Various multivariate statistical methods based on pattern recognitions for the Tennessee Eastman (TE) process have been developed to identify and diagnose the root cause of assumed faults. However, even if the same fault occurs, its patterns or traces generated from such conventional approaches can be different according to fault magnitudes. Thus, the fault magnitude should be considered. In this study, a signed digraph (SDG) based on process knowledge is used to identify the relationships between process variables and conceivable faults. A support vector regression (SVR) and dynamic independent component analysis (DICA) are then applied to construct empirical models as a function of process variables associated with assumed faults and their fault magnitudes for isolating a fault and handling non-Gaussian information. In addition, empirical models for predicting fault magnitudes are constructed. The efficacy of the proposed approach is illustrated by comparing it with the previous studies applied for the benchmark process.