Locating faulty variables by evaluating ratio of variable contribution based on discriminant analysis for online fault diagnosis

A faulty variable isolation algorithm is proposed in this paper using a new Fisher discriminant analysis (FDA) algorithm termed nested-loop FDA (NeLFDA). The NeLFDA model is developed by pairwise analyzing normal and fault process operation data. Some projection directions that are useful for classification between normal and fault cases are extracted along which a quantitative fault evaluation index is defined which is more sensitive to reveal fault effects. Using an iterative variable selection procedure, process variables that are related to the fault are then identified and ordered by checking contribution to the fault evaluation index. Based on variable separation results with discriminant analysis, two different fault diagnosis models are developed to explore different variable correlations of two variables sets separated in each fault class. One is to check the fault effects relative to normal and the other can reveal the similar variations with those at normal status. Each fault is marked with variable selection result, two fault diagnosis models and their associated confidence limits. Online fault diagnosis is then performed by dually checking the characteristics of fault samples using two fault diagnosis models. Its feasibility and performance are illustrated with simulated and pre-programmed faults using data from the Tennessee Eastman (TE) benchmark process.