Correlated and weakly correlated fault detection based on variable division and ICA

Abstract

In many industrial processes, the correlations of multiple variables are complicated. Some variables are correlated and some are weakly correlated with others, which should be considered in process modelling and fault detection. This paper proposes a correlated and weakly correlated fault detection approach, which is mainly based on variable division and independent component analysis (ICA). A few variables are weakly correlated with others and fault detection should be implemented separately for correlated and weakly correlated subspaces. Firstly, variable division based on weighted proximity measure is presented to obtain correlated and weakly correlated variables. Then, ICA is used for fault detection in correlated subspace and weakly correlated subspace, which needs not kernel mapping or kernel parameter setting. Finally, comprehensive statistics are built based on different subspaces. The proposed method considers the correlated and weakly correlated characteristics of variables and the advantages of ICA in handling weakly correlated variables. The monitoring results of the numerical system and Tennessee Eastman (TE) process have been used to demonstrate effectiveness and superiority of the proposed approach.