A fault diagnosis method based on mutual information and canonical variate analysis

Canonical variate analysis (CVA) is a widely used multivariate statistical method for fault diagnosis. CVA features are extracted by maximizing the correlations between two sets of variables. However, traditional CVA extracts features on the basis of the covariance matrix, which only considers variables' linear correlations, and can not reflect general practical production process variables' nonlinear correlations fully. For above problem, we introduce mutual information into CVA and propose a mutual information and CVA based fault diagnosis method. Mutual information is a nonlinear measure, which can describe random variables' dependence degree. The proposed method extracts features based on mutual information, and builds monitoring statistics. The effectiveness of the proposed method is verified by a continuous stirred tank reactor.