LAUSR

Abstract Canonical variate analysis (CVA) has shown its superior performance in statistical process monitoring due to its effectiveness in handling high-dimensional, serially, and cross-correlated dynamic data. A restrictive condition for CVA is that the covariance matrices of dependent and independent variables must be invertible, which may not hold when collinearity between process variables exists or the sample size is small relative to the number of variables. Moreover, CVA often yields dense canonical vectors that impede the interpretation of underlying relationships between the process variables. This article employs a sparse CVA (SCVA) technique to resolve these issues and applies the method to process monitoring. A detailed algorithm for implementing SCVA and its formulation in fault detection and identification are provided. SCVA is shown to facilitate the discovery of major structures (or relationships) among process variables, and assist in fault identification by aggregating the contributions from faulty variables and suppressing the contributions from normal variables. The effectiveness of the proposed approach is demonstrated on the Tennessee Eastman process.