A simple approach to multivariate monitoring of production processes with non-Gaussian data

Abstract

Statistical monitoring of advanced production processes is becoming increasingly challenging due to the large number of key performance variables that characterize a process. These variables often are non-Gaussian, highly correlated and exhibit non-linear dependencies. Traditional multivariate monitoring methods handle non-Gaussian data through combining both independent component analysis (ICA) and support vector data description (SVDD). However, redundant independent components not only increase the modeling complexity of SVDD but also reduce the accuracy of the monitoring. In this article, we solve the above problems by introducing a Laplacian based weighting score to adjust the ICA-SVDD procedure. The key aspect of our model is that independent components are automictically selected using a Laplacian algorithm, which are inputted to a SVDD model to determine the control limits. The advantageous features of the proposed method are demonstrated through a numerical study as well as a case study which concerns an application to a hot rolling process for monitoring steel production processes.