Abstract: Multivariate control charts are often used to monitor process with more than one variable and detect whether there are systematic deviations in the process. When the distribution of the process variables is unknown, nonparametric methods are commonly used alternatively. To the multivariate process with no baseline distribution, a nonparametric multivariate process monitoring method is proposed based on LSSVM (least squares support vector machine). Under the situation where there are in-control data (denoted as reference dataset) only, the incoming process observations are preprocessed using the moving window method and used to dynamically train the LSSVM model together with the reference dataset. Then the distances between the samples in the moving windows and the hyperplane of the trained LSSVM are used as monitoring statistics. The design of the monitoring model and parameter selection are discussed and the performance of the model is evaluated both with simulations and with a real case.

Key words: multivariate process monitoring, least squares support vector machine, moving window, nonparametric method