Hierarchical Support Vector Data Description for Batch Process Monitoring

Abstract

Batch process monitoring remains a challenging task due to the inherent time-varying dynamics. The occurrence of faults results in the variables having three types of variation behavior, that is, univariate variation without interaction effect, correlated variable variation with interaction effect, or all variable variation. The three behavior situations may change at any time under dynamic scenarios. Considering that an effective batch process monitoring method should clearly identify the complex variation behavior of the variables, a hierarchical support vector data description (HSVDD), which integrates univariate monitoring, subspace monitoring, and whole-space monitoring, is proposed to accurately identify the three variation behaviors. First, three-dimensional data are unfolded to batch-wise data with normalization, and the obtained two-dimensional data are separated according to the time axis to obtain time-slice modeling data. Then, the hierarchical structure, which contains the univariate variable...