Information Criterion for Determination Time Window Length of Dynamic PCA for Process Monitoring

摘要

Principal component analysis (PCA) is based on and suitable to analysis of stationary processes. When it is applied to dynamic process monitoring, the moving time window approach is used to construct the data matrix to be analyzed. However, the length of the time window and the moving width between time widows are often empirically tested and selected. In this paper, a criterion for determining the time window length is proposed for dynamic process monitoring. A new algorithm of dynamic monitoring is then presented. The proposed selection criterion is used in the new algorithm. Finally, the proposed approach is successfully applied to a two-input two-output process and Tennessee Eastman process for dynamic monitoring.