Adaptive Selection of Latent Variables for Process Monitoring

Abstract

Latent-variable (LV) methods have been widely applied to multivariate statistical process monitoring. Using LV methods can simplify the process-monitoring problem through a projection of the high-dimensional process data into a low-dimensional LV space that retains most of the important information for fault detection. A key issue is the optimal selection of a subset of LVs to constitute such a low-dimensional LV space. Traditional LV-selection methods do not choose LVs from a fault-detection point of view, which may result in poor process-monitoring performance. To overcome this drawback, a cumulative-percent-contribution (CPC) criterion is proposed to select appropriate LVs for process monitoring. First, contributions of LVs to the T2 value of a sample are computed by the decomposition of the T2 statistic. The importance of LVs to fault detection is then evaluated by their contributions. The larger the contribution is, the more important the LV is. After sorting LVs in order of decreasing contributions,...