Elsevier, Chemometrics and Intelligent Laboratory Systems, (134), p. 10-22, 2014
DOI: 10.1016/j.chemolab.2014.03.002
Full text: Download
In order to ensure operation safety and consistent product quality, multivariate statistical methods have been widely adopted in chemical batch process monitoring. In this paper, a semi-supervised mixture discriminant monitoring (SMDM) scheme is proposed, which integrates the strengths of both supervised and unsupervised techniques. The semi-supervised characteristic enables SMDM to fully make use of both labeled and unlabeled data, leading to more reliable process models. In addition, SMDM is suited to handling non-Gaussian distributed data that are commonly observed in batch processes. Inheriting from supervised learning, SMDM has better online fault diagnosis capability of known faults compared to the unsupervised multivariate statistical process monitoring methods. Meanwhile, the utilization of control charts makes SMDM capable to detect unknown faults. After an unknown fault is detected, the process variables most contributing to the fault can be identified through missing variable analysis. Such information is valuable for process engineers to find out the root cause of the fault. The collected data of the new faults are then used to update the monitoring model. By doing so, the fault diagnosis performance of the monitoring model can be improved online. The proposed method is demonstrated through its application to an injection molding process.