Statistical Process Monitoring Based on Ensemble Structure Analysis

Abstract

Feature extraction with the aim of dimensionality reduction plays a key role in developing statistical monitoring models. The projection directions obtained from the principal component analysis U+0028 PCA U+0029, locality preserving projections U+0028 LPP U+0029, and even multi-manifold projections U+0028 MMP U+0029 algorithms all can provide useful information for fault detection given that the corresponding models are trained in an unsupervised manner. By taking advantages of all these dimensionality reduction algorithms, a novel process monitoring model based on ensemble structure analysis U+0028 ESA U+0029 is presented in this article. The ESA model takes advantage of PCA, LPP and MMP models and combines the multiple solutions within an ensemble result through Bayesian inference. In the developed ESA model, different structure features of the given dataset are taken into account simultaneously, the suitability and reliability of the ESA-based monitoring model are then illustrated through the Tennessee Eastman benchmark process.