Comparison of Linear and Nonlinear Dimension Reduction Techniques for Automated Process Monitoring of a Decentralized Wastewater Treatment Facility

• Activity Number: 264
• Type: Contributed
• Date/Time: Monday, August 10, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Karen Kazor*
• Companies: Colorado School of Mines
• Keywords: multivariate statistical process control ; nonlinear time-varying processes ; principle component analysis ; kernel principal component analysis ; locally linear embedding ; wastewater reuse

Abstract

Decentralized wastewater treatment can increase water availability and reduce energy demands. However, with greater variability in the quality and quantity of influent, decentralized systems require tighter control and faster response to malfunctioning. In this work, we analyze the performance of multivariate statistical methods for on-line monitoring. First, we extend a simulation study proposed by Dong and McAvoy (1996) in order to account for autocorrelation and/or nonstationarity in process data. In this context, we consider the monitoring performance of static, dynamic, adaptive, and adaptive-dynamic versions of principal component analysis (PCA), kernel PCA (KPCA), and locally linear embedding (LLE). In each case, we compare both parametric and nonparametric threshold methods. Then, we apply these methods to real-world process data collected from a membrane bioreactor (MBR) during normal and abnormal operations. We find that adaptive-dynamic versions of all three methods generally improve results when data is nonstationary. When applied to the MBR data, these methods perform similarly, however nonparametric thresholds considerably improve results for all three methods.