A Holistic Evaluation of Multivariate Statistical Process Monitoring in a Biological and Membrane Treatment System

Area of Science:

• Environmental Engineering
• Process Control
• Data Science

Background:

• Unsupervised process monitoring for fault detection in municipal wastewater treatment plants (WWTPs) is challenging due to complex, high-volume sensor data.
• Existing methods often fail to detect subtle or multivariate faults promptly.

Purpose of the Study:

• To extensively test and tune an unsupervised process monitoring method for early fault detection in a full-scale decentralized WWTP.
• To evaluate the efficacy of Adaptive dynamic Principal Component Analysis (AD-PCA) in identifying various fault types.

Main Methods:

• Adaptive dynamic Principal Component Analysis (AD-PCA), a modified dimension reduction technique, was employed.
• Data were subset by treatment processes and operating states; models used week-long training windows and moderate cumulative variance.
• Fault detection thresholds were set high to ensure timely identification of deviations.

Main Results:

• The implemented AD-PCA method successfully detected spike faults, univariate drift faults, and multivariate shift faults.
• Faults were identified prior to existing operational thresholds, demonstrating improved early detection capabilities.
• The system showed effectiveness in monitoring complex, multivariate processes in real-time.

Conclusions:

• AD-PCA is a promising unsupervised method for prompt fault detection in WWTPs.
• Further research is needed to refine outlier removal and detect multivariate drift faults for enhanced real-time monitoring.
• Optimized AD-PCA can improve the operational stability and efficiency of wastewater treatment processes.