Nonlinear Dynamic Process Monitoring Based on Ensemble Kernel Canonical Variate Analysis and Bayesian Inference

Area of Science:

• Process monitoring and control
• Statistical process control
• Machine learning for industrial applications

Background:

• Canonical Variate Analysis (CVA) improves fault detection by considering autocorrelation in process data.
• Kernel CVA (KCVA) extends CVA for nonlinear dynamic processes using kernel principal component analysis (KPCA).
• The performance of Gaussian kernel-based KCVA is highly dependent on the manually selected kernel bandwidth.

Purpose of the Study:

• To develop a novel ensemble kernel canonical variate analysis (EKCVA) method.
• To improve fault detection performance in nonlinear dynamic processes.
• To address the limitations of manual kernel bandwidth selection in KCVA.

Main Methods:

• Integrating ensemble learning with kernel canonical variate analysis (KCVA).
• Establishing multiple KCVA models with different kernel bandwidths.
• Combining T^2 and Q monitoring statistics from base models using Bayesian inference.

Main Results:

• The proposed EKCVA method demonstrates superior fault detection performance compared to single KCVA models.
• Validation through a numerical example and two industrial benchmarks (CSTR and Tennessee Eastman process).
• The ensemble approach effectively leverages multiple learners with varied kernel bandwidths.

Conclusions:

• EKCVA offers enhanced generalization and robustness for fault detection in complex industrial processes.
• The Bayesian combination of monitoring statistics further improves detection accuracy.
• This method provides a more reliable approach to nonlinear process monitoring.