Process monitoring using kernel density estimation and Bayesian networking with an industrial case study.

Area of Science:

• Process monitoring and control
• Data-driven fault detection
• Statistical process control

Background:

• Principal Component Analysis (PCA) and Independent Component Analysis (ICA) are common for detecting process abnormalities.
• These methods rely on Gaussian assumptions and can present interpretation challenges for online fault detection.
• Existing methods like PCA and ICA, while effective in detection and localization, struggle with non-Gaussian data and timely interpretation.

Purpose of the Study:

• To propose a novel dimension reduction technique using Bayesian networks for process fault detection.
• To enhance the interpretability and intelligence of dimension reduction by incorporating process knowledge.
• To address limitations of PCA and ICA, particularly concerning Gaussian assumptions and result interpretation.

Main Methods:

• Utilizing Bayesian networks for dimension reduction, integrating process knowledge for intelligent analysis.
• Combining Bayesian networks with multivariate kernel density estimation for abnormality detection.
• Comparing the performance of Bayesian networks against PCA and ICA using industrial plant data.

Main Results:

• Bayesian networks demonstrated effective dimension reduction and abnormality detection, particularly for non-linear and non-Gaussian data.
• The proposed method offered more intuitive interpretation of results compared to PCA and ICA.
• Performance evaluation on industrial data confirmed the advantages of Bayesian networks in fault detection scenarios.

Conclusions:

• Bayesian networks offer a promising alternative for process fault detection, overcoming limitations of traditional PCA and ICA.
• The integration of process knowledge enhances the interpretability and effectiveness of dimension reduction techniques.
• This approach facilitates quicker and more informed decision-making in industrial process monitoring.