F-Norm-Based Soft LDA Algorithm for Fault Detection in Chemical Production Processes.

Area of Science:

• Chemical Engineering
• Data Science
• Process Monitoring

Background:

• Outliers are common in chemical production, complicating data analysis.
• Existing feature extraction methods are often sensitive to outliers, hindering accurate fault detection.
• Key process features can be overlooked by algorithms focusing on secondary characteristics.

Purpose of the Study:

• To propose a novel algorithm, Frobenius norm based soft linear discriminant analysis (FBSLA), for robust feature extraction and improved fault detection.
• To enhance the reliability of fault detection in chemical processes by addressing outlier sensitivity.
• To improve the identification of critical features in process data.

Main Methods:

• Developed FBSLA utilizing the Frobenius norm for enhanced robustness against outliers.
• Introduced a nonreduced dimensionality projection matrix to clarify training data features.
• Implemented soft constraints to mitigate outlier-induced sensitivity, unlike traditional hard constraints.

Main Results:

• FBSLA demonstrated superior performance in fault detection accuracy compared to existing algorithms.
• Experiments on the Tennessee Eastman Process and Penicillin Fermentation Process data validated FBSLA's effectiveness.
• The algorithm successfully enhanced the prominence of key features despite data outliers.

Conclusions:

• FBSLA offers a significant advancement in fault detection for chemical production processes.
• The algorithm's robustness and focus on key features lead to higher accuracy.
• FBSLA provides a more reliable approach to process monitoring and anomaly identification.