An evolutionary weighted feature influence factor feature selection method for fault detection in the Tennessee...

Area of Science:

• Chemical Engineering
• Process Control
• Data Science

Background:

• Existing fault detection methods for the Tennessee Eastman process exhibit low accuracy (<70%) for complex faults (d03, d09, d15).
• Current approaches often overlook feature subset stability during repeated selections, focusing primarily on discriminative power and dimensionality.

Purpose of the Study:

• To propose an evolutionary weighted feature influence factor (WFIF) method for improved fault detection in the Tennessee Eastman process.
• To enhance the stability and reduce dimensionality of feature subsets for more robust fault diagnosis.

Main Methods:

• Developed an evolutionary weighted feature influence factor (WFIF) algorithm.
• Defined and quantified feature subset stability based on consistency across data partitions.
• Conducted comparative experiments against six existing feature selection techniques.

Main Results:

• The evolutionary WFIF method improved the F1-score for complex faults (d03, d09, d15) from below 73% to over 85%.
• The average F1-score for all 21 TE process faults reached 95.79%, with a significant reduction in selected features to 9.5% of original variables.
• Demonstrated exceptional stability, with a standard deviation of zero for selected feature subset dimensions in 20 out of 21 faults.

Conclusions:

• The evolutionary WFIF method offers superior performance in fault detection accuracy and feature selection stability for the Tennessee Eastman process.
• The selected feature subsets exhibit excellent model generalization capabilities, validated through cross-classifier testing.
• This approach addresses key limitations in existing methods, providing a more reliable tool for industrial process monitoring.