Application of Feature Extraction Methods for Chemical Risk Classification in the Pharmaceutical Industry

Area of Science:

• Data Science
• Chemical Engineering
• Cybersecurity

Background:

• Production sites generate large datasets from toxic and physicochemical sensors, necessitating feature reduction for risk assessment.
• Cybersecurity data also influences risk assessment but requires integration with sensor data.
• Classical feature reduction methods like Kaiser criterion and scree plot analysis have limitations.

Purpose of the Study:

• To develop an advanced feature dimensionality assessment method using correlation and discriminant power.
• To improve the accuracy of risk assessment for chemical threats compared to existing methods.
• To identify key cybersecurity features impacting chemical hazard risk.

Main Methods:

• Feature extraction from sensor and cybersecurity datasets.
• Dimensionality reduction using a novel method based on correlation and discriminant power.
• Classification and comparison with classical methods (Kaiser criterion, scree plot, PCA).
• Factor rotation to class centroids for enhanced risk assessment.

Main Results:

• The proposed method significantly improved classification quality by approximately 7% over no feature extraction and 4% over classical PCA.
• Factor rotation to class centroids proved most effective for chemical threat risk assessment.
• A specific subspace of cybersecurity features (e.g., lost packets, incorrect logins, spam) combined with volatile substance data improves risk assessment.

Conclusions:

• The developed feature reduction method offers superior performance for chemical threat risk assessment.
• Integrating specific cybersecurity metrics with sensor data enhances the accuracy of chemical hazard risk evaluation.
• The findings are applicable to Industry 4.0 systems for real-time risk monitoring.