Active Learning on Dynamic Clustering for Drift Compensation in an Electronic Nose System

Area of Science:

• Sensor Technology
• Machine Learning
• Data Science

Background:

• Electronic noses (E-noses) require drift correction for stable performance.
• Existing laboratory-based machine learning methods for E-nose drift are often insufficient for real-world applications.
• Obtaining labeled data for drifted samples can be costly and difficult.

Purpose of the Study:

• To develop an innovative methodology for E-nose drift correction in scenarios with limited labeled data.
• To address the challenge of selective sample labeling for effective drift correction.
• To improve the efficiency and accuracy of E-nose drift correction under resource constraints.

Main Methods:

• Proposed an active learning (AL) methodology for selective sample labeling.
• Implemented a dynamic clustering process to balance sample categories for labeling.
• Evaluated the methodology in long-term and short-term E-nose drift scenarios.

Main Results:

• The proposed AL methodology demonstrated superior performance compared to state-of-the-art methods.
• Analyzed parameter sensitivity and accuracy trends.
• The Label Efficiency Index (LEI) confirmed the method's efficiency and cost-effectiveness.

Conclusions:

• The developed active learning approach effectively corrects E-nose drift with minimal labeled data.
• The methodology offers a practical solution for real-world E-nose applications where labeling is expensive.
• This work advances the field of online drift correction for E-noses.