Cross-Domain Active Learning for Electronic Nose Drift Compensation.

Area of Science:

• Sensor technology
• Machine learning
• Data science

Background:

• Electronic nose (E-nose) systems are susceptible to data drift caused by sensor ageing, leading to distorted data and reduced detection accuracy.
• Existing active learning methods fail to account for feature distribution misalignment between different domains due to drift when selecting samples.

Purpose of the Study:

• To propose a novel cross-domain active learning (CDAL) method to mitigate data drift in E-nose systems.
• To enhance sample selection by considering inter-domain distribution differences and information content.

Main Methods:

• Developed a CDAL framework integrating active learning and domain adaptation.
• Utilized Hellinger distance (HD) and Maximum Mean Difference (MMD) as a weighted criterion for sample selection.
• Employed a Gaussian kernel function for data distribution alignment across domains.

Main Results:

• The proposed CDAL method effectively combines active learning and domain adaptation to assess inter-domain distribution differences.
• Gaussian kernel mapping successfully aligned data distributions between domains.
• CDAL significantly suppressed time drift effects from sensor ageing, improving E-nose detection accuracy for data collected at different times.
• Comparative analysis demonstrated superior drift compensation compared to recent methodological frameworks.

Conclusions:

• The CDAL framework offers a robust solution for drift compensation in E-nose systems.
• Integrating active learning with domain adaptation improves the resilience of E-nose systems to temporal data variations.
• The method optimizes labeled sample selection for efficient drift management and enhanced detection performance.