Electronic Nose Drift Suppression Based on Smooth Conditional Domain Adversarial Networks.

Area of Science:

• Sensor Technology
• Machine Learning
• Artificial Intelligence

Background:

• Gas sensor drift poses a significant challenge, altering data distributions and causing classification algorithm failure.
• Drift leads to inconsistencies between measured and calibrated data, impacting sensor reliability.

Purpose of the Study:

• To introduce a novel deep transfer learning method to counteract gas sensor drift.
• To ensure consistency between drifted and clean data probability distributions for improved classification.

Main Methods:

• Utilized Conditional Adversarial Domain Adaptation Network (CDAN) with Sharpness Aware Minimization (SAM) optimizer.
• Developed feature extractors and domain discriminators to identify shared features in drift and clean data.
• Employed semi-supervised learning, eliminating the need for drift data labels.

Main Results:

• The proposed method demonstrated effectiveness in aligning drifted and clean data distributions.
• Achieved superior generalization capabilities by amplifying shared features for classification.
• Outperformed state-of-the-art techniques on two public datasets.

Conclusions:

• The CDAN+SAM method offers an efficient and effective solution for gas sensor anti-drift.
• The approach accommodates inter-class correlations and exhibits improved training convergence.
• This methodology significantly enhances the robustness and reliability of gas sensing systems.