Interpretable Calibration Transfer and Drift Compensation for MOS Gas Sensors in Complex Gas Mixtures

Area of Science:

• Sensor Technology
• Machine Learning
• Chemical Sensing

Background:

• Metal oxide semiconductor (MOS) gas sensors require frequent recalibration due to domain shifts like sensor variation and aging.
• Traditional recalibration is time-consuming and labor-intensive, hindering reliable field deployment.
• Existing interpretable machine learning models (FESR) lack inherent capabilities for calibration transfer and drift compensation.

Purpose of the Study:

• To develop a novel, model-based approach for calibration transfer and drift compensation in MOS gas sensors.
• To reduce the effort and time required for recalibration while maintaining or improving model accuracy.
• To enable interpretable quantification of individual volatiles in complex mixtures using MOS sensors.

Main Methods:

• Representing interpretable feature extraction, feature selection, and regression (FESR) models as deep neural networks (IDNNRep).
• Applying transfer learning techniques from deep neural networks (DNNs) to facilitate knowledge reuse across calibration domains.
• Evaluating the IDNNRep approach across multiple gases (e.g., acetone, toluene) under various domain-shift scenarios.

Main Results:

• The proposed IDNNRep method significantly reduced the root mean square error (RMSE) by up to 93% compared to initial models and 89% compared to orthogonal signal correction (OSC).
• Achieved normalized RMSE values of 6.3-9.3% using only 10% of the calibration data.
• The interpretable nature of the FESR structure within IDNNRep provided additional sensor- and gas-specific insights.

Conclusions:

• The IDNNRep approach offers an effective solution for calibration transfer and drift compensation in MOS gas sensors.
• This method substantially reduces recalibration effort and time, enabling more robust and reliable field applications.
• The interpretable framework provides valuable insights into sensor behavior and gas quantification in complex mixtures.