Semi-supervised comparative learning compensation method for chemical gas sensor drift.

Area of Science:

• Chemical sensing
• Machine learning
• Signal processing

Background:

• Sensor drift causes unpredictable chemo-sensory signal variations, posing a significant challenge for chemical sensors.
• Traditional drift compensation methods are costly and labor-intensive due to frequent recalibration requirements.
• Addressing drift is crucial for reliable and accurate chemical sensing applications.

Purpose of the Study:

• To propose a novel algorithm framework, semi-supervised contrastive learning drift compensation (SSCLDC), to address sensor drift.
• To reduce computational load and enhance classifier performance by utilizing a small number of drift calibration samples.
• To overcome data distribution differences between source and target domains caused by sensor drift.

Main Methods:

• Utilizing a multilayer perceptron to extract high-level abstract features for source data representation.
• Incorporating reference sample pairs for semi-supervised learning to handle domain distribution differences.
• Employing a contrastive loss function to effectively represent the matching degree of paired samples.
• Using the Kennard-Stone sequential algorithm for selecting representative reference samples.

Main Results:

• The proposed SSCLDC method demonstrated superior performance compared to classic drift compensation techniques.
• Experiments on a long-term chemical gas sensor drift dataset validated the method's effectiveness.
• The algorithm successfully compensated for sensor drift, improving overall sensor accuracy and reliability.

Conclusions:

• The SSCLDC framework offers an effective and practical solution for sensor drift compensation.
• Semi-supervised contrastive learning provides a robust approach to handle drift-induced data distribution shifts.
• The method reduces the need for extensive recalibration, making drift compensation more efficient.