GT-KANet: Robust Acetone Prediction for the Yttria-Stabilized Zirconia-Based Mixed Potential Type Sensor.

Area of Science:

• Materials Science and Engineering
• Chemical Sensing
• Artificial Intelligence

Background:

• Yttria-stabilized zirconia (YSZ)-based mixed potential gas sensors offer high sensitivity and selectivity for acetone detection.
• Practical application is limited by sensor output drift, noise, and fluctuations due to temperature and humidity variations.

Purpose of the Study:

• To develop a robust deep-learning algorithm for accurate acetone concentration prediction and temperature-humidity compensation in YSZ gas sensors.
• To enhance the feasibility of portable acetone detection systems for real-world applications.

Main Methods:

• Developed GT-KANet, a hybrid deep-learning algorithm integrating Gated Recurrent Unit (GRU) networks, transformer layers with ContraNorm, and a Kolmogorov-Arnold Network (KAN) module.
• Trained the model on a comprehensive dataset from a custom YSZ sensor under controlled environmental conditions.
• Utilized knowledge distillation to reduce model parameter size.

Main Results:

• GT-KANet achieved high predictive accuracy (RMSE = 0.0920 ppm, MAE = 0.0542 ppm) across diverse operating temperatures, humidity levels, and acetone concentrations.
• Demonstrated excellent stability and adaptability, outperforming existing compensation methods.
• Reduced model parameter size by 55% via knowledge distillation while preserving prediction accuracy.

Conclusions:

• GT-KANet provides a competitive and effective approach for temperature and humidity compensation in YSZ-based acetone sensors.
• The algorithm significantly enhances the potential for deploying accurate and stable portable gas sensing systems in challenging environments.
• The optimized model is suitable for resource-constrained embedded platforms, broadening application scope.