Investigation of a Sparse Autoencoder-Based Feature Transfer Learning Framework for Hydrogen Monitoring Using...

Area of Science:

• Materials Science
• Chemical Engineering
• Sensor Technology

Background:

• Governments worldwide are prioritizing alternative fuels like hydrogen-enriched natural gas (HENG) to reduce carbon emissions.
• Safe handling of hydrogen (H2) is critical due to its flammability and explosivity, necessitating sensitive detection sensors.
• Existing microfluidic-based metal-oxide-semiconducting (MOS) sensors lack real-time analysis for precise gas concentration identification.

Purpose of the Study:

• To develop a novel machine learning framework for accurate hydrogen gas concentration estimation in HENG mixtures.
• To address the limitations of current deep learning models that require large datasets and struggle with data shifts.
• To enable real-time safety monitoring of HENG using cost-effective microfluidic gas detectors.

Main Methods:

• Proposed a Sparse Autoencoder-based Transfer Learning (SAE-TL) framework.
• Utilized time-series data from a 3D printed microfluidic detector with two commercial MOS sensors.
• Developed a modular gas detector with coated microchannels for gas selectivity based on diffusion rates.

Main Results:

• The SAE-TL framework achieved dominant performance, with an R-squared value of 94%, outperforming typical machine learning models.
• Successfully estimated hydrogen gas concentrations in simulated HENG mixtures using limited datasets.
• Demonstrated the framework's ability to adapt predictive models to new MOS sensor responses.

Conclusions:

• The SAE-TL framework offers a robust solution for accurate hydrogen gas concentration estimation in HENG.
• This approach overcomes data limitations and instrumental variations common in sensor-based monitoring.
• The framework is highly adaptable for real-world applications in HENG safety and monitoring systems.