A Unified Deep-Learning Framework for Smart Gas Sensing.

Area of Science:

• Artificial Intelligence
• Sensor Technology
• Machine Learning

Background:

• Smart perception systems face challenges in multi-functional inference, cross-scenario deployment, and long-term stability.
• Existing sensing frameworks struggle to meet escalating demands for diverse applications like environmental monitoring and clinical diagnostics.

Purpose of the Study:

• To propose a unified, computationally efficient deep learning framework for intelligent sensing systems.
• To enhance reliability across tasks, scenarios, and time by integrating multi-task learning, transfer learning, and domain adaptation.
• To develop a lightweight, task-aligned model for gas sensing, predicting sensor status, gas identity, and concentration.

Main Methods:

• Developed a lightweight, task-aligned deep learning model using multi-task learning for concurrent prediction.
• Employed SHapley Additive exPlanations (SHAP) for interpretability and sensor-array lightweighting.
• Implemented few-shot structural transfer and semi-supervised adversarial domain adaptation for cross-scenario and cross-period robustness.

Main Results:

• Achieved high accuracy (>0.98 in source, >0.91 in adaptation) with minimal fine-tuning (<2% trainable parameters).
• Demonstrated significantly enhanced robustness against sensor drift (up to 24.7% gain).
• Provided interpretable insights into multi-task synergy and sensor attributions.

Conclusions:

• The proposed framework offers an interpretable and resource-efficient foundation for deployable intelligent sensing systems.
• Successfully addressed fragmented reliability bottlenecks in existing sensing frameworks.
• Enabled cohesive cross-task, cross-scenario, and cross-period reliability in smart perception systems.