Drift-Robust Lightweight Deep Learning on Open Gas Sensor Benchmarks: A Reproducible Architecture Study with CBRN...

Area of Science:

• Edge AI
• Gas Sensor Technology
• Machine Learning for Environmental Monitoring

Background:

• Resource-constrained edge processors on UAVs and wearables need robust gas classification.
• Existing methods are server-grade or degrade under sensor drift.
• CBRN (Chemical, Biological, Radiological, and Nuclear) scenarios require reliable, compact detection.

Purpose of the Study:

• Develop an end-to-end pipeline for drift-robust edge gas classification.
• Address limitations of existing server-grade and drift-prone models.
• Create a standardized benchmark for evaluating edge AI gas classifiers.

Main Methods:

• Proposed LiteSensor-Net: a 1D CNN with depth-wise separable convolutions.
• Integrated INT8 quantization and structured magnitude pruning for model compression.
• Employed knowledge distillation domain adaptation (KD-DM) for sensor drift compensation.
• Mapped UCI gas classes to CBRN behavioral categories using physicochemical analogies.

Main Results:

• LiteSensor-Net achieved 92.63% accuracy and 0.898 macro-F1 on the UCI Gas Sensor Array Drift Dataset.
• Optimized model size: 5.99 kB (INT8 pruned), inference latency: 6.3 ms, energy: 0.04 mJ.
• KD-DM-20 improved accuracy by +9.25 pp over uncompensated models under chronological evaluation.
• Introduced a six-metric framework for standardized edge AI gas classifier evaluation.

Conclusions:

• The proposed pipeline offers an open-source, deployable foundation for edge-class gas classification.
• LiteSensor-Net and KD-DM effectively address model size, latency, and sensor drift challenges.
• The system is suitable for environmental monitoring and CBRN detection applications on edge devices.