TinyML-Based Real-Time Drift Compensation for Gas Sensors Using Spectral–Temporal Neural Networks

Abstract

The implementation of low-cost sensitive and selective gas sensors for monitoring fruit ripening and quality strongly depends on their long-term stability. Gas sensor drift undermines the long-term reliability of low-cost sensing platforms, particularly in precision agriculture. We present a real-time drift compensation framework based on a lightweight Temporal Convolutional Neural Network (TCNN) combined with a Hadamard spectral transform. The model operates causally on incoming sensor data, achieving a mean absolute error below 1 mV on long-term recordings (equivalent to <1 particle per million (ppm) gas concentration). Through quantization, we compress the model by over 70%, without sacrificing accuracy. Demonstrated on a combustion-type gas sensor system (dubbed GMOS) for ethylene monitoring, our approach enables continuous, drift-corrected operation without the need for recalibration or dependence on cloud-based services, offering a generalizable solution for embedded environmental sensing—in food transportation containers, cold storage facilities, de-greening rooms and directly in the field.

Keywords

• gas sensor drift
• TinyML
• convolutional neural networks
• spectral filtering
• hadamard transform
• real-time signal processing