High-fidelity embedded gas sensing: a unified deep learning framework for simultaneous identification and quantification

Abstract Electronic nose systems are a transformative technology, enabling low-cost, real-time gas analysis for critical applications in environmental monitoring, industrial safety, and medical diagnostics. However, their practical performance is fundamentally limited by the computational challenge of analyzing time-series data from their sensor arrays. Conventional machine learning (ML) pipelines rely on lossy feature engineering and employ separate, inefficient models for gas classification and concentration regression, a methodology that fundamentally limits predictive accuracy. This paper presents a unified multitask deep learning model that solves these problems through three key innovations. First, the model learns features directly from the complete, raw sensor time-series using a convolutional neural network, avoiding information loss from pre-processing. Second, a novel attention mechanism links the classification and regression tasks, allowing the model to use the identified gas type to improve its concentration estimate. Third, a custom loss function was developed to ensure high accuracy across a wide range of gas concentrations, a critical challenge in chemical sensing. Using 5-fold cross-validation, the model achieved 100% accuracy in gas classification and a mean absolute percentage error (MAPE) of 6.61% for concentration regression, results that significantly outperform classical ML benchmarks. We also demonstrated the model’s efficiency by deploying it on 10 different STM32 microcontrollers, confirming its suitability for real-time embedded applications. This work provides a robust framework for developing high-performance, AI-based chemical sensors and offers a foundational methodology for future analysis of complex gas mixtures. The source code and data are provided in this GitHub repo: https://github.com/quenx-rd/gas_identifier_public .