Machine learning-enhanced multi-sensor e-nose to quantify and classify low-level ammonia under dynamic environmental condition

This study presents an advanced Enose system equipped with a multichannel MOX sensor array (GM102B, GM302B, GM502B, GM702B) and environmental sensors for temperature and humidity compensation, with the aim of detecting ammonia across low to high concentrations (2.5–100 PPM) under varying environmental conditions. A custom sampling setup was developed to evaluate the system across three distinct cases: Case 1: Fixed temperature and humidity (baseline condition), Case 2: Variable humidity (40%, 50%, 60%) with fixed temperature, Case 3: Variable temperature (30 °C, 40 °C, 50 °C, 60 °C) with fixed humidity. These scenarios simulate real-world environmental fluctuations to test system robustness. Classification performance was rigorously evaluated using SVM, XGBoost, and LightGBM across individual sensors, all sensors per case, and all cases per sensor. XGBoost consistently delivered superior results (up to 1.00 accuracy), particularly with the 702B sensor and in data-rich environments, outperforming both SVM and LightGBM. While SVM showed solid, consistent results with lower computational complexity, LightGBM underperformed in most configurations, except with the 102B sensor. Compared to earlier studies using MOX sensors and classifiers like LDA, MLP, and BP-NN (typically 93%–97% accuracy), this approach achieves significantly higher accuracy, improved reliability, and greater adaptability to environmental changes. By thoroughly analyzing environmental impacts across structured scenarios, this research sets a new benchmark for MOX-based ammonia detection and supports the development of scalable, real-time multi-gas monitoring platforms. • A custom Enose system was developed using a multichannel MOX gas sensor array (GM102B, GM302B, GM502B, GM702B) with integrated temperature and humidity sensors. • Ammonia gas detection was successfully performed across a broad concentration range (2.5 to 100 PPM), covering low, medium, and high exposure levels. • XGBoost consistently outperformed SVM and LightGBM in terms of accuracy, precision, recall, and F1-score across all experimental conditions. • LightGBM was introduced for the first time in Enose-based ammonia classification, offering novel insights into its strengths and limitations in gas detection. • Permutation feature importance analysis revealed “Max” and “Min” as the most critical features, significantly influencing classification accuracy and model interpretability. • XGBoost was identified as the most robust and scalable model, well-suited for practical and cost-effective deployment of real-time Enose systems. • Environmental variations were thoroughly investigated across three structured cases involving controlled changes in temperature and humidity: Case 1: 23 °C at 50% RH; Case 2: 30–60 °C at fixed 50% RH; Case 3: 40%–60% RH at fixed 23 °C; Sensor data were collected under all these conditions to assess MOX sensor sensitivity and system robustness in real-world variability. • Three machine learning models—SVM, XGBoost, and LightGBM—were evaluated using key time-domain features such as RMS, Skewness, Max, and Min. • A six-level classification evaluation framework was employed to analyze model performance across individual sensors, combined cases, and full sensor configurations. • Least influential features such as ZC and RES were identified, aiding in future optimization of feature selection and system design.