Performance Evaluation of Machine Learning Models for Gas Classification Under Sensor Drift on ESP32-S3

Abstract

Machine learning has become an effective approach for electronic nose systems used in gas classification applications. However, most previous studies evaluate classification performance using randomly partitioned datasets, which often fail to represent the long-term sensor drift encountered in practical deployments. Furthermore, embedded implementation of different machine learning models on resource-constrained microcontrollers has received limited attention. In this study, the Gas Sensor Array Drift Dataset from the UCI Machine Learning Repository was used to investigate both classification performance and embedded deployment efficiency. Four supervised machine learning algorithms, namely Multi-Layer Perceptron (MLP), Support Vector Machine (SVM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost), were evaluated under both random stratified and chronological batch-based data partitioning strategies. The chronological evaluation was designed to simulate realistic sensor drift conditions by separating the training and testing data according to acquisition time. Experimental results showed that all models achieved approximately 99% accuracy under random evaluation, whereas chronological testing produced substantially lower accuracies, highlighting the impact of sensor drift on real-world performance. Among the evaluated models, MLP achieved the highest chronological classification accuracy while demonstrating superior robustness against distribution shifts. To investigate practical applicability, all trained models were deployed on an ESP32-S3 microcontroller. Their average inference latency, flash memory usage, RAM consumption, and total firmware size were experimentally measured and compared. Among the evaluated models, the MLP achieved the highest chronological classification accuracy while also exhibiting the lowest inference latency and smallest firmware footprint on the ESP32-S3 platform.