Non-Linear Machine Learning with Active Sampling for MOX Drift Compensation

Abstract

Metal oxide (MOX) gas detectors based on SnO_2 provide low-cost solutions for real-time sensing of complex gas mixtures for indoor ambient monitoring. With high sensitivity under ideal conditions, MOX detectors may have poor long-term response accuracy due to environmental factors (humidity and temperature) along with sensor aging, leading to calibration drifts. Finding a simple and efficient solution to correct such calibration drifts has been the subject of numerous studies but remains an open problem. In this work, we present an efficient approach to MOX calibration using active and transfer sampling techniques coupled with non-linear machine learning algorithms, namely neural networks, extreme gradient boosting (XGBoost) and radial kernel support vector machines (SVM). Applied on the UCI's HT detectors dataset, the study evaluates methods for active sampling, makes an assessment of suitable neural networks architectures and compares the performance of neural networks, XGBoost and radial kernel SVM to classify gas mixtures (banana and wine odours, clean air) in the presence of humidity and temperature changes. The results show high classification accuracy levels (above 90%) and confirm that active sampling can provide a suitable solution.