A Regularized Ensemble of Classifiers for Sensor Drift Compensation

Abstract

Ensemble of classifiers has been recently used in sensor community to address the challenge of identification and fixation of the drift suffered by chemical sensors. Aging and poisoning of the surface of the sensors results in degrading the predictive characteristic of the sensors and exhibits a significant challenge for the sensor developer community. Recent literature suggests ensemble of classifiers with uniform or non-uniform weightage to the participating classifiers in order to improve prediction accuracy in the presence of sensor drift. This paper introduces a new generalized machine-learning approach to deal with this problem by applying the concept of regularization. In this approach, regularization is applied to the weighted ensemble of classifiers to overcome the time-dependent drift occurring in chemical sensors. For the gas discrimination problem, we tested our approach on a publically available time series data set collected over three years using metal–oxide gas sensors. Results clearly indicate the superiority of our approach to recently reported results in achieving higher classification accuracy during testing period with ensemble of classifiers in the presence of sensor drift over time.