An On-line Reconfigurable Classification Algorithm Improves the Long-term Stability of Gas Sensor Arrays in Case of Faulty and Drifting Sensors

In this work, we illustrate an autonomous and real-time reconfigurable classifier. The algorithm starts from a non-adaptive classifier and evolves during the routine operation of sensors providing a dynamic optimization of the feature selection and refinement of classes’ distribution. The model has been tested on an experimental dataset and the results show that the algorithm may improve the resilience of classifiers in case of drifting and/or faulty sensors. The outcome of this studied case suggests that the algorithm might be able to enhance long-term performance almost independently from which classification model is considered.