A Transferable Feature-Based Classifier to Improve Transferability of Electronic Nose Systems

Abstract

Sensor drift is an important issue in electronic nose systems. When sensor drift occurs, the data distribution changes, which affects classification accuracy. Thus, this letter proposes a feature selection method and a classifier combined with weighted features to address the sensor drift problem. In the proposed method, feature scores used to select better features are calculated based on their separability and transferability in feature selection. In addition, to consider the transferability of each feature, the transferability score of each feature is used as a feature weight. For the classifier, to reduce the distribution difference between two domains, an extreme learning machine is employed, which make the output weights of the target domain similar to those of the source domain while considering the importance of each feature in terms of the feature scores. Experimental results demonstrate that the average classification accuracy on an open-access drifted dataset can be improved using the proposed method.