Drift Compensation for Electronic Nose by Semi-Supervised Domain Adaption

Abstract

Drift compensation is an important issue for electronic nose systems. Traditional methods are costly and laborious because they need to frequently recalibrate referred gases or continually provide data labeling. In this paper, a new drift compensation method is proposed. The inspiration of our method is originated from semi-supervised domain adaption that can effectively tackle the mismatches between source domain and target domain. In our approach, a weighted geodesic flow kernel is initially constructed, then the combination of such kind of kernels is proposed considering that there are intermediate unlabeled data between the source and target domains. We will discuss how unlabeled data is selected from the target domain. The selected unlabeled data is used to provide incremental knowledge in order to dynamically adapt classifier to the target domain. Based on the kernel combination and selected unlabeled data, manifold regularization is used to train the classifier. To the best of our knowledge, we are the first to apply domain adaption to deal with the sensor drift problem. The advantages of our method include degrading recalibration rate, requiring few labeled data, and the robustness in handling the drift. Our experiments show that the proposed method significantly outperforms the baseline methods.