Domain Adaptation on Asymmetric Drift Data for an Electronic Nose

Abstract

Domain adaptation, a type of transfer learning, has been theoretically adopted to the electronic nose (EN) drift problem. Current academic achievements in this field are inspired by symmetric data, requiring identical sample categories between source and target domains. Commonly, the source domain comprises prepared EN data with intact ground truth; the target domain contains EN drift data to be recognized without class labels. However, in practical cases, the category number of collected drift data is always less than that of prepared data. Hence, performance degradation would be occurred using traditional domain adaptation. To address this problem, we modified and fused dictionary learning, canonical correlation analysis, and locality preserving projection for the asymmetry domain data. Accordingly, selective first-order, second-order alignments, and topology preservation have been realized via sparse and generalized regulation. Specifically, we presented an associated iterative solution process to gain the projected subspace. Then, we utilized two drift datasets to create several asymmetric EN drift scenarios for performance validation. The experimental results showed the highest recognition accuracies achieved by the proposed method on asymmetric drift data in most scenarios. The proposed methodology was conducted with higher accuracy and stability in drift sample recognition than the other referenced drift compensation methods. It is a suitable choice for EN drift compensation on asymmetric data.