Joint Distributed Manifold Preserving Domain Adaptation for Drift Compensation in E-nose

Gas sensor drift is a critical problem in E-nose since it negatively affects recognition performance. Domain adaptation is currently an advanced method widely used for sensor drift compensation. However, the existing methods are limited by insufficient consideration of both data distribution and structure information, neglect of preserving the inherent manifold structure in both domains, and the negative impact of false pseudo-labels on feature learning. To address these limitations, a novel joint distributed manifold preserving domain adaptation (JD-MPDA) method is proposed for drift compensation in E-nose in this study. First, a joint subspace learning framework is established by simultaneously integrating marginal distribution alignment, conditional distribution alignment, and data manifold preserving for both source and target domains. Second, the supervised and unsupervised manifold preserving strategies are designed for processing the two domains individually based on marginal fisher analysis and locality preserving projection, and the proposed method can effectively capture geometric and discriminative information in subspace learning. In addition, a pseudo-labeling sample selection strategy is constructed for target data using conditional entropy loss. This strategy can identify high-confidence samples to correct conditional distribution inconsistencies and improve the feature learning of JD-MPDA. Finally, the optimization problem can be solved by generalized eigenvalue decomposition, and the subspace is learned by refining the target domain’s pseudo-labels in iterations. The experimental results demonstrate that the JD-MPDA can improve the classification accuracy by 7.41%, 2.30%, and 11.60% in the short-term drift, long-term drift, and drift & shift scenarios, respectively.