A Novel Label Disentangling Subspace Learning Based on Domain Adaptation for Drift E-Nose Data Classification

In sensor-related subjects, sensor drift is an urgent and challenging problem because of its negative impact on the recognition performance and long-term detection of sensors. Earlier methods of pattern recognition failed because of the assumption of a consistent data distribution, and the effect of drift actually creates an inconsistent data distribution. Meanwhile, it is less effective to describe the drift directly due to its nonlinear dynamic characteristics. As a representative method of transfer learning, the domain adaptation (DA) strategy has been used to realize drift compensation by many researchers in recent years. However, these methods ignore the negative impact that label information implicit in features may have on model learning. On the basis of DA, this article proposes a novel label disentangling subspace learning (LDSL) method to tackle the electronic nose (E-nose) sensor drift problem. Before implementing joint DA, we disentangle the implied label information so that the model can learn more essential and transferable features. Then, a cross-coupling strategy is proposed to recover feature recognition for the classification task. The proposed method may provide a new perspective for drift E-nose data classification. Extensive experiments on two public gas sensor drift datasets show that LDSL has good drift compensation effects. Meanwhile, LDSL is proven to be a generalized eigenvalue problem that can be easily solved.