A dual drift compensation framework based on subspace learning and cross-domain adaptive extreme learning machine for gas sensors

Abstract

Sensor drift has been recognized as the root cause of decreased effectiveness in the gas sensor community. To date, most drift compensation strategies have focused on classification accuracy, while neglecting the labelling cost in the real-world. To simultaneously improve classification accuracy and reduce labelling cost, we propose a novel drift compensation framework for gas sensors, which combines cross-domain subspace learning with balanced distribution and cross-domain adaptive extreme learning machine (CSBD-CAELM). First, in a common subspace, the marginal distribution and conditional distribution between the source domain (clean data) and target domain (drift data) are adaptively aligned, which reduces the distribution divergence between domains. Then, the proposed CSBD-CAELM learns a robust classifier by leveraging the guide samples. Finally, both the subspace and the classifier are optimized continually by iteratively refining the soft labels of the target domain. In this way, a dual drift compensation effect combining the feature level and classifier level is achieved. Compared with the existing methods based on transfer samples, the labelling cost is reduced. The experimental results show that, at the same labelling cost, the CSBD-CAELM can improve the accuracy by 19.5% and 26.4% in the long-term drift scenario and short-term drift scenario, respectively.