Calibration transfer and drift compensation of e-noses via coupled task learning

The problems of instrumental variation and sensor drift are receiving increasing concerns in the field of electronic noses. Because the two problems can be uniformly viewed as a variation of the data distribution in the feature space, they can be handled by algorithms such as transfer learning. In this paper, we propose a novel algorithm framework called transfer sample-based coupled task learning (TCTL). It is based on transfer learning and multi-task learning. Given labeled samples in the source domain (i.e. from the master device or without drift) and a small set of transfer samples as inputs, TCTL simultaneously learns a prediction model for data in the source domain and one for data in the target domain (i.e. from the slave device or with drift). The transfer samples are incorporated into a regularization term of the objective function. TCTL is an extensible framework that can apply to various classification and regression models. When combined with the standardization error-based model improvement (SEMI) strategy, its accuracy can be further enhanced. Experiments on a multi-device dataset and a popular long-term drift dataset show that the proposed algorithms achieve better accuracy compared with typical existing methods with much fewer auxiliary samples needed, which proves their efficacy and usability in real-life applications.