Improving the transfer ability of prediction models for electronic noses

Abstract Calibration transfer is attracting more and more attention in the field of electronic noses (e-noses). It aims at making the prediction model trained on one device transferable to other devices, which is important for the large-scale deployment of e-noses, especially when the cost of sample collection is high. In this paper, the transfer ability of prediction models is improved in two steps. First, windowed piecewise direct standardization (WPDS) is used to standardize the slave device, i.e. to transform the variables from the slave device to match the master one. Then, data from the master device are used to develop prediction models with a novel strategy named standardization error based model improvement (SEMI). Finally, the standardized slave data can be predicted by the models with a better accuracy. The proposed WPDS is a generalization of the widely used PDS algorithm. The main idea of SEMI is to make the trained models rely more on variables with small standardization errors, thus less sensitive to the inconsistency of the devices. It links the standardization step and the prediction step. To evaluate the algorithms, three e-noses specialized for breath analysis are adopted to collect a dataset, which contains pure chemicals and breath samples. Experiments show that WPDS outperforms previous methods in the sense of standardization error and prediction accuracy; SEMI consistently enhances the accuracy of the master model applied to standardized slave data. This study provides effective and extensible methods for model transfer of e-noses.