Weighted Domain Adaptation on Unlabeled Electronic-Nose Drift and Interference Data

Electronic nose (EN) drift suppression always needs plenty of prelabeled calibration data, which is very expensive in both time and expenditure. Accordingly, current mathematical methodologies of drift suppression try to use a domain adaptation paradigm that allows fewer or no labels for model updating. Although domain adaptation omits the category labels of drift data, it still requires identical category compositions for training and drift samples. However, it is hard to be satisfied in out-of-laboratory applications because some unexpected categories beyond the known ones may appear in drift data without any signs. The mathematical model will be disturbed, even invalid, under such uncertain category interference. To solve this problem, we have proposed a learning model allowing unexpected category interferences mixed in drift samples. An index vector was introduced to adaptively separate interference samples during a weighted domain adaptation process, and the drift suppression was realized in the purified subspace. Then, two EN drift datasets were adopted to simulate several scenarios with unlabeled interference odors. Based on these datasets, the proposed approach has achieved the highest performance on interference discrimination and drift data classification among all the adopted mathematical models. Meanwhile, we demonstrated the excellent robustness of the proposed method in a parameter sensitivity analysis. As a result, the presented process in this study can successfully enable ENs to recognize drift gas samples against unlabeled interference, which is a constructive step for the robustness enhancement of an EN.