Open Set Domain Adaptation for Electronic Nose Drift Compensation on Uncertain Category Data

An electronic nose (e-nose) is an intelligent sensing device with a gas sensor array and corresponding recognition models. The gas sensor drift is a critical issue of an e-nose system, degrading the recognition performance in long-term detections. Current drift suppression methods have been designed on closed-set data considering identical category compositions between drift and training data. However, this assumption cannot be preserved in out-of-laboratory scenes due to the unexpected appearance of unknown category odors in the training and drift data. The category compositions of drift and training data become uncertain with each other. Thus, we have proposed an open-set domain adaptation (OSDA) model to enhance the robustness of drift suppression models on such open-set data. The proposed methodology can obtain a common subspace with aligning known and separating unknown categories. Accordingly, the first- and second-order statistics are utilized in domain adaptation (DA) considerations. We have also introduced an open-set labeler for drift category recognition and unknown category discrimination. We chose two gas sensor drift datasets to evaluate the model performance. The experimental results indicate that our proposed method performs better drift compensation on open-set data than the other adopted DA methods. We have further explored the learning term effectiveness, parameter sensitivity, and subspace dimension optimization to verify the characteristics of the proposed methodology. As a result, it is proved that the proposed model can cope well with the uncertain category problem in e-nose drift compensation.