Drift Compensation for an Electronic Nose by Adaptive Subspace Learning

An electronic nose (EN) is a bionic system that relies on an array of gas sensors for effective odor recognition. Since the gas-sensor drift would depress the EN performance, we proposed an adaptive Domain based subspace learning method considering both Maximizing label feature Dependency and Minimizing feature Redundancy (DMDMR) to address the EN based drift issue. Considering the inconsistent data distribution caused by drift, the proposed method learns a time-varying common subspace with similar distribution for both regular and recent drifted EN responses. In order to preserve useful classification and topological structure information simultaneously, the Hilbert-Schmidt independence criterion (HSIC) has been adopted to measure the dependence between features and labels while the feature redundancy is reduced according to the PCA criterion. An EN drift dataset acquired from a 16-gas-sensor EN system over 36 months was adopted in the experiments. To verify the effectiveness of the proposed method, we used the variations of both first and second order statistics to measure the movement of data distribution under drift condition. The results show that the data distribution in the subspace learned by DMDMR has stronger stability than the one in original space. Furthermore, the positive effects of DMDMR have been exhibited versus other state-of-the-art methods in recognition. The performance of DMDMR paradigm has demonstrated obvious superiority to other paradigms. These results prove that the proposed subspace learning approach is suitable for EN based drift counteraction and can be successfully implemented.