Measurement ( IF 3.364 ) Pub Date : 2020-11-20 , DOI: 10.1016/j.measurement.2020.108748 Tao Liu; Dongqi Li; Jianjun Chen
Gas-sensor drift is an important issue fading the gas identification performance of an Electronic Nose (E-nose). The most popular drift countermeasure is mathematical-model updating by periodic drift calibrations. Accordingly, drift-calibration-sample formation becomes a challenging issue in online detection because of rare opportunity for collecting drift-calibration samples. However, the class-imbalance problem may occur during such formation. Thus, we proposed an active drift-calibration-sample selection method including a new metric “classifier state” and an associated sample-evaluating procedure. To assess the proposed method, two benchmarks have been adopted. One is a public dataset while the other one is collected from our own E-nose system. Experimental results have demonstrated the superiority of the proposed method in several drift scenarios. Further, we visually explored the behind reason of the proposed method’s high performance. Additionally, a parameter sensitivity analysis was conducted. We conclude that the proposed methodology reduces the negative effect of class-imbalance problem successfully.