I. Introduction
An electronic nose (e-nose) is an intelligent sensing system, which consists of gas sensors and pattern recognition software. When some unknown gas enters into the chamber, gas sensors produce transient response signals and then pattern recognition techniques are applied to identify the gas and estimate its concentration [1], [2]. Ideally, these sensors should indicate identical or similar patterns when exposed to the same gas with the same concentration. However, this is not always the case in practice. For example, these sensors may be poisoned due to exposure to aggressive chemicals, or aged by being used for a long period of time, which are called sensor drift [3]. Formally, sensor drift is defined as the gradual changes of transient response signals, caused by dynamic environment conditions or sensed characteristic modifications over time [4]. Drift can result in serious degradation of the performance on succussive test data by using previously trained classifier/regressor on original labeled data, and so causes e-noses to be unreliable in many real world applications [2], [5]. In sensor research and pattern recognition communities, it is a challenging issue to find robust and adaptive solutions for alleviating the sensor drift problem.