An Aging Drift Calibration and Device-Generality Network With Realistic Transfer Samples for Electronic Nose

The electronic nose (Enose) system can be applied to many applications. Nevertheless, the gas sensor array used in the Enose system can vary from one device to another. Furthermore, sensors suffer from time-related effects. As time passes, the response from sensors in the Enose system may change, affecting the classification accuracy. Many researchers had put effort into dealing with the sensor's drift problem, but most focused on finding representative transfer samples in the target domain, which is not realistic in real life. In the real situation, the Enose system device can be more calibrated for a period of time. In this study, we propose a new Enose dataset with five different odors. For each odor, four different concentrations are tested with four different humidity levels to increase the dataset's representativeness. Additionally, three devices are tested at three different time intervals. All collected data are recorded chronologically to be used in developing algorithms with reasonable transfer samples. A network is presented herein to solve the drift from one device to another and the drift during sensor aging. The proposed network shows good resistance to the drift between devices and solves the drift as time goes on with a relatively little amount of transfer sample, efficiently reducing the gas experiment cost.