Joint Distribution Adaptation for Drift Correction in Electronic Nose Type Sensor Arrays

This research deal with the drift compensation problem in sensor arrays named electronic noses. The drift problem occurs in this kind of sensor when they are exposed to an analyte for long periods, which may cause that the response of the sensor varies with time. Some approaches in the literature have tackled the drift compensation problem from the point of view of signal processing algorithms to obtain high rates of accuracy independently of time. In this work, the drift problem is solved using transfer learning with the joint distribution adaptation (JDA) method, which adapts both marginal and conditional distributions between domains, and requires no labeled data in the target domain to perform a classification task with a machine learning algorithm. The developed methodology for drift compensation is validated by measuring accuracy in the classification process. Validation considers a data set that measured six volatile organic compounds during a period of three years under strongly controlled operating conditions using a series of 16 metal oxide gas (MOX) sensors. JDA and Kernel JDA are used with three different types of kernels to determine the best behavior in terms of accuracy to correct the drift in electronic noses. As a result, it can be concluded that the approach using JDA outperforms standard learners like K-Nearest Neighbor (KNN) method.