Electronic nose sensor drift compensation based on deep belief network

Abstract

Gas sensors drift is a serious limit on the appliance and development of electronic noses. To a certain extent, adaptive methods based on machine learning overcome the impact of drift. However, existing machine learning methods conduct drift compensation from the classification level, rather than the analysis and expression of the depth characteristics of gas sensors. For this reason, deep belief network(DBN) is adopted to preprocess gas sensor data, to make the characteristics of sensor data associated and combined with each other. In this way, the coupling between each characteristic of sensor data will be strengthened. Depth characteristics of the data are extracted and expressed effectively, so as to be classified conductively. At last, through a numerical experiment, this method combined with support vectors machine (SVM) was proved to be effective. Meanwhile principal component analysis (PCA) was applied to the depth characteristics, extracted by DBN, to interpret its advantages in improving gas recognition under drift.