Exploring Informative Response Features of Two Temperature Modulated Gas Sensors at a Wide Range of Relative Humidity

Abstract

The response signals of temperature modulated gas sensors contain essential information about measured target gas that must be separated from other correlated, redundant, or noisy data. This issue becomes more critical when variations in environmental factors such as relative humidity of target gas or background odors affect the sensor response. Conductance values of two electronic noses based on a single TGS-2602 and a single FIS SP-53B sensors to four gases and clean air at a wide range of relative humidity levels were measured for analyzing the response features. The role of each feature and increasing the number of features in the accuracy of an SVM classifier are investigated. A method is proposed based on removing non-informative features and compared to four conventional feature selection techniques. It is shown that our proposed scheme with a simple SVM classifier results in 96.7% detection accuracy for TGS-2602 and 98.8% for FIS SP-53B, which is up to the accuracy value of common or advanced methods of selecting features. It is concluded that employing feature selection techniques is more beneficial for the TGS-2602 dataset, which had more destructive features than FIS SP-53B. In conclusion, when working with an E-Nose dataset, it is first necessary to explore the important features to find out whether feature selection is required or not, and if needed, which feature selection method provides the best accuracy.