Optimization of sensor array in electronic nose by combinational feature selection method

Abstract

Electronic nose (e-nose) is a machine olfaction system and the sensor array is an essential part of the electronic olfaction process. A pattern recognition unit is necessary in electronic nose system to efficiently decide about the output of the test using the responses of all the sensors in the array. The output of a pattern recognition algorithm depends on the quality of the feature set used for training and testing. Relevant and independent feature set improves the performance of a pattern classification algorithm. In some applications of electronic nose, the responses of few sensors are highly corrupted with noise and are either irrelevant or are redundant to the process. These sensors should be identified and eliminated from the sensor system for better accuracy. This paper addresses the selection of sensors in an e-nose system by different feature selection methods and then integrates them to achieve improved classification performance. We have used three types of feature selection methods namely, t-statistics, Fisher's criterion and minimum redundancy maximum relevance (MRMR) technique to select the most informative features. We have tested the proposed method on data obtained from the major aroma producing chemicals of black tea. Multi-class support vector machine (SVM) has been used as a pattern classifier in an electronic nose with black tea samples. The experimental results show that the performance of the e-nose system increased by 6-10% with the use of the proposed combinational feature selection technique.