Sensor Array Optimization For E-nose System By Feature Selection

Electronic noses based on simulation of human olfactory system have been widely used in many aspects, which comprise of sensor array, DAQ system and pattern recognition. Using feature selection to decide the right kinds and number of gas sensors, sensor array optimization is one of the methods to effectively improve the performance of electronic noses. Generally there are two different kinds of means for feature selection, searching methods and non-searching methods. In this paper, searching methods were used and the algorithm program was based on LabVIEW 7.1.In a flammable liquids detection experiment, the algorithm program was used to optimize an original sensor array consisted of six MOS gas sensors, resulting in a optimized four-sensor array. The performances of four searching strategies were compared, which were genetic algorithm (GA), simulated-Annealing algorithm, generalized sequential forward selection and random searching algorithm. Five kinds of class separative criterion were compared here as well. Results showed that generalized sequential forward selection possessed excellent optimization property and spendt the least time in this experiment, thus having higher efficiency than other three methods. And the criterion J3 based on class distance could estimate the class separability relatively appropriately. Optimized feature set got higher correct recognition rate than original feature space, and there was no evident relationship between the number of remained features and the correct recognition rate of optimized feature sets.Quantitative analysis of four kinds of VOCs with four different was performed to further validate the usefulness of feature selection applied in sensor array optimization. Results showed that the optimized feature set gotten by the searching method had the highest correct recognition rate among the original feature set, the optimized feature set gotten by non-searching method and six random feature sets, indicating that the searching method was more reliable than this non-searching method.The paper contained four parts: the first part introduced the principle and system construction of electronic nose, the significance and means of sensor array optimization, and the structure of searching feature selection methods. The second part described the structure of the feature selection program based on LabVIEW in detail. The third part presented the feature selection results in flammable liquids qualitative detecting, while the final part is the results presentation of quantitative analysis of VOCs.