A Study Of Feature Extraction For E-Nose System

Electronic nose is an instrument, which mimics human smelling sense. Electronic nose system is generally composed of sensor array, signal preprocessing process and pattern recognition. Feature extraction is a signal preprocessing method for dimensionality- reduction by projecting high-dimensional space to a low dimensional space. Noises could be reduced or eliminated by the projecting of feature extraction which could make subsequent pattern recognition more efficient. As a first step for data analysis, feature extraction methods could conceive most of the samples information for classification in a new 2 or 3 dimensions object space and show the samples distribution directly. Otherwise, a smaller size input space with larger information, obtained by feature extraction, could improve the computation speed. Problems arised from high dimensions and large data size could also be solved. As the advantage of feature extraction techniques, it plays an important role in improving the ability of electronic nose.Characteristics of four feature extraction methods were separately discussed in this paper through a qualitative analysis of four flammable liquids and a quantitative analysis of four kinds of volatile organic compounds in two experiments. The characteristics, dimensionality reduction power and best dimension of object space of the four feature extraction techniques were compared. Four feature extraction methods involve principle component analysis (PCA), Fisher discriminant analysis (FDA), self-organizing mapping (SOM) and Sammon mapping. Three pattern-recognition methods were applied to judge the property of the different feature extraction techniques by correct recognition rate. The k nearest neighbor method (KNN), probabilistic neural net (PNN) and the error back propagation method (BP) were chosen as the pattern recognition methods.In the experiment of qualitative analysis of four flammable liquids, 6 metal oxide sensors (MOS) constituted the sensor array of the electronic nose. The samples were four flammable liquids (gasoline, alcohol, kerosene and diesel oil) and three normal beverages (orange juice, Coca Cola, black tea). The original sample dataset usually with high dimension was first preprocessed by feature extraction. The ability of feature extraction to recognize the different kinds of the samples was judged by pattern recognition techniques. The results showed that flammable liquids and beverages could be 100% classified. Each sample of the flammable liquids and beverages could be recognized by the FDA, utilizing Hierarchical Classification. The best dimension of object space was depended on the feature extraction techniques, while the best pattern recognition technique had big relationship with the data sets.In the experiment of quantitative analysis of four volatile organic components (VOCs), the sensor array was composed of 14 MOS gas sensors. The sample is four VOCs involving benzene, acetone, carbinol and pentane with four concentrations of 100ppm, 200ppm, 300ppm, 400ppm. The results of pattern recognition techniques indicated that all the samples could be correctly recognized, and FDA showed the better work than other methods. Fewer parameters obtained by feature extraction methods could represent characteristics of the concentration and the species character of the sample efficiently.The paper was divided into four parts: the first part described the principle and the structure of the electronic nose system, the various applications of electronic nose. The research background, the purpose of feature extraction and relevant research were also presented in this part. The second part presented in detail the principles of four feature extractions and the algorithms. The third part showed the characteristics and the dimensionality reduction power of four feature extraction methods through a qualitative analysis of VOCs experiment. The fourth part showed the ability of feature extraction methods applied in quantitative analysis. Characteristics and the dimensionality reduction power of four feature extraction methods were compared.