In the context of the study, the electronic nose and the problems that exist in the odor recognition systems are studied and solutions to those problems with novel hybrid methods are proposed. In this thesis, four different databases are used. While one of them is a database from literature, the other three databases are produced by the electronic nose which is built during the thesis. These four databases consist of different concentration values of n-butanes, 11 different odors, odors of horse mackerel in different days and three different fish (horse mackerel, anchovy and whiting) odors, respectively. A solution to the sensor drift which is the main problem in electronic nose is proposed. The proposed method is to apply the difference signal which is obtained by subtracting the sensor signal of the atmospheric (environment) odor from the sensor signals which are obtained by the applied sample odors. This preprocessing which is different from the literature is decreasing the sensor aliasing for all three databases. In addition, different feature extraction methods are compared. For the sensor data, the sub-sampling method is found to be the method which gets maximum classification performance. For the classification, apart from the literature, a new method which is based on the binary decision tree is proposed. This binary decision tree divides the problem into pieces. In the classification of every piece, the best features are chosen from the extracted features and the best classification method is chosen from the applied Support Vector Machines, k-Nearest Distance, Linear Discriminant Analysis and Bayesian classification methods. The proposed method increases the classification performance when applied to all databases.
Key Words: Signal processing for electronic nose, Feature extraction, Feature selection, Classification, k-NN, LDA, SVM, Bayes classifier, Binary decision tree structure.