Research On Curve-fitting Based Breath Odor Feature Extraction And Analysis

As an important branch of pattern recognition, biometrics has brought far-reaching influence to all aspects of people's life. Besides identity authentication such as fingerprint recognition, biometrics can also be applied in medical diagnosis, especially in the modernization of TCM(Traditional Chinese Medicine). As the development of sensors-based e-nose technology, breath diagnosis, i.e. diagnosis based on the breath of people, is developing quickly and steadily. This convenient, trusty and non-invasive diagnosis method has been the development tide of the international medical diagnosis. However, as a new diagnosis technology, breath diagnosis has many problems to solve, such as improvement of e-nose system, research on the algorithms of feature extraction and classification of odor data.The work of this thesis can be divided into three parts: design and implement of the e-nose system, research on the algorithm of odor data feature extraction, and the classification of the odor data. The design of the e-nose system includes hardware and software. Hardware design includes the optimization of sensor array, the adjustment, collecting, and communication of sensor signals, and building of the sampling system. Software design is coding of the sampling system.The research on the algorithm of odor data feature extraction is the core part of this thesis. Odor data is the electric signals from the output of the sensor array. The traditional feature extraction is to extract the static geometrical features, such as peak value of the curve, position of the peak value, slopes, and the area under the curve, etc. This method has many disadvantages. Because the response of the sensor is not just depend on the types of odors, but also affected by the temperature, humidity and concentration of the odor. The static geometrical features cannot always represent the types of odors. This thesis introduces the curve-fitting technique to extract features. Based on the analysis of many models, a new and high-precision curve-fitting model is proposed. The parameters of this model are extracted as features. These features can represent the class-related information of odors roundly, which make good inputs for classification.Our e-nose system has been used to collect odor samples and now we have over two hundreds healthy samples and nearly one hundred diabetes samples. After feature extraction, a series of classification experiments are carried on those samples, including clustering, principal component analysis(PCA), linear discriminant analysis(LDA), BP neural network(BPNN) and support vector machine(SVM). Experiments show that the feature extraction method based on curve-fitting is effective and the recognition rate using the features extracted by curve-fitting are obviously better than that using the static geometrical features.