Study On Method Of Signal Gathering And Processing Based On Bionic Olfaction

The research of bionic olfaction is slowly advancing for a long time. One main reason is that the hypostasis of the bio-olfaction system is more complicated than the other sense organs. The other is that the signal from organs of smell is not single and their information processing is comparatively complicated. At present, many researchers have investigated the bionic olfaction system based on gas sensor array and pattern recognition, which is becoming an important way in gas or odor analysis. Signal gathering and pattern recognition play the crucial role to the behavior of the bionic olfaction system.The thesis developed the application of the bionic olfaction system on the basis of the principle of biology olfaction system and the correlative results at home and abroad. The main contribution in this thesis can be organized as follows: The whole response of sensor arrays has been applied to identify different odor and solve the problem of low selectivity with single sensor; Based on the review of the existing bionic olfaction signal processing, the study is carried out with two aspects: support vector machines (SVMs) (Qualitative work) and artificial neural networks (ANNs) (Quantitative work); It identified the fish species (classification) and their freshness applying the proposed methods.The support vector machine and artificial neural network provide a powerful tool for us to deal with the measurement photo and data from four fishes. Four most popular fishes in New Zealand market including: Red Snapper, Gurnard, Tarakihi and Trevally have been selected in our cooperative project.Support vector machines (SVMs) are a new-generation machine learning, it has some advantages: (1) SVMs implement the structural risk minimization principle which minimizes an upper bound of the generalization error rather than minimizes the training linearly error and avoid overfitting; (2) Training SVMs is equivalent to solving a linearly constrained quadratic programming, and the solution of SVMs is unique, optimal and global; (3)Using kernel function, SVMs transfer the linearly non-separable samples in the input space to linearly separable samples in the feature space.In order to reduce SVM or ANN complexity and improve the performance of data classification, we used the most efficient features extracted from the image or smell print data as well as pre-processing technique by way of a down-sampled version of the data. The results indicated that we can get the following good performance: the configuration of BP neural network obtained the simplification, the network convergence rate is quicker, moreover, the recognition accuracy is higher.The significance of this work is the extended application and good recognition accuracy by taking the advantages between the SVM operating on the principle of structure risk minimization and the ANN implementing the empirical minimization principle. The experiment results show that the proposed pattern recognition technology is capable of on-line applications.