Research Of Features Extraction And Recognition Algorithm Of Earthquake And Explosion

The wave incurred by explosion can like as that of earthquake spread in the interior and surface of the earth and be recorded by the digital seismic network. At present, the main algorithms to extract discriminative features from wave signals for classifying explosion and earthquake events are Fourier spectral analysis and wavelet transform. The newly near-mature signal processing technique -- Hilbert-Huang Transform (HHT) can also be used for extracting discriminative features from seismic wave signal, which found in some literatures. But rare literature found for classifying explosion and earthquake events by features based on HHT. This thesis mainly discusses the investigation of HHT algorithm and its applications in seismic signal classification. And wavelet packet transform algorithm and its applications a plus.The main contents are listed in three parts as follows:1. Time-frequency spectrum analysis of signals: A small magnitude scale earthquake and an equivalent explosion event are selected, and their wave signals are processed by HHT and wavelet packet transform. By analyzing and comparing HHT time-frequency-energy distribution and wavelet-packet transform energy change with different time and decomposition scale of the 2 types of seismic signals, the possible physical characteristics for classifying earthquake and explosion event extracted by the Hilbert spectrum and wavelet packet transform are explored.2. Parametric feature extraction and feature space formulation: After using HHT we can have finite Intrinsic Mode Function (IMF) components, every IMF component contains the signal's information is not superposition. IMF reflects the signal's instantaneous characteristics, this paper presents an approach ,that is first extracted two parameter features are the peak frequency and average energy from a single IMF component's Hilbert spectrum, and then use different IMF component's the two parameters characteristics to construct different feature spaces. As I know this approach does not appeared in the literatures. The signal after using multi-layer wavelet packet transform, the wavelet coefficients in the last layer include the high and low frequency component information of signal, so we can compute parameter features like coefficients' logarithmic energy entropy, Shannon entropy, and the energy ratio, and then use these parameters separately construct difference feature space.3. Pattern recognition system construction: Randomly selected different characteristics samples for training set, and then with different types of support vector machine using the selected kernel function to design a pattern classifier, finally selected samples for test set which different from raining set, using this test set to test the classification performance of the pattern classifier.The results show that ,for the features which extracted from after HHT, the feature space constructed by the first five IMF components, first use polynomial kernel function to mapping space and then use the v parameter support vector machine(v-SVC) to distinguish the two signals can have a good result; for the features which extracted from after wavelet packet transform, the feature space constructed by the last wavelet coefficients' logarithmic energy entropy first use radial basis kernel function to mapping space and then use the v parameter support vector machine(v-SVC) to distinguish the two signals can have a good result. Integrated use of the two characteristics which have obvious distinguish information can available to better recognition.