A Research On Algorithms For Discriminating Earthquake And Explosion Based Upon HMM And GMM

Earthquake can be regarded as the assemblage of sequential earth medium quake processes due to seismic waves spread outward from the hypocenter. Being observed earthquakes include natural events and explosion events. Explosion is originated by human activities; it may be industrial explosion, mine tremor and nuclear explosion etc. The occurring of explosion is more and more regular. If not being properly treated, explosion event would be frequently erroneously considered as natural earthquake event. Recognizing observed event by seismic waves is the main means of nuclear explosion reconnaissance.Seismic signal is a non-steady non-linear time-variant signal. This thesis extends the feature extraction algorithms and recognition paradigms that having been successfully applied in short-time steady time-variant signal processing (such as speech processing), attempting to recognize earthquake and explosion more robust. Three kinds of features are extracted from seismic waves:Mel-Frequency Cepstrum Coefficient (MFCC) features, Linear Prediction Cepstrum Coefficient (LPCC) features and Hilbert Huang Transform (HHT)-based features. This paper introduces the basic theory and the extractive process of MFCC and LPCC; the paper also details the basic theory of HHT and its applications in seismic signal, and then describes how extract HHT-based features.Two recognition paradigms are utilized in the paper:Hidden Markov Model (HMM) and Gaussian Mixture Model (GMM). HMM is a statistical model based on Markov chain, being consisted of two stages:HMM training and HMM recognition. HMM is constituted by three basic algorithms:the forward-backward algorithm, Viterbi algorithm and Baum-Welch algorithm. GMM is a continuous HMM with only one stateUsing software of Matlabe, this thesis investigates the impact of MFCC's or LPCC's feature number on HMM's correct recognition rate, and the impact of signal sampling valid length on HMM's correct recognition rate. In addition, the impact of model order of GMM on its effect, the impact of MFCC's or LPCC's feature number on GMM's correct recognition rate, and the impact of signal sampling valid length on GMM's correct recognition rateThe results show that:for HMM recognition paradigms and these three kinds of features, MFCC features and LPCC features are better than HHT-based features; for GMM recognition paradigms and these three kinds of features, get the same conclusion:MFCC and LPCC are better than HHT. When comparing the two recognition paradigms, the correct recognition rates of GMM is better than HMM, moreover, the training time of GMM is at least 60% less than that of HMM.