Research Of Seismic Wave Feature Extraction And Recognition Algorithm Of Earthquake And Explosion

The ranges of human activities have being extended to almost every corner of surface in the earth. These activities are continuously intensified, and ever severely deteriorate the qualities of seismic records observed by earthquake observatories, incurring recorded seismic signals may contain a lot of artificial events, such as underground nuclear test and mine blasting. Promptly and accurately recognizing natural earthquake and artificial explosion events is an important task for seismological researches, it is significant to strong earthquake monitoring and small magnitude nuclear test detection.For recognizing earthquake and explosion events, how to acquire efficacious and prominent features is the core problem in the whole recognition process. Comprehensive researches have been carried out, and diverse recognizing criteria(features) have being proposed. However, the best ones of these features, such as:the initial move direction of P-wave, the depth of hypocenter, the ratio of body-wave magnitude scale(mb) and surface-wave magnitude scale(Ms), the existence of characteristic surface wave Lg, etc., are almost event-oriented, not wave recording-oriented, with evident lag-time and poor response time for being acquired only after the confirmation of event's three basic parameters:magnitude, location and time. In addition, it has been revealed that these successful features for large magnitude scale events may severely impair for small magnitude scale events due to inherent deviation, or some of these features may ever not exist.This thesis conducts quite extensive researches about the wave-oriented feature extraction algorithms which being applied in earthquake and explosion recognition, aiming at shortening response time for event type confirmation after the event occurrence and also applicable to small magnitude scale events. This thesis mainly investigates the extraction algorithms of 4 wave features:waveform complexity C, spectral ratio SR, the ratio of integrated SRC, autocorrelation coefficient Rxx. After these 4 features being acquired,3 representative recognition paradigms are investigated:the first is a classical statistical pattern classifier—Fisher's linear classifier, the second is a classifier suitable to non-linear problem—BP neural network classifier, and the last is a classifier suitable to small sample problem—the support vector machine classifier. The 3 classifiers are investigated to verify and evaluate the 4 wave features.Experiments'results show that each of the 4 features:waveform complexity C, spectral ratio SR, the ratio of integrated SRC, autocorrelation coefficient Rxx carries some discriminative information for recognizing earthquake and explosion events, and if being utilized simultaneously, higher correct recognition rate can be achieved:when training sample set for classifier parameters learning and testing sample set for classifier efficacy test are completely independent, correct recognition rate can be larger than 90%.