Researches On Algorithms Of Automatic Partition And Feature Extraction Of Valid Seismic Wave-Section

Seismic activity is potentially hazardous,and seismic events are the result of geophysical evolution.At present,a large number of seismic observation stations have been established nationwide to monitor seismic signals.The seismic monitoring network can not only record natural seismic event signal waveforms,but also monitor the signal waveform of artificial blast events.This dissertation focuses on MFCC maps,interception of active sources,and identification of real-time source types.The event datasets used in the study were ground motion waveforms of 176 man-induced blasting events in 184 natural earthquakes and Mentougou areas around the Capital Circle.This paper firstly extracts the mel-frequency cepstrum coefficient(MFCC)map from the seismic signal waveforms,and then uses the convolutional neural network(CNN)to identify the source types of seismic waveform signals—natural seismic events and blasting events.The waveform signal used for extracting the mel-frequency cepstrum coefficient map is the vertical component waveform among the 3 components of the observation station waveform.In the vertical component waveforms of all the observing stations for each event,the waveforms of some stations submerged by noise are removed by the same sliding window,and only the station waveforms that are not submerged by noise are selected,and different events are left.Waves that are submerged by noise range from several to tens of units(each corresponding to a vertical component in a station).Then extract the remaining Melt Frequency cepstrum coefficient plot of the waveform that is not submerged by the noise,and use the Mel Frequency cepstrum coefficient plot as the CNN input.The CNN output is the waveform source type(natural seismic event or blasting event)..If the recognition unit changes to an event,half the waveforms in each event are used for training,and the remaining half of the waveforms are used for testing.In the effective vertical component waveforms of each station of an event,more than half of the waveforms are identified as an event type.The event was classified as the event type and the correct recognition rate was 96.3%.An effective waveform interception algorithm based on window entropy is proposed,which helps to reduce the redundancy of seismic records.This paper presents a new algorithm that uses window entropy to automatically detect the first arrival time of P and S waves.First,the seismic waveforms covering the entire process of events are normalized;then denoising is performed,and the filtered waveforms are sampled by the integer seconds of the seismic signal sampling time,and the energy entropy of each window is calculated;The threshold estimates the first arrival time of the P-wave and S-wave and estimates the ending time of the coda waveform.The waveform records between the arrival time of primary-wave(P-wave)and the vanished time of coda can be considered as an effective seismic wave section of the event;the rest are invalid wave section and should be discarded,Simulate real-time source waveform recognition,use feature window method to locate feature extraction window,normalize short-window waveform,extract feature vector,use support vector machine(SVM)to train and test,compare different short window length waveforms,Finally,taking 1s as the step length,the length of the short window is 5s,and the length of the long window is 15 s.The event is used as the unit to identify and achieve the ideal recognition effect(In all of our experiments,the recognition rate of any of the two events was almost above 90%,and the total recognition rate was 95% or higher).