Research On Seismic Event Source Type Classification Based On Time-Frequency Analysis And Convolutional Neural Network

In recent years,with the continuous deployment of international seismic stations and China's national seismic stations,and the sustained improvement of the accuracy of the seismic station instruments and the ability to collect and record seismic signals,seismological research has entered the era of seismic signal big data,under this background,there are a series of problems to be solved especially in seismic event source type classification,the accurate classification of seismic event source type is of great significance for the establishment of an accurate and clear catalog of seismic events and strong earthquake early warning,and further research in the field of seismology.Time-Frequency Analysis has excellent ability to analyze nonlinear and non-stationary seismic signals,it's can extract the time-frequency information of seismic signals,due to the successful application of deep learning,especially Convolutional Neural Network,in many fields involving identification and classification problems,how to apply deep learning to the problem of seismic event source type classification has also attracted more and more extensive attention,and relevant research has made some progress.In this thesis,the several different TimeFrequency Analysis methods and Convolutional Neural Network are used to classify the seismic event source types of natural earthquakes and artificial explosions,the main works are as follows.(1)In this thesis,a total of five Time-Frequency Analysis methods in which three with basis functions: Short-Time Fourier Transform,Wavelet Transform and Stockwell Transform and the other two without basis functions: Empirical Mode DecompositionHilbert Transform(EMD-HT)and Ensemble Empirical Mode Decomposition-Hilbert Transform(EEMD-HT)are employed to extracted features of seismic signals originating from the two kinds of seismic event source type: natural earthquake and artificial explosion.The original one-dimensional time-domain waveform signal is converted into a two-dimensional time-frequency domain image(time-frequency domain feature),it's been studied in detail the time-frequency domain characteristics of waveform data of natural earthquake and artificial explosion.(2)The Convolutional Neural Network model was constructed to classify the seismic event source types of the natural earthquake and artificial explosion in the 2D time-frequency domain of the extracted waveform,the time-frequency domain features extracted by the five time-frequency analysis methods were classified by using the Convolutional Neural Network,the classification results of the three time-frequency analysis methods with basis functions,two time-frequency analysis methods without basis functions total five different Time-Frequency Analysis methods are analyzed,respectively.(3)To verify the effectiveness and superiority of the method in this thesis,the following methods are contrastively studied:(1)five Time-Frequency Analysis methods plus Convolutional Neural Network;(2)the time domain waveform data plus Convolutional Neural Network;(3)five Time-Frequency Analysis methods plus Support Vector Machine;(4)five Time-Frequency Analysis methods plus Multilayer Perceptron,their highest classification accuracy was:(1)98.43%,97.01%,97.49%,96.54%,94.97%;(2)90.58%;(3)94.03%,91.21%,93.41%,88.85%,90.27%;(4)94.19%,93.56%,91.21%,89.33%,90.11%,respectively,and analyzed their experimental results,the experimental results proved the combination of TimeFrequency Analysis and Convolutional Neural Network has an excellent classification effect on seismic event source types classification,the highest classification accuracy can reach 98.34% in the framework of Short-Time Fourier Transform plus Convolutional Neural Network.