| With the development of society,various types of engineering blasting operations have increased significantly.If natural seismic waveforms and artificial blasting waveforms cannot be distinguished in a timely manner,the earthquake catalog will be seriously confused,which will affect the study of seismology.Seismic waveforms are non-linear and non-stationary signals.When the Hilbert transform analyzes seismic waveform signals,the instantaneous frequency will show a negative value without any physical meaning.Therefore,in 1998,the National Academy of Engineering N.E Huang et al.proposed an empirical mode decomposition algorithm,which is called the Hilbert-Huang algorithm.This method completely removes the constraints of linearity and stationarity;it does not need to design a basis function,and can adaptively generate a "base" according to the characteristics of the signal itself,which is its most prominent feature-adaptability;this method is no longer subject to Heisenberg The limitation of the uncertainty principle can achieve relatively high accuracy in both time resolution and frequency resolution.Therefore,the Hilbert-Huang transform method is more suitable for analyzing nonlinear and non-stationary signals,and applying it to the classification research of natural seismic and artificial blast waveform data,which can make the signal decomposition and extraction of signal time and frequency domains possible.Time-frequency spectrum characteristics and other aspects have been broken to varying degrees.In this paper,HHT is used to extract the eigenvalue data set of seismic waveform and identify the source type of the event.The main research contents and development work are as follows:(1)EMD is used to decompose the original waveform signal into a limited number of IMF and res.functions,and 26 time-domain statistical features are extracted from the original waveform,the first 7 IMF and the res.function to form 9 feature groups(Named Q0,Q1,...,Q8);Then calculate the amplitude energy ratio of the first 7 IMF to obtain 7 energy ratio features,and select 8 of the 26 time-domain statistical features of the first 4 IMF to form a total of 32 features.Feature group(named Q9).Single-group and multi-group feature combination event type recognition experiments were carried out on these 10 feature sample sets.(2)The original experimental waveform was intercepted by 1000 sampling points forwardand backward at the maximum amplitude to obtain waveform data of 2000 sampling points,which were subjected to HHT to obtain three instantaneous properties(instantaneous frequency,instantaneous Amplitude,instantaneous phase),and then get the Hilbert spectrum(time-frequency-amplitude combination).In this paper,the five time-frequency features of mean,variance,peak,kurtosis,and Shannon entropy of the first three internal model functions after transformation are extracted to form a time-frequency feature group.(3)Use the symmetric KL distance pair to extract the time-domain feature group and time-frequency feature group,Randomly select the part(30%,50%,70%,or 90%)of the events.The features of the corresponding feature group of the 3-component waveforms of all observation stations are used as the training set and also as the test set,the event is used as the recognition unit for classification and recognition,and the experiment is repeated 1000 times.The results show that in the extracted time domain features,the correct recognition rate of the time domain statistical features extracted by 2nd IMF when selecting 90% events is greater than 90%;The experimental results of the time-frequency feature group show that the recognition rate of randomly selected parts(30%,50%,70% or 90%)of the events has reached above 95%.This shows that both the time-domain features extracted by the internal model function and the time-frequency features extracted by the transformed internal model function have good discrimination and recognition capabilities,which can provide more effective features for waveform recognition event source classes. |
PDF Full Text Request