WO2021027026A1

WO2021027026A1 - Seismic wave vibration nature recognition method based on machine learning

Info

Publication number: WO2021027026A1
Application number: PCT/CN2019/107486
Authority: WO
Inventors: 刘昕靓; 任涛; 王柳婷; 杨丹丹; 商冰冰
Original assignee: 东北大学
Priority date: 2019-08-15
Filing date: 2019-09-24
Publication date: 2021-02-18
Also published as: CN110488351A

Abstract

A seismic wave vibration nature recognition method based on machine learning, wherein the method relates to the technical field of machine learning. The implementation of the method is divided into four stages: seismic waveform processing, feature value extraction, model training, and model application, involving: calculating the distance between a seismic source and a station, screening an epicentral distance, reading three-component seismic waveform data of the screened epicentral distance, using a long time-window algorithm/short time-window algorithm (STA/LTA) and an AIC method to accurately find the location of a first movement of seismic waves, and further intercepting a seismic data length; respectively performing time domain analysis and frequency domain analysis on a seismic waveform; extracting three characteristic waveform complexities, a spectrum ratio, and waveform complexity/spectrum ratio as an input of an artificial neural network model; and training an artificial neural network model with two hidden layers used to identify a two-classification problem, and outputting the probability of a category to which the waveform belongs. The model trained by the method can accurately and efficiently determine a category to which a waveform belongs.

Description

Recognition method of seismic wave vibration properties based on machine learning

Technical field

The present invention relates to the technical field of machine learning, in particular to a method for identifying seismic wave vibration properties based on machine learning.

Background technique

Earthquakes can be divided into two categories: natural earthquakes and non-natural earthquakes according to their vibration properties. Natural earthquakes are squeezing and collision between the earth's plates and plates, causing dislocations and ruptures on the edges and inside of the plates. Unnatural earthquakes, also known as induced earthquakes, refer to abnormal seismic activities in a local area caused by human activities, such as artificial nuclear explosion tests or collapses. Observing the seismic wave signals of natural and non-natural earthquakes recorded at each station of the Seismic Network Center, it is found that the waveforms of the two are very similar. If they are not distinguished and screened, the natural and artificial seismic waveforms will be mixed to study the earthquake. The scholars mislead and influence the research work of seismology.

Machine learning theory and application research began in 1986. With the rapid development of artificial intelligence, it has a wide range of applications in the fields of image recognition, signal processing, predictive evaluation, combinatorial optimization, and knowledge engineering. In recent years, machine learning methods have begun to be used to analyze and process seismic waveform data. How to apply the ever-changing computer frontier technology to the classification of seismic events so as to improve the accuracy and stability of recognition is still one of the topics that need to be studied. . Domestic experts and scholars combined genetic algorithm (GA) with BP neural network when identifying natural earthquakes and unnatural earthquakes, and established a genetic BP network (GA-BP network). The genetic algorithm is used to optimize the neural network globally, and then Use the BP backpropagation algorithm to accurately train the neural network. Dowla et al. used a multilayer perceptron (MLP) neural network to identify natural earthquakes and underground explosions. At the beginning of the 21st century, some scholars proposed to use the empirical mode decomposition method to process the original data waveform, extract the eigenvalues from it, preprocess the original waveform data, and then use the support vector machine (SVM) to complete the seismic waveform classification and recognition task, and make the model Has a higher recognition rate. At the same time, some experts proposed to use wavelet transform method to classify and recognize features such as focal scale and focal depth extracted from natural earthquakes and artificial earthquakes. Bi Mingxia, Huang Hanming and others used Hilbert Huang Transform (HHT) to extract waveform features, and SVM for recognition and classification; Bian Yinju and others used wavelet transform to extract seismic waveform features for research; some foreign scholars tried to use P wave and S wave Identify and classify seismic waves when the signal arrives.

Summary of the invention

technical problem

The solution to the problem

Technical solutions

In view of the problems existing in the prior art, the present invention provides a method for identifying seismic wave vibration properties based on machine learning. The present invention classifies waveforms based on waveform data to identify natural earthquakes and non-natural earthquakes. Among them, the Python programming language is mainly used to process the waveform data, and the machine learning method-artificial neural network is used to classify and recognize seismic waves. The invention adopts the supervised learning method to obtain the seismic property classifier, that is, the artificial neural network model is trained with data of known categories. The input of the model is a feature vector, including features that can reflect the nature of the earthquake. When extracting features, the present invention considers both the time domain and the frequency domain to ensure that comprehensive features of the waveform are obtained.

In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for identifying the nature of seismic wave vibration based on machine learning; including the following steps:

Step 1: Read the original seismic waveform data and determine the epicenter distance of the seismic waveform that needs to be classified and identified;

The specific steps of step 1 are

Step 1.1: Use the Python library obspy in Anaconda for the seismic field to read the original seismic waveform data, and select the waveform according to the spherical distance between the seismic source and the station;

Step 1.2: Use the STA/LTA algorithm to find the interval containing the earthquake starting point, and use the AIC algorithm to accurately locate the time of the first arrival of the P wave;

Step 1.3: After determining the first arrival of the P wave, intercept the seismic waveform to a uniform length to facilitate the extraction of waveform characteristics for subsequent analysis and processing;

Step 2: Perform time-frequency domain analysis on the natural earthquake and unnatural seismic waveforms respectively, the time-domain analysis obtains the waveform complexity characteristics, the frequency domain analysis obtains the spectral ratio characteristics, and the two characteristic values are calculated to obtain the time-frequency domain comprehensive characteristics Value waveform complexity/spectrum ratio;

The waveform complexity and spectral ratio characteristics are respectively the characteristic quantities that can characterize the waveform trend in the time domain and the frequency domain;

Step 3: Use the artificial neural network to train the model. The model input is the feature vector of the waveform instance, that is, the feature vector composed of the three feature values obtained in step 2. The output result is calculated by two hidden layers, and the output result is "1 "Indicates that the category of the waveform is natural earthquake, and the output result of "0" indicates that the category of the waveform is non-natural earthquake;

Step 4: Divide the seismic waveform data into two disjoint data sets, the training set and the test set, use the backpropagation algorithm to update the model parameters, continuously fit the training data, and reduce the value of the loss function; test must be used during model training Collect data to evaluate the effect of the model. After a specified number of iterations of training or the model's test accuracy reaches a certain standard, it means that the model training is completed;

Step 5: Save the model. The output result of the saved model can be directly called for subsequent seismic waveform classification problems, without tedious training.

The beneficial effects of the invention

Beneficial effect

The beneficial effects of using the above technical solutions are: the intensity and impact of non-natural earthquakes are smaller than those of natural earthquakes. If the non-natural earthquakes cannot be identified and screened out in time, people will mistakenly believe that these events are strong earthquakes. The precursors of confuses the earthquake catalogue established based on this record, which will affect the future research work of seismology. Therefore, the classification and identification of natural earthquakes and non-natural earthquakes in seismic signals are helpful for monitoring and early warning of destructive tectonic earthquakes, for small-equivalent nuclear test reconnaissance, for the study of seismology, for protecting human property and safeguarding national interests. World peace has important meaning.

Brief description of the drawings

Description of the drawings

Fig. 1 is a flowchart of waveform complexity feature extraction according to the present invention;

Figure 2 is a flow chart of the spectral ratio feature extraction of the present invention;

Fig. 3 is a structural diagram of a model constructed by an artificial neural network used in the present invention.

Invention embodiment

Embodiments of the invention

The specific embodiments of the present invention will be described in further detail below in conjunction with the drawings and embodiments. The following examples are used to illustrate the present invention, but not to limit the scope of the present invention.

The software environment of this embodiment is the WINDOWS 7 system, and the integrated development environment selects Pycharm IDE.

The identification method of seismic wave vibration properties based on machine learning includes the following steps:

The epicenter distance is the spherical distance between the station that recorded the waveform and the epicenter;

The specific steps of step 1 are

Step 1.1: Use the Python library obspy in Anaconda for the seismic field to read the original seismic waveform data, and select the local seismic waveform according to the spherical distance between the source and the station;

Since the energy of the seismic wave signal will be lost during the propagation process, the greater the epicentral distance, the greater the noise doped in the recorded waveform, and the greater the seismic energy loss it represents, that is, the low signal-to-noise ratio, which is not conducive to analysis Waveform data, so this method selects the data within 500km of the epicenter for classification and identification;

There is a lot of noise in the waveform data used in this embodiment, so it is necessary to identify the starting point of the initial motion of the seismic wave, and to facilitate subsequent processing, choose to intercept fixed-length data that can contain complete waveform information;

The seismic wave transmission file ".mseed" is read through the obspy library and used to find the starting point by combining STA/LTA and AIC; the data is intercepted to a uniform length to facilitate the extraction of waveform features. This embodiment determines that the seismic wave determines the P wave first arrival After the earthquake's starting point, the data of 3 seconds before the starting point and the data of 160 seconds after the starting point are intercepted, a total of 163s of data are processed for subsequent analysis;

The waveform complexity and spectral ratio characteristics are the characteristic quantities that can characterize the waveform trend at the time domain level and the frequency domain level, respectively, and the extraction process is shown in Figure 1 and Figure 2;

Step 3: Use the artificial neural network to train the model. As shown in Figure 3, the model input is the feature vector of the waveform instance, that is, the feature vector composed of the three feature values obtained in step 2, and the output result is calculated through two hidden layers , The output result of "1" indicates that the category of the waveform is natural earthquake, and the output result of "0" indicates that the category of the waveform is non-natural earthquake;

Input the three components of the feature vector into the artificial neural network model to train and save the model. Input the feature vector of a single waveform into the model to obtain the classification result; or input all the feature vectors of the waveform of an event into the model, and combine the classification results obtained from multiple seismic waveform data in one event to determine the vibration properties of the current event ( Natural seismic event or non-natural seismic event).

The model structure and use parameter information of the artificial neural network in the step 3 are shown in Table 1:

Table 1 Model structure and usage parameters of artificial neural network

[Table 1]

隐藏层数Hidden layers	22
隐藏层神经元个数Number of hidden layer neurons	66
防止过拟合Prevent overfitting	L2正则化L2 regularization
学习率Learning rate	指数衰减算法更新Exponential decay algorithm update
损失函数Loss function	二元交叉熵损失函数Binary cross entropy loss function
训练迭代次数Number of training iterations	500,000500,000

The setting standard for the number of training iterations in this embodiment is 500,000;

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that; The technical solutions recorded in the foregoing embodiments are modified, or some or all of the technical features are equivalently replaced; therefore, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

A method for identifying seismic properties of seismic waves based on machine learning, which is characterized in that it includes the following steps:

Step 1: Read the original seismic waveform data and determine the epicenter distance of the seismic waveform that needs to be classified and identified;

Step 2: Perform time-frequency domain analysis on the natural earthquake and unnatural seismic waveforms respectively, the time-domain analysis obtains the waveform complexity characteristics, the frequency domain analysis obtains the spectral ratio characteristics, and the two characteristic values are calculated to obtain the time-frequency domain comprehensive characteristics Value waveform complexity/spectrum ratio;

The waveform complexity and spectral ratio characteristics are respectively the characteristic quantities that can characterize the waveform trend in the time domain and the frequency domain;

Step 3: Use artificial neural network to train the model. The input of the model is the eigenvector of the seismic waveform instance, that is, the eigenvector composed of the three eigenvalues obtained in step 2. The output result is calculated by two hidden layers, and the output result is " 1" indicates that the category of the waveform is natural earthquake, and the output result of "0" indicates that the category of the waveform is non-natural earthquake;

Step 4: Divide the seismic waveform data into two disjoint data sets, the training set and the test set, use the backpropagation algorithm to update the model parameters, continuously fit the training data, and reduce the value of the loss function; test must be used during model training Collect data to evaluate the effect of the model. After a specified number of iterations of training or the model's test accuracy reaches a certain standard, it means that the model training is completed;

Step 5: Save the model. For subsequent seismic waveform classification problems, you can directly call the saved model to output the results, eliminating the need for tedious training.
The method for identifying seismic wave vibration properties based on machine learning according to claim 1, characterized in that: the specific steps of step 1 are:

Step 1.1: Use the Python library obspy in Anaconda for the seismic field to read the original seismic waveform data, and select the seismic waveform according to the spherical distance between the seismic source and the station;

Step 1.2: Use the STA/LTA algorithm to find the interval containing the earthquake starting point, and use the AIC algorithm to accurately locate the time of the first arrival of the seismic wave P;

Step 1.3: After determining the first arrival of the P wave, intercept the seismic waveform to a uniform length to facilitate the extraction of waveform characteristics for subsequent analysis and processing.