Seismic waveform classification detection method

Abstract

The application discloses a seismic waveform classification detection method, relates to the technical field of earthquake monitoring, and comprises the following steps: time-frequency feature extraction is performed on waveform data through Hilbert transform and wavelet transform, and latent representation of the waveform data is extracted through a variational autoencoder; a similarity matrix is calculated for waveform data features based on a kernelized cost matrix using a dynamic time warping algorithm; a waveform classifier is constructed based on a convolutional neural network, and a seismic waveform classification result is generated based on a feature fusion matrix; and a time reverse imaging method is used to locate a seismic source. The method can more comprehensively and accurately extract features of seismic waveforms and generate a comprehensive feature matrix by combining various feature extraction technologies and advanced feature fusion methods, a waveform classifier is constructed using a convolutional neural network on the basis, the classification accuracy is further improved, and a time reverse imaging method is introduced to realize accurate positioning of a seismic source.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of seismic monitoring, in particular to a seismic waveform classification detection method. BACKGROUND

[0002] Seismic waveform classification detection technology is one of the key technologies in seismology research and earthquake disaster warning. With the continuous development of seismic monitoring technology, the acquisition of seismic waveform data has become more efficient and accurate. Traditional seismic waveform classification methods mainly rely on human experience, which analyzes the morphological characteristics of seismic waveforms, such as amplitude, frequency, and phase, to determine the type of earthquake. However, this method is not only time-consuming and labor-intensive, but also susceptible to subjective factors, resulting in low accuracy and reliability of the classification results. In recent years, with the rise of machine learning and deep learning technologies, data-driven seismic waveform classification methods have gradually become a research hotspot.

[0003] However, the existing technology still has some shortcomings in processing seismic waveform data. First, seismic waveform data has complex time-frequency characteristics, and traditional feature extraction methods are difficult to fully capture these characteristics, resulting in limited classification accuracy. Second, existing technology in seismic waveform classification relies mainly on single waveform features, lacks comprehensive analysis of multi-source information, and is difficult to accurately classify waveforms. In addition, existing seismic source positioning methods rely mainly on single seismic wave propagation models, which cannot fully consider the propagation delay and uncertainty of waveforms between different receivers, resulting in insufficient accuracy and stability of source positioning. SUMMARY

[0004] In view of the above problems existing in the prior art, the present application is proposed.

[0005] The present application provides a seismic waveform classification detection method, which solves the problem of insufficient feature extraction, low accuracy and robustness of waveform classification, and ineffective positioning of the location of the seismic source in the existing seismic waveform classification method.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a seismic waveform classification detection method, which comprises,

[0008] Obtaining raw waveform data and preprocessing;

[0009] Performing time-frequency feature extraction on the waveform data through Hilbert transform and wavelet transform, and extracting the latent representation of the waveform data through a variational autoencoder to generate a comprehensive feature matrix;

[0010] The radial basis kernel function is used to construct the kernel cost matrix, and the dynamic time warping algorithm is used to calculate the similarity matrix based on the kernel cost matrix, and a feature fusion matrix is generated.

[0011] The waveform classifier is constructed based on the convolutional neural network, and the seismic waveform classification result is generated based on the feature fusion matrix.

[0012] The collected waveform data is back propagated by using the time reverse imaging method, the imaging function is constructed according to the focusing condition of the waveform energy, and the seismic source is located based on the imaging function.

[0013] Based on the seismic waveform classification result and the seismic source positioning result, a seismic waveform analysis report is generated.

[0014] As a preferred scheme of the seismic waveform classification detection method, the original waveform data is obtained and preprocessed, including,

[0015] The original seismic waveform data of k sampling positions is obtained, and denoising, normalization and sliding window segmentation processing are performed.

[0016] As a preferred scheme of the seismic waveform classification detection method, the waveform data is extracted by Hilbert transform and wavelet transform for time-frequency feature extraction, and the latent representation of the waveform data is extracted by a variational autoencoder to generate a comprehensive feature matrix, including,

[0017] The envelope of the waveform is extracted by Hilbert transform based on the preprocessed waveform data;

[0018] The amplitude statistical features including peak amplitude, mean amplitude and change rate are extracted based on the envelope of the waveform;

[0019] The continuous wavelet transform is used to map the waveform data from the time domain signal to the time-frequency domain to generate a two-dimensional time-frequency spectrum;

[0020] Based on the two-dimensional time-frequency spectrum, the energy value of each frequency band is calculated;

[0021] The preprocessed waveform data is input into the encoder of the variational autoencoder, and the encoder structure includes convolutional layers and fully connected layers, the convolutional layers extract local features, and the fully connected layers map the local features to the latent space to obtain the mean and standard deviation of the latent variable;

[0022] The latent variable is generated based on the mean and standard deviation of the latent variable by using a random sampling formula;

[0023] The latent variable is input into the decoder of the variational autoencoder, and the encoder includes fully connected layers and deconvolutional layers, and the reconstructed waveform data is generated through the fully connected layers and the deconvolutional layers.

[0024] The reconstruction error and the regularization error are taken as a loss function, and the parameters of the encoder and the decoder are optimized by gradient descent;

[0025] For a waveform set S = {s1, s2, …, s k}, each waveform data is input into the optimized variational autoencoder to generate a latent variable, and the peak amplitude, mean amplitude, change rate and energy value of each frequency band of each waveform data are calculated, standardized and then the final comprehensive feature matrix F is constructed.

[0026] As a preferred scheme of the seismic waveform classification detection method, wherein: the kernel cost matrix is constructed by using the radial basis kernel function, and the similarity matrix of the waveform data features is calculated based on the kernel cost matrix using the dynamic time warping algorithm, including,

[0027] After preprocessing, the seismic waveform data obtained based on two different sampling positions are represented as time series signals s1 = {s1(1), s1(2), …, s1(N)} and s2 = {s2(1), s2(2), …, s2(M)}, wherein N and M respectively represent the number of sampling points of the two time series signals;

[0028] The radial basis kernel function is used as the kernel to calculate the kernel similarity measure between each point pair in s1 and s2;

[0029] The kernel similarity measure is converted into a kernel cost matrix D;

[0030] The cumulative matching cost matrix C is calculated based on the kernel cost matrix D using the dynamic programming algorithm;

[0031] The final similarity measure is the total distance of the minimum cumulative cost path;

[0032] For a waveform set S = {s1, s2, …, s k}, the similarity measure between each two waveforms is calculated to generate a similarity matrix M.

[0033] As a preferred scheme of the seismic waveform classification detection method, wherein: the feature fusion matrix is generated by splicing the similarity matrix M and the comprehensive feature matrix F.

[0034] As a preferred scheme of the seismic waveform classification detection method, wherein: the waveform classifier is constructed based on the support vector machine, and the seismic waveform classification result is generated based on the feature fusion matrix, including,

[0035] The convolutional neural network is used to construct the waveform classifier, including an input layer, a convolutional layer, a pooling layer and a fully connected layer;

[0036] The waveform classifier is trained using the labeled feature fusion matrix training set, a cross-entropy loss function is selected to calculate the difference between the classification result and the true label, an Adam optimizer is used for gradient descent optimization, the parameters of the classifier are updated, and the iteration is stopped when the loss of the classifier no longer obviously decreases in the continuous iteration process, and the classifier is output.

[0037] The new feature fusion matrix is input into the trained waveform classifier to generate a seismic waveform classification result.

[0038] As a preferred scheme of the seismic waveform classification detection method, wherein: the collected waveform data is reversely propagated by using the time reverse imaging method, the imaging function is constructed according to the focusing condition of the waveform energy, and the seismic source is located based on the imaging function, including,

[0039] After the classification of the seismic waveform is completed, the original waveform data collected at the k sampling positions is denoised;

[0040] The spatial region is divided into grids, each grid point represents a spatial position, the time axis is reversed and discretized into equal interval time steps, and a discretized space-time grid is constructed;

[0041] The finite difference method is used on the discretized space-time grid to solve the wave equation to simulate the propagation process of the wave, the k denoised waveform signals are reversely propagated to the source point, and a three-dimensional velocity model of the region is obtained from a public seismic exploration database to control the propagation speed of the wave;

[0042] After solving the wave equation, the reverse propagation wave field u i (x,y,z,t) is obtained.

[0043] The k reverse propagation wave fields are superimposed in space to form a focused seismic source energy distribution, and the imaging function is constructed according to the focusing condition of the waveform energy;

[0044] Based on the imaging function, the (x,y,z,t) with the maximum imaging function value is selected in the discretized space-time grid as the seismic source point position and the time of earthquake occurrence.

[0045] As a preferred scheme of the seismic waveform classification detection method, wherein: based on the seismic waveform classification result and the seismic source positioning result, a seismic waveform analysis report is generated, including,

[0046] An imaging function distribution map is generated based on the imaging function, and the source point position is marked;

[0047] The arrival time of the waveform is analyzed according to the source point position and the time of earthquake occurrence;

[0048] Based on the seismic waveform classification result, the seismic source point position, the time of occurrence, the arrival time, the imaging function distribution map, and the waveform features recorded at each sampling position, a seismic waveform classification analysis report is generated and sent to relevant personnel.

[0049] In a second aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and wherein the computer program, when executed by the processor, implements any step of the seismic waveform classification detection method according to the first aspect of the present application.

[0050] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any step of the seismic waveform classification detection method according to the first aspect of the present application.

[0051] The present application has the following beneficial effects: the Hilbert transform and the wavelet transform are used to capture the morphological feature information of a single waveform, and the variational autoencoder is used to further extract deep features of the single waveform to obtain local waveform information; the similarity information between different waveforms is integrated by the radial basis function combined with the dynamic time warping algorithm to generate a similarity matrix to obtain global similarity information; the local waveform information and the global similarity information are fused to provide a more comprehensive basis for classification; the convolutional neural network is used to construct a waveform classifier, and the seismic waveforms are accurately classified based on the fused features; and after the classification is completed, the wave propagation process is simulated by the time reverse imaging method to reversely propagate the waveform data and accurately locate the seismic source. BRIEF DESCRIPTION OF DRAWINGS

[0052] To more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort.

[0053] Figure 1 The flowchart of the seismic waveform classification detection method in Example 1.

[0054] Figure 2 The flowchart of the seismic source positioning in Example 1. DETAILED DESCRIPTION

[0055] To make the above-mentioned objects, features and advantages of the present application more apparent and understandable, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0056] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details set forth in this description. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to unnecessarily obscure aspects of the present application.

[0057] Secondly, the "one embodiment" or "embodiment" referred to herein means a specific feature, structure, or characteristic under at least one implementation of the present application. The "in one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments.

[0058] Embodiment 1, reference Figure 1 and Figure 2 The embodiment provides a seismic waveform classification detection method, comprising the following steps:

[0059] S1, obtaining raw waveform data and preprocessing.

[0060] Specifically, raw seismic waveform data of k sampling positions is obtained, and denoising, normalization, and sliding window segmentation processing are performed.

[0061] By obtaining raw seismic waveform data from multiple sampling positions and performing denoising, normalization, and sliding window segmentation processing, the data quality and processing efficiency can be effectively improved. Denoising and normalization processing eliminates environmental noise and unnecessary interference, making the subsequent feature extraction and classification process more accurate. The sliding window segmentation method divides long-time seismic data into multiple short-time periods, facilitating segment-by-segment analysis and ensuring that dynamic changes in seismic waves can be captured in detail. The processed data helps to improve the reliability of subsequent analysis steps and ensures the accuracy of waveform classification and earthquake source positioning results.

[0062] S2, extracting time-frequency features of the waveform data through Hilbert transform and wavelet transform, and extracting latent representation of the waveform data through variational autoencoder to generate a comprehensive feature matrix.

[0063] Specifically, based on the preprocessed waveform data, the envelope line of the waveform is extracted through Hilbert transform, represented as:

[0064]

[0065] Wherein, H(s(t)) represents the result of Hilbert transform, s(t) represents the preprocessed waveform data, and A(t) represents the envelope value of the waveform.

[0066] Based on the envelope line of the waveform, amplitude statistical features including peak amplitude, mean amplitude, and change rate are extracted, represented as:

[0067] A peak = max(A(t)); (2)

[0068]

[0069]

[0070] where A peak represents the peak amplitude, A mean represents the mean amplitude, R A represents the rate of change, and N represents the number of sampling points of the envelope line.

[0071] The waveform data is mapped from the time domain signal to the time-frequency domain using continuous wavelet transform to generate a two-dimensional time-frequency spectrum, denoted as:

[0072]

[0073] where a represents the scale, b represents the time offset, W(a, b) represents the wavelet transform coefficient, and ψ * represents the Morlet mother wavelet, and t represents the length of the time series of the waveform data.

[0074] Based on the two-dimensional time-frequency spectrum, the energy value of each frequency band is calculated, denoted as:

[0075] e i = ∫ T |W(a, b)| 2 db; (6)

[0076] where T represents the time interval, E i represents the wavelet energy of the selected frequency band i.

[0077] The preprocessed waveform data is input into the encoder of the variational autoencoder, and the encoder structure includes convolutional layers and fully connected layers. The convolutional layers extract local features, and the fully connected layers map the local features to the latent space to obtain the mean and standard deviation of the latent variables, denoted as:

[0078] h1 = ρ(W1·s(t) + b1); (7)

[0079] μ,σ = W2·h1 + b2; (8)

[0080] where W1 and W2 are the weight matrices of the convolutional layers and the fully connected layers, respectively, b1 and b2 are the bias terms of the convolutional layers and the fully connected layers, respectively, h1 is the local feature representation output by the convolutional layer, ρ is the ReLU activation function, μ is the mean of the latent variable, and σ is the standard deviation of the latent variable.

[0081] Based on the mean and standard deviation of the latent variable, the latent variable is generated through a random sampling formula, denoted as:

[0082] z = μ + σ · ε; (9)

[0083] where ε is noise sampled from the standard normal distribution, and z is the latent variable;

[0084] The latent variable is input into the decoder of the variational autoencoder, and the encoder includes a fully connected layer and a deconvolutional layer, and the reconstructed waveform data is generated through the fully connected layer and the deconvolutional layer;

[0085] The reconstruction error and the regularization error are taken as the loss function, and the parameters of the encoder and the decoder are optimized through gradient descent;

[0086] For the waveform set S = {s1, s2, …, s k}, each waveform data is input into the optimized variational autoencoder to generate a latent variable, and the peak amplitude, mean amplitude, change rate and energy value of each frequency band of each waveform data are calculated, and after standardization, the final comprehensive feature matrix F is constructed.

[0087] The time-frequency features of the waveform data are extracted by Hilbert transform and wavelet transform, which can provide more abundant waveform information than traditional methods, especially the energy distribution in different frequency bands and time. The envelope feature extracted by Hilbert transform retains the macroscopic morphological information of the waveform, which is indicative for identifying the type of seismic signal. The wavelet transform can effectively process non-stationary signals and extract fine-grained time-frequency features. Further, the latent representation is extracted from the time-frequency features by the variational autoencoder (VAE), which realizes feature dimension reduction and generates a comprehensive feature matrix, realizing efficient and robust feature learning. The self-encoding characteristics of VAE help to extract useful latent information from complex waveform data, while avoiding the limitations of traditional methods relying on manually designed features, thereby improving the accuracy and automation of waveform classification.

[0088] S3, a radial basis kernel function is used to construct a kernel cost matrix, and a dynamic time warping algorithm is used based on the kernel cost matrix to calculate a similarity matrix of the waveform data features, and a feature fusion matrix is generated.

[0089] Specifically, the seismic waveform data obtained based on two different sampling positions are preprocessed and represented as time series signals s1 = {s1(1), s1(2), …, s1(N)} and s2 = {s2(1), s2(2), …, s2(M)}, where N and M represent the number of sampling points of the two time series signals, respectively;

[0090] A radial basis kernel function is used as the kernel to calculate the kernel similarity measure between each pair of points in s1 and s2, denoted as:

[0091]

[0092] where K(s1(i), s2(j)) represents the kernelized similarity measure between two points, σ 2 represents the weighted average of the square of the standard deviation of two waveform signal sampling points;

[0093] The kernelized similarity measure is converted into a kernelized cost matrix D, and the matrix element is represented as:

[0094]

[0095] where i∈[1, N] and j∈[1, M] represent the matrix index, D(i, j) represents the matching degree between two points, and the value range is [0, 1], when the value is 0, it represents that the two points are completely matched, and when the value is 1, it represents that the two points are completely unmatched;

[0096] The cumulative matching cost matrix C is calculated based on the kernelized cost matrix D, represented as:

[0097] C(i, j) = D(i, j) + min{C(i-1, j), C(i, j-1), C(i-1, j-1)};

[0098] where C(i, j) represents the cumulative matching cost of i, j position, which is obtained by recursion, and the initial condition is C(1, 1) = D(1, 1);

[0099] The final similarity measure is the total distance of the minimum cumulative cost path, represented as:

[0100]

[0101] where DTM(s1, s2) represents the similarity measure of s1 and s2 after dynamic time warping, and C(N, M) represents the optimal cumulative cost of the dynamic time warping algorithm in the matching cost matrix from the starting point (1, 1) to the terminal point (N, M);

[0102] For the waveform set S = {s1, s2, …, s k}, the similarity measure between each two waveforms is calculated to generate a similarity matrix M;

[0103] Further, the similarity matrix M is spliced with the comprehensive feature matrix F to generate a feature fusion matrix.

[0104] The radial basis kernel function is used to construct the kernel cost matrix, and the dynamic time warping (DTW) algorithm is used to calculate the similarity matrix between the waveform data features, which can capture the deep similarity between different waveforms. The radial basis kernel function can effectively map the original feature space to a high-dimensional feature space, solving the linear non-separable problem of the original feature space, making the similarity measurement between waveform features more accurate. As a classic time series data similarity measurement method, the DTW algorithm can perform nonlinear alignment on time series, thereby matching waveform features on different time axes and overcoming the defect that the traditional Euclidean distance is sensitive to time shift. The accuracy of similarity judgment is improved. By combining kernel techniques with DTW, the efficiency of the algorithm is maintained while the accuracy of heterogeneous waveform matching is significantly improved.

[0105] The feature fusion matrix contains both local comprehensive features and global similarity information, providing a more comprehensive basis for classification. Through accurate similarity calculation and feature fusion, the robustness of waveform classification is improved, especially when dealing with complex and noisy seismic data, the accuracy of the classification result can be improved.

[0106] S4, a waveform classifier is constructed based on a convolutional neural network, and a seismic waveform classification result is generated based on the feature fusion matrix.

[0107] Specifically, a convolutional neural network is used to construct a waveform classifier, including an input layer, a convolutional layer, a pooling layer, and a fully connected layer.

[0108] The feature fusion matrix training set is used to train the waveform classifier, the cross-entropy loss function is selected to calculate the difference between the classification result and the true label, the Adam optimizer is used for gradient descent optimization, the parameters of the classifier are updated, and the iteration is stopped when the loss of the classifier no longer decreases significantly.

[0109] The new feature fusion matrix is input into the trained waveform classifier to generate a seismic waveform classification result.

[0110] A convolutional neural network (CNN) is used to construct a waveform classifier, and the classifier is trained based on a feature fusion matrix, which can achieve efficient and automated seismic waveform classification. The convolutional neural network can automatically learn the deep representation of waveform features, and through the labeled data of the training set, the classifier can learn the feature patterns of different waveform types. By using the cross-entropy loss function and the Adam optimizer, the classifier can continuously optimize in a large amount of training data, improving the classification accuracy and reducing overfitting. This step can efficiently and automatically classify complex seismic waveforms, and continuously optimize the performance of the classifier to make seismic waveform classification more intelligent and accurate, greatly improving the efficiency and reliability of waveform classification.

[0111] S5, using the time reverse imaging method to propagate the collected waveform data, constructing an imaging function according to the focusing of the waveform energy, and positioning the seismic source based on the imaging function.

[0112] Specifically, after the classification of seismic waveforms is completed, the original waveform data collected at k sampling positions is denoised;

[0113] The spatial region is divided into grids, each grid point representing a spatial position, the time axis is reversed and discretized into equal interval time steps, and a discretized space-time grid is constructed;

[0114] The finite difference method is used to solve the wave equation on the discretized space-time grid to simulate the propagation process of the wave, and the k denoised waveform signals are propagated to the source point, and the propagation velocity of the wave at different depths in the region is obtained from the public seismic exploration database. The wave equation is represented as:

[0115]

[0116] Wherein, u represents the wave field, which represents the wave intensity at the spatial position (x, y, z) and time step t, c represents the wave velocity, represents the spatial second derivative, which represents the spatial propagation characteristics of the wave field;

[0117] After solving the wave equation, the reverse propagation wave field u i (x, y, z, t) is obtained;

[0118] The k reverse propagation wave fields are superimposed in space to form a focused seismic source energy distribution, and an imaging function is constructed according to the focusing of the waveform energy, represented as:

[0119]

[0120] Wherein, I (x, y, z, t) represents the imaging function value, which reflects the focusing degree of the waveform energy at the spatial position (x, y, z) and time step t, T represents the total time step number, u i (x, y, z, t) represents the reverse propagation wave field;

[0121] Based on the imaging function, the (x, y, z, t) with the maximum imaging function value is selected in the discretized space-time grid as the seismic source position and the time of the earthquake.

[0122] The waveform data is reversely propagated by the time reverse imaging method, the imaging function is constructed according to the focusing condition of the waveform energy, and the seismic source is located, the propagation path and the source position of the seismic wave can be accurately inferred, the reverse imaging method can make full use of the collected waveform data, simulate the propagation process of the wave motion, thereby efficiently locating the seismic source, the discrete time-space grid is constructed, and the finite difference method is used to solve the wave motion equation, the most possible seismic source position and the time of occurrence can be determined according to the focused energy of the source point, the reverse propagation of the waveform is accurately modeled through this step, the physical process of the waveform propagation is accurately simulated, high-precision seismic source positioning is realized, the limitation of the single wave propagation model in the traditional method is overcome, and the stability and accuracy of the seismic source positioning are improved.

[0123] S6, generating a seismic waveform analysis report based on the seismic waveform classification result and the seismic source positioning result.

[0124] Specifically, an imaging function distribution map is generated based on the imaging function, and the source point position is marked;

[0125] The arrival time of the waveform is analyzed according to the source point position and the time of occurrence;

[0126] Based on the seismic waveform classification result, the seismic source point position, the time of occurrence, the arrival time, the imaging function distribution map and the waveform features recorded at each sampling position, a seismic waveform classification analysis report is generated and sent to relevant personnel.

[0127] Generating a seismic waveform analysis report based on the seismic waveform classification result and the seismic source positioning result can provide comprehensive and detailed seismic event analysis information for relevant personnel, the report includes key information such as seismic waveform classification, epicenter position, time of occurrence, waveform arrival time and imaging function distribution map, which facilitates seismologists and emergency response personnel to quickly judge the nature of the earthquake and the affected area, this step improves the efficiency of seismic waveform analysis by automatically generating an analysis report, and ensures that decision makers can obtain accurate seismic information in time, which provides comprehensive data support for seismological research and earthquake disaster warning.

[0128] The embodiment also provides a computer device suitable for the case of the seismic waveform classification detection method, which comprises a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the seismic waveform classification detection method proposed in the above embodiment.

[0129] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected by a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is configured to perform wired or wireless communication with an external terminal. The wireless communication can be achieved by WIFI, an operator network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0130] The embodiment also provides a storage medium having a computer program stored thereon, the program being executed by a processor to implement the method for detecting seismic waveform classification according to the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic storage, a flash memory, a magnetic disk or an optical disk.

[0131] To sum up, the present application captures the local and global information of the waveform through Hilbert transform and wavelet transform, further extracts the deep features of the waveform through variational autoencoder, comprehensively combines the similarity information between different waveforms through a radial basis function combined with a dynamic time warping algorithm to generate a similarity matrix, constructs a waveform classifier through a convolutional neural network to quickly and accurately classify the seismic waveforms, simulates the propagation process of the wave through a time reverse imaging method to reversely propagate the waveform data, can accurately locate the seismic source according to the focusing condition of the waveform energy, and generates a seismic waveform analysis report to improve the integrity and systematicness of the seismic waveform analysis.

[0132] Embodiment 2, refer to Table 1, is the second embodiment of the present application, and the experimental simulation data of the seismic waveform classification detection method are given to further verify the technical scheme of the present application.

[0133] To verify the effectiveness of the proposed seismic waveform classification detection method, the research team selected a multi-type seismic monitoring platform for comparative tests, including 328 natural earthquakes: ML 2.0-4.5 events; 152 artificial explosions: 0.5-2.0 kg equivalent explosion experiments; 467 groups of non-seismic signals such as mechanical vibration and vehicle noise. The test used a 32-channel distributed acoustic sensing (DAS) system to collect continuous 30-day seismic monitoring data with a sampling frequency of 500 Hz.

[0134] In the data preprocessing stage, the original waveform is subjected to three-level processing: ① db4 wavelet is used for 5-layer decomposition, and soft threshold method is used to remove power frequency interference; ② 3-second window length (1500 sampling points) is used for sliding segmentation, and the step length is set to 0.5 seconds; ③ each sub-window is normalized by the maximum absolute value, and the comparison method uses the same preprocessing procedure to ensure fairness.

[0135] In the feature extraction link, the following optimization operations are performed: ① Hilbert transform is used to extract the envelope line, and the differential entropy value of the instantaneous frequency is calculated as an additional feature; ② wavelet transform uses a dynamic basis function selection strategy, which automatically matches the optimal wavelet basis for different frequency bands (for natural earthquake low-frequency characteristics 0.1-10 Hz, wavelet transform automatically switches to db12 basis function to improve low-frequency resolution; for artificial explosion high-frequency characteristics 10-100 Hz, cmor3-3 complex wavelet is used to accurately capture transient shock waveform); ③ VAE network is designed as an 8-dimensional latent space, with 4 dimensions allocated for low-frequency band (<5 Hz) and 4 dimensions allocated for medium-high frequency band (5-50 Hz). The encoder uses 3 layers of causal convolution (kernel size 5 / 3 / 3), and the decoder uses an inverse convolution combined with an attention mechanism structure. A spectral similarity constraint term is added during training, and the KL divergence weight is set to 0.01.

[0136] When implementing the dynamic time warping algorithm, an adaptive bandwidth strategy is introduced: the radial basis kernel parameter σ is dynamically adjusted according to the statistical distribution of the distance between feature vectors, and σ is set to 0.5*median_distance. After generating the similarity matrix, it is fused with the VAE feature matrix through the gated attention mechanism, replacing the simple splicing operation.

[0137] The classifier training adopts an improved ResNet-1D architecture, including 5 residual blocks (each block containing 2 convolutional layers) and adding a channel attention module. The optimizer adopts the RAdam algorithm, with an initial learning rate of 0.001, and cooperates with the cosine annealing strategy. The three-dimensional velocity model in the positioning stage is taken from the latest exploration results (grid accuracy of 25m) in the block, and the finite difference solver is realized by CUDA acceleration.

[0138] The comparative test sets two test schemes: the method of the application and the traditional SVM + double difference positioning method

[0139] Table 1 Performance comparison analysis

[0140]

[0141]

[0142] As can be seen from the test data, the method of the application has significant advantages in key performance indicators. The accuracy of 98.7% is improved by 13.4 percentage points compared with the traditional method. The VAE latent space compresses the feature dimension from 256 dimensions to 32 dimensions (reducing by 87.5%), effectively removes noise interference, and breaks through positioning accuracy. The introduction of classification result constraint in time reverse imaging reduces the search space by 72%. The dynamic updating mechanism of the three-dimensional velocity model makes the model error controlled within 1.2%. The GPU accelerated finite difference solver realizes a calculation speed of 0.1ms / step.

[0143] This embodiment verifies the comprehensive breakthrough of the application in multiple technical indicators, especially in the microseismic monitoring scene under complex geological conditions. The comprehensive performance reaches the industry leading level. The test data shows that this method not only solves the core pain points of feature redundancy (dimension reduction of 87.5%) and positioning dependence on artificial experience (error reduction) of the traditional technology, but also creatively realizes the organic integration of physical model and data driven method, providing a new technical paradigm for intelligent seismic monitoring.

[0144] It should be noted that the above embodiments are only used to illustrate the technical solutions of the application and not to limit it. Although the application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the application, which should be covered in the scope of the claims of the application.

Claims

1. A seismic waveform classification and detection method, characterized in that: include, S1. Acquire raw waveform data and perform preprocessing; S2. Time-frequency features are extracted from the waveform data using Hilbert transform and wavelet transform, and the latent representation of the waveform data is extracted using a variational autoencoder to generate a comprehensive feature matrix. include, The waveform envelope is extracted using Hilbert transform based on the preprocessed waveform data; The amplitude statistical features, including peak amplitude, mean amplitude, and rate of change, are extracted based on the waveform envelope. Continuous wavelet transform is used to map waveform data from the time domain signal to the time-frequency domain, generating a two-dimensional time-frequency spectrum. Calculate the energy value of each frequency band based on the two-dimensional time-spectrum diagram; The preprocessed waveform data is input into the encoder of the variational autoencoder. The encoder structure includes convolutional layers and fully connected layers. The convolutional layers extract local features, and the fully connected layers map the local features to the latent space to obtain the mean and standard deviation of the latent variables. Latent variables are generated using a random sampling formula based on the mean and standard deviation of the latent variables; The latent variables are input into the decoder of the variational autoencoder, which contains fully connected layers and deconvolutional layers to generate reconstructed waveform data. The reconstruction error and regularization error are used as loss functions, and the parameters of the encoder and decoder are optimized by gradient descent. For the waveform set S = {s1, s2, ..., s} k Each waveform data is input into the optimized variational autoencoder to generate latent variables, and the peak amplitude, mean amplitude, rate of change, and energy value of each frequency band of each waveform data are calculated. After standardization, the final comprehensive feature matrix F is constructed. S3. A kernelized cost matrix is ​​constructed using a radial basis kernel function. Based on the kernelized cost matrix, a dynamic time warping algorithm is used to calculate the similarity matrix of waveform data features, and a feature fusion matrix is ​​generated. The generation of the feature fusion matrix includes concatenating the similarity matrix and the comprehensive feature matrix to generate the feature fusion matrix; S4. Construct a waveform classifier based on a convolutional neural network, and generate seismic waveform classification results based on the feature fusion matrix; S5. The collected waveform data is backpropagated using the time-reverse imaging method. An imaging function is constructed based on the focusing of the waveform energy, and the earthquake source is located based on the imaging function. S6. Based on the seismic waveform classification results and seismic source location results, generate a seismic waveform analysis report.

2. The seismic waveform classification and detection method as described in claim 1, characterized in that: The process involves acquiring the raw waveform data and performing preprocessing. include, The original seismic waveform data at k sampling locations are acquired and then processed by denoising, normalization, and sliding window segmentation.

3. The seismic waveform classification and detection method as described in claim 2, characterized in that: The process involves constructing a kernelized cost matrix using a radial basis function kernel, and then calculating a similarity matrix for waveform data features using a dynamic time warping algorithm based on the kernelized cost matrix. The earthquake waveform data obtained from two different sampling locations are preprocessed and represented as time series signals s1={s1(1),s1(2),…,s1(N)} and s2={s2(1),s2(2),…,s2(M)}, where N and M represent the number of sampling points of the two time series signals, respectively. Using the radial basis kernel function as the kernel, the kernelized similarity measure between each pair of points in s1 and s2 is calculated; Transform the kernelized similarity measure into a kernelized cost matrix D; The cumulative matching cost matrix C is calculated using a dynamic programming algorithm based on the kernelized cost matrix D. The final similarity metric is the total distance of the path with the minimum cumulative cost; For the waveform set S = {s1, s2, ..., s} k } Calculate the similarity measure between every two waveforms and generate a similarity matrix M.

4. The seismic waveform classification and detection method as described in claim 3, characterized in that: The waveform classifier constructed based on a convolutional neural network, and the seismic waveform classification results generated based on the feature fusion matrix, include: A waveform classifier is built using a convolutional neural network, including an input layer, convolutional layers, pooling layers, and fully connected layers. The waveform classifier is trained using a labeled feature fusion matrix training set. The cross-entropy loss function is selected to calculate the difference between the classification result and the true label. The Adam optimizer is used to optimize the gradient descent and update the parameters of the waveform classifier. If the loss of the waveform classifier no longer decreases significantly during continuous iteration, the iteration is stopped and the waveform classifier is output. Input the new feature fusion matrix into the trained waveform classifier to generate seismic waveform classification results.

5. The seismic waveform classification and detection method as described in claim 4, characterized in that: The method of using time-reverse imaging to backpropagate collected waveform data, constructing an imaging function based on the focusing of waveform energy, and locating the earthquake source based on the imaging function includes: After classifying the seismic waveforms, the raw waveform data collected at k sampling locations are denoised. The spatial region is divided into grids, with each grid point representing a spatial location. The time axis is reversed and discretized into equally spaced time steps to construct a discretized spatiotemporal grid. The wave propagation process is simulated by solving the wave equation using the finite difference method on a discretized spatiotemporal grid. The k denoised waveform signals are backpropagated to the source point, and the three-dimensional velocity model of the area is obtained from a public seismic exploration database to control the wave propagation speed. Solving the wave equation yields the backpropagating wave field u. i (x,y,z,t); The k backpropagating wave fields are superimposed in space to form a focused seismic source energy distribution, and an imaging function is constructed based on the focusing of the waveform energy. Based on the imaging function, the (x,y,z,t) that maximizes the imaging function value is selected as the location of the earthquake source and the time of occurrence in the discretized spatiotemporal grid.

6. The seismic waveform classification and detection method as described in claim 5, characterized in that: The earthquake waveform analysis report generated based on the earthquake waveform classification results and earthquake source location results includes, An imaging function distribution map is generated based on the imaging function, and the location of the earthquake source point is marked. Analyze the arrival time of the waveform based on the location of the earthquake source and the time of occurrence; Based on the seismic waveform classification results, the location of the earthquake source, the time of origin, the time of arrival, the imaging function distribution map, and the waveform characteristics recorded at each sampling location, a seismic waveform classification analysis report is generated and sent to relevant personnel.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the seismic waveform classification and detection method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the seismic waveform classification and detection method according to any one of claims 1 to 6.

Images