A Deep Learning-Based Seismic Wave Identification Method

Abstract

A deep learning-based seismic wave identification method includes the following steps: First, acquiring seismic wave signals to obtain raw data, which is then processed by centering and normalization to obtain a unified signal. Second, decomposing the unified signal using CEEMDAN to obtain several orders of intrinsic mode functions and residual components. Each order of mode function is coarsely processed according to a preset time scale parameter to divide it into subsequences. A similarity criterion is then constructed using a fuzzy membership function, and the fuzzy entropy values ​​of each subsequence are calculated and integrated into a feature matrix. Third, probabilistic principal component analysis is used to reduce the dimensionality of the feature matrix, resulting in a dimensionality-reduced feature matrix. Fourth, the dimensionality-reduced feature matrix is ​​input into a CTCM-1D-CNN model optimized by a tribal competition and member cooperation algorithm to predict the output seismic wave type, thus completing the seismic wave identification. Therefore, this design demonstrates good accuracy and stability in seismic wave identification.

Description

Technical Field

[0001] This invention relates to a seismic wave identification method, belonging to the field of earthquake monitoring, and particularly to a seismic wave identification method based on deep learning. Background Technology

[0002] Natural earthquakes are triggered by tectonic movements within the Earth, and their early warning and rapid reporting directly impact regional disaster prevention and mitigation decisions. Artificial blasting, on the other hand, generates vibrations from mining, infrastructure construction, and other engineering activities. Accurately distinguishing between the two is crucial for ensuring the accuracy of earthquake monitoring and supporting disaster prevention, mitigation, and engineering safety management. Currently, the distinction between natural and non-natural earthquakes relies primarily on signal processing to extract seismic wave features, followed by classification using traditional deep learning models. However, existing technologies, employing single signal processing methods (such as wavelet transform and Hilbert-Huang transform), struggle to fully capture the irregularities of seismic waves across multiple frequency scales, easily leading to feature redundancy or loss of detail. Furthermore, the hyperparameters of traditional deep learning models often rely on pre-set settings based on human experience, and their values ​​cannot be dynamically adjusted according to the actual scenario. This results in fluctuations in learning efficiency and fitting performance, leading to poor accuracy and stability in seismic wave identification.

[0003] Chinese patent application No. 202510371180.2, filed on March 26, 2025, discloses a deep learning-based method for classifying microseismic events. The method includes: acquiring multiple historical microseismic signals and corresponding labeled categories of microseismic events; filtering each historical microseismic signal and then introducing empirical mode decomposition to obtain the microseismic time-domain signal; converting the microseismic time-domain signal into an equivalent two-dimensional spectrogram; and extracting feature vectors V representing the low-frequency, high-frequency, and edge textures of the two-dimensional spectrogram using a deep learning framework. c The image feature descriptor V, representing the shape and texture information of the two-dimensional spectrogram, is extracted using the directional gradient histogram. h V c With V h The high-dimensional feature vectors obtained by splicing or weighted fusion, along with category labels, are used to train a classifier to obtain a classification model for microseismic events. This model is then used to classify and discriminate the microseismic signals to be processed. Although this patent uses an image feature-based classification model to achieve real-time microseismic event classification, it still has the following drawbacks:

[0004] This design suffers from poor accuracy and stability in seismic wave identification due to the inaccuracy of denoising caused by empirical mode decomposition, low efficiency of hyperparameter optimization by grid search, and reliance on experience for feature fusion.

[0005] The information disclosed in this background section is intended only to enhance the understanding of the overall background of this patent application and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings and problems of poor accuracy and stability in seismic wave identification in the existing technology, and to provide a deep learning-based seismic wave identification method with better accuracy and stability.

[0007] To achieve the above objectives, the technical solution of the present invention is: a seismic wave identification method based on deep learning, the identification method comprising the following steps:

[0008] Step 1: First, collect the seismic wave signals generated by the earthquake event to obtain raw data. Then, extract the center of the raw data and normalize the extracted raw data to obtain a unified signal.

[0009] Step 2: The unified signal is decomposed using the CEEMDAN decomposition method to obtain several orders of intrinsic mode functions and corresponding residual components. For each order of intrinsic mode function, coarse-grained processing is first performed based on preset time scale parameters to divide each order of intrinsic mode function into several sub-sequences according to the corresponding time scale. Then, a similarity metric is constructed for the sub-sequences under each time scale using a fuzzy membership function. Based on the similarity metric, the fuzzy entropy value of the sub-sequences under each time scale is calculated. Finally, the fuzzy entropy values ​​are integrated to form a feature matrix.

[0010] Step 3: Use probabilistic principal component analysis to reduce the dimensionality of the above feature matrix to obtain the dimensionality-reduced feature matrix;

[0011] Step 4: Input the dimensionality-reduced feature matrix into the CTCM-1D-CNN convolutional neural network model. The CTCM-1D-CNN convolutional neural network model is constructed by optimizing a one-dimensional convolutional neural network using a tribal competition and member cooperation algorithm. The CTCM-1D-CNN convolutional neural network model performs earthquake event type identification and prediction on the dimensionality-reduced feature matrix and outputs the type of the acquired seismic waves, thereby completing the identification of seismic waves.

[0012] In the first step, acquiring the seismic wave signals generated by the earthquake event to obtain the raw data means: acquiring the acceleration data generated by the earthquake event through earthquake monitoring equipment, and using the acquired acceleration data as the raw data.

[0013] In the first step, the central truncation of the original data means: taking the peak point of the acceleration data waveform in the original data as the center, symmetrically truncating the same number of acceleration data segments to the left and right of the peak point.

[0014] In the second step, the step of using the CEEMDAN decomposition method to decompose the above-mentioned unified signal to obtain several orders of intrinsic mode functions and corresponding residual components means: first, determining the parameters required for CEEMDAN decomposition, including the standard deviation of noise, the amount of noise added, and the maximum number of iterations allowed for decomposition; then, performing the decomposition operation on the unified signal.

[0015] The decomposition operation includes:

[0016] First, a preset number of positive and negative Gaussian white noises with zero expected value and constant standard deviation are added to the above unified signal to generate a noisy signal of the same preset number; the first-order intrinsic mode function is obtained by performing signal decomposition, component screening and mean calculation on each noisy signal; at the same time, the corresponding first-order residual component is obtained by calculating the difference between the unified signal and the first-order intrinsic mode function.

[0017] The signal decomposition, component screening, and mean calculation operations are as follows: by identifying local extreme points in the signal, construct upper and lower envelopes, separate the initial components that meet the conditions of matching the number of extreme points and zero crossing points and zero local mean, and take the arithmetic mean of the same order initial components of all noisy signals to obtain the first order intrinsic mode function.

[0018] Next, dynamically adjusted positive and negative Gaussian white noise is added to the first-order residual component to generate a new noisy residual signal; the above signal decomposition, component filtering and mean calculation operations are repeated on the noisy residual signal to obtain the second-order intrinsic mode function; at the same time, the corresponding second-order residual component is obtained by calculating the difference between the first-order residual component and the second-order intrinsic mode function.

[0019] Similarly, for each order, the residual components of the previous order output are used as the basis for repeatedly adding dynamically adjusted noise, generating noisy residual signals, decomposing signals, filtering components and calculating the mean, extracting the new order intrinsic mode functions, and calculating the corresponding residual components after the update, until the residual components are monotonic functions or the number of extreme points is insufficient to separate components that meet the conditions; finally, the intrinsic mode functions of all orders and the corresponding residual components are output.

[0020] In the second step, the step of first coarsening the intrinsic mode functions (EMFs) based on preset time scale parameters, dividing each EMF into several subsequences according to the corresponding time scale, means: for each EMF obtained from CEEMDAN decomposition, processing is performed sequentially from the first to the last order; during processing, based on multiple preset time scale parameters, coarsening operations at each time scale are performed on the current EMF in sequence.

[0021] In the processing corresponding to each time scale parameter, the original data sequence of the intrinsic mode function of that order is segmented into a group of several consecutive data points. A new data point is generated for each group of data points through statistical aggregation operation. All the new data points of all groups are arranged in sequence to form a new sequence, which is the coarse-grained subsequence of the intrinsic mode function of that order at the current time scale.

[0022] According to the above coarse-grained operation method, all preset time scale parameters are applied sequentially to the same order intrinsic mode function, so that each order intrinsic mode function is decomposed into multiple sub-sequences, and each sub-sequence corresponds to a time scale parameter, reflecting the signal characteristics of the order intrinsic mode function at different time resolutions;

[0023] In the second step, the construction of a similarity metric for subsequences at each time scale using a fuzzy membership function, and the calculation of the fuzzy entropy value of the subsequence at each time scale based on this similarity metric, refers to: defining the core parameters of the fuzzy membership function to convert the similarity between any two data segments in the subsequence into a quantified membership value; when the membership value reaches a preset threshold, the two data segments are determined to have a similar pattern; based on this similarity metric, fuzzy entropy is calculated for the subsequence at each time scale, the probability of repetition of similar patterns of all data segments in the subsequence is statistically analyzed, and the uncertainty of pattern repetition is converted into a quantified index through mathematical transformation, thus obtaining the fuzzy entropy value of the subsequence at the corresponding time scale.

[0024] In the second step, integrating the fuzzy entropy values ​​to form a feature matrix means: taking the set of fuzzy entropy values ​​corresponding to each order of intrinsic mode function as a column of the matrix, and arranging the sets of fuzzy entropy values ​​of all orders of intrinsic mode functions in order of order to form a two-dimensional matrix with fuzzy entropy values ​​as elements; the row dimension of this feature matrix corresponds to the multi-scale features of the signal, and the column dimension corresponds to the intrinsic mode functions of different orders.

[0025] The third step, which involves using probabilistic principal component analysis to reduce the dimensionality of the feature matrix to obtain the dimensionality-reduced feature matrix, involves: first, standardizing the feature matrix to obtain a standardized matrix; then, calculating the covariance or correlation coefficient matrix of the standardized matrix to clarify the degree of linear correlation between features; performing eigenvalue decomposition on the covariance or correlation coefficient matrix to obtain eigenvectors and corresponding eigenvalues, with each eigenvector corresponding to a principal component; then, calculating the ratio of the eigenvalue corresponding to each principal component to the sum of all eigenvalues ​​to obtain the principal component contribution rate; finally, selecting target principal components according to a preset contribution rate threshold and a minimum number of retained components; and projecting the standardized matrix onto the low-dimensional space formed by the selected target principal components to obtain the dimensionality-reduced feature matrix.

[0026] The fourth step, the CTCM-1D-CNN convolutional neural network model, includes a 1D-CNN basic network model and a CTCM optimization module;

[0027] The 1D-CNN basic network model includes 1 input layer, 2 convolutional layers, 2 batch normalization layers, 2 ReLU activation layers, 1 Dropout layer, 1 fully connected layer, and 1 regression layer.

[0028] The CTCM optimization module optimizes the learning rate reduction factor, initial learning rate, and minimum batch size of the 1D-CNN basic network model. The configuration parameters include population size, maximum number of generations, crossover ratio, mutation ratio, mutation rate, number of reference points, and number of decision variables.

[0029] The fourth step, where the CTCM-1D-CNN convolutional neural network model is constructed by optimizing a one-dimensional convolutional neural network using a tribal competition and member cooperation algorithm, involves: first, encoding the hyperparameters of the 1D-CNN basic network model as chromosomes, including the learning rate decrease factor, initial learning rate, and minimum batch size; designing a multi-objective function with mean absolute percentage error, correlation coefficient between predicted and theoretical values, and JS divergence as sub-functions, where mean absolute percentage error corresponds to coefficient α, correlation coefficient between predicted and theoretical values ​​corresponds to coefficient β, and JS divergence corresponds to coefficient γ; and weighting the three sub-functions using α, β, and γ, where α + β + γ = 1; and then running the tribal competition and member cooperation algorithm to iteratively optimize the hyperparameters by simulating the process of cooperation among members within a tribe and resource competition between tribes.

[0030] In the iterative process of the tribal competition and member cooperation algorithm, the 1D-CNN base network model is trained using the stochastic gradient descent momentum method. An initial maximum number of training iterations is set, with the earthquake signal label set to 0 and the artificial blasting signal label set to 1. After inputting the training data into the network, based on the predicted and true labels output during training, the mean absolute error, accuracy, mean deviation error, decision coefficient, and root mean square error of the predicted and true labels are calculated. These serve as feedback indicators for evaluating the hyperparameter suitability of the tribal competition and member cooperation algorithm, dynamically adjusting the tribal cooperation intensity and competition strategy. When the iteration stopping condition is met, the optimal hyperparameters obtained from the tribal competition and member cooperation algorithm are used to replace the corresponding hyperparameters of the 1D-CNN base network model, completing the model construction.

[0031] In the fourth step, the CTCM-1D-CNN convolutional neural network model performs earthquake event type identification and prediction on the aforementioned dimensionality-reduced feature matrix. The output of the acquired seismic wave type refers to: inputting the aforementioned dimensionality-reduced feature matrix into the CTCM-1D-CNN convolutional neural network model for prediction, and outputting a predicted label for the earthquake event type; when the output predicted label is 0, the corresponding seismic wave type is determined to be a natural earthquake; when the output predicted label is 1, the corresponding seismic wave type is determined to be an artificial blast, thereby obtaining the specific type of the seismic wave.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] 1. This invention discloses a seismic wave identification method based on deep learning, comprising the following steps: First, acquiring seismic wave signals to obtain raw data, and then performing centering and normalization processing to obtain a unified signal. Second, using CEEMDAN to decompose the unified signal to obtain several orders of intrinsic mode functions and residual components. For each order of mode function, coarse-grained processing is performed according to a preset time scale parameter to divide it into subsequences. Then, a similarity criterion is constructed using a fuzzy membership function, and the fuzzy entropy values ​​of each subsequence are calculated and integrated into a feature matrix. Third, probabilistic principal component analysis is used to reduce the dimensionality of the feature matrix to obtain a dimensionality-reduced feature matrix. Fourth, the dimensionality-reduced feature matrix is ​​input into a CTCM-1D-CNN model optimized by a tribal competition and member cooperation algorithm to predict the output seismic wave type, thus completing the identification. In application, firstly, seismic wave signals are acquired to obtain raw data, and then the core energy segment of the signal is focused through centering to highlight key components. Waveform features are normalized to eliminate amplitude differences between different signals, resulting in a unified signal with consistent dimensions. CEEMDAN decomposition is then used to decompose this unified signal, suppressing mode aliasing to obtain several orders of intrinsic mode functions and residual components. This accurately separates signal details of different frequency components. For each order of mode function, multi-scale fuzzy entropy values ​​are calculated and integrated into a feature matrix, comprehensively capturing irregular features of the signal at different time resolutions. Dimensionality reduction is performed using probabilistic principal component analysis, retaining core features and eliminating redundant information, reducing computational load while improving feature specificity. Finally, the dimensionality-reduced feature matrix is ​​input into the CTCM-1D-CNN model optimized by tribal competition and member cooperation algorithms. Dynamic optimization of hyperparameters improves model adaptability, ultimately outputting a 0 (natural earthquake) or 1 (artificial blasting) label to achieve seismic wave identification. The advantages of this invention also include:

[0034] First, natural earthquakes are dominated by low-frequency components with slow waveform decay and a high proportion of shear wave energy, while artificial blasting has prominent high-frequency components with rapid waveform decay and a high proportion of P-wave energy. This invention preserves the core waveform features through seismic wave preprocessing, uses CEEMDAN to decompose and separate the feature differences of seismic waves of different frequencies, combines multi-scale feature extraction to quantify the irregular differences of the two types of signals in terms of time resolution, and finally uses the CTCM-1D-CNN model to deeply learn these differences to achieve effective identification of the two types of seismic waves.

[0035] Secondly, CEEMDAN decomposition is used to suppress mode aliasing, accurately separating signal details of different frequency components. Combined with multi-scale coarsening and fuzzy entropy calculation, the irregularity characteristics of seismic waves are quantified from different time resolutions, extracting comprehensive features covering frequency differences and temporal dynamic characteristics. Combined with probabilistic principal component analysis for dimensionality reduction and redundancy removal, the problem of feature redundancy or loss of details caused by single signal processing in existing technologies is overcome. At the same time, the model hyperparameters are dynamically optimized through tribal competition and member cooperation algorithms, improving the model's ability to fit features, thereby improving the accuracy of seismic wave identification.

[0036] Thirdly, the use of tribal competition and member cooperation algorithms to dynamically optimize the hyperparameters of the 1D-CNN model replaces the traditional method of manually fixing parameters. This allows the model to adaptively adjust according to different monitoring scenarios, avoiding performance fluctuations caused by parameter rigidity in existing technologies, improving the stability of seismic wave identification, and ensuring reliable output results even in complex geological or noisy environments.

[0037] Therefore, this invention can not only identify seismic waves, but also has good accuracy and stability in seismic wave identification.

[0038] 2. In the deep learning-based seismic wave identification method of this invention, the feature matrix is ​​first standardized to obtain a standardized matrix. Principal component analysis is then used to reduce the dimensionality of the standardized matrix, resulting in a dimensionality-reduced feature matrix. In application, the feature matrix is ​​first standardized to eliminate dimensional differences between different features, yielding a standardized matrix. Then, the covariance or correlation coefficient matrix of the standardized matrix is ​​calculated. The linear correlation between features is quantified numerically (higher correlation indicates stronger redundancy). After eigenvalue decomposition of the covariance or correlation coefficient matrix, eigenvectors and corresponding eigenvalues ​​are obtained. The resulting eigenvectors are linear combinations of the original features, with each eigenvector corresponding to a corresponding eigenvalue. The process involves identifying independent, non-overlapping principal components, with eigenvalues ​​reflecting the capacity of each principal component to carry the original feature information (larger eigenvalues ​​indicate more core information). Target principal components are then selected by calculating their contribution rate (eigenvalue proportion). High-contribution principal components are retained according to a preset threshold and minimum retention quantity, ensuring no core features are lost while eliminating low-value redundant information. Finally, the standardized matrix is ​​projected onto the low-dimensional space formed by the target principal components to obtain a dimensionality-reduced feature matrix. By quantifying feature correlations and focusing on core information, this method eliminates linear redundancy between features and reduces data size through dimensionality compression, thereby reducing the computational load of subsequent models and improving processing efficiency. Therefore, this invention not only demonstrates good accuracy and stability in seismic wave identification but also improves computational speed.

[0039] 3. In the deep learning-based seismic wave identification method of this invention, in the processing corresponding to each time scale parameter, the original data sequence of the intrinsic mode function of that order is segmented into groups of several consecutive data points. A new data point is generated for each group of data points through statistical aggregation. All the new data points in all groups are arranged sequentially to form a new sequence, thus obtaining the coarse-grained subsequence of the intrinsic mode function of that order at the current time scale. In application, for a certain order intrinsic mode function obtained by CEEMDAN decomposition, multiple time scale parameters are first set. At each time scale, the original data sequence is continuously segmented according to the parameters. Statistical aggregation is performed on each group of data, and the aggregation result of each group is used as the new... Data points, when arranged in their original order, form a coarse-grained subsequence at that time scale. By changing the time scale parameters, multi-resolution resampling of the signal is achieved. Small-scale parameters retain more instantaneous details, while large-scale parameters smooth local fluctuations to highlight the overall trend. This allows for the analysis of signal dynamic characteristics across different time dimensions, accurately capturing the temporal differences between natural earthquakes and artificial blasting, amplifying the short-term high-frequency abrupt changes of artificial blasting, and enhancing the long-term low-frequency stability of natural earthquakes. Multi-scale expansion of time-dimensional features avoids the omission of key information. Statistical aggregation smooths noise, highlights core waveforms, and improves the feature matrix's ability to distinguish seismic wave types, laying a crucial feature foundation for high-precision identification. Therefore, this invention not only improves computational speed but also enhances the ability to distinguish seismic wave types.

[0040] 4. In the deep learning-based seismic wave identification method of this invention, a preset number of positive and negative Gaussian white noises with zero expected value and constant standard deviation are added to the aforementioned unified signal to generate a noisy signal of equal quantity. In application, the standard deviation and preset number of Gaussian white noises are first set according to the energy level of the unified signal. Then, equal quantities of positive and negative Gaussian white noises are alternately superimposed on the unified signal; that is, each time positive noise is added, a noisy signal is generated, followed by the addition of negative noise of the same amplitude to generate a corresponding signal, ultimately obtaining a noisy signal of equal quantity. Because Gaussian white noise has the characteristic of uniform distribution across the entire frequency band, it can provide a reference benchmark for blurred mode boundaries in the signal, causing originally overlapping frequency components (such as the local aliasing of low-frequency signals from natural earthquakes and high-frequency signals from artificial blasting) to separate under noise excitation. Simultaneously, the design of adding positive and negative noise in pairs can cancel out the influence of the noise itself during subsequent averaging, avoiding the introduction of additional biases and noise residues that interfere with feature extraction, while ensuring the stability of the decomposition results. Therefore, this invention not only improves the ability to distinguish seismic wave types but also enhances the accuracy of seismic wave feature extraction. Attached Figure Description

[0041] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0042] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] See Figure 1 A seismic wave identification method based on deep learning, the identification method comprising the following steps:

[0044] Step 1: First, collect the seismic wave signals generated by the earthquake event to obtain raw data. Then, extract the center of the raw data and normalize the extracted raw data to obtain a unified signal.

[0045] Step 2: The unified signal is decomposed using the CEEMDAN decomposition method to obtain several orders of intrinsic mode functions and corresponding residual components. For each order of intrinsic mode function, coarse-grained processing is first performed based on preset time scale parameters to divide each order of intrinsic mode function into several sub-sequences according to the corresponding time scale. Then, a similarity metric is constructed for the sub-sequences under each time scale using a fuzzy membership function. Based on the similarity metric, the fuzzy entropy value of the sub-sequences under each time scale is calculated. Finally, the fuzzy entropy values ​​are integrated to form a feature matrix.

[0046] Step 3: Use probabilistic principal component analysis to reduce the dimensionality of the above feature matrix to obtain the dimensionality-reduced feature matrix;

[0047] Step 4: Input the dimensionality-reduced feature matrix into the CTCM-1D-CNN convolutional neural network model. The CTCM-1D-CNN convolutional neural network model is constructed by optimizing a one-dimensional convolutional neural network using a tribal competition and member cooperation algorithm. The CTCM-1D-CNN convolutional neural network model performs earthquake event type identification and prediction on the dimensionality-reduced feature matrix and outputs the type of the acquired seismic waves, thereby completing the identification of seismic waves.

[0048] In the first step, acquiring the seismic wave signals generated by the earthquake event to obtain the raw data means: acquiring the acceleration data generated by the earthquake event through earthquake monitoring equipment, and using the acquired acceleration data as the raw data.

[0049] In the first step, the central truncation of the original data means: taking the peak point of the acceleration data waveform in the original data as the center, symmetrically truncating the same number of acceleration data segments to the left and right of the peak point.

[0050] In the second step, the step of using the CEEMDAN decomposition method to decompose the above-mentioned unified signal to obtain several orders of intrinsic mode functions and corresponding residual components means: first, determining the parameters required for CEEMDAN decomposition, including the standard deviation of noise, the amount of noise added, and the maximum number of iterations allowed for decomposition; then, performing the decomposition operation on the unified signal.

[0051] The decomposition operation includes:

[0052] First, a preset number of positive and negative Gaussian white noises with zero expected value and constant standard deviation are added to the above unified signal to generate a noisy signal of the same preset number; the first-order intrinsic mode function is obtained by performing signal decomposition, component screening and mean calculation on each noisy signal; at the same time, the corresponding first-order residual component is obtained by calculating the difference between the unified signal and the first-order intrinsic mode function.

[0053] The signal decomposition, component screening, and mean calculation operations are as follows: by identifying local extreme points in the signal, construct upper and lower envelopes, separate the initial components that meet the conditions of matching the number of extreme points and zero crossing points and zero local mean, and take the arithmetic mean of the same order initial components of all noisy signals to obtain the first order intrinsic mode function.

[0054] Next, dynamically adjusted positive and negative Gaussian white noise is added to the first-order residual component to generate a new noisy residual signal; the above signal decomposition, component filtering and mean calculation operations are repeated on the noisy residual signal to obtain the second-order intrinsic mode function; at the same time, the corresponding second-order residual component is obtained by calculating the difference between the first-order residual component and the second-order intrinsic mode function.

[0055] Similarly, for each order, the residual components of the previous order output are used as the basis for repeatedly adding dynamically adjusted noise, generating noisy residual signals, decomposing signals, filtering components and calculating the mean, extracting the new order intrinsic mode functions, and calculating the corresponding residual components after the update, until the residual components are monotonic functions or the number of extreme points is insufficient to separate components that meet the conditions; finally, the intrinsic mode functions of all orders and the corresponding residual components are output.

[0056] In the second step, the step of first coarsening the intrinsic mode functions (EMFs) based on preset time scale parameters, dividing each EMF into several subsequences according to the corresponding time scale, means: for each EMF obtained from CEEMDAN decomposition, processing is performed sequentially from the first to the last order; during processing, based on multiple preset time scale parameters, coarsening operations at each time scale are performed on the current EMF in sequence.

[0057] In the processing corresponding to each time scale parameter, the original data sequence of the intrinsic mode function of that order is segmented into a group of several consecutive data points. A new data point is generated for each group of data points through statistical aggregation operation. All the new data points of all groups are arranged in sequence to form a new sequence, which is the coarse-grained subsequence of the intrinsic mode function of that order at the current time scale.

[0058] According to the above coarse-grained operation method, all preset time scale parameters are applied sequentially to the same order intrinsic mode function, so that each order intrinsic mode function is decomposed into multiple sub-sequences, and each sub-sequence corresponds to a time scale parameter, reflecting the signal characteristics of the order intrinsic mode function at different time resolutions;

[0059] In the second step, the construction of a similarity metric for subsequences at each time scale using a fuzzy membership function, and the calculation of the fuzzy entropy value of the subsequence at each time scale based on this similarity metric, refers to: defining the core parameters of the fuzzy membership function to convert the similarity between any two data segments in the subsequence into a quantified membership value; when the membership value reaches a preset threshold, the two data segments are determined to have a similar pattern; based on this similarity metric, fuzzy entropy is calculated for the subsequence at each time scale, the probability of repetition of similar patterns of all data segments in the subsequence is statistically analyzed, and the uncertainty of pattern repetition is converted into a quantified index through mathematical transformation, thus obtaining the fuzzy entropy value of the subsequence at the corresponding time scale.

[0060] In the second step, integrating the fuzzy entropy values ​​to form a feature matrix means: taking the set of fuzzy entropy values ​​corresponding to each order of intrinsic mode function as a column of the matrix, and arranging the sets of fuzzy entropy values ​​of all orders of intrinsic mode functions in order of order to form a two-dimensional matrix with fuzzy entropy values ​​as elements; the row dimension of this feature matrix corresponds to the multi-scale features of the signal, and the column dimension corresponds to the intrinsic mode functions of different orders.

[0061] The third step, which involves using probabilistic principal component analysis to reduce the dimensionality of the feature matrix to obtain the dimensionality-reduced feature matrix, involves: first, standardizing the feature matrix to obtain a standardized matrix; then, calculating the covariance or correlation coefficient matrix of the standardized matrix to clarify the degree of linear correlation between features; performing eigenvalue decomposition on the covariance or correlation coefficient matrix to obtain eigenvectors and corresponding eigenvalues, with each eigenvector corresponding to a principal component; then, calculating the ratio of the eigenvalue corresponding to each principal component to the sum of all eigenvalues ​​to obtain the principal component contribution rate; finally, selecting target principal components according to a preset contribution rate threshold and a minimum number of retained components; and projecting the standardized matrix onto the low-dimensional space formed by the selected target principal components to obtain the dimensionality-reduced feature matrix.

[0062] The fourth step, the CTCM-1D-CNN convolutional neural network model, includes a 1D-CNN basic network model and a CTCM optimization module;

[0063] The 1D-CNN basic network model includes 1 input layer, 2 convolutional layers, 2 batch normalization layers, 2 ReLU activation layers, 1 Dropout layer, 1 fully connected layer, and 1 regression layer.

[0064] The CTCM optimization module optimizes the learning rate reduction factor, initial learning rate, and minimum batch size of the 1D-CNN basic network model. The configuration parameters include population size, maximum number of generations, crossover ratio, mutation ratio, mutation rate, number of reference points, and number of decision variables.

[0065] The fourth step, where the CTCM-1D-CNN convolutional neural network model is constructed by optimizing a one-dimensional convolutional neural network using a tribal competition and member cooperation algorithm, involves: first, encoding the hyperparameters of the 1D-CNN basic network model as chromosomes, including the learning rate decrease factor, initial learning rate, and minimum batch size; designing a multi-objective function with mean absolute percentage error, correlation coefficient between predicted and theoretical values, and JS divergence as sub-functions, where mean absolute percentage error corresponds to coefficient α, correlation coefficient between predicted and theoretical values ​​corresponds to coefficient β, and JS divergence corresponds to coefficient γ; and weighting the three sub-functions using α, β, and γ, where α + β + γ = 1; and then running the tribal competition and member cooperation algorithm to iteratively optimize the hyperparameters by simulating the process of cooperation among members within a tribe and resource competition between tribes.

[0066] In the iterative process of the tribal competition and member cooperation algorithm, the 1D-CNN base network model is trained using the stochastic gradient descent momentum method. An initial maximum number of training iterations is set, with the earthquake signal label set to 0 and the artificial blasting signal label set to 1. After inputting the training data into the network, based on the predicted and true labels output during training, the mean absolute error, accuracy, mean deviation error, decision coefficient, and root mean square error of the predicted and true labels are calculated. These serve as feedback indicators for evaluating the hyperparameter suitability of the tribal competition and member cooperation algorithm, dynamically adjusting the tribal cooperation intensity and competition strategy. When the iteration stopping condition is met, the optimal hyperparameters obtained from the tribal competition and member cooperation algorithm are used to replace the corresponding hyperparameters of the 1D-CNN base network model, completing the model construction.

[0067] In the fourth step, the CTCM-1D-CNN convolutional neural network model performs earthquake event type identification and prediction on the aforementioned dimensionality-reduced feature matrix. The output of the acquired seismic wave type refers to: inputting the aforementioned dimensionality-reduced feature matrix into the CTCM-1D-CNN convolutional neural network model for prediction, and outputting a predicted label for the earthquake event type; when the output predicted label is 0, the corresponding seismic wave type is determined to be a natural earthquake; when the output predicted label is 1, the corresponding seismic wave type is determined to be an artificial blast, thereby obtaining the specific type of the seismic wave.

[0068] The following are supplementary descriptions of the present invention:

[0069] In the preferred embodiment of the present invention, the fourth step uses 1000 data points for training the CTCM-1D-CNN convolutional neural network model, consisting of 800 natural earthquake events and 200 artificial blasting signals. These data are collected by digital strong-motion seismometers from seismic stations across the country.

[0070] In the second step of the present invention, the parameters required for the CEEMDAN decomposition include a standard deviation of noise of 0.2, a number of noise additions of 24, and a maximum allowed number of iterations of 3600.

[0071] In the preferred embodiment of the present invention, in the fourth step, the average absolute error, accuracy, average deviation error, decision coefficient, and root mean square error between the predicted label and the real label are used as feedback indicators for evaluating the hyperparameter adaptability of the tribal competition and member cooperation algorithm. When calculating the accuracy, the predicted label value is rounded.

[0072] Example 1:

[0073] See Figure 1 A seismic wave identification method based on deep learning, the identification method comprising the following steps:

[0074] Step 1: First, collect the seismic wave signals generated by the earthquake event to obtain raw data. Then, extract the center of the raw data and normalize the extracted raw data to obtain a unified signal.

[0075] Step 2: The unified signal is decomposed using the CEEMDAN decomposition method to obtain several orders of intrinsic mode functions and corresponding residual components. For each order of intrinsic mode function, coarse-grained processing is first performed based on preset time scale parameters to divide each order of intrinsic mode function into several sub-sequences according to the corresponding time scale. Then, a similarity metric is constructed for the sub-sequences under each time scale using a fuzzy membership function. Based on the similarity metric, the fuzzy entropy value of the sub-sequences under each time scale is calculated. Finally, the fuzzy entropy values ​​are integrated to form a feature matrix.

[0076] Step 3: Use probabilistic principal component analysis to reduce the dimensionality of the above feature matrix to obtain the dimensionality-reduced feature matrix;

[0077] Step 4: Input the dimensionality-reduced feature matrix into the CTCM-1D-CNN convolutional neural network model. The CTCM-1D-CNN convolutional neural network model is constructed by optimizing a one-dimensional convolutional neural network using a tribal competition and member cooperation algorithm. The CTCM-1D-CNN convolutional neural network model performs earthquake event type identification and prediction on the dimensionality-reduced feature matrix and outputs the type of the acquired seismic waves, thereby completing the identification of seismic waves.

[0078] Example 2:

[0079] The basic content is the same as in Example 1, except that: in the first step, acquiring the seismic wave signal generated by the earthquake event to obtain the raw data means: acquiring the acceleration data generated by the earthquake event through earthquake monitoring equipment, and using the acquired acceleration data as the raw data; in the first step, centering the raw data means: taking the peak point of the acceleration data waveform in the raw data as the center, symmetrically extracting the same number of acceleration data segments to the left and right of the peak point.

[0080] In application, acceleration data collected by earthquake monitoring equipment can directly reflect the vibration intensity and dynamic characteristics of seismic waves, providing the original physical quantity basis for subsequent analysis; symmetrical interception is performed with the peak point of the acceleration waveform as the center, and the interception formula is as follows: On the one hand, it can accurately locate the core area where seismic wave energy is most concentrated, effectively avoiding interference from ineffective vibration segments at the beginning and end of the original data (such as environmental noise and instrument zero drift), and ensuring that the extracted data segment completely contains the peak point and key fluctuation characteristics before and after it; on the other hand, there are significant differences in the vibration characteristics of natural earthquakes and artificial blasting signals before and after the peak. The former exhibits a continuously decaying low-frequency vibration after the peak, while the latter is often accompanied by steep high-frequency pulses before and after the peak. Symmetrical extraction makes the number of data points on both sides equal, which can completely preserve the differentiated waveform characteristics of the two types of signals in terms of rising slope, falling decay rate, and oscillation period, and establish a unified benchmark for subsequent comparative analysis.

[0081] Based on this, normalization processing is combined to eliminate the amplitude differences of different seismic events. The resulting unified signal retains key waveform features and has consistency in lateral comparison. This avoids feature bias caused by asymmetric truncation length and provides more accurate basic data for multi-scale feature extraction and fuzzy entropy calculation of CEEMDAN decomposition.

[0082] Example 3:

[0083] The basic content is the same as in Example 1, except that in the second step, the CEEMDAN decomposition method is used to decompose the unified signal to obtain several orders of intrinsic mode functions and corresponding residual components. This means: first, determining the parameters required for CEEMDAN decomposition, including the standard deviation of noise, the amount of noise added, and the maximum number of iterations allowed for decomposition; then, performing the decomposition operation on the unified signal; the decomposition operation includes: first, adding a preset number of positive and negative Gaussian white noise with zero expected value and constant standard deviation to the unified signal to generate a noisy signal of the same preset number; performing signal decomposition, component screening, and mean calculation on each noisy signal to obtain the first order intrinsic mode function; and simultaneously calculating the difference between the unified signal and the first order intrinsic mode function to obtain the corresponding first order residual component; the signal decomposition, component screening, and mean calculation operation is as follows: constructing upper and lower envelopes by identifying local extreme points in the signal. The process involves separating initial components that meet the conditions of matching the number of extrema and zero-crossing points and having a local mean of zero. The arithmetic mean of all initial components of the same order in the noisy signal is then taken to obtain the first-order intrinsic mode function (IMF). Next, dynamically adjusted positive and negative Gaussian white noise is added to the first-order residual components to generate a new noisy residual signal. This signal decomposition, component selection, and mean calculation are repeated on the noisy residual signal to obtain the second-order IMF. Simultaneously, the difference between the first-order residual component and the second-order IMF is used to calculate the corresponding second-order residual component. This process continues for each order, based on the residual component output from the previous order. The steps of adding dynamically adjusted noise, generating noisy residual signals, signal decomposition, component selection and mean calculation, extracting new-order IMFs, and calculating the updated residual components are repeated until the residual components are monotonic functions or the number of extrema is insufficient to separate components that meet the conditions. Finally, all IMFs and their corresponding residual components are output.

[0084] When applying this method, first determine the key parameters required for CEEMDAN decomposition based on the actual signal analysis needs (specifically, set the standard deviation of noise to 0.2, the number of noise additions to 24, and the maximum allowed number of decomposition iterations to 3600). Using a unified signal x(t) as the processing object, perform the decomposition according to the following process:

[0085] Add m positive and negative Gaussian white noises with zero expectation and constant standard deviation (m is the set number of noises to be added, corresponding to parameters such as the noise standard deviation) to a uniform signal x(t) to generate m noisy signals. It satisfies the formula: (in (where β is white noise and β is the noise figure); for each EMD decomposition yields m first-order components. The arithmetic mean of the CEEMDAN first-order intrinsic mode functions is obtained: The first-order residual component is calculated by using the difference between the unified signal and the first-order component. For the (k-1)th order residual component Add dynamically adjusted positive and negative Gaussian white noise (dynamically adjust noise figure) Residuals from previous EMD processing Correlation can adapt to the energy attenuation characteristics of residual signals, ensuring the stability of IMF decomposition at each order, and generating m new noisy residual signals. For each Perform signal decomposition (EMD) to obtain m k−1 order components. The arithmetic mean is used to obtain the k-1th order eigenmode function: Update residual components: ;

[0086] Repeat the above iterative steps until the residual components are found. The process stops when the number of iterations reaches the maximum allowed number of iterations or when EMD decomposition is impossible due to insufficient number of monotonic functions or extreme points; ultimately, the original unified signal x(t) can be reconstructed from the residual components and eigenmode functions of each order:

[0087] ;

[0088] This process can accurately output several orders of intrinsic mode functions and their corresponding residual components. This noisy decomposition strategy enhances the stability and robustness of signal decomposition. Even if the original seismic wave signal has weak fluctuations or uneven energy distribution, the consistency of the decomposition results can be improved through the perturbation effect of noise. This provides a more accurate frequency component basis for subsequent multi-scale fuzzy entropy calculation based on IMFs of various orders, thereby ensuring that the extracted feature matrix can truly reflect the essential differences in waveform characteristics between natural earthquakes and artificial blasting.

[0089] Example 4:

[0090] The basic content is the same as in Example 3, except that in the second step, the coarsening process performed on each intrinsic mode function based on preset time scale parameters, dividing each intrinsic mode function into several subsequences according to the corresponding time scale, means that for each intrinsic mode function obtained from CEEMDAN decomposition, it is processed one by one in order from the first to the last; during processing, according to multiple preset time scale parameters, the coarsening operation at each time scale is performed on the current intrinsic mode function in sequence: in the processing corresponding to each time scale parameter, the... The original data sequence of the intrinsic mode functions (IMFs) is segmented into groups of consecutive data points. A new data point is generated for each group through statistical aggregation. All the new data points from each group are then arranged sequentially to form a new sequence, thus obtaining the coarse-grained subsequence of the IMF at the current time scale. Following this coarse-graining method, all preset time scale parameters are applied sequentially to the same IMF, decomposing each IMF into multiple subsequences. Each subsequence corresponds to a time scale parameter, reflecting the IMF at different time scales. The signal characteristics at time resolution; in the second step, the construction of a similarity metric for subsequences at each time scale using a fuzzy membership function, and the calculation of the fuzzy entropy value of the subsequence at each time scale based on the similarity metric, means: by defining the core parameters of the fuzzy membership function, the similarity between any two data segments in the subsequence is converted into a quantized membership value. When the membership value reaches a preset threshold, the two data segments are determined to be similar patterns. Based on the similarity metric, fuzzy entropy is calculated for the subsequence at each time scale, and the probability of repetition of similar patterns of all data segments in the subsequence is statistically analyzed. The uncertainty of pattern repetition is converted into a quantified index through mathematical transformation, thus obtaining the fuzzy entropy value of the subsequence at the corresponding time scale. In the second step, the integration of the fuzzy entropy values ​​into a feature matrix means: the set of fuzzy entropy values ​​corresponding to each order of intrinsic mode function is taken as a column of the matrix, and the fuzzy entropy value sets of all orders of intrinsic mode functions are arranged in order of order to form a two-dimensional matrix with fuzzy entropy values ​​as elements. The row dimension of this feature matrix corresponds to the multi-scale features of the signal, and the column dimension corresponds to the intrinsic mode functions of different orders.

[0091] In application, the Q-order intrinsic mode functions (IMFs) are obtained through CEEMDAN decomposition. For these Q-order IMFs, the multi-scale fuzzy entropy is calculated according to the following procedure:

[0092] Define the phase space dimension m=2, the fuzzy function gradient as 2, and the similarity tolerance r=0.2. Based on the original signal data u(i) (i=1, 2, ..., N, where N is the data length), reconstruct the phase space through coarse-grained processing and construct the following vector: Then calculate the vectors in sequence. and Distance between (j=1, 2, ..., N−m+1 and j=i) And it is transformed into membership values ​​through a fuzzy membership function: For each vector Calculate the membership degree between vectors and other vectors, and then calculate the similarity between vectors in m dimensions. Specifically Further calculate the average similarity of m-dimensional samples. Increase the phase space dimension to m+1, and repeat the above steps to obtain the average similarity of m+1 dimension samples. Finally, the fuzzy entropy corresponding to the sample is obtained. Its value can reflect the degree of irregularity and disorder in the time series;

[0093] The fuzzy entropy values ​​are extracted from the Q-order IMFs to obtain the multi-scale fuzzy entropy (MFE) features of a single signal. If P original signals are processed, the process can be repeated to obtain P sets of CEEMDAN multi-scale fuzzy entropy values, which are arranged according to the signal dimension and IMF order to form a P×Q dimensional feature matrix. This P×Q dimensional feature matrix can accurately characterize the differences in the degree of disorder of the signal at different scales and frequency bands, providing highly discriminative feature inputs for pattern classification tasks (such as seismic wave type discrimination).

[0094] Example 5:

[0095] The basic content is the same as in Example 1, except that in the third step, the dimensionality reduction of the feature matrix using probabilistic principal component analysis to obtain the dimensionality-reduced feature matrix refers to: first, standardizing the feature matrix to obtain a standardized matrix; then calculating the covariance or correlation coefficient matrix of the standardized matrix, clarifying the degree of linear correlation between features through the covariance or correlation coefficient, performing eigenvalue decomposition on the covariance or correlation coefficient matrix to obtain eigenvectors and corresponding eigenvalues, with each eigenvector corresponding to a principal component; then calculating the ratio of the eigenvalue corresponding to each principal component to the sum of all eigenvalues ​​to obtain the principal component contribution rate; finally, selecting target principal components according to a preset contribution rate threshold and minimum retention quantity; and projecting the standardized matrix onto the low-dimensional space formed by the selected target principal components to obtain the dimensionality-reduced feature matrix.

[0096] In application, probabilistic principal component analysis (PPCA) is performed on the CEEMDAN multi-scale fuzzy entropy feature matrix for dimensionality reduction: First, the feature matrix is ​​standardized to eliminate the dimensional differences between different features, resulting in a standardized matrix. Next, the covariance or correlation coefficient matrix of this standardized matrix is ​​calculated to clarify the linear correlation between features. Then, eigenvalue decomposition is performed on the covariance or correlation coefficient matrix to obtain eigenvectors and corresponding eigenvalues, with each eigenvector corresponding to a principal component. The ratio of the eigenvalue corresponding to each principal component to the sum of all eigenvalues ​​is then calculated to obtain the principal component contribution rate. Finally, based on the preset screening criteria of a contribution rate ≥ 80% and retaining no fewer than two composite features, the features are selected... The target principal components are extracted, and the feature vectors corresponding to these target principal components together constitute a low-dimensional projection matrix (i.e., the low-dimensional space formed by the target principal components). The number of synthetic features retained after screening, q, is generally 2-5. Then, the normalized matrix is ​​multiplied by the projection matrix. Through this linear transformation, the high-dimensional normalized matrix is ​​mapped to the low-dimensional space spanned by the target principal components, finally obtaining a P×q dimension reduced feature matrix. This process not only removes redundant information but also retains the core discriminative features, providing concise and highly discriminative data for subsequent model input. By retaining principal components with high contribution rates, the original feature information is preserved to the maximum extent, ensuring the discriminative ability of the feature matrix.

[0097] Example 6:

[0098] The basic content is the same as in Example 1, except that: in the fourth step, the CTCM-1D-CNN convolutional neural network model includes a 1D-CNN basic network model and a CTCM optimization module; the 1D-CNN basic network model includes one input layer, two convolutional layers, two batch normalization layers, two ReLU activation layers, one Dropout layer, one fully connected layer, and one regression layer; the CTCM optimization module optimizes the learning rate reduction factor, initial learning rate, and minimum batch size of the 1D-CNN basic network model, and the configuration parameters include population size, maximum number of generations, crossover ratio, mutation ratio, mutation rate, number of reference points, and number of decision variables; the fourth step, the CTCM-1D-CNN convolutional neural network model constructed by optimizing a one-dimensional convolutional neural network using a tribal competition and member cooperation algorithm, refers to: first, encoding the hyperparameters of the 1D-CNN basic network model into chromosomes, the hyperparameters including the learning rate reduction factor, initial learning rate, and minimum batch size; designing a multi-objective function with mean absolute percentage error, correlation coefficient between predicted and theoretical values, and JS divergence as sub-functions, its The mean absolute percentage error corresponds to coefficient α, the correlation coefficient between predicted and theoretical values ​​corresponds to coefficient β, and the JS divergence corresponds to coefficient γ. These three sub-functions are weighted using α, β, and γ, with α + β + γ = 1. The tribal competition and member cooperation algorithm is then run to iteratively optimize the hyperparameters by simulating the process of cooperation among members within a tribe and resource competition between tribes. During the iteration of the tribal competition and member cooperation algorithm, the 1D-CNN base network model is trained using the stochastic gradient descent momentum method, with an initial maximum number of training iterations. The earthquake signal label is set to 0, and the artificial blasting signal label is set to 1. After inputting the training data into the network, based on the predicted and true labels output during training, the mean absolute error, accuracy, mean deviation error, decision coefficient, and root mean square error between the predicted and true labels are calculated. These serve as feedback indicators for evaluating the hyperparameter suitability of the tribal competition and member cooperation algorithm, dynamically adjusting the tribal cooperation intensity and competition strategy. When the iteration stopping condition is met, the optimal hyperparameters obtained from the tribal competition and member cooperation algorithm are used to replace the corresponding hyperparameters of the 1D-CNN base network model, completing the model construction.

[0099] In application, first build the 1D-CNN basic network model and CTCM optimization module according to the process, encode the hyperparameters (learning rate decrease factor, initial learning rate, minimum batch size) into chromosomes, and design a multi-objective function with mean absolute percentage error, correlation coefficient between predicted and theoretical values, and JS divergence as sub-functions:

[0100] ;

[0101] in, This represents the model's predicted value for the sample. This is the theoretical value of the sample. The average of the predicted set. The average of the set of theoretical values ​​for the sample. This refers to the total number of samples used in the prediction calculation, i.e., the number of samples in the test set. For MAPE (Mean Absolute Percentage Error). The correlation coefficient is... Let JS divergence be used.

[0102] Specific ;

[0103] ;

[0104] And α+β+γ=1; the mean absolute percentage error (MAPE) can determine the existence of absolute value bias in the prediction model. The smaller the value, the smaller the prediction error. The range of MAPE is [0, +∞); the correlation coefficient between the predicted value and the theoretical value is used to evaluate the correlation coefficient between the predicted result vector and the theoretical result vector. The range of the value is [-1, 1]; the value of JS divergence (Jensen-Shannon Divergence) is between 0 and 1. It is 0 when they are exactly the same and 1 when they are opposite. JS divergence measures the similarity between two probability distributions; these three are used as sub-functions to complement each other from different dimensions and comprehensively and accurately optimize the model to adapt to the earthquake and blasting signal prediction tasks;

[0105] Next, the CTCM parameters were configured, with the optimal population size (50), maximum number of iterations (50), crossover percentage (0.5), mutation percentage (0.5), mutation rate (0.02), number of reference points (10), and number of decision variables (3). The CTCM algorithm was run to simulate tribal cooperation and competition, iteratively optimizing the hyperparameters (learning rate descent factor, initial learning rate, and minimum batch size). The 1D-CNN base network model was trained using stochastic gradient descent momentum method (SGD-M), with an initial maximum training iteration count of 50. Label values ​​of 0 and 1 were set to distinguish between earthquake and artificial blasting signals.

[0106] The training data (specifically 883 natural earthquake events and 117 artificial blasting signals, totaling 1000 data points) was input. Based on the predicted and true labels, the following metrics were calculated: Mean Absolute Error (MAE, which measures the absolute error between the predicted and true values, reflecting prediction accuracy), Accuracy (the percentage of correctly classified signals, evaluating the model's discriminative ability), Mean Deviation Error (MBE, reflecting the overall direction of prediction deviation), R-squared (measuring the model's interpretability of the data), and Root Mean Square Error (RMSE, reflecting the degree of dispersion between the predicted and true values). These metrics quantified the model performance from different dimensions and fed back to the multi-objective function optimization. After the stopping condition was met, the corresponding parameters of the basic network were replaced with the optimal hyperparameters to obtain the CTCM-1D-CNN convolutional neural network model. This model can then be used to perform earthquake and blasting predictions on the PPCA-CEEMDAN-MFE feature set. The model's adaptability was verified through metric evaluation, and the signal categories were accurately identified, providing reliable support for earthquake monitoring and artificial blasting discrimination.

[0107] Example 7:

[0108] The basic content is the same as in Example 1, except that in the fourth step, the CTCM-1D-CNN convolutional neural network model performs earthquake event type identification and prediction on the above-mentioned dimensionality-reduced feature matrix. The output of the type of the acquired seismic wave refers to: inputting the above-mentioned dimensionality-reduced feature matrix into the CTCM-1D-CNN convolutional neural network model for prediction, and outputting a predicted label for the earthquake event type; when the output predicted label is 0, the corresponding seismic wave type is determined to be a natural earthquake; when the output predicted label is 1, the corresponding seismic wave type is determined to be an artificial blast, thereby obtaining the specific type of the seismic wave.

[0109] In application, the 1D-CNN basic network model optimized by the CTCM optimization module constitutes the CTCM-1D-CNN convolutional neural network model. After inputting the dimensionality-reduced feature matrix into the model, the model makes predictions based on the learned feature differences between natural earthquakes and artificial blasting, and outputs corresponding prediction labels: when the output label is 0, it corresponds to the typical characteristics of natural earthquake signals such as low-frequency continuous vibration and large shear wave amplitude; when the output label is 1, it corresponds to the significant characteristics of artificial blasting such as high-frequency instantaneous pulse and P-wave dominance. Through this clear label determination rule, the specific type of seismic wave can be directly obtained, providing an automated identification scheme for quickly distinguishing between the two types of events in earthquake monitoring scenarios. This helps to reduce the interference of artificial blasting on the earthquake catalog and improve the accuracy of earthquake rapid reporting and early warning.

[0110] The above description is only a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. Any equivalent modifications or changes made by those skilled in the art based on the content disclosed in the present invention should be included within the scope of protection set forth in the claims.

Claims

1. A deep learning-based seismic wave recognition method, characterized in that: The identification method comprises the following steps: First step: collecting the seismic wave signal generated by the seismic event to obtain original data, then center cutting the original data, and then normalizing the cut original data to obtain a unified signal; Second step: decomposing the unified signal by using the CEEMDAN decomposition method to obtain a plurality of intrinsic mode functions and corresponding residual components; for each intrinsic mode function, first, coarse-grained processing is performed based on a preset time scale parameter, the intrinsic mode functions are divided into a plurality of subsequences according to the corresponding time scale, then a similarity measurement criterion is constructed for each time scale subsequence by using a fuzzy membership function, the fuzzy entropy value of each time scale subsequence is calculated based on the similarity measurement criterion, and then the fuzzy entropy values are integrated to form a feature matrix; Third step: performing dimension reduction processing on the feature matrix by using a probabilistic principal component analysis to obtain a reduced feature matrix; Fourth step: inputting the reduced feature matrix into a CTCM-1D-CNN convolutional neural network model, wherein the CTCM-1D-CNN convolutional neural network model is constructed by optimizing a one-dimensional convolutional neural network by using a tribe competition and member cooperation algorithm; the CTCM-1D-CNN convolutional neural network model performs seismic event type identification and prediction on the reduced feature matrix, and outputs the type of the collected seismic wave, thereby completing the identification of the seismic wave; In the fourth step, the CTCM-1D-CNN convolutional neural network model is constructed by optimizing a one-dimensional convolutional neural network by using a tribe competition and member cooperation algorithm, which means that the hyperparameters of the 1D-CNN basic network model are first encoded into chromosomes, the hyperparameters including a learning rate reduction factor, an initial learning rate, and a minimum batch size; a multi-objective function is designed with the mean absolute percentage error, the correlation coefficient of the predicted value and the theoretical value, and the JS divergence as sub-functions, wherein the mean absolute percentage error corresponds to the coefficient α, the correlation coefficient of the predicted value and the theoretical value corresponds to the coefficient β, and the JS divergence corresponds to the coefficient γ; the three sub-functions are weighted by α, β, and γ, and α+β+γ=1; then the tribe competition and member cooperation algorithm is run, and the hyperparameters are iteratively optimized by simulating the process of member cooperation within the tribe and resource competition between tribes.

2. The seismic wave identification method based on deep learning according to claim 1, wherein: In the first step, the collection of the seismic wave signal generated by the seismic event to obtain the original data means that the acceleration data generated by the seismic event is collected by the seismic monitoring equipment, and the collected acceleration data is taken as the original data.

3. The seismic wave identification method based on deep learning according to claim 2, wherein: In the first step, the center cutting of the original data means that the peak point of the acceleration data waveform in the original data is taken as the center, and the same number of data points of the acceleration data segment is symmetrically cut to the left and right of the peak point.

4. The seismic wave identification method based on deep learning according to claim 1, wherein: In the second step, the signal decomposition of the unified signal by the CEEMDAN decomposition method comprises the following steps: first, determining the parameters required for CEEMDAN decomposition, wherein the parameters required for CEEMDAN decomposition include the standard deviation of noise, the number of noise addition, and the maximum number of iterations allowed for decomposition; and then performing a decomposition operation on the unified signal. The decomposition operation comprises: First, a predetermined number of positive and negative Gaussian white noises with an expected value of zero and a constant standard deviation are added to the unified signal to generate an equal number of noisy signals; the first intrinsic mode function is obtained by performing signal decomposition, component screening, and mean value calculation operations on each noisy signal; and the first-order residual component is obtained by calculating the difference between the unified signal and the first-order intrinsic mode function. The signal decomposition, component screening, and mean value calculation operation comprises the following steps: constructing upper and lower envelope lines by identifying local extreme points in the signal, separating the initial components that meet the conditions of matching the number of extreme points and zero-crossing points and having a local mean value of zero, and taking the arithmetic mean of the initial components of the same order of all noisy signals to obtain the first-order intrinsic mode function. Then, the first-order residual component is added with dynamically adjusted positive and negative Gaussian white noise to generate a new noisy residual signal; the second-order intrinsic mode function is obtained by repeatedly performing the above signal decomposition, component screening, and mean value calculation operations on the noisy residual signal; and the second-order residual component is obtained by calculating the difference between the first-order residual component and the second-order intrinsic mode function. In this way, each order is based on the residual component output by the previous order, and the operations of adding dynamically adjusted noise, generating a noisy residual signal, signal decomposition, component screening, mean value calculation, extracting a new-order intrinsic mode function, and calculating an updated corresponding residual component are repeatedly performed until the residual component is a monotonic function or the number of extreme points is insufficient to separate the components that meet the conditions; and finally, all orders of intrinsic mode functions and corresponding residual components are output.

5. The method of claim 4, wherein: In the second step, for each intrinsic mode function, first perform coarse-grained processing based on a preset time scale parameter to divide the intrinsic mode function into a plurality of sub-sequences according to the corresponding time scale, which comprises the following steps: for each intrinsic mode function obtained by CEEMDAN decomposition, process each intrinsic mode function in order from the first order to the last order; and during processing, perform coarse-grained operation under each time scale on the current order intrinsic mode function according to a plurality of time scale parameters set in advance: In the processing corresponding to each time scale parameter, the original data sequence of the intrinsic mode function is segmented into groups of consecutive data points, a new data point is generated for each group of data points by statistical aggregation operation, and all new data points are arranged in sequence to form a new sequence, i.e., the coarse-grained sub-sequence of the intrinsic mode function under the current time scale is obtained. According to the above coarse-grained operation mode, all preset time scale parameters are sequentially applied to the same order eigenmodal function, so that each order eigenmodal function is decomposed into multiple subsequences, and each subsequence corresponds to a time scale parameter, and reflects the signal characteristics of the order eigenmodal function under different time resolutions. In the second step, the fuzzy membership function is used to construct a similarity measurement criterion for each subsequence under each time scale, and the fuzzy entropy value of each subsequence under each time scale is calculated based on the similarity measurement criterion. The similarity measurement criterion is defined as follows: the similarity between any two data segments in the subsequence is converted into a quantitative membership value by defining the core parameters of the fuzzy membership function. When the membership value reaches a preset threshold, it is determined that the two data segments are similar patterns. Based on the similarity measurement criterion, the fuzzy entropy of each subsequence under each time scale is calculated, and the similarity pattern repetition probability of all data segments in the subsequence is counted. The uncertainty of pattern repetition is converted into a quantitative index through mathematical transformation, and the fuzzy entropy value of the subsequence under the corresponding time scale is obtained.

6. The method according to claim 5, wherein: In the second step, the fuzzy entropy values are integrated to form a feature matrix, which means that the set of fuzzy entropy values corresponding to each order eigenmodal function is taken as a column of the matrix, and the sets of fuzzy entropy values of all order eigenmodal functions are arranged in order to form a two-dimensional matrix with fuzzy entropy values as elements. The row dimension of the feature matrix corresponds to the multi-scale characteristics of the signal, and the column dimension corresponds to the eigenmodal functions of different orders.

7. The method according to claim 1, wherein: In the third step, the feature matrix is reduced in dimension by using probability principal component analysis to obtain a reduced feature matrix, which means that the feature matrix is first standardized to obtain a standardized matrix, then the covariance or correlation coefficient matrix of the standardized matrix is calculated to determine the degree of linear correlation between features, and the covariance or correlation coefficient matrix is decomposed to obtain feature vectors and corresponding eigenvalues, each of which corresponds to a principal component. Then, the ratio of the eigenvalue corresponding to each principal component to the sum of all eigenvalues is calculated to obtain the principal component contribution rate. Finally, the target principal components are selected according to the preset contribution rate threshold and the minimum number of retained components. The standardized matrix is projected into a low-dimensional space composed of the selected target principal components to obtain the reduced feature matrix. 8.The method of claim 1, wherein: In the fourth step, the CTCM-1D-CNN convolutional neural network model includes a 1D-CNN basic network model and a CTCM optimization module. The 1D-CNN basic network model includes one input layer, two convolutional layers, two batch normalization layers, two Relu activation layers, one Dropout layer, one fully connected layer, and one regression layer. The CTCM optimization module takes the learning rate reduction factor, initial learning rate and minimum batch size of the 1D-CNN base network model as the optimization object, and the configuration parameters include population size, maximum evolution generation number, crossover ratio, mutation ratio, mutation rate, reference point number and decision variable number.

9. The method of claim 8, wherein the method is based on deep learning. In the iteration process of the tribe competition and member cooperation algorithm, the 1D-CNN base network model is trained using the stochastic gradient descent momentum method, and the maximum number of training is initially set. The label of the seismic signal is set to 0, and the label of the artificial blasting signal is set to 1. After inputting the training data into the network, the average absolute error, accuracy, average deviation error, determination coefficient and root mean square error of the predicted label and the true label are calculated based on the predicted label and the true label output during the training process, which are used as feedback indicators for evaluating the adaptability of the super parameters of the tribe competition and member cooperation algorithm, and the tribe cooperation strength and competition strategy are dynamically adjusted. When the iteration stopping condition is met, the optimal super parameters obtained by the tribe competition and member cooperation algorithm are used to replace the corresponding super parameters of the 1D-CNN base network model, and the model construction is completed.

10. The method of claim 9, wherein the method is based on deep learning. In the fourth step, the CTCM-1D-CNN convolutional neural network model is used to identify and predict the type of seismic event of the reduced feature matrix, and the type of the collected seismic wave is output. The reduced feature matrix is input into the CTCM-1D-CNN convolutional neural network model for prediction, and the prediction label of the seismic event type is output. When the output prediction label is 0, it is determined that the corresponding seismic wave type is natural earthquake. When the output prediction label is 1, it is determined that the corresponding seismic wave type is artificial blasting, and thus the specific type of the seismic wave is obtained.

Images