Seismic wave recognition method based on deep learning

Abstract

The application relates to a seismic wave recognition method based on deep learning, solves the problem that existing methods rely on fixed base functions or preset parameters to extract features, are difficult to adaptively depict the complex internal structure of non-stationary and time-varying seismic waves, and lead to unstable recognition accuracy and poor generalization ability, and the method comprises the following steps: first, a core signal segment of a seismic wave is intercepted and normalized, and a pure component is obtained through adaptive mode decomposition. Multiscale coarse-grained processing is performed on each component, a fuzzy entropy value is calculated, and a two-dimensional time-frequency feature matrix is constructed. Then, principal component analysis is adopted to reduce the dimension and extract low-dimensional features, and finally, a one-dimensional convolutional neural network is input to complete intelligent recognition. The application has the following effects: through joint representation of multiscale time-frequency features and adaptive feature dimension reduction, the accuracy and reliability of intelligent recognition of seismic waves are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent earthquake event recognition technology, and in particular to a seismic wave recognition method based on deep learning. Background Technology

[0002] Accurately distinguishing between natural earthquakes and non-natural earthquake events such as those caused by artificial blasting is a key technical requirement for improving the accuracy of earthquake monitoring and early warning, as well as the ability to manage engineering vibrations safely. Currently, the commonly used technical approach in this field is to first process seismic wave signals to extract features, and then use classification models to automatically identify the event type.

[0003] In common related technical solutions, fixed signal analysis basis functions or preset decomposition parameters are usually selected to process the signal, and a series of statistical or spectral features are calculated based on these. These features are then fed into a standard classifier with fixed parameters (such as support vector machines, shallow neural networks, etc.) to complete the recognition.

[0004] However, existing methods of this kind share a significant common limitation: the extracted features often fail to comprehensively and adaptively characterize the complex intrinsic structure of seismic wave signals as they change with time and frequency. Due to the significant non-stationarity and event-specific nature of seismic waves, features extracted using fixed analysis patterns are prone to losing key discriminative information or introducing redundancy. This leads to unstable recognition accuracy and limited generalization ability of subsequent classification models in different scenarios, making it difficult to meet the practical needs of high-reliability seismic event identification. Summary of the Invention

[0005] To effectively improve the accuracy and reliability of intelligent seismic wave identification, this application provides a seismic wave identification method based on deep learning.

[0006] This application provides a seismic wave identification method based on deep learning, employing the following technical solution: A deep learning-based seismic wave identification method includes: The raw seismic wave signal is acquired, and the core waveform data segment is symmetrically intercepted with the peak point of the energy concentration interval as the center. Energy normalization processing is then performed to obtain the signal to be analyzed with uniform dimensions. The signal to be analyzed is subjected to fully adaptive noise ensemble empirical mode decomposition. Mode aliasing is suppressed by adaptive injection of positive and negative Gaussian white noise, resulting in pure essential mode components of several order frequencies. For each order component, coarse-grained resampling is performed based on a set of incremental time scale parameters to generate coarse-grained subsequences corresponding to different time resolutions. For each subsequence, the complexity and irregularity of its waveform mode are quantified based on the fuzzy membership function, and the fuzzy entropy value is calculated. The fuzzy entropy values ​​of all components at each time scale are integrated, and a two-dimensional feature matrix is ​​constructed with the component order as the column and the time scale as the row to characterize the frequency domain distribution and time domain multi-scale structure of the seismic wave signal. The two-dimensional feature matrix is ​​reduced by using probabilistic principal component analysis. By maximizing the retained feature variance, low-dimensional feature vectors reflecting the essential differences between signal categories are extracted, resulting in an optimized feature set. The optimized feature set is input into an optimized one-dimensional convolutional neural network classification model for identification, and the classification result is output to complete the intelligent identification of seismic waves.

[0007] By employing the above technical solution, mode aliasing is suppressed through fully adaptive noise ensemble empirical mode decomposition, and multi-scale coarsening and fuzzy entropy quantization are combined to construct a two-dimensional feature matrix that comprehensively characterizes the time-frequency complexity of the signal. After feature purification through probabilistic principal component analysis, accurate classification is achieved by a one-dimensional convolutional neural network with hyperparameter adaptive optimization. The entire process is synergistically optimized, significantly improving the recognition accuracy, stability, and generalization ability of natural earthquakes and artificial blasting. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating a deep learning-based seismic wave identification method according to an embodiment of this application.

[0009] Figure 2 This is a flowchart illustrating the process of acquiring raw seismic wave signals and symmetrically extracting core waveform data segments centered on the peak point of the energy concentration range, according to another embodiment of this application. Detailed Implementation

[0010] The present application will be further described in detail below with reference to the accompanying drawings.

[0011] Reference Figure 1 This application discloses a deep learning-based seismic wave identification method, comprising: Step S1: Acquire the original seismic wave signal, symmetrically extract the core waveform data segment centered on the peak point of the energy concentration interval, and perform energy normalization processing to obtain the dimensionally unified signal to be analyzed.

[0012] The raw seismic wave signal refers to the seismic wave vibration data recorded by seismic instruments (such as seismographs), containing information such as amplitude, frequency, and phase. It is the fundamental data for earthquake research. Acquisition method: Seismographs are deployed at seismic observation stations or exploration sites, and the vibration signals are converted into digitized electrical signals using specialized equipment. Energy normalization processing: The energy of the signal is adjusted to a uniform dimension for easier subsequent analysis. Specifically, the energy value of the signal is divided by its maximum energy value, ensuring that the processed signal energy ranges between 0 and 1.

[0013] The necessary procedures are as follows: 1. Determination of Energy Concentration Range and Peak Point: The energy concentration range is determined through short-time energy calculation. The signal is segmented, the energy of each segment is calculated, and the peak point within the range with the highest energy is found. A core waveform data segment of length 200 is symmetrically extracted using the peak point as the center. For example, the signal can be divided into segments of length 100, the energy of each segment is calculated, the segment with the highest energy is found, the peak point within that segment is determined, and then a core waveform data segment of length 200 is extracted using the peak point as the center.

[0014] 2. Energy Normalization: Calculate the total energy of the signal, divide the value of each sampling point by the square root of the total energy to obtain the normalized signal, ensuring the signal energy range is between 0 and 1. For example, for a seismic wave signal, calculate the sum of the squares of all its sampling points as the total energy, then divide the value of each sampling point by the square root of the total energy to complete the normalization process. Technical Methods and Examples: In practice, the above process can be implemented using professional digital signal processing software (such as MATLAB or related Python tools). These tools allow for convenient calculation of the short-time energy of the signal, determination of peak points, and energy normalization.

[0015] Step S2: Perform fully adaptive noise ensemble empirical mode decomposition on the signal to be analyzed. Adaptively inject positive and negative Gaussian white noise to suppress mode aliasing and obtain essential mode components with pure frequency components of several orders. For each order component, coarse-grained resampling is performed based on a set of incremental time scale parameters to generate coarse-grained subsequences corresponding to different time resolutions. For each subsequence, the complexity and irregularity of its waveform mode are quantified based on the fuzzy membership function, and the fuzzy entropy value is calculated. Integrate the fuzzy entropy values ​​of all components at each time scale, and construct a two-dimensional feature matrix characterizing the frequency domain distribution and time domain multi-scale structure of the seismic wave signal, with the component order as the column and the time scale as the row.

[0016] Among them, the Fully Adaptive Noise Ensemble Empirical Mode Decomposition (EEMD) is an improved Empirical Mode Decomposition (EMD) method that suppresses mode aliasing by adaptively injecting positive and negative Gaussian white noise into the signal, thereby more accurately decomposing the signal into several essential mode components (IMFs). EEMD is an adaptive signal processing technique, particularly suitable for the analysis of nonlinear and non-stationary signals. Essential Mode Components (IMFs): Signal components obtained after EEMD decomposition; each IMF represents a characteristic component of the signal within a specific frequency range, reflecting the local characteristics of the signal. Fuzzy Entropy (FuzzyEn): A method for quantifying the complexity and irregularity of a signal, calculating the fuzzy similarity of a signal sequence based on a fuzzy membership function. The higher the fuzzy entropy value, the more complex and irregular the signal. Two-Dimensional Feature Matrix: A matrix where rows represent different time scales and columns represent essential mode components of different orders. Each element in the matrix is ​​the fuzzy entropy value of the corresponding component at a specific time scale, used to characterize the frequency domain distribution and time-domain multi-scale structural characteristics of the signal.

[0017] The necessary procedures are as follows: 1. EEMD Decomposition Process: First, a fully adaptive ensemble empirical mode decomposition (EEMD) is performed on the seismic wave signal to be analyzed. During this process, positive and negative Gaussian white noise is adaptively injected into the signal. The power spectrum of Gaussian white noise is uniformly distributed across the entire frequency domain. By adding different noise samples multiple times and performing EMD decomposition, the average value is finally taken, thereby suppressing mode aliasing and obtaining several pure intrinsic mode components (IMFs). For example, for a seismic wave signal, EEMD decomposition may yield 5 IMFs, each representing a characteristic component of the signal within a different frequency range.

[0018] 2. Coarse-grained resampling: For each order of IMF, coarse-grained resampling is performed based on a set of incremental time-scale parameters. Specifically, a set of time-scale parameters (such as 1, 2, 4, 8, etc.) is selected, and each order of IMF is resampled to generate coarse-grained subsequences with different time resolutions. For example, for a time-scale parameter of 2, the sampling rate of the IMF is reduced to half of its original value, thereby generating coarse-grained subsequences. These subsequences can capture the characteristic changes of the signal at different time scales.

[0019] 3. Fuzzy Entropy Calculation: For each coarse-grained subsequence, its fuzzy entropy value is calculated based on the fuzzy membership function. The fuzzy membership function is used to define the similarity between signal sequences, and the complexity of the signal is measured by calculating the fuzzy similarity of the signal sequences. For example, setting a similarity threshold of 0.2 and a fuzzy function width of 0.5, the fuzzy entropy value of each coarse-grained subsequence is calculated. The higher the fuzzy entropy value, the more complex and irregular the signal; the lower the fuzzy entropy value, the more regular the signal.

[0020] 4. Construction of the Two-Dimensional Feature Matrix: Integrate the fuzzy entropy values ​​of all components at each time scale, and construct a two-dimensional feature matrix with the component order as columns and the time scale as rows. For example, assuming we have a 5th-order IMF and 4 time scales, the constructed two-dimensional feature matrix will be a 4×5 matrix. Each element in the matrix is ​​the fuzzy entropy value of the corresponding component at a specific time scale. This matrix can characterize the frequency domain distribution and time-domain multi-scale structural characteristics of the seismic wave signal, providing key feature information for subsequent analysis and classification.

[0021] Step S3: Use probabilistic principal component analysis to reduce the dimensionality of the two-dimensional feature matrix. By maximizing the retained feature variance, extract low-dimensional feature vectors that reflect the essential differences between signal categories to obtain the optimized feature set.

[0022] Among them, Probabilistic Principal Component Analysis (PPCA) is a dimensionality reduction technique based on a probabilistic model that extracts the main structure of data by maximizing the variance of preserved features. Compared with traditional Principal Component Analysis (PCA), PPCA provides a more flexible framework, can handle missing data, and can perform more robust feature extraction through probabilistic models. Low-dimensional feature vectors: Feature vectors extracted through dimensionality reduction techniques. These vectors can retain the main information of the original data in a low-dimensional space while removing redundancy and noise, reflecting the essential differences between signal categories. Optimized feature set: The feature set obtained after dimensionality reduction. These features can more effectively represent the category differences of the original data in a low-dimensional space, providing better input for subsequent classification and recognition.

[0023] The specific execution process described above can be found in steps S31 to S34.

[0024] Step S4: Input the optimized feature set into the optimized one-dimensional convolutional neural network classification model for recognition, output the classification result, and complete the intelligent recognition of seismic waves.

[0025] One-dimensional convolutional neural networks (1DCNNs) are deep learning models specifically designed for processing sequential data. They extract local features of signals through one-dimensional convolutional layers, reduce feature dimensionality through pooling layers, and finally perform classification through fully connected layers. 1DCNNs excel at processing time-series data (such as seismic wave signals). Optimized feature set: The feature set obtained after dimensionality reduction. These features can more effectively represent the class differences of the original data in a lower-dimensional space, providing better input for subsequent classification and recognition. Classification result: The class label of the seismic wave signal output by the neural network model, used to indicate the earthquake type or other relevant category information corresponding to the signal.

[0026] The necessary process and key technical details are described below: 1. Construction of a One-Dimensional Convolutional Neural Network (1DCNN): In this step, the optimized feature set is input into an optimized one-dimensional convolutional neural network (1DCNN) classification model for recognition. The construction process of 1DCNN is as follows: Input layer: Receives the optimized feature set as input. The shape of the input data is typically (number of samples, feature length, number of channels). For seismic wave signals, the number of channels is usually 1, indicating a single-channel signal.

[0027] Convolutional layers: These extract local features through one-dimensional convolution operations. The convolution kernel slides across the time series, and the result of the convolution operation is calculated. For example, using a convolution kernel of size 3 with a stride of 1 can extract local feature patterns.

[0028] Activation function: The ReLU (Rectified Linear Unit) activation function is usually used to introduce non-linearity into the model, enabling the network to learn complex feature representations.

[0029] Pooling layers: These reduce feature dimensionality by using max pooling or average pooling, thereby lowering computational complexity while preserving important features. For example, using a max pooling layer of size 2 can halve the feature dimensionality.

[0030] Fully connected layer: Features extracted from convolutional and pooling layers are flattened and input into the fully connected layer for final classification. The output dimension of the fully connected layer is equal to the number of classes. The softmax function is used to convert the output into a probability distribution to obtain the predicted probability of each class.

[0031] 2. Model optimization and training: To improve the model's performance, the 1DCNN needs to be optimized and trained: Optimizer selection: Use the Adam optimizer, which combines the advantages of RMSprop and Momentum, and can automatically adjust the learning rate to accelerate model convergence.

[0032] Loss function: The cross-entropy loss function is used to measure the difference between the probability distribution of the model output and the true label, guiding the training direction of the model.

[0033] Training process: The optimized feature set is divided into a training set and a validation set. The model is trained using the training set and its performance is evaluated using the validation set. During training, hyperparameters such as the learning rate and batch size are adjusted to optimize the model's performance.

[0034] Regularization techniques: To avoid overfitting, Dropout can be used to randomly discard a portion of the neuron outputs during training, thereby enhancing the model's generalization ability.

[0035] 3. Classification Output and Intelligent Recognition: The trained 1DCNN model can classify the optimized input feature set and output the classification results. For example, assuming we have two types of seismic wave signals (seismic signals and non-seismic signals), the model's classification output will indicate which type each input sample belongs to. In this way, intelligent recognition of seismic wave signals can be achieved, providing strong support for earthquake monitoring and early warning.

[0036] Reference Figure 2 The raw seismic wave signal was acquired, and the core waveform data segment was obtained by symmetrically truncating the data around the peak point of the energy concentration range. Step S11: Locate the data point with the largest absolute amplitude value from the collected raw seismic wave signal as the peak point.

[0037] The necessary procedures are as follows: 1. Signal preprocessing: The acquired raw seismic wave signals are preprocessed, such as by removing noise and filtering, to improve signal quality. This step can be accomplished using common signal processing tools.

[0038] 2. Locating the peak point: Calculate the absolute value of the amplitude of each data point in the signal. Traverse the signal, find the data point with the largest absolute amplitude, and record its index position. This index position is the peak point.

[0039] Step S12: Using the peak point as the center, extract data points in the directions of the beginning and end of the original signal respectively. When the extraction operation in any direction reaches the beginning or end of the original signal, the extraction in that direction is stopped, thus forming an initial data segment.

[0040] The necessary procedures are as follows: 1. Determine the peak point location: Based on the result of step S11, the index location of the peak point, peak_index, is known.

[0041] 2. Forward Truncation: Starting from the peak point peak_index, truncate data points one by one towards the beginning of the original signal (in the direction of smaller indices). When the truncation operation reaches the beginning of the signal (index 0), stop truncation.

[0042] 3. Truncation: Starting from the peak point peak_index, truncate data points one by one towards the end of the original signal (in the direction with the larger index). When the truncation operation reaches the end of the signal (index N-1, where N is the total length of the signal), stop truncation.

[0043] 4. Forming the initial data segment: Merge the data points captured from the forward and backward directions to form an initial data segment containing the peak points. The length of the initial data segment depends on the total length of the signal and the location of the peak points.

[0044] Step S13: Calculate the difference between the total number of data points in the initial data segment and a preset fixed length L.

[0045] The necessary procedures are as follows: 1. Determine the length of the initial data segment: Calculate the total number of data points contained in the initial data segment. Assume the initial data segment is `initial_segment`, and its length is N. initial .

[0046] 2. Preset fixed length L: Based on actual needs, preset a fixed length L. This length is the target length that the initial data segment is expected to reach in subsequent processing.

[0047] 3. Calculate the difference: Calculate the length N of the initial data segment. initial The difference Δ between the original length L and the preset fixed length L: Δ = LN initial .

[0048] Step S14: If the difference is zero, the initial data segment is directly used as the core waveform data segment.

[0049] Step S15: If the difference is greater than zero, add the corresponding number of zero-value data points at the front end, back end or both ends of the initial data segment until its total length is equal to the fixed length L, thereby forming the core waveform data segment.

[0050] The necessary procedures are as follows: 1. Determine the difference: Based on the result of step S13, the length N of the initial data segment is known. initial The difference Δ between the initial data segment and the preset fixed length L. If Δ > 0, it means that the length of the initial data segment is less than the preset length L.

[0051] 2. Choose a padding strategy: Based on actual needs, choose to add zero-value data points at the beginning, end, or both ends of the initial data segment. Common padding strategies include: Front-end padding: Add Δ zero-value data points to the front of the initial data segment.

[0052] Back-end padding: Add Δ zero-value data points to the back of the initial data segment.

[0053] Padding at both ends: Add padding to the front and back of the initial data segment. There are zero-value data points (if Δ is odd, an additional zero-value point is added to one end).

[0054] 3. Add zero-value data points: According to the selected filling strategy, add zero-value data points at the corresponding positions until the total length of the initial data segment reaches the preset fixed length L.

[0055] 4. Forming the core waveform data segment: After filling, the length of the initial data segment becomes L, forming the final core waveform data segment.

[0056] For each component, coarse-grained resampling is performed based on a set of incremental time-scale parameters to generate coarse-grained subsequences corresponding to different time resolutions, including: Step S21: For each essential modal component, obtain its corresponding time series and perform spectral analysis to calculate its power spectrum; in the power spectrum, expand from the highest energy frequency point to both sides until the cumulative energy reaches the preset first proportion threshold of the total energy, and determine this frequency range as the dominant frequency band of the component.

[0057] The necessary procedures are as follows: 1. Obtaining Time Series: For each Essential Mode Component (IMF), extract its corresponding time series. These time series are obtained from the EEMD decomposition in step S2. Each IMF is an independent time series, reflecting the characteristics of the signal within a specific frequency range.

[0058] 2. Perform spectrum analysis: Perform spectrum analysis on the time series of each IMF to calculate its power spectrum. Spectrum analysis can be performed using Fast Fourier Transform (FFT) or other spectrum analysis methods to obtain the energy distribution at each frequency point.

[0059] 3. Determine the dominant frequency band: Find the frequency point with the highest energy in the power spectrum, denoted as f. max From f max Expand to both sides and calculate the cumulative energy until the cumulative energy reaches a preset first proportion threshold (e.g., 90%) of the total energy. The frequency range that meets the above conditions is determined as the dominant frequency band of this order component.

[0060] Step S22: Calculate the weighted average frequency of the dominant frequency band as the representative frequency, and determine an initial time scale parameter according to the representative frequency and a preset mapping rule. The higher the representative frequency, the smaller the value of the initial time scale parameter.

[0061] The necessary procedures are as follows: 1. Calculate the weighted average frequency: Within the dominant frequency band, calculate the energy-weighted average frequency at each frequency point. Assume the dominant frequency band is [f...]. min ,f max If the power spectrum is P(f), then the weighted average frequency f is... avg It can be calculated using the following formula: This formula represents the weighted average frequency obtained by multiplying the energy of each frequency point within the dominant frequency band by its frequency value and then dividing by the total energy.

[0062] 2. Determine the initial time scale parameters: based on the weighted average frequency f avg The initial time scale parameter T0 is determined according to a preset mapping rule. The mapping rule is typically a simple inverse relationship, meaning that the higher the frequency, the smaller the initial time scale parameter. For example, the following mapping rule can be used: Where C is a preset constant used to adjust the range of time scale parameters.

[0063] Step S23: Based on the initial time scale parameter, the parameters are multiplied sequentially according to a preset scaling factor to generate an increasing sequence of time scale parameters.

[0064] The necessary procedures are as follows: 1. Determine the initial time scale parameter: Based on the results of step S22, the initial time scale parameter T0 of each IMF is known.

[0065] 2. Preset scaling factor: A preset scaling factor α is used to control the growth rate of the time scale parameter. Typically, the scaling factor α is greater than 1 to ensure that the time scale parameter increases sequentially.

[0066] 3. Generate a timescale parameter sequence: Using the initial timescale parameter T0 as a reference, the sequence is multiplied sequentially according to a preset scaling factor α to generate an increasing timescale parameter sequence {T1, T2, ..., T...}. n The specific calculation formula is as follows: Where i = 1, 2, ..., n. Here, n is the number of preset time scale parameters, determined according to actual needs.

[0067] Step S24: For each parameter in the incremental time scale parameter sequence, the corresponding time series is divided into continuous and non-overlapping data segments according to the parameter value. The arithmetic mean of all data points in each data segment is used to obtain the aggregate value, and all aggregate values ​​are arranged in the original order to form a coarse-grained subsequence.

[0068] The necessary procedures are as follows: 1. Select a time scale parameter: Choose a parameter T from the increasing time scale parameter sequence. i .

[0069] 2. Divide the data into segments: based on the time scale parameter T i The time series is divided into continuous, non-overlapping data segments. The length of each data segment is equal to T. i If the length of the time series cannot be increased by T... i Integer division, the length of the last data segment may be less than T. i .

[0070] 3. Calculate the aggregate value: Calculate the arithmetic mean of all data points within each data segment to obtain the aggregate value for that data segment. The specific formula is: Aggregate value j = Where j is the index of the data segment, and x... k It is the kth data point in the time series.

[0071] 4. Constructing a coarse-grained subsequence: Arrange all aggregated values ​​in their original order to form a coarse-grained subsequence. This subsequence reflects the average characteristics of the time series at that time scale.

[0072] Step S25: For the same essential modal component, traverse all parameters in the incremental time scale parameter sequence, repeat the above division and aggregation operations, and generate a set of coarse-grained subsequences.

[0073] The specific process is as follows: 1. Determine the time scale parameter sequence: Based on the result of step S23, the incremental time scale parameter sequence {T1,T2,…,T} is known. n}

[0074] 2. Traversing the time-scale parameters: For each essential mode component (IMF), traverse each parameter T in the time-scale parameter sequence in turn. i .

[0075] 3. Repeated partitioning and aggregation operations: For each time scale parameter T i Repeat the partitioning and aggregation operations in step S24 to generate a coarse-grained subsequence. The specific steps are as follows: According to T i Divide the time series into continuous and non-overlapping data segments.

[0076] The aggregate value is obtained by taking the arithmetic mean of all data points within each data segment.

[0077] Arrange all aggregated values ​​in their original order to form a coarse-grained subsequence.

[0078] 4. Generate a set of coarse-grained subsequences: After traversing all time scale parameters, a set of coarse-grained subsequences is generated. Each subsequence corresponds to a time scale parameter, reflecting the characteristics of the time series at different time scales.

[0079] For each subsequence, the complexity and irregularity of its waveform pattern are quantified based on the fuzzy membership function, and the fuzzy entropy value is calculated, including: Step S2a: For each coarse-grained subsequence, based on the time scale parameter corresponding to the subsequence, divide the preset first proportional constant by the time scale parameter to obtain the similarity tolerance parameter used for calculating the fuzzy entropy of the subsequence; at the same time, set the embedding dimension to the preset second constant.

[0080] The necessary procedures are as follows: 1. Select coarse-grained subsequence: For each essential mode component (IMF) generated coarse-grained subsequence, select one subsequence for processing.

[0081] 2. Determine the time scale parameter: Obtain the time scale parameter T corresponding to the coarse-grained subsequence. i .

[0082] 3. Calculate the similarity tolerance parameter: A pre-defined scaling constant C1 (the first scaling constant) is used. The similarity tolerance parameter r is then calculated. The similarity tolerance parameter r is used to control the allowable range of similarity in fuzzy entropy calculation.

[0083] Step S2b: Based on the embedding dimension, reconstruct the current coarse-grained subsequence into a series of multi-dimensional vectors.

[0084] The necessary procedures are as follows: 1. Determine the embedding dimension: Based on the result of step S2a, the embedding dimension m is known.

[0085] 2. Reconstructing multidimensional vectors: For the current coarse-grained subsequence {x1, x2, ..., x...} N}, where N is the length of the subsequence. The subsequence is reconstructed into a series of multidimensional vectors using an embedding dimension m. Each vector has a dimension of m, and the vectors are constructed as follows: X i =[x i ,x i+1 ,…,x i+m-1 ], where i is the starting index of the vector, satisfying i=1,2,…,N-m+1.

[0086] 3. Generate a multidimensional vector sequence: Using the above method, generate a series of multidimensional vectors {X1, X2, ..., X...} N-m+1}

[0087] Each vector Xi contains m consecutive data points, reflecting the characteristics of the subsequence within different time windows.

[0088] Step S2c: Calculate the distance between every two multidimensional vectors, and use a preset Gaussian fuzzy membership function to map each distance value to a similarity metric between 0 and 1; the width parameter of the Gaussian function is proportional to the product of the variance of the current subsequence and the preset third coefficient.

[0089] The necessary procedures are as follows: 1. Calculate the distance between multidimensional vectors: For the reconstructed multidimensional vector sequence {X1, X2, ..., X...} N-m+1}, calculate the Euclidean distance between every two vectors.

[0090] The Euclidean distance formula is: Among them, X i and X j There are two multidimensional vectors, where m is the embedding dimension.

[0091] 2. Calculate the variance of the current subsequence: Calculate the variance σ of the current coarse-grained subsequence. 2 : ,in, It is the mean of the subsequence.

[0092] 3. Determine the width parameter of the Gaussian fuzzy membership function: Preset a coefficient C2 (third coefficient). The width parameter σf of the Gaussian fuzzy membership function is directly proportional to the product of the variance of the current subsequence and the preset third coefficient. .

[0093] 4. Calculate the similarity metric: Use a Gaussian fuzzy membership function to map the distance values ​​to a similarity metric between 0 and 1. The Gaussian function formula is: Where d is the distance between the two vectors, and σ f It is the width parameter.

[0094] In step S2d, for each multidimensional vector, count the number of times the similarity metric of the vector with other vectors exceeds a preset threshold, and calculate the proportion of this number to the total number of comparisons, as the local matching probability of the vector.

[0095] The necessary procedures are as follows: 1. Preset similarity threshold: Preset a similarity threshold θ to determine whether two vectors are sufficiently similar. Typically, this threshold is a value between 0 and 1, such as 0.5.

[0096] 2. Calculate the local matching probability for each vector: For each multidimensional vector X i Calculate its relationship with all other vectors X. j ( The similarity metric μ(d(X)) i ,X j The number of times the threshold θ is exceeded.

[0097] Calculate the local matching probability P of this vector. i This refers to the proportion of times the similarity metric value exceeds the threshold out of the total number of comparisons. .in, It is an indicator function that takes the value 1 when the condition is true and 0 otherwise.

[0098] In step S2e, firstly, the first global average similarity is calculated based on the embedding dimension; then, the embedding dimension is increased by a preset fixed increment, and the above vector reconstruction, distance calculation, similarity mapping and local matching probability calculation process is repeated to obtain the second global average similarity.

[0099] The necessary procedures are as follows: 1. Calculate the first global average similarity: Based on the current embedding dimension m, calculate the local matching probability {P1, P2, ..., P} of all multidimensional vectors. N-m+1 The first global average similarity Φ(m) is the average of all local matching probabilities: .

[0100] 2. Increase the embedding dimension: Increase the embedding dimension m by a preset fixed increment Δm. Typically, Δm = 1. The new embedding dimension is m′ = m + Δm.

[0101] 3. Repeated Vector Reconstruction, Distance Calculation, Similarity Mapping, and Local Matching Probability Calculation: Using a new embedding dimension m′, vector reconstruction, distance calculation, similarity mapping, and local matching probability calculation are performed again. A new local matching probability sequence {P1′, P2′, ..., P} is generated. N-m′+1′}

[0102] 4. Calculate the second global average similarity: The second global average similarity Φ(m′) is the average of the new local matching probabilities: .

[0103] Step S2f: The natural logarithm of the ratio of the first global average similarity to the second global average similarity is used as the fuzzy entropy value of the subsequence.

[0104] The necessary procedures are as follows: 1. Calculate the global average similarity ratio: Based on the results of step S2e, the first global average similarity Φ(m) and the second global average similarity Φ(m′) are known.

[0105] Calculate the ratio of these two global average similarities: Ratio = .

[0106] 2. Calculate the natural logarithm: Calculate the natural logarithm of the above ratios to obtain the fuzzy entropy value FuzzyEn: .

[0107] Step S2g: Traverse all coarse-grained subsequences, repeat the above process, obtain the fuzzy entropy values ​​of each subsequence, and form a set.

[0108] The necessary procedures are as follows: 1. Initialize the fuzzy entropy value set: Create an empty set or list to store the fuzzy entropy values ​​of all coarse-grained subsequences.

[0109] 2. Traverse all coarse-grained subsequences: For each order of essential mode component (IMF), process all coarse-grained subsequences generated in sequence.

[0110] 3. Repeat the fuzzy entropy calculation process: For each coarse-grained subsequence, repeat steps S2a to S2f to calculate its fuzzy entropy value. Specific steps include: Calculate the similarity tolerance parameter and set the embedding dimension (step S2a).

[0111] Reconstruct the multidimensional vector (step S2b).

[0112] Calculate the distance between vectors and map it to a similarity metric (step S2c).

[0113] Calculate the local matching probability (step S2d).

[0114] Calculate the first and second global average similarities (step S2e).

[0115] Calculate the fuzzy entropy value (step S2f).

[0116] 4. Store fuzzy entropy values: Store the fuzzy entropy value of each coarse-grained subsequence into an initialized set or list.

[0117] 5. Forming a set of fuzzy entropy values: After traversing all coarse-grained subsequences, a set containing the fuzzy entropy values ​​of all subsequences is obtained.

[0118] By integrating the fuzzy entropy values ​​of all components at various time scales, and using the component order as columns and the time scale as rows, a two-dimensional feature matrix characterizing the frequency domain distribution and time domain multi-scale structure of the seismic wave signal is constructed, including: Step S2A: Based on the total number of essential modal components obtained from the fully adaptive noise set empirical mode decomposition and the total number of incremental time scale parameters used, construct a corresponding two-dimensional data structure.

[0119] The specific process is as follows: 1. Determine the total number of essential modal components: Determine the total number M of the obtained essential modal components (IMFs) through fully adaptive ensemble noise empirical mode decomposition (EEMD). This means that the signal is decomposed into M different frequency components.

[0120] 2. Determine the total number of time scale parameters: Determine the total number N of incremental time scale parameters used. This means that for each essential modal component, we calculated N fuzzy entropy values ​​at different time scales.

[0121] 3. Construct a two-dimensional data structure: Based on M and N, construct an N×M two-dimensional data structure (usually a matrix). The number of rows in this matrix equals the total number of time-scale parameters N, and the number of columns equals the total number of essential modal components M.

[0122] Each element Matrix(i,j) will store the fuzzy entropy value of the j-th essential modal component at the i-th time scale.

[0123] In step S2B, all the fuzzy entropy values ​​calculated for each essential modal component at different time scales are sequentially mapped and filled into the two-dimensional data structure according to the order of their corresponding component order and time scale parameters, so that the data in each row corresponds to the complexity of different frequency components at the same time scale, and the data in each column corresponds to the irregular changes of the same frequency component at different time scales.

[0124] The necessary procedures are as follows: 1. Prepare fuzzy entropy data: Collect fuzzy entropy values ​​of all essential modal components (IMFs) calculated at different time scales. These fuzzy entropy values ​​have been calculated in the previous steps (such as steps S2a to S2f).

[0125] 2. Determine the mapping order: Determine the position of each fuzzy entropy value in the two-dimensional data structure. The position of each fuzzy entropy value is determined by the order (column index) and time scale parameter (row index) of its corresponding essential modal component.

[0126] 3. Fill in the two-dimensional data structure: Following the mapping order, each fuzzy entropy value is filled into the two-dimensional data structure. The specific operation is as follows: For the j-th essential modal component (j=1,2,…,M); For the i-th time scale parameter (i=1,2,…,N); The fuzzy entropy value of the j-th essential modal component at the i-th time scale is FuzzyEn. j,i Fill the i-th row and j-th column of the two-dimensional data structure: Matrix(i,j)=FuzzyEn j,i .

[0127] 4. Check data integrity: Ensure that all fuzzy entropy values ​​have been correctly filled into the two-dimensional data structure, without any omissions or errors.

[0128] In step S2C, after all data is filled, a complete feature matrix is ​​obtained. The row direction of this matrix represents the structural characteristic spectrum of the signal under multiple time resolutions, and the column direction represents the irregular distribution of the signal in multiple frequency domain components, thus forming a structured feature representation that can jointly reflect the comprehensive time-frequency domain characteristics of the seismic wave signal.

[0129] The necessary procedures are as follows: 1. Check data integrity: Ensure that all locations in the two-dimensional data structure have been correctly filled with fuzzy entropy values, without any omissions or errors.

[0130] 2. Forming a complete feature matrix: The two-dimensional data structure now constitutes a complete feature matrix. The number of rows in this matrix equals the total number of time-scale parameters N, and the number of columns equals the total number of essential modal components M.

[0131] 3. Explain the meaning of the characteristic matrix: Row direction: The data in each row corresponds to the complexity of different frequency components at the same time scale. This reflects the structural characteristic spectrum of the signal at multiple time resolutions.

[0132] Column direction: The data in each column corresponds to the irregular changes of the same frequency component at different time scales. This reflects the irregular distribution of the signal across multiple frequency domain components.

[0133] 4. Summarize the role of the feature matrix: This feature matrix jointly reflects the comprehensive characteristics of seismic wave signals in the time and frequency domains, providing a structured feature representation for subsequent signal analysis and classification.

[0134] All fuzzy entropy values ​​calculated for each essential modal component at different time scales are sequentially mapped and filled into a two-dimensional data structure according to their corresponding component order and time scale parameters, including: Step S2B1: Construct the column coordinate axes of the feature matrix in the order of the essential modal components. The order from low to high strictly corresponds to the physical distribution of frequency components in the signal from high to low, thus forming an ordered index characterizing the frequency domain characteristics of the signal in the column dimension.

[0135] The necessary procedures are as follows: 1. Determine the order of essential modal components: Determine the total number M of the obtained essential modal components (IMFs) through fully adaptive noise set empirical mode decomposition (EEMD).

[0136] They are arranged in ascending order according to the intrinsic modal components. Generally, lower-order IMFs correspond to higher-frequency components in the signal, and higher-order IMFs correspond to lower-frequency components.

[0137] 2. Construct the column coordinate axes of the feature matrix: Construct the column coordinate axes of the feature matrix so that each column corresponds to the order of an essential modal component.

[0138] The column coordinate axes are ordered as follows: column 1 corresponds to the first-order IMF, column 2 corresponds to the second-order IMF, and so on, until column M corresponds to the M-order IMF.

[0139] This arrangement strictly corresponds to the physical distribution of frequency components in the signal from high to low.

[0140] 3. Form an ordered index representing the frequency domain characteristics of the signal: Each column of the column coordinate axis forms an ordered index, characterizing the signal's properties in the frequency domain. In this way, the column dimensions of the feature matrix can reflect the signal's characteristics at different frequency components.

[0141] Step S2B2: Construct the row coordinate axes of the feature matrix in order of the numerical values ​​of the time scale parameters. The arrangement of parameters from small to large strictly corresponds to the physical change from fine to coarse in the time resolution of the signal analysis, thus forming an ordered index in the row dimension that characterizes the multi-scale characteristics of the signal in the time domain.

[0142] The necessary procedures are as follows: 1. Determine the order of timescale parameters: Determine the total number N of incremental timescale parameters to be used. Arrange the timescale parameters from smallest to largest according to their numerical value. Generally, smaller timescale parameters correspond to higher time resolution (more refined analysis), and larger timescale parameters correspond to lower time resolution (coarser analysis).

[0143] 2. Constructing the row axes of the feature matrix: Construct the row axes of the feature matrix so that each row corresponds to a time scale parameter. The order of the row axes is as follows: the first row corresponds to the smallest time scale parameter, the second row corresponds to the second smallest time scale parameter, and so on, until the Nth row corresponds to the largest time scale parameter. This arrangement strictly corresponds to the physical change in the time resolution of signal analysis from fine to coarse.

[0144] 3. Forming an ordered index that characterizes the multi-scale characteristics of the signal in the time domain: Each row of the row coordinate axis forms an ordered index that characterizes the signal's characteristics in the multi-scale time domain.

[0145] In this way, the row dimension of the feature matrix can reflect the characteristics of the signal at different time resolutions.

[0146] Step S2B3: Based on the rules determined by the constructed row and column coordinate axes, perform coordinate positioning for each fuzzy entropy value to be filled: determine its position on the column coordinate axis according to the order of the essential modal component from which the fuzzy entropy value originates; determine its position on the row coordinate axis according to the time scale parameters used in the calculation of the fuzzy entropy value.

[0147] The necessary procedures are as follows: 1. Determine the source of the fuzzy entropy value: For each fuzzy entropy value to be filled, determine its source: The order of the essential modal component (IMF): determines which order of IMF the fuzzy entropy value comes from.

[0148] Time scale parameter: Determines under which the fuzzy entropy value was calculated.

[0149] 2. Determine the position on the column coordinate axis: The position of the fuzzy entropy value on the column coordinate axis is determined based on the IMF order from which it originates.

[0150] For example, if the fuzzy entropy value comes from the j-th order IMF, then its column coordinate is j.

[0151] 3. Determine the position on the row coordinate axis: The position of the value on the row coordinate axis is determined based on the time scale parameter used in the fuzzy entropy value calculation.

[0152] For example, if the fuzzy entropy value is calculated under the i-th time scale parameter, then its row coordinate is i.

[0153] 4. Perform coordinate positioning: Position the fuzzy entropy value to the corresponding cell in the feature matrix, i.e., the i-th row and j-th column.

[0154] Step S2B4: Fill each fuzzy entropy value into a matrix cell uniquely specified by its row and column coordinates, thereby completing the one-to-one mapping from physical features to matrix elements.

[0155] The necessary procedures are as follows: 1. Prepare the feature matrix: Ensure that the feature matrix has been constructed according to steps S2B1 and S2B2, and that the row and column axes have been determined.

[0156] 2. Fill in the fuzzy entropy values: For each fuzzy entropy value FuzzyEn j,i Based on its corresponding row coordinate i and column coordinate j, fill it into the corresponding cell of the feature matrix.

[0157] The specific steps are as follows: Matrix(i,j)=FuzzyEn j,i .

[0158] 3. Check data integrity: Ensure that all fuzzy entropy values ​​have been correctly filled into the feature matrix, without omissions or errors.

[0159] 4. Complete the mapping: Through the above filling operation, complete the one-to-one mapping from physical features (fuzzy entropy values) to matrix elements.

[0160] Principal component analysis (PCA) is used to reduce the dimensionality of the two-dimensional feature matrix. By maximizing the retained feature variance, low-dimensional feature vectors reflecting the essential differences between signal categories are extracted, resulting in an optimized feature set including: Step S31: Based on the preset signal-to-noise ratio level of the time-frequency components corresponding to each feature column, assign preset weight coefficients to each column of the two-dimensional feature matrix and perform weighted standardization processing.

[0161] The necessary procedures are as follows: 1. Determine the signal-to-noise ratio (SNR) level of each feature column: Analyze each column (i.e., each feature) of the two-dimensional feature matrix and determine the SNR level of its corresponding time-frequency component. The SNR level reflects the ratio of signal components to noise components in the feature and can usually be determined based on domain knowledge or prior analysis.

[0162] 2. Assign weight coefficients: Based on the signal-to-noise ratio level of each feature column, assign a preset weight coefficient w to each column. j The principle for allocating weight coefficients is: feature columns with high signal-to-noise ratios (SNR) are assigned larger weights, and feature columns with low SNRs are assigned smaller weights. Weight coefficients can be predefined or determined through an optimization method.

[0163] 3. Perform weighted standardization: Perform weighted standardization on each column of the two-dimensional feature matrix. The purpose of standardization is to ensure that all feature columns have the same scale, thus enabling comparability in subsequent analyses. The weighted standardization formula is: Where Matrix(i,j) is the element in the i-th row and j-th column of the original feature matrix. j σ is the mean of the j-th column. j is the standard deviation of the j-th column. j It is the weight coefficient of the j-th column.

[0164] Step S32: Calculate the weighted covariance matrix based on the weighted standardized data, and obtain each principal component and its corresponding weighted variance value through eigenvalue decomposition.

[0165] The specific process is as follows: 1. Calculate the weighted covariance matrix: Calculate the weighted covariance matrix using the weighted and standardized data. The formula for calculating the weighted covariance matrix is: Among them, X weighted It is the weighted and standardized feature matrix (size N×M, where N is the number of samples and M is the number of features). It is X weighted The transpose of . C is an M×M weighted covariance matrix.

[0166] 2. Eigenvalue decomposition: Perform eigenvalue decomposition on the weighted covariance matrix C to obtain eigenvalues ​​and eigenvectors.

[0167] The eigenvalue decomposition formula is: .

[0168] Where, λ i It is the i-th eigenvalue, v i It is the i-th eigenvector, with eigenvalue λ. i This represents the weighted variance of the corresponding principal component.

[0169] 3. Sort the eigenvalues ​​and eigenvectors: Arrange the eigenvalues ​​in descending order and the corresponding eigenvectors in the same order.

[0170] Eigenvectors with larger eigenvalues ​​correspond to the main direction of variation (i.e., principal components).

[0171] Step S33: Calculate the weighted variance contribution rate of each principal component, and perform preliminary screening of principal components based on the preset first contribution rate threshold; for the initially screened principal components, calculate the preset correlation threshold between their direction vectors and the high-weight feature set, retain the principal components that meet the correlation condition, and form the final target principal component subset.

[0172] The necessary procedures are as follows: 1. Calculate the weighted variance contribution rate: For each principal component, calculate its weighted variance contribution rate. The weighted variance contribution rate is the proportion of the principal component's eigenvalue to the sum of all eigenvalues. Contribution rate i = , where λ i is the eigenvalue of the i-th principal component, and M is the total number of principal components.

[0173] 2. Preliminary screening of principal components: Based on a preset first contribution rate threshold (e.g., 85%), select principal components whose cumulative contribution rate reaches this threshold. The specific steps are as follows: Principal components are sorted from largest to smallest based on their eigenvalues.

[0174] The contribution rate of the principal components is calculated cumulatively until the cumulative contribution rate reaches or exceeds a preset threshold.

[0175] These principal components were retained as preliminary screening results.

[0176] 3. Calculate the correlation between the direction vector and the high-weight feature set: For the initially selected principal components, calculate the correlation between their direction vectors and the high-weight feature set. The high-weight feature set refers to the feature columns with larger weight coefficients.

[0177] The correlation can be calculated using the Pearson correlation coefficient or other correlation measures.

[0178] 4. Retain principal components that meet the correlation criteria: Based on a preset correlation threshold (e.g., 0.7), principal components with a correlation higher than the threshold to the high-weight feature set are retained.

[0179] These retained principal components constitute the final target principal component subset.

[0180] Step S34: Project the weighted and standardized feature matrix onto a low-dimensional space spanned by the target principal component subsets to obtain the optimized feature set.

[0181] The specific process is as follows: 1. Determine the target principal component subset: Based on the results of step S33, determine the final target principal component subset to be retained. The direction vectors of these principal components will be used to construct the low-dimensional space.

[0182] 2. Construct the projection matrix: A projection matrix P is formed by combining the direction vectors of the target principal component subsets. Each column of the projection matrix is ​​a direction vector of a target principal component.

[0183] Suppose that the target principal component subset contains K principal components, and the size of the projection matrix P is M×K, where M is the number of columns in the original feature matrix.

[0184] 3. Perform the projection operation: The weighted and standardized feature matrix X weighted Projecting to a lower-dimensional space. The projection operation is performed using matrix multiplication: X projected =X weighted P. Wherein, X projected X is the projected low-dimensional feature matrix of size N×K, where N is the number of samples and K is the number of target principal components. weighted P is the weighted and standardized feature matrix, with a size of N×M. P is the projection matrix, with a size of M×K.

[0185] 4. Obtain the optimized feature set: the projected low-dimensional feature matrix X projected This is the optimized feature set. This low-dimensional feature set retains the main structural features of the original data while removing redundancy and noise, making it suitable for subsequent analysis and classification tasks.

[0186] For the initially selected principal components, calculate the preset correlation threshold between their direction vectors and the high-weight feature set, and retain the principal components that meet this correlation condition to form the final target principal component subset, which includes: Step S331: Sort all weight coefficients from largest to smallest, and select the feature columns corresponding to the top K weight coefficients to form a high-weight feature set.

[0187] The specific process is as follows: 1. Determine the weighting coefficients: Collect the weighting coefficients corresponding to all feature columns. These weighting coefficients have already been assigned according to the signal-to-noise ratio level of each feature column in step S31.

[0188] 2. Sort by weight coefficients: Sort all weight coefficients from largest to smallest value. The purpose of sorting is to determine which feature columns contribute the most to the signal.

[0189] 3. Select high-weight feature columns: Based on a preset quantity K, select the feature columns corresponding to the top K weight coefficients. These feature columns constitute the high-weight feature set. K is a preset parameter that can be determined based on actual needs or domain knowledge.

[0190] 4. Form a high-weight feature set: Combine the selected K feature columns into a high-weight feature set. This feature set will be used for subsequent correlation calculations.

[0191] Step S332: For each principal component that has been preliminarily selected, calculate the projection energy of its direction vector on the dimension corresponding to the high-weight feature set, and use the ratio of the projection energy to the energy of the entire direction vector as the correlation measure between the principal component and the high-weight feature set.

[0192] The necessary procedures are as follows: 1. Determine the high-weight feature set: Based on the results of step S331, the feature columns contained in the high-weight feature set and their corresponding weight coefficients are known.

[0193] 2. Calculate the projected energy of the direction vector: For each initially selected principal component, calculate the projected energy of its direction vector on the dimension corresponding to the high-weight feature set.

[0194] Assuming the direction vector of the principal component is v, and the feature column indices of the high-weight feature set are {j1,j2,…,j...} K}, then the projected energy E proj for: Among them, v jk It is the direction vector v at the j-th k Components on each feature column.

[0195] 3. Calculate the energy of the entire direction vector: Calculate the total energy E of the principal component direction vectors. total : Where M is the total number of columns in the feature matrix, v j It is the component of the direction vector v on the j-th feature column.

[0196] 4. Calculate the correlation metric: Calculate the ratio of projected energy to total energy, which serves as a correlation measure between the principal component and the high-weight feature set: Correlation metric = .

[0197] Step S333: Compare the correlation metric of each principal component with a preset second correlation threshold, and retain only the principal components whose correlation metric is greater than or equal to the threshold to form the final target principal component subset.

[0198] The necessary procedures are as follows: 1. Determine the preset second relevance threshold: A preset relevance threshold θ2 is used to determine whether the relevance between the principal component and the high-weight feature set is high enough. This threshold is usually determined based on actual needs or domain knowledge, for example, it can be set to 0.7 or 0.8.

[0199] 2. Compare correlation metrics: For each initially selected principal component, compare its correlation metric with the preset second correlation threshold θ2.

[0200] 3. Retain the principal components that meet the conditions: If the correlation measure of a principal component is greater than or equal to θ2, then the principal component is retained.

[0201] If the correlation measure is less than θ2, then discard the principal component.

[0202] 4. Form the final target principal component subset: All principal components that meet the criteria are combined into a final target subset of principal components. These principal components will be used for subsequent feature extraction and analysis.

[0203] The optimized one-dimensional convolutional neural network classification model is constructed through the following steps: Step SA00: Preset the key structural parameters and their value ranges of the one-dimensional convolutional neural network to form the basic architecture search space; encode each possible configuration of the key structural parameters as an individual in the tribal competition and member cooperation algorithm, and initialize the algorithm running parameters.

[0204] The necessary procedures are as follows: 1. Determine the key structural parameters and their value ranges: Number of convolutional layers: Preset range for the number of convolutional layers, such as 1 to 5 layers.

[0205] Kernel size: Preset range for kernel size, such as 3, 5, 7, etc.

[0206] Number of convolutional kernels: Preset range of the number of convolutional kernels per layer, such as 16, 32, 64, etc.

[0207] Pooling layer type: Preset pooling layer type, such as max pooling or average pooling.

[0208] Pooling layer size: Preset pooling layer size range, such as 2, 3, etc.

[0209] Number of fully connected layers: Preset range for the number of fully connected layers, such as 1 to 3 layers.

[0210] Number of neurons in fully connected layers: Preset range for the number of neurons in each fully connected layer, such as 128, 256, 512, etc.

[0211] Activation function: Preset activation function type, such as ReLU, LeakyReLU, etc.

[0212] Learning rate: Preset learning rate range, for example, 0.001 to 0.01.

[0213] 2. Construct the basic infrastructure search space: The aforementioned key structural parameters and their value ranges are combined to form a basic architecture search space. Each possible configuration combination of each parameter constitutes a candidate network architecture.

[0214] 3. Encoding as an individual in the tribal competition and member cooperation algorithm: Each candidate network architecture is encoded as an individual in the tribal competition and member cooperation algorithm. Each individual represents a possible network structure configuration.

[0215] 4. Initialize algorithm running parameters: Initialize the running parameters for the tribal competition and member cooperation algorithm, including: Number of tribes: The preset number of tribes, for example, 5 tribes.

[0216] Number of tribe members: The preset number of members for each tribe, for example, 10 members per tribe.

[0217] Maximum number of iterations: The maximum number of iterations for the preset algorithm, for example, 100 times.

[0218] Convergence condition: The convergence condition of the algorithm is preset, such as the average fitness no longer significantly improving in 30 consecutive iterations.

[0219] Step SB00 uses historical seismic wave data and corresponding labels as the training and validation sets, and the classification accuracy of the one-dimensional convolutional neural network on the validation set as the fitness function. It runs a tribal competition and member cooperation algorithm. The algorithm simulates the iterative process of internal tribal cooperation and resource competition and inter-tribal resource competition, dynamically optimizes within the search space, and continuously evaluates and updates the hyperparameter combination according to the fitness function until the preset convergence condition is met or the maximum number of iterations is reached. Finally, it outputs the hyperparameter combination that achieves the optimal classification accuracy.

[0220] The specific process is as follows: 1. Prepare the dataset: Use historical seismic wave data and their corresponding labels as the training and validation sets. These datasets will be used to train and evaluate the performance of the one-dimensional convolutional neural network.

[0221] 2. Define the fitness function: The classification accuracy of the one-dimensional convolutional neural network on the validation set is used as the fitness function. The higher the classification accuracy, the larger the value of the fitness function, indicating better model performance.

[0222] 3. Initialize the tribal competition and member cooperation algorithm: Based on the infrastructure search space defined in step SA00, each network architecture configuration is encoded as an individual in the algorithm.

[0223] The initialization parameters of the algorithm include the number of tribes, the number of tribe members, the maximum number of iterations, and the convergence criteria.

[0224] 4. Run the algorithm: The algorithm dynamically optimizes within the search space by simulating the iterative process of cooperation and sharing within tribes and resource competition between tribes.

[0225] In each iteration, the algorithm evaluates the performance of each individual (i.e., the network structure configuration) based on the fitness function and updates the hyperparameter combination.

[0226] 5. Evaluation and Update of Hyperparameter Combinations: The algorithm continuously evaluates the performance of each hyperparameter combination based on the fitness function and updates the tribe members according to the evaluation results. If a member of a tribe fails to significantly improve its fitness in multiple iterations, it may be replaced by a better member from another tribe to promote global optimization.

[0227] 6. Check convergence conditions: The algorithm checks whether preset convergence conditions are met, such as the average fitness no longer significantly improving after multiple consecutive iterations. If the convergence conditions are met or the maximum number of iterations is reached, the algorithm terminates.

[0228] 7. Output the optimal hyperparameter combination: Finally, the algorithm outputs the hyperparameter combination that achieves the optimal classification accuracy. These hyperparameters will be used to configure the final one-dimensional convolutional neural network classification model.

[0229] Step SC00: Based on the optimal combination of hyperparameters, configure and construct an optimized one-dimensional convolutional neural network classification model.

[0230] The specific process is as follows: 1. Extract the optimal hyperparameter combination: Based on the results of step SB00, extract the optimal hyperparameter combination output by the tribal competition and member cooperation algorithm. These hyperparameters include the number of convolutional layers, kernel size, pooling layer type and size, number of fully connected layers and neurons, activation function type, learning rate, etc.

[0231] 2. Configure a one-dimensional convolutional neural network: Configure the structure of a one-dimensional convolutional neural network using the optimal combination of hyperparameters. Specific steps include: Constructing convolutional layers: Construct convolutional layers based on the number of convolutional layers, kernel size, and number of convolutional kernels in the optimal hyperparameter combination.

[0232] Add activation functions: Add a specified activation function, such as ReLU or LeakyReLU, to each convolutional layer.

[0233] Add pooling layers: Add pooling layers based on the pooling layer type and size in the optimal hyperparameter combination.

[0234] Constructing fully connected layers: Construct fully connected layers based on the number of fully connected layers and neurons in the optimal combination of hyperparameters.

[0235] Configure learning rate: Set the learning rate during network training.

[0236] 3. Initialize network weights: Initialize the weights of the configured one-dimensional convolutional neural network. Random initialization methods, such as Xavier initialization or He initialization, are typically used.

[0237] 4. Construct a classification model: The configured network structure is combined with the training and validation sets to build a complete classification model. The model is ensured to receive historical seismic wave data as input and output classification results.

[0238] 5. Verify model performance: Use a validation set to validate the performance of the constructed classification model. Ensure that the model's classification accuracy on the validation set reaches the performance level corresponding to the optimal hyperparameter combination.

[0239] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A seismic wave identification method based on deep learning, characterized in that, include: The raw seismic wave signal is acquired, and the core waveform data segment is symmetrically intercepted with the peak point of the energy concentration interval as the center. Energy normalization processing is then performed to obtain the signal to be analyzed with uniform dimensions. The signal to be analyzed is subjected to fully adaptive noise set empirical mode decomposition. Mode aliasing is suppressed by adaptively injecting positive and negative Gaussian white noise, and pure essential mode components with several order frequency components are obtained. For each component, coarse-grained resampling is performed based on a set of incremental time scale parameters to generate coarse-grained subsequences with different time resolutions. For each subsequence, the complexity and irregularity of its waveform pattern are quantified based on the fuzzy membership function, and the fuzzy entropy value is calculated. The fuzzy entropy values ​​of all components at each time scale are integrated, and a two-dimensional feature matrix is ​​constructed with the component order as the column and the time scale as the row to characterize the frequency domain distribution and time domain multi-scale structure of the seismic wave signal. The two-dimensional feature matrix is ​​reduced by using probabilistic principal component analysis. By maximizing the retained feature variance, low-dimensional feature vectors reflecting the essential differences between signal categories are extracted, resulting in an optimized feature set. The optimized feature set is input into an optimized one-dimensional convolutional neural network classification model for identification, and the classification result is output to complete the intelligent identification of seismic waves. For each component, coarse-grained resampling is performed based on a set of incremental time-scale parameters to generate coarse-grained subsequences corresponding to different time resolutions, including: For each essential modal component, its corresponding time series is obtained and spectral analysis is performed to calculate its power spectrum. In the power spectrum, the frequency range is determined as the dominant frequency band of the component by expanding from the highest energy frequency point to both sides until the cumulative energy reaches the preset first proportion threshold of the total energy. The weighted average frequency of the dominant frequency band is calculated as the representative frequency, and an initial time scale parameter is determined according to the representative frequency and a preset mapping rule. The higher the representative frequency, the smaller the value of the initial time scale parameter. Based on the initial time scale parameter, the parameters are multiplied sequentially according to a preset scaling factor to generate an increasing sequence of time scale parameters. For each parameter in the increasing time scale parameter sequence, the corresponding time series is divided into continuous and non-overlapping data segments according to the parameter value. The arithmetic mean of all data points in each data segment is used to obtain the aggregate value, and all aggregate values ​​are arranged in the original order to form a coarse-grained subsequence. For the same essential modal component, traverse all parameters in the increasing time-scale parameter sequence, repeat the above division and aggregation operations, and generate a set of coarse-grained subsequences; Principal component analysis (PCA) is used to reduce the dimensionality of the two-dimensional feature matrix. By maximizing the retained feature variance, low-dimensional feature vectors reflecting the essential differences between signal categories are extracted, resulting in an optimized feature set including: Based on the preset signal-to-noise ratio level of the time-frequency components corresponding to each feature column, preset weight coefficients are assigned to each column of the two-dimensional feature matrix, and weighted standardization is performed. The weighted covariance matrix is ​​calculated based on the weighted standardized data, and the principal components and their corresponding weighted variance values ​​are obtained through eigenvalue decomposition. Calculate the weighted variance contribution rate of each principal component, and perform preliminary screening of principal components based on the preset first contribution rate threshold; for the initially screened principal components, calculate the preset correlation threshold between their direction vectors and the high-weight feature set, retain the principal components that meet the correlation condition, and form the final target principal component subset. The weighted and standardized feature matrix is ​​projected onto a low-dimensional space spanned by the target principal component subsets to obtain the optimized feature set; The optimized one-dimensional convolutional neural network classification model is constructed through the following steps: The key structural parameters and their value ranges of a one-dimensional convolutional neural network are preset to form the basic architecture search space; each possible configuration of the key structural parameters is encoded as an individual in the tribal competition and member cooperation algorithm, and the algorithm running parameters are initialized. Using historical seismic wave data and corresponding labels as training and validation sets, and the classification accuracy of a one-dimensional convolutional neural network on the validation set as the fitness function, the algorithm runs a tribal competition and member cooperation algorithm. The algorithm dynamically optimizes within the search space by simulating the iterative process of internal cooperation and sharing and inter-tribal resource competition, and continuously evaluates and updates the hyperparameter combination according to the fitness function until the preset convergence condition is met or the maximum number of iterations is reached. Finally, it outputs the hyperparameter combination that achieves the optimal classification accuracy. Based on the optimal combination of hyperparameters, an optimized one-dimensional convolutional neural network classification model is configured and constructed.

2. The seismic wave identification method based on deep learning according to claim 1, characterized in that, The raw seismic wave signal was acquired, and symmetrically truncated around the peak point of the energy concentration range to obtain the core waveform data segment, which includes: The peak point is identified by locating the data point with the largest absolute amplitude from the collected raw seismic wave signal. Centered on the peak point, data points are extracted towards the beginning and end of the original signal respectively. When the extraction operation in any direction reaches the beginning or end of the original signal, the extraction in that direction is stopped, thus forming an initial data segment. Calculate the difference between the total number of data points in the initial data segment and a preset fixed length L; If the difference is zero, the initial data segment will be used directly as the core waveform data segment. If the difference is greater than zero, then a corresponding number of zero-value data points are added simultaneously at the front, back, or both ends of the initial data segment until its total length equals the fixed length L, thus forming the core waveform data segment.

3. The seismic wave identification method based on deep learning according to claim 1, characterized in that, For each subsequence, the complexity and irregularity of its waveform pattern are quantified based on the fuzzy membership function, and the fuzzy entropy value is calculated, including: For each coarse-grained subsequence, based on the time scale parameter corresponding to the subsequence, the preset first proportional constant is divided by the time scale parameter to obtain the similarity tolerance parameter used for calculating the fuzzy entropy of the subsequence; at the same time, the embedding dimension is set to a preset second constant. Based on the embedding dimension, the current coarse-grained subsequence is reconstructed into a series of multi-dimensional vectors; Calculate the distance between every two multidimensional vectors, and use a preset Gaussian fuzzy membership function to map each distance value to a similarity metric between 0 and 1; the width parameter of the Gaussian function is directly proportional to the product of the variance of the current subsequence and the preset third coefficient. For each multidimensional vector, count the number of times the similarity metric between the vector and other vectors exceeds a preset threshold, and calculate the proportion of this number to the total number of comparisons, as the local matching probability of the vector; First, the first global average similarity is calculated based on the embedding dimension. Then, the embedding dimension is increased by a preset fixed increment, and the above vector reconstruction, distance calculation, similarity mapping and local matching probability calculation process is repeated to obtain the second global average similarity. The natural logarithm of the ratio of the first global average similarity to the second global average similarity is taken as the fuzzy entropy value of the subsequence; Repeat the above process to traverse all coarse-grained subsequences and obtain the fuzzy entropy values ​​of each subsequence, forming a set.

4. The seismic wave identification method based on deep learning according to claim 1, characterized in that, By integrating the fuzzy entropy values ​​of all components at various time scales, and using the component order as columns and the time scale as rows, a two-dimensional feature matrix characterizing the frequency domain distribution and time domain multi-scale structure of the seismic wave signal is constructed, including: Based on the total number of essential modal components obtained from the fully adaptive noise set empirical mode decomposition and the total number of incremental time scale parameters used, a corresponding two-dimensional data structure is constructed. All fuzzy entropy values ​​calculated for each essential modal component at different time scales are sequentially mapped and filled into a two-dimensional data structure according to the order of their corresponding component order and time scale parameters. This makes the data in each row correspond to the complexity of different frequency components at the same time scale, and the data in each column correspond to the irregular changes of the same frequency component at different time scales. After all data is filled in, a complete feature matrix is ​​obtained. The row direction of this matrix represents the structural characteristic spectrum of the signal at multiple time resolutions, and the column direction represents the irregular distribution of the signal in multiple frequency domain components. Thus, a structured feature representation that can jointly reflect the comprehensive time-frequency domain characteristics of the seismic wave signal is formed.

5. The seismic wave identification method based on deep learning according to claim 4, characterized in that, All fuzzy entropy values ​​calculated for each essential modal component at different time scales are sequentially mapped and filled into a two-dimensional data structure according to their corresponding component order and time scale parameters, including: The column coordinate axes of the feature matrix are constructed according to the order of the essential modal components. The order from low to high strictly corresponds to the physical distribution of frequency components in the signal from high to low, thus forming an ordered index characterizing the frequency domain characteristics of the signal in the column dimension. The row axes of the feature matrix are constructed in order of the numerical magnitude of the time scale parameters. The arrangement of the parameters from small to large strictly corresponds to the physical change from fine to coarse in the time resolution of the signal analysis, thus forming an ordered index in the row dimension that characterizes the multi-scale characteristics of the signal in the time domain. Based on the rules defined by the constructed row and column coordinate axes, coordinate positioning is performed for each fuzzy entropy value to be filled: its position on the column coordinate axis is determined according to the order of the essential modal component from which the fuzzy entropy value originates; its position on the row coordinate axis is determined according to the time scale parameters used in the calculation of the fuzzy entropy value. Each fuzzy entropy value is filled into a matrix cell uniquely specified by its row and column coordinates, thus completing a one-to-one mapping from physical features to matrix elements.

6. The seismic wave identification method based on deep learning according to claim 1, characterized in that, For the initially selected principal components, calculate the preset correlation threshold between their direction vectors and the high-weight feature set, and retain the principal components that meet this correlation condition to form the final target principal component subset, which includes: Sort all weight coefficients from largest to smallest, and select the feature columns corresponding to the top K weight coefficients to form a high-weight feature set. For each principal component that has been initially selected, calculate the projection energy of its direction vector on the dimension corresponding to the high-weight feature set, and use the ratio of the projection energy to the energy of the entire direction vector as the correlation measure between the principal component and the high-weight feature set. The correlation metric of each principal component is compared with a preset second correlation threshold, and only principal components with correlation metrics greater than or equal to the threshold are retained to form the final target principal component subset.

Images