Seismic first arrival picking method and device based on multi-input attention mechanism network

Abstract

The application discloses a method and equipment for picking up a seismic first arrival wave based on a multi-input attention mechanism network. The application realizes automatic picking up of the first arrival by fusing multi-domain data and the advantages of various networks. Compared with manual picking up, the efficiency is greatly improved and the result is more objective. Compared with a traditional automatic picking up method, the accuracy and stability are higher under different signal-to-noise ratios. Meanwhile, after hyperparameter optimization, the generalization ability is strong, reliable data can be provided for subsequent seismic processing, the advantage is prominent under complex geological conditions, and the method has great significance for geological analysis and resource exploration.

Description

Technical Field

[0001] This invention relates to the field of earthquake data processing technology, specifically to a method and device for picking first arrival waves of earthquakes based on a multi-input attention mechanism network. Background Technology

[0002] In seismic data processing, first-arrival travel time data is crucial for near-surface inversion and forms the foundation for subsequent analyses. While manual first-arrival acquisition, a traditional method, maintains high accuracy, it is not only time-consuming and labor-intensive, but also highly susceptible to the influence of interpreters' experience. With the continuous expansion of seismic exploration operations, the number of coverage events has increased dramatically, and the volume of seismic data has exploded, making the shortcomings of manual acquisition methods in terms of efficiency and accuracy increasingly prominent.

[0003] Existing semi-automatic or automatic first arrival (IHA) techniques have many areas for improvement. IHA methods based on the cross-correlation characteristics of adjacent traces are significantly affected by the selection of the standard trace, making it difficult to guarantee IHA accuracy in seismic data processing scenarios with low signal-to-noise ratios (SNR). The Short-to-Long Time Window Energy Ratio (STA / LTA) method, with its energy-based algorithm, exhibits good noise resistance and automatic IHA performance; however, the window length setting greatly influences IHA accuracy, leading to the development of various improved methods based on multiple time windows. The Akaike Information Criterion (AIC) method determines first arrivals by exploiting the differences in statistical characteristics between seismic signals and noise, but its IHA results are often unsatisfactory in low SNR environments. While fractal dimension algorithms possess some noise resistance, they neglect the similarities between seismic traces. Furthermore, methods that transform signals to other domains or spaces to highlight the differences between signal and noise, such as τ-p transform and wavelet transform, while showing some advantages in IHA picking across different trace sets, still struggle to achieve high-precision IHA picking for complex seismic data.

[0004] With the rise of deep learning, various network models have been introduced into the field of earthquake first arrival picking. However, current first arrival picking efforts are generally limited to the time-space domain or a specific transform domain, failing to fully explore the characteristic value of data across multiple domains. In fact, the amplitude, phase, and frequency characteristics of the data are all key elements for identifying first arrivals. The lack of identification of these characteristics will lead to inaccurate and insufficient picking of seismic wave first arrivals. Summary of the Invention

[0005] Based on the problems raised in the background technology above, the purpose of this invention is to provide a method and device for seismic first arrival picking based on a multi-input attention mechanism network. The fused features after being fused by the coordinate attention module are passed through two convolutional layers and activation functions in the output module to obtain the final first arrival position prediction result. This overcomes the current limitation of seismic first arrival picking to a specific domain, which ignores the characteristic value of seismic data in multiple domains.

[0006] This invention is achieved through the following technical solution: The first aspect of this invention provides a method for picking first arrival waves of earthquakes based on a multi-input attention mechanism network, comprising the following steps: Step S1: Acquire seismic data and preprocess the seismic data to obtain preprocessed seismic data; the preprocessed seismic data includes time-space domain seismic data, instantaneous amplitude, frequency, and phase characteristics; Step S2: Construct a multi-input attention mechanism network architecture, which includes four branch modules, a coordinate attention module, and an output layer; Step S3: The time-space domain seismic data, instantaneous amplitude, frequency, and phase features are processed using the four branch modules respectively to obtain four multi-domain features; Step S4: Use the coordinate attention module to perform coordinate attention processing on the four multi-domain features to obtain fused features, and output the fused features through the output layer to obtain the picking result.

[0007] In the above technical solution, seismic data from different regions and under different geological conditions are first collected. This data covers various complex geological structures and noisy environments. Since the collected seismic data suffers from noise interference and large amplitude differences, preprocessing is used to denoise the data, removing random noise and anomalous signals. Simultaneously, a normalization method is employed to unify the amplitude of the seismic data to a specific range, thereby improving data quality and providing a reliable data foundation for subsequent network training.

[0008] Then, a multi-input attention mechanism network architecture is constructed. Specifically, the multi-input attention mechanism network architecture adopts the MI-Attn-PrecisePickNet network. The MI-Attn-PrecisePickNet network is used for automatic first arrival picking of seismic signals, deeply mining the potential information of seismic data from multiple dimensions to improve the accuracy and reliability of first arrival picking. This multi-input attention network architecture includes four branch modules, which are used to process four types of two-dimensional input data: time-space domain seismic data, instantaneous amplitude, frequency, and phase characteristics, respectively, to obtain four multi-domain features. This allows for the acquisition of the time-space characteristics of seismic signals from different perspectives, providing a foundation for subsequent feature extraction and analysis.

[0009] After processing by four branch modules, the four types of two-dimensional input data yield features from different perspectives. In order to make full use of these multi-domain features, this method applies different attention to them through the coordinate attention module to achieve feature fusion. At the same time, the features are recalibrated to highlight important features and suppress secondary features.

[0010] The fused features obtained by the coordinate attention module are then processed by two convolutional layers and activation functions in the output module to obtain the final first arrival location prediction result. This overcomes the current limitation of earthquake first arrival picking to a specific domain, which ignores the feature value of earthquake data in multiple domains.

[0011] In one optional embodiment, the branching module includes: an m-layer residual module and an n-layer residual long short-term memory module; Specifically, the time-space domain seismic data, the instantaneous amplitude, the frequency, and the phase characteristics are used as two-dimensional input data. The m-layer residual module is used to process the two-dimensional input data to capture the spatial characteristics of the seismic signal. The spatial features of the seismic signal are converted into a time series, and the n-layer residual long short-term memory module is used to capture the features of the time series to obtain dynamic features in the time dimension.

[0012] In one optional embodiment, the two-dimensional input data is processed using the m-layer residual module, including:

[0013] In the above formula, For the residual module The output of the layer, Representing the Input data for the layer residual module, For the first Nonlinear transformation of layer residual modules, For the first Layer residual module parameters.

[0014] In one optional embodiment, the time series is used to capture features using the n-layer residual long short-term memory module, including:

[0015] In the above formula, For the first Layer in The hidden state of the output after residual connection at any given moment. For LSTM networks, For the first Layer residual long short-term memory module in The input sequence at time step.

[0016] In an optional embodiment, the coordinate attention module is used to perform coordinate attention processing on four multi-domain features, including: The four multi-domain features are concatenated to obtain the concatenated features. Coordinate attention is calculated on the spliced ​​features to obtain the fused features; the formula for calculating coordinate attention is as follows:

[0017] In the above formula, and Typically, these are indices in the height and width directions of the feature map; and This represents the result of global pooling in the horizontal and vertical directions. It's about height. It is the width; It is the ReLU activation function. yes Convolutional layer; and It is The result of the splitting is feature maps in the horizontal and vertical directions; It is the Sigmoid activation function. and It is used to generate attention weights in the horizontal and vertical directions. Convolutional layer; This indicates element-wise multiplication; This is a splicing feature.

[0018] In one optional embodiment, training the multi-input attention mechanism network architecture includes: Samples are selected from the earthquake data as a training set, and the training set is input into the multi-input attention mechanism network architecture; The multi-input attention mechanism network architecture is trained using binary cross-entropy as the loss function and an adaptive matrix derived from backpropagation; wherein the loss function is expressed as follows:

[0019] In the above formula, is the loss value, y is the binary label, which takes the value 0 or 1; p(y) represents the probability that the output belongs to the label y; This represents the total number of training samples.

[0020] In one optional embodiment, after obtaining the picking result, the method further includes: Obtain the test set labels, compare the picking results with the test set labels, and obtain comparison data; The comparison data is marked with pixel precision to obtain the correct pixel percentage; the calculation process for marking the comparison data with pixel precision is as follows:

[0021] In the above formula, For the correct pixel ratio, This represents the number of pixels of category i that are inferred to be category i. This represents all pixels.

[0022] A second aspect of the present invention provides a seismic first arrival picking system based on a multi-input attention mechanism network, comprising: The data processing module is used to acquire seismic data and preprocess the seismic data to obtain preprocessed seismic data; the preprocessed seismic data includes time-space domain seismic data, instantaneous amplitude, frequency, and phase characteristics; An architecture building module is used to build a multi-input attention mechanism network architecture, which includes four branch modules, a coordinate attention module, and an output layer. The branch feature module is used to process the time-space domain seismic data, the instantaneous amplitude, the frequency, and the phase features using the four branch modules respectively, to obtain four multi-domain features; The feature fusion module is used to perform coordinate attention processing on four multi-domain features using the coordinate attention module to obtain fused features, and output the fused features through the output layer.

[0023] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a seismic first arrival wave picking method based on a multi-input attention mechanism network.

[0024] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for picking up first arrival waves of earthquakes based on a multi-input attention mechanism network.

[0025] Compared with the prior art, the present invention has the following advantages and beneficial effects: This invention integrates multi-domain data and the advantages of various networks to achieve automatic arrival picking, significantly improving efficiency and providing more objective results compared to manual picking. Compared to traditional automatic picking methods, it offers higher accuracy and stability under different signal-to-noise ratios. Through hyperparameter optimization, it exhibits strong generalization ability, providing reliable data for subsequent seismic processing. Its advantages are particularly pronounced under complex geological conditions, making it significant for geological analysis and resource exploration. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is a schematic diagram of the multi-input attention mechanism network architecture provided in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of a hyperparameter experiment provided in Embodiment 1 of the present invention; wherein, Figure 2 (a) is a schematic diagram of the accuracy curves for different learning rates. Figure 2 (b) is a schematic diagram of the accuracy curves for different kernel matrix sizes. Figure 2 (c) is a schematic diagram of the accuracy curves for different batch sizes. Figure 2 (d) is a schematic diagram of the accuracy curves for different model depths; Figure 3 This is a comparison diagram of the initial arrival picking results provided in Embodiment 1 of the present invention; wherein, Figure 3 (a) is the original earthquake record diagram. Figure 3 (b) Comparison chart of picking results using different methods; Figure 4 This is a comparison chart of absolute errors provided in Embodiment 1 of the present invention; Figure 5 The projection of the first arrival time provided in Embodiment 1 of the present invention onto seismic data (a), instantaneous amplitude (b), frequency characteristic profile (c), and phase characteristic profile (d); Figure 6 This is a partial enlarged view of the seismic trace provided in Embodiment 1 of the present invention; wherein, Figure 6 (a) is a magnified view of a local seismic trace from the original seismic record. Figure 6 (b) is a magnified view of a local seismic trace with instantaneous amplitude. Figure 6 (c) is a magnified view of the seismic traces with frequency characteristics. Figure 6 (d) is a magnified view of the seismic trace with phase characteristics. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0028] Example 1 Embodiment 1 of the present invention provides a method for picking first arrival waves of earthquakes based on a multi-input attention mechanism network. The method includes the following steps: Step S1: Acquire seismic data and preprocess the seismic data to obtain preprocessed seismic data; the preprocessed seismic data includes time-space domain seismic data, instantaneous amplitude, frequency, and phase characteristics; Step S2: Construct a multi-input attention mechanism network architecture, which includes four branch modules, a coordinate attention module, and an output layer; Step S3: The time-space domain seismic data, instantaneous amplitude, frequency, and phase features are processed using the four branch modules respectively to obtain four multi-domain features; Step S4: Use the coordinate attention module to perform coordinate attention processing on the four multi-domain features to obtain fused features, and output the fused features through the output layer to obtain the picking result.

[0029] It should be noted that current technologies for first arrival acquisition are generally limited to the time-space domain or a specific transform domain, failing to fully explore the characteristic value of data across multiple domains. In fact, the amplitude, phase, and frequency of data are all key elements for first arrival identification. The lack of identification of these characteristics will lead to inaccurate and insufficient acquisition of seismic wave first arrivals.

[0030] This embodiment proposes a seismic first-arrival picking method based on a multi-input attention mechanism network. This method first collects seismic data from different regions and under varying geological conditions, encompassing complex geological structures and noisy environments. Due to noise interference and large amplitude variations in the acquired seismic data, preprocessing is used to denoise the data, removing random noise and anomalous signals. Simultaneously, a normalization method is employed to unify the amplitude of the seismic data to a specific range, improving data quality and providing a reliable data foundation for subsequent network training.

[0031] Then, a multi-input attention mechanism network architecture is constructed. Specifically, the multi-input attention mechanism network architecture adopts the MI-Attn-PrecisePickNet network. The MI-Attn-PrecisePickNet network is used for automatic first arrival picking of seismic signals, deeply mining the potential information of seismic data from multiple dimensions to improve the accuracy and reliability of first arrival picking. This multi-input attention network architecture includes four branch modules, which are used to process four types of two-dimensional input data: time-space domain seismic data, instantaneous amplitude, frequency, and phase characteristics, respectively, to obtain four multi-domain features. This allows for the acquisition of the time-space characteristics of seismic signals from different perspectives, providing a foundation for subsequent feature extraction and analysis.

[0032] After processing by four branch modules, the four types of two-dimensional input data yield features from different perspectives. In order to make full use of these multi-domain features, this method applies different attention to them through the coordinate attention module to achieve feature fusion. At the same time, the features are recalibrated to highlight important features and suppress secondary features.

[0033] The fused features obtained by the coordinate attention module are then processed by two convolutional layers and activation functions in the output module to obtain the final first arrival location prediction result. This overcomes the current limitation of earthquake first arrival picking to a specific domain, which ignores the feature value of earthquake data in multiple domains.

[0034] Furthermore, seismic data preprocessing also includes: using the Hilbert-Huang transform to calculate seismic data and obtain three-dimensional seismic profiles. These calculations can reveal various characteristics of the seismic data, specifically: Analytical signals of seismic signals It can be represented as:

[0035] in, The real part of the complex traces represents the actual seismic record traces; is the imaginary part of the complex number path, and is the orthogonal part of the real number path.

[0036] Instantaneous amplitude is a measure of reflection intensity and can be expressed as:

[0037] Instantaneous phase is a measure of the continuity of in-phase axes on a seismic profile:

[0038] The instantaneous frequency is the rate of change of phase over time.

[0039] Significant changes in the three types of instantaneous information at the same location can reveal details that cannot be observed in the original time domain information. Since the first arrival is the first strong amplitude signal received by the detector in the seismic record, amplitude is its primary characteristic. However, seismic data may also contain noise with frequencies and phases similar to the effective signal. To suppress this noise, the instantaneous frequency and instantaneous phase are constrained using amplitude (i.e., seismic data) to obtain phase characteristic data and frequency characteristic data.

[0040]

[0041] In one optional embodiment, the branching module includes: an m-layer residual module and an n-layer residual long short-term memory module; Specifically, the time-space domain seismic data, the instantaneous amplitude, the frequency, and the phase characteristics are used as two-dimensional input data. The m-layer residual module is used to process the two-dimensional input data to capture the spatial characteristics of the seismic signal. The spatial features of the seismic signal are converted into a time series, and the n-layer residual long short-term memory module is used to capture the features of the time series to obtain dynamic features in the time dimension.

[0042] It should be noted that the purpose of constructing the MI-Attn-PrecisePickNet network architecture is to automatically pick first arrivals of seismic signals and to deeply mine the potential information in seismic data from multiple dimensions, thereby improving the accuracy and reliability of first arrival picking. The MI-Attn-PrecisePickNet network architecture is as follows: Figure 1 As shown, it includes a branch module, a coordinate attention module, and an output module. The branch module (dual residual feature learning module) comprises an m-layer residual module (ResNet) and an n-layer long short-term memory module with residual structures (residual LSTM). These four branch modules process four types of two-dimensional input data: time-space domain seismic data, instantaneous amplitude, and frequency and phase features. This allows for the acquisition of time-space features of seismic signals from different perspectives, providing a foundation for subsequent feature extraction and analysis.

[0043] In this embodiment, the different branches of the constructed MI-Attn-PrecisePickNet network architecture share weights, which can effectively reduce the number of training parameters; the multi-feature data input network after Hilbert-Huang transform assigns different weights to amplitude, frequency, and phase features respectively by the attention module; the network structure is flexible and can be adapted to tasks of different complexity by adjusting the number of layers of the feature extraction unit.

[0044] Specifically, the residual module directly transmits the original 2D seismic trace data to the output end through a shortcut path, fully preserving spatial information such as amplitude and phase; the convolution kernel slides to perceive local spatial relationships, maintain spatial order and position, and capture multi-scale features; at the same time, the network learns residual information, integrates original spatial information and convolution features, and retains key details such as geological structure boundaries, thereby achieving effective preservation of seismic trace spatial information.

[0045] In one optional embodiment, the two-dimensional input data is processed using the m-layer residual module, including:

[0046] In the above formula, For the residual module The output of the layer, Representing the Input data for the layer residual module, For the first Nonlinear transformation of layer residual modules, For the first Layer residual module parameters.

[0047] Among them, the The nonlinear transformation of the layer residual module includes convolution and activation operations.

[0048] After processing by the m-layer residual module, the output is converted into a sequence suitable for processing by the residual long short-term memory module. The residual long short-term memory module, with its unique gating mechanism (input gate, forget gate, output gate), can effectively process time series data. By selectively retaining or forgetting historical information, it accurately captures the waveform, amplitude, and phase changes at different times in the input data, thereby uncovering the dynamic patterns of the data in the time dimension.

[0049] In one optional embodiment, the time series is used to capture features using the n-layer residual long short-term memory module, including:

[0050] In the above formula, For the first Layer in The hidden state of the output after residual connection at any given moment. For LSTM networks, For the first Layer residual long short-term memory module in The input sequence at time step.

[0051] Among them, the Layer in Output at time That is, the entire layer residual LSTM module in The output result at each moment, and the final output result Converted to two-dimensional form This is used for subsequent feature fusion and analysis.

[0052] The features after splicing are .

[0053] In an optional embodiment, the coordinate attention module is used to perform coordinate attention processing on four multi-domain features, including: The four multi-domain features are concatenated to obtain the concatenated features. Coordinate attention is calculated on the spliced ​​features to obtain the fused features; the formula for calculating coordinate attention is as follows:

[0054] In the above formula, and Typically, these are indices in the height and width directions of the feature map; and This represents the result of global pooling in the horizontal and vertical directions. It's about height. It is the width; It is the ReLU activation function. yes Convolutional layer; and It is The result of the splitting is feature maps in the horizontal and vertical directions; It is the Sigmoid activation function. and It is used to generate attention weights in the horizontal and vertical directions. Convolutional layer; This indicates element-wise multiplication; This is a splicing feature.

[0055] Feature map after fusion by coordinate attention module After passing through two convolutional layers and an activation function, the final initial arrival position prediction result is obtained:

[0056] In the above formula, and These are the parameters of the two convolutional layers, respectively.

[0057] Using the above formula, the network sequentially completes feature extraction from the input data, multi-domain feature capture, multi-domain feature fusion, and final classification output, with the output size matching the input size, thus meeting the requirements of seismic signal processing.

[0058] In one optional embodiment, training the multi-input attention mechanism network architecture includes: Samples are selected from the earthquake data as a training set, and the training set is input into the multi-input attention mechanism network architecture; The multi-input attention mechanism network architecture is trained using binary cross-entropy as the loss function and an adaptive matrix derived from backpropagation; wherein the loss function is expressed as follows:

[0059] In the above formula, is the loss value, y is the binary label, which takes the value 0 or 1; p(y) represents the probability that the output belongs to the label y; This represents the total number of training samples.

[0060] Where y is 0, it represents a non-initial position; and y is 1, it represents an initial position.

[0061] Representative samples were selected from the collected seismic data, and first arrivals were precisely picked manually by professional seismic interpreters to serve as the real label data for network training. During the picking process, strict adherence to unified standards ensured accuracy and consistency. The time-space domain data, instantaneous amplitude features, frequency and phase feature data, and their corresponding first arrival picking information were compiled into a training dataset, which was then divided into training, validation, and test sets at a ratio of 80%, 10%, and 10%, respectively. The training set was used for parameter learning of the network model, the validation set was used to evaluate model performance and adjust hyperparameters during training, and the test set was used to ultimately evaluate the model's generalization ability and picking accuracy.

[0062] The training set data is input into the constructed MI-Attn-PrecisePickNet network. Supervised learning is employed, using binary cross-entropy as the loss function. Binary cross-entropy effectively measures the difference between the network's predictions and the true initial labels. In the loss function calculation, for a label y of 1, if the predicted value p(y) approaches 1, then the loss function should approach 0; conversely, if the predicted value p(y) approaches 0, then the loss function should be very large.

[0063] The network is trained using the Adaptive Moment Estimation (Adam) method via backpropagation, with the goal of minimizing the loss function, thus updating the network model. Training stops when the validation set loss decreases and remains stable over a certain iteration period, and the weights are then adjusted. Save the results of the training run. The saved optimal weights are used to predict the initial arrival on the test set, as shown in the following formula:

[0064] This represents the role of the network.

[0065] During training, closely monitor metrics such as the validation set loss and picking accuracy, and dynamically adjust network training parameters, such as learning rate, number of iterations, and batch size, based on these metrics to prevent overfitting and ensure good generalization ability. When the validation set loss remains stable and no longer decreases within a certain iteration period, stop training and save the network weights at this point as the training result for subsequent initial picking predictions on the test set.

[0066] Systematic experimental analysis was conducted on key hyperparameters of the MI-Attn-PrecisePickNet network, such as network depth, learning rate, kernel size, and batch size. During the experiments, only one hyperparameter to be tested was changed at a time, while other parameters remained constant (e.g., initial settings included a learning rate of 0.005, a kernel size of 5, a batch size of 4, and a network structure of...). By comparing the network performance under different hyperparameter combinations on a validation set, and using pixel accuracy as the basis, the hyperparameter model that enables the network to achieve optimal performance was determined. In this embodiment, experiments were conducted on learning rate, kernel matrix size, batch size, and network depth. Only one experimental parameter was changed each time, while the other parameters remained constant. The optimal parameters were determined based on the pixel accuracy of the validation set. The experimental results are as follows: Figure 2 As shown, from Figure 2 (a) It can be seen that when the learning rate is too large, the network is difficult to converge and will oscillate. For this experimental data, a learning rate of 0.005 has the best effect. Figure 2 (b) shows that when the kernel matrix size is 5, the network achieves the highest accuracy and fastest convergence speed; Figure 2 (c) It can be seen that while a larger batch size can improve generalization ability, it also increases computation time and memory requirements; Figure 2 (d) Display, and The network structures described above all achieve good accuracy, but increasing network depth increases memory usage. This experiment effectively confirms the parameters used during model training.

[0067] A well-trained and hyperparameter-optimized MI-Attn-PrecisePickNet network model is used to process the temporal-spatial and temporal-frequency domain data of actual seismic data. The network output is a probability value for each data point as the first arrival location. Thresholding is applied to the network output results: seismic traces with probability values ​​less than 0.4 are discarded and not picked for first arrival; for seismic traces with probability values ​​greater than or equal to 0.4, the point with the highest probability is selected as the final first arrival location and marked as 1, thus achieving automatic picking of seismic first arrivals.

[0068] This embodiment focuses on three-dimensional seismic data collected in the North China Plain, such as... Figure 3 As shown in (a), the 3D seismic data contains 250 shot gather records, each shot gather has 10 permutations, each permutation has 60 channels, each channel has 501 sampling points, the sampling rate is 4ms, and the surface elevation is between 92 and 160m. To reduce training memory consumption, the original data and amplitude, frequency, and phase feature profiles are cut into 256×256 data blocks along the spatiotemporal direction, with some overlap between adjacent data blocks. The input batch size is set to 4, and the Adam algorithm is used for optimization. The data is normalized before being input into the network, and the training, validation, and test sets are divided into 80%, 10%, and 10% sets, respectively. The results of first arrival picking are compared using the energy ratio method, manual picking, and the MI-Attn-PrecisePickNet method. The comparison results are shown in the figure. Figure 3 As shown in (b), in areas with high signal-to-noise ratio (SNR) for seismic traces, the results of the MI-Attn-PrecisePickNet method and manual picking are basically consistent. However, when there is strong anomalous noise in the seismic record, MI-Attn-PrecisePickNet fails to pick the first arrival (IAM) because the training labels do not cover the IAM location information of such seismic traces. In low SNR conditions, manual picking seems to be more effective based on human inference of the results of surrounding traces, but it has subjective factors. Although the energy ratio method picks each trace, the results are chaotic in low SNR conditions. Considering that erroneous IAMs have a significant impact on velocity modeling, while a small number of missing IAMs have a relatively small impact on subsequent processing, it is more reasonable to abandon picking than to pick incorrectly in low SNR conditions.

[0069] Figure 4 The absolute error of the seismic traces picked by this method relative to the results of manual picking is provided. The absolute error shows that the energy ratio method is less effective than MI-Attn-PrecisePickNet.

[0070] The projection of the first arrival time onto the seismic data, instantaneous amplitude, frequency characteristic profile, and phase characteristic profile is as follows: Figure 5 As shown.

[0071] like Figure 6 As shown in the enlarged view of channels 29-35, the first arrival is located at the position of the first continuous strong amplitude and has similar phase and frequency characteristics. The method proposed in this embodiment can accurately pick out the first arrival through these characteristics.

[0072] In one optional embodiment, after obtaining the picking result, the method further includes: Obtain the test set labels, compare the picking results with the test set labels, and obtain comparison data; The comparison data is marked with pixel precision to obtain the correct pixel percentage; the calculation process for marking the comparison data with pixel precision is as follows:

[0073] In the above formula, For the correct pixel ratio, This represents the number of pixels of category i that are inferred to be category i. This represents all pixels.

[0074] This metric comprehensively evaluates the model's picking performance. Simultaneously, it tests the model on real seismic data with different signal-to-noise ratios, analyzing the model's stability and accuracy under varying data quality levels to fully verify its reliability and applicability.

[0075] Based on the evaluation results, the network model is further validated and optimized. If the model is found to perform poorly under certain data characteristics or geological scenarios, such as low accuracy in picking data with low signal-to-noise ratios, or unsatisfactory picking performance for seismic data from specific geological structures (such as near faults or complex lithological areas), the network structure can be adjusted accordingly. This includes increasing or decreasing the number of residual module layers, adjusting the number of hidden layer nodes in the residual long short-term memory module, and optimizing the weight allocation method of the attention mechanism. Training data can also be reselected, supplementing it with data samples possessing specific geological characteristics to enhance the model's adaptability to various complex situations. Training parameters can be optimized, such as adjusting the learning rate and changing the batch size, to improve the model's convergence speed and accuracy. Through continuous validation and optimization, the model's initial arrival picking accuracy and stability can be continuously improved, enabling it to be better applied to actual seismic data processing and providing more reliable initial arrival data support for seismic exploration and geological research.

[0076] The method provided in this embodiment achieves the following effects: 1. Construct a multi-domain fusion earthquake first arrival picking network: By integrating time-space domain and time-frequency domain data, establish characteristic response relationships that can accurately identify earthquake first arrivals; 2. Achieve high-precision first arrival picking: Utilize a multi-domain fusion network to efficiently and accurately pick first arrivals from massive seismic data, eliminating picking errors of traditional methods in complex data environments; 3. Enhance the overall efficiency of seismic processing: Provide reliable first arrival data for subsequent seismic processing procedures such as near-surface inversion and seismic imaging, optimize seismic data processing results, and promote the intelligent advancement of seismic exploration.

[0077] Example 2 Embodiment 2 of the present invention provides a seismic first arrival picking system based on a multi-input attention mechanism network. The system includes: The data processing module is used to acquire seismic data and preprocess the seismic data to obtain preprocessed seismic data; the preprocessed seismic data includes time-space domain seismic data, instantaneous amplitude, frequency, and phase characteristics; An architecture building module is used to build a multi-input attention mechanism network architecture, which includes four branch modules, a coordinate attention module, and an output layer. The branch feature module is used to process the time-space domain seismic data, the instantaneous amplitude, the frequency, and the phase features using the four branch modules respectively, to obtain four multi-domain features; The feature fusion module is used to perform coordinate attention processing on four multi-domain features using the coordinate attention module to obtain fused features, and output the fused features through the output layer.

[0078] Example 3

[0079] The present invention provides an electronic device in embodiment 3, which includes a processor, a memory, an input device, and an output device; the number of processors in the computer device can be one or more, taking one processor as an example; the processor, memory, input device, and output device in the electronic device can be connected by a bus or other means, taking connection by a bus as an example.

[0080] Memory, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules. Processors execute various functional applications and data processing of electronic devices by running the software programs, instructions, and modules stored in memory, thereby implementing the seismic first-arrival wave pickup method based on a multi-input attention mechanism network as described in Example 1.

[0081] The memory may primarily comprise a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function; the data storage area may store data created based on terminal usage. Furthermore, the memory may include high-speed random access memory (RAM) and non-volatile memory, such as at least one disk storage device, flash memory, or other non-volatile solid-state storage device. In some instances, the memory may further include memory remotely located relative to the processor, which can be connected to the electronic device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks (LANs), mobile communication networks, and combinations thereof. The input device can be used to receive user input such as ID and password. The output device is used to output the network configuration page.

[0082] Example 4 Embodiment 4 of the present invention also provides a computer-readable storage medium, wherein the computer-executable instructions, when executed by a computer processor, are used to implement the seismic first arrival wave picking method based on a multi-input attention mechanism network as provided in Embodiment 1.

[0083] The storage medium containing computer-executable instructions provided in the embodiments of the present invention is not limited to the method operation provided in Embodiment 1, but can also perform related operations in the seismic first arrival wave picking method based on multi-input attention mechanism network provided in any embodiment of the present invention.

[0084] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for picking first arrival waves of earthquakes based on a multi-input attention mechanism network, characterized in that, Includes the following steps: Step S1: Acquire seismic data and preprocess the seismic data to obtain preprocessed seismic data; the preprocessed seismic data includes time-space domain seismic data, instantaneous amplitude, frequency, and phase characteristics; Step S2: Construct a multi-input attention mechanism network architecture, which includes four branch modules, a coordinate attention module, and an output layer; the branch modules include: an m-layer residual module and an n-layer residual long short-term memory module; Specifically, the time-space domain seismic data, the instantaneous amplitude, the frequency, and the phase characteristics are used as two-dimensional input data. The m-layer residual module is used to process the two-dimensional input data to capture the spatial characteristics of the seismic signal. The spatial features of the seismic signal are converted into a time series, and the n-layer residual long short-term memory module is used to capture the features of the time series to obtain dynamic features in the time dimension. Step S3: The time-space domain seismic data, instantaneous amplitude, frequency, and phase features are processed using the four branch modules respectively to obtain four multi-domain features; Step S4: Perform coordinate attention processing on the four multi-domain features using the coordinate attention module to obtain fused features, and output the fused features through the output layer to obtain the picking result; the coordinate attention processing on the four multi-domain features using the coordinate attention module includes: The four multi-domain features are concatenated to obtain the concatenated features. Coordinate attention is calculated on the spliced ​​features to obtain the fused features; the formula for calculating coordinate attention is as follows: In the above formula, and It is an index in the height and width directions of the feature map; and This represents the result of global pooling in the horizontal and vertical directions. It's about height. It is the width; It is the ReLU activation function. yes Convolutional layer; and It is The result of the splitting is feature maps in the horizontal and vertical directions; It is the Sigmoid activation function. and It is used to generate attention weights in the horizontal and vertical directions. Convolutional layer; This indicates element-wise multiplication; This is a splicing feature.

2. The seismic first arrival wave picking method based on a multi-input attention mechanism network according to claim 1, characterized in that, The two-dimensional input data is processed using the m-layer residual module, including: In the above formula, For the residual module The output of the layer, Representing the Input data for the layer residual module, For the first Nonlinear transformation of layer residual modules, For the first Layer residual module parameters.

3. The seismic first arrival wave picking method based on a multi-input attention mechanism network according to claim 1, characterized in that, The time series is used to capture features using the n-layer residual long short-term memory module, including: In the above formula, For the first Layer in The hidden state of the output after residual connection at any given moment. For LSTM networks, For the first Layer residual long short-term memory module in The input sequence at time step.

4. The seismic first arrival wave picking method based on a multi-input attention mechanism network according to claim 1, characterized in that, The training of the multi-input attention mechanism network architecture includes: Samples are selected from the earthquake data as a training set, and the training set is input into the multi-input attention mechanism network architecture; The multi-input attention mechanism network architecture is trained using binary cross-entropy as the loss function and an adaptive matrix derived from backpropagation; wherein the loss function is expressed as follows: In the above formula, is the loss value, y is the binary label, which takes the value 0 or 1; p(y) represents the probability that the output belongs to the label y; This represents the total number of training samples.

5. The seismic first arrival wave picking method based on a multi-input attention mechanism network according to claim 1, characterized in that, After obtaining the picking result, it also includes: Obtain the test set labels, compare the picking results with the test set labels, and obtain comparison data; The comparison data is marked with pixel precision to obtain the correct pixel percentage; the calculation process for marking the comparison data with pixel precision is as follows: In the above formula, For the correct pixel ratio, This represents the number of pixels of category i that are inferred to be category i. This represents all pixels.

6. A seismic first arrival wave picking system based on a multi-input attention mechanism network, characterized in that, include: The data processing module is used to acquire seismic data and preprocess the seismic data to obtain preprocessed seismic data. The preprocessed seismic data includes time-space domain seismic data, instantaneous amplitude, frequency, and phase characteristics; An architecture building module is used to construct a multi-input attention mechanism network architecture, which includes four branch modules, a coordinate attention module, and an output layer; the branch modules include: an m-layer residual module and an n-layer residual long short-term memory module. Specifically, the time-space domain seismic data, the instantaneous amplitude, the frequency, and the phase characteristics are used as two-dimensional input data. The m-layer residual module is used to process the two-dimensional input data to capture the spatial characteristics of the seismic signal. The spatial features of the seismic signal are converted into a time series, and the n-layer residual long short-term memory module is used to capture the features of the time series to obtain the dynamic features in the time dimension. The branch feature module is used to process the time-space domain seismic data, the instantaneous amplitude, the frequency, and the phase features using the four branch modules respectively, to obtain four multi-domain features; The feature fusion module is used to perform coordinate attention processing on four multi-domain features using the coordinate attention module to obtain fused features, and output the fused features through the output layer; the coordinate attention processing on the four multi-domain features using the coordinate attention module includes: The four multi-domain features are concatenated to obtain the concatenated features. Coordinate attention is calculated on the spliced ​​features to obtain the fused features; the formula for calculating coordinate attention is as follows: In the above formula, and It is an index in the height and width directions of the feature map; and This represents the result of global pooling in the horizontal and vertical directions. It's about height. It is the width; It is the ReLU activation function. yes Convolutional layer; and It is The result of the splitting is feature maps in the horizontal and vertical directions; It is the Sigmoid activation function. and It is used to generate attention weights in the horizontal and vertical directions. Convolutional layer; This indicates element-wise multiplication; This is a splicing feature.

7. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the seismic first arrival wave picking method based on a multi-input attention mechanism network as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the seismic first arrival wave picking method based on a multi-input attention mechanism network as described in any one of claims 1 to 5.

Images