Seismic first-arrival wave pickup method based on symmetric double-decoding structure

Abstract

The invention discloses a seismic first arrival wave pickup method based on a symmetric double decoding structure, and relates to the field of seismic monitoring. The method comprises the following steps: acquiring pre-stack seismic waveform data; the pre-stack seismic waveform data comprises multi-channel single-channel seismic data; performing convolution operation and average pooling processing on the pre-stack seismic waveform data to obtain a plurality of coded feature maps; performing up-sampling on the plurality of coding feature maps in sequence to obtain first decoding feature data and second decoding feature data, and splicing the first decoding feature data and the second decoding feature data in a channel dimension to obtain splicing features corresponding to the pre-stack seismic waveform data; and fusing the splicing features to obtain a probability matrix, and extracting first-arrival wave information of each piece of single-channel seismic data from the probability matrix. According to the method, the confirmation precision of the seismic first arrival wave information is improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method for picking up first arrival waves of earthquakes based on a symmetrical dual-decoding structure. Background Technology

[0002] First-arrival wave acquisition is a crucial step in seismic data processing, and its accuracy directly impacts the reliability of subsequent interpretation work such as source location and geological structure inversion, thus possessing significant engineering importance. Traditional first-arrival wave acquisition methods rely on manual acquisition. However, with the deepening of exploration work and the ever-increasing volume of exploration data, the efficiency of manual acquisition is limited, necessitating the development of high-precision automated acquisition technology.

[0003] In recent years, data-driven methods, represented by deep learning technology, have developed rapidly. These methods establish a nonlinear mapping relationship between raw seismic records and first arrival waves through hierarchical feature learning mechanisms, effectively avoiding errors introduced by humans. However, these methods heavily rely on supervised learning with labeled data. Existing research typically focuses on model architecture innovation and input feature enhancement, neglecting the impact of output task constraints on the direction of feature learning. This results in low accuracy in picking first arrival waves. Summary of the Invention

[0004] Therefore, it is necessary to provide a seismic first-arrival acquisition method based on a symmetrical dual-decoding structure to address the aforementioned technical problems. This method improves the acquisition accuracy of seismic first-arrival information.

[0005] The present invention adopts the following technical solution: This invention provides a method for picking up seismic first arrival waves based on a symmetrical dual-decoding structure, comprising: Acquire pre-stack seismic waveform data; upsampled pre-stack seismic waveform data includes multiple single-channel seismic data; each single-channel seismic data includes multiple sampling points, with each sampling point spaced at a preset time interval; Multiple convolution operations were performed on the pre-stack seismic waveform data to obtain multiple encoded feature maps; Multiple encoded feature maps are upsampled sequentially by two multi-level transposed convolutions to obtain the first and second decoded feature data, respectively. The upsampled first and second decoded feature data are then concatenated along the channel dimension to obtain the concatenated features corresponding to the pre-stack seismic waveform data. After convolution and high-dimensional embedding space mapping of the upsampled concatenated features, the global context is modeled through a multi-head self-attention mechanism. The modeling results are then subjected to multi-scale convolution, and the results of the multi-scale convolution are fused to obtain the probability matrix. First arrival information for each single-channel seismic data is extracted from the upsampling probability matrix; the upsampling probability matrix includes first arrival information for multiple channels of seismic data; the first arrival information includes the arrival time of the first arrival.

[0006] Preferably, the first arrival information of each single-channel seismic data is extracted using a first arrival picking model; the upsampled first arrival picking model includes a shared coding module, a dual-path symmetrical decoding module, and a feature fusion module. Multiple convolution operations are performed on the pre-stack seismic waveform data through a shared coding module to obtain multiple coded feature maps; the upsampling shared coding module includes multiple cascaded convolutional groups; each convolutional group includes two stacked convolutional blocks; each convolutional block includes a concatenated convolutional layer, a normalization layer and a LeakyReLU activation layer in sequence; The upsampling dual-path symmetric decoding module sequentially upsamples multiple encoded feature maps to obtain the first decoded feature data and the second decoded feature data. The upsampling dual-path symmetric decoding module includes decoding branch one and decoding branch two. Both upsampling decoding branch one and upsampling decoding branch two include four levels of transposed convolutional layers. The probability matrix is ​​generated by the upsampling feature fusion module. The upsampling feature fusion module includes a first 1×1 convolutional layer, a multi-head attention mechanism, a second 1×1 convolutional layer, a multi-scale convolutional layer and a 3×3 convolutional layer connected in series. The convolution result of the upsampling first 1×1 convolutional layer and the convolution result of the upsampling second 1×1 convolutional layer are summed and then input into the upsampling multi-scale convolutional layer.

[0007] Preferably, the training process of the upsampled first-arrival picking model includes a global stage, a local stage, and a fusion stage. Specifically, the training process of the first-arrival picking model includes: Obtain the global segmentation labels and local focus labels of the sample seismic waveform data; In the global upsampling phase, the upsampling shared coding module and the upsampling decoding branch 1 are trained using global segmentation labels until the value of the binary cross-entropy loss function is minimized. In the local upsampling phase, the parameters of the upsampling shared coding module and the upsampling decoding branch one are frozen, and the decoding branch two is trained by local focusing labels until the value of the binary cross-entropy loss function is minimized. During the upsampling fusion stage, the parameters of the shared encoding module are frozen, and the upsampling decoding branch one, upsampling decoding branch two, and upsampling feature fusion module are jointly optimized through local focus labels until the value of the joint loss function is minimized. After completing the training of the global upsampling stage, the local upsampling stage, and the upsampling fusion stage, the upsampling seismic first arrival wave picking model is obtained.

[0008] Preferably, the upsampling joint loss function is: ; in, , It is the weighting balance coefficient between the dominant loss term and the regularization term. For the binary cross-entropy loss in the global phase, For the binary cross-entropy loss in the local phase, As the dominant loss item, For regularization terms, For the joint loss value, N This represents the number of training samples; The dominant loss term for upsampling is: ; in, For the first i The true label of each sample For the first i The model predicts the probability value of each sample.

[0009] Preferably, the upsampling process performs multiple convolution operations on the pre-stack seismic waveform data to obtain multiple encoded feature maps, specifically including: The pre-stack seismic waveform data is input into the upsampling shared coding module. In each convolutional group of the shared coding module, the upsampling pre-stack seismic waveform data is convolved and downsampled to obtain multiple feature maps. All feature maps generated by multiple convolutional groups are identified as multiple encoded feature maps.

[0010] Preferably, the upsampling is performed sequentially on multiple encoded feature maps through two multi-level transposed convolutions to obtain the first decoded feature data and the second decoded feature data, specifically including: The first decoded feature data is obtained by sequentially upsampling multiple encoded feature maps through the upsampling decoding branch; The second decoding feature data is obtained by upsampling multiple encoded feature maps sequentially through the second upsampling decoding branch.

[0011] Preferably, after upsampling concatenates the features and performs convolution and high-dimensional embedding space mapping, the global context is modeled through a multi-head self-attention mechanism. The modeling results are then subjected to multi-scale convolution, and the results of the multi-scale convolution are fused to obtain a probability matrix, specifically including: The upsampled spliced ​​features are channel-mapped to a high-dimensional embedding space using 1×1 convolution, and the spliced ​​feature data mapped to the high-dimensional embedding space is sampled to a 32×32 dimension using adaptive average pooling. A linear transformation is performed on the concatenated features mapped to a high-dimensional embedding space to generate input parameters for a multi-head self-attention mechanism; A global feature representation is obtained by modeling the input parameters in a global context using a multi-head self-attention mechanism. The upsampled global feature representation is restored to spatial resolution through bilinear interpolation. A 1×1 convolution is used to map the global feature representation after restoring spatial resolution back to the original number of channels. The global feature representation mapped back to the original number of channels is added to the residual of the concatenated feature data to obtain the self-attention intermediate feature map. The upsampled self-attention intermediate feature map is convolved with 3×3 convolution, 3×3 dilated convolution with an expansion rate of 2 and 1×1 convolution respectively to obtain the first convolution result, the second convolution result and the third convolution result; The upsampling first convolution result, the upsampling second convolution result, and the upsampling third convolution result are channel-wise spliced ​​and fused, and the channel-wise spliced ​​and fused result is then convolved with 3×3 to obtain the upsampling probability matrix. The dimension of the upsampling probability matrix is ​​the same as that of the pre-stack seismic waveform data, and the value of the probability matrix is ​​between [0, 1].

[0012] Preferably, upsampling extracts the first arrival information of each single-channel seismic data from the upsampling probability matrix, specifically including: The maximum value in each column of the upsampling probability matrix is ​​used to determine the first arrival wave information of each single-channel seismic data.

[0013] This invention provides a seismic first arrival wave pickup device based on a symmetrical dual-decoding structure, comprising: The acquisition module is used to acquire pre-stack seismic waveform data; the pre-stack seismic waveform data includes multi-channel single-channel seismic data; each single-channel seismic data includes multiple sampling points, and each sampling point is spaced at a preset time interval; A shared coding module is used to perform multiple convolution operations on pre-stack seismic waveform data to obtain multiple coded feature maps. The dual-path symmetric decoding module is used to upsample multiple encoded feature maps sequentially through two-path multi-level transposed convolutions to obtain the first decoded feature data and the second decoded feature data respectively. The first decoded feature data and the second decoded feature data are then concatenated in the channel dimension to obtain the concatenated features corresponding to the pre-stack seismic waveform data. The feature fusion module is used to perform convolution and high-dimensional embedding space mapping on the concatenated features, model the global context through a multi-head self-attention mechanism, perform multi-scale convolution on the modeling results, and fuse the results of the multi-scale convolution to obtain a probability matrix. The picking module is used to extract the first arrival information of each single-channel seismic data from the probability matrix; the probability matrix includes the first arrival information of multiple seismic data; the first arrival information includes the arrival time of the first arrival.

[0014] This invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for picking up first arrival waves of a seismic wave based on a symmetrical dual-decoding structure.

[0015] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-described method for picking up the first arrival wave of a seismic wave based on a symmetrical dual-decoding structure.

[0016] The above-mentioned at least one technical solution adopted in this invention can achieve the following beneficial effects: Convolution and average pooling operations are performed on pre-stack seismic waveform data to obtain multiple encoded feature maps. These operations enhance sensitivity to initial arrival times and reduce noise interference. The encoded feature maps are then sequentially upsampled to obtain first and second decoded feature data. These first and second decoded feature data are then concatenated along the channel dimension to obtain the concatenated features corresponding to the pre-stack seismic waveform data. Upsampling gradually restores spatial resolution and reconstructs seismic wave details, ensuring accurate first-arrival position information. Concatenation along the channel dimension improves the completeness of this representation. The concatenated features are then fused to obtain a probability matrix. First-arrival information for each single-channel seismic data point is extracted from the probability matrix, and the arrival time of the first-arrival is accurately located based on the peak value of the probability matrix. This method improves the accuracy of first-arrival confirmation. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This invention provides a schematic flowchart of a seismic first arrival wave picking method based on a symmetrical dual-decoding structure. Figure 2 This is a schematic diagram of the earthquake first arrival wave picking model provided by the present invention; Figure 3 This invention provides a schematic diagram of the multi-head attention mechanism and multi-scale convolution. Figure 4 This is a schematic diagram of the field data picking structure of the earthquake first arrival picking model provided by the present invention; wherein, Figure 4 Figure (a) in the figure is a schematic diagram of the picking results of the pre-stack seismic profile. Figure 4 Figure (b) is a magnified view of the seismic waveform data of trace 32 in the profile; Figure 5A schematic diagram of global segmentation and local focusing labels for sample seismic waveform data provided by the present invention; wherein, Figure 5 Figure (a) shows pre-stack seismic waveform data from multiple channels. Figure 5 Figure (b) in the figure is a schematic diagram of different types of labels for the 32nd single-channel seismic waveform data; Figure 6 A schematic diagram of a seismic first arrival wave pickup device based on a symmetrical dual-decoding structure provided by the present invention; Figure 7 This is a schematic diagram of the structure of a computer device for executing a seismic first arrival wave picking method based on a symmetrical dual decoding structure, as provided by the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0019] Devices such as desktop computers, servers, and laptops are capable of executing the solutions of this invention. For ease of explanation, the following description will focus on servers as the executing entity.

[0020] To address the limitations of traditional single-task models, this invention proposes a multi-task collaborative first-arrival picking framework based on a symmetric dual-decoders structure (MT-SDD Picking). MT-SDD, the seismic first-arrival picking model in this invention, focuses on the guiding role of the output task in feature learning. It innovatively constructs a collaborative learning mechanism between global segmentation and local focusing tasks, and drives the model to simultaneously capture both macroscopic structural features of the wavefield and microscopic details of the waveform through a multi-stage progressive optimization strategy of "global → local → fusion." To address the need for cross-task feature fusion, a feature fusion module combining Transformer self-attention and multi-scale convolutional pyramids is further designed, enabling the interaction and enhancement of features from different tasks and effectively improving the overall representational capability of the model.

[0021] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Figure 1This is a schematic diagram of a seismic first arrival wave picking method based on a symmetrical dual-decoding structure in this invention, which specifically includes the following steps: S101: Acquire pre-stack seismic waveform data; the pre-stack seismic waveform data includes multiple single-channel seismic data; each single-channel seismic data includes multiple sampling points, and each sampling point is spaced at a preset time interval.

[0023] In an exemplary embodiment, the first arrival information of each single-channel seismic data is extracted using a first arrival picking model. The first arrival picking model includes a shared coding module, a dual-path symmetric decoding module, and a feature fusion module. The shared coding module performs multiple convolution operations on the pre-stack seismic waveform data to obtain multiple coded feature maps. The shared coding module includes multiple cascaded convolutional groups; each convolutional group includes two stacked convolutional blocks; each convolutional block includes a concatenated convolutional layer, a normalization layer, and a LeakyReLU activation layer. The dual-path symmetric decoding module... Multiple encoded feature maps are upsampled sequentially to obtain the first decoded feature data and the second decoded feature data; the dual-path symmetric decoding module includes decoding branch one and decoding branch two; both decoding branch one and decoding branch two include four-level transposed convolutional layers; a probability matrix is ​​generated through the feature fusion module; the feature fusion module includes a first 1×1 convolutional layer, a multi-head attention mechanism, a second 1×1 convolutional layer, a multi-scale convolutional layer and a 3×3 convolutional layer connected in series, wherein the convolution result of the first 1×1 convolutional layer and the convolution result of the second 1×1 convolutional layer are summed and then input to the multi-scale convolutional layer.

[0024] Specifically, the encoding and decoding module of the seismic first-arrival picking model is model-independent and can adapt to first-arrival picking models with various encoding and decoding structures. Considering the spatial continuity of seismic profiles, this invention adopts the widely used U-Net model as an example implementation of MT-SDD.

[0025] This invention provides, for example Figure 2 The earthquake first arrival wave picking model shown consists of a shared coding module, a dual-path symmetrical decoding module, and a feature fusion module. Through a three-stage optimization strategy, it gradually realizes the transition from single feature learning to multi-view feature fusion.

[0026] In an exemplary embodiment, the training process of the earthquake first-arrival picking model includes a global stage, a local stage, and a fusion stage. Specifically, the training process includes: acquiring global segmentation labels and local focus labels for sample earthquake waveform data; in the global stage, training the shared coding module and decoding branch one using the global segmentation labels until the binary cross-entropy loss function is minimized to capture the macroscopic distribution and global contextual features of the earthquake first-arrival wavefield; in the local stage, freezing the parameters of the shared coding module and decoding branch one, and training decoding branch two using the local focus labels until the binary cross-entropy loss function is minimized to capture the local waveform details of the earthquake first-arrival; in the fusion stage, freezing the parameters of the shared coding module, and jointly optimizing decoding branch one, decoding branch two, and the feature fusion module using the local focus labels until the joint loss function is minimized; after completing the training of the global, local, and fusion stages, the earthquake first-arrival picking model is obtained.

[0027] Specifically, to reduce mutual interference between learning tasks with different labels, the earthquake first arrival wave picking model divides training into three stages and adopts a progressive optimization strategy of "global-local-fusion" to help the earthquake first arrival wave picking model learn features from different dimensions and perspectives.

[0028] Phase 1: Global Feature Modeling.

[0029] By utilizing global segmentation labels to simultaneously optimize the encoding module and the decoding branch, the macroscopic distribution and global contextual features of the earthquake first arrival wave field are captured, ensuring the model's ability to capture both the underlying and global features of the data.

[0030] Phase Two: Focusing on Local Features.

[0031] By freezing the parameters of the encoding module and the first decoding branch, and training the second decoding branch using only local focus labels, the model is encouraged to focus on the local waveform details of the first arrival wave while avoiding destroying the global knowledge learned by the first local branch.

[0032] Phase 3: Dynamic Feature Fusion The parameters of the encoding module are frozen, and the two decoding branches and the feature fusion module are jointly optimized using the local focus label task to achieve information interaction and complementary enhancement of global and local features.

[0033] For training in stages one and two, the classic binary cross-entropy loss function is used to optimize the difference between the stage output and the true label.

[0034] For stage three, a joint loss function is used that combines the cross-entropy loss function with the cosine similarity loss based on global branch features and local branch features.

[0035] For the joint loss function, the cross-entropy term is the dominant loss term, ensuring the direction of model optimization; the cosine similarity term is the loss regularization term, guiding the feature distributions of the global branch and the local branch to tend to be consistent, enhancing the synergy between the two branches, and avoiding excessive differences in features between the global branch and the local branch.

[0036] The binary cross-entropy loss function is shown in Equation (1): (1); in, N The number of training samples. For the first i The true label of each sample For the first i The model predicts the probability value of each sample.

[0037] The joint loss function is shown in equation (2): (2); in, , It is the weighting balance coefficient between the dominant loss term and the regularization term. For the binary cross-entropy loss in the global phase, For the binary cross-entropy loss in the local phase, As the dominant loss item, For regularization terms, This is the combined loss value.

[0038] S102: Perform multiple convolution operations on the pre-stack seismic waveform data to obtain multiple encoded feature maps.

[0039] In an exemplary embodiment, multiple convolution operations are performed on the pre-stack seismic waveform data to obtain multiple encoded feature maps. Specifically, this includes: inputting the pre-stack seismic waveform data into a shared coding module; performing convolution and downsampling on the pre-stack seismic waveform data in each convolution group of the shared coding module to obtain multiple feature maps; and determining all feature maps generated by the multiple convolution groups as multiple encoded feature maps.

[0040] Specifically, the encoding module consists of four cascaded convolutional groups. Each convolutional group contains two stacked convolutional layers, followed by a batch normalization layer and a LeakyReLU activation layer. The convolutional layers use asymmetric 3×9 convolutional kernels to fully capture the horizontal continuity features of the seismic profile. The feature map output by each convolutional group is spatially downsampled using an average pooling layer. After four levels of compression, a high-dimensional abstract feature representation of the input seismic data is formed as shown in Equation (3): (3); Where X is the input seismic data, X' is the output of the encoding module, B is the batch sample, and C is the number of convolution channels.

[0041] S103: Multiple encoded feature maps are upsampled sequentially through two-way multi-level transposed convolution to obtain the first decoded feature data and the second decoded feature data respectively. The first decoded feature data and the second decoded feature data are then concatenated in the channel dimension to obtain the concatenated features corresponding to the pre-stack seismic waveform data.

[0042] Each path in the dual-path symmetric decoding module includes four transposed convolutional layers. The four transposed convolutional layers are connected in series.

[0043] In an exemplary embodiment, multiple encoded feature maps are sequentially upsampled through two multi-level transposed convolutions to obtain first decoded feature data and second decoded feature data, respectively. Specifically, this includes: sequentially upsampling multiple encoded feature maps through decoding branch one to obtain first decoded feature data; and sequentially upsampling multiple encoded feature maps through decoding branch two to obtain second decoded feature data.

[0044] Specifically, the dual decoding modules have identical structures and are designed as mirror images of the encoding modules. Each decoding module also contains four levels of transposed convolutional layers, which progressively reconstruct the spatial dimensions of the feature maps. During the reconstruction process, the features at each scale of the decoding and encoding modules are strictly aligned, and skip connections are used to achieve the splicing and fusion of encoding-decoding features at the same scale. This design effectively preserves the high-frequency details of the original input waveform data, providing important spatial feature support for pixel-level initial arrival picking tasks.

[0045] S104: After convolution and high-dimensional embedding space mapping of the concatenated features, the global context is modeled through a multi-head self-attention mechanism. The modeling results are then subjected to multi-scale convolution, and the results of the multi-scale convolution are fused to obtain the probability matrix.

[0046] In an exemplary embodiment, after convolution and high-dimensional embedding space mapping of the concatenated features, a global context is modeled using a multi-head self-attention mechanism. The modeling results are then subjected to multi-scale convolution, and the results of the multi-scale convolution are fused to obtain a probability matrix. Specifically, this includes: mapping the concatenated features to a high-dimensional embedding space using 1×1 convolution, and sampling the concatenated feature data mapped to the high-dimensional embedding space to a 32×32 dimension using adaptive average pooling; performing a linear transformation on the concatenated features mapped to the high-dimensional embedding space to generate input parameters for the multi-head self-attention mechanism; modeling the global context using the multi-head self-attention mechanism to obtain a global feature representation; and performing bilinear interpolation on the global feature representation. To restore spatial resolution, a 1×1 convolution is used to map the global feature representation after spatial resolution restoration back to the original number of channels. The global feature representation mapped back to the original number of channels is then added to the residual of the stitched feature data to obtain a self-attention intermediate feature map. The self-attention intermediate feature map is then convolved with a 3×3 convolution, a 3×3 dilated convolution with a dilation rate of 2, and a 1×1 convolution to obtain the first, second, and third convolution results. The first, second, and third convolution results are then channel-stitched and fused, and the channel-stitched and fused result is then subjected to a 3×3 convolution to obtain a probability matrix. The dimension of the probability matrix is ​​the same as that of the pre-stack seismic waveform data, and the values ​​of the probability matrix are between [0, 1].

[0047] The formula corresponding to the multi-head self-attention mechanism is shown in formula (4): (4); in, , , and For a learnable parameter matrix, For global feature representation, Indicates splicing, Q , K and V The input parameters represent the multi-head attention mechanism, and Attention represents the attention mechanism.

[0048] Specifically, this invention provides a feature fusion module, whose multi-head attention mechanism and multi-scale convolution implementation are as follows: Figure 3 As shown, to achieve adaptive fusion of global wavefield structure features guided by global segmentation annotations and local waveform detail features guided by local focus annotations, this invention proposes a feature fusion module that combines a Transformer attention mechanism and a multi-scale convolutional pyramid. This module takes the outputs of the dual decoding module concatenated along the channel dimension as input, adaptively balancing the importance of local and global features while preserving specific task information, and finally outputs fused features for seismic wave first arrival picking.

[0049] The feature fusion module first uses 1×1 convolutions for channel mapping, mapping the concatenated features to a high-dimensional embedding space (2→32). To balance the module's computational complexity and feature representation capability, adaptive average pooling is used to sample the feature image to a 32×32 dimension. Then, an 8-head self-attention mechanism is applied for global context modeling. The structure of the self-attention mechanism is as follows: Figure 3 As shown in Figure (a), after the spatial resolution of the attention output is restored by bilinear interpolation, it is mapped back to the original number of channels using a 1×1 convolution in the last layer, and finally added to the original concatenated feature residual.

[0050] Subsequently, a multi-scale pyramid convolutional structure is employed to capture feature information at different scales. This structure is as follows: Figure 3 As shown in Figure (b), there are three parallel convolutional branches: (1) a standard 3×3 convolution for local feature extraction; (2) a 3×3 dilated convolution with an expansion rate of 2 to expand the receptive field; and (3) a 1×1 convolution for cross-channel information interaction and dimensionality reduction. The outputs of each branch are fused by channel splicing and finally output as the final fused feature tensor by a 1×1 convolution.

[0051] The feature fusion module effectively balances the need for local detail features and global structure awareness, which helps the model fuse waveform detail features learned from local branches and wave field structure features learned from global branches.

[0052] S105: Extract the first arrival information of each single-channel seismic data from the probability matrix; the probability matrix includes the first arrival information of multiple seismic data; the first arrival information includes the arrival time of the first arrival.

[0053] In an exemplary embodiment, extracting the first arrival information of each single-channel seismic data from the probability matrix specifically includes: determining the maximum value of each column in the probability matrix as the first arrival information of each single-channel seismic data.

[0054] This invention proposes a multi-task collaborative first-arrival picking framework based on a symmetric dual-decoding structure, which effectively utilizes the complementary advantages of global segmentation tasks and local focusing tasks in feature representation. Through multi-stage optimization strategies and feature fusion modules, this invention gradually realizes the transformation from single-feature learning to multi-view feature fusion, successfully overcoming the perspective limitations of single-task models and breaking through the bottlenecks in picking accuracy and stability of first-arrival picking models. Compared with single-task models, this invention can reduce first-arrival picking errors by approximately 20% and improve picking accuracy within a 2ms range by approximately 1%.

[0055] Figure 4 This is a schematic diagram of the field data picking structure of the earthquake first arrival picking model provided by the present invention. Figure 4 The signal-to-noise ratio of the example data is 2. Figure 4Figure (a) in the figure is a schematic diagram of the picking results of the pre-stack seismic profile. Figure 4 Figure (b) is an enlarged view of the seismic waveform data of the 32nd trace in the profile; as shown in the picking results, the first arrival wave picking accuracy of the present invention has achieved results that are almost consistent with those of manual picking.

[0056] Figure 5 This is a schematic diagram of global segmentation and local focusing labels for sample seismic waveform data provided by the present invention. (See diagram below.) Figure 5 Figure (a) shows pre-stack seismic waveform data from multiple channels. Figure 5 Figure (b) in the diagram illustrates the different types of labels for the 32nd single-channel seismic waveform data. For example... Figure 5 As shown in Figure (b), the global segmentation labels divide the sampling points before and after the initial arrival time of the earthquake into two categories, such as... Figure 5 As shown by the red line in Figure (b), for locally focused labels, the label at the first arrival time is explicitly defined as 1, while the label values ​​of adjacent waveform regions gradually decay to 0 as the distance from the first arrival time increases. Depending on the decay function, this type can be further subdivided into matrix, triangular, and Gaussian forms, among others.

[0057] When applying the seismic first arrival wave picking method based on a symmetrical dual-decoding structure provided by this invention, it is not necessary to consider... Figure 1 The steps shown are executed in sequence. The specific execution order of each step can be determined as needed, and this invention does not impose any restrictions on it.

[0058] The above describes a seismic first-arrival pickup method based on a symmetrical dual-decoding structure, provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding seismic first-arrival pickup device based on a symmetrical dual-decoding structure, such as... Figure 6 As shown.

[0059] Figure 6 A schematic diagram of a seismic first arrival wave pickup device based on a symmetrical dual-decoding structure provided by the present invention includes: The acquisition module 601 is used to acquire pre-stack seismic waveform data; the pre-stack seismic waveform data includes multiple single-channel seismic data; each single-channel seismic data includes multiple sampling points, and each sampling point is spaced at a preset time interval.

[0060] The shared coding module 602 is used to perform multiple convolution operations on the pre-stack seismic waveform data to obtain multiple coded feature maps.

[0061] The dual-channel symmetric decoding module 603 is used to upsample multiple encoded feature maps sequentially through two-channel multi-level transposed convolution to obtain first decoded feature data and second decoded feature data respectively, and then concatenate the first decoded feature data and second decoded feature data in the channel dimension to obtain the concatenated features corresponding to the pre-stack seismic waveform data.

[0062] The feature fusion module 604 is used to perform convolution and high-dimensional embedding space mapping on the concatenated features, model the global context through a multi-head self-attention mechanism, perform multi-scale convolution on the modeling results, and fuse the results of the multi-scale convolution to obtain a probability matrix.

[0063] The picking module 605 is used to extract the first arrival information of each single-channel seismic data from the probability matrix; the probability matrix includes the first arrival information of multiple seismic data; the first arrival information includes the arrival time of the first arrival.

[0064] Specific limitations regarding the seismic first-arrival pickup device based on a symmetrical dual-decoding structure can be found in the limitations of the seismic first-arrival pickup method based on a symmetrical dual-decoding structure described above, and will not be repeated here. Each module in the aforementioned seismic first-arrival pickup device based on a symmetrical dual-decoding structure can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the corresponding operations of each module.

[0065] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 A seismic first arrival wave picking method based on a symmetrical dual-decoding structure is provided.

[0066] The present invention also provides Figure 7 The schematic diagram of the computer device shown is as follows: Figure 7 As shown, at the hardware level, this computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above. Figure 1 A seismic first arrival wave picking method based on a symmetrical dual-decoding structure is provided.

[0067] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0068] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

Claims

1. A method for picking up first arrival waves of earthquakes based on a symmetrical dual-decoding structure, characterized in that, include: Acquire pre-stack seismic waveform data; The pre-stack seismic waveform data includes multiple single-channel seismic data; each single-channel seismic data includes multiple sampling points, and each sampling point is spaced at a preset time interval. Convolution and average pooling operations were performed on the pre-stack seismic waveform data to obtain multiple encoded feature maps. Multiple coded feature maps are sequentially decoded and upsampled to obtain first decoded feature data and second decoded feature data. The first decoded feature data and second decoded feature data are then concatenated along the channel dimension to obtain the concatenated features corresponding to the pre-stack seismic waveform data. The stitched features are fused to obtain a probability matrix, and the first arrival information of each single-channel seismic data is extracted from the probability matrix; the probability matrix includes the first arrival information of multiple seismic data; the first arrival information includes the arrival time of the first arrival. Earthquake first-arrival information is extracted using an earthquake first-arrival picking model, which includes a shared coding module, a dual-path symmetric decoding module, and a feature fusion module. The shared coding module performs convolution and average pooling operations on the pre-stack seismic waveform data. The shared coding module includes multiple cascaded convolutional groups and average pooling layers. The dual-path symmetric decoding module sequentially decodes and upsamples multiple coded feature maps. The dual-path symmetric decoding module includes a decoding branch one and a decoding branch two. Both decoding branch one and decoding branch two include four-level transposed convolutional layers. The feature fusion module fuses the spliced ​​features. The feature fusion module includes a first 1×1 convolutional layer, a multi-head attention mechanism, a second 1×1 convolutional layer, a multi-scale convolutional layer, and a 3×3 convolutional layer, sequentially connected. The convolution results of the first 1×1 convolutional layer and the second 1×1 convolutional layer are summed and then input to the multi-scale convolutional layer. The training process of the earthquake first-arrival picking model includes a global stage, a local stage, and a fusion stage. Specifically, the training process includes: acquiring global segmentation labels and local focus labels for sample earthquake waveform data; in the global stage, training the shared encoding module and the first decoding branch using the global segmentation labels until the binary cross-entropy loss function is minimized to capture the macroscopic distribution and global context features of the earthquake first-arrival wavefield; in the local stage, freezing the parameters of the shared encoding module and the first decoding branch, and training the second decoding branch using the local focus labels until the binary cross-entropy loss function is minimized to capture the local waveform details of the earthquake first-arrival; in the fusion stage, freezing the parameters of the shared encoding module, and jointly optimizing the first decoding branch, the second decoding branch, and the feature fusion module using the local focus labels until the joint loss function is minimized; after completing the training of the global stage, the local stage, and the fusion stage, the earthquake first-arrival picking model is obtained.

2. The method as described in claim 1, characterized in that, The joint loss function is: ； in, , It is the weighting balance coefficient between the dominant loss term and the regularization term. For the binary cross-entropy loss in the global phase, For the binary cross-entropy loss in the local phase, As the dominant loss item, For regularization terms, This is the joint loss value; The dominant loss term is: ; in, N The number of training samples. For the first i The true label of each sample For the first i The model predicts the probability value of each sample.

3. The method as described in claim 1, characterized in that, The pre-stack seismic waveform data is subjected to convolution and average pooling operations to obtain multiple encoded feature maps, specifically including: The pre-stack seismic waveform data are respectively input into multiple cascaded convolutional groups and average pooling layers; In each convolutional group and average pooling layer, the pre-stack seismic waveform data is convolved and downsampled to obtain multiple feature maps; all feature maps generated by multiple cascaded convolutional groups and average pooling layers are determined as multiple encoded feature maps.

4. The method as described in claim 3, characterized in that, The convolutional group comprises two stacked convolutional layers; each convolutional layer is followed by a normalization layer and a LeakyReLU activation layer in series.

5. The method as described in claim 1, characterized in that, The step of fusing the spliced ​​features to obtain a probability matrix specifically includes: mapping the spliced ​​features to a high-dimensional embedding space through 1×1 convolution, and sampling the spliced ​​feature data mapped to the high-dimensional embedding space to a 32×32 dimension using adaptive average pooling. A linear transformation is performed on the concatenated features mapped to the high-dimensional embedding space to generate the input parameters for the multi-head self-attention mechanism. Q , K and V ; A global feature representation is obtained by modeling the input parameters using a multi-head self-attention mechanism; the formula corresponding to the multi-head self-attention mechanism is: ; in, , , and For a learnable parameter matrix, For global feature representation, Indicates splicing, , and The input parameters represent the multi-head attention mechanism, and Attention represents the attention mechanism. The global feature representation is restored to spatial resolution through bilinear interpolation. After the global feature representation after restoring spatial resolution is mapped back to the original number of channels using 1×1 convolution, the global feature representation mapped back to the original number of channels is added to the residual of the concatenated feature data to obtain the self-attention intermediate feature map. The self-attention intermediate feature map is input into a 3×3 convolution, a 3×3 dilated convolution with an expansion rate of 2, and a 1×1 convolution, respectively, to obtain the first convolution result, the second convolution result, and the third convolution result; The first convolution result, the second convolution result, and the third convolution result are concatenated and fused through channels. The result of the channel concatenation and fusion is then convolved with a 3×3 convolution to obtain the probability matrix. The dimension of the probability matrix is ​​the same as that of the pre-stack seismic waveform data, and the value of the probability matrix is ​​between [0, 1].

6. The method as described in claim 1, characterized in that, The extraction of the first arrival information of each single-channel seismic data from the probability matrix specifically includes: The maximum value in each column of the probability matrix is ​​used to determine the first arrival wave information of each single-channel seismic data.

Images