Multi-scale feature extraction and scale restoration method for DAS seismic phase picking

Abstract

The application provides a multi-scale feature extraction and scale recovery method for DAS seismic phase picking, and is applied to the technical field of DAS seismic phase picking. Multi-scale decomposition is performed on the basis of input DAS data of a seismic phase picking model to obtain multi-scale DAS data features of the DAS data, multi-scale convolution suitable for multi-scale decomposition is performed on the DAS data to obtain scale weights in a scale recovery module, different direction sorting is used to model seismic signals propagated to a DAS device from different directions to obtain multi-scale, and scale weight guided feature scale recovery is performed on the basis of the obtained scale weights and multi-scale. The method is helpful to solve the problem that traditional DAS seismic phase picking methods usually rely on single-scale feature extraction or time domain feature extraction, and the traditional DAS seismic phase picking network has insufficient feature extraction capability for DAS data.

Description

Technical Field

[0001] This invention is applied to the field of DAS phase picking technology, and particularly relates to a multi-scale feature extraction and scale recovery method for DAS phase picking. Background Technology

[0002] In distributed acoustic sensing seismic monitoring scenarios, DAS (Distributed Acoustic Sensing) data typically features a large number of channels, high sampling frequency, long continuous recording time, and complex background noise. The arrival time, amplitude, and waveform morphology of the same seismic phase at different spatial locations can also vary due to the influence of propagation path, coupling conditions, and local interference.

[0003] Traditional phase picking methods, relying on waveform abrupt changes in a single gather, fixed time window features, or manual thresholding, struggle to reliably extract effective features of weak, overlapping, and low signal-to-noise ratio phases from large-scale continuous DAS records. To improve the automation and accuracy of DAS phase picking, existing methods typically introduce multi-scale feature extraction and scale recovery mechanisms. The basic workflow is as follows: First, the continuous vibration data acquired by distributed acoustic sensing is preprocessed, converting signals from different gathers or time windows into a spatiotemporal data matrix suitable for model input. Then, convolution, pooling, dilated convolution, or encoder structures are used to extract local waveform variations, inter-trace continuity, and phase propagation features over long time periods from the DAS data at multiple scales to obtain feature representations at different temporal resolutions and spatial receptive fields. Based on this, By enhancing the correlation between shallow detail features and deep semantic features through feature fusion, skip connections, or attention weighting, the model can simultaneously retain abrupt change information near the arrival time of the seismic phase and wavefield evolution information over a larger range. Finally, deconvolution, interpolation upsampling, or decoder structures are used to scale the downsampled feature map, so that the output results correspond to the time sampling scale and gather location scale of the original DAS data, and the picking probabilities or arrival times of seismic phases such as P-waves, S-waves, or noise backgrounds are generated, thereby realizing the automatic identification and picking of seismic phase events in large-scale continuous DAS records.

[0004] Traditional DAS phase picking methods typically rely on single-scale feature extraction or temporal feature extraction, which results in insufficient feature extraction capabilities of traditional DAS phase picking networks for DAS data. Summary of the Invention

[0005] Therefore, embodiments of the present invention provide a method for multi-scale feature extraction and scale recovery for DAS phase picking.

[0006] The technical solution of this invention is implemented as follows: DAS data based on the input seismic phase picking model is decomposed into multiple scales to obtain multi-scale DAS data features; the DAS data is subjected to multi-scale convolution adapted to the multi-scale decomposition to obtain scale weights in the scale recovery module; different directional sorting is used to model seismic signals propagating to the DAS device from different directions to obtain multiple scales; the multi-scale convolution represents processing DAS data based on depth-separable convolution branches; based on the obtained scale weights and multiple scales, feature scale recovery guided by scale weights is performed to extract multi-directional information at different scales, obtaining three multi-directional features; and upsampling and remapping are performed based on the multi-directional features.

[0007] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0008] 1. This invention performs multi-scale decomposition on DAS data based on the input seismic phase picking model to obtain multi-scale DAS data features, which helps to characterize local abrupt changes, cross-channel continuous propagation, and overall propagation trends of seismic phase signals from different time sampling ranges and different spatial gathers. By performing multi-scale convolution on DAS data adapted to multi-scale decomposition, scale weights are obtained in the scale recovery module. Different directional sorting is used to model seismic signals propagating to the DAS device from different directions to obtain multi-scale features, which helps to enhance the network's ability to express oblique propagation texture, curved energy bands, and seismic phase responses in different directions. Based on the obtained scale weights and multi-scale features, feature scale recovery guided by scale weights is performed, which helps to achieve accurate correspondence between deep semantic features and the original time sampling scale and spatial gather scale, thereby helping to achieve stable picking of weak phases, overlapping phases, and phases with changing propagation directions in complex DAS records.

[0009] 2. By employing three depthwise separable convolutions with different strides, corresponding DAS features are extracted at different sampling scales, resulting in DAS features at three scales and multi-scale DAS data features. The DAS data is then processed using at least three depthwise separable convolution branches, each with a different kernel size, and each outputs DAS features at its corresponding scale. Existing technologies suffer from drawbacks such as using a single convolution scale or a fixed receptive field to extract DAS phase features, making it difficult to simultaneously capture local details and long-distance propagation trends. This approach is more conducive to simultaneously capturing short-term phase boundaries, the continuity of adjacent gathers, and wavefield evolution information over a larger area, thus improving the completeness of the DAS phase picking network's representation of multi-scale phase features and the stability of feature extraction.

[0010] 3. By using a multi-scale Mamba module, depthwise separable convolutions of a preset size are used to reduce the dimensionality of the input feature channels. The feature maps are arranged sequentially from four directions and four VSS Blocks with independent parameters are used to extract the directional features of the four directions. The directional features of the four directions are concatenated in the channel dimension, and the output of the multi-directional Mamba module is obtained through a fully connected layer. In contrast, existing technologies mainly model along a single time direction or a single spatial direction, which is difficult to fully express the oblique energy bands and multi-directional propagation textures in the DAS spatiotemporal map. This scheme is more conducive to independently modeling and fusing the phase response in different propagation directions, and is more conducive to enhancing the network's ability to identify phases with changing directions, cross phases, and curved propagation trajectories.

[0011] 4. By extracting multi-directional information at different scales from three multi-scale features through three independent multi-directional Mamba modules, three multi-directional features are obtained. Existing technologies have the disadvantages of sharing the same feature extraction structure for features at different scales, which can easily lead to feature aliasing between scales or insufficient expression of directional information. Compared with existing technologies, this approach is more conducive to preserving the corresponding directional propagation features and spatiotemporal dependencies at each scale, thereby helping to improve the discriminability between multi-scale features, the effectiveness of fusion, and the accuracy of seismic phase boundary location in complex backgrounds.

[0012] 5. When aftershocks are frequent after a main shock or small earthquakes occur densely along active fault lines, multiple seismic events may occur consecutively within a short period of time. When the arrival time distances between different seismic phases are relatively short, time overlap interference parameters are generated based on the crossover or overlap of multi-scale DAS data features within the same time window. Propagation direction change parameters are generated based on the arrival time changes of multi-scale DAS data features between adjacent gathers. When the preset stability conditions are not met, stability repair of time overlap interference and propagation direction change interference changes is activated. Compared with the shortcomings of existing technologies that directly extract features from DAS data with overlapping seismic phases or abrupt changes in propagation direction, and are prone to mixing noise pulses or adjacent seismic phases into the same seismic phase response, this technology helps to make targeted corrections to unstable time windows and abnormal propagation areas before subsequent convolutional feature extraction, thereby improving the identification of weak seismic phases, the continuity of seismic phase boundaries, and the accuracy of arrival time picking. Attached Figure Description

[0013] Figure 1 This is a flowchart of a multi-scale feature extraction and scale recovery method for DAS phase picking provided in an embodiment of the present invention;

[0014] Figure 2 This is a schematic diagram of phase pickup provided in an embodiment of the present invention;

[0015] Figure 3This is a schematic diagram of scale recovery provided in an embodiment of the present invention;

[0016] Figure 4 This is a multi-directional Mamba processing diagram provided in an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0018] Example 1: This embodiment of the invention provides a method for multi-scale feature extraction and scale recovery for DAS phase picking. For example... Figure 1 The flowchart shown is for a multi-scale feature extraction and scale recovery method for DAS phase picking. The processing flow of this method may include the following steps:

[0019] In the established DAS phase picking process, DAS data based on the input phase picking model is decomposed into multiple scales to obtain multi-scale DAS data features. This enables feature extraction at different scales while preserving local and global information from the original DAS data. By performing multi-scale decomposition on the DAS data at three scales, DAS phase signal features can be extracted from different time sampling ranges and spatial gathers. This preserves local amplitude abrupt changes and phase boundary details at small scales, as well as cross-channel propagation trends and global wavefield changes at large scales. This avoids the problem of losing weak phase details or insufficient background trend expression caused by relying solely on single-scale features, thereby improving the ability to characterize multi-scale phase features in complex DAS data.

[0020] Applying multi-scale convolution to DAS data to adapt to multi-scale decomposition and obtain scale weights in the scale recovery module helps avoid excessive influence of noise at a single scale on the overall feature extraction, promoting the representation of effective features in the final output features. Different directional sorting is used to model seismic signals propagating to the DAS device from different directions to obtain multi-scale data. Multi-scale convolution is based on depth-separable convolution branches to process DAS data, fully utilizing the spatiotemporal characteristics of DAS data to enhance the module's feature extraction capabilities. By applying multi-scale convolution to DAS data to adapt to multi-scale decomposition and obtaining scale weights for multiple scales in the scale recovery module, adaptive adjustments can be made based on the contribution of different scale features to the current phase picking task, reducing the interference of noisy or redundant scales on the final features, and enhancing scale features related to phase boundaries, propagation continuity, and effective energy response. Meanwhile, by modeling seismic signals propagating to the DAS device from different directions through sorting in four different directions, the phase texture features in the time direction, spatial gather direction, and oblique propagation direction can be fully extracted, improving the network's ability to identify changes in propagation direction, energy band intersections, and curved propagation trajectories.

[0021] Based on the obtained scale weights and multi-scale data, scale-weighted feature recovery is performed to extract multi-directional information at different scales, resulting in three multi-directional features. Upsampling and remapping are then performed based on these features, which helps ensure the flexibility and plug-and-play nature of the multi-scale Mamba module. S2 is used to perform multi-scale convolution on DAS data adapted to multi-scale decomposition, obtaining the scale weights of multiple scales in the scale recovery module. This allows for adaptive adjustment based on the contribution of different scale features to the current phase picking task, reducing the interference of noisy or redundant scales on the final features and enhancing scale features related to phase boundaries, propagation continuity, and effective energy response. Simultaneously, by sorting the seismic signals propagating to the DAS device from four different directions, modeling can fully extract phase texture features in the temporal direction, spatial gather direction, and oblique propagation direction, improving the network's ability to identify changes in propagation direction, energy band intersections, and curved propagation trajectories.

[0022] In this embodiment, the above method helps to improve the DAS phase picking network's ability to extract and fuse complex spatiotemporal features at multiple scales and directions, and reduces the impact of single-scale noise, superposition of adjacent phases, and changes in propagation direction on the picking results. This helps to improve the accuracy of weak phase identification, the stability of phase arrival time picking, and the continuity of adjacent gather picking results.

[0023] Furthermore, multi-scale decomposition is performed on the DAS data based on the input phase picking model to obtain the multi-scale DAS data features. The specific process is as follows: For the DAS data... Multi-scale feature extraction is performed, where C represents the number of feature channels in the DAS data, H represents the size of the feature map in the height direction, and W represents the size of the feature map in the width direction.

[0024] Multi-scale feature extraction is performed, specifically including: employing three depthwise separable convolutions with different strides to extract corresponding DAS features at different sampling scales, obtaining DAS features at three scales, and thus multi-scale DAS data features, including features F1, F2, and F3 at the three scales; wherein the stride settings for each depthwise separable convolution are as follows: =2, =4, =8, the specific method for obtaining the feature F of multi-scale DAS data is as follows:

[0025]

[0026] Among them, K p K represents a depthwise convolution kernel that performs a convolution operation in a single channel. d This represents pointwise convolution, used for cross-channel information fusion. This represents the stride of the convolution for the i-th feature. For convolution operations, This represents the ReLU activation function, used to introduce nonlinear features.

[0027] By performing multi-scale decomposition of DAS data at three scales and using depthwise separable convolutions with different amplitudes to extract DAS features, it is helpful to improve the completeness of the DAS phase picking network in extracting phase features under weak phases, overlapping phases, and complex background noise.

[0028] The scale weights in the scale recovery module are obtained by performing multi-scale convolution on the DAS data to adapt to multi-scale decomposition. Specifically, multi-scale convolution processing is performed on the DAS data; the scale weights in the scale recovery module, i.e., scale attention weights, are obtained through remapping and max pooling.

[0029] The multi-scale convolution processing includes: processing DAS data using at least three depthwise separable convolutional branches, each with a different kernel size, and outputting DAS features at the corresponding scale. By setting multiple depthwise separable convolutional branches adapted to multi-scale decomposition and making each branch have a different kernel size, DAS feature responses can be further extracted from different receptive fields, making the kernel size match the temporal sampling range and spatial gather range at the corresponding scale. For example, this invention chooses to use strides of 2, 4, and 8 in the multi-scale decomposition stage, so 2x2, 4x4, and 8x8 convolutional kernels are used to extract the importance of features at different scales in the original features. Then, scale attention weights in the format of sx1 are obtained through remapping and max pooling, where s is the number of scales, which is 3 here. In addition, by remapping and max pooling the DAS convolutional features at different scales, scale attention weights in the format of s×1 are obtained, which can adaptively represent the contribution of each scale feature in scale recovery. In the subsequent feature scale recovery process, more effective scale features can be picked up for the current seismic phase by enhancing the scale attention weight.

[0030] In this embodiment, the above processing helps to improve the DAS phase picking network's ability to express multi-scale phase features and its feature selection ability in the scale recovery process, enhances the identification stability of weak phases, overlapping phases and phase boundaries under complex noise backgrounds, thereby improving the accuracy and continuity of the final phase picking results.

[0031] Furthermore, different directional sorting is used to model seismic signals propagating to the DAS device from different directions to obtain multi-scale data, specifically:

[0032] Based on the channel attention mechanism, channel attention weights representing the importance of different channels in the original features are obtained from the DAS data. This is used to standardize the subsequent splicing of outputs from different VSS blocks, avoiding the omission of important information in the input DAS data. Simultaneously, it can adaptively evaluate the effectiveness of different channels in the original features, assigning high weights to channel features related to phase-to-time boundaries, energy abrupt changes, cross-channel continuous propagation, and directional texture, while weakening channel features related to background noise, local impulse interference, or redundant responses. Specifically:

[0033] ;

[0034] Among them W c These are channel attention weights or scale attention weights, where W1 and W2 represent the weights passed through two fully connected networks, and the remaining weights represent the weights for X. m Global average pooling was performed, X m(h,w) represents the feature value of the m-th input feature map or the m-th channel feature at spatial position (h,w), where H represents the size of the feature map in the height direction, W represents the size of the feature map in the width direction, h represents the position index in the height direction, and the value ranges from 1 to H; w represents the position index in the width direction, and the value ranges from 1 to W.

[0035] In the multi-scale Mamba module, a depthwise separable convolution with a preset size, such as 1x1, is used to reduce the dimensionality of the input feature channels. The feature maps are arranged sequentially in four directions, and four VSS Blocks with independent parameters are used to extract directional features in the four directions. The four directions are: 1. Row-first from left to right, then from top to bottom; 2. Row-first from right to left, then from bottom to top; 3. Column-first from top to bottom, then from left to right; 4. Column-first from bottom to top, then from right to left. The directional features in the four directions are concatenated along the channel dimension, and the four concatenated directional features are normalized by channel attention. The output of the multi-directional Mamba module is obtained through a fully connected layer.

[0036] like Figure 2 The diagram shown illustrates the phase picking method provided in this application. By performing multi-scale decomposition on the input DAS data features, feature information within different resolution ranges is extracted. The decomposed data is then processed using a multi-directional Mamba module to extract and analyze directional depth features. Combined with the scale weights output by the scale attention module, the extracted features are reconstructed into a format matching the original distribution using a scale recovery module.

[0037] like Figure 3 The diagram illustrates the scale recovery provided in this application. The upper part of the diagram represents the main branch of scale recovery. Three multi-scale features are input into a multi-directional selective state space module. This module models the temporal, spatial gather, and oblique propagation features at different scales, yielding multi-directional enhanced features at the corresponding scales. Subsequently, the scale weights output by the scale attention module are used to weight the multi-directional features at different scales. The lower part of the diagram represents the scale attention module, which performs multi-scale convolution processing on the input C×H×W features using 2×2, 4×4, and 8x8 depthwise separable convolutions to obtain feature responses at different scales. The responses at each scale are then fused and channel remapping is performed using a 1×1 depthwise separable convolution. Max pooling is then used to generate scale attention weights with a size equal to one scale. Finally, multiple features restored to the same scale are concatenated along the channel dimension and then channel remapping is performed using a 1×1 depthwise separable convolution to obtain output features of size C×H×W.

[0038] like Figure 4The diagram shows the multi-directional selective state-space processing method provided in this application: The input is a DAS input feature of size C×H×W. The module first generates C×1 channel attention weights through average pooling and activation operations to characterize the importance of different channel features. Simultaneously, the input features are fed into four 1×1 depthwise separable convolutional branches for channel dimensionality reduction, resulting in dimensionality-reduced features of size C / 4×H×W. These features are then arranged according to four different scanning orders (direction 1, direction 2, direction 3, and direction 4) and input into the corresponding visual state-space modules for directional feature extraction to model the phase texture, cross-channel continuity, and diagonal energy band changes in different propagation directions of the DAS spatiotemporal features. The directional features extracted by each directional branch are then concatenated along the channel dimension and normalized using the channel attention weights. Features consistent with the effective phase propagation direction are enhanced, while noise or redundant directional features are suppressed. Afterward, nonlinear mapping and feature fusion are performed through a fully connected layer, outputting enhanced features of size C×H×W. The lower part of the figure further shows the internal structure of the visual state space module, which includes layer normalization, a two-dimensional selective scanning module, and a fully connected layer. The two-dimensional selective scanning module realizes the state space modeling of two-dimensional DAS features through linear layers, depthwise separable convolution, activation functions, two-dimensional selective scanning, layer normalization, and linear layer output mapping, thereby improving the module's ability to extract multi-directional phase propagation features and long-distance spatiotemporal dependencies.

[0039] By using depthwise separable convolutions of a preset size in the multi-scale Mamba module to reduce the dimensionality of the input feature channels, the number of feature channels can be reduced while preserving the main phase response information. This reduces the computational cost and parameter count when the subsequent VSS Block (Visual State Space Block) performs multi-directional sequence modeling, and also reduces the interference of redundant channels on the extraction of propagation direction features. By arranging the feature maps sequentially from four directions and using four VSS Blocks with independent parameters for feature extraction, the phase picking model can model the spatiotemporal features of DAS from different scanning directions, capturing phase texture and energy band changes in the time direction, spatial gather direction, and different oblique propagation directions.

[0040] It should be added that the phase picking model for DAS phase picking and scale recovery provided in this application can be constructed using one or more of the following: convolutional neural network, U-Net-like encoder-decoder network, feature pyramid network, Transformer-like spatiotemporal modeling network, Mamba-like state space model, VMamba model, VSS Block, PhaseNet-like phase picking network, Earthquake Transformer-like phase detection network, and attention enhancement network. It can realize multi-scale feature extraction, multi-directional spatiotemporal modeling, scale weight generation, and feature scale recovery of DAS data.

[0041] By concatenating the directional features from the four directions along the channel dimension and using channel attention to normalize the concatenated directional features, we can further adaptively filter them based on their contribution to the current seismic phase picking task. This enhances directional features consistent with the actual seismic phase propagation direction, while suppressing directional features that are inconsistent or heavily affected by noise. Furthermore, by obtaining the output of the multi-directional Mamba module through a fully connected layer, we can fuse the seismic phase propagation information from the four directions into a unified spatiotemporal enhancement feature, providing a more stable feature foundation for subsequent scale recovery and seismic phase picking.

[0042] It's important to explain that because Mamba is sensitive to sequence order, single-type serialization imposes a fixed directional bias on DAS data, a feature map containing spatiotemporal information. This makes it difficult for the phase picking model to simultaneously capture cross-channel and temporal feature information. Arranging DAS data in four directions—row-first, column-first, and forward / backward—can cover as many anisotropic patterns as possible in DAS data, such as propagation along the fiber, folding back, and delay, thus enhancing the phase picking model's ability to model spatiotemporal features. Furthermore, combining the four parameter-independent VSS Blocks used in the phase picking model enables modeling of large-scale spatiotemporal and directional information, overcoming the limitation of traditional two-dimensional convolution primarily modeling local neighborhoods.

[0043] The modeling process involves obtaining multi-scale data and also includes verifying the effectiveness of the multi-directional Mamba module based on ablation experiments. Table 1 shows the ablation experiment results provided in this application. The specific experimental results are as follows:

[0044] Table 1 Ablation Experiment Results

[0045]

[0046] The models in the table all use the phase picking model presented in this paper, with modifications only to the multi-directional Mamba module. The single-directional Mamba module replaces the multi-directional Mamba module with a single-directional Mamba module. Shared weights are achieved by using the same parameters for the four VSSB blocks in the model but without channel attention. Independent weights are achieved by using four independent VSS blocks. Four-directional Mamba plus channel attention constitutes the complete multi-directional Mamba module. For model performance calculation, statistics are performed at the channel level, meaning each channel is treated as a sample for metric calculation. Furthermore, a threshold of 0.5s is used in the experiment; a phase picking error of no more than 0.5s is considered correct. All phase picking models in the table use the same locally acquired and labeled training and test sets.

[0047] In ablation experiments, as the phase picking model gradually expanded from a single-direction Mamba to four-direction shared weights, then to four-direction independent weights, and finally combined with channel attention, the precision, recall, and F1 scores of both P-waves and S-waves showed a steady improvement. The multi-directional arrangement effectively alleviated the directional bias caused by a single direction, reducing false negatives in the phase picking model. By setting the four VSS blocks as independent parameters, the phase picking model could more fully fit seismic signals from different directions, thus further reducing false positives while maintaining high precision. Introducing channel attention further highlighted consistent seismic signals from multiple directions and suppressed noise inconsistencies between directions after stitching together features from the four directions, resulting in the best overall performance. Ablation experiments showed that both P-wave and S-wave phase picking results benefited from the multi-directional Mamba, with the gain being more significant for S-waves.

[0048] In addition, the multi-directional, multi-scale Mamba module was compared with the traditional phase picking model, as shown in Table 2, which is a comparison table of the multi-directional, multi-scale Mamba module provided in this application and the traditional phase picking model:

[0049] Table 2 Comparison between the multi-directional, multi-scale Mamba module and the traditional phase picking model

[0050]

[0051] The multi-directional, multi-scale Mamba module represents the module added to each skip connection point in PhaseNet-DAS. PhaseNet-DAS and the multi-directional, multi-scale Mamba module were trained and tested using the same DAS data. EQTransformer and PhaseNet performed phase picking every three channels in the DAS data. The final results show that traditional models designed for three-channel seismic data are not well-suited for phase picking in DAS data.

[0052] In this embodiment, the above processing enables the DAS phase picking network to no longer rely solely on single-direction or local temporal features, but instead combines channel attention, channel dimensionality reduction, four-directional VSS block modeling, and directional feature fusion to jointly characterize phase responses at different propagation directions, different spatial gathers, and different temporal sampling locations in DAS data. This helps improve the network's ability to identify oblique energy bands, curved propagation trajectories, intersections of adjacent phases, and weak phase boundaries, while reducing the impact of background noise and missing single-direction features on the phase picking results.

[0053] Furthermore, based on the obtained scale weights and multi-scale data, feature scale recovery guided by the scale weights is performed. The specific process is as follows: The DAS data... Multiscale decomposition yields three multiscale features F1, F2, and F3, with dimensions respectively. , as well as The DAS data is input into the scale attention module to obtain a 3-row, 1-column scale attention weight W. c The three multi-scale features are extracted using three independent multi-directional Mamba modules to obtain multi-directional information at different scales. Seismic signals with good characterization across different directions are preserved, while noise inconsistencies between directions are suppressed, resulting in three multi-directional features, including F... d1 F d2 and F d3 In the DAS (Digital Seismic Analysis) phase picking scenario, it can be used to characterize phase response information in different directions, such as waveform variation features in the time direction, continuous propagation features in the spatial gather direction, and phase texture features in the oblique propagation direction. Multi-directional features are weighted based on scale attention weights, and the weighted features are upsampled and remapped. The upsampling uses a combination of bilinear interpolation and 3x3 convolution. Each bilinear interpolation amplifies the feature by a factor of two, followed by remapping using 3x3 convolution. For features at three different scales, upsampling is performed 1, 2, and 3 times respectively, sampling all three multi-scale features to... Scale. Subsequently, features of the three same scale are concatenated at the channel level and remapped using a 1x1 depthwise separable convolution to restore the feature scale, ensuring the module's flexibility and plug-and-play capability. Based on scale attention weights, multi-directional features are weighted, which can adaptively enhance the phase texture in the effective propagation direction and suppress interfering features in noise or redundant directions according to the importance of phase features in different directions in the current DAS data. Specifically:

[0054] ;

[0055] ;

[0056] ;

[0057] In the formula, F D1 F D2 and F D3 W represents the weighted multi-directional features. c [0]、W c [1] and W c [2] represent the first, second and third weight components in the attention weights of this scale, respectively.

[0058] In this embodiment, by decomposing DAS data into three multi-scale features of different sizes, the network can focus on both local boundary details near the arrival time of the seismic phase and characterize cross-channel propagation trends and wavefield evolution relationships over a larger area. Since the three scale features have different spatial dimensions, direct fusion can easily lead to scale misalignment or unbalanced feature contributions. Therefore, by obtaining scale attention weights through a scale attention module, the importance of different scale features in the current seismic phase picking task can be adaptively evaluated.

[0059] By inputting the three multi-scale features into three independent multi-directional Mamba modules, phase propagation information in different directions can be extracted at each scale. Since each scale is processed by an independent multi-directional Mamba module, feature interference between different scales can be reduced, and the directional features at different scales can be fully expressed, thereby improving the ability to identify obliquely propagating phases, curved energy bands, and multi-directional phase intersection areas.

[0060] By weighting the three multi-directional features based on scale attention weights, the multi-directional features at different scales can participate in scale reconstruction according to their actual contribution to phase picking, avoiding the problem of weakening effective phase features caused by simple splicing or averaging fusion. This weighting process can highlight phase signals with stable responses in multiple directions and suppress noise interference that only appears in local directions or a single scale, thereby improving the reliability and discriminativeness of the reconstructed features.

[0061] By employing a combination of bilinear interpolation and 3×3 convolution to upsample and remap the weighted features, spatial differences between scales can be smoothed while gradually restoring the feature size. Furthermore, 3×3 convolution is used to reorganize the interpolated local neighborhood features, reducing boundary blurring or detail loss that may result from simple interpolation. Upsampling is performed a corresponding number of times for features at three different scales, ultimately restoring them to a unified C×H×W scale. This helps ensure the correspondence between the restored features and the original DAS data in terms of temporal sampling points and spatial gather locations, thereby improving the accuracy of seismic phase boundary localization and arrival picking.

[0062] The above processing can establish an effective connection between multi-scale decomposition, multi-directional modeling and scale attention weighting, so that the DAS phase picking network can retain both the global propagation semantics in deep features and the local phase boundaries and gather positional relationships in the original DAS data when restoring the output scale. This helps to improve the weak phase identification capability, phase arrival time positioning accuracy, adjacent gather picking continuity and the stability of phase picking results under complex noise background.

[0063] Example 2, based on Example 1, describes how, during DAS phase acquisition, when aftershocks are frequent after the mainshock or small earthquakes occur densely along active fault zones, multiple source events may occur consecutively within a short period. The short arrival times between different phases can easily lead to the superposition of P-waves, S-waves, or phases from multiple events. This results in multiple energy band intersections within the same time window in the DAS record, blurred local phase boundaries, decreased continuity of arrival times between adjacent gathers, and weak phase characteristics being masked by strong phases or noise. Specifically, DAS data typically constitutes a two-dimensional spatiotemporal signal with both temporal and spatial gather dimensions. Different phases in this two-dimensional signal not only exhibit abrupt amplitude changes but also show oblique textures, curved trajectories, or multi-directional energy distributions formed along the fiber optic direction. When multiple phases are close in time or space, different directions... Different scales of seismic phase features may overlap in local areas, making it difficult for waveform changes within a single time window to accurately correspond to specific phase boundaries. This paper addresses this issue by performing multi-scale decomposition on DAS data input to the phase picking model to obtain multi-scale DAS data features. It also includes a phase picking stability analysis to address the impact of adjacent phase overlap and changes in phase propagation direction on DAS phase picking. The specific process is as follows: Based on the intersection or overlap of multi-scale DAS data features within the same time window, a time overlap interference parameter is generated; based on the time-to-arrival changes of multi-scale DAS data features between adjacent gathers, a propagation direction change parameter is generated; before proceeding to depthwise separable convolutional feature extraction, it is first determined whether there are unstable factors such as adjacent phase overlap or abrupt changes in propagation direction in the current DAS data. Compared to directly performing uniform convolution processing on all DAS data, this method can identify abnormal time windows that significantly affect phase boundary location and cross-gather continuity in advance, avoiding overlapping phases, strong noise, or local propagation direction changes from being directly input as effective features into subsequent networks, thereby improving the reliability of subsequent multi-scale feature extraction.

[0064] The time overlap interference parameter specifically refers to a parameter used to characterize the degree of overlap of different candidate seismic phase energy bands in terms of time sampling points, spatial gather positions, and local energy responses within the same time window. It includes at least the overlap time length of the seismic phase energy bands. The overlap time length of the seismic phase energy bands represents the time sampling length of different seismic phase energy bands monitored by timers and counters that have overlapping responses within the same time window. A seismic phase energy band represents a band-shaped characteristic region in the DAS spatiotemporal data matrix, formed by continuous time sampling points and continuous spatial gather positions, exhibiting a seismic phase propagation trend. The DAS spatiotemporal data matrix represents a data matrix formed by arranging DAS data in two dimensions according to the sampling time order and the spatial arrangement order of fiber optic gathers. One dimension of the matrix corresponds to the time sampling point, and the other dimension corresponds to different fiber optic gathers or sensing channels. Matrix elements represent the vibration amplitude at the corresponding time sampling point and the corresponding gather position. For example, within a time window, the first seismic phase energy band is sampled at time points 1000 to 1080. A continuous high-energy band-shaped region is formed within the range of sample points and gather numbers 20 to 45. The second phase energy band forms another continuous high-energy band-shaped region within the range of time sampling points 1040 to 1120 and gather numbers 35 to 60. The interval corresponding to the 1040th to 1080th time sampling points can be regarded as the overlap time range of the two phase energy bands, and its continuous sampling length can be regarded as the overlap time length of the phase energy bands. The propagation direction change parameter specifically refers to the parameter used to characterize the degree of change in the propagation direction of the phase energy band between adjacent gathers. It includes at least the slope of the arrival time change of adjacent gathers. Adjacent gathers refer to two or more sampling channels arranged in the DAS fiber optic sensing channel according to the fiber optic deployment direction or spatial position and whose gather numbers or spatial distances are adjacent. The slope of the arrival time change of adjacent gathers represents the ratio between the arrival time difference of the same phase energy band between adjacent gathers and the gather spacing, as monitored by DAS acquisition equipment, such as a distributed fiber optic acoustic wave sensor demodulator or a DAS data acquisition host.

[0065] The stability of phase picking is determined by conditions for temporal overlap interference parameters and propagation direction change parameters. When the preset stability conditions are not met, it indicates that there is a risk of instability in phase picking caused by the superposition of adjacent phases, abrupt changes in propagation direction, or abnormal slope fluctuations within the current time window. Stability repair for temporal overlap interference and propagation direction change interference is then initiated. After repair, the DAS data is subjected to depthwise separable convolution with different kernel sizes that are adapted to multi-scale decomposition. When the preset stability conditions are met, it indicates that the degree of overlap of phase energy bands and the degree of change in propagation direction within the current time window are within the allowable range, and the current multi-scale DAS data features do not show obvious phase superposition or abrupt changes in propagation direction that would affect subsequent feature extraction. The DAS data is subjected to depthwise separable convolution with different kernel sizes that are adapted to multi-scale decomposition. The preset stability conditions include: the temporal overlap interference parameter is less than or equal to the preset temporal overlap threshold; and the propagation direction change parameter is less than or equal to the preset propagation direction change threshold.

[0066] Among them, preset thresholds such as preset time overlap threshold and preset propagation direction change threshold are stored in advance in the database of this application. Taking the preset time overlap threshold as an example, the setting process of the preset threshold is explained:

[0067] Historical time overlap interference parameters are obtained, including at least the overlap time length of seismic phase energy bands. Based on manually marked seismic phase boundaries or manually marked seismic phase arrival times, the historical time window is divided into a stable picking sample set and an unstable picking sample set. The stable picking sample set indicates that seismic phase overlap within the corresponding time window does not significantly affect the seismic phase picking results, while the unstable picking sample set indicates that seismic phase superposition, weak seismic phase obscuring, or picking offset exist within the corresponding time window. The distribution range of the historical seismic phase energy band overlap time length and the historical overlap energy percentage in the stable and unstable picking sample sets are statistically analyzed, and a critical value that can distinguish between the stable and unstable picking sample sets is selected as the threshold. The candidate time overlap threshold is used as the basis for determining the stability decision based on the candidate time overlap threshold. This threshold is then applied to the validation sample set to calculate the degree of consistency between the stability decision result obtained based on the candidate time overlap threshold and the manually labeled stability result. This is the proportion of samples whose stability decision result is consistent with the manually labeled stability result to the total number of validation samples. When the degree of consistency meets the preset consistency condition, the candidate time overlap threshold is determined as the preset time overlap threshold and associated with the corresponding sampling frequency, time window length, gather spacing, and application scenario identifier in the database. When the degree of consistency does not meet the preset consistency condition, the candidate time overlap threshold is updated, and the validation is re-executed until the preset time overlap threshold that meets the preset consistency condition is obtained.

[0068] For example, in an earthquake DAS monitoring scenario, the DAS sampling frequency is 1000Hz, the length of a single analysis time window is 200ms, and the gather spacing is 1m. Historical time overlap interference parameters were obtained for 500 historical time windows, with the overlap time length of the seismic phase energy bands statistically analyzed for each window. Based on manually annotated seismic phase boundaries and arrival times, 320 historical time windows were divided into a stable picking sample set, and 180 historical time windows were divided into an unstable picking sample set. Statistical analysis revealed that in the stable picking sample set, the overlap time length of the seismic phase energy bands mainly ranged from 0ms to 28ms; in the unstable picking sample set, the overlap time length mainly ranged from 35ms to 80ms. Therefore, 30ms was selected as a candidate overlap time length threshold.

[0069] Subsequently, the candidate time overlap threshold was applied to 100 validation samples that were not included in the threshold statistics. When the overlap time of the seismic energy band within a certain validation time window was 42ms, it was determined to be an unstable picking time window because it was greater than the candidate overlap time length threshold of 30ms. If this determination result was consistent with the manually labeled stability result, it was recorded as a correct determination. After statistical analysis of the 100 validation samples, if the consistency rate between the stability judgment result obtained based on the candidate time overlap threshold and the manually labeled stability result reached 92%, which was greater than the preset consistency rate threshold of 90%, and the average value of the seismic phase picking offset obtained after stability judgment based on the candidate time overlap threshold was 2.5ms, which was less than the preset picking offset threshold of 5ms, then the candidate overlap time length threshold of 30ms was set, and it was associated with a sampling frequency of 1000Hz, a time window length of 200ms, and a gather spacing of 1m and stored in the database.

[0070] If the consistency rate is only 82% during the verification process, or the average value of the phase picking offset is greater than the preset picking offset threshold, it indicates that the current candidate time overlap threshold is not effective in distinguishing between stable and unstable samples. In this case, the candidate overlap time length threshold can be adjusted from 30ms to 32ms or 28ms, and the stability judgment of the verification sample set can be re-performed until the preset time overlap threshold that meets the preset consistency condition and the preset picking offset condition is obtained.

[0071] Specifically, the stability repair process for time overlap interference and propagation direction change interference is initiated as follows: Based on the time overlap interference parameters and propagation direction change parameters, the target time window range is adjusted according to the time window correction parameters, including the time window start point, time window end point, window length, and sliding step size. By adjusting the target time window range according to the time window correction parameters, the time window to be repaired can more accurately cover the overlapping area of ​​the seismic phase energy band and its preceding and following propagation transition areas, avoiding incomplete analysis of overlapping seismic phases due to a time window that is too short, and avoiding the introduction of too much irrelevant noise due to a time window that is too long. By adaptively adjusting the start and end points of the time window, its length, and the sliding step, the network's tracking accuracy in areas with overlapping local seismic phases, areas where weak seismic phases are obscured, and areas where the propagation direction changes can be improved. For example, the original time window is from the 1000th to the 1100th time sampling point, the window length is 100 sampling points, and the sliding step is 50 sampling points. When overlapping energy bands of seismic phases are detected within the 1030th to 1070th time sampling points, and the preset stability conditions are not met, the start point of the time window can be adjusted to the 1010th time sampling point, the end point can be adjusted to the 1090th time sampling point, or it can be further expanded to the 1000th to the 1120th time sampling point, and the sliding step can be reduced from 50 sampling points to 25 sampling points.

[0072] Based on the time overlap interference parameters and propagation direction change parameters, the slope of the time change between adjacent gathers is smoothed according to the propagation direction smoothing parameters, so that the repaired phase energy band maintains a relatively continuous propagation trend along the fiber optic channel. The fitting window length is used to determine the number of continuous gathers participating in the propagation direction curve fitting; the target time window range represents the time sampling interval corresponding to when the preset stability conditions are not met.

[0073] The arrival time variation slope between adjacent gathers is smoothed according to the propagation direction smoothing parameter. Specifically, taking the current gather as the center, adjacent gathers before and after the current gather are selected according to the fitting window length. Based on the neighborhood weighted smoothing algorithm, the arrival time variation slope within the fitting window length is smoothed. This helps to reduce the influence of abnormal arrival times, noise pulses, or local strong phase interference from a single gather on the propagation direction judgment, so that the restored phase energy band maintains a more continuous propagation trend along the fiber optic channel. Especially in the case of dense aftershocks after the mainshock or continuous small earthquakes in active fault zones, this processing helps to reduce problems such as local boundary jumps, fault picking by adjacent gathers, and misjudgment of propagation direction. For example, when the current gather is gather 50 and the fitting window length is 5, gathers 48, 49, 50, 51, and 52 are selected as the local fitting range. The neighborhood weighted smoothing algorithm can cover local weighted regression, weighted moving average, Gaussian weighted smoothing, and other methods.

[0074] The time window correction parameters are obtained by inputting the time overlap interference parameters and the propagation direction change parameters into the output of the time window correction model.

[0075] Taking the gradient boosting tree model as an example, the construction process of the time window correction model is explained as follows: Historical time overlap interference parameters, propagation direction change parameters, and corresponding manually labeled time window correction parameters are obtained. The historical time overlap interference parameters and historical propagation direction change parameters are normalized, and each sample number is matched with its corresponding manually labeled time window correction parameter to form a training sample set. The normalized historical time overlap interference parameters and historical propagation direction change parameters are used as input features, and the manually labeled time window correction parameters are used as output labels for supervised training of the gradient boosting tree model. During training, an initial regression tree is first constructed to output the initial time window correction result. Then, based on the residual between the initial time window correction result and the manually labeled time window correction parameters, subsequent regression trees are constructed round by round, and the residual is fitted, allowing the model to gradually learn the mapping relationship between the degree of time overlap interference, the degree of propagation direction change, and the start point, end point, window length, and sliding step size of the time window. When the model is validated... When the correction error on the samples meets the preset error condition, the trained time window correction model is obtained. During the verification process, the historical time overlap interference parameters and propagation direction change parameters that were not involved in the training are used as the verification sample set and input into the trained time window correction model to obtain the predicted time window correction parameters. The predicted time window correction parameters are compared with the corresponding manually labeled time window correction parameters, and the correction errors of the time window start point, the time window end point, the window length, and the sliding step size are calculated respectively. The weighted summation or mean square error of each error is then calculated to obtain the model correction error. When the model correction error is less than or equal to the preset correction error threshold, the time window correction model is determined to meet the verification requirements. When the model correction error is greater than the preset correction error threshold, the number of trees, tree depth, learning rate, or minimum number of sample splits in the gradient boosting tree model are adjusted, and the model is retrained and verified until the model correction error meets the preset error condition, resulting in a time window correction model that can be stored in the database.

[0076] The propagation direction smoothing parameters are obtained by inputting the temporal overlap interference parameters and propagation direction variation parameters into the output of the propagation direction smoothing model. By inputting the temporal overlap interference parameters and propagation direction variation parameters into the time window correction model and the propagation direction smoothing model respectively, the model can adaptively output the time window correction parameters and propagation direction smoothing parameters according to the current interference level. This allows stability repair to dynamically adjust based on the degree of phase superposition, the degree of propagation direction variation, and the local data state, rather than relying on fixed empirical parameters. Therefore, it helps improve adaptability to different seismic activity scenarios, different noise backgrounds, and different DAS deployment conditions.

[0077] The aforementioned time window correction model and propagation direction smoothing model are both models that are pre-trained and constructed based on machine learning models. The machine learning models include at least one of the following: random forest model, gradient boosting tree model, support vector regression model, multilayer perceptron model, convolutional neural network model, long short-term memory network model, and extreme gradient boosting model.

[0078] The construction of the propagation direction smoothing model is similar to that of the time window correction model. Its training samples include historical time overlap interference parameters, propagation direction change parameters, and corresponding manually labeled propagation direction smoothing parameters.

[0079] In this embodiment, the above processing can identify, determine, and repair the picking instability risks caused by the superposition of adjacent seismic phases and changes in propagation direction, so that the DAS data entering the subsequent deep separable convolution has clearer seismic phase boundaries, more continuous cross-channel propagation relationships, and more stable multi-scale feature expressions. This helps to reduce the missed picking, mispicking, and picking offset of seismic phases, improve the ability to identify weak seismic phases, the continuity of picking adjacent gathers, and the stability of seismic phase picking results under complex DAS records.

[0080] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0081] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0082] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0083] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0084] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other.

[0085] Artificial intelligence (AI) is the study of how computers can simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It encompasses both hardware and software technologies. AI hardware technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, and big data processing. AI software technologies mainly include computer vision, speech recognition, natural language processing, machine learning / deep learning, big data processing, and knowledge graph technologies.

[0086] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0087] The above description is only an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for multi-scale feature extraction and scale recovery in DAS phase picking, characterized in that, The method includes: Multi-scale decomposition of DAS data based on input seismic phase picking model is performed to obtain multi-scale DAS data features. Multi-scale convolution adapted to multi-scale decomposition is applied to DAS data to obtain scale weights in the scale recovery module. Seismic signals propagating to the DAS device from different directions are modeled using sorting in different directions to obtain multi-scale data. The multi-scale convolution represents DAS data processing based on depth-separable convolution branches. Based on the obtained scale weights and multiple scales, feature scale recovery guided by scale weights is performed to extract multi-directional information at different scales, resulting in three multi-directional features. Upsampling and remapping are then performed based on these multi-directional features.

2. The multi-scale feature extraction and scale recovery method for DAS phase picking as described in claim 1, characterized in that, The DAS data based on the input seismic phase picking model is decomposed into multi-scale features to obtain the multi-scale DAS data features. The specific process is as follows: Multi-scale feature extraction of DAS data specifically includes: using three depthwise separable convolutions with different strides to extract corresponding DAS features at different sampling scales, obtaining DAS features at three scales, and thus obtaining multi-scale DAS data features.

3. The multi-scale feature extraction and scale recovery method for DAS phase picking as described in claim 1, characterized in that, The process of adapting DAS data to multi-scale convolution for multi-scale decomposition to obtain scale weights in the scale recovery module is as follows: Perform multi-scale convolution processing on the DAS data; The scale weights, or scale attention weights, in the scale recovery module are obtained through remapping and max pooling.

4. The multi-scale feature extraction and scale recovery method for DAS phase picking as described in claim 3, characterized in that, The multi-scale convolution processing includes: processing the DAS data using at least three depthwise separable convolution branches, each depthwise separable convolution branch having a different convolution kernel size, and outputting DAS features at the corresponding scale.

5. The multi-scale feature extraction and scale recovery method for DAS phase picking as described in claim 1, characterized in that, The method of using different directional sorting to model seismic signals propagating to the DAS device from different directions to obtain multi-scale data specifically involves: Based on the channel attention mechanism, obtain the channel attention weights corresponding to the DAS data; In the multi-scale Mamba module, depthwise separable convolutions of a preset size are used to reduce the dimensionality of the input features channels; The feature maps are arranged sequentially from four directions, and feature extraction is performed using four VSS Blocks with independent parameters to obtain directional features in the four directions. The directional features of the four directions are concatenated in the channel dimension, and the output of the multi-directional Mamba module is obtained through a fully connected layer.

6. The multi-scale feature extraction and scale recovery method for DAS phase picking as described in claim 1, characterized in that, The feature scale recovery guided by the scale weights, based on the obtained scale weights and multiple scales, is performed as follows: The DAS data is decomposed into three multi-scale features; The DAS data is input into the scale attention module to obtain the scale attention weights; The three multi-scale features are processed by three independent multi-directional Mamba modules to extract multi-directional information at different scales, resulting in three multi-directional features. Multi-directional features are weighted based on scale attention weights, and the weighted features are then upsampled and remapped.

7. The multi-scale feature extraction and scale recovery method for DAS phase picking as described in claim 2, characterized in that, The DAS data based on the input phase picking model is decomposed into multiple scales to obtain the multi-scale DAS data features. This also includes a phase picking stability analysis to address the impact of adjacent phase superposition and changes in phase propagation direction on DAS phase picking. The specific process is as follows: Based on the crossover or overlap of multi-scale DAS data features within the same time window, a time overlap interference parameter is generated. Based on the time-varying characteristics of multi-scale DAS data between adjacent gathers, propagation direction variation parameters are generated.

8. The multi-scale feature extraction and scale recovery method for DAS phase picking as described in claim 7, characterized in that, The time overlap interference parameter specifically refers to the parameter used to characterize the degree of overlap of different candidate seismic phase energy bands in terms of time sampling points, spatial gather locations, and local energy responses within the same time window, and includes at least the overlap time length of the seismic phase energy bands. The propagation direction change parameter specifically refers to the parameter used to characterize the degree of change in the propagation direction of the seismic energy band between adjacent gathers, including at least the slope of the change when adjacent gathers arrive. The stability of phase picking is determined by the time overlap interference parameter and the propagation direction change parameter. When the preset stability condition is not met, the stability repair of the time overlap interference and the propagation direction change interference is started. After the repair is completed, the DAS data is subjected to depth separable convolution with different kernel sizes that are adapted to multi-scale decomposition. When the preset stability condition is met, the DAS data is subjected to depth separable convolution with different kernel sizes that are adapted to multi-scale decomposition. The preset stability conditions include: The time overlap interference parameter is less than or equal to the preset time overlap threshold; The propagation direction change parameter is less than or equal to the preset propagation direction change threshold.

9. The multi-scale feature extraction and scale recovery method for DAS phase picking as described in claim 8, characterized in that, The stability repair process for interference caused by overlapping start times and changes in propagation direction is as follows: Based on the time overlap interference parameters and propagation direction change parameters, the target time window range is adjusted according to the time window correction parameters, including the time window start point, time window end point, window length, and sliding step size; Based on the time overlap interference parameter and the propagation direction change parameter, the arrival time change slope between adjacent gathers is smoothed according to the propagation direction smoothing parameter.

10. The multi-scale feature extraction and scale recovery method for DAS phase picking as described in claim 9, characterized in that, The smoothing of the time-of-arrival slope between adjacent gathers according to the propagation direction smoothing parameter specifically means: taking the current gather as the center, selecting the gathers adjacent to the current gather before and after the current gather according to the fitting window length; Based on the neighborhood weighted smoothing algorithm, the slope of the time change within the fitting window length is smoothed. The time window correction parameters are obtained by inputting the time overlap interference parameters and the propagation direction change parameters into the output of the time window correction model. The propagation direction smoothing parameter is obtained by inputting the time overlap interference parameter and the propagation direction change parameter into the output of the propagation direction smoothing model.

Images