Deep fusion network microseismic positioning method and system based on velocity model constraint

Abstract

The application belongs to the technical field of microseismic event positioning, and in particular to a deep fusion network microseismic positioning method and system based on a velocity model constraint. A self-attention mechanism branch is used to extract branch global dependency features of seismic signal data. A convolutional neural network branch is used to extract branch local space-time features of the seismic signal data. The branch global dependency features and the branch local space-time features are subjected to weighted normalization fusion to obtain fusion features. The fusion features are subjected to a plurality of fully connected layers, and the three-dimensional position of a microseismic event is regressed. The application comprehensively utilizes global and local information, and greatly improves the microseismic event positioning accuracy.

Description

Technical Field

[0001] This application belongs to the field of microseismic event localization technology, specifically a deep fusion network microseismic localization method and system based on velocity model constraints. Background Technology

[0002] Microseismic event location plays a crucial role in rockburst risk early warning and fracturing production enhancement assessment. Although deep learning methods such as convolutional neural networks (CNN) can extract local spatial features from waveforms and use them for event location prediction, their location accuracy is highly sensitive to low signal-to-noise ratio (SNR) and incomplete data in complex geological environments, and their robustness is limited.

[0003] In recent years, deep learning has been introduced into the field of microseismic (MS) for denoising, event recognition, and first wave picking, and has overcome some of the limitations of traditional methods in MS localization tasks. However, existing works mostly rely on a single convolutional kernel for local feature extraction. The local receptive field limits the network's ability to capture global features and temporal dependencies of seismic waveforms. 2D convolutional isotropic feature learning is prone to introducing systematic localization biases in surface monitoring data. Methods such as U-Net also have a strong dependence on complete waveforms and are difficult to deal with missing or damaged records. Summary of the Invention

[0004] This application provides a deep fusion network microseismic location method and system based on velocity model constraints, which solves the problems of insufficient location accuracy and poor robustness of traditional convolutional neural networks in microseismic source location.

[0005] According to an embodiment of the first aspect of this application, a deep fusion network microseismic localization method based on velocity model constraints is provided, the method comprising:

[0006] The self-attention mechanism is used to extract the branch global dependency features of seismic signal data, including: linear mapping and adding position encoding to form sequence embedding to obtain input features, aggregating long-range dependencies and inter-station phase similarity in the time dimension through multiple multi-head self-attention and feedforward sub-layers to obtain time-series features, and then obtaining the branch global dependency features through pooling.

[0007] The branch local spatiotemporal features of seismic signal data are extracted using a convolutional neural network. The process includes: first, the initial convolution is projected onto a multi-channel feature space, then the local texture of time and frequency and cross-station is extracted through multiple coded residual blocks via local convolution, normalization and residual skip connection, and the receptive field is expanded by several layers of downsampling to obtain the branch local feature vector.

[0008] The fused feature is obtained by weighting and normalizing the global dependency features of branches and the local spatiotemporal features of branches;

[0009] By fusing features through several fully connected layers, the three-dimensional location of microseismic events can be regressed.

[0010] Furthermore, by aggregating long-range dependencies and inter-station phase similarities in the time dimension through multiple layers of multi-head self-attention and feedforward sub-layers, temporal features are obtained, including:

[0011] The multi-head self-attention mechanism employs an attention sub-layer, adding the attention sub-layer input and the attention sub-layer output as a residual connection, and then performing normalization processing.

[0012] After processing each time step independently with two fully connected layers and a ReLU function in the feedforward sublayer, the residuals are added to the output of the attention sublayer, and finally normalized to obtain the normalized temporal features.

[0013] Furthermore, the outputs of all encoded residual blocks are aggregated along the channel dimension and globally converged.

[0014] Furthermore, the fused features are obtained by weighted normalization fusion of the branch global dependency features and the branch local spatiotemporal features. This includes mapping the signal-to-noise ratio, mask, geometric information, and velocity voxels near the prediction location of each station to adaptive fusion weights through a lightweight network or gating unit, and then weighted normalization fusion of the branch global dependency features and the branch local spatiotemporal features to obtain the fused features.

[0015] A deep fusion network microseismic location system based on velocity model constraints, according to a second aspect of this application, includes: a deep fusion network model, after training, for processing seismic signal data and outputting the three-dimensional location of microseismic events;

[0016] The deep fusion network model includes:

[0017] The self-attention mechanism branch is used to extract the branch global dependency features of seismic signal data, including: linear mapping and adding position encoding to form sequence embedding to obtain input features, aggregating long-range dependencies and inter-station phase similarity in the time dimension through multiple multi-head self-attention and feedforward sub-layers to obtain temporal features, and obtaining branch global dependency features through pooling.

[0018] The convolutional neural network branch is used to extract the local spatiotemporal features of the seismic data. This includes: first, the initial convolution is projected onto the multi-channel feature space, and then the local texture of time and frequency and cross-station is extracted through multiple encoded residual blocks via local convolution, normalization and residual skip connection. The receptive field is expanded by several layers of downsampling to obtain the local feature vector of the branch.

[0019] A learnable weighted fusion layer is used to obtain fused features by weighted normalization fusion of global dependency features of branches and local spatiotemporal features of branches;

[0020] Several fully connected layers will fuse features to regress the three-dimensional location of microseismic events.

[0021] Furthermore, the multi-head self-attention mechanism employs an attention sub-layer, adding the attention sub-layer input and the attention sub-layer output as a residual connection for normalization processing;

[0022] After processing each time step independently with two fully connected layers and a ReLU function in the feedforward sublayer, the residuals are added to the output of the attention sublayer, and finally normalized to obtain the normalized temporal features.

[0023] Furthermore, convolutional neural network branches include:

[0024] The initial convolutional layer is used to project seismic signal data onto a multi-channel feature space;

[0025] Multiple coded residual blocks are concatenated to extract local textures of time and frequency and across stations through local convolution, normalization and residual skip connections;

[0026] Residual connection convolutional layers are used to connect the input and output residuals of encoded residual blocks.

[0027] Furthermore, the fused features are obtained by weighted normalization fusion of the global dependency features of the branch and the local spatiotemporal features of the branch. This includes mapping the signal-to-noise ratio, mask, geometric information of each station and the velocity voxels near the predicted location to adaptive fusion weights through a lightweight network or gating unit, and then weighted normalizing fusion of the global dependency features of the branch and the local spatiotemporal features of the branch to obtain the fused features.

[0028] Furthermore, during the training process, the deep fusion network model uses a three-dimensional velocity model to perform elastic wave forward modeling to obtain clean seismic signal data, and superimposes real-acquired background noise to construct training samples. By adjusting the signal-to-noise ratio and the noise weight factor of each detector, training samples with multiple noise levels are obtained.

[0029] Furthermore, the loss function of the deep fusion network model during training is:

[0030] ,

[0031] in This represents the number of training samples; No. The true coordinates of each sample; For predicting coordinates; For speed constraint coefficient, For indicator functions, This represents the wave velocity value at the coordinate position of the velocity model.

[0032] Compared with existing technologies, this application offers the following advantages: It comprehensively utilizes global and local information to significantly improve the accuracy of microseismic event location. By incorporating a three-dimensional velocity model into the loss function to create physical constraints, the prediction results conform to the propagation laws of seismic waves, maintaining high consistency even under complex velocity structures. When the signal-to-noise ratio decreases to −20dB, the average location error remains less than 4m; even with 20% of the geophones missing, the error remains within 7.3m, demonstrating significantly better robustness than traditional CNN methods. This application can locate 57 microseismic events in real time, with the location results highly consistent with the geological model. Attached Figure Description

[0033] Figure 1 This is a network architecture diagram of a deep fusion network microseismic location system based on velocity model constraints provided in an embodiment of this application.

[0034] Figure 2 The locations of the seismic sources in this application are derived from actual data. (a) is a three-dimensional perspective view, (b) is a three-dimensional point projection onto an east-west depth plane view, (c) is a three-dimensional point projection onto a north-south depth plane view, and (d) is a three-dimensional point projection onto an east-west-north-south plane view.

[0035] Figure 3 This is a spatial distribution map of well trajectories and microseismic events provided in the embodiments of this application. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0037] like Figure 1 As shown, this application provides a deep fusion network microseismic location system based on velocity model constraints, including: a deep fusion network model that, after training, is used to process seismic signal data and output the three-dimensional location of microseismic events;

[0038] The deep fusion network model includes:

[0039] The self-attention mechanism branch is used to extract the branch global dependency features of seismic signal data, including: linear mapping and adding position encoding to form sequence embedding to obtain input features, multiple Transformer layers to aggregate long-range dependencies and inter-station phase similarity in the time dimension to obtain temporal features, and pooling to obtain the branch global dependency features;

[0040] The convolutional neural network branch is used to extract the local spatiotemporal features of the seismic data. This includes: first, the initial convolution is projected onto the multi-channel feature space, and then the local texture of time and frequency and cross-station is extracted through multiple encoded residual blocks via local convolution, normalization and residual skip connection. The receptive field is expanded by several layers of downsampling to obtain the local feature vector of the branch.

[0041] A learnable weighted fusion layer is used to obtain fused features by weighted normalization fusion of global dependency features of branches and local spatiotemporal features of branches;

[0042] Several fully connected layers will fuse features to regress the three-dimensional location of microseismic events.

[0043] The deep fusion network model combines parallel convolutional neural network (CNN) branches and self-attention mechanism branches to form a comprehensive feature representation. Through this parallel fusion structure, global relevance and local structural features can be preserved at the same time, avoiding the limitations of traditional single-branch network feature learning.

[0044] The fusion output formula is: in, The three-dimensional spatial coordinates predicted by the model; The feature output of the self-attention mechanism branch; The feature output of the convolutional neural network branch; and These are the learnable fusion weights for the self-attention mechanism branch and the convolutional neural network branch, respectively; (symbol) This indicates a feature concatenation operation. This indicates a merged output.

[0045] In one embodiment, the Transformer layer in the self-attention mechanism branch adopts an attention sub-layer, and the input and output of the attention sub-layer are added together as a residual connection for normalization processing;

[0046] After processing each time step independently with two fully connected layers and a ReLU function in the feedforward sublayer, the residuals are added to the output of the attention sublayer, and finally normalized to obtain the normalized temporal features.

[0047] The input seismic signal is represented as a matrix. The input features are obtained after passing through the embedding layer and the positional encoding layer: , in, This is the embedding layer weight matrix; It is the bias vector; For learnable location encoding; It is the result of linearly mapping the original features of each time step to the model; In order to be in Temporal location information is added as input to the Transformer layer to learn long-range temporal dependencies and inter-station relationships. Subsequently, temporal features are extracted through an attention sub-layer and a feedforward sub-layer, calculated using the following formula:

[0048] ,

[0049] in, For layer normalization function; For multi-head self-attention mapping functions; For feedforward network weights, For the feedforward network bias; To output regression parameters; Features after global pooling; The temporal feature matrix of the attention sublayer is obtained by summing the residuals and performing layer normalization; This is the final temporal feature matrix after passing through the attention sub-layer and the feedforward sub-layer. The final output matrix of the feedforward sublayer The Line, i.e., the first Feature vectors at each time step; This represents the number of sampling points; Indexed by time; final output , For the next level (the first) The temporal feature matrix of each time step in the time series (layer), with shape as follows: (d is the model dimension). When hour, (or This refers to the sequence input after embedding and adding position encoding. .

[0050] For linear projection, the original input channels are projected onto the model dimension to facilitate subsequent calculations. To add location encoding, inject location information.

[0051] For multi-head self-attention computation, calculate the attention with other time steps at each time step and output the result. The context representation.

[0052] The attention sublayer computation process is as follows: First, the input and output of the attention sublayer are added together as a residual connection. Then, the output of the attention sublayer is normalized to obtain the normalized temporal features.

[0053] For the feedforward sublayer computation process: Perform two fully connected layers independently at each time step. The function is connected to the input residual of the feedforward sublayer, normalized, and then the final output matrix of the feedforward sublayer is obtained. This indicates a fully connected first layer. This indicates a fully connected second layer. express function.

[0054] This indicates that global average pooling is applied to the time dimension to obtain the global feature vector, which is obtained by passing through the average pooling layer.

[0055] This indicates a linear mapping to the task output dimension.

[0056] In one embodiment, the number of Transformer layers is 8.

[0057] The seismic signal undergoes preprocessing including instrument response removal, mean / trend removal, bandpass filtering, and resampling, and is then aligned according to a fixed time window to obtain the input matrix. ( (Represented as a D-row, G-column real matrix), then the embedding is obtained through linear projection. And add positional coding form The sequence is used as input to the self-attention mechanism branch; in multiple Transformer layers, the sequence is first mapped through a multi-head self-attention mechanism. The contextual information from different time steps and channels is aggregated according to attention weights, and the output of the attention sub-layer is generated by residual summation and layer normalization. That is, the temporal feature matrix of the attention sub-layer after residual summation and layer normalization. By using a time-progressive feedforward sublayer and performing residual and normalization operations, the final output matrix of the feedforward sublayer is obtained. Global pooling is performed at all time steps to obtain sequence-level global pooled features. The output is mapped to localization or regression through a fully connected layer. , This is the final time series feature matrix after passing through the attention sublayer and the feedforward sublayer. To enhance physical consistency and robustness, the three-dimensional velocity model can be calculated during training using a differentiable interpolation / lookup table operator to calculate the theoretical arrival time. The residuals of the theoretical arrival time and the observed arrival time can be used as additional constraint terms for joint optimization, so that the model can still maintain global consistency in satisfying the earthquake propagation law even under low signal-to-noise ratio or missing measurement conditions.

[0058] In one embodiment, the convolutional neural network branch (CNN branch) includes:

[0059] The initial convolutional layer is used to project seismic signal data onto a multi-channel feature space;

[0060] Multiple cascaded coded residual blocks (ERBs) are used to extract local textures of time and frequency and across stations through local convolution, normalization and residual skip connections;

[0061] Residual connection convolutional layers are used to connect the input and output residuals of encoded residual blocks.

[0062] The computation process of a branch in a convolutional neural network is represented as follows:

[0063] ,

[0064] in, For batch normalization function; This is a two-dimensional convolution operation; It is a 1×1 convolution; This is the residual splicing function; This is the parameter set for the residual block. Input feature map; It is an intermediate feature obtained by first batch normalizing the input and then performing ReLU activation; Yes Features after performing a single 2D convolution and batch normalization; Yes The features are then convolved again and batch-normalized. It is the mapping function for the branches of the convolutional neural network; This is the output feature map of the convolutional neural network branch. The convolutional neural network branch starts with an initial convolutional layer, followed by nine encoded residual blocks (ERBs). Within each ERB, features are processed in the following order: normalization → ReLU function → local convolution (3×3) → normalization → local convolution (3×3) → normalization → pooling downsampling. Finally, the input and processing result are added together through residual skip connections (1×1 convolution is used to align the number of channels if necessary) to preserve low-level information. After multiple ERBs, feature fusion and downsampling are performed, and the final 3D coordinate position is output by regression through several fully connected layers.

[0065] In the middle, stride convolution or pooling downsampling is used to expand the receptive field to capture longer-range propagation features. Different filters in the channel dimension learn complementary representations such as energy, phase, envelope or spectral components. All coded residual block outputs are aggregated in the channel dimension and then globally converged before entering several fully connected layers for regression, outputting the three-dimensional coordinate position of the event.

[0066] By preserving gradient information between input and output through cross-layer connections, the vanishing gradient phenomenon during deep network training is prevented, thereby improving network stability.

[0067] The fused features are obtained by weighted normalizing and fusing the global dependency features of the branch and the local spatiotemporal features of the branch. This includes mapping the signal-to-noise ratio, mask, geometric information and velocity voxels near the prediction location of each station to adaptive fusion weights through a lightweight network or gating unit, and then weighted normalizing and fusing the global dependency features of the branch and the local spatiotemporal features of the branch to obtain the fused features.

[0068] During the training process, the deep fusion network model uses a three-dimensional velocity model to perform elastic wave forward modeling to obtain clean seismic signal data, and superimposes real-acquired background noise to construct training samples. By adjusting the signal-to-noise ratio and the noise weight factor of each detector, training samples with multiple noise levels are obtained.

[0069] In one embodiment, a three-dimensional velocity model is used to perform forward modeling of elastic waves to obtain clean seismic signal data, and real-acquired background noise is superimposed on this data to construct training samples.

[0070] Let the original seismic signal matrix and noise matrix be respectively in Indicates the first The pure signal from each detector Indicates the first Noise signal of each detector This represents the total number of detectors.

[0071] The formulas for calculating signal power and noise power are: in, Indicates the total number of sampling points. This is the noise scaling factor. For the first The first channel in the The instantaneous amplitude of the signal at each sampling time. For the first The first channel in the Noise samples at each time point.

[0072] Signal-to-noise ratio is defined as: The noise weighting factor of each detector is calculated using the above formula, and the resulting composite signal is: All detector outputs form a noisy sample matrix: Each element in the matrix represents the synthesized signal plus noise value at a specific time point for a specific channel, where the matrix dimension is... , Indicates the time sampling point. Indicates the first Each detector has a column representing the complete time series of one detector. The signal-to-noise ratio is adjusted accordingly. This enables the generation of training samples with multiple noise levels, providing a foundation for the robust learning of subsequent neural networks.

[0073] The loss function for training a deep fusion network model is:

[0074] ,

[0075] in This represents the number of training samples; No. The true coordinates of each sample; For predicting coordinates; For speed constraint coefficient, For indicator functions, This represents the wave velocity value at the coordinate position of the velocity model.

[0076] Training employed the AdamW optimizer with a batch size of 32 and an initial learning rate of 0.01, combined with adaptive learning rate decay and early stopping strategies. When the validation set loss no longer decreased after several consecutive rounds, the learning rate was automatically reduced or training was terminated. This optimization scheme effectively improved the model's convergence speed and generalization performance.

[0077] The deep fusion network model is a parallel structure. Starting with a raw seismic record, it first reads the waveform and metadata. For each record, it performs instrument response removal, mean / trend removal, bandpass filtering, resampling, and truncation or zero-padding according to a fixed time window. It extracts noise segments, estimates the signal-to-noise ratio and energy, and injects noise with scaling factors as needed for data augmentation. Finally, it concatenates each time series column-wise into an input matrix. One copy of the input matrix enters a branch of the convolutional neural network. It is first projected onto a multi-channel feature space by the initial convolutional layer, and then passes through a series of coded residual blocks (ERBs). Through local convolution, normalization, and residual skip connections, it extracts time-frequency and cross-station local textures and downsamples at several layers to expand the receptive field, thus obtaining the local spatiotemporal features of the branch. Another branch enters the self-attention mechanism, first linearly mapping and adding positional encoding to form a sequence embedding. After passing through multiple layers of serial multi-head self-attention mechanism and feedforward sub-layer, long-range dependencies, inter-station phase consistency, and waveform similarity are aggregated in the time dimension. Finally, pooling is used to obtain the branch's global dependency features. Simultaneously, the signal-to-noise ratio, mask, geometric information, and even velocity voxels near the predicted location of each station are mapped to adaptive fusion weights (weights can be based on dimension or channel) through a lightweight network or gated unit. The branch features are then weighted, normalized, and fused to obtain the fused features. The fused features are regressed through several fully connected layers to obtain the three-dimensional location of the event. During training, the regression loss function and the arrival consistency constraint based on the three-dimensional velocity model are jointly optimized. The velocity constraint obtains the theoretical arrival time by trilinear interpolation of the predicted location on the pre-calculated travel time surface and compares it with the observed arrival time (or cross-correlation peak time), which serves as an additional differentiable loss term, thereby ensuring that the final output minimizes the data-driven error and satisfies the physical consistency of earthquake propagation.

[0078] On the other hand, embodiments of this application provide a deep fusion network microseismic localization method based on velocity model constraints, the method comprising:

[0079] The self-attention mechanism is used to extract the branch global dependency features of seismic signal data, including: linear mapping and adding position encoding to form sequence embedding to obtain input features, aggregating long-range dependencies and inter-station phase similarity in the time dimension through multiple multi-head self-attention and feedforward sub-layers to obtain time-series features, and then obtaining the branch global dependency features through pooling.

[0080] The branch local spatiotemporal features of seismic signal data are extracted using a convolutional neural network. The process includes: first, the initial convolution is projected onto a multi-channel feature space, then the local texture of time and frequency and cross-station is extracted through multiple coded residual blocks via local convolution, normalization and residual skip connection, and the receptive field is expanded by several layers of downsampling to obtain the branch local feature vector.

[0081] The fused feature is obtained by weighting and normalizing the global dependency features of branches and the local spatiotemporal features of branches;

[0082] By fusing features through several fully connected layers, the three-dimensional location of microseismic events can be regressed.

[0083] In one embodiment, time-series features are obtained by aggregating long-range dependencies and inter-station phase similarities in the time dimension through multiple layers of multi-head self-attention and feedforward sub-layers, including:

[0084] The multi-head self-attention mechanism employs an attention sub-layer, adding the attention sub-layer input and the attention sub-layer output as a residual connection, and then performing normalization processing.

[0085] After processing each time step independently with two fully connected layers and a ReLU function in the feedforward sublayer, the residuals are added to the output of the attention sublayer, and finally normalized to obtain the normalized temporal features.

[0086] In one embodiment, the outputs of all coded residual blocks are aggregated along the channel dimension and globally converged.

[0087] In one embodiment, the fused features are obtained by weighted normalization fusion of the branch global dependency features and the branch local spatiotemporal features. This includes mapping the signal-to-noise ratio, mask, geometric information, and velocity voxels near the predicted location of each station to adaptive fusion weights through a lightweight network or gating unit, and then performing weighted normalization fusion of the branch global dependency features and the branch local spatiotemporal features to obtain the fused features.

[0088] To verify the superiority of the method in this application, comparative experiments were conducted with traditional convolutional neural networks (CNN), parallel networks without speed constraints (PTCNets1), and several heuristic / evolutionary optimization methods. Simulated datasets (600 test events in total, with a training / validation / test ratio of 7:2:1) and field measured data (11th fracturing site, 57 events in total) were used as evaluation benchmarks. Evaluation indicators included eastward error, northward error, depth error, and mean absolute error (MAE).

[0089] The mean absolute error (MAE) is used as the primary performance metric, and it is defined as follows: ,in, The number of samples; For the first The true coordinates of each sample; For the first The predicted coordinates of each sample.

[0090] To ensure that the training process is guided by physical priors, the loss function adopts a composite form based on MSE and incorporates velocity model constraints (i.e., a velocity value weighting term is introduced when the distance between the predicted point and the true point is less than 20m) to improve physical consistency and convergence stability.

[0091] Comparative experiments based on 600 test events show that the overall mean absolute error (MAE) of the traditional CNN is 39.241m (25.416m east, 16.616m north, and 20.618m depth); the overall MAE of PTCNets1 (without velocity constraints) is 8.174m (4.267m east, 3.891m north, and 4.089m depth); while the overall MAE of the embodiment using velocity model constraints is further reduced to 3.502m (1.744m east, 1.731m north, and 1.771m depth). Therefore, this embodiment significantly outperforms the comparative algorithms in terms of positioning accuracy and error balance, with a more concentrated error distribution and no obvious directional deviation.

[0092] To examine the stability of the model under different signal-to-noise ratio (SNR) conditions, tests were conducted with SNR values ​​of {20dB, 10dB, 5dB, −5dB, −10dB, −15dB, −20dB}. Experimental results show that as the SNR decreases, the errors of all models tend to increase, but PTCNets1 exhibits the smallest increase in error. Even under the extreme condition of SNR = −20dB, the average localization error remains less than 4m. In contrast, traditional CNNs show large error fluctuations and numerous outliers under negative SNR conditions, and PTCNets1 still exhibits significant instability at low SNR. Furthermore, the embodiments of this application demonstrate that velocity constraints play a crucial role in convergence to physically reasonable solutions under low SNR conditions, thereby significantly enhancing the model's noise robustness.

[0093] Random masking experiments were conducted at different missing rate (e.g., 2%, 5%, 10%, 15%) to evaluate the degradation characteristics of each model under incomplete observations. The experimental results show that when the missing rate is low (2%), the overall MAE of PTCNets1 is only about 3.8008m; when the missing rate increases to 15%, the overall mean absolute error of PTCNets1 only increases to about 7.3105m; while the traditional CNN already exhibits an overall MAE of about 40.16m at 2% missing rate, rising to about 48.90m at 15% missing rate; the degradation of PTCNets1 with increasing missing rate is also much greater than that of PTCNets1 with 3D velocity model constraints. Therefore, in engineering environments with incomplete observations or certain channel failures, this embodiment significantly improves the robustness and reliability of localization through physical constraints.

[0094] The method described in this application was experimentally verified in the 11th stage of hydraulic fracturing in the Southwest Shale Gas Project, such as... Figure 2 The image shows the location of the earthquake source. Figure 2 (a) in the figure is a three-dimensional perspective view. Figure 2 (b) in the image is a planar view projected onto the east-west depth of a 3D point. Figure 2 (c) in the figure is a planar view projected onto the north-south depth of a 3D point. Figure 2 (d) in the figure represents a three-dimensional point projection onto an east-west-north-south plane view, used to visually demonstrate the spatial distribution, depth range, and temporal evolution of the event relative to the well trajectory, source, and perforation points. Figure 3The image shows the spatial distribution of well trajectory and microseismic events, with points A and B representing the beginning and end of the construction section, respectively. After extracting events and processing them using a sliding window from the one-minute raw data (60,000×38) collected on-site, 57 microseismic events were identified and located. On-site results indicate that the spatial distribution of the predicted events is highly consistent with the geological structure and fracture development areas. The predicted events generally extend along the perforation points. The estimated stimulated reservoir volume (SRV) also spatially overlaps with the high-permeability zone obtained through seepage / fracture inversion, further demonstrating the feasibility and practical value of this embodiment in engineering deployment.

[0095] Based on the above comparisons and field verifications, it is evident that the parallel structure of the deep fusion network model provides complementary representations of global temporal information and local spatiotemporal information; the three-dimensional velocity model constraint, as a physical prior in the loss function, helps guide the network to converge to a physically acceptable solution, especially in scenarios with low signal-to-noise ratio and incomplete data, significantly improving positioning accuracy and suppressing outlier errors. In summary, the embodiments of this application demonstrate the accuracy advantages, robustness, and engineering adaptability of the proposed method and system under complex geological and harsh observation conditions.

[0096] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A velocity model constraint based deep fusion network microseismic positioning system, characterized in that, Comprise: After training, a deep fusion network model for processing seismic signal data and outputting the three-dimensional position of microseismic events; The deep fusion network model comprises: A self-attention mechanism branch for extracting branch global dependency features of seismic signal data, comprising: linear mapping and adding position coding to form sequence embedding to obtain input features, aggregating long-range dependencies and inter-station phase similarities in the time dimension through multiple layers of Transformer layers to obtain time sequence features, and obtaining branch global dependency features through pooling; A convolutional neural network branch for extracting branch local spatio-temporal features of seismic data, comprising: first projecting to a multi-channel feature space by an initial convolution, then extracting local textures in time and across stations through local convolution, normalization and residual skip connection by multiple encoding residual blocks, and expanding the receptive field by several layers of down-sampling to obtain a branch local feature vector; A learnable weight fusion layer for weighting and normalizing the fusion of branch global dependency features and branch local spatio-temporal features to obtain fusion features; A plurality of fully connected layers for regressing the three-dimensional position of microseismic events from the fusion features; The loss function of the deep fusion network model in the training process is: ， wherein is the number of training samples; the true coordinates of the the true coordinates of the is the predicted coordinates; is the velocity constraint coefficient, is the indicator function, is the wave velocity value of the velocity model at the coordinate position. 2.The deep fusion network microseismic positioning system based on velocity model constraint according to claim 1, wherein, Each layer of the Transformer layer comprises: An attention sublayer, which adds the input of the attention sublayer and the output of the attention sublayer as a residual connection, and performs normalization processing; A feedforward sublayer, which independently processes two layers of full connection and ReLU function at each time step, then adds the output of the attention sublayer to the residual, and finally performs normalization processing to obtain normalized time sequence features. 3.The deep fusion network microseismic positioning system based on velocity model constraint of claim 1, wherein, The convolutional neural network branch comprises: An initial convolutional layer for projecting seismic signal data to a multi-channel feature space; A plurality of encoding residual blocks connected in series for extracting local textures in time and across stations through local convolution, normalization and residual skip connection; A residual connection convolutional layer for connecting the input and output of the encoding residual block through residual skip connection.

4. The deep fusion network microseismic positioning system based on velocity model constraints according to claim 1, wherein The fusion features obtained by weighting and normalizing the fusion of branch global dependency features and branch local spatio-temporal features comprise: mapping the signal-to-noise ratio, mask, geometric information of each station, and the speed voxel near the predicted position through a light network or a gating unit to obtain adaptive fusion weights, and weighting and normalizing the fusion of branch global dependency features and branch local spatio-temporal features to obtain fusion features. 5.The deep fusion network microseismic positioning system based on velocity model constraint of claim 1, wherein, In the training process of the deep fusion network model, a three-dimensional velocity model is used to perform elastic wave forward calculation to obtain pure seismic signal data, and real background noise is superimposed to construct training samples. By adjusting the signal-to-noise ratio and the noise weight factor of each geophone, multiple noise level training samples are obtained.

6. A velocity model constraint based deep fusion network microseismic positioning method, implemented by using the system of any one of claims 1-5, characterized in that, The method comprises: Extracting branch global dependency features of seismic signal data using a self-attention mechanism branch, comprising: linear mapping and adding position coding to form sequence embedding to obtain input features, aggregating long-range dependencies and inter-station phase similarities in the time dimension through multiple layers of Transformer layers to obtain time sequence features, and obtaining branch global dependency features through pooling; The branch local space-time features of the seismic signal data are extracted by using a convolutional neural network branch, including: first, projecting to a multi-channel feature space by an initial convolution, then extracting local textures of time-frequency and cross-station by local convolution, normalization and residual skip connection through a plurality of encoding residual blocks, and expanding the receptive field by using a plurality of down-sampling layers to obtain a branch local feature vector; The branch global dependent features and the branch local space-time features are fused by weighted normalization to obtain fused features; The fused features are regressed to three-dimensional positions of microseismic events by a plurality of fully connected layers.

7. The deep fusion network microseismic positioning method based on velocity model constraint according to claim 6, characterized in that, Each layer of the Transformer layer includes: An attention sub-layer, which adds the attention sub-layer input and the attention sub-layer output as a residual connection, and performs normalization processing; A feedforward sub-layer, which independently processes two fully connected layers and a ReLU function at each time step, then adds the output of the attention sub-layer to the residual, and finally performs normalization processing to obtain normalized time sequence features. 8.The method of claim 6, wherein, The outputs of all the encoding residual blocks are aggregated in the channel dimension and globally pooled. 9.The method of claim 7, wherein, The branch global dependent features and the branch local space-time features are fused by weighted normalization to obtain fused features, including: mapping the signal-to-noise ratio, mask, geometric information and speed voxel near the predicted position of each station to adaptive fusion weights by a light network or a gating unit, and fusing the branch global dependent features and the branch local space-time features by weighted normalization to obtain fused features.

Images