Microseismic source positioning method, device and system, and storage medium

Abstract

The invention discloses a microseismic source positioning method, device and system, and a storage medium. The method comprises the following steps: dividing a microseismic data sample data set into a training set and a test set; according to the training set, constructing a microseism inversion subnet containing a Swin Transform encoder, and according to the training set, constructing a microseism inversion subnet containing a Swin Transform encoder; the micro-seismic forward modeling subnet realizes seismic wave field continuation by inputting a micro-seismic source position and a speed model and utilizing a recurrent neural network structure and a convolution operator, and establishes a forward modeling subnet based on a wave equation; constructing an inversion-forward modeling closed-loop neural network according to the micro-seismic inversion subnet and the forward modeling subnet based on the wave equation; and inputting the test set into the inversion-forward closed-loop neural network to carry out micro-seismic source positioning. By adopting the technical scheme of the invention, the limitations on physical constraint, feature modeling and anti-noise capability in the prior art are overcome.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of microseismic monitoring, and particularly relates to a microseismic source positioning method and device, system and storage medium. BACKGROUND

[0002] Microseismic source positioning technology has key application value in the fields of oil and gas field hydraulic fracturing monitoring, geological disaster warning, etc., and its core goal is to accurately invert the source location by analyzing seismic waveform data. In the existing positioning methods based on deep learning, although the convolutional neural network (CNN) can learn the nonlinear mapping relationship between microseismic waveform and source location through a large amount of data training, it is essentially a data-driven "black box" model, which has significant technical bottlenecks: on the one hand, this kind of model lacks effective modeling of the physical properties of underground media, and cannot incorporate the physical laws of seismic wave propagation (such as wave field continuation described by wave equation, reflection and refraction mechanism), which leads to the possibility that the positioning result may violate the actual wave field propagation logic and has poor interpretability; on the other hand, the receptive field of CNN is limited by the size of the convolution kernel, and it is difficult to capture the long-distance dependence relationship in microseismic records across spatial scales (such as the phase delay correlation of the same source signal received by different geophones), which has insufficient modeling ability for global features under complex geological structures or noise interference, resulting in a decrease in positioning accuracy.

[0003] In recent years, the Swin Transformer architecture has shown strong global information modeling capability in natural language processing and computer vision fields due to its hierarchical window attention mechanism, which can effectively capture long-distance dependencies in sequence data. However, there are still limitations when it is directly applied to microseismic data processing: the ability of Swin Transformer to capture local spatial details is weaker than CNN, while the local features such as high-frequency waveform mutations and phase reversals near the source in microseismic data are critical information for accurate positioning; in addition, the data-driven Swin Transformer model lacks physical constraints like traditional CNN, and cannot guarantee that the inversion result meets the physical properties of underground media (such as velocity distribution, attenuation characteristics), which may lead to the positioning result deviating from the actual source location.

[0004] The traditional U-Net network as a classic image segmentation architecture has an encoder-decoder structure that extracts multi-scale features through convolution operations, and has certain advantages in local feature extraction of microseismic data. However, due to the dependence of the encoder on CNN, the same difficulties as traditional CNNs are encountered when processing microseismic data: limited by the locality of convolution operations, it is difficult to effectively model the propagation characteristics of seismic waves in the entire space, especially in a noisy background, and it is impossible to suppress irrelevant signal interference through global features, resulting in insufficient robustness of the positioning results. For example, in a simple layered model, although the traditional U-Net can recover some features through down-sampling and up-sampling structures, it is insufficient to model the long-distance correlation of cross-layer waveforms, resulting in a significant increase in positioning error with increasing model complexity.

[0005] From the physical nature, the propagation process of microseismic waves is the dynamic response of the underground medium system to the excitation of the source, and the observation data is the result of the joint action of the medium structure and the source characteristics. Therefore, both the fitting based on the data space (such as traditional deep learning methods) and the physical inversion based on the model space (such as full waveform inversion) have limitations: the former lacks physical constraints, and the latter is limited by the accuracy of the initial model and the computational efficiency. Existing research attempts to build a "forward-inversion" closed-loop structure to combine physical models with neural networks, for example, by generating simulated data through full waveform inversion as the supervision signal for neural network training, but in such methods, the traditional forward model (such as finite difference method) and the neural network are independent of each other, and end-to-end joint optimization is not achieved, resulting in the physical constraints that cannot be dynamically adjusted during network training, and the fixed wave equation model is difficult to adapt to complex media (such as anisotropy and non-uniform velocity distribution), which limits the physical reasonableness and generalization ability of the positioning results.

[0006] Under complex geological conditions (such as strong non-uniformity in the Marmousi velocity model), the defects of traditional methods are further highlighted: the CNN-based model has an increased waveform matching error due to the lack of global features, and the physical inversion-based model is difficult to solve in real time due to the explosion of computational load. In addition, microseismic data in actual engineering is often disturbed by Gaussian noise, engineering vibration, etc. In a low SNR environment, it is difficult for traditional methods to distinguish the feature differences between effective signals and noise, and it is easy to fall into a local optimal solution, resulting in divergent positioning results. For example, in microseismic data with 60% Gaussian noise added, the positioning error of the traditional U-Net increases significantly, and the pure Swin Transformer model lacks sufficient suppression ability for high-frequency noise due to the lack of local feature constraints SUMMARY

[0007] The technical problem to be solved by the present application is to provide a microseismic source positioning method and device, system, and storage medium, aiming to overcome the limitations of existing technology in physical constraints, feature modeling, and noise resistance.

[0008] To achieve the above object, the application adopts the following technical solutions:

[0009] A microseismic source positioning method, comprising:

[0010] Step S1, dividing the microseismic data sample data set into a training set and a test set;

[0011] Step S2, constructing a microseismic inversion subnetwork containing a Swin Transformer encoder according to the training set; wherein the microseismic inversion subnetwork comprises a S-Unet network composed of a Swin Transformer encoder and a U-Net decoder;

[0012] Step S3, the microseismic forward modeling subnetwork realizes seismic wave field continuation by inputting the microseismic source location and the velocity model, using a recurrent neural network structure and a convolution operator, and establishes a forward modeling subnetwork based on wave equation;

[0013] Step S4, constructing an inversion-forward closed-loop neural network according to the microseismic inversion subnetwork and the forward modeling subnetwork based on wave equation; wherein the source spatial position parameters output by the inversion subnetwork are taken as the source term input of the forward modeling subnetwork, and the wave motion law is taken as the constraint, so that the data-driven inversion result can be verified by the forward modeling process; the forward modeling subnetwork simulates the wave field based on the input source parameters, and the difference between its output and the measured data can drive the inversion subnetwork to optimize the source location parameters in the reverse direction;

[0014] Step S5, inputting the test set into the inversion-forward closed-loop neural network for microseismic source positioning.

[0015] As a preferred, the inversion subnetwork comprises a S-Unet network composed of a Swin Transformer encoder and a U-Net decoder; the Swin Transformer encoder adopts a 4-layer hierarchical structure, each layer contains a plurality of Swin TransformerBlock, which is used to extract the global features of the microseismic data; the U-Net decoder contains 4 up-sampling stages, each stage restores the feature map resolution through transposed convolution; the encoder and the decoder are connected by a jump connection, and an attention gate mechanism is added to weight and fuse the feature maps.

[0016] As a preferred, the loss function of the inversion-forward closed-loop neural network is:

[0017]

[0018] Wherein, N is the total number of samples, y i ∈{0,1} is the true source location label, is the probability value predicted by the network; λ is the physical constraint weight coefficient, d i is the synthetic seismic data, To observe data.

[0019] The application also provides a microseismic source positioning device, comprising:

[0020] The first processing module is configured to divide the microseismic data sample dataset into a training set and a test set.

[0021] The second processing module is configured to construct a microseismic inversion subnetwork comprising a Swin Transformer encoder according to the training set; wherein the microseismic inversion subnetwork comprises a S-Unet network composed of the Swin Transformer encoder and a U-Net decoder.

[0022] The third processing module is configured to make the microseismic forward subnetwork input the microseismic source position and the velocity model, and realize seismic wave field continuation by using a recurrent neural network structure and a convolution operator to establish a wave equation-based forward subnetwork.

[0023] The fourth processing module is configured to construct an inversion-forward closed-loop neural network according to the microseismic inversion subnetwork and the wave equation-based forward subnetwork; wherein the spatial position parameter of the source output by the inversion subnetwork is taken as the source term input of the forward subnetwork, and the wave motion law is taken as a constraint to enable the data-driven inversion result to be verified through the forward process; the forward subnetwork simulates the wave field based on the input source parameter, and the difference between the output and the measured data can reversely drive the inversion subnetwork to optimize the source position parameter.

[0024] The fifth processing module is configured to input the test set into the inversion-forward closed-loop neural network for microseismic source positioning.

[0025] Preferably, the inversion subnetwork comprises a S-Unet network composed of a Swin Transformer encoder and a U-Net decoder; the Swin Transformer encoder adopts a 4-layer hierarchical structure, each layer comprising a plurality of Swin Transformer Blocks for extracting global features of the microseismic data; the U-Net decoder comprises 4 up-sampling stages, each stage recovering the resolution of the feature map through transposed convolution; a skip connection is arranged between the encoder and the decoder, and an attention gate mechanism is added to weight and fuse the feature maps.

[0026] Preferably, the loss function of the inversion-forward closed-loop neural network is as follows:

[0027]

[0028] wherein N is the total number of samples, y i is a true source position label, is a probability value predicted by the network; λ is a physical constraint weight coefficient, and di for synthetic seismic data, for observed data.

[0029] The application also provides a microseismic source positioning system, comprising a memory and a processor, wherein the memory stores a computer program which is run by the processor, and the computer program performs the microseismic source positioning method when run by the processor.

[0030] The application also provides a storage medium, wherein the storage medium stores a computer program which performs the microseismic source positioning method when run.

[0031] The application has the following innovations:

[0032] 1. An S-UNet inversion subnetwork is designed, wherein a Swin Transformer is used to replace the encoder of a traditional U-Net, and a hierarchical attention mechanism is used to capture the global long-distance dependence relationship of microseismic data, while the convolution structure of the U-Net decoder is retained to extract local detailed features, and the complementary of global and local information is realized through multi-scale feature fusion;

[0033] 2. A closed-loop neural network is constructed, wherein a forward subnetwork based on full waveform inversion is connected in series with the S-UNet inversion subnetwork, and the inversion result meets both the data fitting accuracy (such as binary cross-entropy loss) and the wave field propagation physical law (such as content loss constrained by wave equation) through joint optimization, so that the "black box" problem of a pure data-driven model is solved;

[0034] 3. The feature interaction of the encoder and the decoder is optimized through an attention gate mechanism, the interference of irrelevant areas is suppressed, and a waveform fitting term is introduced into the loss function, so that the robustness of the model to noise and velocity model errors is enhanced.

[0035] The application effectively solves the problems of lack of physical constraints and insufficient global feature capture of traditional deep learning models by combining the global feature modeling capability of Swin Transformer with the local detail extraction capability of U-Net to construct a closed-loop neural network containing a full waveform inversion forward subnet. Specifically, Swin Transformer replaces the traditional CNN as an encoder, significantly improving the modeling capability of long-distance waveform dependency in microseismic data through hierarchical attention mechanisms and multi-scale feature output, while the U-Net decoder realizes the fine recovery of local features through skip connections and attention gate mechanisms. The combination of the two improves the positioning accuracy by more than 80% compared to traditional U-Net, with a single source positioning error of only 1 grid point (10m) in a simple layered model and an average error of less than 1 grid distance in a complex Marmousi model. By introducing full waveform inversion constraints, the physical laws of wave equations are embedded into the training process, making the model output conform to the logic of seismic wave propagation and avoiding the "black box" defect of pure data-driven models. The positioning results can be explained by the wave field continuation principle, and the interpretability is significantly enhanced. In terms of noise resistance and robustness, the model uses content loss (L_content) to constrain waveform fitting accuracy and combines NMS method to filter non-maximum points. When adding 60% Gaussian noise, the vertical error is still controlled within 10m, and the horizontal error of velocity model error (such as 4x4 and 6x6 smoothing operator interference) is ≤3m and the vertical error is ≤5m. In actual engineering tests, experiments based on L-26 well data in Songliao Basin show that the positioning error of the model for real noise data is within 1 grid, and for low signal-to-noise ratio data (effective signal ratio <20%), the model can still accurately identify the source location, and the training efficiency is improved by more than 50% compared with traditional methods, and it can converge within 100 iterations, significantly reducing the computational cost, providing a high-precision, interpretable, and robust microseismic source positioning solution for oil and gas development, geological disaster monitoring, and other scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, brief descriptions will be given below of the drawings needed in the embodiments or prior art descriptions. Obviously, the drawings in the following description are only embodiments of the present application, and those skilled in the art can obtain other drawings according to the provided drawings without creative labor.

[0037] Figure 1 Flow chart of the microseismic source positioning method of the embodiment of the present application;

[0038] Figure 2 Schematic diagram of the forward network structure based on RNN;

[0039] Figure 3Fig. 1 is a network structure diagram of the S-UNet inversion network;

[0040] Figure 4 Fig. 2 is a network structure diagram of the closed-loop neural network;

[0041] Figure 5 Fig. 3 is a comparison diagram of loss functions of different network models;

[0042] Figure 6 Fig. 4 is a positioning diagram of the S-UNet inversion network and a positioning diagram of the SWIN_UR-Net network; wherein, (a) is the positioning result of the S-UNet, and (b) is the positioning result of the SWIN_UR-Net;

[0043] Figure 7 Fig. 5 is seismic data under different noise levels;

[0044] Figure 8 Fig. 6 is the positioning result under different noise levels; wherein, (a) is the result of adding 20% Gaussian noise, (b) is the result of adding 40% Gaussian noise, and (c) is the result of adding 60% Gaussian noise. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0046] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0047] Embodiment 1

[0048] As shown in Fig. 1, the microseismic source positioning method provided by the embodiments of the present application comprises: Figures 1 to 4

[0049] Step 1, data set preparation:

[0050] I. Velocity model sample organization

[0051] ​The present application constructs two types of velocity models to cover different geological scenarios. First, a simple layered velocity model is generated by superimposing sine, cosine and linear functions to generate a 4-7 layered medium. The simple layered velocity range is set to 1500-4500 m / s, the model space size is 2240x2240, and the horizontal and vertical grid size is 10 m. This type of model is mainly used to simulate the wave field propagation characteristics in horizontal layered strata, and constitutes the basic data set. Second, a complex Marmousi velocity model is used. The Marmousi model in the field of exploration geophysics is used and down-sampled to retain its complex geological structure characteristics such as faults and salt domes, as well as a wide range of velocity changes from 1500 to 5500 m / s. It is used to verify the positioning robustness of the algorithm in strong heterogeneous media. These two types of models together constitute a multi-scenario velocity model library from simple to complex, providing diverse physical constraints for subsequent microseismic data synthesis.

[0052] II. Microseismic data sample generation and processing

[0053] The generation and processing of microseismic data samples includes three key steps. The first step is data generation and label construction. The wave field propagation process is simulated using the finite difference method. The point source excitation main frequency is 20 Hz Ricker wavelet. The source position is randomly distributed in the model to cover the entire regional geological structure. The label construction uses Gaussian distribution to describe the probability distribution of the source position, forming a two-dimensional probability image with the true source coordinates as the peak value. The second step is the preprocessing procedure. To address the problem of high dimensionality and loss of position information, the microseismic data time series is resampled to match the depth direction grid points of the velocity model, and a dimension reduction strategy is used to reduce the data dimension while preserving the spatio-temporal characteristics of the microseismic signal. The third step is the data augmentation strategy. To enhance the robustness of the model, 20-60% Gaussian noise is added to the microseismic data to simulate a low signal-to-noise ratio environment, and random noise is added to the Marmousi model to generate training data with parameter uncertainty.

[0054] III. Actual engineering data integration

[0055] To verify the applicability of the algorithm in real scenarios, the present application uses the actual data of L-26 well in Songliao Basin to construct a test set. First, the actual data is collected to obtain 27 channels of microseismic signals with a duration of 15s. These signals contain engineering noise and strata heterogeneity interference. Then, the initial velocity model is constructed. Based on the P-wave and S-wave velocity, density curves are processed layer by layer, and the perforation signal with known position is used to correct the model error, improving the matching degree of the model and the actual stratum velocity. Finally, the simulation data is generated. The forward simulation is performed in a monitoring area of 810m x 2000m. From the 81 simulated signals, 27 effective signals are selected, and the actual collected noise is injected to form a test sample containing real interference characteristics.

[0056] Step two: Microseismic source location network setting based on closed-loop structure:

[0057] The full-waveform inversion driven neural network is used as a forward subnetwork, which is connected in series with the data-driven S-Unet inversion network to form a closed-loop neural network. The closed-loop neural network takes the predicted source location output by the inversion subnetwork as the trainable parameter input of the forward network, and the microseismic data obtained by the forward network is input into the inversion network, so that the two networks form a closed-loop structure.

[0058] The network model is set as:

[0059] I. Inversion subnetwork design: S-UNet network architecture

[0060] (I) Encoder: Global feature modeling of Swin Transformer

[0061] The Swin Transformer replaces the traditional CNN encoder of U-Net, and captures the long-distance waveform dependence in microseismic data through a hierarchical window attention mechanism. Specifically, it adopts a 4-layer encoding-decoding structure, and each layer is composed of multiple Swin Transformer Blocks. During the calculation process, each layer reduces the resolution of the feature map (such as from 224x224 to 14x14) and increases the number of channels (such as from 64 to 512) through Patch Merging operation, thereby completing multi-scale feature extraction. The multi-scale feature maps (containing global features of different resolutions) output by each layer of the encoder are transmitted to the decoder, and are fused with the local features of the corresponding level to enhance the richness and hierarchy of feature expression.

[0062] (II) Decoder: Local detail recovery of U-Net

[0063] The decoder adopts the traditional U-Net architecture, which gradually recovers the spatial resolution through upsampling operations (such as deconvolution or interpolation), and fuses the multi-scale features of the encoder with the corresponding level features of the decoder through skip connections. In the skip connection, an attention gate is introduced to suppress the interference of irrelevant areas (such as noise background) by applying a weight matrix to the encoder feature map, highlighting the feature response of the source-related area, thereby improving the recovery accuracy of local details (such as waveform mutations near the source).

[0064] (III) Output layer and activation function

[0065] The last layer of the network adopts a sigmoid activation function to map the output range to 0-1, and the output result is a two-dimensional probability distribution image, which represents the possibility of each position in the input microseismic data being the source. The probability distribution is compared with the true source label through the binary cross-entropy loss function (BCELoss), and the network parameters are optimized to minimize the difference between the predicted position and the true position, so as to realize the probabilistic positioning of the source position.

[0066] Step three: closed-loop neural network, embedding of physical constraints

[0067] (Four) Forward subnet: wave field simulation driven by full waveform inversion

[0068] An RNN-based forward subnet is used to simulate the propagation of seismic waves in the underground medium. The forward subnet takes the spatial coordinates (x, z) of the source as trainable parameters, and replaces the traditional adjoint state method through the automatic differentiation function of the neural network to efficiently solve the gradient of the source parameters. The forward subnet can generate simulated microseismic data (including wave field continuation, reflection, refraction, etc.) according to the input source position, and its output result is used as the supervision signal for the training of the inversion subnet, forcing the inversion result to comply with the physical laws of seismic wave propagation.

[0069] (Five) Joint optimization of closed-loop structure

[0070] The predicted source position output by the inversion subnet is used as the input parameter of the forward subnet to generate corresponding simulated microseismic data. By calculating the waveform difference (content loss L_content) between the simulated data and the true observed data, together with the classification loss (BCELoss) of the inversion subnet, the total loss function of the closed-loop neural network is formed: Loss = BCELoss + λL content . Where λ is the weight coefficient (usually 0.1-10), used to balance the weights of data fitting (BCELoss) and physical constraints (L_content), ensuring that the two types of losses are in the same order of magnitude. Through end-to-end joint optimization, the network learns the mapping relationship between microseismic data and source position while complying with the physical laws described by the wave equation, improving the explainability and physical reasonableness of the positioning results.

[0071] Step four: loss function design

[0072] The inversion subnet loss is used to measure the difference between the predicted source position probability distribution and the true label, and the formula is: Where N is the total number of samples, y i ∈{0,1} is the true source position label (1 represents the source position and 0 represents the non-source position), is the predicted probability value of the network. This loss function maximizes the predicted probability of the true source position, driving the network to focus on the true source area.

[0073] The content loss measures the waveform difference between the forward simulation data and the real data by mean square error (MSE), and the formula is: Where d i is the simulated microseismic data generated by the forward subnetwork, is the actual observation data. This loss function ensures that the predicted source location corresponds to the simulated waveform and the real waveform, thereby constraining the physical consistency of the inversion result.

[0074] Step five: network model training

[0075] The closed-loop neural network training is based on the collaborative optimization mechanism of data-driven and physical constraints. First, the input data is preprocessed: the theoretical synthetic samples are cropped to 224x224 to retain the complete waveform features, and the actual engineering samples are cropped to 128x128 to adapt to the low-resolution scene. The AdamW optimizer is used, with an initial learning rate of 10 -5 (decreases with iterations), and 8 samples are input per batch to balance the calculation efficiency. The data set is divided into 80% training set (including layered, Marmousi model and 0%-60% noise data) and 20% validation set (including velocity model error and real engineering noise data). The training process triggers a validation every 10 iterations, optimizes the inversion subnetwork (S-UNet) parameters through the loss function Loss = BCELoss + 0.5xLcontent (BCELoss measures the positioning probability error, and Lcontent constrains the difference between the simulated data and the real data), forms a "prediction-forward-correction" closed loop, and completes 300 iterations and 30 validations to realize the deep integration of global feature capture and physical law constraints.

[0076] Step six: microseismic data positioning and robustness test

[0077] The model test is carried out from the multiple dimensions of accuracy, noise resistance, generalization and efficiency. In the noise-free layered model, the vertical error of the closed-loop network (Swin UR-Net) is reduced to 0 from 8m compared with the traditional S-UNet, and the average error of the complex Marmousi model is less than 10m; the anti-noise test shows that the error is less than 3m under 20% Gaussian noise, the error is less than 10m under 60% noise, and the error is less than 1 grid point (10m) in the real engineering noise; the generalization test shows that the horizontal error is less than 3m and the vertical error is less than 5m in the face of the velocity model error introduced by 4x4 / 6x6 smoothing operator, and it is compatible with 224x224 and 128x128 multi-scale input; in terms of calculation efficiency, the single inference time is less than 1 second (GPU acceleration), which is improved by 3 orders of magnitude compared with the traditional full waveform inversion. Experiments prove that the method has high positioning accuracy, strong robustness and real-time calculation ability in the scene of complex geology, noise interference and model mismatch, and provides a reliable solution for actual microseismic monitoring.

[0078] As an embodiment of the embodiment of the application, the closed-loop neural network model training includes data forward propagation and error back propagation, the parameters are randomly initialized and updated by the Adam optimizer, and the initial learning rate is 10 -5 . During forward propagation, the theoretical synthesis (224x224) and actual engineering (128x128) microseismic data are input into the network after being divided into blocks, the inversion subnet (S-UNet) extracts global features through the Swin Transformer encoder, the U-Net decoder fuses local information to output the source location probability distribution, the forward subnet (full waveform inversion RNN) inputs the predicted source location as a parameter to generate simulation data, and the total loss is calculated by combining the binary cross-entropy loss (BCELoss) and the content loss (Lcontent, weight 0.5); the network parameters are updated through the automatic differentiation mechanism during back propagation, the loss and positioning error are evaluated every 10 iterations using a 20% validation set, until the total loss converges, and finally a closed-loop optimization mechanism of "data-driven + physical constraint" is formed, which improves the accuracy and interpretability of microseismic source positioning.

[0079] The experimental platform of the embodiment of the application is configured: Inter(R) Core(TM) CPU i7-10700K, 16GB RAM, the selected GPU is GeForce RTX-3080TI, the framework used for building the positioning model is Pytorch1.6, and the CUDA version is 11.0. The initial learning rate is set to 1x10 -5 , the network is updated using the AdamW optimizer, and the network is iterated for 300 times.

[0080] 1. Seismic data training set preprocessing

[0081] The data loading and conversion process is as follows: read the.mat format data file through the scipy.io.loadmat function, and obtain the input data x_data (seismic waveform) and label data y_data from it. The input data needs to be converted into a PyTorch tensor, specifically adding a channel dimension based on the original shape (500, 224, 224) to convert it into a standard image tensor format (500, 1, 224, 224); the label data remains unchanged in the original shape (500, 224, 224).

[0082] x_data = loadmat(x_path)['f']

[0083] seismic_data_tensor = torch.Tensor(x_data).unsqueeze(1) # (500, 224, 224) -> (500, 1, 224, 224) standard image tensor format (B, C, H, W)

[0084] self.data = seismic_data_tensor

[0085] y_data = loadmat(y_path)['f']

[0086] self.labels = y_data # (500, 224, 224)

[0087] 2. Design network model structure

[0088] The neural network driven by full waveform inversion is used as a forward subnetwork, and the data-driven S-Unet inversion network is connected in series to form a closed-loop neural network. In this closed-loop structure, the predicted source location output by the inversion subnetwork is input as a trainable parameter to the forward network, and the microseismic data generated by the forward network is input to the inversion network, so that the two networks form a closed loop.

[0089] I. Inversion subnetwork design: S-UNet network architecture

[0090] (1) Encoder: Global feature modeling of Swin Transformer

[0091] Swin Transformer replaces the traditional CNN encoder of U-Net, and captures the long-range waveform dependence in microseismic data through a hierarchical window attention mechanism. Specifically, a 4-layer encoding-decoding structure is adopted, and each layer is composed of multiple Swin Transformer Blocks. During the calculation process, each layer reduces the resolution of the feature map by merging adjacent feature map blocks through Patch Merging operation, and increases the number of channels by 512, thereby completing multi-scale feature extraction. The multi-scale feature maps output by each layer of the encoder contain different resolutions of global features, which are passed to the decoder and fused with the local features of the corresponding level to enhance the richness and hierarchy of feature expression.

[0092] (II) Decoder: Local detail recovery of U-Net

[0093] The decoder adopts the traditional U-Net architecture, which gradually recovers the spatial resolution through upsampling operations such as deconvolution or interpolation, and fuses the multi-scale features from the encoder with the corresponding level features from the decoder through skip connections. In the skip connection, an attention gate is introduced to suppress the interference of irrelevant areas (such as noise background) by applying a weight matrix to the encoder feature map, highlighting the feature response of the source-related area, thereby improving the recovery accuracy of local details (such as waveform mutations near the source).

[0094] (III) Output layer and activation function

[0095] The last layer of the network uses a sigmoid activation function to map the output range to 0-1, and the output result is a two-dimensional probability distribution image, representing the likelihood of each position in the input microseismic data being a source. This probability distribution is compared with the true source label through the binary cross-entropy loss function (BCELoss), and the network parameters are optimized to minimize the difference between the predicted position and the true position, thereby achieving probabilistic positioning of the source location.

[0096] Closed-loop neural network, embedding of physical constraints

[0097] (IV) Forward subnetwork: Wavefield simulation driven by full waveform inversion

[0098] A recurrent neural network (RNN) based on the two-dimensional wave equation is used as the forward subnetwork to simulate the propagation of seismic waves in the underground medium. This forward subnetwork can generate simulated microseismic data (including wavefield continuation, reflection, refraction, etc.) according to the input source location, and its output is used as a supervision signal for the training of the inversion subnetwork, forcing the inversion result to comply with the physical laws of seismic wave propagation.

[0099] (V) joint optimization of closed-loop structure

[0100] The predicted source position output by the inversion subnet is taken as the input parameter of the forward subnet to generate the corresponding simulated microseismic data. By calculating the waveform difference (content loss L_content) between the simulated data and the real observed data, together with the classification loss (BCELoss) of the inversion subnet, the overall loss function of the closed-loop neural network is formed: where λ is the weight coefficient (usually 0.1-10), used to balance the weights of data fitting (BCELoss) and physical constraints (L_content), to ensure that the two types of losses are in the same order of magnitude. Through end-to-end joint optimization, the network learns the mapping relationship between microseismic data and source position while following the physical laws described by the wave equation, improving the explainability and physical rationality of the positioning results.

[0101] 3. Design loss function

[0102] The construction of the loss function includes: on the basis of the binary cross-entropy loss function of the inversion subnet, the waveform matching loss expression between the synthetic seismic data output by the forward subnet and the observed data is superimposed:

[0103]

[0104] 4. Train and save network model:

[0105] The sample is input into the closed-loop neural network model using the closed-loop neural network model. During the model training process, 1 validation is performed every 10 iteration training. The overall training is performed 300 times of training and 30 times of validation. Finally, the trained model is saved, torch.save(model.state_dict(),os.path.join(save_model,f"model_epoch_{epoch}.pt")), where the save address is save_model, and the model parameters are saved with model_epoch_{epoch}.pt as the file name.

[0106] The optimization function used in this implementation is the AdamW function for optimizing network parameters. The function call statement optimizer = torch.optim.AdamW(model.parameters(), lr = cfg['training']['lr'], weight_decay = 1e-4) obtains all trainable parameters of the model, the learning rate is read from the configuration file cfg['training']['lr'], and the weight decay is set to 1e-4 for L2 regularization to prevent overfitting. The learning rate is 0.00001. This implementation sets the number of training iterations to 300, and sets epoch = 300.

[0107] 5. Test network model performance

[0108] Model testing is carried out from multiple dimensions such as accuracy, noise resistance, generalization and efficiency. In terms of positioning accuracy, in the noise-free layered model, the vertical error of the closed-loop network is reduced from 8m to 0 compared with S-UNet, and the average error of the complex Marmousi model is less than 10m; the noise resistance test shows that the error is less than 3m under 20% Gaussian noise, less than 10m under 60% noise, and less than 1 grid point (10m) in real engineering noise; the generalization test shows that in the face of velocity model error introduced by 4x4 / 6x6 smoothing operator, the horizontal error is less than 3m and the vertical error is less than 5m, and it is compatible with 224x224 and 128x128 multi-scale input; in terms of computational efficiency, the single inference time is less than 1 second (GPU acceleration), which is 3 orders of magnitude higher than traditional full waveform inversion. Experiments show that this method has high positioning accuracy, strong robustness and real-time computing ability in complex geology, noise interference and model mismatch scenarios, providing a reliable solution for actual microseismic monitoring.

[0109] Implementation effect:

[0110] I. Synergistic advantage of physical constraints and data-driven

[0111] Through the closed-loop structure, the physical constraints of full waveform inversion and the data modeling capability of Swin Transformer are fused, the loss of the network decreases faster in training (as shown in Figure 5 , the loss value of Swin_UR-Net is lower than that of S-UNet), and the positioning result conforms to the law of seismic wave propagation, and the interpretability is significantly improved. Experiments show that this method balances noise suppression, feature extraction and physical consistency, providing a solution for microseismic monitoring with accuracy, efficiency and reliability.

[0112] II. Numerical experiment verifies positioning accuracy and noise resistance

[0113] In the simple layered model, the closed-loop neural network shows a significant advantage over S-UNet: in the noise-free scenario, the closed-loop neural network positioning error is 0, while the S-UNet vertical direction error is 8m; Figure 6 The test data are used to show the source location map obtained by S-UNet inversion network and the source location map obtained by SWIN_UR-Ne network, respectively, wherein the dark circles represent the true source location, and the light circles represent the predicted location. As can be seen from the experimental display result map, in the absence of noise, the positioning predicted by SWIN_UR-Ne coincides with the true source location, while the source location map obtained by S-UNet is horizontally deviated from the source distribution obtained by the closed-loop neural network Figure 1 , and in the vertical direction, the error is 8m. Figure 7 The seismic data with different noise are given, Figure 7 (a) Under 20% Gaussian noise, the predicted source location is basically consistent with the true source location, and the average error in the vertical direction is less than 3m; Figure 7 (b) When the noise level is 40%, the average error in the horizontal direction is less than 3m, and the average error in the vertical direction is less than 5m; Figure 7 (c) When the noise level is 60%, the horizontal and vertical errors are both less than 10m; therefore, it can be concluded that the microseismic source positioning method based on the closed-loop neural network proposed in this paper has good anti-noise performance. Figure 8 The positioning results under different noise levels are shown. The complex Marmousi model experiment shows that the network has an average positioning error of less than 10m under the error-free condition, and when facing the velocity model error introduced by 4x4 and 6x6 smoothing operators, the horizontal error is less than 3m and the vertical error is less than 5m, proving its robustness to complex geological structures and model mismatch.

[0114] The microseismic source positioning method of the full waveform inversion constrained Swin Transformer architecture of the application comprises the following steps: 1. Construction of a data set: The application prepares a data set and constructs two types of velocity models to cover different geological scenarios: a simple layered velocity model is generated by function superposition and is used to simulate horizontal layered stratum wave field propagation; a complex Marmousi velocity model retains complex geological structures and verifies the robustness of the algorithm. Microseismic data samples are generated (finite difference method, Ricker wavelet), preprocessed (time sampling matching, dimension reduction), and augmented (20% to 60% Gaussian noise is added). The actual data of Well L-26 in the Songliao Basin is used to construct a test set, which is acquired, modeled, and simulated data generated to form a test sample containing real interference. 2. Inversion subnetwork construction: Based on the U-Net network, the encoder is replaced with a Swin Transformer, which captures global features and local details of microseismic data through hierarchical design and multi-scale feature output. Each layer in the hierarchical structure is composed of multiple Swin Transformer modules, which reduce the resolution and increase the number of channels through feature map block merging operations, the decoder adopts the traditional U-Net architecture and introduces attention gates, weights the encoder feature maps, and then fuses them with the decoder features to finally output the source location probability distribution. 3. Closed-loop neural network construction: The neural network driven by full waveform inversion is used as a forward subnetwork and connected in series with the above inversion subnetwork. The predicted source location output by the inversion subnetwork is input as a trainable parameter into the forward subnetwork, which simulates wave field propagation based on RNN to generate microseismic data, and then inputs the data into the inversion subnetwork to form a closed-loop structure of "prediction-forward-correction", realizing the organic combination of data-driven and physical constraints. 4. Loss function and training mechanism: The total loss function is composed of a weighted binary cross-entropy loss and a content loss, the former measures the difference between the predicted source location and the true label, and the latter measures the waveform difference between the simulated microseismic data and the real data. The AdamW optimizer is used to update the network parameters, and the training set and validation set are set, and the validation is triggered after a certain number of iterations of training is completed. The overall training process optimizes the parameters through an end-to-end method to form a learning mechanism with physical interpretability. 5. Multi-scenario experimental verification: In the numerical experiments of the simple layered model and the complex Marmousi model, the method shows high positioning accuracy in the noise-free and different levels of Gaussian noise environment, and has strong robustness to velocity model errors. In the actual engineering data test, facing the microseismic data containing real noise, the positioning result is close to the real source location, and the calculation efficiency is significantly better than that of the traditional method, which can meet the real-time monitoring demand.

[0115] Implementation effect: By combining the global feature modeling capability of Swin Transformer and the physical constraints of full waveform inversion, this method effectively improves the explainability and accuracy of microseismic source positioning, and shows good robustness and generalization ability in complex geological conditions, noise interference and model mismatch scenarios. The closed-loop structure realizes the cooperative optimization of data feature learning and physical law correction, and the improvement of computing efficiency makes it suitable for real-time monitoring requirements in practical engineering, providing a reliable technical solution for underground resource development and geological disaster monitoring

[0116] Embodiment 2:

[0117] The embodiment of the application also provides a microseismic source positioning device, comprising:

[0118] The first processing module is used for dividing the microseismic data sample data set into a training set and a test set;

[0119] The second processing module is used for constructing a microseismic inversion subnetwork containing a Swin Transformer encoder according to the training set; wherein the microseismic inversion subnetwork comprises a S-Unet network composed of a Swin Transformer encoder and a U-Net decoder;

[0120] The third processing module is used for the microseismic forward subnetwork to input the microseismic source location and the velocity model, realize seismic wave field continuation by using a recurrent neural network structure and a convolution operator, and establish a forward subnetwork based on wave equation;

[0121] The fourth processing module is used for constructing an inversion-forward closed-loop neural network according to the microseismic inversion subnetwork and the forward subnetwork based on wave equation; wherein the source space position parameter output by the inversion subnetwork is taken as the source term input of the forward subnetwork, and the wave law is taken as the constraint, so that the data-driven inversion result can be verified by the forward process; the forward subnetwork simulates the wave field based on the input source parameter, and the difference between its output and the measured data can drive the inversion subnetwork to optimize the source position parameter in reverse;

[0122] The fifth processing module is used for inputting the test set into the inversion-forward closed-loop neural network for microseismic source positioning.

[0123] As an embodiment of the present application, the inversion subnetwork is composed of a Swin Transformer encoder and a U-Net decoder to form an S-Unet network; the Swin Transformer encoder adopts a 4-layer hierarchical structure, each layer containing multiple Swin Transformer Blocks, for extracting global features of microseismic data; the U-Net decoder contains 4 up-sampling stages, each stage recovering the feature map resolution through transposed convolution; a skip connection is set between the encoder and the decoder, and an attention gate mechanism is added to weight and fuse the feature maps.

[0124] As an embodiment of the present application, the loss function of the inversion-forward closed-loop neural network is:

[0125]

[0126] wherein N is the total number of samples, y i is a true source location label, is a probability value predicted by the network; λ is a physical constraint weight coefficient, d i is synthetic seismic data, is observed data.

[0127] Embodiment 3:

[0128] The embodiment of the present application also provides a microseismic source positioning system, comprising a memory and a processor, the memory has a computer program run by the processor stored thereon, and the computer program executes the microseismic source positioning method when run by the processor.

[0129] Embodiment 4:

[0130] The embodiment of the present application also provides a storage medium, the storage medium has a computer program stored thereon, and the computer program executes the microseismic source positioning method when run.

[0131] The above-described embodiments only describe the preferred modes of the present application, and do not limit the scope of the present application, and various modifications and improvements to the technical solutions of the present application made by those skilled in the art without departing from the design spirit of the present application shall fall within the protection scope determined by the claims of the present application.

Claims

1. A microseismic source location method, characterized in that, The method comprises the following steps: Step S1, dividing a microseismic data sample data set into a training set and a test set; Step S2, constructing a microseismic inversion subnetwork comprising a Swin Transformer encoder according to the training set; wherein the microseismic inversion subnetwork comprises an S-Unet network composed of the Swin Transformer encoder and a U-Net decoder; Step S3, a microseismic forward modeling subnetwork, by inputting a microseismic source location and a velocity model, uses a recurrent neural network structure and a convolution operator to realize seismic wave field continuation, and establishes a forward modeling subnetwork based on wave equation; Step S4, constructing an inversion-forward closed-loop neural network according to the microseismic inversion subnetwork and the forward modeling subnetwork based on wave equation; wherein the source spatial location parameter output by the inversion subnetwork is taken as the source term input of the forward modeling subnetwork, and the wave motion law is taken as the constraint, so that the data-driven inversion result can be verified by the forward modeling process; the forward modeling subnetwork simulates the wave field based on the input source parameter, and the difference between the output and the measured data can reversely drive the inversion subnetwork to optimize the source location parameter; Step S5, inputting the test set into the inversion-forward closed-loop neural network for microseismic source positioning.

2. The microseismic source location method of claim 1, wherein, The inversion subnetwork comprises an S-Unet network composed of a Swin Transformer encoder and a U-Net decoder; the Swin Transformer encoder adopts a 4-layer hierarchical structure, each layer comprising a plurality of Swin Transformer Blocks for extracting global features of the microseismic data; The U-Net decoder comprises 4 upsampling stages, each stage recovering the resolution of the feature map through transposed convolution; a skip connection is arranged between the encoder and the decoder, and an attention gate mechanism is added to weight and fuse the feature maps.

3. The microseismic source location method of claim 2, wherein, The loss function of the inversion-forward closed-loop neural network is: where N is the total number of samples, y i ∈ {0,1} is the true source location label, is the network predicted probability value; λ is the physical constraint weight coefficient, d i is the synthetic seismic data, is the observed data.

4. A microseismic source location device, characterized by, The method comprises the following steps: A first processing module is configured to divide a microseismic data sample data set into a training set and a test set; A second processing module is configured to construct a microseismic inversion subnetwork comprising a Swin Transformer encoder according to the training set; wherein the microseismic inversion subnetwork comprises an S-Unet network composed of the Swin Transformer encoder and a U-Net decoder; A third processing module is configured to use a recurrent neural network structure and a convolution operator to realize seismic wave field continuation by inputting a microseismic source location and a velocity model, and establish a forward modeling subnetwork based on wave equation; A fourth processing module is configured to construct an inversion-forward closed-loop neural network according to the microseismic inversion subnetwork and the forward modeling subnetwork based on wave equation; wherein the source spatial location parameter output by the inversion subnetwork is taken as the source term input of the forward modeling subnetwork, and the wave motion law is taken as the constraint, so that the data-driven inversion result can be verified by the forward modeling process; the forward modeling subnetwork simulates the wave field based on the input source parameter, and the difference between the output and the measured data can reversely drive the inversion subnetwork to optimize the source location parameter; A fifth processing module is configured to input the test set into the inversion-forward closed-loop neural network for microseismic source positioning.

5. The microseismic source location device of claim 4, wherein, The inversion subnetwork is composed of a SwinTransformer encoder and a U-Net decoder to form an S-Unet network; the Swin Transformer encoder adopts a 4-layer hierarchical structure, each layer containing multiple Swin Transformer Blocks, for extracting global features of microseismic data; The U-Net decoder contains 4 up-sampling stages, each stage restoring the resolution of the feature map through transposed convolution; a jump connection is set between the encoder and the decoder, and an attention gate mechanism is added to weight and fuse the feature maps.

6. The microseismic source location device of claim 5, wherein, The loss function of the inversion-forward closed-loop neural network is: where N is the total number of samples, y i ∈ {0,1} is the true source location label, is the network predicted probability value; λ is the physical constraint weight coefficient, d i is the synthetic seismic data, is the observed data.

7. A microseismic source location system characterized in that, including: A memory and a processor, the memory has a computer program run by the processor stored thereon, the computer program executes the microseismic source positioning method as claimed in any one of claims 1-3 when being run by the processor.

8. A storage medium, characterized by The storage medium has a computer program stored thereon, the computer program executes the microseismic source positioning method as claimed in any one of claims 1-3 when being run.

Images