A microseismic source positioning method based on convolutional neural network

Abstract

The present application belongs to the cross field of geoscience and artificial intelligence, and particularly relates to a microseismic source positioning method based on a convolutional neural network, 1, training data set preparation; 2, sample label preparation; 3, network input and label setting; 4, deep learning network model G structure setting; 5, loss function setting; 6, network model training; 7: source positioning accuracy test. The powerful function fitting relationship of the convolutional neural network is used to convert the inversion problem of solving the source position into the mapping problem of constructing the probability distribution of the source position from the microseismic data. The U-shaped structure network is used as the backbone feature extraction network, the attention mechanism and the spatial hollow multi-scale pooling module are fused, the feature extraction and fusion process is strengthened, the noise factor interference is reduced, the high-level semantic information of the network is effectively utilized, the robust performance is good, the velocity model error, noise interference and other factors are not sensitive, and the accurate source position can be provided with high calculation efficiency.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of earth science and technology and artificial intelligence, and specifically relates to a method for locating microseismic sources based on convolutional neural networks. Background Technology

[0002] With the development of the global oil and gas industry from conventional to unconventional oil and gas, the exploration and development of unconventional oil and gas has attracted increasing attention. Currently, the main technology for developing unconventional oil and gas is hydraulic fracturing. Rock fracturing generates microseismic events, which propagate as seismic waves. Microseismic monitoring technology is used to monitor the seismic waves generated by microseismic events, and then the effective waveforms are analyzed to determine relevant parameters such as the hypocenter location. Each location within the fracture can be considered a hypocenter, and accurate hypocenter location is the foundation of all quantitative analysis. Analyzing and processing the information collected by the geophones yields the spatial morphology of the hydraulic fracturing fractures, which helps fracturing engineers adjust injection optimization strategies in a timely manner.

[0003] Traditional location methods can be categorized into travel-time-based and waveform-based algorithms based on their principles. However, due to complex geological conditions and acquisition techniques, traditional algorithms face numerous challenges: Firstly, there's the challenge of accuracy. Microseismic signals are weak and susceptible to noise interference during propagation, resulting in low signal-to-noise ratios in the obtained microseismic records and making initial arrival acquisition difficult, rendering travel-time-based algorithms unsuitable. Furthermore, velocity model errors and the layout of the observation system both affect the microseismic data processing results, limiting location accuracy. Secondly, there's the challenge of efficiency. Waveform-based algorithms, while achieving a certain level of accuracy, require significant computational power, and the process of determining the source coordinates is time-consuming, failing to meet the real-time analysis requirements of microseismic monitoring. Therefore, further optimizing the accuracy of microseismic source location by incorporating new technologies and methods is a crucial prerequisite for improving shale gas fracturing productivity using microseismic information features, possessing significant engineering and research value.

[0004] Deep learning-based seismic source location algorithms transform the solution of complex inversion problems in traditional methods into establishing a mapping relationship between seismic gathers and source locations. This avoids complex computational problems while significantly improving computational efficiency, making it a promising research direction. Based on their implementation mechanisms, they can be divided into two types: one involves identifying microseismic data using deep learning before location; the other utilizes neural networks to construct a mapping directly from seismic waveforms to source location coordinates. The advantages of deep learning-based location algorithms lie in their high efficiency and the fact that they automatically select and extract features from observed microseismic data, solving the problems of manual and model dependence inherent in traditional algorithms. However, the feature extraction process does not fully consider the characteristics of microseismic signals. Currently, when using gather data as network input, the algorithm often ignores the influence of noise during the mapping to source locations. Limited support features lead to high ambiguity in the model, affecting the location accuracy, and the robustness of the model needs further improvement.

[0005] As mentioned above, combining the velocity inversion method driven by the physical model with the data-driven deep learning method constitutes a velocity inversion method driven by both data and models. Deep learning methods address the problems in velocity inversion, while the combination of data and model-driven approaches leverages seismic exploration methods to provide more feature engineering information, thereby reducing the learning difficulty of deep neural network models and improving generalization ability and result accuracy. Summary of the Invention

[0006] The purpose of this invention is to provide a microseismic source location method based on deep learning data-driven principles. A network model is established, mapping microseismic waveform information to the probability distribution of source location. This model integrates an attention mechanism module and a spatial void multi-scale module to extract effective features from microseismic data, reducing noise interference and improving the accuracy of support features. This allows the neural network to fully learn the mapping relationship between microseismic event waveform features and source location. The location error is set as the evaluation metric, representing the distance between the location of the maximum value in the network output features and the location of the maximum value in the sample labels. A loss function is used to evaluate the training effect of the network model.

[0007] The technical solution adopted in this invention is: a microseismic source localization method based on convolutional neural networks, which includes the following steps:

[0008] Step 1: Preparing the Training Dataset

[0009] The network input consists of observation data acquired by the detector. For multichannel microseismic records, each signal is a continuous time function. The two-dimensional microseismic data is sampled along the time direction at a certain ratio, with the number of sampling points equal to the number of grid points in the z-direction of the velocity model, to maintain the same size for both input and output data. To enhance the robustness and applicability of the model, the impact of noise interference and velocity model errors on positioning accuracy is considered when constructing the dataset. Gaussian smoothing operators are used to smooth the original velocity model in both the x and z directions. The smoothed velocity model is used to generate the training dataset for the initial velocity model, and the original velocity model is used to generate the test dataset. Gaussian noise of level r is added to the microseismic data to simulate the susceptibility of microseismic data to noise interference in practical applications. Level r is the ratio of Gaussian noise to seismic data, resulting in several microseismic data y samples with varying degrees of noise interference.

[0010] Step 2: Sample Label Preparation

[0011] The probability distribution image labeled with the location of the earthquake source is shown, and the peak position of the probability is the location of the earthquake source corresponding to the microseismic data. In order to reduce the size of the network output and make it easier for the network to learn the relationship between the input data and the corresponding output labels, the size of the monitoring area is represented by the number of grid points in the monitoring area, and each pixel represents the probability value of the earthquake source at the corresponding grid position. A Gaussian distribution is used to describe the probability distribution of the earthquake source, and the maximum point of the earthquake source probability distribution is the location of the earthquake source. The probability distribution image of the earthquake source location is set as l.

[0012] Step 3: Network Input and Tag Setting

[0013] Using the microseismic waveform data y obtained in step one as input, and the corresponding source location probability distribution image l from step two as label, the localization network model G is trained.

[0014] Step 4: Setting the structure of the deep learning network model G

[0015] The deep learning network model G uses a U-shaped network architecture, combined with an attention mechanism and a spatially hollow pyramid pooling module, to strengthen the mapping relationship from microseismic data to seismic source distribution images. The network input is microseismic data, which is reduced in feature dimension through convolution and pooling operations in the encoding network. The final output features of the encoding network are passed to the spatially hollow multi-scale pooling module, which then passes the information-mined and reused features to the decoding network. The attention module highlights effective features during the upsampling process through feature weighting, and the deconvolution operation enables the low-resolution seismic source mapping map to be restored layer by layer to a high-resolution seismic source location probability distribution image with the same spatial size as the input data. The network learns the mapping relationship between microseismic waveforms and seismic source locations, and uses the trained network to predict the seismic source locations of microseismic events that did not participate in the training process.

[0016] The network model is set as follows

[0017] (1) Backbone network for extracting effective microseismic signals

[0018] The core feature extraction network is used to map microseismic data to the probability distribution of seismic source locations. The network model is set as an n-layer encoder-decoder structure, using ReLU as the activation function for both the encoder and decoder networks. Batch normalization is applied to each layer to normalize the output features, increasing training stability. In the last layer, the Sigmoid activation function is used to map the output range to 0-1, representing the probability distribution of seismic source locations. Since the localization model's task is to map the temporal information of microseismic data to the spatial information of the seismic source location probability distribution, there is a weak correspondence between the data and the elements. The details extracted by the shallow convolutional kernels, including noise, are less relevant to the localization task. Therefore, no feature concatenation operation is performed on the shallow encoder-decoder information of U-Net when constructing the localization model.

[0019] (2) Microseismic data effective feature extraction and enhancement module

[0020] Considering the low signal-to-noise ratio of microseismic data, to enhance the extraction of effective signal features, a spatial attention mechanism is used to assign different weights to different features in the seismic trace, thereby highlighting effective waveform information and reducing the interference of background noise on the positioning task. The task focusing process is reflected in the allocation of weight coefficients. The feature map of the encoding part is activated by the ReLU function and normalized to obtain the attention weight. The feature in the decoder is multiplied by the weight to obtain the final output activation feature. That is, the weight represents the importance of information, so as to achieve the screening and focusing of key features.

[0021] (3) Microseismic data effective feature fusion module

[0022] The U-Net network uses a spatially dilated multi-scale pooling module to replace the original bridging module at the bottom layer. After feature maps from the encoder are extracted by the spatially dilated multi-scale pooling module, they are provided to the decoder part of the network. This can preserve important features in high-level convolutions to a greater extent and provide multi-scale and multi-level information for subsequent upsampling operations. The dilation rates of the three parallel convolutional layers are k1, k2, and k3, respectively. The global average pooling layer is used to obtain global feature information, and a 1×1 convolutional layer is used to fuse the obtained features and feed them into the decoding network.

[0023] Step 5: Setting the loss function

[0024] The MSE loss function is used to measure the difference between the predicted location and the actual source location. The expression for the loss function is:

[0025]

[0026] N is the total number of training samples, l i This represents the source location information of the i-th microseismic data predicted by the network. This represents the corresponding true earthquake source location label; the smaller the loss function, the closer the network output is to the characteristics of the sample label, and the higher its positioning accuracy.

[0027] Step Six: Network Model Training

[0028] Input the microseismic waveform features obtained from steps one to five above into y, and input the probability distribution image of the labeled seismic source location into the network for training. Use the Adam optimizer to update the network and learn to minimize the loss function. When the model training converges, save the trained network parameters.

[0029] Step 7: Seismic source location accuracy test

[0030] Microseismic data were generated using forward modeling with the original velocity model. After sampling and noise reduction, the test dataset y′ was obtained and input into a trained and tuned convolutional neural network model G to obtain the probability distribution image of the seismic source location. The maximum probability value is the location of the earthquake source predicted by the network.

[0031] The beneficial effects of this invention are as follows: It provides a method and process for microseismic source localization based on deep learning data-driven principles. A network model is established, which maps microseismic waveform information to the probability distribution of source location. This model integrates an attention mechanism module and a multi-scale module for spatial voids to extract effective features from microseismic data, reducing noise interference and improving the accuracy of support features. This allows the neural network to fully learn the mapping relationship between microseismic event waveform features and source location. After network training, using untrained microseismic waveforms as input, the corresponding source location can be predicted in milliseconds. This process requires no manual intervention, achieving intelligent and automated microseismic data localization. Attached Figure Description

[0032] Figure 1 This is the network model diagram of Implementation Example 1;

[0033] Figure 2 This is a schematic diagram of feature transformation during the network training process in Example 1;

[0034] Figure 3 This is a comparison chart of the microseismic data before and after downsampling in Example 1;

[0035] Figure 4 Figure 1 shows one of the test data samples selected from Example 1;

[0036] Figure 5 This refers to the positioning result of sample 1 in Example 1;

[0037] Figure 6 Figure 2 is a sample of test data selected from Example 1;

[0038] Figure 7 This refers to the location result of sample image 2 in Example 1;

[0039] Figure 8 This describes the process of creating the sample dataset in Example 2;

[0040] Figure 9 This is a comparison chart of the results after adding different levels of noise to the same sample data in Example 2;

[0041] Figure 10 This is a sample image of any one of the training data from Example 2;

[0042] Figure 11 This is Figure 1, which is one of the test data samples selected from Example 2;

[0043] Figure 12 This is Figure 2, which is one of the test data samples selected from Example 2;

[0044] Figure 13 This refers to the positioning result of sample 1 in Example 2;

[0045] Figure 14 Figure 2 shows the positioning result of sample 1 in Example 2. Detailed Implementation

[0046] Example 1

[0047] Reference Figures 1-7 A microseismic source localization method based on convolutional neural networks, the source localization method comprising the following steps:

[0048] Step 1: Preparing the training dataset:

[0049] Design a simple underlying medium velocity model where the velocity increases linearly with depth, with 256 grid points in the horizontal direction and 200 grid points in the vertical direction. The grid size is 5m×5m, and the velocity range is 2km / s~3km / s. Detectors are placed on the ground surface, with one detector at each grid point in the horizontal direction. The sampling interval is 0.25ms, and the monitoring time is 1800ms.

[0050] Based on this simple layered dataset, experiments were conducted using single-shot and multi-shot data. In the multi-shot data, 3 to 4 seismic sources were used to simulate elongated fracture morphology. The first seismic source was randomly excited within the monitoring area, and the remaining sources were excited sequentially along the designed fracture morphology. Gaussian noise with a maximum amplitude of 20%, 30%, 40%, 50%, and 60% was added to the microseismic data to simulate the susceptibility of microseismic data to noise interference in practical applications, resulting in several noisy microseismic data sets y.

[0051] Step 2: Sample Label Preparation

[0052] Each event is labeled with a two-dimensional Gaussian distribution with a peak value of 1, and the peak position represents the source location corresponding to the microseismic event. For multi-source data, the labels generated by each source are superimposed and normalized to represent the sample labels for multi-source data. The size of the monitoring area is represented by the number of grid points, and each pixel represents the probability value of the source at the corresponding grid position. In this example, the size of the sample label is 200×256, resulting in several probability distribution images l of source locations corresponding to the microseismic data.

[0053] Step 3: Network Input and Tag Setting:

[0054] Using the microseismic waveform data y obtained in step one as input, and the corresponding source location probability distribution image l from step two as label, the localization network model G is trained.

[0055] Step 4: Setting the structure of the deep learning network model G:

[0056] The deep learning network model G is based on a U-shaped network architecture and further strengthens the mapping relationship from microseismic data to seismic source distribution images by combining an attention mechanism and a spatially hollow pyramid pooling module. The network input is microseismic data. In the encoding network, the microseismic data is reduced in feature dimensionality through convolution and pooling operations. The final output features of the encoding network are passed to the spatially hollow multi-scale pooling module. The features after information mining and reuse are then passed to the decoding network. The attention module further highlights effective features during the upsampling process by weighting features. The deconvolution operation enables the low-resolution seismic source mapping map to be restored layer by layer to a high-resolution seismic source location probability distribution image with the same spatial size as the input data. Finally, the network learns the mapping relationship between microseismic waveforms and seismic source locations, and can use the trained network to predict the seismic source locations of microseismic events that did not participate in the training process.

[0057] The network model is set as follows:

[0058] (1) Backbone network for extracting effective microseismic signals

[0059] The core feature extraction network maps microseismic data to the probability distribution of seismic source locations. This network model is a 6-layer encoder-decoder structure, using 3×3 convolutional kernels in each convolutional layer. ReLU is used as the activation function for both the encoder and decoder networks, and a batch normalization module is applied to each layer to normalize the output features, increasing training stability. The last layer uses a Sigmoid activation function to map the output range to 0-1, representing the probability distribution of seismic source locations. Since the localization model's task is to map the temporal information of microseismic data to the spatial information of the seismic source location probability distribution, and since there is a weak correspondence between the data elements, the noise and other details extracted by the shallow convolutional kernels are less relevant to the localization task. Therefore, feature concatenation is not performed on the shallow encoder-decoder information of the U-Net when constructing the localization model.

[0060] (2) Microseismic data effective feature extraction and enhancement module

[0061] Considering the low signal-to-noise ratio of microseismic data, this invention enhances the extraction of effective signal features by assigning different weights to different features in the seismic trace using a spatial attention mechanism. This highlights effective waveform information and reduces the interference of background noise on the positioning task. The task focusing process is reflected in the allocation of weight coefficients. The feature map of the encoded part is activated by the ReLU function and normalized to obtain the attention weight. The feature in the decoder is multiplied by the weight to obtain the final output activation feature. That is, the weight represents the importance of the information, so as to achieve the screening and focusing of key features.

[0062] (3) Microseismic data effective feature fusion module

[0063] This invention replaces the original bridging module at the bottom layer of the U-Net network with a spatially dilated multi-scale pooling module. Feature maps from the encoder, after feature extraction by the spatially dilated multi-scale pooling module, are provided to the decoder part of the network. This allows for greater preservation of important features from higher-level convolutions, providing multi-scale and multi-level information for subsequent upsampling operations. The dilation rates of the three parallel convolutional layers are set to 6, 12, and 18, respectively. A global average pooling layer is used to obtain global feature information, and a 1×1 convolutional layer fuses the obtained features and feeds them into the decoding network.

[0064] Step 5: Setting the loss function:

[0065] The MSE loss function is used to measure the difference between the predicted location and the actual source location. The expression for the loss function is:

[0066]

[0067] N is the total number of training samples, l iThis represents the source location information of the i-th microseismic data predicted by the network. This represents the corresponding true earthquake source location label; the smaller the loss function, the closer the network output is to the characteristics of the sample label, and the higher its positioning accuracy.

[0068] Step Six: Network Model Training

[0069] The microseismic waveform features obtained from steps one through five above are input into y, and the probability distribution image of the labeled seismic source location is input into the network for training. The Adam optimizer is used to update the network and minimize the loss function. The network undergoes 400 iterations. The initial learning rate of the training model is 0.01. When the performance on the validation set does not improve after 4 iterations, the learning rate is automatically adjusted to half of its original value, and the trained network parameters are saved.

[0070] Step 7: Seismic source location accuracy test:

[0071] Microseismic data were generated using forward modeling with the original velocity model. After sampling and noise reduction, the test dataset y′ was obtained and input into a trained and tuned convolutional neural network model G to obtain the probability distribution image of the seismic source location. The maximum probability value is the location of the earthquake source predicted by the network.

[0072] When creating the dataset, the impact of noise and velocity model errors on the accuracy of seismic source location was considered. A smoothed velocity model and the original velocity model were used as the initial velocity models for the training and testing data, respectively. After forward numerical simulation, microseismic waveform data was obtained. Then, a downsampling operation was performed at a certain ratio, and Gaussian noise was introduced. The processed noisy microseismic data was used as the input to a convolutional neural network. The network label was a two-dimensional Gaussian distribution image of the seismic source location, where any pixel in the image represented the probability of the seismic source location at that grid point. The localization network was essentially a mapping from microseismic data to the probability distribution of seismic source locations. During the network model construction, attention mechanisms and spatially voided multi-scale pooling were combined to reduce the interference of irrelevant features such as noise, enabling further feature extraction. The network loss was measured using the MSE loss function.

[0073] Example 2

[0074] Reference Figures 3-12The training process of the localization model constructed by this method mainly includes forward data propagation and backward iteration of network loss. The forward propagation process of the model is as follows: First, the preprocessed microseismic waveform signal y is used as the feature input of the network. In the encoding network, features are extracted through convolutional pooling operations to generate a series of feature maps. When the microseismic data y is input into the first layer, the height and width of the feature map are consistent with the input, and the number of channels becomes 32. Each time downsampling is performed, the width and height of the feature map are halved, the number of feature channels is doubled, the parameters of the upper layer are passed to the next layer, and then these features are encoded. The features finally output by the encoder are passed to the spatial void pyramid pooling module. After information mining and reuse, the features are then passed to the decoder part of the network. In this process, an attention mechanism module is introduced to further highlight the effective features through feature weighting. The backward iteration operation of the model is as follows: MSE is used as the loss function to measure the difference between the network output and the sample label. The Adam optimization algorithm is used to minimize the error to update the network until the convergence condition is met and the training ends. The trained network model is then saved.

[0075] The experimental platform configuration for this method is as follows: the GPU used in the experiment is a GeForce RTX-3080TI; the framework used to build the localization model is PyTorch 1.6 with CUDA version 11.0; and some processing operations are performed using Matlab 2017b. The specific implementation process is as follows:

[0076] 1. Preprocessing of microseismic data training dataset

[0077] 1.1 Smoothing Operation of Velocity Model

[0078] The original velocity model is smoothed using the built-in Matlab function `smooth2a.m`. The function is defined as: `function matrixOut = smooth2a(matrixIn, Nr, Nc)`, where `matrixIn` is the original velocity model, and `Nr` and `Nc` are the sizes of the smoothing operators in the horizontal and vertical directions, respectively. In this example, the Marmousi velocity model is used for numerical simulation to generate a test dataset with 288 grid points in the horizontal direction and 300 grid points in the vertical direction. Smoothing operators of sizes 4×4 and 6×6 are used to smooth the original velocity model, respectively. The smoothed model is then used to generate the training dataset.

[0079] 1.2 Microseismic data excitation source

[0080] The source location is excited using a Ricker wavelet, which is defined as follows:

[0081]

[0082] Where lt represents the monitoring duration, fm represents the frequency of the Ricker wavelet, PI is pi, dt is the sampling time interval, and wave_length represents the wavelength. In this example, the monitoring duration is set to 2700ms.

[0083] 1.3 Simulation of Noisy Microseismic Data

[0084] After generating microseismic data through numerical simulation, Gaussian noise with an intensity of 20% to 60% is added to the microseismic data. The percentage of noise added is measured by the ratio of Gaussian noise to seismic data. The specific implementation is as follows: First, the microseismic data file is read and its variable is defined as "shot". The definition of the noisy microseismic data "noise_shot" is: noise_shot = shot + max(shot(:)) * r * randn(nz, nx); where nx and nz are the size of the microseismic data, and r is the noise ratio, which is taken as 0.2, 0.3, 0.4, 0.5, and 0.6 in this example.

[0085] 1.4 Downsampling Operation of Microseismic Data

[0086] First, read the noisy microseismic data file noise_shot. The monitoring time (time) and trace number (trace) of this microseismic data are [time,trace] = size(noise_shot). Then, perform uniform sampling in the time direction: n = 1:9:time; array = noise_shot(n,:); array is the sampled microseismic data.

[0087] 2. Sample label preparation

[0088] Given that the location of the earthquake source is (x0, z0) and the radius of the Gaussian distribution is 70, the process of creating the sample label is as follows:

[0089]

[0090] 3. Design the network structure of the deep learning model G: The deep learning network model G uses a U-shaped network as the basic architecture for feature extraction, and combines attention mechanism and spatial void pyramid pooling module to further strengthen the mapping relationship from microseismic data to seismic source distribution characteristics.

[0091] 3.1 Design of the backbone network for extracting effective microseismic signals

[0092] The U-Net network consists of an encoding network and a decoding network.

[0093] The main code in the encoding network is as follows:

[0094]

[0095] The main code in the decoding network is as follows:

[0096]

[0097] 3.2 Design of an Enhancement Module for Effective Feature Extraction of Microseismic Data

[0098] If the attention mechanism module is defined as Attention_block, then the module is defined as follows:

[0099]

[0100]

[0101] The key code for the forward function of the model is as follows:

[0102]

[0103] Where psi is the weighting coefficient, after processing by the attention module, the weighted encoded features are obtained. 3.3 Design of the effective feature fusion module for microseismic data

[0104] The attention mechanism module is defined as ASPP, and the 1×1 convolution and three parallel convolutions with dilation rates of 6, 12, and 18 are defined as follows:

[0105] self.atrous_block1=nn.Conv2d(in_channel,depth,1,1)

[0106] self.atrous_block6=nn.Conv2d(in_channel,depth,3,1,padding=6,dilation=6)

[0107] self.atrous_block12=nn.Conv2d(in_channel,depth,3,1,padding=12,dilation=12)

[0108] self.atrous_block18=nn.Conv2d(in_channel,depth,3,1,padding=18,dilation=18)

[0109] The resulting features are then fused using a 1×1 convolutional layer: self.conv_1x1_output = nn.Conv2d(depth*5, depth, 1, 1).

[0110] The key code for passing the forward function in this function block is as follows:

[0111] atrous_block1=self.atrous_block1(x)

[0112] atrous_block6=self.atrous_block6(x)

[0113] atrous_block12=self.atrous_block12(x)

[0114] atrous_block18=self.atrous_block18(x)

[0115] cat=torch.cat([image_features,atrous_block1,atrous_block6,

[0116] atrous_block12,atrous_block18],dim=1)

[0117] net=self.conv_1x1_output(cat)

[0118] return net

[0119] 3.4 Definition of overall network structure:

[0120] On the U-shaped network infrastructure, taking the 4th layer encoder as an example, the specific implementation process of the attention mechanism is as follows: x4 = self.Att5(g = d5, x = x4); d5 = torch.cat((x4, d5), dim = 1). The encoder and decoder are connected using the spatially dilated pyramid pooling module: self.SPP = ASPP(1024); d6 = self.SPP(x6). Where x6 is the feature in the last layer encoder; d6 is the corresponding decoder feature. Finally, the sigmoid function is used to map the features to the interval [0, 1]: d1 = self.sigmoid(d1).

[0121] 4. Design the loss function:

[0122] Set the network output and sample labels as variables `hires_batch`, and the loss function as `content_criterion`. Define the loss function as the common MSE loss function: `content_criterion = nn.MSELoss()`, and then use this loss function to measure the difference between the network output and the sample labels: `loss = content_criterion(hires_batch, hits_batch)`.

[0123] 5. Train and save the network model:

[0124] The Adam loss function is used for optimization during training, and the calling statement is:

[0125] optim_G=optim.Adam(generator.parameters(),lr=args.lr)

[0126] scheduler_g=torch.optim.lr_scheduler.ReduceLROnPlateau(

[0127] optimizer=optim_G,patience=args.scheduler_patience,factor=0.5,verbose=True)

[0128] The learning rate is set to 0.001, and the number of iterations is 400 epochs.

[0129] 6. Test the performance of the network model:

[0130] The network model saved in step 5 is obtained by calling the `torch.load` function, defined as `model = torch.load(f"{args.model_folder} / {args.model_name}", map_location = device)`. The test dataset, after sampling and noise addition, is input into the trained and tuned convolutional neural network model G to obtain the source location probability distribution `output`. The predicted probability is `pr = np.max(output)`, and the predicted source locations `p_x` and `p_z` are output as follows:

[0131]

[0132]

[0133] Implementation results:

[0134] The specific process of creating the sample dataset is as follows: Figure 8 As shown, Figure 9 The results show the addition of different levels of noise to any microseismic data sample. It can be seen that as the noise level gradually increases, the effective signal is gradually submerged in the noise, making it difficult to distinguish the effective waveform features.

[0135] Figure 10 For any training data sample image, the velocity model used for this sample is a velocity model smoothed using a 6×6 size operator, with a noise level of 20%; Figure 11 and Figure 12 The test sample data obtained using the original Marmousi model forward modeling has noise levels of 30% and 60%, respectively. It can be seen that compared with the training samples, the obtained microseismic signal waveforms are more complex due to the complex fault layers of the velocity model.

[0136] The network model's prediction results for the location of earthquake sources in the test samples are given, and its probability distribution diagram is shown below. Figure 13 and Figure 14 As shown, the predicted probabilities are 0.9742 and 0.9621, respectively, and the predicted source locations are at grid points (170, 160) and (190, 81), with errors in the horizontal x and z directions not exceeding one grid point. This fully demonstrates the beneficial results of this method, which can predict the source distribution with a relatively high probability and obtain a relatively accurate source location for microseismic data contaminated by noise.

Claims

1. A method for locating microseismic sources based on convolutional neural networks, characterized in that: The method for locating the source of a microseismic event includes the following steps: Step 1: Preparing the Training Dataset In two-dimensional microseismic data, samples are taken along the time direction at a certain ratio, with the number of sampling points equal to the number of grid points in the z-direction of the velocity model. Gaussian smoothing operators are used to smooth the original velocity model in both the x and z directions. The smoothed velocity model is used to generate a training dataset for the initial velocity model, and the original velocity model is used to generate a test dataset. Gaussian noise of level r is added to the microseismic data to simulate the situation where microseismic data is susceptible to noise interference in practical applications. Level r is the ratio of Gaussian noise to seismic data, resulting in several microseismic data y with different degrees of noise interference. Step 2: Sample Label Preparation The probability distribution image labeled with the epicenter location is the epicenter location corresponding to the microseismic data. The size of the monitoring area is represented by the number of grid points in the monitoring area, and each pixel represents the probability value of the epicenter at the corresponding grid location. The probability distribution of the epicenter is described using a Gaussian distribution, and the maximum point of the epicenter probability distribution is the epicenter location. The probability distribution image of the epicenter location is set as l. Step 3: Network Input and Tag Setting Using the microseismic waveform data y obtained in step one as input, and the corresponding source location probability distribution image l from step two as label, the localization network model G is trained. Step 4: Setting the structure of the deep learning network model G The deep learning network model G uses a U-shaped network architecture, combined with an attention mechanism and a spatially hollow pyramid pooling module, to strengthen the mapping relationship from microseismic data to seismic source distribution images. The network input is microseismic data, which is reduced in feature dimension through convolution and pooling operations in the encoding network. The final output features of the encoding network are passed to the spatially hollow multi-scale pooling module, which then passes the information-mined and reused features to the decoding network. The attention module highlights effective features during the upsampling process through feature weighting, and the deconvolution operation enables the low-resolution seismic source mapping map to be restored layer by layer to a high-resolution seismic source location probability distribution image with the same spatial size as the input data. The network learns the mapping relationship between microseismic waveforms and seismic source locations, and uses the trained network to predict the seismic source locations of microseismic events that did not participate in the training process. The network model is set as follows (1) Backbone network for extracting effective microseismic signals The core feature extraction network is used to map microseismic data to the probability distribution of seismic source locations. The network model is set as an n-layer encoder-decoder structure, using ReLU as the activation function for both the encoder and decoder networks. A batch normalization module is applied in each layer to normalize the output features of each layer to increase training stability. In the last layer of the network, the Sigmoid activation function is used to map the output range to between 0 and 1 to represent the probability distribution of seismic source locations. (2) Microseismic data effective feature extraction and enhancement module To enhance the extraction of effective signal features, different weights are assigned to different features in the seismic traces using a spatial attention mechanism. The task focusing process is reflected in the allocation of weight coefficients. The feature map of the encoding part is activated by the ReLU function and normalized to obtain the attention weight. The feature in the decoder is multiplied by the weight to obtain the final output activation feature. That is, the weight represents the importance of the information, so as to achieve the screening and focusing of key features. (3) Microseismic data effective feature fusion module The original bridging module at the bottom layer of the U-Net network is replaced by a spatially dilated multi-scale pooling module. The feature map from the encoder is extracted by the spatially dilated multi-scale pooling module and then provided to the decoder part of the network. The dilation rates of the three parallel convolutional layers are k1, k2 and k3, respectively. The global average pooling layer is used to obtain global feature information. The obtained features are fused by a 1×1 convolutional layer and fed into the decoding network. Step 5: Setting the loss function The MSE loss function is used to measure the difference between the predicted location and the actual source location. The expression for the loss function is: N is the total number of training samples, l i This represents the source location information of the i-th microseismic data predicted by the network. This represents the corresponding true earthquake source location label; the smaller the loss function, the closer the network output is to the characteristics of the sample label, and the higher its positioning accuracy. Step Six: Network Model Training Input the microseismic waveform features obtained from steps one to five above into y, and input the probability distribution image of the labeled seismic source location into the network for training. Use the Adam optimizer to update the network and learn to minimize the loss function. When the model training converges, save the trained network parameters. Step 7: Seismic source location accuracy test Microseismic data were generated using forward modeling with the original velocity model. After sampling and noise reduction, the test dataset y′ was obtained and input into a trained and tuned convolutional neural network model G to obtain the probability distribution image of the seismic source location. The maximum probability value is the location of the earthquake source predicted by the network.

Images