Vibration P wave arrival time pickup method of lightweight transfer learning model

Abstract

A vibration P wave arrival time pickup method of a lightweight transfer learning model belongs to the technical field of micro-seismic monitoring, and comprises the following steps: 1, importing earthquake and micro-seismic event data; 2, preprocessing the data in the step 1; and 3, making a training set and a test set by using the data in the step 2. And 4, constructing a seismic data arrival time pickup model. And 5, inputting data of the seismic data training set in the step 3 to the pickup model in the step 4 for training. And 6, inputting data of the seismic data test set in the step 3 to the pickup model in the step 4 for P-wave arrival time pickup, and evaluating the pickup effect of the model. And 7, inputting the micro-seismic data to the qualified pickup model in the step 6 for transfer learning, and finely adjusting the model to realize the arrival time pickup of the micro-seismic data. According to the method, the P-wave arrival time picking method with high accuracy and reliability is provided.

Description

Technical Field

[0001] The technical field to which the present invention belongs is the field of microseismic monitoring technology, and specifically is a vibration P-wave arrival time picking method of a lightweight transfer learning model. Background Art

[0002] Earthquakes are major natural disasters caused by sudden and large-scale ruptures of crustal rocks, resulting in extremely destructive vibrations. Mine tremors are surface vibrations in mining areas caused by rock instability due to mining. The two are related but not completely the same in terms of their occurrence mechanisms and disaster prevention. They have both similarities and differences.

[0003] The occurrence of mine earthquakes is related to the regional earthquake background. Considering that the layout of mine earthquake stations is difficult to meet the low noise conditions in practical applications, mine earthquake data are usually accompanied by strong noise and complex waveform characteristics. The low accuracy of traditional mine earthquake arrival time picking leads to poor positioning accuracy, which makes the traditional picking method face challenges in accuracy and robustness. Due to the early research on earthquakes, a large number of public data sets have been accumulated, while most of the mine earthquake data are preserved and studied by mining enterprises, making the public data sets far less than seismic data sets. Due to the inherent connection between earthquakes and mine earthquake waveforms, the seismic signal arrival time picking method can be transferred to the mine earthquake arrival time picking. For this reason, a vibration P wave arrival time picking method based on a lightweight transfer learning model is proposed. The purpose is to learn a large amount of P wave arrival time knowledge of earthquake events that have occurred, and then transfer it to the mine earthquake event arrival time picking to achieve better results, so as to provide a technical solution for microseismic monitoring and early warning. At the same time, due to the lightweight model, real-time arrival time picking can be realized, which is expected to enhance the accuracy of mine earthquake positioning and provide reliable data support for scientific research on mine earthquakes. Summary of the invention

[0004] In view of the problem in the prior art that the positioning accuracy is poor due to the low accuracy of picking up the arrival of mine earthquakes, the present invention provides a vibration P-wave arrival picking method based on a lightweight transfer learning model.

[0005] The technical solution adopted by the present invention is as follows:

[0006] A vibration P-wave arrival time picking method based on a lightweight transfer learning model, the steps are as follows:

[0007] 1) Import earthquake and microseismic event data;

[0008] Specifically: the time when the earthquake occurs, which generates shock waves. After multiple sensors collect the shock waves, the program calculates the source coordinates and the manually marked arrival time. The source coordinates are composed of three components: X, Y, and Z. They are the relative offset of the coordinate origin defined by the earthquake, and the unit is m. The manually marked arrival time is the manually marked earthquake P wave arrival pickup point.

[0009] 2) Preprocessing of earthquake and microseismic data;

[0010] Specifically: seismic data is processed. Since the waveform of the original data of the earthquake event is in txt format, it is not suitable for subsequent input model calculation. Therefore, the data is converted to convert the waveform of the event into its corresponding hdf5 array format for subsequent processing.

[0011] 3) Using the processed earthquake data to create earthquake training data sets and earthquake test data sets;

[0012] The specific method is: use the first 70% of the constructed data set as the training data set, 10% as the test data set, and the last 20% as the validation set.

[0013] 4) Construct a P-wave arrival time picking model based on dual-branch deep convolution and attention mechanism;

[0014] The specific picking model is:

[0015] Model structure: The dual-branch deep convolution extracts different features by splitting the correlation between the spatial dimension and the channel (depth) dimension, and uses different convolution kernels to extract different features, significantly reducing the number of parameters required for convolution calculation, effectively integrating multiple features, and improving feature expression capabilities. The input is the three-component waveform data obtained in step 3, and the output is the intermediate feature after the waveform data is convolved. The adaptive attention mechanism is used to capture the local changes and overall trends of the features, highlight important features, including average pooling, maximum pooling, and standard deviation pooling, and distribute the extracted data features. The results of the three pooling operations are added after the full connection operation, and the weights are assigned by the activation function, and then multiplied with the dual-branch convolution result to obtain the final attention feature. The output of the dual-branch deep convolution is spliced ​​with the output of the adaptive attention mechanism and multiple upsampling operations are performed to obtain the final picking result. The final picking result is compared with the actual picking result, and the weight matrix is ​​updated by backpropagation using the Adam optimization algorithm. The trained model is used for picking in subsequent steps.

[0016] The construction method is as follows:

[0017] 4.1.1) Use dual-branch deep convolution to extract different features;

[0018] Use dual-branch deep convolution to extract different features from seismic data through different convolution kernels. The function of dual-branch deep convolution is:

[0019] branch1:ReLU(Pointwise)Conv(ReLU(DepthwiseConv(X,K1=7×1)),K2=1×1)(1)

[0020] branch2:ReLU(Depthwise)Conv(ReLU(PointwiseConv(X,K3=1×1)),K4=7×1)(2)

[0021] Y1=Concat(branch1,branch2)(3)

[0022] Among them, branch1 is branch 1, branch2 is branch 2, PointwiseConv is point-wise convolution, DepthwiseConv is depth convolution, X is input data, K is the convolution kernel size, and Y is the output result of dual-branch depth convolution.

[0023] 4.1.2) Use dual-branch depth convolution to reduce the number of parameters;

[0024] Compared with traditional convolution, the number of parameters and the amount of computation in depth convolution are significantly reduced. The following will analyze and compare depthwise separable convolution with traditional convolution in terms of the number of parameters and the amount of computation. in , where H and W are the height and width of the feature map, respectively, and C in is the number of input channels. The size of the convolution kernel is K×K, and the number of output channels is C out , the output feature map size is H′×W’.

[0025] From the perspective of parameter quantity, the number of parameters of traditional convolution is K 2 ×C in ×C out , the number of parameters of the depth convolution is K 2 ×C in , the number of parameters of point-by-point convolution is C in ×C out , so the total number of parameters is K 2 ×C in +C in ×C out , calculated by formula (4), the number of parameters of depth-separable convolution is 1 / K of the traditional convolution 2 +1 / C out times.

[0026]

[0027] From the perspective of computational complexity, the computational complexity of traditional convolution is: H′×W'×C in ×C out ×K 2 The computational complexity of the depthwise convolution is H′×W'×C in ×K 2 , the computational complexity of point-by-point convolution is H'×W'×C in ×Cout , adding the computational complexity of depthwise convolution and pointwise convolution, the total computational complexity of depthwise separable convolution is H'×W'×C in ×(K 2 +C out ), calculated by formula (5), the computational complexity of depth-separable convolution is 1 / C of that of traditional convolution. out +1 / K 2 times.

[0028]

[0029] Therefore, depthwise separable convolution can optimize both computational efficiency and storage requirements, and has significant advantages in building lightweight neural networks.

[0030] 4.2) Use adaptive attention mechanism to capture local changes and overall trends of features;

[0031] Seismic data is three-component waveform data. These channels represent seismic signals in different directions, and the information contained in each channel may be significantly different at certain moments. Channel attention can automatically learn the importance of different channels through the aggregation and weighting of global information, thereby enhancing the characteristics of key channels and ignoring irrelevant or noisy channels. The adaptive attention mechanism combines three pooling operations: the maximum pooling layer, the average pooling layer, and the standard deviation pooling layer. The result obtained by adding the full connection operation is weighted by the activation function, and then multiplied by the double-branch convolution result to obtain the final attention feature. The adaptive attention mechanism function is:

[0032] MaxPool(i,j)=max 0≤m≤k,0≤n＜k f(i+m,j+n)(6)

[0033]

[0034] Y2=ReLU(Concat(FC(MaxPool,AvgPool,StdPool)))·Y1(10)

[0035] Where f represents the input feature map, (i, j) represents the position of the pooling window, k represents the size of the pooling window, and μ(i, j) is the mean of the elements in the pooling window.

[0036] 4.3) The results of 4.1.1) and 4.2) are concatenated and upsampled to obtain the final arrival-time pickup output:

[0037] Through the encoding and decoding structure, the input vibration signal is subjected to layer-by-layer feature extraction. In the encoder step 4.1.1), the double-branch deep convolution structure is described. The decoder restores the compressed feature map extracted by the encoder to the same spatial resolution as the input, thereby generating high-quality output. At each layer of the decoder, the model gradually restores the original size of the signal through the deconvolution (TransposeConvolution) operation. The deconvolution layer can not only gradually enlarge the spatial size of the feature map, but also retain the key features extracted by the encoder. At the same time, the deconvolution can effectively fuse the multi-scale information extracted by the encoder, so that the decoder maintains high fidelity of the input signal during the restoration process.

[0038] 5) Input the seismic data into the picking model in step 4, train and test it, and save the best model;

[0039] The specific method is:

[0040] 5.1) Optimize the model hyperparameters in 4):

[0041] Minimize the error, specify the mean square error of the validation set as the objective function of the optimization algorithm, and make the objective function continuously decrease to find the optimal hyperparameter value of the model. The objective function is expressed as:

[0042]

[0043] Where θ represents a set of hyperparameters, including learning rate and number of network layers, ξ val represents the loss function value on the validation set, θ opt represents the optimal hyperparameters.

[0044] 5.2) Load the optimal hyperparameter training step 4) to pick up the model weight matrix:

[0045] The Adam optimization algorithm is used to train the model weight matrix, and the loss function is:

[0046]

[0047] where y i Indicates the actual time point. Indicates the time point when the model is picked up, and N indicates the number of samples.

[0048] 6) Input the actual seismic data into the picking model trained in step 5 for timely picking;

[0049] The specific method is:

[0050] In order to better analyze the performance of the model, the accuracy of the test results was calculated during the training process. The accuracy rate indicates the correctness of the overall prediction of the model at that time. Although the accuracy rate can reflect the overall performance of the model, in scenarios such as earthquake monitoring where there may be class imbalance, using the accuracy rate alone may mask the model's inadequacy in identifying microseismic events. Therefore, in addition to the accuracy rate, the loss value (Loss), recall rate (Recall) and F1 score are introduced for a more comprehensive performance evaluation. The formula is as follows:

[0051]

[0052] Among them, TP is the number of positive samples detected, that is, the number of correctly detected waveforms. FP is the number of positive samples detected, that is, the number of negative samples, that is, the number of wrong waveforms mistakenly detected as correct waveforms. TN is the number of negative samples detected, that is, the number of wrong waveforms correctly detected. FN is the number of positive samples detected as negative, that is, the number of correct waveforms missed.

[0053] 7) Using the processed microseismic data to create a microseismic training dataset and a microseismic test dataset;

[0054] The specific method is: use the first 70% of the constructed data set as the training data set, 10% as the test data set, and the last 20% as the validation set.

[0055] 8) Input the microseismic data to the qualified picking model in step 5) for transfer learning, fine-tune the model, and realize the timely picking of microseismic data.

[0056] The specific method is:

[0057] 8.1) Load the model proposed in step 5), input the microseismic training set data, and minimize the loss function L on the source domain through pre-training S , L S The formula is:

[0058]

[0059] Among them, D s represents the seismic dataset, θ represents the parameters of different layers, and S represents the source domain.

[0060] This process enables the model to learn some common underlying features that are still applicable to similar tasks in the target domain.

[0061] 8.2) Freeze the downsampling layer parameters of the model;

[0062] In order to reduce the amount of calculation and maintain the general features of pre-training, the downsampling layer parameters of the model are frozen, and the upsampling layer parameters are fine-tuned for microseismic data. Fine-tuning refers to optimizing high-level parameters on a small amount of data in the target domain, so that the model can better adapt to the characteristics of microseismic data. T The formula is:

[0063]

[0064] Among them, θ s,low represents the downsampling layer parameters of the model, θ T,high represents the upsampling layer parameters, D T represents the microseismic dataset, and T represents the target domain.

[0065] 8.3) Save the fine-tuned model, read the microseismic test data set, perform time-based pick-up on it, and save the results.

[0066] The beneficial effects of the present invention are:

[0067] 1. By introducing depthwise separable convolution, the computational complexity and number of parameters are greatly reduced, enabling the model to run efficiently on resource-constrained terminal devices such as embedded systems or mobile devices.

[0068] 2. The adaptive mixed-scale attention mechanism retains the overall trend and local variability by adjusting the weights of features at different scales, improving the model's ability to process complex waveform data, especially in noisy environments, and can still accurately pick up the arrival time of P waves. This invention solves the current common problem that terminal equipment resources are limited and it is difficult to quickly and accurately pick up the arrival time of vibration signals.

[0069] 3. The picking method can be transferred to microseismic data to establish a picking model for mine microseismic data, so as to realize the accurate arrival time picking of P waves of mine microseismic data, thus solving the problems of low accuracy and poor robustness of traditional microseismic picking models. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 The overall structure diagram of the model;

[0071] Figure 2 2D plan view of the model;

[0072] Figure 3 The overall flow chart of model training;

[0073] Figure 4 Structure diagram of adaptive mixed-scale attention mechanism;

[0074] Figure 5 This is the structure diagram of the transfer learning model;

[0075] Figure 6 Compare the actual value of the test set picked up at the time with the value picked up by the model. DETAILED DESCRIPTION

[0076] The present invention provides a method for picking up the arrival time of a vibration P wave based on a lightweight transfer learning model, which specifically comprises the following steps:

[0077] 1) Import earthquake data.

[0078] In the step 1), the earthquake data specifically includes: the time when the earthquake occurs, the vibration wave is generated, the source coordinates are calculated by the program after multiple sensors collect the vibration waves, and the arrival time of the manually marked part. The source coordinates are composed of three components: X, Y, and Z, which are the relative offset of the coordinate origin defined by the earthquake, and the unit is m. The manually marked arrival time is the manually marked earthquake P wave arrival picking point.

[0079] 2) Preprocess the seismic data.

[0080] The preprocessing method in step 2) is specifically as follows: processing the seismic data. Since the waveform of the original data of the seismic event is in txt format, it is not suitable for subsequent input model calculation. Therefore, the data is converted to convert the waveform of the event into its corresponding hdf5 array format for subsequent processing.

[0081] 3) Use the processed earthquake data to create earthquake training data sets and earthquake test data sets.

[0082] The specific method of step 3) is: the first 70% of the constructed data set is used as a training data set, and the last 30% is used as a test data set.

[0083] 4) Construct a P-wave arrival time picking model based on dual-branch deep convolution and attention mechanism.

[0084] The picking model of step 4) is specifically:

[0085] Model structure: The dual-branch deep convolution extracts different features by splitting the correlation between the spatial dimension and the channel (depth) dimension, and uses different convolution kernels to extract different features, significantly reducing the number of parameters required for convolution calculation, effectively integrating multiple features, and improving feature expression capabilities. The input is the three-component waveform data obtained in step 3, and the output is the intermediate feature after the waveform data is convolved. The adaptive attention mechanism is used to capture the local changes and overall trends of the features, highlight important features, including average pooling, maximum pooling, and standard deviation pooling, and distribute the extracted data features. The results of the three pooling operations are added after the full connection operation, and the weights are assigned by the activation function, and then multiplied with the dual-branch convolution result to obtain the final attention feature. The output of the dual-branch deep convolution is spliced ​​with the output of the adaptive attention mechanism and multiple upsampling operations are performed to obtain the final picking result. The final picking result is compared with the actual picking result, and the weight matrix is ​​updated by backpropagation using the Adam optimization algorithm. The trained model is used for picking in subsequent steps.

[0086] 4.1.1) Use dual-branch deep convolution to extract different features;

[0087] Use dual-branch deep convolution to extract different features from seismic data through different convolution kernels. The function of dual-branch deep convolution is:

[0088] branch1:ReLU(Pointwise)Conv(ReLU(DepthwiseConv(X,K1=7×1)),K2=1×1)(1)

[0089] branch2:ReLU(Depthwise)Conv(ReLU(PointwiseConv(X,K3=1×1)),K4=7×1)(2)

[0090] Y1=Concat(branch1,branch2)(3)

[0091] Among them, branch1 is branch 1, branch2 is branch 2, PointwiseConv is point-wise convolution, DepthwiseConv is depth convolution, X is input data, K is the convolution kernel size, and Y is the output result of dual-branch depth convolution.

[0092] 4.1.2) Use dual-branch depth convolution to reduce the number of parameters;

[0093] Compared with traditional convolution, the number of parameters and the amount of computation in depth convolution are significantly reduced. The following will analyze and compare depthwise separable convolution with traditional convolution in terms of the number of parameters and the amount of computation. in , where H and W are the height and width of the feature map, respectively, and C inis the number of input channels. The size of the convolution kernel is K×K, and the number of output channels is C out , the output feature map size is H'×W'.

[0094] From the perspective of parameter quantity, the number of parameters of traditional convolution is K 2 ×C in ×C out , the number of parameters of the depth convolution is K 2 ×C in , the number of parameters of point-by-point convolution is C in ×C out , so the total number of parameters is K 2 ×C in +C in ×C out , calculated by formula (4), the number of parameters of depth-separable convolution is 1 / K of the traditional convolution 2 +1 / C out times.

[0095]

[0096] From the perspective of computational complexity, the computational complexity of traditional convolution is: H′×W'×C in ×C out ×K 2 The computational complexity of the depthwise convolution is H′×W′×C in ×K 2 , the computational complexity of point-by-point convolution is H'×W'×C in ×C out , adding the computational complexity of depthwise convolution and pointwise convolution, the total computational complexity of depthwise separable convolution is H'×W'×C in ×(K 2 +C out ), calculated by formula (5), the computational complexity of depth-separable convolution is 1 / C of that of traditional convolution. out +1 / K 2 times.

[0097]

[0098] Therefore, depthwise separable convolution can optimize both computational efficiency and storage requirements, and has significant advantages in building lightweight neural networks.

[0099] 4.2) Use adaptive attention mechanism to capture local changes and overall trends of features;

[0100] Seismic data is three-component waveform data. These channels represent seismic signals in different directions, and the information contained in each channel may be significantly different at certain moments. Channel attention can automatically learn the importance of different channels through the aggregation and weighting of global information, thereby enhancing the characteristics of key channels and ignoring irrelevant or noisy channels. The adaptive attention mechanism combines three pooling operations: the maximum pooling layer, the average pooling layer, and the standard deviation pooling layer. The result obtained by adding the full connection operation is weighted by the activation function, and then multiplied by the double-branch convolution result to obtain the final attention feature. The adaptive attention mechanism function is:

[0101] MaxPool(i,j)=max 0≤m＜k,0≤n＜k f(i+m,j+n)(6)

[0102]

[0103] Y2=ReLU(Concat(FC(MaxPool,AvgPool,StdPool)))·Y1(10)

[0104] Where f represents the input feature map, (i, j) represents the position of the pooling window, k represents the size of the pooling window, and μ(i, j) is the mean of the elements in the pooling window.

[0105] 4.3) The results of 4.1.1) and 4.2) are concatenated and upsampled to obtain the final arrival-time pickup output:

[0106] Through the structure of encoding and decoding, the input vibration signal is subjected to layer-by-layer feature extraction. The encoder is the dual-branch deep convolution structure described in step 4.1.1). The decoder restores the compressed feature map extracted by the encoder to the same spatial resolution as the input, thereby generating high-quality output. At each layer of the decoder, the model gradually restores the original size of the signal through the deconvolution (TransposeConvolution) operation. The deconvolution layer can not only gradually enlarge the spatial size of the feature map, but also retain the key features extracted by the encoder. At the same time, deconvolution can effectively fuse the multi-scale information extracted by the encoder, so that the decoder maintains high fidelity of the input signal during the restoration process.

[0107] 5) Input the seismic data into the picking model in step 4, train and test it, and save the best model;

[0108] 5.1) Pick up the model hyperparameters optimization in step 4:

[0109] Minimize the error, specify the mean square error of the validation set as the objective function of the optimization algorithm, and make the objective function continuously decrease to find the optimal hyperparameter value of the model. The objective function is expressed as:

[0110]

[0111] Where θ represents a set of hyperparameters, including learning rate and number of network layers, ξ val represents the loss function value on the validation set, θ opt represents the optimal hyperparameters.

[0112] 5.2) Load the optimal hyperparameter training step 4) to pick up the model weight matrix:

[0113] The Adam optimization algorithm is used to train the model weight matrix, and the loss function is:

[0114]

[0115] where y i Indicates the actual time point. Indicates the time point when the model is picked up, and N indicates the number of samples.

[0116] 6) Inputting the actual seismic data into the picking model trained in step 5) respectively for timely picking;

[0117] The specific method of step 6) is:

[0118] In order to better analyze the performance of the model, the accuracy of the test results was calculated during the training process. The accuracy rate indicates the correctness of the overall prediction of the model at that time. Although the accuracy rate can reflect the overall performance of the model, in scenarios such as earthquake monitoring where there may be class imbalance, using the accuracy rate alone may mask the model's inadequacy in identifying microseismic events. Therefore, in addition to the accuracy rate, the loss value (Loss), recall rate (Recall) and F1 score are introduced for a more comprehensive performance evaluation. The formula is as follows:

[0119]

[0120] Among them, TP is the number of positive samples detected, that is, the number of correctly detected waveforms. FP is the number of positive samples detected, that is, the number of negative samples, that is, the number of wrong waveforms mistakenly detected as correct waveforms. TN is the number of negative samples detected, that is, the number of wrong waveforms correctly detected. FN is the number of positive samples detected as negative, that is, the number of correct waveforms missed.

[0121] 7) Using the processed microseismic data to create a microseismic training dataset and a microseismic test dataset;

[0122] The specific method of step 7) is: the first 70% of the constructed data set is used as a training data set, and the last 30% is used as a test data set.

[0123] 8) Input the microseismic data to the qualified picking model in step 5) for transfer learning, fine-tune the model, and realize the timely picking of microseismic data.

[0124] The specific method of step 8) is:

[0125] 8.1) Load the model proposed in step 5), input the microseismic training set data, and minimize the loss function L on the source domain through pre-training S , L S The formula is:

[0126]

[0127] Among them, D s represents the seismic dataset, θ represents the parameters of different layers, and S represents the source domain.

[0128] This process enables the model to learn some common underlying features that are still applicable to similar tasks in the target domain.

[0129] 8.2) Freeze the downsampling layer parameters of the model;

[0130] In order to reduce the amount of calculation and maintain the general features of pre-training, the downsampling layer parameters of the model are frozen, and the upsampling layer parameters are fine-tuned for microseismic data. Fine-tuning refers to optimizing high-level parameters on a small amount of data in the target domain, so that the model can better adapt to the characteristics of microseismic data. T The formula is:

[0131]

[0132] Among them, θ s,low represents the downsampling layer parameters of the model, θ T,hig represents the upsampling layer parameters, D T represents the microseismic dataset, and T represents the target domain.

[0133] 8.3) Save the fine-tuned model, read the microseismic test data set, perform time-based pick-up on it, and save the results.

[0134] Embodiment 1:

[0135] The Italian INSTANCE earthquake dataset and a mine earthquake dataset in Liaoning were used for training and testing. The earthquake dataset selected 200,000 earthquake events, and the microseismic dataset selected 10,100 data.

[0136] Step 1: Import earthquake and microseismic event data, preprocess the data, convert the file format of the original data into hdf5 array format; divide the data into training set, validation set and test set with a ratio of 7:1:2.

[0137] Step 2: Set up the model running environment and parameters

[0138] The model evaluation experiment was implemented in Pytorch and Python 3.8, and the hardware used was a 13th Gen Intel(R) Core(TM) i9-13900HX 2.20GHz processor, an NVIDIA GeForce RTX 4060Laptop GPU, and a memory size of 128.0GB. The relevant parameters of the model were set, including 100 iterations of the model, 16 batches, and an optimizer learning rate of 0.001.

[0139] Step 3: Construct a P-wave arrival time picking model based on dual-branch deep convolution and attention mechanism. Model structure: The dual-branch deep convolution splits the correlation between the spatial dimension and the channel (depth) dimension, uses different convolution kernels to extract different features, significantly reduces the number of parameters required for convolution calculation, effectively integrates multiple features, and improves the feature expression ability. The input is the three-component waveform data obtained in step 2, and the output is the intermediate feature after the waveform data is convolved. The adaptive attention mechanism is used to capture the local changes and overall trends of the features, highlight the important features, including average pooling, maximum pooling and standard deviation pooling, and distribute the extracted data features. The results of the three pooling operations are added after the full connection operation, and the weights are assigned by the activation function, and then multiplied with the dual-branch convolution result to obtain the final attention feature. The output of the dual-branch deep convolution is spliced ​​with the output of the adaptive attention mechanism and multiple upsampling operations are performed to obtain the final picking result. The final picking result is compared with the actual picking result, and the weight matrix is ​​updated by backpropagation using the Adam optimization algorithm. The trained model is used for picking in the subsequent steps. The results are compared with other different picking models. The results are shown in Table 1, and the best results are marked in bold.

[0140] Step 4: Use the model, input microseismic data to fine-tune the model, and save the fine-tuned model. Read the test data and pick it. Compare the results with the model without fine-tuning. The microseismic arrival time picking model is the model used directly without migration, the initial model is the model obtained by using microseismic data as the training set, and the transfer learning model is the model proposed in this paper. The comparison results are shown in Table 2, and the best result is marked in bold.

[0141] Table 1 Comparison of picking effects of different models

[0142]

[0143] Table 2 Table 4-3 Comparative experimental results of the on-time pickup transfer learning model using INSTANCE data as the source domain

[0144]

[0145] It can be seen from the results that the model proposed in the present invention is superior to other models in terms of accuracy and loss value on the adopted seismic data set, and has better stability; in transfer learning, all indicators of the transfer learning model proposed in the present invention are higher than those of the other two models, and has better accuracy and robustness.

[0146] The P-wave arrival time picking method proposed in the present invention integrates lightweight and transfer learning models, uses deep learning and transfer learning technologies to realize the P-wave arrival time picking of earthquake and microseismic data, has high accuracy and stability, and has good application value.

Claims

1. A method for picking up vibration P-wave arrival time based on a lightweight transfer learning model, characterized in that: The steps are: 1) Import earthquake and microseismic event data; 2) Preprocessing of earthquake and microseismic data; 3) Using the processed earthquake data to create earthquake training data sets and earthquake test data sets; 4) Construct a P-wave arrival time picking model based on dual-branch deep convolution and attention mechanism; 5) Input the seismic data into the picking model in step 4, train and test it, and save the best model; 6) Input the actual seismic data into the picking model trained in step 5 for timely picking; 7) Using the processed microseismic data to create a microseismic training dataset and a microseismic test dataset; 8) Input the microseismic data to the qualified picking model in step 5) for transfer learning, adjust the model, and realize the timely picking of microseismic data.

2. According to the method for picking up the arrival time of vibration P wave of a lightweight transfer learning model according to claim 1, it is characterized in that: In the above 1), the earthquake data specifically include: the time when the earthquake occurs, the vibration wave is generated, the source coordinates are calculated by the program after multiple sensors collect the vibration waves, and the arrival time of the manually marked part. The source coordinates are composed of three components of X, Y, and Z, which are the relative offset of the coordinate origin defined by the earthquake, in meters. The manually marked arrival time is the manually marked earthquake P wave arrival picking point.

3. According to the method for picking up the arrival time of vibration P wave of a lightweight transfer learning model according to claim 1, it is characterized in that: The preprocessing method in 2) is specifically: processing the seismic data, and converting the txt format waveform of the original data of the seismic event into the corresponding hdf5 array format.

4. According to the method for picking up the arrival time of vibration P wave of a lightweight transfer learning model according to claim 1, it is characterized in that: In the above 4), the structure of the model is: dual-branch deep convolution extracts different features by splitting the correlation between spatial dimension and channel dimension using different convolution kernels; the input is the three-component waveform data obtained in 3), and the output is the intermediate features after the waveform data is convolved; The adaptive attention mechanism is used to capture local changes and overall trends of features and highlight important features, including average pooling, maximum pooling and standard deviation pooling, and distributes and extracts data features. The results of the three pooling operations are added together after full connection, and weighted by the activation function, and then multiplied with the double-branch convolution result to obtain the final attention feature; The output of the dual-branch deep convolution is concatenated with the output of the adaptive attention mechanism and then upsampled multiple times to obtain the final picking result.

5. According to the method for picking up the arrival time of vibration P wave of a lightweight transfer learning model according to claim 1, it is characterized in that: In the above 4), the specific steps are: 4.1.1) Use dual-branch deep convolution to extract different features; Use dual-branch deep convolution to extract different features from seismic data through different convolution kernels. The function of dual-branch deep convolution is: branch1:ReLU(Pointwise)Conv(ReLU(DepthwiseConv(X,K1=7×1)), K2=1×1)(1)branch2:ReLU(Depthwise)Conv(ReLU(PointwiseConv(X,K3=1×1)), K4=7×1)(2)Y1=Concat(branch1,branch2)(3)where branch1 is branch 1, branch2 is branch 2, PointwiseConv is pointwise convolution, DepthwiseConv is depth convolution, X is input data, K is convolution kernel size, and Y is the output result of dual-branch depth convolution; 4.1.2) Use dual-branch depth convolution to reduce the number of parameters; Assume H×W×C in , where H and W are the height and width of the feature map, respectively, and C in is the number of input channels, the size of the convolution kernel is K×K, and the number of output channels is C out , the output feature map size is H'×W'''; From the perspective of parameter quantity, the parameter quantity of point-by-point convolution is C in ×C out , so the total number of parameters is K 2 ×C in +C in ×C out , calculated by formula (4), the number of parameters of depth-separable convolution is 1 / K of the traditional convolution 2 +1 / C out times; From the perspective of computational complexity, the computational complexity of deep convolution is H′×W'×C in ×K 2 , the computational complexity of point-by-point convolution is H′×W'×C in ×C out , adding the computational complexity of depthwise convolution and pointwise convolution, the total computational complexity of depthwise separable convolution is H′×W'×C in ×(K 2 +C out ); 4.2) Use adaptive attention mechanism to capture local changes and overall trends of features; Seismic data is three-component waveform data. These channels represent seismic signals in different directions. The information contained in each channel may be significantly different at different times. Channel attention can automatically learn the importance of different channels through the aggregation and weighting of global information, enhance the characteristics of key channels, and ignore irrelevant or noise-affected channels. The adaptive attention mechanism combines three pooling operations: the maximum pooling layer, the average pooling layer, and the standard deviation pooling layer. The result obtained by adding the full connection operation is weighted by the activation function and then multiplied by the double-branch convolution result to obtain the final attention feature. The adaptive attention mechanism function is: MaxPool(i,j)=max 0≤m＜k,0≤n＜k f(i+m,j+n)(6) Y2=ReLU(Concat(FC(MaxPool,AvgPool,StdPool)))·Y1(10)where f represents the input feature map, (i,j) represents the position of the pooling window, k represents the size of the pooling window, and μ(i,j) is the mean of the elements in the pooling window. 4.3) The results of 4.1.1) and 4.2) are concatenated and upsampled to obtain the final arrival-time pickup output: Through the encoding and decoding structure, the input vibration signal is subjected to layer-by-layer feature extraction. The encoder is the dual-branch deep convolution structure described in step 4.1.1). The decoder restores the compressed feature map extracted by the encoder to the same spatial resolution as the input and generates an output. At each layer of the decoder, the model gradually restores the original size of the signal through deconvolution operations.

6. According to the method for picking up the arrival time of vibration P wave of a lightweight transfer learning model according to claim 1, it is characterized in that: In the above 5), the specific steps are: 5.1) Optimize the hyperparameters of the model picked in 4): Minimize the error, specify the mean square error of the validation set as the objective function of the optimization algorithm, and make the objective function continuously decrease to find the optimal hyperparameter value of the model. The objective function is expressed as: Where θ represents a set of hyperparameters, including learning rate and number of network layers, ξ val represents the loss function value on the validation set, θ opt represents the optimal hyperparameters; 5.2) Load the optimal hyperparameter training step 4) to pick up the model weight matrix: The Adam optimization algorithm is used to train the model weight matrix, and the loss function is: where y i Indicates the actual time point. Indicates the time point when the model is picked up, and N indicates the number of samples.

7. The method for picking up the arrival time of vibration P waves of a lightweight transfer learning model according to claim 1, characterized in that: In the above 6), the specific steps are: The loss value Loss, recall rate Recall and F1 score are introduced for a more comprehensive performance evaluation. The formula is as follows: Among them, TP is the number of positive samples detected, that is, the number of correctly detected waveforms; FP is the number of negative samples detected, that is, the number of incorrect waveforms mistakenly detected as correct waveforms; TN is the number of negative samples detected, that is, the number of correctly detected incorrect waveforms; FN is the number of negative samples detected, that is, the number of correct waveforms missed.

8. The method for picking up the arrival time of vibration P waves of a lightweight transfer learning model according to claim 1, characterized in that: In the above 8), the specific steps are: 8.1) Load the model in step 5), input the microseismic training data, and minimize the loss function L on the source domain through pre-training S , L S The formula is: Among them, D s represents the seismic data set, θ represents the parameters of different layers, and S represents the source domain; thus, the model learns the common underlying features; 8.2) Freeze the downsampling layer parameters of the model; Select the downsampling layer parameters of the frozen model and fine-tune the upsampling layer parameters for the microseismic data. Fine-tuning means optimizing the high-level parameters on a small amount of data in the target domain so that the model can adapt to the characteristics of the microseismic data. T The formula is: Among them, θ s,low represents the downsampling layer parameters of the model, θ T,high represents the upsampling layer parameters, D T represents the microseismic dataset, T represents the target domain; 8.3) Save the fine-tuned model, read the microseismic test data set, and pick and save the results when it is done.

Images