First-arrival picking method and system for seismic data, and computer device

Abstract

A first-arrival picking method for seismic data, which relates to the technical field of seismic monitoring. The method comprises: collecting seismic data; preliminarily picking a first-arrival moment of a seismic signal in the seismic data by using an automatic first-arrival picking algorithm, and generating a training sample set; performing manual correction on the first-arrival moment picked by the first-arrival picking algorithm, so as to generate a validation sample set; respectively making a noise label and a clean label by using the training sample set and the validation sample set, so as to generate a meta-learning data set; on the basis of features of the meta-learning data set, designing a corresponding deep neural network model; on the basis of a double-loop training method for meta-learning, performing iterative training on the deep neural network model by using the meta-learning data set; and performing first-arrival picking on the seismic data by using the trained deep neural network model. By means of an inner and outer double-loop training method and an online gradient updating method, the precision and cost effectiveness of a data driving method are improved. The present invention further relates to a first-arrival picking system for seismic data, and a computer device.

Description

A method, system and computer equipment for picking first arrival of seismic data Technical Field

[0001] The present invention relates to the technical field of seismic data processing and seismic monitoring, and in particular to a method, system and computer equipment for picking first arrivals of seismic data. Background Art

[0002] In the fields of seismic data processing and seismic monitoring, accurate first arrival selection is of vital importance.

[0003] Traditionally, first-arrival picking is done manually. However, with the explosive growth of data volumes, manual operations can no longer meet the practical needs of seismic data processing. Therefore, data-driven automatic first-arrival picking methods have emerged.

[0004] However, data-driven first arrival picking methods rely heavily on a large amount of manually annotated data, requiring significant manpower and material resources. Furthermore, improper labeling can cause errors between the annotated first arrival time and the actual first arrival time, leading to label noise and affecting method performance. Furthermore, data-driven methods face generalization issues when faced with different noise characteristics and geological environments. Summary of the Invention

[0005] In view of the generalization problem of existing data-driven picking technology when facing different noise characteristics and geological environments, the present invention proposes a first-arrival picking method, system and computer equipment for seismic data. Based on the double-loop training method of meta-learning, the meta-learning dataset is used to iteratively train the deep neural network model; the trained deep neural network model is used to pick the first arrival of seismic data, thereby solving the generalization problem of existing data-driven picking technology.

[0006] A method for picking first arrivals of seismic data comprises the following steps:

[0007] Acquisition of seismic data;

[0008] Using an automatic first-break picking algorithm to preliminarily pick the first-break moments of seismic signals in seismic data to generate a meta-learning dataset; the meta-learning dataset includes a training sample set and a validation sample set;

[0009] Design a corresponding deep neural network model based on the characteristics of the meta-learning dataset;

[0010] A double-loop training method based on meta-learning, which iteratively trains deep neural network models using meta-learning datasets;

[0011] Use the trained deep neural network model to pick the first arrival of the earthquake data to be detected;

[0012] The meta-learning-based double-loop training process uses a meta-learning dataset to iteratively train a deep neural network. Each iterative training includes the following steps:

[0013] Sample N small batches of samples from the training sample set;

[0014] The meta-network weights Give independent sub-networks;

[0015] Take M small batches of samples from the validation sample set;

[0016] Use the online gradient update method to calculate the sub-network loss value and back-propagate to optimize the weight matrix ω;

[0017] The weight matrix ω processed by using N small batches of training samples and the optimized weight matrix ω * , using the loss function to calculate the gradient and backpropagate to optimize the weights of the meta-network

[0018] Furthermore, after the seismic data is collected, filtering is performed on the seismic data, and then clipping, normalization and first arrival picking operations are performed on the processed seismic data.

[0019] Furthermore, the input data of the deep neural network model is earthquake records, each earthquake record contains 64 channels of earthquake data, each channel contains 512 sampling points, and each sampling point is separated by 10 milliseconds; the output of the deep neural network model is a 0-1 matrix.

[0020] Furthermore, the deep neural network model adopts a U-Net model.

[0021] Furthermore, the method of using the automatic first arrival picking algorithm to preliminarily pick the first arrival time of the seismic signal in the seismic data to generate a meta-learning dataset specifically includes the following steps:

[0022] The sampling points before and after the first arrival are divided into two categories: before the first arrival and after the first arrival, and different labels are assigned to each category.

[0023] The first arrival picking results with errors from the automatic first arrival picking algorithm are used as noise labels;

[0024] The manually revised first arrival time is used as the clean label;

[0025] Generate meta-learning datasets based on noisy labels and clean labels.

[0026] Furthermore, the adopted loss function is expressed as:

[0027] Loss=ωf i (θ)

[0028] Where ω is the weight matrix, f i (θ) represents the loss function associated with the noise label i, where θ is a trainable parameter in the deep neural network model.

[0029] Furthermore, the weight matrix ω processed by using N small batches of training samples and optimizing the weight matrix ω is * , using the loss function to calculate the gradient and backpropagate to optimize the weights of the meta-network Its specific expression is:

[0030]

[0031] Among them, the symbol * represents the optimal value of the parameter, f j represents the loss function associated with the clean label j.

[0032] Furthermore, the online gradient update method is used to calculate the sub-network loss value, and the weight matrix ω is optimized by back propagation. The estimated value of the weight matrix ω is expressed as:

[0033]

[0034] Where η is the learning rate of ω, t represents the current iteration number, and m is the number of samples in the M mini-batch samples;

[0035] The estimated value After normalization, it is expressed as:

[0036]

[0037]

[0038] Here, δ is the minimum value that prevents ω from being completely zero.

[0039] Furthermore, a first arrival picking system for seismic data includes:

[0040] An acquisition module, used for acquiring seismic data;

[0041] A data set generation module is used to preliminarily pick the first arrival time of the seismic signal in the seismic data using an automatic first arrival picking algorithm to generate a meta-learning data set; the meta-learning data set includes a training sample set and a validation sample set;

[0042] The training module is used for a double-loop training method based on meta-learning, and iteratively trains the deep neural network model using the meta-learning dataset. Each training iteration includes the following steps:

[0043] Sample N small batches of samples from the training sample set;

[0044] The meta-network weights Give independent sub-networks;

[0045] Take M small batches of samples from the validation sample set;

[0046] Use the online gradient update method to calculate the sub-network loss value and back-propagate to optimize the weight matrix ω;

[0047] The weight matrix ω processed by using N small batches of training samples and the optimized weight matrix ω * , using the loss function to calculate the gradient and backpropagate to optimize the weights of the meta-network

[0048] The first-arrival picking module is used to input the seismic data to be detected into the trained deep neural network model and output the first-arrival picking results of the seismic data to be detected.

[0049] Furthermore, a computer device for picking first arrival of seismic data includes: a memory, a processor and a computer program stored in the memory, and when the processor executes the computer program, the steps of the method for picking first arrival of seismic data described in any one of claims 1-8 are implemented.

[0050] The present invention provides a method, system, and computer device for picking first arrivals of seismic data, which have the following beneficial effects:

[0051] The present invention uses an automatic first-arrival picking algorithm to pick the first-arrival moment of seismic signals in seismic data, generates a meta-learning data set, and uses a double-loop training method based on meta-learning to train a deep neural network model; the trained deep neural network model is used to pick the first-arrival of seismic data, which significantly reduces the manpower and resources required for the data-driven method, reduces costs, and improves efficiency and accuracy, and successfully solves problems such as noise labels and training costs through the internal and external double-loop training method and the online gradient update method; compared with existing methods, the present method can be seamlessly adapted to various deep learning models, meet various end-to-end requirements, and has higher efficiency than the model-driven first-arrival picking method, while solving the generalization problem of existing data-driven picking technology and improving the picking accuracy and cost-effectiveness of the data-driven method. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] FIG1 is a schematic diagram of a meta-learning structure according to an embodiment of the present invention;

[0053] FIG2 is a schematic diagram of the structure of a U-Net network model according to an embodiment of the present invention;

[0054] FIG3 is a diagram showing a data label according to an embodiment of the present invention;

[0055] FIG4 is a flow chart of internal and external double-loop training based on meta-learning in an embodiment of the present invention;

[0056] FIG5 is a schematic diagram of field data collection using a trained U-Net model in an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0058] An automatic first arrival picking method based on meta-learning includes the following steps:

[0059] Collect seismic data.

[0060] The automatic first arrival picking algorithm is used to preliminarily pick the first arrival time of the seismic signal in the seismic data to generate a training sample set.

[0061] The first arrival time picked by the first arrival picking algorithm is manually corrected to generate a verification sample set.

[0062] The training sample set and the validation sample set are used to produce noisy labels and clean labels respectively to generate a meta-learning dataset.

[0063] Design the corresponding deep neural network model according to the characteristics of the meta-learning dataset.

[0064] A double-loop training method based on meta-learning is used to iteratively train deep neural network models using meta-learning datasets.

[0065] The trained deep neural network model is used to perform first arrival picking on seismic data.

[0066] The meta-learning-based double-loop training process uses a meta-learning dataset to iteratively train a deep neural network. Each iterative training includes the following steps:

[0067] Sample N small batches of samples from the training sample set.

[0068] Set the initial weight matrix ω0, calculate the gradient according to the loss function, and backpropagate to optimize the meta-network weights

[0069] The meta-network weights Assign an independent sub-network.

[0070] Take M small batches of samples from the validation sample set.

[0071] The online gradient update method is used to calculate the sub-network loss value and backpropagate to optimize the weight matrix ω.

[0072] Processing the optimized weight matrix ω maintains the stability of training.

[0073] Using N small batches of training samples and the processed weight matrix ω * , calculate the gradient according to the loss function, and back-propagate to optimize the weights of the meta-network

[0074] Meta-learning, a method within deep learning, enables models to quickly adapt to new tasks with limited data and iterations. This paper uses the concept of meta-learning, as shown in Figure 1. By training a model through a series of different tasks, the model acquires the "ability to learn," also known as the "meta-learning knowledge" in Figure 1. This knowledge empowers the model with greater adaptability, enabling it to effectively and quickly handle new, previously unseen tasks. Specifically, the process involves the following steps:

[0075] Step 1: First, seismic data from a mining area in China was collected. The data was then initially filtered to remove noise and improve the signal-to-noise ratio and quality of the field seismic data. The processed data was then cropped, normalized, and first-break picked. Non-picked seismic traces were deleted, and data with no valid time was truncated to create a dataset that conforms to the input paradigm of the deep neural network model: each sample in the dataset contains 64 traces, each with 512 sampling points, and each sampling point is sampled 1 millisecond apart.

[0076] Step 2: This invention does not involve model improvements. The choice of deep neural network depends on the complexity of the data and the specific requirements of the application. CNN and U-Net are two widely used models for first-arrival picking. Given that first-arrival picking requires capturing pixel-level details and preserving spatial context, U-Net is used as an example model for this invention.

[0077] The specific number of layers and structure of U-Net are designed based on data and experience. A schematic diagram of U-Net is shown in Figure 2. U-Net includes convolution, pooling, deconvolution, and skip connection operations. The U-Net used in the example consists of six blocks, including three compression path blocks and three expansion path blocks. Each compression path block includes three convolution layers and one maximum pooling layer. As the feature map passes through each module, the number of channels doubles and the size is halved. Conversely, each module in the expansion path doubles the size of the input feature map and reduces the number of channels by half through deconvolution. The modified feature map is then fused with the feature map saved in the corresponding compression path through a skip connection operation.

[0078] Step 3 and Step 4: Use STA / LTA algorithm and manual seismic event detection to perform preliminary first arrival picking and calibration.

[0079] First, we used an automatic first-break picking algorithm, the STA / LTA algorithm, to perform a preliminary search for the first arrival times of seismic signals in the data. Due to the variability of seismic signals and the complexity of the subsurface, this first-break picking algorithm suffers from accuracy issues. Therefore, we manually corrected a small number of first-break times picked by the STA / LTA algorithm (5% of the total sample size in the training example).

[0080] Step 5: Create labels and generate meta-learning datasets.

[0081] As a pixel-level segmentation task, the first arrival picking task usually assigns a binary value of 0 or 1 to each pixel in the sample. There are two different ways to implement this method. The first method is to classify the sampling points at or near the first arrival moment into one category, and the remaining pixels into another category. Because the above-mentioned labeling method has the problem of sample imbalance, it is necessary to use methods such as cost-sensitive learning functions to solve this problem. Therefore, another labeling method is adopted, that is, the sampling points before and after the first arrival picking moment are classified into two categories and assigned different labels (0 or 1) respectively.

[0082] Specifically, the potentially inaccurate first arrival times picked up by the STA / LTA algorithm are used as noise labels, while clean labels are generated based on the first arrival times corrected by experts. Figure 3 shows the seismic data and its corresponding clean and noisy labels, clearly contrasting the difference between the clean and noisy labels for the seismic data and those with and without human correction. The manually corrected dataset serves as the validation set. The remaining uncorrected first arrival times, those automatically picked up by the STA / LTA algorithm, serve as the training set. Both the validation and training sets are used together when training the model.

[0083] Step 6: Double-loop training process based on meta-learning.

[0084] In order to solve the error problem of the STA / LTA algorithm in picking the first arrival, a validation dataset with clean labels is used to guide the deep neural network during the training phase to encourage the network to adaptively learn from the training dataset samples with label noise, so that the network can effectively assign appropriate weights to the gradient of each sample. Specifically, a weighted loss function is designed:

[0085] Loss=ωf i (θ) (1)

[0086] In formula (1), ω is the weight matrix, which is used to eliminate the loss error caused by model prediction and label inaccuracy during training due to label noise. Operationally, the weight matrix will be adaptively adjusted during training, assigning lower weights to noisy labels and higher weights to noise-free labels. i (θ) represents the loss function associated with the noisy label i, where θ is a trainable parameter in the model.

[0087] The model training in this invention adopts an internal and external double-loop training process based on meta-learning. The double-loop training diagram is shown in Figure 4. Specifically, during training, the model parameters θ are updated using samples from the noisy dataset, and the weight matrix ω is updated using a validation set with clean labels:

[0088]

[0089] In formula (2), M and N represent the number of samples in the training set and validation set, respectively. In one iterative training, the update of the model parameter θ depends on the gradient calculated from a small batch of samples in the training set. Taking the SGD optimization algorithm as an example:

[0090]

[0091] In formula (3), α is the learning rate. For the calculation of the gradient of parameter ω, the present invention adopts an online gradient update method to reduce the computing power burden brought by the internal and external double loop iterative training, and uses the gradient value of a small batch of samples in the validation set as the estimated value

[0092]

[0093] In formula (4), η is the learning rate of parameter ω.

[0094] During the continuous training process, the model learns to assign greater weight to the loss provided by samples with similar gradient directions in the training set and the test set, and assign smaller weight to the loss provided by samples with different gradient directions. It is processed to ensure that it is a positive number and normalized to eliminate the influence of the learning rate η:

[0095]

[0096]

[0097] Here, δ is the minimum value that prevents ω from being all zero. After updating ω, we solve equation (3) and finally optimize the model parameters θ again using the samples in the training dataset.

[0098] Step 7: Figure 5 shows an example of using the trained U-Net model on field data, compared with the widely used STA / LTA first-arrival picking algorithms. This intuitively demonstrates the effectiveness and excellence of the present invention in first-arrival picking tasks, and its ability to accurately and rapidly pick first-arrival locations in complex geological conditions and low signal-to-noise ratio environments.

[0099] Based on the same inventive concept, the present invention provides a seismic data first arrival picking system, comprising:

[0100] The acquisition module is used to acquire seismic data.

[0101] The data set generation module is used to use the automatic first arrival picking algorithm to preliminarily pick the first arrival time of the seismic signal in the seismic data to generate a meta-learning data set; the meta-learning data set includes a training sample set and a verification sample set.

[0102] The model generation module is used to design the corresponding deep neural network model based on the characteristics of the meta-learning dataset.

[0103] The training module is used for a double-loop training method based on meta-learning, which iteratively trains the deep neural network model using the meta-learning dataset. Each training iteration includes the following steps:

[0104] Sample N small batches of samples from the training sample set.

[0105] The meta-network weights Assign an independent sub-network.

[0106] Take M small batches of samples from the validation sample set.

[0107] The online gradient update method is used to calculate the sub-network loss value and backpropagate to optimize the weight matrix ω.

[0108] The weight matrix ω processed by using N small batches of training samples and the optimized weight matrix ω * , using the loss function to calculate the gradient and backpropagate to optimize the weights of the meta-network

[0109] The first arrival picking module is used to pick the first arrival of the seismic data to be detected using the trained deep neural network model.

[0110] Based on the same inventive concept, the present invention also proposes a computer device for first-arrival picking of seismic data, comprising: a memory, a processor, and a computer program stored in the memory. When the processor executes the computer program, the steps of the first-arrival picking method of seismic data are implemented.

[0111] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for picking the first arrivals of seismic data, characterized in that, It includes the following steps: Collect seismic data; Use an automatic first arrival picking algorithm to preliminarily pick the first arrival times of seismic signals in the seismic data and generate a meta-learning dataset; the meta-learning dataset includes a training sample set and a validation sample set; the use of the automatic first arrival picking algorithm to preliminarily pick the first arrival times of seismic signals in the seismic data and generate a meta-learning dataset specifically includes the following steps: Divide the sampling points before and after the first arrival time picking into two categories, namely before the first arrival and after the first arrival, and assign different labels respectively; regard the first arrival picking results with errors of the automatic first arrival picking algorithm as noise labels; regard the first arrival times revised manually as clean labels; generate a meta-learning dataset according to the noise labels and clean labels; Based on the double-loop training method of meta-learning, use the meta-learning dataset to iteratively train the deep neural network model; Input the seismic data to be detected into the trained deep neural network model and output the first arrival picking results of the seismic data to be detected; Among them, the double-loop training process based on meta-learning uses the meta-learning dataset to iteratively train the deep neural network. Each iterative training includes the following steps: Sample N small batches of samples from the training sample set; The meta-network weights Assign to an independent sub-network; Adopt M small batches of samples from the validation sample set; Use the online gradient update method to calculate the sub-network loss value and backpropagate to optimize the weight matrix ω; Using N small batches of training samples and the weight matrix ω processed by optimizing the weight matrix ω * , calculating the gradient using the loss function, and backpropagating to optimize the weights of the meta-network The adopted loss function is expressed as: Loss=ωf i (θ) where ω is the weight matrix, and f i (θ) represents the loss function associated with the noisy label i, where θ are the trainable parameters in the deep neural network model; The weight matrix ω processed using N small batches of training samples and the optimized weight matrix ω * , calculate the gradient using the loss function, and backpropagate to optimize the weights of the meta-network Specifically, it is expressed as: Among them, the symbol * represents the optimal value of the parameter, and f j represents the loss function associated with the clean label j.

2. The first arrival picking method of seismic data according to claim 1, characterized in that, After collecting the seismic data, perform filtering processing on the seismic data, and then perform cropping, normalization, and first arrival picking operations on the processed seismic data.

3. A method for picking the first arrival of seismic data according to claim 1, characterized in that, The input data of the deep neural network model is a seismic record. Each seismic record contains 64 channels of seismic data, each channel contains 512 sampling points, and the interval between each sampling point is 10 milliseconds; the output of the deep neural network model is a matrix of 0-1.

4. A method for picking the first arrival of seismic data according to claim 1, characterized in that, The deep neural network model adopts a U-Net model.

5. A method for picking the first arrival of seismic data according to claim 1, characterized in that, Using the online gradient update method, calculate the loss value of the sub-network, and backpropagate to optimize the weight matrix ω. The estimated value of the weight matrix ω is expressed as: Among them, η is the learning rate of ω, t represents the current iteration number, and m is the number of samples in M small batches of samples; For the estimated value Perform normalization processing, which is expressed as: Among them, 6 is the minimum value to prevent ω from all being zero, and the estimated value Is a positive number.

6. An initial arrival picking system for seismic data, characterized in that, Includes: A collection module for collecting seismic data; A dataset generation module for using an automatic first arrival picking algorithm to preliminarily pick the first arrival times of seismic signals in the seismic data and generate a meta-learning dataset; the meta-learning dataset includes a training sample set and a validation sample set; the use of the automatic first arrival picking algorithm to preliminarily pick the first arrival times of seismic signals in the seismic data and generate a meta-learning dataset specifically includes the following steps: Divide the sampling points before and after the first arrival time picking into two categories, namely before the first arrival and after the first arrival, and assign different labels respectively; regard the first arrival picking results with errors of the automatic first arrival picking algorithm as noise labels; regard the first arrival times revised manually as clean labels; generate a meta-learning dataset according to the noise labels and clean labels; A training module for iteratively training the deep neural network model using the meta-learning dataset based on the double-loop training method of meta-learning; among them, each iterative training includes the following steps: Sample N small batches of samples from the training sample set; The meta-network weights Assign to an independent sub-network; Adopt M small batches of samples from the validation sample set; Using the online gradient update method, calculate the sub-network loss value, and backpropagate to optimize the weight matrix ω; Using the training samples of N small batches and the weight matrix ω processed by optimizing the weight matrix ω … , calculating the gradient using the loss function, and backpropagating to optimize the weights of the meta-network The loss function adopted is expressed as: Loss=ωf i (θ) where ω is the weight matrix, and f i (θ) represents the loss function associated with the noisy label i, where θ are the trainable parameters in the deep neural network model; The weight matrix ω processed using N small batches of training samples and the optimized weight matrix ω * , calculate the gradient using the loss function, and backpropagate to optimize the weights of the meta-network Specifically, it is expressed as: where the symbol * represents the optimal value of the parameter, and f j represents the loss function associated with the clean label j; The first arrival picking module is used to input the seismic data to be detected into the trained deep neural network model and output the first arrival picking result of the seismic data to be detected.

7. A computer device for picking the first arrivals of seismic data, characterized in that, Including: A memory, a processor, and a computer program stored in the memory. When the processor executes the computer program, the steps of the first arrival picking method for seismic data according to any one of claims 1-5 are implemented.

Images