Deep learning dataset making method and apparatus for first arrival picking

Abstract

The application discloses a deep learning data set manufacturing method and device for first arrival picking. The application adopts a first arrival picking data set generation method based on a STA / LTA algorithm, generates a data set through an automatic processing method, strengthens first arrival features in training data and label data, improves data set manufacturing efficiency, and has practical value for efficiently picking first arrival data of seismic exploration by applying a convolutional neural network.

Description

Technical Field

[0001] This invention belongs to the field of seismic exploration data processing, and more specifically, relates to a method and apparatus for creating deep learning datasets for first arrival picking. Background Technology

[0002] During seismic exploration, after seismic waves are generated at the shot point, geophones placed on or beneath the surface begin to receive the signals. The first valid seismic wave received by the geophone is called the "first arrival wave." First arrival wave acquisition is a fundamental and crucial task in seismic data processing. As the exploration terrain becomes increasingly complex, the waveform of the first arrival wave varies considerably, and various waves interfere with each other, rendering existing first arrival wave acquisition algorithms ineffective. Often, a human-computer interactive approach is used, relying on manually added auxiliary lines to acquire the first arrival wave. However, with massive amounts of seismic data, the workload of first arrival wave acquisition is becoming increasingly enormous, severely limiting the speed of data processing.

[0003] With the deepening application of the rapidly developing deep learning technology in the field of seismic exploration, significant progress has been made in the research of first arrival picking using convolutional neural networks. In the preparation of datasets, the first arrival picking problem is usually treated as a binary classification problem, and the first arrival and the "background" other than the first arrival are distinguished by different categories. When creating training and label data using Segy seismic data, multiple seismic data are usually selected as a set of training data, and these seismic data are manually labeled as a set of label data. This method of taking the whole data cannot focus on the features before and after the arrival of the first arrival, and the generated training and label data are large in size, which increases the consumption of computational and memory resources during the training process and reduces training efficiency. Summary of the Invention

[0004] In view of this, the present disclosure provides a method, apparatus, electronic device and medium for creating a deep learning dataset for first arrival picking, which can generate datasets through automated processing, reduce manpower consumption and improve dataset production efficiency, and has practical value for the efficient picking of first arrival data in seismic exploration using convolutional neural networks.

[0005] In a first aspect, embodiments of this disclosure provide a method for creating a deep learning dataset for initial picking, including:

[0006] Step 1: Preprocess the raw seismic data to obtain the original seismic profile grayscale map based on the preprocessed seismic data.

[0007] Step 2: The STA / LTA algorithm is used to process the raw seismic data to obtain the predicted first arrival position;

[0008] Step 3: Generate a grayscale image of the label data based on the predicted initial arrival position;

[0009] Step 4: Using the predicted initial arrival position as a reference, cut multiple original data blocks from the original seismic profile grayscale image as training sample data for the convolutional neural network.

[0010] Step 5: Using the predicted initial arrival position as a reference, crop multiple label data blocks from the grayscale image of the label data to serve as the label sample data for the convolutional neural network.

[0011] In some implementations, step 1 specifically includes:

[0012] Read each seismic trace from the raw seismic data; each seismic trace includes multiple sample point data.

[0013] For each seismic trace read, normalize all sample data of that seismic trace based on the maximum and minimum values ​​among all sample data of that seismic trace.

[0014] Multiply all sample data points of all seismic traces after normalization by 255 to perform grayscale processing, and obtain the original seismic profile grayscale map.

[0015] In some implementations, step 3 specifically includes:

[0016] A matrix is ​​constructed with the number of seismic traces in the original seismic data as columns and the number of sample data points in each seismic trace as rows.

[0017] In the matrix, for each column of data, all data above the predicted initial arrival position are set to 1, and all data below the predicted initial arrival position are set to 0.

[0018] Multiply all data in the matrix by 255 to generate the grayscale image of the label data.

[0019] In some implementations, step 4 specifically includes:

[0020] Using the preset first arrival position of a specified seismic trace in the original seismic profile grayscale image as a reference, a predetermined number of seismic traces are selected to the right from the specified seismic trace, and the same length is cut upwards and downwards from the predicted first arrival position of the specified seismic trace to form a time window centered on the predicted first arrival position. The selected predetermined number of seismic traces are cropped according to the time window to generate the training sample data.

[0021] In some implementations, step 5 specifically includes:

[0022] Using the preset first arrival position of the specified seismic trace in the grayscale image of the label data as a reference, a predetermined number of seismic traces are selected to the right from the specified seismic trace, and the same length is cut upwards and downwards from the predicted first arrival position of the specified seismic trace to form a time window centered on the predicted first arrival position. The selected predetermined number of seismic traces are cropped according to the time window to generate the label sample data.

[0023] Secondly, embodiments of this disclosure also provide an electronic device, the electronic device comprising:

[0024] Memory, which stores executable instructions;

[0025] A processor that executes the executable instructions in the memory to implement the method for creating a deep learning dataset for initial picking.

[0026] Thirdly, embodiments of this disclosure also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for creating a deep learning dataset for initial picking.

[0027] Fourthly, embodiments of this disclosure also provide an apparatus for creating deep learning datasets for initial picking, comprising:

[0028] The original grayscale image generation unit is used to preprocess the original seismic data and obtain the original seismic profile grayscale image based on the preprocessed seismic data.

[0029] The first arrival location prediction unit is used to calculate the predicted first arrival location from the raw seismic data using the STA / LTA algorithm;

[0030] A label grayscale image generation unit is used to generate a label data grayscale image based on the predicted initial arrival position;

[0031] The training sample cropping unit is used to crop multiple original data blocks from the original seismic profile grayscale image based on the predicted initial arrival position, as training sample data for the convolutional neural network.

[0032] The label sample cropping unit is used to crop multiple label data blocks from the grayscale image of the label data based on the predicted initial arrival position, as the label sample data of the convolutional neural network.

[0033] In some implementations, the training sample pruning unit is specifically used for:

[0034] Using the preset first arrival position of a specified seismic trace in the original seismic profile grayscale image as a reference, a predetermined number of seismic traces are selected to the right from the specified seismic trace, and the same length is cut upwards and downwards from the predicted first arrival position of the specified seismic trace to form a time window centered on the predicted first arrival position. The selected predetermined number of seismic traces are cropped according to the time window to generate the training sample data.

[0035] In some implementations, the label sample cropping unit is specifically used for:

[0036] Using the preset first arrival position of the specified seismic trace in the grayscale image of the label data as a reference, a predetermined number of seismic traces are selected to the right from the specified seismic trace, and the same length is cut upwards and downwards from the predicted first arrival position of the specified seismic trace to form a time window centered on the predicted first arrival position. The selected predetermined number of seismic traces are cropped according to the time window to generate the label sample data.

[0037] To optimize the process of creating first-arrival picking deep learning datasets, this application proposes a first-arrival picking dataset generation method based on the STA / LTA algorithm. The dataset is generated through automated processing, which enhances the first-arrival features in the training and label data, thereby improving the efficiency of dataset creation. This method has practical value for the efficient picking of first-arrival data in seismic exploration using convolutional neural networks.

[0038] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0039] The above and other objects, features and advantages of the present invention will become more apparent from the more detailed description of exemplary embodiments of the invention in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments of the invention.

[0040] Figure 1 A flowchart illustrating a method for creating a deep learning dataset for initial picking according to an embodiment of this application is shown.

[0041] Figure 2 A schematic diagram of the original seismic profile grayscale image is shown according to an exemplary embodiment of the present invention.

[0042] Figure 3 A comparison diagram of predicted first arrival position and manually picked first arrival position according to an exemplary embodiment of the present invention is shown.

[0043] Figure 4 A schematic diagram of a grayscale image of label data according to an exemplary embodiment of this application is shown.

[0044] Figure 5 (a) and (b) respectively illustrate schematic diagrams of training sample images and label sample images cropped according to an exemplary embodiment of this application. Detailed Implementation

[0045] Preferred embodiments of the invention will now be described in more detail. While preferred embodiments of the invention are described below, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0046] Figure 1 A flowchart of a method for creating a deep learning dataset for initial picking according to an embodiment of this application is shown. The method includes steps 1 to 5.

[0047] Step 1: Preprocess the raw seismic data to obtain the raw seismic profile grayscale map based on the preprocessed seismic data.

[0048] In some implementations, step 1 specifically includes:

[0049] Read each seismic trace from the raw seismic data; each seismic trace includes multiple sample point data.

[0050] For each seismic trace read, normalize all sample data of that seismic trace based on the maximum and minimum values ​​among all sample data of that seismic trace.

[0051] Multiply all sample data points of all seismic traces after normalization by 255 to perform grayscale processing, and obtain the original seismic profile grayscale map.

[0052] Convolutional neural networks (CNNs) have achieved great success in image semantic segmentation and edge detection. To apply CNNs to first-arrival picking, the raw seismic data first needs to undergo image preprocessing, transforming the first-arrival picking problem into an image segmentation and edge detection problem. Raw seismic data, i.e., conventional seismic profiles, is typically formed by arranging amplitude curves from seismic trace samples at certain intervals. According to this application, the raw seismic data can be preprocessed image-wise. First, the sample data for each trace can be normalized. Then, the normalized sample data is multiplied by 255 and processed into grayscale, converting the raw sample data into grayscale data, thus forming the raw seismic profile grayscale image. Subsequent training data is generated from the seismic profile grayscale image. The grayscale seismic data can be saved in CSV format.

[0053] Figure 2 A schematic diagram of the original seismic profile grayscale image is shown according to an exemplary embodiment of the present invention.

[0054] Back Figure 1Step 2: The STA / LTA algorithm is used to process the original seismic data to obtain the predicted first arrival position.

[0055] The STA / LTA algorithm is used to predict the initial arrival position, which significantly reduces manual labor and improves the efficiency of generating training sets compared to the existing method of manually picking the initial arrival position.

[0056] Figure 3 A comparison diagram of the predicted first arrival position and the manually picked first arrival position according to an exemplary embodiment of the present invention is shown. The curve with fewer spikes is the manually picked curve, and the curve with more spikes is the predicted curve obtained by the STA / LTA algorithm. It can be seen that the two have a good overlap and basically coincide.

[0057] Back Figure 1 Step 3: Generate a grayscale image of label data based on the predicted initial arrival position.

[0058] In some possible implementations, step 3 specifically includes:

[0059] A matrix is ​​constructed with the number of seismic traces in the original seismic data as columns and the number of sample data points in each seismic trace as rows.

[0060] In the matrix, for each column of data, all data above the predicted initial arrival position are set to 1, and all data below the predicted initial arrival position are set to 0.

[0061] Multiply all data in the matrix by 255 to generate the grayscale image of the label data.

[0062] Each column of the matrix corresponds to a seismic trace, and the data points in the column correspond to the sample points of that seismic trace. Using the predicted first arrival position as the boundary, values ​​above the first arrival position are assigned 1, and values ​​below the first arrival position are assigned 0. The entire matrix data is then multiplied by 255 for grayscale processing, converting the label sample point data into grayscale data, thus obtaining the grayscale image of the first arrival wave label data. Figure 4 A schematic diagram of a grayscale image of label data according to an exemplary embodiment of this application is shown. The label data subsequently used for training is also generated from this grayscale image of label data. The grayscale label matrix data can also be saved in CSV format.

[0063] Back Figure 1 Step 4: Using the predicted initial arrival position as a reference, cut multiple original data blocks from the original seismic profile grayscale image as training sample data for the convolutional neural network.

[0064] In some possible implementations, step 4 specifically includes:

[0065] Using the preset first arrival position of a specified seismic trace in the original seismic profile grayscale image as a reference, a predetermined number of seismic traces are selected to the right from the specified seismic trace, and the same length is cut upwards and downwards from the predicted first arrival position of the specified seismic trace to form a time window centered on the predicted first arrival position. The selected predetermined number of seismic traces are cropped according to the time window to generate the training sample data.

[0066] Figure 5 (a) shows a schematic diagram of training sample data cropped according to an exemplary embodiment of this application.

[0067] Those skilled in the art can set the cutting intervals and sizes as needed.

[0068] Back Figure 1 Step 5: Based on the predicted initial arrival position, cut multiple label data blocks from the grayscale image of the label data to serve as the label sample data of the convolutional neural network.

[0069] In some possible implementations, step 5 specifically includes:

[0070] Using the preset first arrival position of the specified seismic trace in the grayscale image of the label data as a reference, a predetermined number of seismic traces are selected to the right from the specified seismic trace, and the same length is cut upwards and downwards from the predicted first arrival position of the specified seismic trace to form a time window centered on the predicted first arrival position. The selected predetermined number of seismic traces are cropped according to the time window to generate the label sample data.

[0071] Figure 5 (b) shows a schematic diagram of label sample data cropped according to an exemplary embodiment of this application.

[0072] Those skilled in the art will understand that the position, step size, and size of the cropped label sample data are consistent with the position, step size, and size of the cropped training sample data.

[0073] In the above embodiments, a first arrival picking dataset generation method based on the STA / LTA algorithm is adopted. The dataset is generated through automated processing, which enhances the first arrival features in the training data and label data, improves the efficiency of dataset production, and has practical value for the efficient picking of first arrival data in seismic exploration using convolutional neural networks.

[0074] An electronic device according to an embodiment of the present disclosure includes a memory and a processor.

[0075] This memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0076] The processor may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device to perform desired functions. In one embodiment of this disclosure, the processor is used to execute computer-readable instructions stored in the memory.

[0077] Those skilled in the art will understand that, in order to solve the technical problem of how to achieve a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this disclosure.

[0078] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0079] This disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for creating a deep learning dataset for initial picking.

[0080] A computer-readable storage medium according to embodiments of the present disclosure stores non-transitory computer-readable instructions. When these non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the methods described in the foregoing embodiments of the present disclosure are performed.

[0081] The aforementioned computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or portable hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge).

[0082] According to an embodiment of this application, a deep learning dataset creation apparatus for initial arrival picking is also disclosed, including an original grayscale image generation unit, an initial arrival position prediction unit, a label grayscale image generation unit, a training sample cropping unit, and a label sample cropping unit.

[0083] The original grayscale image generation unit is used to preprocess the original seismic data and obtain the original seismic profile grayscale image based on the preprocessed seismic data.

[0084] The first arrival location prediction unit is used to calculate the predicted first arrival location from the raw seismic data using the STA / LTA algorithm;

[0085] A label grayscale image generation unit is used to generate a label data grayscale image based on the predicted initial arrival position;

[0086] The training sample cropping unit is used to crop multiple original data blocks from the original seismic profile grayscale image based on the predicted initial arrival position, as training sample data for the convolutional neural network.

[0087] The label sample cropping unit is used to crop multiple label data blocks from the grayscale image of the label data based on the predicted initial arrival position, as the label sample data of the convolutional neural network.

[0088] In some implementations, the training sample pruning unit is specifically used for:

[0089] Using the preset first arrival position of a specified seismic trace in the original seismic profile grayscale image as a reference, a predetermined number of seismic traces are selected to the right from the specified seismic trace, and the same length is cut upwards and downwards from the predicted first arrival position of the specified seismic trace to form a time window centered on the predicted first arrival position. The selected predetermined number of seismic traces are cropped according to the time window to generate the training sample data.

[0090] In some implementations, the label sample cropping unit is specifically used for:

[0091] Using the preset first arrival position of the specified seismic trace in the grayscale image of the label data as a reference, a predetermined number of seismic traces are selected to the right from the specified seismic trace, and the same length is cut upwards and downwards from the predicted first arrival position of the specified seismic trace to form a time window centered on the predicted first arrival position. The selected predetermined number of seismic traces are cropped according to the time window to generate the label sample data.

[0092] For further details and advantages of this embodiment, please refer to the relevant description above.

[0093] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A method for creating a deep learning dataset for initial picking, characterized in that, include: Step 1: Preprocess the raw seismic data to obtain the original seismic profile grayscale map based on the preprocessed seismic data. Step 2: The STA / LTA algorithm is used to process the raw seismic data to obtain the predicted first arrival position; Step 3: Generate a grayscale image of the label data based on the predicted initial arrival position; Step 4: Using the predicted initial arrival position as a reference, cut multiple original data blocks from the original seismic profile grayscale image to serve as training sample data for the convolutional neural network. Specifically, this includes: using the predicted initial arrival position of a specified seismic trace in the original seismic profile grayscale image as a reference, selecting a predetermined number of seismic traces to the right from the specified seismic trace, and taking the predicted initial arrival position of the specified seismic trace as the center, cutting them upwards and downwards by the same length to form a time window centered on the predicted initial arrival position. The selected predetermined number of seismic traces are then cut according to this time window to generate the training sample data. Step 5: Using the predicted initial arrival position as a reference, crop multiple label data blocks from the grayscale image of the label data to serve as label sample data for the convolutional neural network. Specifically, this includes: using the predicted initial arrival position of a specified seismic trace in the grayscale image of the label data as a reference, selecting a predetermined number of seismic traces to the right from the specified seismic trace, and using the predicted initial arrival position of the specified seismic trace as the center, cropping upwards and downwards by the same length to form a time window centered on the predicted initial arrival position. Cropping the selected predetermined number of seismic traces according to this time window generates the label sample data.

2. The method according to claim 1, characterized in that, Step 1 specifically includes: Read each seismic trace from the raw seismic data; each seismic trace includes multiple sample point data. For each seismic trace read, normalize all sample data of that seismic trace based on the maximum and minimum values ​​among all sample data of that seismic trace. Multiply all sample data points of all seismic traces after normalization by 255 to perform grayscale processing, and obtain the original seismic profile grayscale map.

3. The method according to claim 1, characterized in that, Step 3 specifically includes: A matrix is ​​constructed with the number of seismic traces in the original seismic data as columns and the number of sample data points in each seismic trace as rows. In the matrix, for each column of data, all data above the predicted initial arrival position are set to 1, and all data below the predicted initial arrival position are set to 0. Multiply all data in the matrix by 255 to generate the grayscale image of the label data.

4. An electronic device, characterized in that, The electronic device includes: Memory, which stores executable instructions; A processor that executes the executable instructions in the memory to implement the method for creating a deep learning dataset for initial picking as described in any one of claims 1-3.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method for creating a deep learning dataset for initial picking as described in any one of claims 1-3.

6. A device for creating deep learning datasets for initial picking, characterized in that, include: The original grayscale image generation unit is used to preprocess the original seismic data and obtain the original seismic profile grayscale image based on the preprocessed seismic data. The first arrival location prediction unit is used to calculate the predicted first arrival location from the raw seismic data using the STA / LTA algorithm; A label grayscale image generation unit is used to generate a label data grayscale image based on the predicted initial arrival position; The training sample pruning unit is used to prune multiple original data blocks from the original seismic profile grayscale image based on the predicted first arrival position, as training sample data for the convolutional neural network. Specifically, the training sample pruning unit is used to: select a predetermined number of seismic traces to the right from the predicted first arrival position of a specified seismic trace in the original seismic profile grayscale image, and prune the same length upwards and downwards from the predicted first arrival position of the specified seismic trace as the center, forming a time window centered on the predicted first arrival position. The selected predetermined number of seismic traces are pruned according to the time window to generate the training sample data. The label sample cropping unit is used to crop multiple label data blocks from the grayscale image of the label data, based on the predicted initial arrival position, as the label sample data of the convolutional neural network. Specifically, the label sample cropping unit is used to: select a predetermined number of seismic traces to the right from the predicted initial arrival position of a specified seismic trace in the grayscale image of the label data, and crop the same length upwards and downwards from the predicted initial arrival position of the specified seismic trace as the center, forming a time window centered on the predicted initial arrival position. The selected predetermined number of seismic traces are cropped according to the time window to generate the label sample data.

Images