First arrival picking method and device based on deep learning network

Abstract

The present invention discloses a first arrival picking method and device based on a deep learning network, the method comprising: collecting a density display image of original seismic shot gather data under preset display parameters; cutting the density display image into a plurality of first arrival sample images of preset sizes; inputting the plurality of first arrival sample images of preset sizes into a pre-trained deep network model, outputting a plurality of first arrival prediction result images; synthesizing the plurality of first arrival prediction result images to generate a first arrival picking result image corresponding to the original seismic shot gather data; extracting the first arrival time from the first arrival picking result image. The first arrival picking accuracy of the present invention is high and is not affected by noise.

Description

Technical Field

[0001] The present invention relates to the technical field of oil and gas exploration, and in particular to a first arrival picking method and device based on a deep learning network. Background Art

[0002] At present, high-efficiency and high-density acquisition is gradually being applied in geophysical exploration, and the exploration target area is gradually shifting to complex surface areas. Due to the large changes in the surface medium, the acoustic wave velocity changes greatly, resulting in distortion of the first arrival of a single shot, serious layer stringing, difficulty in picking up the first arrival, and poor accuracy of picking up the first arrival. At the same time, due to the large number of receiving channels, the amount of data for a single shot becomes large, and a lot of manpower and time are required for picking up the first arrival. The quality of the first arrival picking directly affects the structural morphology and imaging quality of the stacked section. Therefore, it is very important to pick up the first arrival accurately.

[0003] There are many methods for first-arrival picking in the prior art, for example, time window seismic attribute feature method, correlation method, digital image processing method, etc. The above methods have a certain suppression effect on noise, but when the noise of the seismic record is serious, it becomes very difficult to accurately pick the first arrival, and the accuracy of first arrival picking is low. Summary of the invention

[0004] The embodiment of the present invention provides a first arrival picking method based on a deep learning network, which has a high first arrival picking accuracy and is not affected by noise. The method includes:

[0005] Acquire density display images of raw seismic shot gather data under preset display parameters;

[0006] Cutting the density display image into a plurality of first-arrival sample images of preset sizes;

[0007] Input multiple first-arrival sample images of preset sizes into a pre-trained deep network model, and output multiple first-arrival prediction result images;

[0008] Synthesize multiple first arrival prediction result images to generate first arrival picking result images corresponding to the original seismic shot gather data;

[0009] Extracting the first arrival time from the first arrival picking result image;

[0010] Before acquiring the density display image of the original seismic shot gather data under the preset display parameters, the method further comprises:

[0011] Acquire seismic shot gather sample data and first arrival label data under preset display parameters, wherein the first arrival label data is used to identify the first arrival wave position in the seismic shot gather sample data;

[0012] Training a deep network model using the seismic shot gather sample data and the first arrival label data;

[0013] Also includes:

[0014] Under the same display parameters, the first arrival information is used to generate the first arrival label data. The first arrival information is saved as a two-dimensional point coordinate vector, and each point consists of two parts: the channel number and the first arrival depth time information;

[0015] The deep network model is trained using the seismic shot gather sample data and the first arrival label data, including:

[0016] The output of the excitation discriminant layer of the deep network model is used to discriminate the label results and detect the error between the generated error and the initial target.

[0017] The embodiment of the present invention further provides a first arrival picking device based on a deep learning network, which has a high first arrival picking accuracy and is not affected by noise. The device comprises:

[0018] An image acquisition module, used for acquiring a density display image of the original seismic shot gather data under preset display parameters;

[0019] An image cutting module, used for cutting the density display image into a plurality of initial arrival sample images of preset sizes;

[0020] A first arrival prediction result image output module is used to input a plurality of first arrival sample images of preset sizes into a pre-trained deep network model and output a plurality of first arrival prediction result images;

[0021] A first arrival prediction result image synthesis module is used to synthesize multiple first arrival prediction result images to generate a first arrival picking result image corresponding to the original seismic shot gather data;

[0022] A first arrival time extraction module is used to extract the first arrival time from the first arrival picking result image;

[0023] Wherein, the device further comprises:

[0024] A data acquisition module, used to acquire seismic shot gather sample data and first arrival label data under preset display parameters, wherein the first arrival label data is used to identify the first arrival wave position in the seismic shot gather sample data;

[0025] A deep network model training module, used to train a deep network model using the seismic shot gather sample data and the first arrival label data;

[0026] The data acquisition module is also used to:

[0027] Under the same display parameters, the first arrival information is used to generate the first arrival label data. The first arrival information is saved as a two-dimensional point coordinate vector, and each point consists of two parts: the channel number and the first arrival depth time information;

[0028] The deep network model training module is specifically used for:

[0029] The output of the excitation discriminant layer of the deep network model is used to discriminate the label results and detect the error between the generated error and the initial target.

[0030] An embodiment of the present invention further provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above method when executing the computer program.

[0031] An embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program implements the above method when executed by a processor.

[0032] In an embodiment of the present invention, a density display image of original seismic shot gather data under preset display parameters is collected, the density display image is cut into a plurality of first arrival sample images of preset sizes, and the plurality of first arrival sample images of preset sizes are input into a pre-trained deep network model, a plurality of first arrival prediction result images are output, and then the plurality of first arrival prediction result images are synthesized to generate a first arrival picking result image corresponding to the original seismic shot gather data, and the first arrival time is extracted from the first arrival picking result image, and the first arrival picking accuracy is high and is not affected by noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0034] Figure 1 Flow chart of a first arrival picking method based on a deep learning network in an embodiment of the present invention;

[0035] Figure 2 This is a flow chart of training a deep network model in an embodiment of the present invention;

[0036] Figure 3 Schematic diagram of the structure of a first arrival picking device based on a deep learning network in an embodiment of the present invention;

[0037] Figure 4 A schematic diagram of a structure for training a deep network model in an embodiment of the present invention;

[0038] Figure 5 is a schematic diagram of a computer device according to an embodiment of the present invention;

[0039] Figure 6 This is a specific example diagram of deep network model training in an embodiment of the present invention;

[0040] Figure 7 This is a specific example diagram of first arrival picking using a deep network model according to an embodiment of the present invention;

[0041] Figure 8 This is an effect diagram of selecting 5 groups of similar small data to perform 10 iterations of training on a small depth grid model in an embodiment of the present invention, and performing first arrival picking on the test data after the training is completed;

[0042] Fig. 9 This is an effect diagram of selecting 10 groups of similar small data to train a small deep grid model 30 times in an embodiment of the present invention, and performing first arrival picking on the test data after the training is completed;

[0043] Fig.10 In the embodiment of the present invention, Fig. 9 The trained deep mesh model Figure 8 The effect diagram of first arrival picking of the test data in;

[0044] Fig.11 This is an effect diagram of selecting 5 shot data to train a small neural grid model 20 times in an embodiment of the present invention, and performing first arrival picking on the test data after the training is completed;

[0045] Fig.12 This is an effect diagram of selecting 5 shot data to perform 50 trainings on a small depth grid model in an embodiment of the present invention, and performing first arrival picking on the test data after the training is completed. DETAILED DESCRIPTION

[0046] To make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings. Here, the exemplary embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0047] Figure 1 A flowchart of a first arrival picking method based on a deep learning network provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the method includes:

[0048] Step 101: Acquire a density display image of original seismic shot gather data under preset display parameters.

[0049] Step 102: cutting the density display image into a plurality of first-arrival sample images of preset sizes.

[0050] In the specific implementation, the density display image of the original shot data under the specified display parameters is saved, and then it is cut into small images of size X×Y, and these first-arrival sample image sets are used as input data for subsequent deep network model prediction.

[0051] Step 103: Input a plurality of first-arrival sample images of preset sizes into a pre-trained deep network model, and output a plurality of first-arrival prediction result images.

[0052] In the specific implementation, the trained deep network model is loaded, including setting the deep network model structure and parameters, etc. The loaded trained deep network model is used to predict the input first arrival sample image to obtain the corresponding first arrival detection data set, which is composed of first arrival detection result images corresponding to the input small first arrival sample images. The numerical value in the first arrival detection result image represents the predicted probability value of the position being identified as the first arrival wave, and its value range is 0 to 1.

[0053] Step 104: synthesize the multiple first-arrival prediction result images to generate a first-arrival picking result image corresponding to the original seismic shot gather data.

[0054] Based on step 103, during specific implementation, the first arrival prediction result images can be combined into a large first arrival result image corresponding to the original shot gather data according to their numbers and positions.

[0055] Step 105: extract the first arrival time from the first arrival picking result image.

[0056] In an embodiment, extracting the first arrival time from the first arrival picking result image includes:

[0057] The first arrival picking result data smaller than a preset value is deleted from the first arrival picking result image to extract the first arrival time.

[0058] In the specific implementation, points with predicted probability values ​​less than a certain threshold are deleted from the first arrival picking result image, and points with very small predicted probability values ​​are considered not to be first arrival wave points. Then, the pure first arrival position information is extracted from the first arrival picking result image, that is, the maximum value of each column that is not 0 is found column by column in the matrix, and the y coordinate of the point is recorded. After that, the actual first arrival time corresponding to each track is calculated from the first arrival position in the first arrival picking result image using information such as the starting position of the starting track in the image and the horizontal and vertical display resolution, and the results are exported for use by other software.

[0059] The first-arrival picking method based on deep learning network provided by the embodiment of the present invention collects the density display image of the original seismic shot gather data under preset display parameters, cuts the density display image into multiple first-arrival sample images of preset sizes, inputs the multiple first-arrival sample images of preset sizes into a pre-trained deep network model, outputs multiple first-arrival prediction result images, and then synthesizes the multiple first-arrival prediction result images to generate a first-arrival picking result image corresponding to the original seismic shot gather data, extracts the first-arrival time from the first-arrival picking result image, and has a high first-arrival picking accuracy and is not affected by noise.

[0060] Figure 2 This is a flow chart of training a deep network model according to an embodiment of the present invention, such as Figure 2 As shown, before acquiring the density display image of the original seismic shot gather data under the preset display parameters, the method further includes:

[0061] Step 201: Acquire seismic shot gather sample data and first arrival label data under preset display parameters, wherein the first arrival label data is used to identify the first arrival wave position in the seismic shot gather sample data.

[0062] In an embodiment, the seismic shot gather sample data are samples provided to the deep network model for learning. By learning the content and context of these seismic shot gather sample data, the deep network model can abstract and extract the first arrival features and location information from different levels.

[0063] In order to facilitate the computer and deep network model to learn more easily, the present invention uses the density image of seismic data as the data source of seismic shot collection samples, so that the first arrival features in such image data are more prominent, more continuous, easier to learn and identify for the computer. Before generating sample data, it is necessary to set the data display parameters, display resolution, horizontal channel number display interval, vertical time display interval and other parameters. These parameters are mainly used for the actual positioning of the sampling time from the image to the seismic channel. At the same time, it is also necessary to generate the corresponding first arrival sample data at this parameter, and the display parameters of the data source and sample data need to be consistent.

[0064] For the first arrival label data, the first arrival information can be used to generate the first arrival label data under the same display parameters. The first arrival is the position where the seismic wave first arrives. For a piece of seismic data, there is at most one point to mark the position of the first arrival wave, or there may be none. Therefore, the first arrival information is usually saved as a two-dimensional point coordinate vector, and each point consists of two parts: the channel number and the first arrival depth time information. The method adopted by the present invention is an end-to-end training method, so the image composed of the seismic shot set sample data needs to be consistent in size with the image composed of the first arrival label data, and the position where the first arrival wave arrives is represented by 1, and the other parts are represented by 0.

[0065] Step 202: Train a deep network model using the seismic shot gather sample data and the first arrival label data.

[0066] Furthermore, in the target data of deep learning, the first arrival is a curve, which is a very small part relative to the entire seismic data. In order to narrow the data range, lock the learning target, and improve the learning efficiency, the seismic shot sample data and the first arrival label data can be cut and screened. Specifically, for example, first, the seismic shot sample data and the first arrival label data can be cut into several small matrices of size X×Y, and the insufficient data are supplemented with 0, and a total of R×C small matrices are divided. The small first arrival attribute matrix and the small first arrival label matrix after cutting correspond to each other. Then, based on the small first arrival label matrix after segmentation, the number of first arrival data in each small first arrival matrix is ​​counted respectively. If the number of first arrivals contained is less than c, the small first arrival label data and its corresponding small first arrival attribute data are deleted.

[0067] In addition, since the first arrival data corresponds to the surface topography, when the terrain tilts, the corresponding first arrival data also tilts. In order to better utilize the limited training data selected to simulate the complexity and diversity of the first arrival data caused by the diversity of surface morphology, it is necessary to perform some small-angle rotation transformations (including left and right flipping, etc.) on the selected data (the first arrival attribute data and its corresponding first arrival label data) to simulate the first arrival changes caused by the changes in the surface topography to achieve the purpose of expanding the training samples. After the data expansion training, the generated first arrival attribute data and first arrival mark data are placed in the specified directory respectively, and then the paired first arrival attribute data and first arrival mark data are read in turn, normalized respectively, and loaded into the deep network model for training.

[0068] In the embodiment of the present invention, the deep network model used is a deep network model of an encoder-decoder structure, which is mainly composed of two parts. The first part is the feature extraction and extraction part of the first arrival, which is used to capture the data context information for first arrival feature extraction, and the second part is the first arrival position restoration part. Specifically:

[0069] The first part is a feature abstraction network composed of multiple first-arrival feature extraction layers of decreasing size, which is used to extract and compress the first-arrival features. Each first-arrival feature extraction layer is composed of a convolution layer, a pooling layer, and a data normalization layer. The output of the previous layer is the input of the next layer.

[0070] The second part is the network part used to solve the problem of first arrival wave position restoration and high-precision positioning. This part consists of multiple layers with increasing sizes. Each layer is composed of a deconvolution layer, a data merging layer, a convolution layer, and a data normalization layer. The output of one layer is used as input, and the pre-output of the same scale layer of the encoding part is added to improve the accuracy of first arrival feature position recovery. Finally, the output of the excitation discriminant layer is used to discriminate with the label result, and the error between the generated error and the first arrival target is detected, so as to reflect and correct the weight.

[0071] Furthermore, a binary classification method can be used to detect and distinguish first arrivals, and the loss function of the entire deep network model is calculated using a logarithmic loss function, namely, binary cross entropy. The cross entropy loss function can be used to measure the similarity between a given first arrival and the first arrival learned by the neural network.

[0072] Among them, the loss function used by the deep network model is the following cross entropy:

[0073]

[0074] Where y is the expected output, a is the actual output of the neural network, a = σ(z) where z = ∑W j *X j +b where σ is the neuron activation function, W j is the neuron weight, X j is the neuron input and b is a constant.

[0075] In the specific implementation, after defining the deep network model, the normal distribution initialization method is used to initialize the weights of each neuron. Conventional deep learning optimization algorithms such as Adam are used as learning adjustment algorithms for deep network models. The weights of each neuron node are adjusted through continuous iteration and training. These methods have the characteristics of fast learning speed and strong anti-interference ability.

[0076] The present invention is described below with a specific embodiment:

[0077] First, set the display parameters and save the seismic data into the first-arrival sample image data in the specified form and format. Generate the corresponding first-arrival label image using the given first-arrival information, whose display parameters and size are consistent with the first-arrival sample image. Cut the first-arrival sample data and first-arrival label data according to the specified size, and remove the data according to the number of first-arrival identifiers in the first-arrival label data, and delete the first-arrival sample data and first-arrival label data that do not contain first-arrival identifiers or have very few first-arrival identifiers. Perform operations such as small-angle rotation or left-right flipping on the paired first-arrival sample data and first-arrival label data to simulate terrain changes to expand the training samples.

[0078] Secondly, an end-to-end deep network model structure is used for first-arrival learning and training. The deep network model consists of a first-arrival feature extraction part and a first-arrival position recovery part. The first-arrival feature extraction part extracts the first-arrival features layer by layer. The first-arrival position recovery part recovers the first-arrival features and their positions through layer-by-layer deconvolution. At the same time, in order to improve the first-arrival positioning accuracy, the output results of the same-size network layer of the first-arrival feature extraction part are directly superimposed, while maintaining the abstract feature information and high-precision position information of the first arrival. For a specific example diagram of deep network model training, please refer to Figure 6 .

[0079] Finally, the seismic data to be detected is saved as a first-arrival sample image of a specified form and format according to the specified display parameters, and then the first-arrival sample image is cut according to the specified size to form a set of small-sized first-arrival sample data. Define and load the trained first-arrival detection network model, and then use this model to predict the generated first-arrival sample image to obtain a set of corresponding first-arrival prediction result data, and then merge these prediction results to obtain a large first-arrival result data that matches the first-arrival sample image, delete the small first-arrival prediction values ​​in the prediction results, and then find the largest non-zero value column by column, use this position as the first-arrival wave position to extract the first-arrival information, and use the display parameters to reversely calculate the first-arrival time of each channel. For a specific example diagram of first-arrival picking using a deep network model, see Figure 7 .

[0080] Specifically, for example, Figure 8 To select 5 groups of similar small data, a small deep grid model is trained for 10 iterations. After the training is completed, the effect of first arrival picking on the test data is shown. For high signal-to-noise ratio data, the current method can effectively pick it. Fig. 9 In order to select 10 groups of similar small data, a small deep grid model is trained 30 times. After the training is completed, the effect diagram of the first arrival picking of the test data is shown. Fig.10 To use Fig. 9 The trained deep mesh model Figure 8 The effect diagram of first arrival picking based on the test data in . Fig.11 To select 5 shot data, a small neural grid model is trained 20 times. After the training is completed, the effect of first arrival picking is performed on the test data. Fig.12 In order to select 5 shot data, a small depth grid model is trained 50 times respectively. After the training is completed, the effect diagram of the first arrival picking of the test data is shown.

[0081] Based on the same inventive concept, a first-arrival picking device based on a deep learning network is also provided in an embodiment of the present invention, as described in the following embodiments. Since the principle of solving the problem by the first-arrival picking device based on a deep learning network is similar to the first-arrival picking method based on a deep learning network, the implementation of the first-arrival picking device based on a deep learning network can refer to the implementation of the first-arrival picking method based on a deep learning network, and the repetitions will not be repeated. As used below, the term "unit" or "module" can implement a combination of software and / or hardware for a predetermined function. Although the device described in the following embodiments is preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceived.

[0082] Figure 3 A schematic diagram of the structure of a first arrival picking device based on a deep learning network provided by an embodiment of the present invention, such as Figure 3 As shown, the device comprises:

[0083] An image acquisition module 301 is used to acquire a density display image of the original seismic shot gather data under preset display parameters;

[0084] An image cutting module 302, used for cutting the density display image into a plurality of first-arrival sample images of a preset size;

[0085] A first arrival prediction result image output module 303 is used to input a plurality of first arrival sample images of a preset size into a pre-trained deep network model, and output a plurality of first arrival prediction result images;

[0086] A first arrival prediction result image synthesis module 304 is used to synthesize a plurality of first arrival prediction result images to generate a first arrival picking result image corresponding to the original seismic shot gather data;

[0087] The first arrival time extraction module 305 is used to extract the first arrival time from the first arrival picking result image.

[0088] Figure 4 Schematic diagram of the structure of training a deep network model according to an embodiment of the present invention. Figure 4 As shown, the device also includes:

[0089] The data acquisition module 401 is used to acquire seismic shot gather sample data and first arrival label data under preset display parameters, wherein the first arrival label data is used to identify the first arrival wave position in the seismic shot gather sample data;

[0090] The deep network model training module 402 is used to train the deep network model using the seismic shot gather sample data and the first arrival label data.

[0091] In this embodiment of the present invention, the loss function used by the deep network model is the following cross entropy:

[0092]

[0093] Where y is the expected output, a is the actual output of the neural network, a = σ(z) where z = ∑W j *X j +b where σ is the neuron activation function, W j is the neuron weight, X j is the neuron input and b is a constant.

[0094] In the embodiment of the present invention, the first arrival time extraction module 305 is further used for:

[0095] The first arrival picking result data smaller than a preset value is deleted from the first arrival picking result image to extract the first arrival time.

[0096] In order to achieve the above object, according to another aspect of the present application, a computer device is also provided. Figure 5 As shown, the computer device includes a memory, a processor, a communication interface and a communication bus. A computer program that can be executed on the processor is stored in the memory. When the processor executes the computer program, the steps in the above-mentioned embodiment method are implemented.

[0097] The processor may be a central processing unit (CPU). The processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or a combination of the above chips.

[0098] The memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer executable programs and units, such as the corresponding program units in the above method embodiments of the present invention. The processor executes various functional applications of the processor and works data processing by running the non-transitory software programs, instructions and modules stored in the memory, that is, implementing the method in the above method embodiments.

[0099] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created by the processor, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0100] The one or more units are stored in the memory, and when executed by the processor, the method in the above embodiment is performed.

[0101] An embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program for executing the above method.

[0102] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0103] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0104] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1A function specified in one or more boxes.

[0105] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0106] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A first arrival picking method based on deep learning network, characterized in that: include: Acquire density display images of raw seismic shot gather data under preset display parameters; Cutting the density display image into a plurality of first-arrival sample images of preset sizes; Input multiple first-arrival sample images of preset sizes into a pre-trained deep network model, and output multiple first-arrival prediction result images; Synthesize multiple first arrival prediction result images to generate first arrival picking result images corresponding to the original seismic shot gather data; Extracting the first arrival time from the first arrival picking result image; Before acquiring the density display image of the original seismic shot gather data under the preset display parameters, the method further comprises: Acquire seismic shot gather sample data and first arrival label data under preset display parameters, wherein the first arrival label data is used to identify the first arrival wave position in the seismic shot gather sample data; Training a deep network model using the seismic shot gather sample data and the first arrival label data; Also includes: Under the same display parameters, the first arrival information is used to generate the first arrival label data. The first arrival information is saved as a two-dimensional point coordinate vector, and each point consists of two parts: the channel number and the first arrival depth time information; The deep network model is trained using the seismic shot gather sample data and the first arrival label data, including: The output of the excitation discriminant layer of the deep network model is used to discriminate the label results and detect the error between the generated error and the initial target.

2. The method according to claim 1, characterized in that The loss function used by the deep network model is the following cross entropy: Where y is the expected output, a is the actual output of the neural network, a = σ(z) where z = ∑W j *X j +b where σ is the neuron activation function, W j is the neuron weight, X j is the neuron input and b is a constant.

3. The method according to claim 1, characterized in that Extract the first arrival time from the first arrival picking result image, including: The first arrival picking result data smaller than a preset value is deleted from the first arrival picking result image to extract the first arrival time.

4. A first arrival picking device based on a deep learning network, characterized in that: include: An image acquisition module, used for acquiring a density display image of the original seismic shot gather data under preset display parameters; An image cutting module, used for cutting the density display image into a plurality of initial arrival sample images of preset sizes; A first arrival prediction result image output module is used to input a plurality of first arrival sample images of preset sizes into a pre-trained deep network model and output a plurality of first arrival prediction result images; A first arrival prediction result image synthesis module is used to synthesize multiple first arrival prediction result images to generate a first arrival picking result image corresponding to the original seismic shot gather data; A first arrival time extraction module is used to extract the first arrival time from the first arrival picking result image; Wherein, the device further comprises: A data acquisition module, used to acquire seismic shot gather sample data and first arrival label data under preset display parameters, wherein the first arrival label data is used to identify the first arrival wave position in the seismic shot gather sample data; A deep network model training module, used to train a deep network model using the seismic shot gather sample data and the first arrival label data; The data acquisition module is also used to: Under the same display parameters, the first arrival information is used to generate the first arrival label data. The first arrival information is saved as a two-dimensional point coordinate vector, and each point consists of two parts: the channel number and the first arrival depth time information; The deep network model training module is specifically used for: The output of the excitation discriminant layer of the deep network model is used to discriminate the label results and detect the error between the generated error and the initial target.

5. The device according to claim 4, characterized in that The loss function used by the deep network model is the following cross entropy: Where y is the expected output, a is the actual output of the neural network, a = σ(z) where z = ∑W j *X j +b where σ is the neuron activation function, W j is the neuron weight, X j is the neuron input and b is a constant.

6. The device according to claim 4, characterized in that The first arrival time extraction module is further used to: The first arrival picking result data smaller than a preset value is deleted from the first arrival picking result image to extract the first arrival time.

7. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 3 is implemented.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 3 is implemented.

Images