Intelligent first arrival pickup method and system

Abstract

The invention provides an intelligent first arrival pickup method and system, and belongs to the technical field of geophysical exploration. According to the method, energy similar attributes of seismic data and the seismic data are used as dual-channel data of an input layer, and first arrival is picked up by using a network model. The method comprises the following steps: (1) shot gather data preprocessing: reading seismic data and performing seismic trace supplementation processing; (2) preparing network input data: performing dynamic correction processing, seismic data normalization and energy similarity attribute calculation; (3) preparing label data: performing first-arrival travel time interpolation processing at a bad track and a missing track, and mapping from the first-arrival travel time to a three-dimensional data body; (4) network model training: training the 3DUnet network to obtain a trained network model; and (5) first arrival pickup: inputting the preprocessed seismic data, and obtaining first arrival travel time by using the trained network model. According to the method, the first arrival picking efficiency is high, the manual intervention degree is reduced, and the first arrival picking efficiency and precision are improved.

Description

Technical Field

[0001] The present invention belongs to the field of geophysical exploration technology, and specifically relates to an intelligent first-arrival picking method and system, which realizes first-arrival picking through shot gather data preprocessing technology oriented to first-arrival picking and a dual-channel 3DUnet network, and can be used for first-arrival picking processing in the preprocessing process of petroleum geophysical exploration. Background Art

[0002] First arrival picking is an important part of seismic data processing. The accuracy of first arrival picking directly affects the quality of static correction and imaging effect. Traditional first arrival picking methods mainly include energy ratio method, instantaneous intensity ratio method of seismic traces, wavelet amplitude ratio method, etc., which use the characteristics of waveform in energy, frequency and other aspects to identify first arrival, but the application effect is poor on seismic data with medium and low signal-to-noise ratio, and most of them are picked one by one, with low utilization of correlation between multi-channel data and poor lateral continuity. In addition, for different exploration areas, it is often necessary to select different first arrival picking methods suitable for the exploration area according to the actual situation of the exploration area, and manual parameter adjustment is required, with much manual intervention, time-consuming and labor-intensive first arrival picking of massive data, and low efficiency.

[0003] In recent years, deep learning technology has developed rapidly and has been widely used in the field of seismic data processing, showing great application prospects. If seismic data is regarded as an image, first-arrival picking can be regarded as an image segmentation problem. Under this assumption, first-arrival picking methods based on fully convolutional networks FCN, residual neural networks ResNet and U-type fully convolutional networks Unet have been proposed one after another, among which the first-arrival picking based on Unet has been studied the most. Unet splices shallow features and deep features through skip-layer connections, so that the output of the network combines the global features of the image and the local features of the texture. It is simple, efficient and has a small demand for training samples. Although there are many studies on first-arrival picking based on Unet, there are few studies on its application in actual data processing. Summary of the invention

[0004] The purpose of the present invention is to solve the above-mentioned difficulties in the prior art, to provide an intelligent first-break picking method and system, to use 3DUnet to realize intelligent first-break picking on shot gather data, to reduce the degree of manual intervention in first-break picking, to improve the efficiency and accuracy of first-break picking, to better serve the static correction in seismic data processing, and to provide strong technical support for shortening the production cycle and accelerating the seismic exploration process.

[0005] The present invention is achieved through the following technical solutions:

[0006] A first aspect of the present invention provides an intelligent first arrival picking method, which uses the energy similarity attribute of seismic data and the seismic data as dual-channel data of the input layer, and uses a network model to pick the first arrival.

[0007] Preferably, the method comprises:

[0008] (1) Shot gather data preprocessing: reading seismic data and completing seismic traces;

[0009] (2) Prepare network input data: dynamic correction processing, seismic data normalization and energy similarity attribute calculation;

[0010] (3) Prepare label data: interpolation of first arrival travel time at bad and missing tracks, and mapping of first arrival travel time to three-dimensional data volume;

[0011] (4) Network model training: Train the 3DUnet network to obtain a trained network model;

[0012] (5) First arrival picking: Input the preprocessed seismic data and use the trained network model to obtain the first arrival travel time.

[0013] Preferably, the operation of step (1) includes:

[0014] (11) Reading seismic data;

[0015] (12) dividing the data into multiple three-dimensional shot gather data volumes according to the shot number and the detection line number;

[0016] (13) Based on the maximum number of traces N set for each arrangement, seismic trace padding processing is performed on each three-dimensional shot gather data volume.

[0017] Preferably, the operation of step (2) includes:

[0018] (21) performing dynamic correction processing on each three-dimensional shot gather data volume using the shot offset and the dynamic correction speed to obtain shot gather data after dynamic correction;

[0019] (22) On the shot data after dynamic correction, the seismic data between time zero and the latest arrival time of the first arrival wave are intercepted as the data to be processed, and the data to be processed are normalized to between [-1, 1] as the first channel data of the input layer;

[0020] (23) Calculate the energy similarity attribute of the data to be processed as the second channel data of the input layer.

[0021] Preferably, the operation of step (3) includes:

[0022] Read the first arrival travel time, and use the first arrival travel time of the adjacent seismic traces for interpolation processing at the bad and empty traces;

[0023] The label data is formed according to the first arrival travel time of each seismic trace: each seismic trace is processed in turn as follows: the data with a time less than the first arrival travel time is assigned a value of 0, and the data with a time greater than or equal to the first arrival travel time is assigned a value of 1; after all seismic traces are processed, the first arrival travel times of all seismic traces are mapped into a three-dimensional data body that is either 0 or 1, namely the label data.

[0024] Preferably, the operation of step (4) includes: performing network training on the 3DUnet network using a training sample label set to obtain a trained network model; the training sample label set includes first channel data, second channel data and label data.

[0025] Preferably, the operation of step (5) includes:

[0026] (51) Processing the seismic data to be processed by the same method as steps (1) and (2) to obtain first channel data and second channel data;

[0027] (52) inputting the first channel data and the second channel data into the trained network model, and the trained network model outputs a predicted three-dimensional data volume;

[0028] (53) Map the predicted three-dimensional data volume back to the first arrival travel time data.

[0029] Preferably, the operation of step (53) includes:

[0030] The following processing is performed on each seismic trace in the predicted three-dimensional data volume in turn:

[0031] (531), the moment when the first value on the seismic trace is greater than 0.5 is taken as the first arrival travel time picked;

[0032] (532), add the dynamic correction value to the picked first arrival travel time to obtain the final picked first arrival travel time of the seismic trace.

[0033] A second aspect of the present invention provides an intelligent first arrival picking system, the system comprising:

[0034] Preprocessing unit, used for reading seismic data and completing seismic traces;

[0035] An input data preparation unit, connected to the preprocessing unit, is used for dynamic correction processing, seismic data normalization and energy similarity attribute calculation;

[0036] The label data preparation unit is connected to the preprocessing unit and the input data preparation unit respectively, and is used for interpolation processing of the first arrival travel time at the bad track and the missing track, and mapping of the first arrival travel time to the three-dimensional data volume;

[0037] The network model training unit is connected to the input data preparation unit and the label data preparation unit respectively, and is used to train the 3DUnet network to obtain a trained network model;

[0038] The first arrival picking unit is connected to the preprocessing unit, the input data preparation unit and the network model training unit respectively, and is used to input the preprocessed seismic data and obtain the first arrival travel time by using the trained network model.

[0039] The third aspect of the present invention provides a computer-readable storage medium, which stores at least one computer-executable program, and when the at least one program is executed by the computer, the computer executes the steps of an intelligent first-arrival picking method of the present invention.

[0040] Compared with the prior art, the beneficial effects of the present invention are: the first arrival picking efficiency is high using the method of the present invention, the degree of manual intervention is reduced, the first arrival picking efficiency and accuracy are improved, the stability is good, and a good application effect is shown under moderately complex geological conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flowchart of the steps of the method of the present invention;

[0042] Figure 2 A schematic diagram of the network structure used in the present invention;

[0043] Figure 3a The results of 3DUnet first arrival picking using the method of the present invention in a certain work area in the west are shown;

[0044] Figure 3b The first arrival picking results of a work area in the west of China were obtained using commercial software.

[0045] Figure 3c A local comparison of the picking results of the two methods. DETAILED DESCRIPTION

[0046] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0047] Aiming at the problem of first arrival picking in actual data, the present invention proposes a fast and efficient 3DUnet intelligent first arrival picking process. The present invention ensures data continuity by interpolating first arrivals on bad tracks and missing tracks, extracts energy similarity attributes of seismic data, and uses it and seismic data as dual-channel data of the input layer, and uses the network model to pick first arrivals, thereby improving the first arrival picking accuracy, and reducing the amount of calculation by performing linear dynamic correction on seismic data.

[0048] like Figure 2As shown, the 3DUnet network used for first arrival picking of the present invention (refer to the literature: 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation) consists of two parts, an encoder and a decoder, wherein the basic network structure of the encoder includes: convolution layer, batch normalization and maximum pooling layer, the decoder includes: upper convolution layer, batch normalization and convolution layer, the convolution kernel size is 3*3*3, except that the activation function of the output layer is sigmoid, the activation functions of other layers are ReLU, and there is a skip layer connection between the encoder and the decoder to realize the splicing of shallow features and deep features.

[0049] In actual data processing, the specific network structure (the number of convolutional layers and the number of channels in each convolutional layer, etc.) can be adjusted according to the gun data situation.

[0050] The loss function used in the present invention is the dice function, and its formula is as follows:

[0051]

[0052] Among them, Y pred is the network output value, Y label is the label, ε is the smoothing coefficient, which is set to 1.0 in the present invention. The Adam algorithm is used to minimize the loss function and continuously update the network model until the set number of iterations is reached or the loss function value tends to be stable, and the final trained 3DUnet network is obtained.

[0053] When applied to real data, data preprocessing and preparation of training sample label sets are important components of first arrival picking in 3DUnet network. The training sample label set consists of two parts: input and label. The input includes two channels: seismic data and energy similarity attributes. The label is a three-dimensional data volume mapped from the first arrival travel time.

[0054] In order to reduce the amount of calculation, linear dynamic correction is performed on the shot data. The dynamic correction calculation formula is as follows:

[0055]

[0056] Among them, x is the shot distance; v0 is the dynamic correction speed, which is a constant value.

[0057] Each seismic trace in the shot collection is moved by Δt along the time axis in the direction of decreasing time (moving the seismic trace in the direction of decreasing time means changing the time corresponding to each sampling point on the seismic trace) to obtain the shot collection data after dynamic correction. On the shot collection data after dynamic correction, the seismic data between the zero time and the latest arrival time of the first arrival wave are intercepted (a complete shot collection data contains not only the first arrival wave, but also the reflection wave, etc. Theoretically, the first arrival wave is the first wave to jump in the seismic trace. The goal of the present invention is to identify the first arrival, so the seismic data corresponding to the first arrival waveform can be intercepted as the network input), and it is normalized to between [-1,1] as the first channel data of the input layer, and the energy similarity attribute of this segment of data is calculated at the same time as the second channel data of the input layer. The label size is consistent with the input layer size ("size" refers to the number of sampling points in the three directions of x, y, and z of the data body), where the value above the first arrival is 0, and the value below the first arrival is 1.

[0058] In order to ensure the consistency of data size in the training sample label set, the maximum number of channels in each arrangement is set to N. For single-arrangement data with an actual channel number less than N, the seismic channels from right to left are used to fill them. In addition, in order to ensure the continuity of the first arrival travel time and eliminate outliers, the first arrival data at bad channels and empty channels are interpolated.

[0059] The method of the present invention comprises a series of technologies such as shot gather data preprocessing for first arrival picking, preparation of training sample label set, network model training and reasoning, so as to realize intelligent picking of first arrival on shot gather data.

[0060] like Figure 1 As shown, the method of the present invention comprises the following steps:

[0061] (1) Shot gather data preprocessing: reading seismic data and completing seismic traces;

[0062] (2) Prepare network input data: dynamic correction processing, seismic data normalization and energy similarity attribute calculation;

[0063] (3) Prepare label data: interpolation of first arrival travel time at bad and missing tracks, and mapping of first arrival travel time to three-dimensional data volume;

[0064] (4) Network model training: Use the sample label set to train the 3DUnet network and obtain a trained network model;

[0065] (5) First arrival picking: Input the preprocessed seismic data and use the trained network model to obtain the first arrival travel time.

[0066] Compared with the prior art, the present invention has the advantages of high first-arrival picking efficiency, low manual intervention and good stability.

[0067] The embodiments of the method of the present invention are as follows:

[0068] Example 1:

[0069] The operations in step (1) include:

[0070] (11) Read seismic data;

[0071] (12) According to the shot number and geophone line number, split the data into multiple 3D shot gather data volumes;

[0072] (13) According to the maximum number of traces N set for each array, perform seismic trace padding processing on each 3D shot gather data volume.

[0073] Among them, the operations in step (12) include:

[0074] Select all seismic traces with the same shot number to form a 3D shot gather data volume, where seismic traces with the same geophone line number are placed in the same row to form an array; each seismic trace in each array corresponds to a different trace number, and a 3D shot gather data volume is composed of multiple arrays arranged side by side. The three dimensions of this 3D shot gather data volume are geophone line number, trace number, and time respectively.

[0075] Among them, the operations in step (13) include:

[0076] Assume that the single-array data has a total of M traces. If M < N, then select N - M seismic traces from right to left in this array in sequence to perform padding processing on this array, so that the total number of seismic traces of the processed single-array data is N.

[0077] Example 2:

[0078] The operations in step (2) include:

[0079] (21) Use the offset and NMO velocity to perform NMO processing on each 3D shot gather data volume to obtain the NMO-processed shot gather data, specifically including:

[0080] (211), obtain the NMO correction amount Δt of each seismic trace using the following formula:

[0081]

[0082] Among them, x is the offset; v0 is the NMO velocity, which is a constant value and can be estimated by finding the slope of the first arrival wave on the single-shot record;

[0083] (212), move each seismic trace in the shot gather along the time axis in the direction of decreasing time by the NMO correction amount Δt corresponding to this seismic trace in sequence to obtain the NMO-processed shot gather data.

[0084] (22) On the shot data after dynamic correction, the seismic data between time zero and the latest arrival time of the first arrival wave are intercepted as the data to be processed, and the data to be processed are normalized to between [-1, 1] as the first channel data of the input layer, as follows:

[0085] The operations to normalize the data to be processed to [-1,1] include:

[0086] Use the following formula to normalize the data to be processed to between [-1,1]:

[0087]

[0088] Among them, X is the original seismic data, ||X|| ∞ is the infinite norm of the seismic data, X norm is the normalized seismic data.

[0089] (23) Calculate the energy similarity attribute of the data to be processed as the second channel data of the input layer, thus forming dual-channel data used as network input, as follows:

[0090] The energy similarity property is calculated using the following formula:

[0091]

[0092] Among them, X(i,j,l) is the seismic data at the position (i,j,l), W is the smoothing window size, and Y(i,j,k) is the energy similarity attribute at the position (i,j,k).

[0093] The energy similarity attribute describes the energy intensity variation relationship of seismic data, and the first arrival waveform corresponds to the position of strong energy. Introducing this attribute is more conducive to picking the first arrival than using only seismic data, and can improve the accuracy of first arrival picking.

[0094] The network input used for first-arrival picking only needs to contain the seismic data segment of the first-arrival wave. After linear dynamic correction of the seismic data, the number of longitudinal sampling points of the seismic data containing the first-arrival wave is reduced, which can reduce the number of convolution operations of each convolution layer in the network model, thereby reducing the amount of calculation.

[0095] Embodiment 3:

[0096] The operation of step (3) includes:

[0097] Read the first arrival travel time, and interpolate the bad track and empty track using the first arrival travel time of the adjacent seismic track; form label data based on the first arrival travel time of each seismic track.

[0098] The operations for forming label data based on the first arrival travel time of each seismic trace include:

[0099] Perform the following processing on each seismic trace in turn:

[0100] The data whose time is less than the first arrival time is assigned a value of 0, and the data whose time is greater than or equal to the first arrival time is assigned a value of 1.

[0101] The first arrival travel time here can be obtained through commercial software or manual picking.

[0102] After processing all seismic traces, the first arrival travel times of all seismic traces are mapped into a three-dimensional data volume that is either 0 or 1, namely, label data.

[0103] Embodiment 4:

[0104] The operation of step (4) includes: based on the 3DUnet network, using the training sample label set, setting the learning rate, using the Adam algorithm to perform network training, and saving the trained network model. The training sample label set includes the first channel data, the second channel data and the label data.

[0105] Embodiment 5:

[0106] The operation of step (5) includes:

[0107] (51) Processing the seismic data to be processed by the same method as steps (1) and (2) to obtain first channel data and second channel data;

[0108] (52) inputting the first channel data and the second channel data into the trained network model, and the trained network model outputs a predicted three-dimensional data volume;

[0109] (53) Map the predicted three-dimensional data volume back to the first arrival travel time data.

[0110] The operation of step (53) includes:

[0111] The following processing is performed on each seismic trace in the predicted three-dimensional data volume in turn:

[0112] (531), the moment when the first value on the seismic trace is greater than 0.5 is taken as the first arrival travel time picked;

[0113] (532), the picked first arrival travel time is added with the dynamic correction amount to obtain the final picked first arrival travel time of the seismic trace, that is, the first arrival travel time picked by 3DUnet for each seismic trace is added with the dynamic correction amount of the seismic trace to obtain the final picked first arrival travel time.

[0114] The second aspect of the present invention provides an intelligent first arrival picking system, and the embodiments of the system are as follows:

[0115] Embodiment 6:

[0116] The system comprises:

[0117] Preprocessing unit, used for reading seismic data and completing seismic traces;

[0118] An input data preparation unit, connected to the preprocessing unit, is used for dynamic correction processing, seismic data normalization and energy similarity attribute calculation;

[0119] The label data preparation unit is connected to the preprocessing unit and the input data preparation unit respectively, and is used for interpolation processing of the first arrival travel time at the bad track and the missing track, and mapping of the first arrival travel time to the three-dimensional data volume;

[0120] The network model training unit is connected to the input data preparation unit and the label data preparation unit respectively, and is used to train the 3DUnet network to obtain a trained network model;

[0121] The first arrival picking unit is connected to the preprocessing unit, the input data preparation unit and the network model training unit respectively, and is used to input the preprocessed seismic data and obtain the first arrival travel time by using the trained network model.

[0122] The intelligent first arrival picking method proposed in the present invention is applied to the shot gathering data of a certain work area in western my country. The 100 shot gathering data of the work area and their first arrival travel time are used to make a training sample label set, and the network structure designed in the present invention is trained. The training cycle is set to 200. After the network training is completed, the other shot gathering data of the work area are preprocessed (linear dynamic correction, seismic trace filling, seismic data normalization and energy similarity attribute calculation), and the preprocessed 3D data volume and energy similarity attribute are input into the network model to obtain the picked first arrival (such as Figure 3a ) and the results picked up by commercial software (such as Figure 3b The first arrival continuity of the present invention based on 3DUnet network is better ( Figure 3c The left picture in the figure is the first arrival travel time result picked by the method of the present invention, and the right picture is the result picked by commercial software. The black line is the first arrival travel time picked. Figure 3c As shown by the black line in the figure, the results picked up by the commercial software are disturbed by noise, and the picked up first arrival travel time suddenly decreases, which does not conform to the propagation law of seismic wave travel time. The first arrival picking method based on the 3DUnet network of the present invention can make full use of the seismic wave information of adjacent channels, has a certain noise resistance, and the picked up first arrival travel time changes more smoothly and has better continuity). The picking results at far offset distances are also more accurate, which verifies the effectiveness and accuracy of the present invention.

[0123] The present invention realizes first-break picking through shot gather data preprocessing oriented to first-break picking and a dual-channel 3DUnet network, and can be used for first-break picking processing in the preprocessing process of petroleum geophysical exploration.

[0124] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "connected" and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0125] In the description of the present invention, unless otherwise specified, the terms "upper", "lower", "left", "right", "inside", "outside", etc. indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, they cannot be understood as limitations on the present invention.

[0126] The above technical solution is only one implementation mode of the present invention. For those skilled in the art, it is easy to make various types of improvements or modifications based on the principles disclosed in the present invention, and it is not limited to the technical solution described in the above specific embodiments of the present invention. Therefore, the above description is only preferred and not restrictive.

Claims

1. An intelligent first arrival picking method, characterized in that: The method uses the energy similarity attribute of seismic data and the seismic data as dual-channel data of the input layer, and uses a network model to pick up the first arrival.

2. The intelligent first arrival picking method according to claim 1, characterized in that: The method comprises: (1) Shot gather data preprocessing: reading seismic data and completing seismic traces; (2) Prepare network input data: dynamic correction processing, seismic data normalization and energy similarity attribute calculation; (3) Prepare label data: interpolation of first arrival travel time at bad and missing tracks, and mapping of first arrival travel time to three-dimensional data volume; (4) Network model training: Train the 3DUnet network to obtain a trained network model; (5) First arrival picking: Input the preprocessed seismic data and use the trained network model to obtain the first arrival travel time.

3. The intelligent first arrival picking method according to claim 2, characterized in that: The operation of step (1) includes: (11) Reading seismic data; (12) dividing the data into multiple three-dimensional shot gather data volumes according to the shot number and the detection line number; (13) Based on the maximum number of traces N set for each arrangement, seismic trace padding processing is performed on each three-dimensional shot gather data volume.

4. The intelligent first arrival picking method according to claim 2, characterized in that: The operation of step (2) includes: (21) performing dynamic correction processing on each three-dimensional shot gather data volume using the shot offset and the dynamic correction speed to obtain shot gather data after dynamic correction; (22) On the shot data after dynamic correction, the seismic data between time zero and the latest arrival time of the first arrival wave are intercepted as the data to be processed, and the data to be processed are normalized to between [-1, 1] as the first channel data of the input layer; (23) Calculate the energy similarity attribute of the data to be processed as the second channel data of the input layer.

5. The intelligent first arrival picking method according to claim 2, characterized in that: The operation of step (3) includes: Read the first arrival travel time, and use the first arrival travel time of the adjacent seismic traces for interpolation processing at the bad and empty traces; The label data is formed according to the first arrival travel time of each seismic trace: each seismic trace is processed in turn as follows: the data with a time less than the first arrival travel time is assigned a value of 0, and the data with a time greater than or equal to the first arrival travel time is assigned a value of 1; after all seismic traces are processed, the first arrival travel times of all seismic traces are mapped into a three-dimensional data body that is either 0 or 1, namely the label data.

6. The intelligent first arrival picking method according to claim 2, characterized in that: The operation of step (4) includes: using the training sample label set to perform network training on the 3DUnet network to obtain a trained network model; the training sample label set includes first channel data, second channel data and label data.

7. The intelligent first arrival picking method according to claim 2, characterized in that: The operation of step (5) includes: (51) Processing the seismic data to be processed by the same method as steps (1) and (2) to obtain first channel data and second channel data; (52) inputting the first channel data and the second channel data into the trained network model, and the trained network model outputs a predicted three-dimensional data volume; (53) Map the predicted three-dimensional data volume back to the first arrival travel time data.

8. The intelligent first arrival picking method according to claim 7, characterized in that: The operation of step (53) includes: The following processing is performed on each seismic trace in the predicted three-dimensional data volume in turn: (531), the moment when the first value on the seismic trace is greater than 0.5 is taken as the first arrival travel time picked; (532), add the dynamic correction value to the picked first arrival travel time to obtain the final picked first arrival travel time of the seismic trace.

9. An intelligent first arrival picking system, characterized in that: The system comprises: Preprocessing unit, used for reading seismic data and completing seismic traces; An input data preparation unit, connected to the preprocessing unit, is used for dynamic correction processing, seismic data normalization and energy similarity attribute calculation; The label data preparation unit is connected to the preprocessing unit and the input data preparation unit respectively, and is used for interpolation processing of the first arrival travel time at the bad track and the missing track, and mapping of the first arrival travel time to the three-dimensional data volume; The network model training unit is connected to the input data preparation unit and the label data preparation unit respectively, and is used to train the 3DUnet network to obtain a trained network model; The first arrival picking unit is connected to the preprocessing unit, the input data preparation unit and the network model training unit respectively, and is used to input the preprocessed seismic data and obtain the first arrival travel time by using the trained network model.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one computer-executable program, and when the at least one program is executed by the computer, the computer executes the steps in the intelligent first-arrival picking method described in any one of claims 1-8.

Images