Deep learning-based DAS-VSP four-wave-field separation method and system
By constructing the RMTU-Net model based on the U-Net network, the problem of four-wavefield separation in DAS-VSP technology was solved, achieving high-precision and fast wavefield processing, and improving the accuracy and imaging effect of underground detection.
Patent Information
- Application Number
- CN202511908862.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-17
- Publication Date
- 2026-03-17
AI Technical Summary
In existing DAS-VSP technology, the complex mixing of P-waves, S-waves, and up and down waves leads to wavefield energy overlap and mode confusion. Traditional methods and existing deep learning methods are difficult to achieve synchronous and high-precision separation of the four wavefields, which affects the accuracy of underground detection.
A deep learning approach based on the U-Net network is adopted. By constructing the RMTU-Net model and utilizing skip connections and residual blocks, four-wave field separation of DAS-VSP data is achieved, including uplink and downlink separation of P-waves and S-waves. Iterative training is performed using a high-quality training set to generate the final wave field separation model.
It achieves high-precision separation of four wave fields, improves the accuracy and imaging quality of underground detection, reduces wave field residue and spurious frequency phenomena, and improves the efficiency and accuracy of separation.
Smart Images
Figure CN121679685A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data processing, and in particular to a deep learning-based DAS-VSP four-wavefield separation method and system. Background Technology
[0002] Currently, vertical seismic profiling (VSP) technology, especially DAS-VSP technology based on distributed fiber acoustic sensing (DAS), has been widely used in seismic exploration due to its advantages of high sampling rate, high resolution, and low cost. The wavefield information contained in DAS-VSP data, including downflow P-waves, downflow S-waves, upflow P-waves, and upflow S-waves, is crucial for reservoir characterization and subsurface structure interpretation. However, these wavefields intersect and interfere with each other in the records, forming complex couplings that pose a significant challenge to wavefield separation. Traditional wavefield separation methods, such as median filtering, FK filtering, Radon transform, and polarization filtering, typically utilize apparent velocity or polarization differences for separation, but are prone to mixing and aliasing. Although researchers have combined multiple traditional methods (such as the fusion of Radon transform and median filtering) to improve separation results, these methods still rely on manual parameter adjustments, resulting in significant human intervention and difficulty in avoiding wavefield residues and distortions. In recent years, deep learning has been introduced into DAS-VSP data processing, such as improved methods based on attention networks or U-Net. However, these methods mostly focus on separating uplink and downlink wave fields and have failed to solve the cross-coupling problem of P-waves and S-waves at the same time.
[0003] The DAS-VSP seismic record contains a complex mixture of P-waves, S-waves, and up- and down-going waves, leading to wavefield energy overlap and mode confusion. Existing deep learning methods often ignore the polarization and velocity differences between P-waves and S-waves, making it impossible to achieve simultaneous and high-precision separation of the four wavefields. This severely limits the accuracy of subsequent migration imaging, inversion, and interpretation.
[0004] Therefore, how to develop a separation method that can effectively separate four wave fields simultaneously to overcome complex coupling interference and improve the accuracy of underground detection is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] In view of this, the present invention provides a DAS-VSP four-wavefield separation method and system based on deep learning, which overcomes the above-mentioned defects.
[0006] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a deep learning-based DAS-VSP four-wavefield separation method, the specific steps of which are as follows: Acquire VSP data and separate the VSP data into multiple wavefield types; Multiple types of wave fields are superimposed to generate a hybrid wave field, and a training set is generated based on the hybrid wave field and the multiple types of wave fields. An initial wavefield separation model is constructed based on the U-Net network. The initial wavefield separation model is then iteratively trained using the training set to generate the final wavefield separation model. Acquire the VSP data to be processed, input the VSP data to be processed into the wave field separation model, and output the prediction results of various wave fields.
[0007] Optionally, the separation steps for the multiple types of wave fields are as follows: The VSP data is subjected to P-wave and S-wave field separation; The P-wave field and the S-wave field are separated into uplink and downlink waves using a filtering method to obtain the uplink P-wave field, the uplink S-wave field, the downlink P-wave field, and the downlink S-wave field.
[0008] Optionally, the steps for generating the training set are as follows: Multiple types of wave fields are superimposed to generate a hybrid wave field, and the hybrid wave field and multiple types of wave fields are divided into patch blocks of a preset size. After normalizing the patch blocks, the training set is constructed.
[0009] Optionally, the wave field separation model includes an input layer, an encoder, a decoder, and an output layer connected in sequence; the encoder and the decoder are connected in a skip connection. The input layer is used to receive input data; The encoder is used to perform feature extraction and feature downsampling operations on the input data and output encoded data; The decoder consists of multiple cascaded feature reconstruction modules, used to upsample and reconstruct features from the encoded data, and output reconstructed data. The output layer contains multiple prediction branches, and each prediction branch independently outputs a prediction result based on the reconstructed data.
[0010] Optionally, the encoder includes multiple cascaded coding layers, each coding layer including a superimposed feature extraction module and a residual unit.
[0011] Optionally, the feature extraction module includes a convolutional layer, a batch normalization layer, and a LeakyReLU activation function connected in sequence.
[0012] Optionally, the residual unit includes: a main path containing two convolutional layers and an auxiliary path containing one convolutional layer.
[0013] Optionally, the expression for the LeakyReLU activation function is: ; In the formula, The output value of the Leaky ReLU activation function; For input signals; It is the slope coefficient of the negative half-axis.
[0014] Optionally, the prediction branch generates a linear prediction result through a 1×1 convolution.
[0015] Secondly, this application provides a deep learning-based DAS-VSP four-wavefield separation system, comprising: The wave field separation module is used to acquire VSP data and separate the VSP data into multiple types of wave fields; The training set composition module is used to superimpose multiple types of wave fields to generate a hybrid wave field, and generate a training set based on the hybrid wave field and the multiple types of wave fields. The model training module is used to construct an initial wavefield separation model based on the U-Net network, and to iteratively train the initial wavefield separation model using the training set to generate the final wavefield separation model. The result output module is used to acquire the VSP data to be processed, input the VSP data to be processed into the wave field separation model, and output the prediction results of various wave fields.
[0016] According to the specific embodiments provided in this application, this application has the following technical effects: This application discloses a deep learning-based DAS-VSP four-wavefield separation method and system, which can achieve fast and efficient wavefield processing while ensuring high-precision separation. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a schematic diagram of the wavefield separation model in one embodiment of this application; Figure 2 This is a schematic diagram of the structure of a growth fault model in one embodiment of this application; Figure 3 This is a schematic diagram of synthesized VSP data in one embodiment of this application; Figure 4The following are schematic diagrams of the wavefield separation results of the wavefield separation model on the synthetic data in one embodiment of this application: (a) is a schematic diagram of the downlink P-wave (DP) wavefield separated from the synthetic data; (b) is a schematic diagram of the downlink S-wave (DS) wavefield separated from the synthetic data; (c) is a schematic diagram of the uplink P-wave (UP) wavefield separated from the synthetic data; and (d) is a schematic diagram of the uplink S-wave (US) wavefield separated from the synthetic data. Figure 5 The diagram below shows a comparison of the method provided in one embodiment of this application and the Kirchhoff migration imaging results of PP waves and PS waves without separated wavefields; (a) a schematic diagram of the P-wave velocity model; (b) a comparison diagram of the Kirchhoff migration imaging results of PP waves; and (c) a comparison diagram of the Kirchhoff migration imaging results of PS waves. Figure 6 This is a schematic diagram of actual DAS-VSP data for a single gun in one embodiment of this application; Figure 7 The following are schematic diagrams of wavefield separation results of actual data in one embodiment of this application: (a) is a schematic diagram of downlink P-wave (DP) wavefield separation of actual data; (b) is a schematic diagram of downlink S-wave (DS) wavefield separation of actual data; (c) is a schematic diagram of uplink P-wave (UP) wavefield separation of actual data; and (d) is a schematic diagram of uplink S-wave (US) wavefield separation of actual data. Figure 8 This is a schematic diagram of a method flow in one embodiment of this application. Detailed Implementation
[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0020] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0021] This embodiment discloses a deep learning-based DAS-VSP four-wavefield separation method, such as... Figure 8 As shown, the specific steps are as follows: Step 1: Acquire VSP data and separate the VSP data into multiple wavefield types; Step 2: Superimpose the multiple types of wave fields to generate a hybrid wave field, and generate a training set based on the hybrid wave field and the multiple types of wave fields; Step 3: Construct an initial wavefield separation model based on the U-Net network, and use the training set to iteratively train the initial wavefield separation model to generate the final wavefield separation model; Step 4: Obtain the VSP data to be processed, input the VSP data to be processed into the wave field separation model, and output the prediction results of various wave fields.
[0022] Furthermore, the processing flow disclosed in this embodiment is as follows: First, a high-quality DAS-VSP dataset is constructed, which uses the FK filtering method to further separate the P-wave and S-wave fields into uplink and downlink waves, obtaining uplink P-wave, uplink S-wave, downlink P-wave, and downlink S-wave as label data; then, the four types of wave fields are superimposed to form a new mixed wave field, which is used as training input. Second, a wave field separation model is constructed, specifically by designing a U-Net multi-task network based on residual blocks—RMTU-Net. Through encoding and decoding training of synthetic data (i.e., training samples), a mapping relationship between the total VSP wave field and the uplink, downlink, P-wave, and S-wave fields is established. This model can achieve four-wave field separation of the input wave field. Finally, the model is tested on synthetic data and actual DAS-VSP data.
[0023] Furthermore, the wavefield separation model employs the classic convolutional neural network U-Net as its backbone. The skip connections of U-Net can transfer high-resolution details between corresponding layers in the encoder and decoder, allowing the decoder to retain more image features during reconstruction. To further improve network performance, the standard convolutional blocks in the U-Net encoder are replaced with residual blocks from a residual network (Res-Net), enhancing feature extraction capabilities through identity mapping. Combining the advantages of both, a residual multi-task U-Network (RMTU-Net) based on residual blocks is finally constructed, with the structure as follows: Figure 1 As shown, it includes: a four-layer encoder on the left (Conv+ResBlock+Downsampling); a four-layer decoder on the right (UpConv+Skipconnection); a central bottleneck layer connecting the two ends; and then four parallel output branches, which output four types of wave fields respectively.
[0024] In one embodiment, the separation steps for the multiple types of wave fields are as follows: The VSP data is subjected to P-wave and S-wave field separation; The P-wave field and the S-wave field are separated into uplink and downlink waves using a filtering method to obtain the uplink P-wave field, the uplink S-wave field, the downlink P-wave field, and the downlink S-wave field.
[0025] In one embodiment, the step of generating the training set is as follows: Multiple types of wave fields are superimposed to generate a hybrid wave field, and the hybrid wave field and multiple types of wave fields are divided into patch blocks of a preset size. After normalizing the patch blocks, the training set is constructed.
[0026] Furthermore, RMTU-Net adopts a single-input, four-output architecture, which places higher demands on the quality of the training set. The quality of the training set directly determines the network's separation performance and overall performance. Therefore, in this embodiment, a high-quality open-source dataset is selected as the base data, and after a series of processing steps such as wavefield separation, the training set for this embodiment is constructed.
[0027] This dataset contains P-wave and S-wave wavefields, extracted using finite-difference elastic wave numerical simulation and Helmholtz decomposition. Subsequently, based on the separated wavefields, suitable traditional filtering methods were employed to separate the uplink and downlink waves. Among numerous traditional methods, this embodiment selected the FK filtering method as a traditional method for further separating the uplink and downlink waves, ensuring a balance of accuracy, efficiency, and stability in the separation results, providing a reliable wavefield foundation for subsequent deep learning model training and validation. Ultimately, four wavefield components—uplink P-wave, downlink P-wave, uplink S-wave, and downlink S-wave—were successfully extracted.
[0028] To ensure consistency in amplitude between input and output data, further processing of the input data is required. Specifically, the extracted four wavefield components are superimposed to generate a hybrid wavefield, which is then used to replace the original synthesized DAS-VSP wavefield as input data for network training.
[0029] Furthermore, the training set contains 240 data sets, each including the VSP uplink P-wave, downlink P-wave, uplink S-wave, downlink S-wave, and the superimposed total wavefield. Each data set is divided into segments of size [size missing]. The patches were generated with a step size of 32 in both the vertical and horizontal directions. This was to ensure that no data overlapped during training, resulting in 61,440 training patches. Given the significant differences in earthquake amplitude, all patch data underwent normalization. During training, RMTU-Net was used, with the total wavefield data as the network input and the four wavefields as different output components for training.
[0030] In one embodiment, the wave field separation model includes an input layer, an encoder, a decoder, and an output layer connected in sequence; the encoder and the decoder are connected in a skip connection. The input layer is used to receive input data; The encoder is used to perform feature extraction and feature downsampling operations on the input data and output encoded data; The decoder consists of multiple cascaded feature reconstruction modules, used to upsample and reconstruct features from the encoded data, and output reconstructed data. The output layer contains multiple prediction branches, each of which independently outputs a prediction result based on the reconstructed data.
[0031] In one embodiment, the encoder includes multiple cascaded coding layers, each coding layer including a superimposed feature extraction module and a residual unit.
[0032] In one embodiment, the feature extraction module includes a convolutional layer, a batch normalization layer, and a LeakyReLU activation function connected in sequence.
[0033] In one embodiment, the residual unit includes: a main path containing two convolutional layers and an auxiliary path containing one convolutional layer.
[0034] Furthermore, RMTU-Net uses a single-channel seismic data patch of size H×W×C as input, where H=W=32 and C=1. The network has four output branches with the same size as the input, corresponding to the prediction results of downlink P-wave, downlink P-wave, uplink P-wave, and uplink S-wave. The input tensor is represented by the symbol... The predictions for the four output branches are as follows: (1); In the formula, These represent the prediction results for the downward P-wave, downward S-wave, upward P-wave, and upward S-wave, respectively.
[0035] The convolutional neural network structure adopts an encoder-decoder architecture, which includes feature extraction path and feature reconstruction path, with a total of 34 trainable network layers; the input layer receives a 32×32×1 feature map.
[0036] The encoder contains a four-level feature extraction module. It first extracts initial features using a 3×3 convolutional layer (64 channels), followed by batch normalization and LeakyReLU. =0.01) activation to enhance nonlinear expressive power. Four residual units are then stacked sequentially, increasing the number of channels from 64 to 512. After each residual unit, feature downsampling is performed using convolutions with a stride of 2, gradually reducing the spatial resolution of the feature map to 16×16, 8×8, 4×4, and 2×2. The core pathway within each residual unit consists of two 3×3 convolutions, while the bypass pathways use 1×1 convolutions for dimension alignment to ensure consistency in information fusion. The decoding part consists of four feature reconstruction modules. The network recovers spatial resolution step-by-step through 3×3 deconvolution operations (stride=2), expanding from 2×2 to 32×32. Each upsampling stage is accompanied by batch normalization and LeakyReLU activation layers to stabilize gradient propagation and suppress training oscillations. Simultaneously, skip connections are established between corresponding layers in the decoding and encoding ends to fuse shallow spatial details with deep semantic features; the number of channels is adjusted using 1×1 convolutions to make the fused features more compact. The model's output layer has four branch prediction heads, each generating a linear output (corresponding to the four wavefield components) through a 1×1 convolution. Dropout (scale 0.15) is inserted between convolutional layers to prevent overfitting. The entire network uses the Adam optimizer for parameter updates, and the loss function is the mean squared error (MSE), which improves reconstruction accuracy while ensuring training stability.
[0037] In one embodiment, the convolutional neural network uses LeakyReLU as the activation function, which introduces a small non-zero slope in the negative region. This ensures that the output is a small negative value, guaranteeing that the gradient is still propagated during backpropagation, thus effectively mitigating the neuron death problem. The mathematical form of LeakyReLU can be expressed as: (2); In the formula, The output value of the Leaky ReLU activation function; For input signals; This is the negative half-axis slope coefficient (leak parameter), which is set to 0.01 in this embodiment.
[0038] Its advantages lie in the fact that LeakyReLU preserves non-zero gradients in the negative region, improving gradient backpropagation efficiency and enhancing the stability of deep structures. At the same time, it helps maintain the integrity of feature information, making it particularly suitable for deep encoder-decoder networks.
[0039] In one embodiment, MSE is selected as the loss function of the network model. MSE is used to measure the mean squared error between the model's predicted values and the true values, and its calculation formula is as follows: (3); In the formula, It is the sample size; It is the first The true value of each sample; The model is for the first The predicted value for each sample.
[0040] In one embodiment, the separation results are visualized and subjected to spectral analysis to evaluate wavefield continuity, amplitude preservation, and separation accuracy, ensuring that the separation results approximate the ideal wavefield.
[0041] This invention provides a deep learning-based DAS-VSP four-wavefield separation method that can achieve fast and efficient wavefield processing while ensuring high-precision separation.
[0042] In one embodiment, a simple velocity model was designed to test the model's performance, such as... Figure 2 As shown, the model has a depth of 1500 m and a width of 2000 m, describing a shallow growing fault model. The elastic parameters of the model gradually increase with depth, with a total of 6 gradient layers. Specific model parameters are shown in Table 1. Twenty seismic sources were placed from 0 m to 760 m on the model surface, with a fixed spacing of 40 m. A well was located at 1000 m on the model surface. Twelve geophones with a spacing of 10 m were placed in the well from 200 m to 1390 m. The maximum reception time was defined as 2 s, and the sampling interval was 1 ms.
[0043] Table 1 Parameters of the growth fault model
[0044] Based on the 2.5D elastic wave equation, a forward modeling of the velocity model was performed using a 15 Hz Ricker wavelet. Figure 2 VSP data of the three red sources in the middle are as follows Figure 3 As shown, each shot has 120 tracks, for a total of 360 tracks, which is different from the training set constructed above.
[0045] To more comprehensively and objectively evaluate the separation effect, the intelligent wavefield separation results of RMTU-Net were analyzed. The separation results are as follows: Figure 4 As shown. It can be seen that this method works in four types of wave fields (downward P-wave (…)). Figure 4 (a) and downward S-wave ( Figure 4 (b) and rising P wave ( Figure 4 (c) and rising S wave ( Figure 4The wavefield characteristics in both the middle (d) and lower (d) images are continuous and clear, with only slight descending first arrivals remaining at a few locations. The overall separation results are highly consistent with the ideal wavefield, with no obvious residues or artifacts, fully demonstrating the advantages of this method in terms of wavefield separation accuracy and integrity. In addition, this method supports parallel separation of multi-shot data, maintaining high accuracy while also achieving high computational efficiency, demonstrating excellent comprehensive performance.
[0046] Since the velocity model is known, the separation results were further compared using imaging to verify the effectiveness of different methods. Figure 5 The results of PP and PS wave imaging using Kirchhoff migration on the original VSP wavefield and the UP and US waves separated by the method of this embodiment are presented. The migration region is 300–1000 m laterally and 0–1500 m longitudinally. The P-wave velocity model of the imaging region is as follows. Figure 5 As shown in (a).
[0047] In PP wave imaging results (such as) Figure 5 As shown in (b), the shallow co-directional axis position in the unseparated VSP wave field is shallower; after processing by the method of this embodiment, the energy of the shallow co-directional axis is significantly enhanced. PS wave imaging results (as shown in...) Figure 5 As shown in (c), the pattern is similar to that of PP waves. This embodiment effectively corrects for unidirectional axes in the VSP that significantly deviate from geological interfaces, making the imaging results closer to the actual structural morphology. This indicates that wavefield separation has a significant effect on improving the quality of migration imaging. The ascending P-waves and S-waves separated by this method can accurately reconstruct the stratigraphic structure in migration imaging, with high positioning accuracy and clear imaging results.
[0048] In one embodiment, a set of actual DAS-VSP data was selected to verify the generalization ability of the method. This data contains 30 shots with a sampling interval of 2 ms. For ease of explanation, the non-zero offset data of the first shot is used as an example (e.g., Figure 6 (As shown). After automatic gain compensation (AGC) processing, complex uplink and downlink waves and P- and S-waves can be clearly identified. Meanwhile, real-world data inevitably contains noise interference, which provides a basis for verifying the robustness of the method. To maintain consistency, the same comparison method as the synthetic data test described above was introduced on this data. It should be noted that this well data was not used for model training in the deep learning method.
[0049] Separation test results on actual data are as follows Figure 7As shown in (a)-(d), the method provided in this embodiment achieves clear and continuous separation results in various wave fields. The slope of the same direction axis changes reasonably over time, and is highly consistent with the characteristics of the original wave field. The separation results have no obvious residuals or false frequencies, the wave field energy distribution is uniform, and the structure is complete. In actual data, it balances separation accuracy and wave field authenticity, and has excellent adaptability and stability.
[0050] This embodiment also discloses a deep learning-based DAS-VSP four-wavefield separation system, including: The wave field separation module is used to acquire VSP data and separate the VSP data into multiple types of wave fields; The training set composition module is used to superimpose multiple types of wave fields to generate a hybrid wave field, and generate a training set based on the hybrid wave field and the multiple types of wave fields. The model training module is used to construct an initial wavefield separation model based on the U-Net network, and to iteratively train the initial wavefield separation model using the training set to generate the final wavefield separation model. The result output module is used to acquire the VSP data to be processed, input the VSP data to be processed into the wave field separation model, and output the prediction results of various wave fields.
Claims
1. A DAS-VSP four-wave field separation method based on deep learning, characterized in that, The specific steps are: acquire VSP data, and separate the VSP data into multiple types of wave fields; superimpose the multiple types of wave fields to generate a mixed wave field, and generate a training set based on the mixed wave field and the multiple types of wave fields; construct an initial wave field separation model based on a U-Net network, iteratively train the initial wave field separation model using the training set, and generate a final wave field separation model; acquire VSP data, and separate the VSP data into multiple types of wave fields; 2. The DAS-VSP four-wavefield separation method based on deep learning according to claim 1, wherein, The separation steps of the multiple types of wave fields are: perform P-wave field and S-wave field separation on the VSP data; use a filtering method to separately perform upgoing and downgoing wave separation on the P-wave field and the S-wave field, and obtain upgoing P-wave field, upgoing S-wave field, downgoing P-wave field, and downgoing S-wave field.
3. The DAS-VSP four-wavefield separation method based on deep learning according to claim 1, characterized in that, The generation steps of the training set are: superimpose the multiple types of wave fields to generate a mixed wave field, and cut the mixed wave field and the multiple types of wave fields into patch blocks of a preset size, perform normalization processing on the patch blocks, and construct the training set.
4. The DAS-VSP four-wavefield separation method based on deep learning according to claim 1, characterized in that, The wave field separation model comprises an input layer, an encoder, a decoder, and an output layer connected in sequence; and the encoder and the decoder are jump-connected. The input layer is configured to receive input data. The encoder is configured to perform feature extraction and feature down-sampling operations on the input data, and output encoded data. The decoder is composed of multiple cascaded feature reconstruction modules, and is configured to perform up-sampling and feature reconstruction on the encoded data, and output reconstructed data. The output layer comprises multiple prediction branches, and each prediction branch independently outputs a prediction result according to the reconstructed data.
5. The DAS-VSP four-wavefield separation method based on deep learning according to claim 4, characterized in that, The encoder comprises multiple cascaded encoding layers, and each encoding layer comprises a feature extraction module and a residual unit arranged in superposition.
6. The DAS-VSP four-wavefield separation method based on deep learning according to claim 5, characterized in that, The feature extraction module comprises a convolution layer, a batch normalization layer, and a LeakyReLU activation function connected in sequence.
7. The DAS-VSP four-wavefield separation method based on deep learning according to claim 5, characterized in that, The residual unit comprises a main path comprising two convolution layers and an auxiliary path comprising one convolution layer.
8. The DAS-VSP four-wavefield separation method based on deep learning according to claim 6, characterized in that, The expression of the LeakyReLU activation function is: ； wherein is an output value of the Leaky ReLU activation function; is an input signal; is a negative half-axis slope coefficient.
9. The DAS-VSP four-wavefield separation method based on deep learning according to claim 4, characterized in that, The prediction branch generates a linear prediction result through 1x1 convolution.
10. A deep learning based DAS-VSP four-wavefield separation system, characterized in that, It comprises: a wave field separation module configured to acquire VSP data and separate the VSP data into multiple types of wave fields; a training set construction module configured to superimpose the multiple types of wave fields to generate a mixed wave field, and generate a training set based on the mixed wave field and the multiple types of wave fields; a model training module configured to construct an initial wave field separation model based on a U-Net network, iteratively train the initial wave field separation model using the training set, and generate a final wave field separation model; and a result output module configured to acquire VSP data, input the VSP data into the wave field separation model, and output prediction results of the multiple types of wave fields.