A low signal-to-noise ratio first arrival wave data preprocessing method, application, device and medium based on deep learning

Abstract

The application belongs to the technical field of seismic data processing, and specifically discloses a low signal-to-noise ratio first arrival wave data preprocessing method based on deep learning, application, equipment and medium. The method disclosed by the application comprises the following steps: obtaining forward single-shot data of existing data; adding noise to the forward single-shot data to obtain noise single-shot data; performing curvelet transform on the forward single-shot data and the noise single-shot data to generate curvelet coefficient subgraphs respectively, and using the same to train a deep network model; inputting the curvelet-transformed single-shot data to be preprocessed into the deep network model for detection, converting the same into curvelet coefficients, and then performing inverse curvelet transform. The application effectively removes the noise in the first arrival wave data by combining the deep learning technology and the curvelet transform, and improves the quality of the low signal-to-noise ratio first arrival wave data. The application can be widely applied to the preprocessing of low signal-to-noise ratio first arrival wave data.

Description

Technical Field

[0001] This invention belongs to the field of seismic data processing technology, specifically relating to a method, application, equipment, and medium for preprocessing low signal-to-noise ratio first arrival wave data based on deep learning. Background Technology

[0002] In the current field of oil and gas seismic exploration, rapid urbanization and industrialization have brought unprecedented challenges to exploration work. As exploration areas gradually expand to the outskirts of towns and industrial zones, noise interference increases significantly. These interference sources severely affect the quality of seismic data, resulting in a generally low signal-to-noise ratio (SNR) for single-shot data. This problem is particularly pronounced in "dual-complex" areas (i.e., areas with extremely complex surface and subsurface geological structures). Complex surface lithology and its spatial variations make the propagation path of seismic waves exceptionally complex, often resulting in chaotic and disordered first-arrival fields that are difficult to clearly identify. At the same time, the complexity of subsurface geological structures further exacerbates the scattering and reflection of seismic waves, causing the first-arrival signal to be submerged in noise, with violent amplitude fluctuations and extremely low first-arrival identifiability. This low SNR, violent amplitude fluctuation, and low identifiability of first-arrival data makes efficient and accurate first-arrival acquisition extremely difficult.

[0003] Traditional first-arrival acquisition methods in the industry, such as the instantaneous feature method, neural network method, and edge detection method, while performing reasonably well in conventional exploration areas, have poor applicability and low accuracy in "double-complex" areas. The instantaneous feature method identifies first-arrivals by analyzing the instantaneous properties of seismic waves, but under low signal-to-noise ratio conditions, instantaneous features are often difficult to extract accurately. The neural network method automatically identifies first-arrivals using machine learning algorithms, but its training requires a large amount of high-quality data, and its performance in complex areas is often unsatisfactory. The edge detection method identifies first-arrivals by detecting the edges of seismic waveforms, but this method is sensitive to noise and prone to false positives and false negatives. Therefore, these traditional methods are ineffective in complex areas, usually requiring significant manual intervention and correction, which is time-consuming and labor-intensive, and cannot meet the needs of modern efficient exploration.

[0004] The accuracy of first-arrival pickup directly impacts the precision and accuracy of subsequent seismic data processing tasks, such as near-surface velocity modeling and static correction. Therefore, a new method is needed to improve the quality of first-arrival pickup and address noise issues in complex scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based preprocessing method for low signal-to-noise ratio first-arrival data, which improves the quality of low signal-to-noise ratio first-arrival data by combining deep learning technology with curve transform. The second objective of this invention is to provide an application of denoised single-shot data obtained based on the above method in first arrival quality control, surface velocity modeling, static correction, seismic tomography, and seismic inversion. The third objective of this invention is to provide a terminal device that, when executing its own program, can implement a deep learning-based low signal-to-noise ratio first-arrival data preprocessing method. A fourth objective of this invention is to provide a computer-readable storage medium for storing a corresponding computer program for a deep learning-based method for preprocessing low signal-to-noise ratio first-arrival data.

[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: A deep learning-based preprocessing method for low signal-to-noise ratio first-arrival wave data includes: Obtain the forward modeling single-shot data from existing resources; Noise is added to the forward model single-shot data to obtain noisy single-shot data; Curvature transform is performed on forward modeling single-shot data and noisy single-shot data to generate curve shape coefficient sub-maps, which are then used to train a deep network model. The preprocessed single-shot data after curvelet transform is input into the deep network model for detection. After being converted into curvelet coefficients, inverse curvelet transform is then performed.

[0007] As a limitation, the method includes the following steps performed sequentially: S1. Collect existing data to construct a geological model and obtain forward modeling single-shot data through forward modeling; S2. Collect noise data and load the collected noise data onto the forward modeling single-shot data to obtain noise single-shot data; S3. Perform curvelet transform on the forward modeling single-shot data, the noisy single-shot data, and the single-shot data to be preprocessed, respectively, calculate the curvelet coefficients for each scale and angle of the three types of data, and convert the curvelet coefficients into images, crop them into n×m sub-images, and obtain the curvelet coefficient sub-images of the forward modeling single-shot data, the noisy single-shot data, and the single-shot data to be preprocessed; n and m are positive integers; S4. Establish an end-to-end convolutional deep network model. Use the curve coefficient sub-map of noisy single-shot data as input data and the curve coefficient sub-map of forward modeling single-shot data as target result data. Train the deep network model so that the input data and target result data are the same size, and obtain the trained deep network model. S5. Based on the trained deep network model, the curve coefficient sub-map of the single-shot data to be preprocessed is accurately detected to obtain the noise detection result image. S6. The noise detection result image is converted and processed to obtain the curve coefficients after noise removal. The noise-removed single-shot data is obtained through inverse curve transformation.

[0008] As a further limitation, step S4 involves training the deep network model, which includes: Set the learning rate, the number of samples per training session, and the number of training rounds for the deep network model; The input data is randomly fed into the deep network model in batches for training. The model is saved once every few training epochs during the training process; The loss function and the accuracy of noise reduction are calculated during training.

[0009] As a further limitation, the operation of loading the collected noise data onto the forward modeling single-shot data in step S2 is characterized by: normalizing the amplitude of both the forward modeling single-shot data and the noise data, and then multiplying the noise value by a series of coefficients.

[0010] As a further limitation, the noise data mentioned in step S2 is noise collected in different areas in the field or noise extracted before the initial arrival of single-shot data to be preprocessed.

[0011] As a second limitation, the deep network model includes multiple cascaded noise feature extraction layers, each of which includes a convolutional layer, a pooling layer, and a data normalization layer.

[0012] As a further limitation, the deep network model is equipped with a channel attention mechanism for extracting noise feature components.

[0013] As a third limitation, the forward modeling single-shot data is two-dimensional or three-dimensional data.

[0014] As a further definition, the curved wave transformation is divided into two-dimensional curved wave transformation and three-dimensional curved wave transformation; When the forward modeling single-shot data is two-dimensional, the curvelet transform is a two-dimensional curvelet transform; When the forward modeling single-shot data is three-dimensional, the curvelet transform is either a two-dimensional curvelet transform or a three-dimensional curvelet transform.

[0015] An application of denoised single-shot data obtained based on the above method in first arrival quality control, surface velocity modeling, static correction, seismic tomography and seismic inversion.

[0016] A terminal device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned deep learning-based low signal-to-noise ratio first-arrival data preprocessing method.

[0017] A computer-readable storage medium storing a computer program, which, when executed by a processor, is used to implement the above-described deep learning-based method for preprocessing low signal-to-noise ratio first-arrival wave data.

[0018] The present invention, by adopting the above-described technical solution, achieves the following technical advancements compared to existing technologies: (1) The method of the present invention, through the combination of deep learning and curve transform, can effectively remove noise in the first arrival data, improve the quality of the first arrival data, and provide strong data support for the accurate location and evaluation of oil and gas reservoirs; (2) The method of the present invention designs training content for the deep network model, and reasonably sets the learning rate, batch size and number of training rounds, which can effectively improve the training efficiency and stability, increase the generalization ability of the model, and improve the denoising performance of the model through accuracy monitoring and loss optimization. (3) In the method of the present invention, noise is loaded onto single shot data, amplitude normalization is used, and the noise coefficient is adjusted. Amplitude normalization ensures that the two data are on the same order of magnitude, avoiding problems caused by amplitude differences. Secondly, by adjusting the noise coefficient, the intensity of noise can be flexibly controlled, simulating different signal-to-noise ratio conditions, thereby more realistically reflecting the actual exploration environment, and providing diverse test data for the training of deep network models. (4) In the method of the present invention, multiple noise feature extraction layers are set in the deep network model. The layer-by-layer extraction method is adopted to capture the multi-scale features of noise from the shallow layer to the deep layer. The shallow layer extracts local details and simple patterns, and the deep layer extracts global structure and complex patterns, thereby separating noise more comprehensively and accurately, and improving the denoising ability and generalization performance of the model. (5) The method of the present invention adds a channel attention mechanism to the deep network model, which enables the deep network model to pay more attention to the feature components related to noise, suppress irrelevant or redundant information, thereby extracting and removing noise more accurately, while retaining the key features of the effective signal, significantly improving the denoising performance and the robustness of the model. (6) In the method of the present invention, the corresponding curve transform is selected according to the dimension of the forward modeling single shot data, which can make full use of the spatial structure information of the data, thereby more accurately describing the geometric characteristics of the data, improving the effect of noise separation and signal reconstruction, and enhancing the adaptability and computational efficiency of the algorithm.

[0019] This invention belongs to the field of seismic data processing technology. By combining deep learning and curve transform, noise in first arrival data can be effectively removed, thus improving the quality of first arrival data. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0021] In the attached diagram: Figure 1 This is a flowchart of the processing in Embodiment 1 of the present invention; Figure 2 This is the forward modeling single-shot dataset exemplified in Embodiment 1 of the present invention; Figure 3 A represents a single-shot forward modeling data in Embodiment 1 of this invention; Figure 3 B is Figure 3 A. Noise data for a single gun after adding noise data; Figure 4 This is a schematic diagram of the AU-Net deep network model in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of the channel attention mechanism in Embodiment 1 of the present invention; Figure 6 This refers to a single-shot data to be preprocessed in Embodiment 1 of the present invention; Figure 7 To use Embodiment 1 of the present invention for Figure 6 The processed, noise-reduced single-shot data. Detailed Implementation

[0022] The preferred embodiments of the present invention will now be described with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of the invention.

[0023] Example 1 This embodiment presents a deep learning-based method for preprocessing low signal-to-noise ratio first-arrival data. The method includes: Obtain the forward modeling single-shot data from existing resources; Noise is added to the forward model single-shot data to obtain noisy single-shot data; Curvature transform is performed on forward modeling single-shot data and noisy single-shot data to generate curve shape coefficient sub-maps, which are then used to train a deep network model. The preprocessed single-shot data after curvelet transform is input into the deep network model for detection. After being converted into curvelet coefficients, inverse curvelet transform is then performed.

[0024] Specifically, such as Figure 1 As shown, the steps are performed sequentially: S1. Collect existing data to construct a geological model, and obtain forward modeling single-shot data through forward modeling.

[0025] The forward modeling single-shot data here can be localized data extracted from Marmousi, Overthrust, Kuche, or Sigsbee2A velocity models, simulated seismic data obtained through multiple forward simulations, publicly available synthetic data from SEG, or other 2D or 3D data. The resulting forward modeling single-shot data varies widely; example forward modeling single-shot datasets include... Figure 2 As shown.

[0026] S2. Collect noise data and load the collected noise data onto the forward modeling single-shot data to obtain noise single-shot data.

[0027] When collecting noise data, noise data can be collected in different areas in the field, or noise data can be obtained by capturing the noise before the initial arrival of the single shot data to be preprocessed.

[0028] After collecting the noise data, the amplitude of both the forward modeling single-shot data and the noise data is normalized. Then, the noise value is multiplied by a series of coefficients, so that the resulting noise single-shot data have different signal-to-noise ratio variations.

[0029] A comparison chart of the before and after adding noise data to a single shot in a forward modeling exercise is shown below. Figure 3 A and Figure 3 As shown in Figure B, noise can mask the effective signal, increasing the difficulty of initial pickup and leading to a decrease in pickup accuracy.

[0030] S3. Perform curvelet transform on the forward modeling single-shot data, the noisy single-shot data, and the single-shot data to be preprocessed, respectively, calculate the curvelet coefficients for each scale and angle of the three types of data, and convert the curvelet coefficients into images, crop them into 64×64 sub-images, and obtain the curvelet coefficient sub-images of the forward modeling single-shot data, the noisy single-shot data, and the single-shot data to be preprocessed.

[0031] There are significant differences in coefficient distribution between forward modeling single-shot data and noise data in the curve domain: the effective signal of forward modeling single-shot data exhibits strong coefficient characteristics with regular geometric shape, while noise data exhibits high-energy anomalies in random or specific directions.

[0032] In this embodiment, based on the multi-scale and multi-directional characteristics of curvelet transform, the above three types of data are decomposed into curvelet coefficients corresponding to different scales and angles, forming a multi-dimensional feature representation. Then, utilizing the multi-resolution characteristics of curvelet transform, the curvelet coefficients of the three types of data are converted into 64×64 sub-maps respectively, resulting in the curvelet coefficient sub-maps of forward modeling single-shot data, noisy single-shot data, and single-shot data to be preprocessed.

[0033] In this embodiment, the curvelet transform is a two-dimensional curvelet transform, which can be adjusted according to the actual situation: when the forward modeling single-shot data is two-dimensional data, the curvelet transform is a two-dimensional curvelet transform; when the forward modeling single-shot data is three-dimensional data, the curvelet transform can be a two-dimensional curvelet transform or a three-dimensional curvelet transform.

[0034] S4. Establish an AU-Net deep network model, using the curve coefficient submap of noisy single-shot data as input data and the curve coefficient submap of forward modeling single-shot data as target result data. Train the deep network model so that the input data and target result data are of the same size, and obtain a trained deep network model.

[0035] There are many end-to-end convolutional deep network models. In this embodiment, the AU-Net deep network model is selected, and a channel attention mechanism is added to the deep network model. The AU-Net deep network model is a deep network model with an encoder-decoder structure. It mainly uses curvelet transform to extract and dereference noise features, thereby capturing data context information for noise feature extraction. The channel attention mechanism complements the multi-scale decomposition characteristics of curvelet transform, and is used to extract the main components reflecting noise features, which can significantly improve the feature selection ability of the deep network model for curvelet coefficient subgraphs.

[0036] The AU-Net deep network model in this embodiment includes multiple noise feature extraction layers. Each noise feature extraction layer comprises a convolutional layer, a pooling layer, and a data normalization layer. These noise feature extraction layers are interconnected, with the output of one layer serving as the input to the next. This hierarchical processing mechanism complements the multi-scale decomposition characteristics of curvelet transform, preserving the physical interpretability of noise features while enhancing the modeling capability of complex noise patterns through deep learning. AU-Net deep network model such as Figure 4 As shown in the figure, the green arrows represent the channel attention mechanism in the network model. The specific structure of the channel attention mechanism is as follows: Figure 5 As shown.

[0037] When training the deep network model, the learning rate, the number of samples per training session, and the number of training epochs are set. The input data is a submap of curvesweep coefficients from noisy single-shot data, and the target result data is a submap of curvesweep coefficients from forward-modeled single-shot data. The input data is randomly fed into the deep network model in batches to ensure that the input and target result data are of similar size. During training, the model is saved every few training epochs, and the loss function and denoising accuracy are calculated and validated. The convergence of the loss function and the improvement in denoising accuracy reflect the denoising precision of the deep network model. After training, the trained deep network model is obtained.

[0038] In this embodiment, a large learning rate can be given in the early stage of training. As the number of training iterations increases and the accuracy of the predicted initial arrival position matching the actual position increases, the learning rate is gradually reduced to accelerate training efficiency. The number of samples per training iteration can be adjusted according to the device configuration. The default number of training iterations is 600, but it can also be adjusted based on the signal-to-noise ratio.

[0039] S5. Based on the trained deep network model, the curve coefficient sub-map of the single-shot data to be preprocessed is accurately detected to obtain the noise detection result image.

[0040] After obtaining the trained deep network model, the preprocessed single-shot data can be processed. The obtained curve coefficient submap of the preprocessed single-shot data is input into the trained deep network model for accurate detection, and the noise detection result image corresponding to the input curve coefficient submap is obtained.

[0041] S6. The noise detection result image is converted and processed to obtain the curve coefficients after noise removal. The noise-removed single-shot data is obtained through inverse curve transformation.

[0042] Figure 6 The figure shows a single shot data to be preprocessed, which was collected in practice. It can be seen from the figure that there are many sources of noise interference, the signal-to-noise ratio of the single shot is low, the first arrival wave is submerged by noise, the signal-to-noise ratio of the first arrival wave data is low, and the first arrival time is difficult to determine. Figure 7 To use the method of this embodiment to Figure 6 The denoised single-shot data after processing is then... Figure 7 and Figure 6 The comparison shows that random noise is suppressed, the characteristics of the first arrival wave are enhanced, the first arrival waves of adjacent channels become smooth and continuous, and the number of first arrival reciprocals that can be picked up also increases.

[0043] In this embodiment, the image is cropped to 64×64 when obtaining the curve coefficient sub-image. The two numbers 64 can be adjusted according to the actual situation, as long as both numbers are positive integers.

[0044] In this embodiment, the AU-Net deep network model was selected when choosing an end-to-end convolutional deep network model. The selection of the deep network model can be adjusted according to the actual situation, as long as the deep network model used is an end-to-end convolutional deep network model and can achieve the function of the deep network model in this embodiment.

[0045] In summary, this embodiment, through the combination of deep learning and curve transform, can effectively remove noise from the first arrival data, improve the quality of the first arrival data, and provide strong data support for the accurate location and evaluation of oil and gas reservoirs.

[0046] Example 2 This embodiment describes the application of denoised single-shot data obtained based on the method described in Embodiment 1 in first arrival quality control, surface velocity modeling, static correction, seismic tomography, and seismic inversion.

[0047] By improving the quality of first-arrival data, denoised single-shot data aids in the identification of first-arrival anomalies; provides reliable data support for surface velocity modeling; and allows for a more detailed characterization of near-surface velocity structures, thus supporting the calculation of static corrections. In seismic tomography and seismic inversion, denoised data enhances the convergence stability of velocity field reconstruction by reducing time-difference interference in ray path inversion, providing a high-fidelity initial model for full-waveform inversion. Example 3 This embodiment is a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a deep learning-based low signal-to-noise ratio first-arrival data preprocessing method according to Embodiment 1.

[0048] Example 4 This embodiment is a computer-readable storage medium. The computer-readable storage medium in this embodiment stores a computer program. When the computer program is executed by a processor, it is used to implement a low signal-to-noise ratio first-arrival data preprocessing method based on deep learning according to Embodiment 1.

[0049] The computer-readable storage medium can be a computer storage medium or a communication medium. A communication medium includes any medium that facilitates the transfer of computer programs from one location to another. A computer storage medium can be any available medium accessible to a general-purpose or special-purpose computer. For example, a computer-readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the computer-readable storage medium. Of course, the computer-readable storage medium can also be a component of the processor. The processor and the computer-readable storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the ASIC can reside in a user equipment. Of course, the processor and the computer-readable storage medium can also exist as discrete components in a communication device. Specifically, the computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk, etc. The storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

Claims

1. A deep learning-based preprocessing method for low signal-to-noise ratio first-arrival wave data, characterized in that, include: Obtain the forward modeling single-shot data from existing resources; Noise is added to the forward model single-shot data to obtain noisy single-shot data; Curvature transform is performed on forward modeling single-shot data and noisy single-shot data to generate curvelet coefficient sub-maps, which are then used to train a deep network model. The preprocessed single-shot data after curvelet transform is input into the deep network model for detection. After being converted into curvelet coefficients, inverse curvelet transform is then performed.

2. The method for preprocessing low signal-to-noise ratio first-arrival wave data based on deep learning according to claim 1, characterized in that, The method includes the following steps performed sequentially: S1. Collect existing data to construct a geological model and obtain forward modeling single-shot data through forward modeling; S2. Collect noise data and load the collected noise data onto the forward modeling single-shot data to obtain noise single-shot data; S3. Perform curvelet transform on the forward modeling single-shot data, the noisy single-shot data, and the single-shot data to be preprocessed, respectively, calculate the curvelet coefficients for each scale and angle of the three types of data, and convert the curvelet coefficients into images, crop them into n×m sub-images, and obtain the curvelet coefficient sub-images of the forward modeling single-shot data, the noisy single-shot data, and the single-shot data to be preprocessed; n and m are positive integers; S4. Establish an end-to-end convolutional deep network model. Use the curve coefficient sub-map of noisy single-shot data as input data and the curve coefficient sub-map of forward modeling single-shot data as target result data. Train the deep network model so that the input data and target result data are the same size, and obtain the trained deep network model. S5. Based on the trained deep network model, the curve coefficient sub-map of the single-shot data to be preprocessed is accurately detected to obtain the noise detection result image. S6. The noise detection result image is converted and processed to obtain the curve coefficients after noise removal. The noise-removed single-shot data is obtained through inverse curve transformation.

3. The method for preprocessing low signal-to-noise ratio first-arrival wave data based on deep learning according to claim 2, characterized in that, Step S4 involves training the deep network model, which includes: Set the learning rate, the number of samples per training session, and the number of training rounds for the deep network model; The input data is randomly fed into the deep network model in batches for training. The model is saved once every few training epochs during the training process; The loss function and the accuracy of noise reduction are calculated during training.

4. A method for preprocessing low signal-to-noise ratio first-arrival wave data based on deep learning according to claim 2 or 3, characterized in that, The operation of loading the collected noise data onto the forward modeling single-shot data in step S2 is as follows: normalize the amplitude of both the forward modeling single-shot data and the noise data, and then multiply the noise value by a series of coefficients.

5. A method for preprocessing low signal-to-noise ratio first-arrival data based on deep learning according to claim 2 or 3, characterized in that, The noise data mentioned in step S2 is noise collected in different areas in the field or noise extracted from the initial arrival of single-shot data to be preprocessed.

6. The method for preprocessing low signal-to-noise ratio first-arrival wave data based on deep learning according to claim 4, characterized in that, The noise data mentioned in step S2 is noise collected in different areas in the field or noise extracted from the initial arrival of single-shot data to be preprocessed.

7. A method for preprocessing low signal-to-noise ratio first-arrival wave data based on deep learning according to any one of claims 1, 2, 3, or 6, characterized in that, The deep network model includes multiple cascaded noise feature extraction layers, each of which includes a convolutional layer, a pooling layer, and a data normalization layer.

8. A method for preprocessing low signal-to-noise ratio first-arrival wave data based on deep learning according to claim 7, characterized in that, The deep network model is equipped with a channel attention mechanism to extract noise feature components.

9. A method for preprocessing low signal-to-noise ratio first-arrival wave data based on deep learning according to claim 1 or 2, characterized in that, The forward modeling single-shot data is either two-dimensional or three-dimensional.

10. A method for preprocessing low signal-to-noise ratio first-arrival wave data based on deep learning according to claim 9, characterized in that, The curve wave transformation is divided into two-dimensional curve wave transformation and three-dimensional curve wave transformation; When the forward modeling single-shot data is two-dimensional, the curvelet transform is a two-dimensional curvelet transform; When the forward modeling single-shot data is three-dimensional, the curvelet transform is either a two-dimensional curvelet transform or a three-dimensional curvelet transform.

11. An application of denoised single-shot data obtained by the method according to any one of claims 1 to 10 in first arrival quality control, surface velocity modeling, static correction, seismic tomography and seismic inversion.

12. A terminal 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, it implements a deep learning-based low signal-to-noise ratio first-arrival data preprocessing method as described in any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, is used to implement the deep learning-based low signal-to-noise ratio first-arrival data preprocessing method according to any one of claims 1 to 10.

Images