Data denoising method and device, electronic equipment, readable storage medium and computer program product
By combining asymmetric downsampling and multi-branch convolutional networks with random masking matrices, the problems of incomplete noise suppression and loss of signal details in seismic data are solved, achieving efficient noise removal and signal recovery, and adapting to complex noise environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-01
- Publication Date
- 2026-05-01
AI Technical Summary
Existing seismic data denoising methods are insufficient in handling spatial correlation of noise, resulting in incomplete noise suppression, easy loss of effective signal details, and high dependence on clean data, making them difficult to apply on a large scale.
An asymmetric downsampling strategy is adopted, using different sampling strides in the horizontal and vertical directions. Combined with a multi-branch convolutional network and adaptive unsupervised learning, noise removal is achieved by replacing and fusing data through a random mask matrix.
It effectively reduces spatial correlation of noise, preserves seismic signal details to the greatest extent, improves denoising accuracy, reduces dependence on clean data, and adapts to complex noise environments.
Smart Images
Figure CN121956149A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence technology, and in particular to a data denoising method, apparatus, electronic device, readable storage medium, and computer program product. Background Technology
[0002] Seismic data denoising has become a key technology in the field of geophysical exploration, primarily aiming to improve the quality of seismic data to support more accurate geological structure analysis and resource exploration. Existing seismic data denoising schemes suffer from unsatisfactory noise removal results, specifically exhibiting the following problems: 1. Insufficient handling of spatial correlation of noise, resulting in incomplete denoising: Noise in actual seismic data (such as surface waves and random interference) generally exhibits spatial correlation in the horizontal or vertical directions. However, traditional denoising methods (such as SURE-TCNN based on tensor convolution and basic blind spot networks) often assume that the noise satisfies statistical independence, failing to specifically break this correlation. This leads to incomplete noise suppression, and residual noise can interfere with subsequent geological structure identification, especially in complex geological areas (such as mountainous and marine exploration), where this problem is more prominent.
[0003] 2. Effective signal details are easily lost, and data integrity is poor: Some methods (such as denoising schemes that rely on TV regularization) may over-smooth the data when suppressing noise, resulting in the loss of key details in the seismic signal (such as stratigraphic boundaries, small-scale faults, and thin reservoir reflection information); other methods train by segmenting data blocks (such as self-supervised adjacent data block pairing strategies), which disrupts the continuity of seismic data, further reduces the integrity of the denoised data, and affects the accuracy of geological interpretation.
[0004] 3. Dependence on clean data or poor performance in unsupervised learning, and high application threshold: Supervised learning methods (such as two-stage attention mechanism networks) require a large amount of paired clean seismic data as labels, but clean data needs to be generated through repeated explorations or manual annotation, which is extremely costly and difficult to apply on a large scale; Traditional unsupervised / self-supervised methods (such as basic unsupervised loss function models) lack adaptation to the characteristics of seismic data, and have problems such as slow convergence and overfitting noise. The denoising accuracy is far lower than that of supervised learning methods, which cannot meet the actual exploration needs. Summary of the Invention
[0005] This application provides a data denoising method, apparatus, electronic device, readable storage medium, and computer program product that can solve the problem of poor noise removal effect in seismic data in the prior art.
[0006] To solve the above-mentioned technical problems, this application is implemented as follows: Firstly, a data denoising method is provided, including: Acquire the seismic data to be denoised, perform noise characteristic analysis on the seismic data to be denoised, and obtain the spatial correlation information of the noise in the seismic data to be denoised; Based on the spatial correlation information, an asymmetric downsampling strategy is used to process the seismic data to be denoised, resulting in denoised downsampled data; wherein, the asymmetric downsampling strategy employs different sampling step sizes in the horizontal and vertical directions; The downsampled data to be denoised is input into a pre-trained blind spot network to obtain preliminary denoised data; The preliminary denoised data is subjected to inverse downsampling to restore the original data dimension of the seismic data to be denoised, thus obtaining preprocessed denoised data. Multiple random mask matrices are generated, and the preprocessed denoised data and the seismic data to be denoised are replaced using the random mask matrices to obtain multiple replacement data. The multiple replacement data are respectively input into the pre-trained blind spot network to obtain multiple intermediate denoising results. The multiple intermediate denoising results are then fused to obtain the final denoised data.
[0007] Secondly, a data denoising device is provided, comprising: The data reconstruction module is used to acquire seismic data to be denoised, perform noise characteristic analysis on the seismic data to be denoised, and obtain spatial correlation information of noise in the seismic data to be denoised. An asymmetric downsampling module is used to process the seismic data to be denoised using an asymmetric downsampling strategy based on the spatial correlation information, to obtain the downsampled data to be denoised; wherein, the asymmetric downsampling strategy uses different sampling step sizes in the horizontal and vertical directions; The blind spot network denoising module is used to input the downsampled data to be denoised into a pre-trained blind spot network to obtain preliminary denoised data; The preprocessing module is used to perform inverse downsampling on the preliminary denoised data to restore the original data dimension of the seismic data to be denoised, thereby obtaining preprocessed denoised data. A random substitution module is used to generate multiple random mask matrices, and to use the random mask matrices to replace the preprocessed denoised data and the seismic data to be denoised, thereby obtaining multiple replacement data. The blind spot network denoising module is further configured to input the multiple replacement data into the pre-trained blind spot network to obtain multiple intermediate denoising results, and to fuse the multiple intermediate denoising results to obtain the final denoised data.
[0008] Thirdly, an electronic device is provided, comprising a processor and a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor to implement the above-described data denoising method.
[0009] Fourthly, a readable storage medium is provided, wherein at least one computer program is stored in the readable storage medium, the computer program being loaded and executed by a processor to implement the above-described data denoising method.
[0010] Fifthly, a computer program product is provided, the computer program product comprising at least one computer program, the computer program being loaded and executed by a processor to implement the data denoising method provided in the various optional implementations described above.
[0011] The data denoising method, apparatus, electronic device, readable storage medium, and computer program product provided in this application introduce an asymmetric downsampling strategy. By using different sampling step sizes in the horizontal and vertical directions, the spatial correlation between adjacent noisy sampling points is effectively reduced, enabling the preservation of effective details of the seismic signal to the greatest extent while suppressing noise. Furthermore, this application improves upon existing blind spot networks by employing a multi-branch convolutional network structure combined with dilated convolutional modules of different dilation rates. This allows for the extraction of seismic signal features at multiple scales, effectively solving the problem of signal information loss during seismic data denoising and significantly improving the accuracy of the blind spot network in handling complex noise. To further enhance the denoising effect, this application proposes a random substitution refinement strategy, using a random mask matrix to refine and remove noise, reducing the spatial correlation of noise signals, enhancing the recovery of effective signals, and ensuring data integrity. Finally, this application employs an adaptive unsupervised learning strategy. By introducing a loss function based on the asymmetric downsampling strategy, the blind spot network can effectively learn and recover clean seismic signals from noisy seismic data without relying on expensive or difficult-to-obtain clean label data.
[0012] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description
[0013] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0014] Figure 1 A flowchart of a data denoising method provided in an exemplary embodiment of this application is shown; Figure 2A flowchart illustrating the specific steps of step S101 provided in an exemplary embodiment of this application is shown. Figure 3 A noise feature map of seismic data provided in an exemplary embodiment of this application is shown; Figure 4 A flowchart illustrating a method for training a blind spot network according to an exemplary embodiment of this application is shown. Figure 5 A schematic diagram of the structure of a denoising model for a blind spot network provided in an exemplary embodiment of this application is shown; Figure 6 A flowchart illustrating the method for training a blind spot network provided in an application example of this application is shown. Figure 7 This illustration shows a schematic diagram of a process for replacing and fusing data using a random mask matrix to obtain the final denoised data, provided by an exemplary embodiment of this application. Figure 8 A flowchart of the data denoising method provided in an application example of this application is shown; Figure 9 This invention illustrates a structural block diagram of a data denoising apparatus provided in an exemplary embodiment of this application; Figure 10 A structural block diagram of an electronic device provided in an exemplary embodiment of this application is shown. Detailed Implementation
[0015] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.
[0016] To address the problem of poor noise removal performance in existing seismic data, this application provides a data denoising method.
[0017] The execution subject of this method can be various types of computing devices, or it can be an application or app installed on a computing device. The computing device can be a user terminal such as a mobile phone, tablet computer, or smart wearable device, or it can be a server. The server can be a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing, consisting of a large number of computers and network servers.
[0018] The following will combine Figures 1 to 8The data denoising methods provided in the embodiments of this application are explained and described in detail. These embodiments are only used to explain this application and do not constitute a limitation thereof.
[0019] Figure 1 A flowchart illustrating a data denoising method according to an exemplary embodiment of this application is shown. Figure 1 As shown, this data denoising method mainly includes the following steps (S101-S106): S101. Obtain the seismic data to be denoised, perform noise characteristic analysis on the seismic data to be denoised, and obtain the spatial correlation information of noise in the seismic data to be denoised. In this step, the seismic data to be denoised is acquired and its characteristics are analyzed. The core is to clarify the distribution patterns of noise and effective signals through multi-dimensional analysis, providing a basis for subsequent downsampling strategies and network parameter design. The implementation method of step S101 is described below.
[0020] In some embodiments, noise characteristic analysis is performed on the seismic data to be denoised to obtain spatial correlation information of noise in the seismic data to be denoised, including: calculating the noise mean and variance of each seismic trace in the seismic data to be denoised, and calculating the correlation between each sampling point in the seismic data to be denoised and surrounding sampling points, to determine the spatial correlation distribution law of noise in the horizontal and vertical directions in the seismic data to be denoised.
[0021] In some embodiments, calculating the correlation between each sampling point in the seismic data to be denoised and surrounding sampling points includes: selecting a target sub-data block in the seismic data to be denoised, calculating the two-dimensional correlation coefficient between the target sub-data block and sub-data blocks of the same size in different directions and at different distances, and obtaining the average correlation between each sampling point and surrounding sampling points as a function of direction and distance by repeating the calculation a preset number of times and taking the average.
[0022] For example, such as Figure 2 As shown, the specific steps of step S101 are as follows: S1011. Using a real noise block containing 768 milliseconds and 128 seismic traces, first calculate the noise mean and variance for each seismic trace; S1012. Then, using noise blocks of the same size, calculate the correlation between each sampling point and its surrounding sampling points; S1013. Finally, randomly select a 30×30 noise block and calculate the two-dimensional correlation coefficient between the noise block and noise blocks of the same size in different directions and at different distances. S1014. By repeating the process multiple times and taking the average, analyze how the average correlation between the noise sampling point and its surrounding sampling points changes with direction and distance.
[0023] In this embodiment of the application, by performing noise characteristic analysis on the seismic data to be denoised in S101, the spatial correlation of noise in the actual data is significantly reduced, enabling the blind spot network to cope with spatially correlated noise problems and expanding the application scope of the blind spot network.
[0024] Figure 3 The noise characteristics in the seismic data are presented from different dimensions. Figure 3(a) shows a schematic diagram of a two-dimensional profile of seismic noise, as... Figure 3 (b) shows a schematic diagram of the statistical characteristics analysis of noise. Figure 3 (c) shows a schematic diagram of spatial correlation analysis of noise. As shown in Figure 3(a), the horizontal axis represents the number of seismic traces, and the vertical axis represents time (in milliseconds). The image exhibits obvious noise textures (such as oblique and layered interference patterns), intuitively reflecting the spatial distribution of noise in the seismic data. This type of noise is mostly surface waves and random interference, which has strong spatial correlation and is a core pain point that traditional denoising methods struggle to completely suppress. The asymmetric downsampling strategy mentioned later in this application is designed specifically for this type of spatially distributed noise, breaking its correlation through differentiated step sizes. Figure 3 (b) and Figure 3 As shown in (c), although the noise has zero mean, it exhibits varying degrees of spatial correlation in both the horizontal and vertical directions, indicating that traditional blind spot networks cannot effectively remove this type of noise. Therefore, this application proposes a novel asymmetric downsampling strategy, as detailed in step S102.
[0025] S102. Based on spatial correlation information, an asymmetric downsampling strategy is used to process the seismic data to be denoised, resulting in downsampled data to be denoised. The asymmetric downsampling strategy involves using different sampling step sizes in the horizontal and vertical directions. Existing blind spot networks, as a method trained on individual noisy samples, reduce the need for clear data, but still have limitations in denoising real-world seismic data. This method is primarily suited for removing noise with zero mean and statistical independence. However, noise in real-world seismic data often exhibits strong spatial correlation, not fully meeting the assumptions of blind spot networks. Therefore, this application proposes an effective asymmetric downsampling strategy by analyzing the statistical characteristics and spatial correlation of noise in real-world seismic data, aiming to expand the application scope of blind spot networks and enable them to handle non-statistically independent real-world noise data.
[0026] The implementation method of step S102 will be described below.
[0027] The asymmetric downsampling strategy employed in this application reduces the spatial correlation of noise by increasing the distance between adjacent noise sampling points. In some embodiments, the asymmetric downsampling strategy includes a first sampling step size parameter and a second sampling step size parameter. The first sampling step size parameter is used to process the seismic data to be denoised obtained in step S101, and the second sampling step size parameter is used to process the noisy seismic data used for training the blind spot network. The vertical sampling step size in the first sampling step size parameter is smaller than the vertical sampling step size in the second sampling step size parameter.
[0028] For example, in some embodiments, the horizontal sampling step size in the first sampling step size parameter is ≥1, and the vertical sampling step size in the first sampling step size parameter is ≥2; the horizontal sampling step size in the second sampling step size parameter is ≥1, and the vertical sampling step size in the second sampling step size parameter is ≥5.
[0029] For example, in some embodiments, in the asymmetric downsampling strategy during the actual denoising stage, the vertical sampling step size in the first sampling step size parameter is 2. In the asymmetric downsampling strategy during the training stage, the vertical sampling step size in the second sampling step size parameter is 5.
[0030] Asymmetric downsampling strategies can suppress spatial correlation of noise while preserving the detailed information of seismic signals to the maximum extent, achieving a balance between noise suppression and signal leakage. However, a large downsampling step size can induce aliasing artifacts, especially when the frequency components of the original signal are not effectively limited. To address this issue, the blind spot network is further trained to identify and remove aliasing artifacts caused by a large step size factor. For example, in the asymmetric downsampling strategy of this application embodiment, the vertical sampling step size during training is set to 5, making the noise information appear spatially uncorrelated; the vertical sampling step size during the testing phase is set to 2, reducing both the impact of aliasing artifacts and the spatial correlation between adjacent noise signals. Blind spot networks based on different sampling step sizes can remove noise in an unsupervised manner without requiring any clean samples, while preserving the details of the seismic signal.
[0031] In this embodiment, an asymmetric downsampling strategy is introduced. By using different sampling step sizes in the horizontal and vertical directions, the spatial correlation between adjacent noise sampling points is effectively reduced, which can suppress noise while preserving the effective details of the seismic signal to the greatest extent.
[0032] S103. Input the downsampled data to be denoised into the pre-trained blind spot network to obtain preliminary denoised data; In some embodiments, before inputting the downsampled data to be denoised into a pre-trained blind spot network to obtain preliminary denoised data, the method provided in this application embodiment further includes a method for pre-training the blind spot network. Figure 4 A flowchart illustrating a method for training a blind spot network, as shown in an exemplary embodiment of this application, is presented. Figure 4 As shown, the method for training a blind spot network provided in this application embodiment mainly includes the following steps (S201-S204): S201. Obtain noisy seismic data for training, perform noise characteristic analysis on the noisy seismic data for training, and determine the spatial correlation distribution of noise during the training phase. S202. An asymmetric downsampling strategy is used to process the noisy seismic data used for training to obtain the downsampled data used for training. The specific implementation methods of steps S201 to S202 are the same as those of steps S101-S102. For details, please refer to the detailed description of steps S101-S102 above, which will not be repeated here.
[0033] It is important to note that, for example, in some embodiments, the vertical sampling step size in the first sampling step size parameter of the asymmetric downsampling strategy in the actual denoising stage of step S102 is 2. However, in this step S202, i.e., the asymmetric downsampling strategy in the training stage, the vertical sampling step size in the second sampling step size parameter is 5, making the noise information appear spatially uncorrelated. Although aliasing artifacts have low spatial correlation in the data subplot, they still follow the zero-mean constraint, which is comparable to the statistical mean of the noise data. Therefore, the optimized blind spot network can effectively learn and remove these artifacts, thereby improving the quality of the denoised data. Moreover, blind spot networks based on different sampling step sizes can remove noise in an unsupervised manner without requiring any clean samples, while preserving the details of the seismic signal.
[0034] S203. Construct the initial blind spot network by inputting the downsampled training data into the initial blind spot network and calculating the loss value using an adaptive unsupervised loss function. This application optimizes and adjusts the existing blind spot network architecture to provide a denoising model for seismic data. Figure 5 A schematic diagram of the structure of a denoising model for a blind spot network provided in an embodiment of this application is shown. In some embodiments, such as Figure 5 As shown, the blind spot network constructed in this embodiment includes: A1. Set the input layer as a 1×1 convolutional layer with 64 convolutional kernels, and connect a ReLU activation function after the 1×1 convolutional layer; In this embodiment, the model first includes a 1×1 convolutional layer with 64 convolutional kernels, followed by the introduction of nonlinearity through the ReLU activation function to enhance the data feature extraction capability. This convolutional layer serves as the initial layer of the model, primarily used to capture low-dimensional features and lay the foundation for subsequent multi-scale feature extraction.
[0035] A2. Set up a feature extraction layer. The feature extraction layer includes three branches. The three branches use mask convolution kernels of different sizes. Each branch is configured with 64 convolution kernels. Multiple dilated convolution modules are deployed inside each branch. Each dilated convolution module is set with a different dilation rate. Next, the model is designed with three independent branches, each focusing on feature extraction at a different scale. Specifically, the three branches employ mask convolution techniques of different sizes: the first branch uses a 3×3 mask convolution kernel; the second branch uses a 5×5 mask convolution kernel; and the third branch uses a 7×7 mask convolution kernel.
[0036] Each branch contains 64 convolutional kernels, further enhancing the model's feature extraction capabilities. Within each branch, 12 dilated convolutional modules are deployed with dilation rates of 2, 3, and 4. By setting different dilation rates, this structure can adapt to feature extraction needs at various scales, ensuring effective capture of seismic signals at different spatial scales. Each dilated convolutional module employs a residual learning framework, combining 1×1 convolutional layers with corresponding 3×3 convolutional layers at appropriate dilation rates. The residual learning framework improves the model's training efficiency by reducing the gradient vanishing problem. Furthermore, each convolutional module is equipped with two ReLU activation functions to further enhance the model's non-linear expressive power, thereby increasing the depth of feature extraction.
[0037] A3. Set up the output layer. The output layer fuses the feature maps extracted from the three branches. The fusion process includes passing through three 1×1 convolutional layers and two ReLU activation functions in sequence, and finally outputs the intermediate denoising result.
[0038] After feature extraction is completed in the three branches, all extracted feature maps will be fused and refined through three 1×1 convolutional layers and two ReLU activation functions to finally generate denoised seismic data output.
[0039] In this embodiment, a complementary masking technique is introduced based on the blind spot network. The noisy seismic data is divided into two opposing complementary subsets, which are then denoised separately and recombined, thereby enhancing the model's ability to handle blind spot noise. Compared to the single-view processing method in existing technologies, this innovative grouped denoising method can comprehensively separate noise and signal from multiple perspectives, effectively solving the problem of signal information loss during seismic data denoising and significantly improving the accuracy of the blind spot network when processing complex noise.
[0040] In some embodiments, calculating the loss value using an adaptive unsupervised loss function includes: The loss value is calculated using the following formula (1) for the adaptive unsupervised loss function. : (1) in, Asymmetric downsampling operation, This is the inverse operation of asymmetric downsampling. Sampling rate, This is the mapping function for the initial blind spot network. These are the parameters of the initial blind spot network. Noisy seismic data was used for training. For expectation operators.
[0041] This application employs an adaptive unsupervised learning strategy aimed at learning from noise-contaminated seismic data. Restore its clean data To achieve this goal, a denoising model was designed. By learning model parameters Mapping noisy data to denoised data Formula (1) of the adaptive unsupervised loss function is used to measure the difference between the model output and the original noisy data.
[0042] In this embodiment, the sampling rate s=5. By introducing downsampling operations with different sampling rates, the model can adapt to different levels of noise and sampling rates, thereby improving the denoising effect. Based on the above unsupervised loss function, the denoising model can effectively learn and recover clean seismic signals from noisy seismic data without relying on expensive or difficult-to-obtain clean labeled data. This loss function enhances the robustness of the denoising model, especially when dealing with seismic noise of different complexities and types, ensuring a stable improvement in denoising performance. The introduced asymmetric downsampling strategy further improves the adaptability and denoising performance of the blind spot network by breaking the spatial correlation of noise in the actual data.
[0043] S204. Based on the loss value, the parameters of the initial blind spot network are iteratively updated using the gradient descent method until the adaptive unsupervised loss function converges, thus obtaining the pre-trained blind spot network.
[0044] This completes the training of the blind spot network.
[0045] Below, in an application example, the method for training a blind spot network provided in the embodiments of this application will be explained by way of example. Figure 6 A flowchart illustrating a method for training a blind spot network, as shown in an exemplary embodiment of this application, is presented. Figure 6 As shown, the method for training a blind spot network provided in this application embodiment mainly includes the following steps (S301-S304): S301. An asymmetric downsampling strategy with a step size operator of 5 is used to process the training data. Decompose and reassemble to generate data blocks. To eliminate the spatial correlation of noise; S302, Use the initial blind spot network to reconstruct the data Denoising is performed to generate denoised data. ; S303. An inverse downsampling strategy with a step size of 5 is adopted to denoise the data. Combined to obtain preprocessing results ; S304, Calculate the preprocessing results With noisy training data To mitigate the losses between different network parameters, optimize network parameters to improve model performance.
[0046] In this embodiment, the downsampled data to be denoised obtained in step S102 is input into the trained blind spot network. The blind spot network suppresses noise at different scales through multi-branch feature extraction layers, outputting preliminary denoised data with a dimension of 512×128. In some examples, the noise mean of the preliminary denoised data is reduced from 0.085 in the original data to 0.012, the noise energy is reduced by 85%, and the peak retention rate of the effective signal reaches 98%.
[0047] In this embodiment, a pre-trained blind spot network extracts features from seismic data at different scales to effectively remove noise. This significantly improves the denoising effect while reducing complexity.
[0048] S104. Perform inverse downsampling on the preliminary denoised data to restore the original data dimension of the seismic data to be denoised, and obtain the preprocessed denoised data. S105. Generate multiple random mask matrices, and use the random mask matrices to replace the preprocessed denoised data and the seismic data to be denoised, to obtain multiple replacement data. While a downsampling strategy with a step size operator of 2 can effectively reduce signal artifacts, this step size also leads to the loss of some high-frequency signals. Therefore, in order to optimize the recovery of seismic signal details, this application proposes an optimization method for the mask matrix. Figure 7 This illustration shows a schematic diagram of data replacement and fusion using a random mask matrix, as provided in an embodiment of this application, to obtain the final denoised data. This mask matrix optimization method reduces signal loss and optimizes denoising by performing local data replacement.
[0049] In some embodiments, the mask matrix is a binary mask matrix, and satisfies Where T is the time dimension length of the seismic data to be denoised, and S is the number of seismic traces in the seismic data to be denoised; It is a structured binary matrix generated by a moving window with a step size of 2, and satisfies .
[0050] In this embodiment, a random replacement refinement strategy is proposed, which further reduces the correlation between noise signals by improving the rule sampling method. For example... Figure 7 As shown, the random replacement refinement strategy adopts A randomly generated binary mask Replace the original fixed mask This yields a random mask matrix.
[0051] like Figure 7 As shown, the seismic data to be denoised is rendered using i mask matrices. Some of the noise in the data was replaced with pre-processed denoised data. The corresponding data in the table is used to complete a partial replacement operation.
[0052] In some embodiments, a random mask matrix is used to perform data replacement on the preprocessed denoised data and the seismic data to be denoised, resulting in multiple replacement data, including: The replacement data is obtained by using the random mask matrix to replace the preprocessed denoised data and the seismic data to be denoised using the following formula (2). : (2) in, For the i-th replacement data, Let y be the i-th random mask matrix, y be the seismic data to be denoised, and I be the identity matrix. To preprocess the denoised data, This is for the Hadamard product operation.
[0053] In this embodiment, since the noise data is randomly placed in the first... Replacement data Therefore, the expected correlation between adjacent noise signals will be multiplied by the probability. This significantly reduces the correlation between noise levels. Compared to the regular masking strategy, the random replacement refinement strategy can significantly reduce the expected correlation between noise levels, thereby enhancing the denoising effect.
[0054] By combining a random replacement refinement strategy with an asymmetric downsampling blind spot network, the data after random replacement... By directly inputting the blind spot network for processing, the spatial correlation of input noise is greatly reduced, further optimizing the denoising effect.
[0055] In this embodiment, a random replacement refinement strategy is introduced. By introducing a random mask matrix, the correlation between replacement data is reduced, thereby further reducing the spatial correlation of noisy data and further improving the accuracy of data denoising and detail preservation, especially with significant effect on the recovery of high-frequency signals.
[0056] S106. Input multiple replacement data into the pre-trained blind spot network to obtain multiple intermediate denoising results. Fuse the multiple intermediate denoising results to obtain the final denoised data.
[0057] like Figure 7 As shown, in some embodiments, multiple intermediate denoising results are fused to obtain the final denoised data, including: The final denoised data is calculated using the following fusion formula (3). : (3) in, The mapping function for the pre-trained blind spot network. Let represent the i-th intermediate denoising result, and n represent the number of random mask matrices.
[0058] In this embodiment, after the replacement is completed, a denoising model is used. For each A second denoising process is performed. The final denoising result is reconstructed by averaging each processed data point. .
[0059] Below, an application example will be provided to illustrate the data denoising method provided in the embodiments of this application. Figure 8 A flowchart illustrating a data denoising method according to an exemplary embodiment of this application is shown. Figure 8 As shown, the data denoising method provided in this application mainly includes the following steps (S401-S406): S401. A downsampling strategy with a step size operator of 2 is adopted for the actual data with noise reduction. Decompose and generate data blocks ; S402, Utilizing a pre-trained blind spot network to... Denoising is performed to obtain the denoising result. ; S403. Employ an inverse downsampling strategy with a step size of 2 to denoise the data. Reassemble to generate preprocessed results ; S404, generated based on Bernoulli sampling A random mask matrix Extracting partial data from the original noisy data using a mask matrix. and replace the original denoised data. ,get The final result is the reconstructed data after replacement. ; S405, will Data The reconstructed matrix is input into the blind spot network and denoised again to obtain the denoised result. Since the step size is 2, the calculation yields... ; S406, to Denoising results Averaging is performed to generate the final denoised output. .
[0060] The data denoising method provided in this application introduces an asymmetric downsampling strategy. By using different sampling step sizes in the horizontal and vertical directions, it effectively reduces the spatial correlation between adjacent noisy sampling points, enabling the preservation of effective details of the seismic signal to the greatest extent while suppressing noise. Furthermore, this application improves upon existing blind spot networks by employing a multi-branch convolutional network structure combined with dilated convolutional modules of different dilation rates. This allows for the extraction of seismic signal features at multiple scales, effectively solving the problem of signal information loss during seismic data denoising and significantly improving the accuracy of the blind spot network in handling complex noise. To further enhance the denoising effect, this application proposes a random substitution refinement strategy. This strategy uses a random mask matrix to refine and remove noise, reducing the spatial correlation of noise signals, enhancing the recovery of effective signals, and ensuring data integrity. Finally, this application employs an adaptive unsupervised learning strategy. By introducing a loss function based on the asymmetric downsampling strategy, the blind spot network can effectively learn and recover clean seismic signals from noisy seismic data without relying on expensive or difficult-to-obtain clean label data.
[0061] Figure 9 This diagram illustrates a block diagram of a data denoising apparatus 100 provided in an exemplary embodiment of this application, which can achieve the following: Figures 1 to 8All or part of the contents of any of the embodiments shown. The following is only a brief description of the structure and function of the data denoising device 100. For other matters not covered, please refer to the relevant descriptions in the above data denoising method. The embodiments of the data denoising device 100 correspond to the embodiments of the above data denoising method. All implementation processes and methods of the above method embodiments can be applied to the embodiments of the data denoising device and can achieve the same technical effect.
[0062] In some embodiments, such as Figure 9 As shown, the data denoising device includes: The data reconstruction module 101 is used to acquire the seismic data to be denoised, perform noise characteristic analysis on the seismic data to be denoised, and obtain the spatial correlation information of noise in the seismic data to be denoised. The asymmetric downsampling module 102 is used to process the seismic data to be denoised based on spatial correlation information using an asymmetric downsampling strategy to obtain the downsampled data to be denoised; wherein, the asymmetric downsampling strategy is to use different sampling step sizes in the horizontal and vertical directions; The blind spot network denoising module 103 is used to input the downsampled data to be denoised into the pre-trained blind spot network to obtain preliminary denoised data; Preprocessing module 104 is used to perform inverse downsampling on the preliminary denoised data to restore the original data dimension of the seismic data to be denoised, thus obtaining preprocessed denoised data; The random substitution module 105 is used to generate multiple random mask matrices, and to use the random mask matrices to replace the preprocessed denoised data and the seismic data to be denoised, thereby obtaining multiple replacement data. The blind spot network denoising module 103 is also used to input multiple replacement data into the pre-trained blind spot network to obtain multiple intermediate denoising results, and to fuse the multiple intermediate denoising results to obtain the final denoised data.
[0063] In some embodiments, the data reconstruction module 101 performs noise characteristic analysis on the seismic data to be denoised in the following manner to obtain spatial correlation information of noise in the seismic data to be denoised: calculates the noise mean and variance of each seismic trace in the seismic data to be denoised, and calculates the correlation between each sampling point in the seismic data to be denoised and the surrounding sampling points, thereby determining the spatial correlation distribution law of noise in the horizontal and vertical directions in the seismic data to be denoised.
[0064] In some embodiments, the data reconstruction module 101 calculates the correlation between each sampling point in the seismic data to be denoised and the surrounding sampling points in the following manner: selects a target sub-data block in the seismic data to be denoised, calculates the two-dimensional correlation coefficient between the target sub-data block and sub-data blocks of the same size in different directions and at different distances, and obtains the average correlation between each sampling point and the surrounding sampling points as a function of direction and distance by repeating the calculation a preset number of times and taking the average.
[0065] In some embodiments, before inputting the downsampled data to be denoised into the pre-trained blind spot network to obtain preliminary denoised data, the blind spot network denoising module 103 is also used to pre-train the blind spot network.
[0066] The blind spot network denoising module 103 pre-trains the blind spot network in the following ways: acquiring noisy seismic data for training, performing noise characteristic analysis on the noisy seismic data for training, and determining the spatial correlation distribution of noise during the training phase; processing the noisy seismic data for training using an asymmetric downsampling strategy to obtain downsampled data for training; constructing an initial blind spot network by inputting the downsampled data for training into the initial blind spot network and calculating the loss value using an adaptive unsupervised loss function; and iteratively updating the parameters of the initial blind spot network based on the loss value using the gradient descent method until the adaptive unsupervised loss function converges, thereby obtaining the pre-trained blind spot network.
[0067] In some embodiments, the asymmetric downsampling strategy includes a first sampling step size parameter and a second sampling step size parameter. The first sampling step size parameter is used for processing the seismic data to be denoised, and the second sampling step size parameter is used for processing the noisy seismic data for training. The vertical sampling step size in the first sampling step size parameter is smaller than the vertical sampling step size in the second sampling step size parameter.
[0068] In some embodiments, the horizontal sampling step size in the first sampling step size parameter is ≥1, and the vertical sampling step size in the first sampling step size parameter is ≥2; the horizontal sampling step size in the second sampling step size parameter is ≥1, and the vertical sampling step size in the second sampling step size parameter is ≥5.
[0069] In some embodiments, the blind spot network denoising module 103 constructs the blind spot network in the following manner: Set up an input layer, which is a 1×1 convolutional layer with 64 convolutional kernels. The 1×1 convolutional layer is followed by a ReLU activation function. A feature extraction layer is set up, which includes three branches. Each branch uses a mask convolution kernel of a different size. Each branch is configured with 64 convolution kernels, and multiple dilated convolution modules are deployed inside each branch. Each dilated convolution module is set with a different dilation rate. The output layer is set up to fuse the feature maps extracted from the three branches. The fusion process includes passing through three 1×1 convolutional layers and two ReLU activation functions in sequence, and finally outputting the intermediate denoising result.
[0070] In some embodiments, the blind spot network denoising module 103 calculates the loss value using an adaptive unsupervised loss function by: calculating the loss value using the following formula for the adaptive unsupervised loss function. :
[0071] in, Asymmetric downsampling operation, This is the inverse operation of asymmetric downsampling. Sampling rate, This is the mapping function for the initial blind spot network. These are the parameters of the initial blind spot network. Noisy seismic data was used for training. For expectation operators.
[0072] In some embodiments, the random mask matrix is a binary mask matrix, and satisfies Where T is the time dimension length of the seismic data to be denoised, and S is the number of seismic traces in the seismic data to be denoised; In some embodiments, the random substitution module 105 performs data substitution on the preprocessed denoised data and the seismic data to be denoised using a random mask matrix in the following manner to obtain multiple replacement data: The replacement data is obtained by using a random mask matrix to replace the preprocessed denoised data and the seismic data to be denoised using the following formula. :
[0073] in, For the i-th replacement data, Let y be the i-th random mask matrix, y be the seismic data to be denoised, and I be the identity matrix. To preprocess the denoised data, This is for the Hadamard product operation.
[0074] In some embodiments, the blind spot network denoising module 103 fuses multiple intermediate denoising results to obtain the final denoised data in the following manner: The final denoised data is calculated using the following fusion formula. :
[0075] in, The mapping function for the pre-trained blind spot network. Let represent the i-th intermediate denoising result, and n represent the number of random mask matrices.
[0076] The data denoising device provided in this application introduces an asymmetric downsampling strategy. By using different sampling step sizes in the horizontal and vertical directions, it effectively reduces the spatial correlation between adjacent noisy sampling points, thus suppressing noise while preserving the effective details of the seismic signal to the greatest extent. Furthermore, this application improves upon existing blind spot networks by employing a multi-branch convolutional network structure combined with dilated convolutional modules of different dilation rates. This enables the extraction of seismic signal features at multiple scales, effectively solving the problem of signal information loss during seismic data denoising and significantly improving the accuracy of the blind spot network in handling complex noise. To further enhance the denoising effect, this application proposes a random substitution refinement strategy. This strategy refines and removes noise using a random mask matrix, reducing the spatial correlation of noise signals, enhancing the recovery of effective signals, and ensuring data integrity. Finally, this application employs an adaptive unsupervised learning strategy. By introducing a loss function based on the asymmetric downsampling strategy, the blind spot network can effectively learn and recover clean seismic signals from noisy seismic data without relying on expensive or difficult-to-obtain clean label data.
[0077] Figure 10 The diagram illustrates a structural block diagram of an electronic device 1000 according to an exemplary embodiment of this application. The electronic device 1000 can be implemented as the aforementioned data denoising device, which can be configured in electronic devices such as smartphones, tablets, laptops, desktop computers, smartwatches, televisions, or servers.
[0078] Typically, electronic device 1000 includes a processor 1001 and a memory 1002.
[0079] Processor 1001 may include one or more processing cores, such as a quad-core processor, a deca-core processor, etc. Processor 1001 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). Processor 1001 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 1001 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 1001 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.
[0080] The memory 1002 may include one or more computer-readable storage media, which may be non-transitory. The memory 1002 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 1002 are used to store at least one instruction, which is executed by the processor 1001 to implement all or part of the steps in the data denoising method shown in the method embodiments of this application.
[0081] Those skilled in the art will understand that Figure 10 The structure shown does not constitute a limitation on the electronic device 1000, and may include more or fewer components than shown, or combine certain components, or use different component arrangements.
[0082] In one exemplary embodiment, a readable storage medium is also provided, which stores a program or instructions that, when executed by a processor, implement all or part of the steps in the data denoising method described above. For example, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, or optical data storage device, etc.
[0083] In one exemplary embodiment, a computer program product is also provided, comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions that, when executed by a computer, cause the computer to perform the aforementioned... Figures 1 to 8 All or part of the steps of the data denoising method shown in any embodiment.
[0084] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.
[0085] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.
Claims
1. A data denoising method, characterized in that, include: Acquire the seismic data to be denoised, perform noise characteristic analysis on the seismic data to be denoised, and obtain the spatial correlation information of the noise in the seismic data to be denoised; Based on the spatial correlation information, an asymmetric downsampling strategy is used to process the seismic data to be denoised, resulting in denoised downsampled data; wherein, the asymmetric downsampling strategy employs different sampling step sizes in the horizontal and vertical directions; The downsampled data to be denoised is input into a pre-trained blind spot network to obtain preliminary denoised data; The preliminary denoised data is subjected to inverse downsampling to restore the original data dimension of the seismic data to be denoised, thus obtaining preprocessed denoised data. Multiple random mask matrices are generated, and the preprocessed denoised data and the seismic data to be denoised are replaced using the random mask matrices to obtain multiple replacement data. The multiple replacement data are respectively input into the pre-trained blind spot network to obtain multiple intermediate denoising results. The multiple intermediate denoising results are then fused to obtain the final denoised data.
2. The method according to claim 1, characterized in that, The step of performing noise characteristic analysis on the seismic data to be denoised to obtain spatial correlation information of the noise in the seismic data to be denoised includes: The mean and variance of noise in each seismic trace of the seismic data to be denoised are calculated, and the correlation between each sampling point in the seismic data to be denoised and the surrounding sampling points is calculated to determine the spatial correlation distribution law of noise in the horizontal and vertical directions of the seismic data to be denoised.
3. The method according to claim 2, characterized in that, The calculation of the correlation between each sampling point in the seismic data to be denoised and its surrounding sampling points includes: Select the target sub-data block in the seismic data to be denoised, calculate the two-dimensional correlation coefficient between the target sub-data block and sub-data blocks of the same size in different directions and at different distances, and obtain the average correlation between each sampling point and surrounding sampling points as a function of direction and distance by repeating the calculation a preset number of times and taking the average.
4. The method according to claim 1, characterized in that, Before inputting the downsampled data to be denoised into the pre-trained blind spot network to obtain preliminary denoised data, the method further includes: Acquire noisy seismic data for training, perform noise characteristic analysis on the noisy seismic data for training, and determine the spatial correlation distribution of noise during the training phase; The asymmetric downsampling strategy is used to process the noisy seismic data used for training to obtain downsampled data used for training. Construct the initial blind spot network by inputting the training downsampled data into the initial blind spot network and calculating the loss value using an adaptive unsupervised loss function; Based on the loss value, the parameters of the initial blind spot network are iteratively updated using gradient descent until the adaptive unsupervised loss function converges, thus obtaining the pre-trained blind spot network.
5. The method according to claim 4, characterized in that, The asymmetric downsampling strategy includes a first sampling step size parameter and a second sampling step size parameter. The first sampling step size parameter is used to process the seismic data to be denoised, and the second sampling step size parameter is used to process the noisy seismic data for training. The vertical sampling step size in the first sampling step size parameter is smaller than the vertical sampling step size in the second sampling step size parameter.
6. The method according to claim 5, characterized in that, The horizontal sampling step size in the first sampling step size parameter is ≥1, and the vertical sampling step size in the first sampling step size parameter is ≥2; The horizontal sampling step size in the second sampling step size parameter is ≥1, and the vertical sampling step size in the second sampling step size parameter is ≥5.
7. The method according to claim 4, characterized in that, The construction of the blind spot network includes: An input layer is set up, which is a 1×1 convolutional layer with 64 convolutional kernels. The 1×1 convolutional layer is followed by a ReLU activation function. A feature extraction layer is set up, which includes three branches. The three branches use mask convolution kernels of different sizes. Each branch is configured with 64 convolution kernels, and multiple dilated convolution modules are deployed inside each branch. Each dilated convolution module is set with a different dilation rate. An output layer is set up, which fuses the feature maps extracted from the three branches. The fusion process includes passing through three 1×1 convolutional layers and two ReLU activation functions in sequence, and finally outputs the intermediate denoising result.
8. The method according to claim 4 or 7, characterized in that, The calculation of the loss value using an adaptive unsupervised loss function includes: The loss value is calculated using the following formula for the adaptive unsupervised loss function. : in, This is an asymmetric downsampling operation. This is the inverse operation of asymmetric downsampling. Sampling rate, The mapping function for the initial blind spot network. The parameters of the initial blind spot network are... The training data is noisy seismic data. For expectation operators.
9. The method according to claim 1, characterized in that, The random mask matrix is a binary mask matrix, and satisfies the following conditions: Where T is the time dimension length of the seismic data to be denoised, and S is the number of seismic traces of the seismic data to be denoised; The process involves using the random mask matrix to perform data replacement on the preprocessed denoised data and the seismic data to be denoised, resulting in multiple replacement data sets, including: The replacement data is obtained by performing data replacement on the preprocessed denoised data and the seismic data to be denoised using the random mask matrix according to the following formula. : in, For the i-th replacement data, Let y be the i-th random mask matrix, y be the seismic data to be denoised, and I be the identity matrix. For the preprocessed denoised data, This is for the Hadamard product operation.
10. The method according to claim 9, characterized in that, The multiple intermediate denoising results are fused to obtain the final denoised data, including: The final denoised data is calculated using the following fusion formula. : in, This is the mapping function for the pre-trained blind spot network. Let be the i-th intermediate denoising result, and n represent the number of random mask matrices.
11. A data denoising device, characterized in that, include: The data reconstruction module is used to acquire seismic data to be denoised, perform noise characteristic analysis on the seismic data to be denoised, and obtain spatial correlation information of noise in the seismic data to be denoised. An asymmetric downsampling module is used to process the seismic data to be denoised using an asymmetric downsampling strategy based on the spatial correlation information, to obtain the downsampled data to be denoised; wherein, the asymmetric downsampling strategy uses different sampling step sizes in the horizontal and vertical directions; The blind spot network denoising module is used to input the downsampled data to be denoised into a pre-trained blind spot network to obtain preliminary denoised data; The preprocessing module is used to perform inverse downsampling on the preliminary denoised data to restore the original data dimension of the seismic data to be denoised, thereby obtaining preprocessed denoised data. A random substitution module is used to generate multiple random mask matrices, and to use the random mask matrices to replace the preprocessed denoised data and the seismic data to be denoised, thereby obtaining multiple replacement data. The blind spot network denoising module is further configured to input the multiple replacement data into the pre-trained blind spot network to obtain multiple intermediate denoising results, and to fuse the multiple intermediate denoising results to obtain the final denoised data.
12. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing programs or instructions that can run on the processor, the programs or instructions being executed by the processor to implement the steps of the data denoising method as described in any one of claims 1 to 10.
13. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the data denoising method as described in any one of claims 1 to 10.
14. A computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions that, when executed by a computer, cause the computer to perform the steps of the data denoising method as claimed in any one of claims 1 to 10.