A blind denoising method for seismic data with adaptive distribution estimation constraints

Abstract

The present application discloses a blind denoising method for seismic data with adaptive distribution estimation constraints, comprising the following steps: S1, constructing an initial noise distribution prediction network model and an initial generation diffusion denoising network model; S2, collecting noisy seismic data to obtain an ideal noise distribution prediction network model; S3, inputting a test data set in the noisy seismic data into the ideal noise distribution prediction network model for processing to obtain noise distribution estimation data; S4, passing the noise distribution estimation data through an adaptive noise distribution matching algorithm to obtain a matching time step; S5, inputting the test data set, the noise distribution estimation data and the matching time step together into the initial generation diffusion denoising network model to obtain an ideal generation diffusion denoising network model; S6, inputting the test data set into the ideal generation diffusion denoising network model to obtain high signal-to-noise ratio seismic data.

Description

Technical Field

[0001] The present application relates to the technical field of geophysics and artificial intelligence, and in particular to a blind denoising method for seismic data with adaptive distribution estimation constraints. Background Art

[0002] Seismic noise suppression is a classic problem in the field of seismic data processing, and seismic data with a high signal-to-noise ratio is also an important guarantee for improving the quality of subsequent processing. Although many studies have focused on this and proposed a variety of effective methods to improve the signal-to-noise ratio of seismic data, as seismic exploration shifts to deep formations, the complexity and diversity of seismic data increases, and finding a more effective seismic noise suppression method remains an important research topic.

[0003] According to the mathematical characteristics of seismic signals, a variety of traditional seismic data denoising methods have been developed. Classical methods can be roughly divided into filtering-based, sparse transformation-based and rank reduction-based methods from the underlying logic. The filtering-based seismic data denoising method uses the sorting of effective signals and noise in the time domain or transform domain to separate signals and noise by designing filters. However, its denoising effect is overly dependent on the sorting of effective signals and noise in the corresponding domain of the filter. The sparse transformation-based seismic denoising method is based on the theory of compressed sensing, projects the data into the transform domain, and uses the difference in the properties of signals and noise in it to separate them through a threshold function. However, for complex seismic data, a good denoising effect generally requires transformation from multiple scales and directions, which usually brings greater computational complexity. The rank reduction-based method assumes that seismic signals are low-rank, while random noise will increase the rank of the data matrix. By constructing seismic data into matrices such as Hankel or Toeplitz, different characteristics of signals and noise in the matrix are obtained, and then they are separated. In general, traditional seismic data denoising methods usually have good interpretability and low computational complexity, but complex parameter adjustments, limited generalization, and performance fluctuations when faced with complex distributions limit the application of such methods.

[0004] With the development of computer hardware, deep learning methods have made great progress in various fields. In order to better suppress the noise in the signal, convolutional neural networks with strong fitting ability have been introduced into the field of seismic data processing. However, seismic data lacks noise-free true values, which leads to that many deep learning methods cannot be well applied to seismic data noise suppression. The generative diffusion model, as a deep learning method that has developed rapidly in recent years, has great application prospects in the field of seismic signal processing. However, the high computational complexity of the inverse diffusion process of the diffusion model and the need for true value data constraints in the training process make its application in the field of seismic denoising also challenging.

[0005] Therefore, in order to solve the problems existing in the above-mentioned prior art, a seismic data blind denoising method with adaptive distribution estimation constraints is proposed. Summary of the invention

[0006] An embodiment of the present application provides a blind denoising method for seismic data with adaptive distribution estimation constraints, which is used to solve the problem that seismic data lacks noise-free true values, resulting in many deep learning methods not being well applied to seismic data noise suppression. In addition, the inverse diffusion process of the existing diffusion model has high computational complexity and the training process requires true value data constraints, resulting in poor technical problems in terms of seismic data denoising performance.

[0007] In view of this, the present application provides a blind denoising method for seismic data with adaptive distribution estimation constraints, comprising the following steps:

[0008] S1, constructing the initial noise distribution prediction network model and the initial generation diffusion denoising network model;

[0009] S2, collecting noisy seismic data, and dividing the noisy seismic data into a training data set and a test data set, inputting the training data set into the initial noise distribution prediction network model for training, and obtaining an ideal noise distribution prediction network model;

[0010] S3, inputting the test data set into the ideal noise distribution prediction network model for processing to obtain noise distribution estimation data;

[0011] S4, matching the noise distribution estimation data to the planned time step of the initial generation diffusion denoising network model in the generation diffusion denoising network diffusion process through an adaptive noise distribution matching algorithm to obtain a matching time step;

[0012] S5, inputting the training data set, the noise distribution estimation data and the matching time step into the initial generation diffusion denoising network model to obtain an ideal generation diffusion denoising network model;

[0013] S6. Input the test data set into the ideal generation diffusion denoising network model to obtain high signal-to-noise ratio seismic data.

[0014] Optionally, in step S2, the training data set is input into the initial noise distribution prediction network model for training, and the method steps for obtaining the ideal noise distribution prediction network model are:

[0015] Inputting the training data set into the initial noise distribution prediction network model, and performing data enhancement processing on the training data set using a snake sampling mechanism to obtain an enhanced training data set;

[0016] Then, through the optimization strategy without truth value constraints, the intrinsic data characteristics of the enhanced training data set are adaptively extracted and the parameters of the noise distribution prediction network model are gradually optimized to obtain the ideal noise distribution prediction network model.

[0017] Optionally, the formula of the optimization strategy without truth value constraints is:

[0018]

[0019] in represents the objective function, represents a parameterizable initial noise distribution prediction network model, represents the training target data, Represents the training input data, where express and They never intersect.

[0020] Optionally, in step S4, the noise distribution estimation data is matched to the planned time step of the initial generation diffusion denoising network model in the generation diffusion denoising network diffusion process through an adaptive noise distribution matching algorithm, and the method steps for obtaining the matching time step are:

[0021] The probability density function of the planned noise addition and the noise distribution estimation is calculated by the kernel density estimation method;

[0022] Then, by optimizing the relative entropy of the probability density function, the matching time step is obtained. The optimization process formula is:

[0023]

[0024] in, represents the relative entropy, represents the kernel density estimate, represents the training target data, represents the training input data, represents the ideal noise distribution prediction network model, Represents the cumulative noise attenuation factor defined in the generated diffusion denoising network model, represents noise that follows a multivariate normal distribution, Used to calculate the time step in the Markov chain Ideal noise distribution prediction network model An estimate of the noise distribution, To calculate the Markov chain Always plan to add noise.

[0025] Optionally, in step S5, the training data set, the noise distribution estimation data and the matching time step are input together into the initial generative diffusion denoising network model, and the method steps for obtaining the ideal generative diffusion denoising network model are as follows:

[0026] Inputting the training data set, the noise distribution estimation data and the matching time step into the initial generation diffusion denoising network model;

[0027] Then, the initial generative diffusion denoising network model is effectively constrained by an adaptive optimization algorithm without explicit label constraints and the adaptive noise distribution matching algorithm, and the generative diffusion denoising network model is optimized and trained in combination with the multi-scale characteristics of the diffusion process to obtain an ideal generative diffusion denoising network suitable for high noise environments.

[0028] Optionally, the adaptive optimization algorithm formula without explicit label constraints is:

[0029]

[0030] in represents the objective function, represents a parameterizable initial generation diffusion denoising network model, Represents the target data, Represents the input data in the diffusion process The specific calculation method is ,in Represents the cumulative noise attenuation factor defined in the generated diffusion denoising network model, represents the ideal noise distribution prediction network model, represents the noise distribution corresponding to the matching time step, Represents input data.

[0031] Optionally, in step S6, the test data set is input into the ideal generative diffusion denoising network model to obtain high signal-to-noise ratio seismic data by the following method steps:

[0032] Inputting the test data set into the ideal generative diffusion denoising network model;

[0033] The sampling operation is performed from the matching time step, and high signal-to-noise ratio seismic data is obtained through an adaptive constrained sampling process, wherein the adaptive constrained sampling process formula is:

[0034]

[0035] in represents the ideal generated diffusion denoising network obtained through training in step S5, Represents the time step The signal retention ratio is Represents the time step The ratio of noise added when Represents the cumulative noise attenuation factor, used to represent the time step To the current time step The accumulated signal attenuation, represents a random variable that follows a standard normal distribution, It is the additive noise part, which is used to introduce randomness to maintain sample diversity.

[0036] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0037] The present application provides a blind denoising method for seismic data with adaptive distribution estimation constraints. The method generates a diffusion denoising network model through adaptive noise distribution prediction constraints, thereby adding conditional information to the diffusion model in training without truth value constraints. The data features in the original seismic data can be more effectively utilized, thereby improving the denoising performance of the overall method, and improving the effectiveness and significance of seismic data processing. The method shows outstanding advantages in protecting and restoring detailed information of seismic data, and shows extremely high reliability when processing weak signals in complex noise environments, providing a strong guarantee for accurately capturing key feature data information in seismic data. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly express the technical solutions of the embodiments of the present application, the drawings required for describing the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0039] Figure 1 A step diagram of a blind denoising method for seismic data with adaptive distribution estimation constraints provided in an embodiment of the present application;

[0040] Figure 2 A data flow diagram between the steps of a blind denoising method for seismic data with adaptive distribution estimation constraints provided in an embodiment of the present application;

[0041] Figure 3 Schematic diagram of the noise distribution prediction network training process;

[0042] Figure 4 Schematic diagram of the training process for the diffusion denoising network;

[0043] Figure 5 Schematic diagram of the noise distribution prediction network and the diffusion denoising network structure;

[0044] Figure 6 Schematic diagram of the residual module composition in the noise distribution prediction network;

[0045] Figure 7 Schematic diagram of the residual module composition in the generated diffusion denoising network;

[0046] Figure 8 Schematic diagram of the structure of the self-attention module in the noise distribution prediction network and the generation diffusion denoising network;

[0047] Fig. 9 Comparison of the experimental results of synthetic seismic records, (a) is a pure signal, (b) is a noisy signal, (c)-(g) correspond to the denoising results of SOSVMF, K-SVD, GAN, DIP and the method of the present invention respectively;

[0048] Fig.10 Comparison diagram of the residuals corresponding to the denoising results of synthetic seismic records, (a) is the actual added noise, (b)-(f) correspond to the denoising residuals of SOSVMF, K-SVD, GAN, DIP and the method of the present invention respectively;

[0049] Fig.11 Comparison of the experimental results of a single profile of synthetic seismic records, (a) is a pure signal, (b) is a noisy signal, (c)-(g) correspond to the denoising results of SOSVMF, K-SVD, GAN, DIP and the method of the present invention, respectively;

[0050] Fig.12 This is an enlarged comparison diagram of the local area 1 of the single-section experimental results of synthetic seismic records, (a) is a pure signal, (b) is a noisy signal, (c)-(g) correspond to the denoising results of SOSVMF, K-SVD, GAN, DIP and the method of the present invention respectively;

[0051] Fig.13 This is an enlarged comparison diagram of the local area 2 of the single-section experimental results of synthetic seismic records, (a) is a pure signal, (b) is a noisy signal, (c)-(g) correspond to the denoising results of SOSVMF, K-SVD, GAN, DIP and the method of the present invention respectively;

[0052] Fig.14 The comparison diagram of denoising results of real seismic records, (a) is a noisy signal, (b)-(f) correspond to the denoising results of SOSVMF, K-SVD, GAN, DIP and the method of the present invention respectively;

[0053] Fig.15 Comparison diagram of residuals corresponding to denoising results of real seismic records, (a)-(e) correspond to the denoising residuals of SOSVMF, K-SVD, GAN, DIP and the method of the present invention, respectively;

[0054] Fig.16 The comparison diagram of the experimental results of a single profile of a real earthquake record, (a) is a noisy signal, (b)-(f) correspond to the denoising results of SOSVMF, K-SVD, GAN, DIP and the method of the present invention respectively;

[0055] Fig.17 This is a magnified comparison diagram of the local area 1 of the single-profile experimental results of real earthquake records. (a) is a noisy signal, and (b)-(f) correspond to the denoising results of SOSVMF, K-SVD, GAN, DIP and the method of the present invention, respectively;

[0056] Fig.18 This is an enlarged comparison diagram of the local area 2 of the single-profile experimental results of real earthquake records. (a) is a noisy signal, and (b)-(f) correspond to the denoising results of SOSVMF, K-SVD, GAN, DIP and the method of the present invention, respectively. DETAILED DESCRIPTION

[0057] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0058] In order to understand the technical solution provided by the present invention, please refer to Figure 1 To further understand the relationship between the steps and the data flow, see Figure 2 The present invention provides a blind denoising method for seismic data with adaptive distribution estimation constraints, comprising the following steps:

[0059] S1, constructing the initial noise distribution prediction network model and the initial generation diffusion denoising network model;

[0060] S2, collecting noisy seismic data, and dividing the noisy seismic data into a training data set and a test data set, inputting the training data set into the initial noise distribution prediction network model for training, and obtaining an ideal noise distribution prediction network model;

[0061] Specifically, Figure 3 As shown in the figure, the training data set is input into the initial noise distribution prediction network model, and the serpentine sampling mechanism is used to perform data enhancement processing on the training data set to obtain the enhanced training data set; then, the intrinsic data characteristics of the enhanced training data set are adaptively extracted through the optimization strategy without truth value constraints, and the parameters of the noise distribution prediction network model are gradually optimized to obtain the ideal noise distribution prediction network; the formula of the optimization strategy without truth value constraints is:

[0062]

[0063] in represents the objective function, which is the mean square error between the model output and the target value. represents a parameterizable initial noise distribution prediction network model, represents the training target data, Represents the training input data, where express and The purpose of the formula is to learn the parameters , minimize the objective function , so that the model The output is as close to the target value as possible . and The basis of constant non-intersection is collection and The samplers of are always not in the same partition in the training dataset. Represents the division of the training data set. Defined as ,in Represents the sampler. During the training process, in order to ensure and Constantly do not intersect, when Select a specific As training target data When training the input data Sampler The selection formula is:

[0064]

[0065] in represents the Euclidean distance, Represents the training target data The sampler, Represents the training input data A sampler, where Indicates the selectable sampling range, express Must come from the partitioning of the training set , but not itself, so and The purpose of this formula is to pass the sampler from Optimize selection samples to the target The sum of the Euclidean distances is minimal.

[0066] The serpentine sampling mechanism is as follows: the sampler moves along a three-dimensional serpentine path in the three data dimensions of the training data set; when the sampler completes the sampling of a main survey line, it only moves one sampling step along the direction of the connecting survey line to the adjacent main survey line, and samples in the opposite direction; when a time profile is sampled, it moves one sampling step along the two-way time axis to the next time profile for sampling; this sampling method ensures that the data blocks obtained from the previous and next samples are continuous in the training data set, enhancing the potential statistical characteristics of the input data.

[0067] The selection of target data and input data must satisfy the requirement that they are always disjoint. Figure 3 As shown in Figure 1, when the target data is determined, the input data must come from the same training data set, but cannot be the target data itself, and the sum of the Euclidean distances from the selected input data to the target data must be minimized.

[0068] S3, then input the test data set into the ideal noise distribution prediction network model for processing to obtain noise distribution estimation data;

[0069] S4, matching the noise distribution estimation data to the planned time step of the initial generation diffusion denoising network model in the generation diffusion denoising network diffusion process through an adaptive noise distribution matching algorithm, and obtaining a matching time step;

[0070] Specifically, in S4, the noise distribution estimation data is matched to the planned time step of the generative diffusion denoising network model in the diffusion process of the generative diffusion denoising network through an adaptive noise distribution matching algorithm, and the method steps for obtaining the matching time step are as follows:

[0071] The probability density function of the planned noise addition and the noise distribution estimation is calculated by the kernel density estimation method;

[0072] Then, by optimizing the relative entropy of the probability density function, the matching time step is obtained. The optimization process formula is:

[0073]

[0074] in, represents the relative entropy, represents the kernel density estimate, represents the training target data, represents the training input data, represents the ideal noise distribution prediction network model, Represents the cumulative noise attenuation factor defined in the generated diffusion denoising network model, represents noise that follows a multivariate normal distribution, Used to calculate the time step in the Markov chain Ideal noise distribution prediction network model An estimate of the noise distribution, To calculate the Markov chain The purpose of the formula is to find the time step that best matches the estimated data with the noise distribution by minimizing the relative entropy. , that is, matching time steps.

[0075] S5, inputting the training data set, the noise distribution estimation data and the matching time step into the initial generation diffusion denoising network model to obtain the ideal generation diffusion denoising network model;

[0076] Specifically, Figure 4 As shown, the training data set, the noise estimation data obtained by the noise distribution prediction network and the matching time step calculated by S4 are input together into the generative diffusion denoising network model; then the initial generative diffusion denoising network model is effectively constrained by the adaptive optimization algorithm without explicit label constraints and the adaptive noise distribution matching algorithm, and the generative diffusion denoising network model is optimized and trained in combination with the multi-scale characteristics of the diffusion process to obtain an ideal generative diffusion denoising network suitable for high noise environments.

[0077] The formula of the adaptive optimization algorithm without explicit label constraints is:

[0078]

[0079] in represents the objective function, which is the mean square error between the model output and the target value. represents the parameterizable initial generation diffusion denoising network model, Represents the target data, Represents the input data in the diffusion process The diffusion state of The specific calculation formula is:

[0080]

[0081] in Represents the cumulative noise attenuation factor defined in the generated diffusion denoising network model, represents the ideal noise distribution prediction network model, represents the noise distribution corresponding to the matching time step, represents the input data. The purpose of the formula is to learn the parameters , minimize the objective function , so that the model The output is as close to the target value as possible .in The calculation is as follows Figure 4As shown in the process starting from the matching time step, it involves inverse noise distribution, spatial scrambling, The specific calculation formula is:

[0082]

[0083] in represents the inverse noise distribution in the Markov chain, represents the ideal noise distribution prediction network model, and the subtraction of the numerator corresponds to Figure 4 The residual in . Represents the time step The signal retention ratio is represents the matching time step calculated in S4. represents the Markov chain from the initial to the time step The cumulative noise attenuation factor. It represents a spatial random scrambling operation, which is used to prevent the denoising diffusion probability model from falling into the solution space of the noise distribution prediction network, so that the denoising diffusion probability model learns the noise distribution characteristics instead of simply learning the explicit noise model.

[0084] S6, then input the test data set into the ideal generation diffusion denoising network model to obtain high signal-to-noise ratio seismic data, and achieve noise suppression and denoising of the three-dimensional seismic data;

[0085] Specifically, in step S6, the test data set is input into the ideal generative diffusion denoising network model to obtain high signal-to-noise ratio seismic data in the following steps:

[0086] Inputting the test data set into the ideal generative diffusion denoising network model;

[0087] The sampling operation is performed from the matching time step, and high signal-to-noise ratio seismic data is obtained through an adaptive constrained sampling process, wherein the adaptive constrained sampling process formula is:

[0088]

[0089] in represents the ideal generated diffusion denoising network obtained through training in step S5, Represents the time step The signal retention ratio is Represents the time step The ratio of noise added when Represents the cumulative noise attenuation factor, used to represent the time step To the current time step The accumulated signal attenuation, represents a random variable that follows a standard normal distribution, It is the additive noise part, which is used to introduce randomness to maintain sample diversity.

[0090] Furthermore, the initial noise distribution prediction network model and the initial generation diffusion denoising network model constructed in step S1 and step S4 are composed of:

[0091] like Figure 5 As shown in the figure, the initial noise distribution prediction network model and the initial generation diffusion denoising network model are both composed of a residual module, a downsampling module, a self-attention module and an upsampling module. Jump connections are introduced in the network sampling and reconstruction stages to enhance the ability to represent features.

[0092] like Figure 6 As shown in the figure, the residual module of the initial noise distribution prediction network model is connected by two convolutional layers. The input features of the first convolutional layer and the output features of the second convolutional layer are fused as the output of the residual module. Before the feature fusion operation, a channel number consistency check is performed. If inconsistent, the input channel number is adjusted through a 1x1 convolutional layer to be consistent with the output channel number before feature fusion.

[0093] like Figure 7 As shown in the figure, the residual module of the initial generation diffusion denoising network model is composed of a time encoding module and two convolutional layers. The input features of the first convolutional layer and the output features of the second convolutional layer are fused as the output of the residual module. Before the feature fusion operation, a channel number consistency check is performed. If inconsistent, the input channel number is adjusted through a 1x1 convolutional layer to be consistent with the output channel number before feature fusion; the time encoding module accepts integer time steps, maps the time steps to high-dimensional continuous space through sine and cosine position encoding to obtain the representation of the time steps in high-dimensional space. After the representation passes through the bottleneck layer and the Swish activation function, it is connected to a fully connected layer and feature fused with the output features of the first convolutional layer of the residual module, so that the generation diffusion denoising network model can introduce global time information in different time steps.

[0094] like Figure 8 As shown in the figure, the self-attention module of the initial noise distribution prediction network model and the initial generated diffusion denoising network model maps the input features into the query matrix Q, the key matrix K and the value matrix V through a 1x1 convolution layer, and then calculates the relevance score according to the dot product of the query matrix Q and the key matrix K, and applies a scale factor transformation to the relevance score to control the numerical range, and then normalizes the relevance score through the Softmax function to generate an attention weight matrix, and then performs a weighted summation on each distribution of the value matrix V according to the attention weight matrix to generate output features containing context information, and fuses the output features with the input features, thereby introducing self-attention information into the input features.

[0095] The downsampling modules of the initial noise distribution prediction network model and the initial generation diffusion denoising network model include 1 convolution layer, and the sampling stride is set to 2.

[0096] The upsampling modules of the initial noise distribution prediction network module and the generated diffusion denoising network module are connected by 1 upsampling layer and 1 convolution layer, and the scale of the upsampling layer is set to 2.

[0097] Furthermore, the composition of the initial noise distribution prediction network model and the generated diffusion denoising network model is specifically as follows:

[0098] Both the initial noise distribution prediction network model and the initial generation diffusion denoising network model are composed of a 43-layer structure. The first layer is the input convolution layer. The 2nd to 15th layers are connected in sequence by two residual modules and one downsampling module, and one self-attention module is added after the 11th and 13th layers. The 16th to 23rd layers are connected in sequence by residual modules, and one self-attention module is added after the 18th layer. The 24th to 42nd layers are connected in sequence by one upsampling module and three residual modules, and one self-attention module is added after the 25th, 27th and 29th layers, and the 43rd layer is the output convolution layer.

[0099] Furthermore, the skip connections used by the initial noise distribution prediction network model and the initial generation diffusion denoising network model are as follows:

[0100] The jump connections of the initial noise distribution prediction network model and the initial generation diffusion denoising network model are that the inputs of the 2nd to 10th layers are respectively connected with the outputs of the 41st to 39th, 37th to 35th, and 33th to 31st layers to form jump connections as the inputs of the 42nd to 40th, 38th to 36th, and 34th to 32nd layers, and the inputs of the 11th, 13th, 15th to 18th layers are respectively connected with the outputs of the 28th, 26th, 24th, and 22th to 20th layers to form jump connections as the inputs of the 29th, 27th, 25th, and 23rd to 21st layers.

[0101] Furthermore, the jump connection adopts the form of channel splicing to splice features of different levels in the channel dimension to form new features.

[0102] Furthermore, the sizes of the convolution filters in the convolution layers of the initial noise distribution prediction network model and the initial generated diffusion denoising network model are both set to 3×3, the Swish activation function is used after the convolution layer, and random inactivation operations and normalization processing are introduced after the convolution layer.

[0103] Furthermore, the feature fusion method adopted by the residual structure in the initial noise distribution prediction network model and the initial generation diffusion denoising network model is element-by-element addition.

[0104] Furthermore, the input features of the residual structure in the initial noise distribution prediction network model and the initial generated diffusion denoising network model are first batch normalized, passed through the Swish activation function, and then input into a 3×3 convolutional layer after random inactivation to alleviate the gradient vanishing problem, improve the nonlinear expression ability of the model, and reduce the risk of model overfitting.

[0105] Furthermore, the parameters of the initial noise distribution prediction network model and the initial generation diffusion denoising network model are adjusted through the L2 loss function.

[0106] Furthermore, the training process of the initial noise distribution prediction network model and the initial generation diffusion denoising network model both uses Adam with a fixed The learning rate is used to optimize the parameters.

[0107] Furthermore, in each training iteration, batch sample data is used to calculate the gradient as the basis for parameter update, and the batch size is 16.

[0108] Furthermore, the input training data set data is normalized before model training to reduce the impact of the input data range on model performance.

[0109] In order to prove the denoising ability of the seismic data blind denoising method with adaptive distribution estimation constraints in the present invention, structure-oriented space-varying median filtering (SOSVMF), K-singular value decomposition (K-SVD), generative adversarial network (GAN), and deep image prior (DIP) are selected for comparison with the method described in the present invention.

[0110] The comparative experiments were conducted on synthetic seismic datasets and real seismic datasets. For the synthetic dataset, a karst dataset simulating the ancient karst system was selected. Its data form is as follows: Fig. 9 As shown in (a), its reflection signal usually appears as depression or discontinuity, which makes it difficult to recover its details using seismic denoising methods.

[0111] The synthetic seismic data set is originally noise-free. In the experiment, random noise is added to it to construct a noisy synthetic seismic record. Relying on the existence of noise-free true value data, quantitative evaluation indicators can be calculated to quantitatively measure the performance of each method. In the experiment, the signal-to-noise ratio (SNR), structural similarity (SSIM), and mean squared error (MSE) of the denoised data of each method relative to the noisy data were calculated.

[0112] Table 1 Comparison of denoising effects of different methods (SNR unit: dB)

[0113]

[0114] Table 1 shows the quantitative comparison results of the denoising effects of various methods on synthetic seismic records with added noise. It can be seen that the method of the present invention can effectively improve the signal-to-noise ratio of noisy records, and better protect the effective signal morphology during the denoising process, with good structural similarity. Among the three quantitative evaluation indicators, the method of the present invention has obtained better performance indicators.

[0115] In order to more comprehensively compare the denoising performance of each method under different noise intensities and prove the blind denoising property of the present invention, that is, without prior knowledge or adjustment of algorithm parameters, its performance is still robust under the condition of noise intensity fluctuations. Noise of different intensities is added to the synthetic seismic records to obtain multiple sets of comparative experimental results.

[0116] Table 2 Comparison of denoising effects of different methods on seismic records with relatively slight noise (SNR unit: dB)

[0117]

[0118] Table 3 Comparison of denoising effects of different methods on seismic records with strong noise (SNR unit: dB)

[0119]

[0120] Table 2 shows the quantitative comparison of the denoising effect of each method in seismic records with relatively mild noise, and the signal-to-noise ratio of the noisy record is 1.988 dB. Table 3 shows the quantitative comparison of the denoising effect of each method in seismic records with relatively strong noise, and the signal-to-noise ratio of the noisy record reaches -8.251 dB. It can be seen that the method proposed in the present invention can achieve effective noise suppression when processing seismic data containing noise of different intensities.

[0121] Fig. 9 The denoising results of each method on earthquake records with strong noise are shown. Fig. 9The seismic signal processed by the method of the present invention shown in (g) is compared with Fig. 9 The noisy signal shown in (b) has an obvious noise removal effect, and the denoised data shape is similar to Fig. 9 The pure signal shown in (a) is the closest, proving that it effectively removes noise from the signal while protecting the valid information in the signal.

[0122] The protection effect of the method of the present invention on the effective signal can be Fig.10 This was further confirmed in . Fig.10 The residual graphs corresponding to the denoising results of each method are shown. It can be seen that there is no obvious effective signal leakage in the data residual after denoising by the method of the present invention.

[0123] Since the seismic records used in the experiment are three-dimensional data, a single sub-image can display its three sections, but it is not convenient to observe the details. In order to more intuitively display the details of the data after denoising by each method, the single section data in the three-dimensional data is displayed separately, such as Fig.11 As shown. Fig.11 The two regions of interest in the image are enlarged to observe the details of the denoised data. The results are as follows: Fig.12 and Fig.13 The denoising performance of the method of the present invention can be more clearly observed from the cross-sectional view and the detailed magnified view.

[0124] For the real seismic data set, the present invention selects the F3 Netherlands data set, which is a public seismic data set. The original data consists of 650 main survey lines and 950 connecting survey lines to form a three-dimensional seismic record. The seismic records of the experiment mainly contain reflection layers of the Miocene, Pliocene and Pleistocene.

[0125] The real earthquake record form is as follows Fig.14 As shown in (a), Fig.14 (b)-(f) show the qualitative denoising results of each comparison method. It can be seen that the method of the present invention effectively removes noise while retaining rich stratigraphic result details.

[0126] Fig.15 The residual graphs corresponding to the denoising results of each method are shown. The data denoised by the method of the present invention has less signal leakage and no obvious structural signal leakage.

[0127] In order to more intuitively show the details of the effects of each denoising method, the single profile record in the real 3D seismic record is displayed separately, such as Fig.16 As shown. Fig.16 The two regions of interest in the image are enlarged to observe the details of the denoised data. The results are as follows: Fig.17 and Fig.18As shown in the enlarged figure, it can be more clearly observed that the method of the present invention retains the most detailed stratum structure, and can maintain the integrity of the signal morphology while effectively suppressing the noise in the seismic signal, without blurring or deformation of the effective signal.

[0128] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A blind denoising method for seismic data with adaptive distribution estimation constraints, characterized in that: The following steps are involved: S1, constructing the initial noise distribution prediction network model and the initial generation diffusion denoising network model; S2, collecting noisy seismic data, and dividing the noisy seismic data into a training data set and a test data set, inputting the training data set into the initial noise distribution prediction network model for training, and obtaining an ideal noise distribution prediction network model; S3, inputting the test data set into the ideal noise distribution prediction network model for processing to obtain noise distribution estimation data; S4, matching the noise distribution estimation data to the planned time step of the initial generation diffusion denoising network model in the generation diffusion denoising network diffusion process through an adaptive noise distribution matching algorithm to obtain a matching time step; S5, inputting the training data set, the noise distribution estimation data and the matching time step into the initial generation diffusion denoising network model to obtain an ideal generation diffusion denoising network model; S6. Input the test data set into the ideal generation diffusion denoising network model to obtain high signal-to-noise ratio seismic data.

2. The method for blind denoising of seismic data with adaptive distribution estimation constraints according to claim 1, characterized in that: In step S2, the training data set is input into the initial noise distribution prediction network model for training, and the method steps for obtaining the ideal noise distribution prediction network model are as follows: Inputting the training data set into the initial noise distribution prediction network model, and performing data enhancement processing on the training data set using a snake sampling mechanism to obtain an enhanced training data set; Then, through the optimization strategy without truth value constraints, the intrinsic data features of the enhanced training data set are adaptively extracted and the parameters of the noise distribution prediction network model are gradually optimized to obtain the ideal noise distribution prediction network model.

3. The blind denoising method for seismic data with adaptive distribution estimation constraints according to claim 2, characterized in that: The formula of the optimization strategy without truth value constraints is: in represents the objective function, represents a parameterizable initial noise distribution prediction network model, represents the training target data, Represents the training input data, where express and They never intersect.

4. The method for blind denoising of seismic data with adaptive distribution estimation constraints according to claim 1, characterized in that: In step S4, the noise distribution estimation data is matched to the planned time step of the initial generation diffusion denoising network model in the generation diffusion denoising network diffusion process through an adaptive noise distribution matching algorithm, and the method steps for obtaining the matching time step are as follows: The probability density function of the planned noise addition and the noise distribution estimation is calculated by the kernel density estimation method; Then, by optimizing the relative entropy of the probability density function, the matching time step is obtained. The optimization process formula is: in, represents the relative entropy, represents the kernel density estimate, represents the training target data, represents the training input data, represents the ideal noise distribution prediction network model, Represents the cumulative noise attenuation factor defined in the generated diffusion denoising network model, represents noise that follows a multivariate normal distribution, Used to calculate the time step in the Markov chain Ideal noise distribution prediction network model An estimate of the noise distribution, To calculate the Markov chain Always plan to add noise.

5. The method for blind denoising of seismic data with adaptive distribution estimation constraints according to claim 1, characterized in that: In step S5, the training data set, the noise distribution estimation data and the matching time step are inputted into the initial generation diffusion denoising network model together, and the method steps for obtaining the ideal generation diffusion denoising network model are as follows: Inputting the training data set, the noise distribution estimation data and the matching time step into the initial generation diffusion denoising network model; Then, the initial generative diffusion denoising network model is effectively constrained by an adaptive optimization algorithm without explicit label constraints and the adaptive noise distribution matching algorithm, and the generative diffusion denoising network model is optimized and trained in combination with the multi-scale characteristics of the diffusion process to obtain an ideal generative diffusion denoising network suitable for high noise environments.

6. The method for blind denoising of seismic data with adaptive distribution estimation constraints according to claim 5, characterized in that: The formula of the adaptive optimization algorithm without explicit label constraints is: in represents the objective function, represents the parameterizable initial generation diffusion denoising network model, Represents the target data, Represents the input data in the diffusion process The specific calculation method is ,in Represents the cumulative noise attenuation factor defined in the generated diffusion denoising network model, represents the ideal noise distribution prediction network model, represents the noise distribution corresponding to the matching time step, Represents input data.

7. The method for blind denoising of seismic data with adaptive distribution estimation constraints according to claim 1, characterized in that: In step S6, the test data set is input into the ideal generation diffusion denoising network model to obtain high signal-to-noise ratio seismic data in the following steps: Inputting the test data set into the ideal generative diffusion denoising network model; The sampling operation is performed from the matching time step, and high signal-to-noise ratio seismic data is obtained through an adaptive constrained sampling process, wherein the adaptive constrained sampling process formula is: in represents the ideal generated diffusion denoising network obtained through training in step S5, Represents the time step The signal retention ratio is Represents the time step The ratio of noise added when Represents the cumulative noise attenuation factor, used to represent the time step To the current time step The accumulated signal attenuation, represents a random variable that follows a standard normal distribution, It is the additive noise part, which is used to introduce randomness to maintain sample diversity.

Images