Random noise denoising method for deep reflection seismic data based on multi-scale feature fusion

Abstract

The present application relates to the field of geophysical technology, and specifically to a method for random noise denoising of deep reflection seismic data based on multi-scale feature fusion, which is executed by a processor, and includes: generating a training set based on a deep reflection random medium forward model and seismic random noise in different frequency bands; building a training network, training the training set, and obtaining trained model parameters; performing transfer learning on the trained model parameters, and performing actual random noise denoising of deep reflection seismic data. The present application solves the problems of low quality of random noise denoising of deep reflection seismic data and overly complex and low efficiency in the prior art.

Description

Technical Field

[0001] The present application relates to the field of geophysical technology, and in particular to a method for denoising random noise of deep reflection seismic data based on multi-scale feature fusion. Background Art

[0002] The deep earth provides most of the resources and energy for human survival and development, is the material basis for the reproduction of all things, and is the root cause of major natural disasters (earthquakes, volcanoes, etc.). These are closely related to the composition, structure and dynamic processes of the earth's interior. Therefore, various methods are needed to detect the interior of the earth, especially the deep earth. Among them, deep reflection seismic detection technology is recognized as an effective means to detect deep problems such as the fine structure of the lithosphere, the basement, and the fine structure of the upper mantle. However, due to the large detection depth of the deep reflection seismic profile, the weak effective signal in the deep layer, the long profile, spanning multiple structural units, complex geological conditions and the absorption and attenuation of the strata, the signal energy in the deep layer is weak, the background interference is strong, and the signal-to-noise ratio of the data is very low. Therefore, it is of great significance to use reasonable methods to remove noise, enhance the weak reflection signal in the deep, and improve the signal-to-noise ratio to interpret the characteristics of deep structures.

[0003] In recent years, the theory of deep learning has been applied to seismic noise suppression. The deep neural network model has a powerful feature learning ability and can effectively mine rich abstract feature expressions from big data, fit the abstract features of the target information in a short computing time, and achieve accurate extraction of target features. However, this method has the following disadvantages in denoising: (1) The choice of network structure affects the denoising effect; (2) The supervised algorithm requires a large amount of noise-free and noise-containing data as training sets; (3) Noise-free data is difficult to obtain in practical applications. Therefore, it is necessary to provide a simpler and more efficient noise suppression method to solve the above problems. Summary of the invention

[0004] The present invention overcomes the defects of the prior art and provides a method for random noise denoising of deep reflection seismic data based on multi-scale feature fusion, an important purpose of which is to improve the quality and efficiency of deep reflection seismic data denoising.

[0005] To achieve the above object, the present invention provides a method for random noise denoising of deep reflection seismic data based on multi-scale feature fusion, which is executed by a processor and includes:

[0006] Generate a training set based on the deep reflection random medium forward model and seismic random noise in different frequency bands;

[0007] Build a training network, perform training on the training set, and obtain the trained model parameters;

[0008] Transfer learning is performed on the trained model parameters to perform actual random noise denoising of deep reflection seismic data.

[0009] Furthermore, the training set is generated based on the deep reflection random medium forward model and seismic random noise in different frequency bands, specifically:

[0010] Based on geological structure, a deep reflection random medium forward model is constructed;

[0011] Perform forward simulation on the deep reflection random medium forward model to obtain synthetic deep reflection seismic data;

[0012] Normalizing the synthesized deep reflection seismic data and adding seismic random noises of different frequency bands to obtain the noisy deep reflection seismic data;

[0013] The deep reflection seismic data after noise addition are subjected to edge zero padding, cutting and data enhancement to obtain a training set.

[0014] Furthermore, the training network is constructed, and training is performed on the training set to obtain trained model parameters, specifically:

[0015] Based on the training set, we build a training network with convolutional autoencoders, variational autoencoders, and attention mechanisms.

[0016] Based on the training network, construct the loss function and perform parameter updates;

[0017] Input the training set into the training network for training to obtain the trained model parameters.

[0018] Furthermore, the training network is constructed based on the training set in conjunction with a convolutional autoencoder, a variational autoencoder and an attention mechanism, and further includes:

[0019] Based on the shared layers, feature extraction is performed on the training set and local patterns in deep reflection seismic data are identified;

[0020] The training set is input into the convolution branch to extract deterministic geological features from the training set;

[0021] Input the training set into the variational branch, perform mean and log variance encoding on the data in the training set, and obtain the latent vector;

[0022] Based on geological features and latent vectors, multi-scale features are obtained;

[0023] The multi-scale features are fed into the attention mechanism from the output layer, and the multi-scale latent feature vector is output;

[0024] The decoder performs reconstruction on the multi-scale latent feature vectors to obtain reconstructed data.

[0025] Furthermore, the method of merging the convolution branch and the variational branch into an output layer further includes:

[0026] Divide the output layer into upper branch, middle branch and lower branch;

[0027] The lower branch and the middle branch transmit the feature information in the multi-scale features to the attention mechanism to obtain the feature importance weight data;

[0028] The upper branch directly inputs multi-scale features to subsequent layers.

[0029] Furthermore, the loss function is constructed based on the training network, specifically:

[0030] Based on the convolutional autoencoder CAE, the MSE loss function is constructed;

[0031] Based on variational autoencoder VAE, KL divergence loss function is constructed;

[0032] Based on the training network, the MSE loss function and the KL divergence loss function are combined to form a composite loss function. The formula of the composite loss function is:

[0033]

[0034] is the original data, is the reconstructed data, n is the amount of training set data, and Represent the probability distribution functions of the original and reconstructed data respectively; , is the weight, where:

[0035] .

[0036] Furthermore, the transfer learning is performed on the trained model parameters to perform actual random noise denoising of deep reflection seismic data, specifically:

[0037] Migrate the trained model parameters to the training network;

[0038] Perform iterative training on the training network to obtain an optimized training network;

[0039] Get the actual data;

[0040] Based on the optimized trained network, denoising is performed on the actual random noise of deep reflection seismic data.

[0041] The technical solution of the present application has at least the following advantages and beneficial effects: by generating a training set, training and migrating the training set, and then reconstructing the deep reflection seismic data, the purpose of denoising is achieved. Among them, a multi-scale feature fusion autoencoder is provided to extract the spatial features of deep reflection seismic data and capture statistical features, so as to realize the extraction of multi-scale and multi-angle features of deep reflection seismic data, and the fused feature representation can more comprehensively describe the complex patterns of deep reflection seismic data, and thus can more comprehensively describe the complex patterns of deep reflection seismic data. In addition, the features of the two branches are fused by adding an attention mechanism and a joint loss function is constructed, so that the network allocates limited learning ability to the learning of effective features, highlights the most relevant parts, enhances the features extracted by the two branches, and at the same time makes the reconstructed data undistorted and the extracted features tend to the standard normal distribution, and finally obtains a better denoising effect, thereby obtaining a better denoising effect, thereby solving the problems of low quality of random noise denoising of deep reflection seismic data and overly complex and low efficiency in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A schematic diagram of the process of random noise denoising method for deep reflection seismic data provided in this application;

[0043] Figure 2 This is a schematic diagram of the principle of an embodiment of the present application;

[0044] Figure 3 This is one of the example diagrams of the training set in the embodiment of this application;

[0045] Figure 4 This is the second example diagram of the training set in the embodiment of the present application;

[0046] Figure 5 This is a schematic diagram of the network structure of a multi-scale feature fusion autoencoder in an embodiment of the present application;

[0047] Figure 6 This is a schematic diagram of the loss function results in the embodiment of the present application;

[0048] Figure 7 This is a schematic diagram of the original data used in the comparison of denoising results in this application;

[0049] Figure 8 This is a schematic diagram of the results after wavelet transformation in the denoising results comparison of this application;

[0050] Fig. 9 This is a schematic diagram of the results after mean filtering in the denoising result comparison of this application;

[0051] Fig.10 This is a schematic diagram of the results of multi-scale feature fusion provided in this embodiment for comparison of denoising results in this application. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0053] Application Overview:

[0054] In deep reflection seismic, the types of noise are diverse and complex, which can be divided into coherent noise and random noise. Coherent noise mainly includes surface waves, acoustic waves and multiple waves, which show regularity on the seismic profile. However, random noise has no fixed frequency and apparent velocity, and shows irregularity on the profile. It mainly includes low-frequency interference, microseismic and high-frequency interference, with large intensity variation and wide frequency band. The highest frequency can reach about 150Hz. The frequency range of low-frequency interference is roughly 10~30Hz, and the frequency range of high-frequency interference is roughly 80~200Hz. The existence of random noise reduces the coherence and signal-to-noise ratio of the signal. Although superposition can suppress some random noise, the residual noise will still interfere with the interpretation of seismic data, especially for deep reflection seismic data. The existence of random noise hinders the identification of signals from seismic data. Therefore, eliminating random noise in deep reflection seismic has become a research focus.

[0055] For random noise in deep reflection seismic data, some methods use FX domain to remove random noise. This method first enhances the continuity of the event axis, and then predicts the linear event axis. The coherence of random noise is weaker than that of the effective signal. According to the difference in coherence, the random noise can be separated and suppressed. However, this method will enhance the coherent noise, and under complex geological conditions, it does not meet the conditions for the use of this method. The random noise can be removed by using the polynomial fitting method based on the similarity of the effective signal and multi-channel correlation. In addition, the combination of FX deconvolution and median filtering and the combination of wavelet transform and spatial frequency domain prediction denoising technology have also achieved good results in removing random noise. In addition, the deep reflection seismic data is decomposed with high precision by Shearlet sparse transform, and then the random noise is removed by the adaptive threshold method of angle and scale. This process does not destroy the effective signal. The stacked profile imaging effect obtained by this method is good, and the structural details of the medium and deep layers and the trend of the Moho surface are clearly presented. These methods design appropriate thresholds based on the characteristic differences between effective signals and interference noise in the transform domain, thereby suppressing the random noise in seismic data to a certain extent. However, they are highly dependent on factors such as the parameters used. Especially when dealing with seismic data under complex geological conditions, this method of manually controlling threshold parameters will have deficiencies in denoising performance and efficiency, and the denoising effect is not very ideal.

[0056] In recent years, the theory of deep learning has been applied to seismic noise suppression. The deep neural network model has a strong feature learning ability and can effectively mine rich abstract feature expressions from big data, fit the abstract features of target information in a short computing time, and achieve accurate extraction of target features. However, this method has the following disadvantages in denoising: (1) The choice of network structure affects the denoising effect; (2) The supervised algorithm requires a large amount of noise-free and noisy data as training sets; (3) It is difficult to obtain noise-free data in practical applications.

[0057] Based on the foregoing, the present application provides a random noise denoising method for deep reflection seismic data based on multi-scale feature fusion, wherein a multi-scale feature fusion autoencoder is proposed to extract the features of deep reflection seismic data to reconstruct data to achieve the purpose of denoising. In detail, the encoder can perform both convolution and variational encoding at the same time. The convolutional encoder is responsible for extracting the spatial features of deep reflection seismic data, and the variational encoder captures statistical features. These two encoders can complementarily extract the multi-scale and multi-angle features of deep reflection seismic data. The fused feature representation can more comprehensively describe the complex pattern of deep reflection seismic data, thereby obtaining a better denoising effect. On the other hand, the random noise denoising method for deep reflection seismic data of the present application also adds an attention mechanism to fuse the features of the two branches, so that the network allocates limited learning capacity to the learning of effective features, highlights the most relevant parts, and enhances the features extracted by the two branches. At the same time, a joint loss function is constructed to ensure that the reconstructed data is not distorted and the extracted features tend to be standard normal distribution.

[0058] Example

[0059] Please refer to Figure 1-Figure 10 The present application provides a random noise denoising method for deep reflection seismic data based on multi-scale feature fusion, the method being executed by a processor and comprising:

[0060] Generate a training set based on the deep reflection random medium forward model and seismic random noise in different frequency bands;

[0061] Build a training network, perform training on the training set, and obtain the trained model parameters;

[0062] Transfer learning is performed on the trained model parameters to perform actual random noise denoising of deep reflection seismic data.

[0063] It is worth noting that this embodiment provides a random noise denoising method for deep reflection seismic data based on multi-scale feature fusion, which achieves the purpose of denoising by generating a training set, training the training set, and then reconstructing the deep reflection seismic data. In detail, by extracting the spatial characteristics of the deep reflection seismic data and capturing the statistical characteristics, the multi-scale and multi-angle features of the deep reflection seismic data are extracted, which can more comprehensively describe the complex patterns of the deep reflection seismic data and finally obtain a better denoising effect.

[0064] The following is a detailed description of the application method:

[0065] In some embodiments, the training set is generated based on the deep reflection random medium forward model and seismic random noise in different frequency bands, specifically:

[0066] Based on geological structure, a deep reflection random medium forward model is constructed;

[0067] Perform forward simulation on the deep reflection random medium forward model to obtain synthetic deep reflection seismic data;

[0068] Normalizing the synthesized deep reflection seismic data and adding seismic random noises of different frequency bands to obtain the noisy deep reflection seismic data;

[0069] The deep reflection seismic data after noise addition are subjected to edge zero padding, cutting and data enhancement to obtain a training set.

[0070] It should be noted that there are at least two or more geological structures. By performing normalization in the synthetic deep reflection seismic data and then adding random noise of different frequency bands, the complex situation in the actual exploration environment can be simulated, so that the trained model performs more accurately in actual applications. Among them, when the synthetic deep reflection seismic data is normalized, the value of the synthetic deep reflection seismic data is converted to between 0 and 1, thereby reducing the difficulty of network training. By adding random noise of earthquakes of different frequency bands to the normalized deep reflection seismic data, the complex situation in the actual exploration environment can be simulated, so that the trained model performs better in real applications.

[0071] Furthermore, the training network is constructed, and training is performed on the training set to obtain trained model parameters, specifically:

[0072] Based on the training set, we build a training network with convolutional autoencoders, variational autoencoders, and attention mechanisms.

[0073] Based on the training network, construct the loss function and perform parameter updates;

[0074] Input the training set into the training network for training to obtain the trained model parameters.

[0075] It should be noted that in seismic exploration, the forward simulation of earthquakes plays a key role. It provides a reference for inferring the characteristics and properties of underground geological structures, and can help reduce the multi-solution problem in seismic data interpretation and inversion. Therefore, an accurate seismic forward model is the basis for simulation. In the prior art, in order to reduce complexity, the underground geological structure of the target area is often simplified into a uniform medium, and a forward model is constructed on this basis for research. However, according to a large number of actual case studies and drilling verification, it is found that crustal materials generally have multi-scale heterogeneity. Therefore, the forward model established based on the assumption of a uniform medium has a large deviation from the actual result and low accuracy. In addition, the heterogeneity of the underground medium is usually manifested as disturbances in seismic records, so analyzing the relationship between these disturbances and observed data is crucial for data processing and geological interpretation. This embodiment constructs a deep reflection random medium forward model, the purpose of which is to generate an accurate deep reflection seismic data set and ensure the generalization ability of the model to different structural types. The deep reflection random medium forward model is trained to obtain an inhomogeneous medium seismic forward model. Compared with the homogeneous medium seismic model in the prior art, the inhomogeneous medium seismic forward model provided by this embodiment can obtain results that are closer to reality and have higher accuracy.

[0076] It should be noted that if Figure 3 and Figure 4 , Figure 3 and Figure 4 All of them are example graphs of the training set. Specifically, the deep reflection seismic data after noise addition is cut by a sliding window. In order to effectively process the data boundary, the edge filling is performed on the deep reflection seismic data after noise addition, that is, the edge is filled by zero filling. Then, with each point on the profile as the center, a small data block of size p×p is intercepted with a step size of P / 2 to obtain n data blocks. This ensures that all the cut data blocks remain the same size, which is conducive to subsequent batch processing and model training. Each small data block can reflect local changes and trends, which helps the model learn more local details. Finally, a training set consisting of multiple data blocks is obtained. In addition, a variety of data enhancement techniques including data rotation, mirror flipping and other methods are used to further enrich the training set data, that is, the training set. Through these enhancement techniques, the training data covers more potential deformation conditions and improves the wide adaptability and generalization ability of the model. In detail, by constructing a deep reflection random medium forward model, clean deep reflection seismic data are generated, normalized, and then seismic random noise of different frequency bands is added to obtain noisy deep reflection seismic data, that is, noisy deep reflection seismic data, edge zero padding and cutting and splitting are performed to obtain small data blocks, and data enhancement technology is applied to the small data blocks to improve the generalization ability of the network model.

[0077] In some embodiments, a variety of data enhancement techniques are also used, including at least data rotation, mirroring, and flipping, to further enrich the above-mentioned optimized training data, so that the optimized training data covers more potential deformation conditions, thereby improving the wide adaptability and generalization ability of the inhomogeneous medium seismic forward model.

[0078] In some embodiments, the training network is built based on the training set, in conjunction with a convolutional autoencoder, a variational autoencoder, and an attention mechanism, and an encoder and a decoder are generated, and further includes:

[0079] Based on the shared layers, feature extraction is performed on the training set and local patterns in deep reflection seismic data are identified;

[0080] The training set is input into the convolution branch to extract deterministic geological features from the training set;

[0081] Input the training set into the variational branch, perform mean and log variance encoding on the data in the training set, and obtain the latent vector;

[0082] Based on geological features and latent vectors, multi-scale features are obtained;

[0083] The multi-scale features are fed into the attention mechanism from the output layer, and the multi-scale latent feature vector is output;

[0084] The decoder performs reconstruction on the multi-scale latent feature vectors to obtain reconstructed data.

[0085] It is worth noting that first, the small data blocks in the training set are subjected to feature extraction through the shared layer to identify local patterns in deep reflection seismic data. The shared layer includes convolutional layers, batch normalization, and ReLU activation functions: the convolutional layer extracts convolutional features, and the number of channels increases from 32 to 64, allowing the network to gradually learn more complex geological features and capture spatial structures and local features; batch normalization is applied after each convolutional layer to reduce internal covariate shifts, which helps to accelerate the training process and improve the stability of the network model; the ReLU activation function introduces nonlinearity, which allows the network to learn more complex patterns, enhances the network's expressiveness, and alleviates the gradient vanishing problem. It should be noted that the shared layer is in parallel with the convolutional branch and the variational branch.

[0086] Furthermore, the features extracted from the shared layer are input into two parallel branches respectively. The convolution branch contains multiple convolution layers, and the number of channels starts from 128 and finally decreases to 32. This branch focuses on extracting deterministic geological features in the training set. In the variational branch, the mean and logarithmic variance of the input features are encoded, and the latent vector is obtained. Specifically, the latent vector is generated by using the reparameterization technique, thereby reducing the number of channels from 128 to 64. The introduced randomness helps to deal with the inherent uncertainty of seismic data, such as noise.

[0087] Furthermore, based on geological features and latent vectors, multi-scale features are obtained. Specifically, the features extracted from the convolution branch and the variational branch are combined and fused into a multi-scale feature through a concatenation operation. The dual-branch structure provided in this embodiment allows the network to simultaneously capture the deterministic and probabilistic distribution features in seismic data.

[0088] Furthermore, the convolution branch and the variational branch are merged into the output layer, and the multi-scale features are fed into the attention mechanism from the output layer to output a multi-scale potential feature vector. In detail: After the convolution branch and the variational branch are merged into the output layer, the fused multi-scale features are processed by the attention mechanism. The attention mechanism can help the network model identify and focus on important seismic signal features in the feature space, such as strong reflections and fault structures, and further enhance the fusion of multi-scale features by learning to assign importance weights to features of different scales. After being processed by the attention mechanism, the multi-scale features are passed through the fully connected layer, batch normalization layer, and ReLU activation function, and finally the output multi-scale potential feature vector is obtained through the output. The output multi-scale potential feature vector is a 32-dimensional multi-scale potential feature vector.

[0089] Further, the decoder reconstructs the multi-scale potential feature vector to obtain reconstructed data. Specifically, the decoder reconstructs the multi-scale potential feature vector output by the encoder back to the original data space to finally obtain the reconstructed data. In detail: starting from the 32-dimensional multi-scale potential feature vector, the deep reflection seismic data with random noise removed is gradually reconstructed through a series of operations. The spatial dimensions of the multi-scale features are gradually increased by using transposed convolution (TransConv) and ReLU activation functions. The last layer of the decoder uses a Sigmoid activation function to ensure that the output value is within a reasonable range and is suitable for the amplitude representation of seismic data. The multi-scale feature fusion method provided in this embodiment enables the network to effectively process the complex geological structure in deep reflection seismic data, and everything from small-scale details (such as thin layers) to large-scale features (such as fault systems) can be fully represented and reconstructed. Through the method provided in this embodiment, the network can better capture and retain key geological information in seismic data. Please refer to Figure 5 , Figure 5 Schematic diagram of the multi-scale feature fusion autoencoder network structure.

[0090] In some embodiments, the step of merging the convolution branch and the variational branch into an output layer further includes:

[0091] Divide the output layer into upper branch, middle branch and lower branch;

[0092] The lower branch and the middle branch transmit the feature information in the multi-scale features to the attention mechanism to obtain the feature importance weight data;

[0093] The upper branch directly inputs multi-scale features to subsequent layers.

[0094] It is worth noting that after the outputs of the convolution branch and the variational branch are combined through the concatenation operation, Concatenate has three output branches, namely the upper branch, the middle branch and the lower branch: the middle branch is input to the attention mechanism to provide the main feature information to the attention mechanism and calculate the importance weight of the feature. The attention mechanism will generate a weighted feature according to the content of the input feature map, highlighting the important features and suppressing the unimportant features; the lower branch is also input to the attention mechanism to assist the attention mechanism in calculating the weights, considering a wider context or specific guidance information. By combining the information of this branch, the attention mechanism can more accurately identify and emphasize important features, and enhance the decision-making ability of the attention mechanism by providing global information; the upper branch is a skip attention mechanism, which directly passes the fused multi-scale features obtained after concatenation to the subsequent layers without passing through the attention mechanism. The purpose is to retain the original feature information and prevent possible information loss in the attention mechanism. The skip mechanism used in this embodiment is similar to the residual connection, which can help the flow of information and gradients and ensure that the network remains stable during the training process. This design provides a flexible feature processing method by combining the attention mechanism and the skip connection. The attention mechanism can focus on important features, while the jump connection ensures the integrity of information and the effective propagation of gradients. The multi-branch structure set in this embodiment allows the network to better capture and utilize multi-level information when processing complex data.

[0095] In some embodiments, the loss function is constructed based on the training network, specifically:

[0096] Based on the convolutional autoencoder CAE, the MSE loss function is constructed;

[0097] Based on variational autoencoder VAE, KL divergence loss function is constructed;

[0098] Based on the training network, the MSE loss function and the KL divergence loss function are combined to form a composite loss function. The formula of the composite loss function is:

[0099]

[0100] is the original data, is the reconstructed data, n is the amount of training set data, and Represent the probability distribution functions of the original and reconstructed data respectively; , is the weight, where:

[0101] .

[0102] It is worth noting that when using a multi-scale autoencoder, the choice of loss function is particularly critical, and a balance needs to be struck between structure preservation and computational efficiency. In the prior art, the loss function of CAE mostly uses mean squared error (MSE), and VAE uses KL divergence loss function. However, using a single loss function often fails to meet the target requirements. The seismic data denoising method based on multi-scale feature fusion provided in this embodiment, the network structure constructed by it integrates CAE and VAE, and the loss function of the network structure is constructed by combining MSE and KL divergence loss, that is, the formula of the above-mentioned composite loss function is obtained.

[0103] In detail, based on the above content, the MSE used can directly measure the difference between the reconstructed data and the original input data, providing a clear optimization goal for the network, namely minimizing the reconstruction error. The network adjusts parameters through back propagation to reduce this error. In the denoising task, MSE encourages the network to learn to distinguish between signals and noise, and the network tends to reconstruct the main features while ignoring random noise. Moreover, the gradient of MSE is not prone to disappearing or exploding, which is beneficial to the training process of the network. The loss function constructed in this embodiment not only takes into account the reconstruction quality of the overall data, but is also sensitive to local details, which helps to capture detailed information while retaining the overall structure, and is easy to be used in combination with other loss functions (such as regularization terms).

[0104] In detail, based on the above, the KL divergence loss used forces the distribution of the latent variables output by the encoder to be close to the predefined prior distribution (i.e., the standard normal distribution), which can effectively ensure that the latent space has good structure and continuity. By restricting the distribution of latent variables, the KL divergence loss plays a regularization role, which helps prevent the network model from overfitting the training data. However, the shortcomings of the KL divergence loss should also be noted: excessive regularization may cause the latent space to be too regular and lose some useful feature information. Over-emphasizing the distribution of the latent space may affect the reconstruction accuracy. It may cause gradient instability in the early stages of training. Lack of sensitivity to local structure, focusing mainly on global distribution, may ignore local details.

[0105] Furthermore, based on the above content, after combining the MSE function and the KL divergence loss function, KL divergence provides regularization to prevent overfitting, and MSE ensures high-quality reconstruction, achieving moderate regularization while maintaining the reconstruction quality; KL divergence promotes learning more abstract and decoupled features, and MSE ensures that these features can accurately reconstruct the data; MSE provides a stable gradient, KL divergence increases the regularization effect of the network model, and the training process is more stable, reducing the risk of mode collapse; MSE focuses on local details, and KL divergence helps capture global structure, which can retain both local details and global structure.

[0106] In some embodiments, the transfer learning of the trained model parameters is performed to perform random noise denoising of actual deep reflection seismic data, specifically:

[0107] Migrate the trained model parameters to the training network;

[0108] Perform iterative training on the training network to obtain an optimized training network;

[0109] Get the actual data;

[0110] Based on the optimized trained network, denoising is performed on the actual random noise of deep reflection seismic data.

[0111] It is worth noting that in order to increase the universality of this application, a transfer learning module is added after the trained network model. Through transfer learning, the trained model parameters can be transferred to the training network, and a new model is obtained. The shared knowledge between the relevant data or tasks is used to fine-tune the iterative training network, accelerate the training speed, obtain the optimized training network, and improve the generalization ability of the new model. Obtain actual data, input the actual data into the new model generated by the optimized training network, and perform deep reflection seismic data random noise denoising on the actual data. Specifically: First, the area containing key geological feature signals is intercepted from the deep reflection seismic profile. The intercepted seismic profile time is 8-8.8s, the number of sampling points is 200, and the number of channels is from 8600-8699. According to the method of constructing a training set, 40,000 data blocks are obtained; then the training batch size is set to 128, and 31,300 iterations are performed, and the learning rate is set to 0.001; this value not only helps to ensure the stable convergence of the new model in the early stage of training, but also can achieve more refined weight adjustment in subsequent iterations. By applying the network to real deep reflection seismic random noise denoising.

[0112] It should be noted that based on the above iterative process, please refer to Figure 6 , Figure 6 The loss function results obtained from training are shown in the figure. From 0 to 500 iterations: the loss value shows an extremely steep downward trend. In this stage, the loss value drops sharply to about 0.1, indicating that the model learns very quickly in the early stage and can quickly adjust parameters to reduce errors. Transition period: Between about 500 and 5000 iterations, the curve is still declining, but the speed is obviously slowing down; some small fluctuations can be observed in this stage, reflecting the fine-tuning process of the model when searching for the optimal solution. From 5000 iterations to 30000 iterations, the curve enters a long-term slow decline stage, and the decline rate of the loss value slows down further in this stage. When approaching 30000 iterations, the curve is almost completely flat, and the loss value is very close to 0, indicating that the model has reached a convergence state. There is no obvious fluctuation or rise in the entire curve, indicating that the training process is relatively stable. There is a slight oscillation in the later stage, but the amplitude is very small, and there is no obvious overfitting phenomenon.

[0113] In some embodiments, please refer to Figure 7-10 , compare different denoising methods, and evaluate the seismic data denoising method provided in this application. Specifically, wavelet transform and mean filtering denoising are performed on the same data, and the denoising results are compared with those of this application.

[0114] For details, please refer to Figure 7, is the original data used. The original data contains more noise and messy signals. The formation reflection signal is not clear enough. There is a lot of small-scale random noise. The details are covered by the noise and are difficult to distinguish. Please refer to Figure 8 , wavelet transform denoising, random noise is significantly reduced, the main stratigraphic structure and strong reflection signals are retained, the image becomes smoother as a whole, retaining more details, but some subtle geological features may be over-smoothed, but the overall signal-to-noise ratio is improved. Please refer to Fig. 9 Although the mean filter denoising reduces random noise, the effect is not as obvious as the wavelet transform. The image becomes slightly blurred, some high-frequency details may be lost, and the stratum boundaries are not clear enough, but the main stratum structure is still clearly visible. Please refer to Fig.10 The multi-feature fusion autoencoder denoising provided in this embodiment significantly reduces noise, and the stratigraphic structure, boundaries and continuity are well preserved and enhanced. The profile has high clarity and good detail preservation, and random noise is effectively removed while retaining useful information.

[0115] In summary: wavelet transform has a good denoising effect and can effectively reduce random noise, but it may over-smooth some details; mean filter denoising is relatively weak, although it reduces some random noise, it may cause the overall profile to become blurred; the multi-feature fusion autoencoder provided in this embodiment has the best effect in removing random noise, not only effectively removing random noise, but also well preserving the stratigraphic structure and details, and striking a balance between removing random noise and retaining useful geological information. The random noise denoising method for deep reflection seismic data based on multi-scale feature fusion provided in this embodiment not only provides a clearer and more detailed seismic profile, but also effectively suppresses random noise, improves the interpretability of the data, and does not introduce obvious artificial distortion.

[0116] The above shows and describes the basic principles and main features of the present application and the advantages of the present application. For those skilled in the art, it is obvious that the present application is not limited to the details of the above exemplary embodiments, and the present application can be implemented in other specific forms without departing from the spirit or basic features of the present application. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present application is defined by the attached claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present application. Any figure mark in the claims should not be regarded as limiting the claims involved.

[0117] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A random noise denoising method for deep reflection seismic data based on multi-scale feature fusion, the method being executed by a processor and characterized in that: include: Based on the deep reflection random medium forward model and seismic random noise in different frequency bands, a training set is generated, specifically: Based on geological structure, a deep reflection random medium forward model is constructed; Perform forward simulation on the deep reflection random medium forward model to obtain synthetic deep reflection seismic data; Normalizing the synthesized deep reflection seismic data and adding seismic random noises of different frequency bands to obtain the noisy deep reflection seismic data; Perform edge zero padding, cutting and data enhancement on the deep reflection seismic data after noise addition to obtain a training set; Build a training network, perform training on the training set, and obtain the trained model parameters, specifically: Based on the training set, we build a training network with convolutional autoencoders, variational autoencoders, and attention mechanisms. Based on the training network, construct the loss function and perform parameter updates; specifically: Based on the convolutional autoencoder CAE, the MSE loss function is constructed; Based on variational autoencoder VAE, KL divergence loss function is constructed; Based on the training network, the MSE loss function and the KL divergence loss function are combined to form a composite loss function. The formula of the composite loss function is: ； In the formula, is the original data, is the reconstructed data, n is the amount of training set data, and Represent the probability distribution functions of the original and reconstructed data respectively; , is the weight, where: ； Input the training set into the training network for training to obtain the trained model parameters; Transfer learning is performed on the trained model parameters to perform actual random noise denoising of deep reflection seismic data.

2. The method for denoising random noise of deep reflection seismic data according to claim 1, characterized in that: The training network constructed based on the training set, in conjunction with the convolutional autoencoder, the variational autoencoder and the attention mechanism, also includes: Based on the shared layers, feature extraction is performed on the training set and local patterns in deep reflection seismic data are identified; The training set is input into the convolution branch to extract deterministic geological features from the training set; Input the training set into the variational branch, perform mean and log variance encoding on the data in the training set, and obtain the latent vector; Based on geological features and latent vectors, multi-scale features are obtained; The multi-scale features are fed into the attention mechanism from the output layer, and the multi-scale latent feature vector is output; The decoder performs reconstruction on the multi-scale latent feature vectors to obtain reconstructed data.

3. The method for denoising random noise of deep reflection seismic data according to claim 2, characterized in that: The convolution branch and the variation branch are combined into an output layer, further comprising: Divide the output layer into upper branch, middle branch and lower branch; The lower branch and the middle branch transmit the feature information in the multi-scale features to the attention mechanism to obtain the feature importance weight data; The upper branch directly inputs multi-scale features to subsequent layers.

4. The method for denoising random noise of deep reflection seismic data according to claim 1, characterized in that: The transfer learning of the trained model parameters is performed to perform random noise denoising of actual deep reflection seismic data, specifically: Migrate the trained model parameters to the training network; Perform iterative training on the training network to obtain an optimized training network; Get the actual data; Based on the optimized trained network, denoising is performed on the actual random noise of deep reflection seismic data.

Images