Pre-stack seismic data strong scattering noise suppression method and system based on residual network

Abstract

The present invention discloses a method and system for suppressing strong scattering noise of pre-stack seismic data based on a residual network. The CABDCN network is constructed, and 5% is randomly selected from several pre-stack seismic data for training, and an unsupervised deep learning method is used for denoising processing to obtain corresponding label data. The noisy seismic data and the corresponding clean labels constitute the training set of the network, and the data not selected into the training set is used as the test set. For each seismic trace gather in the training set, several data with a size of 60*60 are randomly selected and input into the network for training. After the training is completed, the network is used to process the seismic data in the test set to complete the suppression of various random noises in the pre-stack seismic data. The present invention solves the problem of random noise interference in post-stack three-dimensional seismic data, effectively suppresses random noise in pre-stack seismic data, can obtain satisfactory results using less data for training, and at the same time adopts an end-to-end processing method, which can greatly improve the data processing efficiency and has good self-adaptability, meeting the large-scale computing requirements in the industry.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing, and particularly relates to a method and system for suppressing strong scattering noise of prestack seismic data based on a residual network. Background Art

[0002] With the continuous improvement of the degree of oil and gas exploration and the continuous extension of the exploration field, the main targets of oil and gas exploration are also different from those before. The main targets of oil and gas exploration have changed to fractured, subtle and deep oil and gas reservoirs. Such oil and gas reservoirs belong to the most difficult types of reservoirs to explore and develop, and face a series of common exploration challenges such as complex surface conditions and underground structures, thin reservoir thickness and strong heterogeneity. The above challenges force us to extract as much useful information as possible from seismic exploration data, which puts forward higher requirements for seismic exploration technology, seismic data processing methods and interpretation accuracy. At the same time, high signal-to-noise ratio, high resolution and high fidelity have increasingly become the pursuit goals of seismic exploration. Therefore, suppressing random noise in seismic data is an important step in the process of seismic signal processing and interpretation.

[0003] Traditional methods can be divided into the following categories. The first category is mainly based on the continuity of seismic signal isochrones, which has predictable characteristics. For example, f-x deconvolution and t-x predictive filtering techniques. However, large-amplitude prestack noise is likely to destroy the continuity of isochrones, and the denoising effect will be affected when the noise is relatively severe. The second category is based on the low-rank prior of seismic data, mainly including low-rank decomposition and nuclear norm minimization methods. Such methods mainly achieve seismic data denoising based on optimization iteration, and the processing efficiency is very low. The third category is to use sparse prior for denoising. However, due to the uneven energy of prestack noise, it is difficult for the data to meet the sparse assumption. In recent years, deep learning methods have been widely used in the processing of seismic data and have achieved a series of results. For post-stack data, the similarity of the underground structure makes it well meet the assumption condition that the training data and test data required by deep learning are independently and identically distributed. In addition, its signal-to-noise ratio is relatively high and the signal amplitude changes gently, making it relatively easy to train the network. However, for prestack seismic data, the attenuation of the signal amplitude and the significant waveform changes of the reflected waves make the network training unstable. In addition, near-surface scattering noise will seriously damage the effective signal in the f-k domain and t-x domain, and it is difficult to filter it using traditional methods. At the same time, the processing of seismic data also faces the problems of large data volume and long processing time.

[0004] The Cadowz filtering method based on SVD decomposition is widely used in the existing seismic random noise suppression: first, transform the data in the t-xy domain to the Fourier domain f-xy to obtain different frequency slices, then use different frequency slices to construct the Hankle matrix, then perform SVD decomposition on the Hankle matrix, and then select the first K principal components to restore the data. Because in the SVD decomposition, the components corresponding to small singular values are often high-frequency components in the data, mainly including noise and some edges in the signal. A clean signal is obtained by discarding these high-frequency noises.

[0005] The Cadowz method transforms the time domain waveform into the frequency domain. Although the positions of the seismic waves in the time domain are very different, the waveforms of the signals are similar, which means that the frequency components contained in the signals are similar. When the data in the t-xy domain is transformed into the f-xy domain, the self-similarity of the signal will definitely be enhanced. Furthermore, different frequency slices are photographed into a Hankle matrix, and the rank of the matrix must be small. Then, the signal components corresponding to the largest K singular values are selected to reconstruct the original data.

[0006] However, the Cadowz method has the problem of selecting the number of intercepts, which needs to be determined through continuous trials when processing data in practice. For different types of seismic signals, it is still difficult to select the appropriate number of intercepts. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide a method and system for suppressing strong scattering noise of pre-stack seismic data based on a residual network in view of the deficiencies in the above-mentioned prior art, so as to solve the technical problems that random noise in pre-stack seismic data is difficult to eliminate and the processing efficiency is low.

[0008] The present invention adopts the following technical solutions:

[0009] The method for suppressing strong scattering noise of pre-stack seismic data based on residual network includes the following steps:

[0010] S1. Randomly extract several noisy seismic traces X1, X2, X3, ... X from a survey line of seismic records. n , take the seismic gathers that have not been extracted as test samples X test1 ,X test2 ,X test3 ,…X testm ;

[0011] S2, use the unsupervised prestack noise suppression method based on depth image prior to process the noisy seismic gathers X1, X2, X3, ... X in step S1 n , get the corresponding label data The noisy seismic gathers X1, X2, X3, ... Xn and label data are divided into several data pairs of size 60*60 that correspond to each other

[0012] S3. Construct the CABDCN network, and use the data pairs obtained in step S2 as input to the constructed CABDCN network for training. Use the test sample X obtained in step S1 test1 , X test2 , X test3 , … X testm as input to the trained CABDCN network. Process the synthetic seismic data and actual seismic data in the test set through the CABDCN network to complete the suppression of random noise in the pre-stack seismic records.

[0013] Specifically, in step S1, the seismic record X is:

[0014] X = Y + N

[0015] where Y is the effective signal and N is the random noise.

[0016] Specifically, in step S2, use the unsupervised pre-stack noise suppression model based on deep image prior to process the noisy seismic trace sets X1, X2, X3, … X obtained in step S1 n Specifically:

[0017] Use a deep generative convolutional network to learn a mapping f from a latent vector z to clean seismic data θ (z); regard the noisy seismic data as the superposition of the effective signal and random noise, determine the loss function based on the least mean square error criterion, and use the earlystop method to stop the iteration before the network fits to the noise to achieve unsupervised denoising of the seismic data.

[0018] Furthermore, use R(f θ (z) to represent the regularization term constraint of the network structure, and the loss function is:

[0019] E(f θ (z), Y) = ||f θ (z) - Y|| 2 + R(f θ (z))

[0020] where R(f θ (z) is the effective signal, f θ (z) is the mapping from the latent vector z to clean seismic data, and Y is the effective signal.

[0021] Specifically, in step S2, the noisy seismic trace sets X1, X2, X3, … X n and label data Divided into several pairs of data with a size of 60*60 corresponding to each other Specifically:

[0022] Determine the number of samples taken in the seismic trace gather according to the different number of traces in the seismic trace gather. When the number of traces in the seismic trace gather is greater than 900, 2000 training samples are selected from it. When the number of traces is between 500 and 900, it is 1000 training samples. When the number of traces is less than 500, it is 500 samples; Denote the number of samples as m, the number of sampling points in the seismic trace gather as p, the number of traces as q, and the size of the sampled data block as s×s. The sampling process is as follows:

[0023] Randomly extract m integers in [0, q - s) to form a vector v vertical , randomly extract m integers in [0, p - s) to form a vector v horizontal , calculate v vertical and v horizontal 's Cartesian product v; Then use the positions corresponding to each element in v as the positions of the upper left corner of the sampled data block, and extract m seismic data blocks with a size of s×s from the seismic trace gather. The noisy seismic data and its corresponding unsupervised denoising result are sampled using the same group of random sampling methods to obtain training samples

[0024] Furthermore, the number of sampling points in the seismic trace gather for training is 3001, and the sampling period is 2 ms.

[0025] Specifically, in step S3, the CABDCN network is a fully convolutional network, including 15 hidden layers, using the ReLU function as the activation function, and adding a batch normalization layer in front of each activation function.

[0026] Specifically, in step S3, the parameters of the CABDCN network are adjusted using the mini-batch gradient descent method to minimize the empirical loss function, and the optimization objective function is:

[0027]

[0028] Among them, Y is the effective signal in the noisy seismic data, represents the label data, N is the random noise and scattering noise, h represents the mapping from the noisy data to the effective signal, and L represents the loss function adopted for network training;

[0029] Use the empirical loss function of the training samples to train the CABDCN network, and the objective function is as follows:

[0030]

[0031] Among them, h θ(·) is the mapping from noisy data to effective signals learned by the CABDCN network, θ is the parameter of the CABDCN network, and X i is the seismic data to be denoised, and i is the corresponding label data of X is the sum of the mean square error losses between the denoising results obtained by passing N three-dimensional noisy seismic data through the 3D-SNACNN network and their corresponding label data.

[0032] Specifically, in step S3, the CABDCN network uses the SE channel attention mechanism to screen features. The SE channel attention mechanism is specifically as follows:

[0033] The Squeeze operation: Through a global average pooling, the features of h×w×c are changed into features of 1×1×c, obtaining features with a global receptive field;

[0034] The Excitation operation: Input the 1×1×c features obtained above into a multi-layer perceptron and then normalize the output to obtain the weights of each feature channel. It should be noted that the weights here can be dynamically learned;

[0035] The Scale operation: The weights obtained through Excitation represent the importance of the corresponding feature channels. Multiply the learned weights with the previous features respectively to complete the recalibration of the original data in the channel dimension.

[0036] In a second aspect, an embodiment of the present invention provides a pre-stack seismic data strong scattering noise suppression system based on a residual network, including:

[0037] An acquisition module, which randomly extracts several noisy seismic trace gathers X1, X2, X3,... X n from a seismic line of a seismic record, and uses the unextracted seismic trace gathers as test samples X test1 , X test2 , X test3 , … X testm ;

[0038] A partitioning module, which processes the noisy seismic trace gathers X1, X2, X3, … X in the acquisition module using an unsupervised pre-stack noise suppression method based on deep image prior n to obtain the corresponding label data and partitions the noisy seismic trace gathers X1, X2, X3, … X n and the label data into several corresponding data pairs of size 60*60

[0039] Suppression module, construct the CABDCN network, and use the data pair obtained by the partitioning module Input it into the constructed CABDCN network for training, and input the test sample X obtained by the acquisition module test1 , X test2 , X test3 ,… X testm Input it into the trained CABDCN network, and process the synthetic seismic data and actual seismic data in the test set through the CABDCN network to complete the suppression of random noise in the pre-stack seismic record.

[0040] Compared with the prior art, the present invention has at least the following beneficial effects:

[0041] A method for suppressing strong scattering noise in pre-stack seismic data based on a residual network

[0042] All the training data and the data to be processed come from the same survey line, and there is a great similarity between the data, thus meeting the independent and identically distributed hypothesis conditions of deep learning. Select fewer samples and first use the unsupervised learning model U-SNANet for denoising to obtain labeled data, then use the obtained labels to train the supervised learning model CABDCN, and then use the trained CABDCN network to perform noise suppression on the data to be processed; compared with supervised learning, which has the problem of weak generalization ability, unsupervised learning has better adaptability to different data, and the obtained labeled data has better effects. This patent adopts the cascaded method of unsupervised and supervised learning for seismic signal processing, which can not only improve the processing efficiency, but also has better generalization ability for seismic gathers with similar statistical characteristics.

[0043] Furthermore, during the process of collecting seismic data, due to the influence of various factors such as human, environment, and equipment, various noises including random noise will inevitably be introduced. The random noise in the noisy seismic data can usually be regarded as additive noise, and then the superposition of the effective signal and random noise is used for modeling.

[0044] Furthermore, the supervised model proposed by the present invention essentially learns a mapping h, and through the mapping h, the noisy data distribution is transferred to the clean data distribution. Since the data distributions on the same survey line are relatively similar, the generalization ability can be guaranteed at this time, and then only a very small number of training samples need to be selected from them for training to achieve better effects. By adopting this method of dataset selection, on the one hand, the pressure of network training is reduced, on the other hand, the processing efficiency is improved, and the generalization ability of the model is also guaranteed.

[0045] Furthermore, in a deep convolutional network, by reasonably setting the network, a certain prior can be imposed on the fitting process of the model, which is called the deep network prior. Two-dimensional seismic signals have local smoothness, and there is a relatively large correlation between adjacent data. Since the convolution operation acts on the entire field of view, it will impose a certain stationarity and self-similarity on the output. By setting the network structure, the deep generative model tends to extract components with relatively large correlations in the signal. The useful signals in seismic records have large correlations, while the correlations between random noise data are relatively small. During the iteration process, the deep generative model tends to extract useful signals and has weak ability to extract noise in the data. Therefore, the early stop method can be used to stop the iteration before the network fits the noise, and unsupervised denoising of seismic data can be achieved. Compared with other denoising methods, the unsupervised learning method gives the model a larger hypothesis space. The unsupervised denoising method based on the deep network prior uses the structure of the model itself as a regularization term of the network loss function. By imposing prior constraints, the distance between the output of the network and the clean seismic data is relatively close, and it is easier to optimize to the vicinity of the real clean data. In addition, compared with other manually set priors, the constraint of the network structure is weaker, giving the model a larger hypothesis space, enabling the model to work according to the set prior and not being restricted by too strong a prior.

[0046] Furthermore, when dividing seismic trace gathers into small data blocks for training the network, deep-layer data is removed, mainly to prevent the unsatisfactory denoising of deep-layer data by the unsupervised learning method. If the supervised learning model learns this part of the data, it will reduce the denoising performance of the model. The random sampling method is adopted during data extraction mainly to prevent the problem that the rules set by humans may lead to non-uniform sampling of all possible situations.

[0047] Furthermore, since it is difficult to obtain the true distribution of seismic data, from the perspective of feasibility, the distribution of finite seismic record samples in the training set is used to approximate the true distribution of seismic data. The mean square error between the denoising result and the label data is used as the loss of the neural network to train the network model and learn an explicit mapping function. Since both the training model and the data to be processed come from the same survey line, the data distributions are very similar. Therefore, it can be considered that the training set and the test set satisfy the assumption of independent and identical distribution. The trained CABDCN model can obtain good generalization ability and thus achieve good results on the test data.

[0048] Furthermore, the constructed CABDCN network is a fully convolutional network. Therefore, under the condition that the data size does not cause the computer memory to explode, the input of the network can be of any size, thus avoiding the use of the overlap operation that wastes computing power. Through actual comparison, it is found that the structure of sampling residual learning enables the network to fit noise better than directly learning the effective signal. The CABDCN network contains 15 hidden layers, uses the ReLU function as the activation function, and adds the batch normalization (BatchNorm) layer in front of each activation function. In addition, in order to improve the denoising effect, we embed the channel attention mechanism in the designed network.

[0049] Furthermore, using the gradient descent algorithm, in each iteration process, calculate the derivatives of the loss function with respect to each parameter, and update the parameters along the gradient direction of the current loss function. As the number of iterations increases, the loss function will gradually decrease. Although it may not reach a global minimum, the final result is still good, and it will not completely overfit the training set data and lead to a decline in generalization ability.

[0050] Furthermore, the convolutional neural network uses different convolutional kernels to process the input two-dimensional data to extract various features of the data. There are differences between the outputs of different channels, with different structures, different scales, and many other differences. In the denoising task, due to the differences in the statistical characteristics of the effective signal and random noise. When fitting the noise, different channels have different contributions. Based on this idea, the CABDCN network adopts the SE channel attention mechanism to emphasize some features and suppress some features, so that the features that contribute greatly to the output are propagated deeper into the network, and those with little contribution are reduced in their impact on the result.

[0051] It can be understood that the beneficial effects of the second aspect can refer to the relevant descriptions in the first aspect above, and will not be elaborated here.

[0052] To sum up, the present invention uses the unsupervised learning model U-SNANet to perform denoising processing to obtain labeled data, then uses the obtained labels to train the supervised learning model (CABDCN), and then uses the trained CABDCN to perform denoising processing on the test set data, thereby improving the efficiency of seismic signals and having good generalization ability for seismic trace gathers with similar statistical characteristics.

[0053] Next, through the drawings and embodiments, the technical solutions of the present invention will be further described in detail. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is the network structure diagram of the CABDCN network;

[0055] Figure 2It is the network structure diagram of U-SNANet;

[0056] Figure 3 It is the processing result of synthetic seismic data. Among them, (a) is the label of synthetic seismic data, with 96 traces and 750 sampling points. Adding random Gaussian noise gives the noisy seismic data in Figure (b). (c) is the effective signal obtained by processing (b) using the CABDCN network, and (d) is the noise suppressed in (b).

[0057] Figure 4 It is the result of unsupervised denoising of prestack seismic data through the U-SNANet network. Among them, (a) is the original seismic data contaminated by noise, (b) is the effective signal y obtained by processing with the U-SNANet network, and (c) is the noise suppressed after processing the data in (a) through the U-SNANet network;

[0058] Figure 5 It is the processing result of actual seismic data. Among them, (a) is the effective signal obtained by processing Figure 4 (a) using the U-SNANet network, (b) is the effective signal obtained by processing Figure 4 (a) using the CABDCN network with a residual structure, (c) is the effective signal obtained by processing Figure 4 (a) using the CABDCN network with a direct mapping structure, and (d), (e), (f) are the Figure 4 corresponding noise data suppressed in (a), (b), (c) of (a);

[0059] Figure 6 It is the result of the ablation experiment on the channel attention module in the CABDCN network. Figure (a) shows the effective signals obtained by processing Figure 4 (a) using the CABDCN network with and without the channel attention module respectively, and Figures (c) and (d) are the corresponding noise data of Figures (a) and (b) respectively.

[0060] Figure 7 It is the schematic diagram of the process of the present invention. Detailed implementation manners

[0061] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0062] In the description of the present invention, it should be understood that the terms "include" and "comprise" indicate the presence of the described features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their combinations.

[0063] It should also be understood that the terms used in the specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly dictates otherwise, the singular forms "a", "an" and "the" are intended to include the plural forms.

[0064] It should be further understood that the term "and / or" used in the specification of the present invention and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations. For example, A and / or B can represent: the case where A exists alone, the case where A and B exist simultaneously, and the case where B exists alone. In addition, the character " / " herein generally indicates that the objects before and after are in an "or" relationship.

[0065] It should be understood that although terms such as first, second, third, etc. may be used in the embodiments of the present invention to describe preset ranges, etc., these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from each other. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0066] Depending on the context, the word "if" as used herein can be interpreted as "when" or "while" or "in response to determining" or "in response to detecting". Similarly, depending on the context, the phrase "if determined" or "if detected (stated condition or event)" can be interpreted as "when determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)".

[0067] Schematic diagrams of various structures according to the disclosed embodiments of the present invention are shown in the drawings. These figures are not drawn to scale, where certain details are enlarged for the purpose of clear expression, and certain details may be omitted. The shapes of the various regions and layers shown in the figures and their relative sizes and positional relationships are merely exemplary, and may actually deviate due to manufacturing tolerances or technical limitations, and those skilled in the art can additionally design regions / layers with different shapes, sizes and relative positions according to actual needs.

[0068] The present invention provides a method for suppressing strong scattered noise in pre-stack seismic data based on a residual network, constructs a CABDCN (Channel Attention Boosting Deep Convolutional Networks) network, randomly extracts 5% from several pre-stack seismic data for training, and uses an unsupervised deep learning method for denoising processing to obtain corresponding labeled data. The noisy seismic data and the corresponding clean labels constitute the training set of the network, and the data not selected into the training set is used as the test set. For each seismic trace gather in the training set, several data of size 60*60 are randomly extracted and input into the network for training. After the training is completed, the network is used to process the seismic data in the test set to complete the suppression of various random noises in the pre-stack seismic data. The present invention solves the problem of interference from random noise in post-stack 3D seismic data, effectively suppresses random noise in pre-stack seismic data, can obtain satisfactory results using less data for training, and at the same time adopts an end-to-end processing method, which can greatly improve the data processing efficiency and has good adaptability, meeting the large-scale computing requirements in the industry.

[0069] Please refer to Figure 7 , a method for suppressing strong scattered noise in pre-stack seismic data based on a residual network according to the present invention, comprising the following steps:

[0070] S1. Randomly extract several noisy seismic trace gathers X1, X2, X3, … X n from a seismic line in a seismic record, and use the seismic trace gathers not extracted as test samples X test1 , X test2 , X test3 , … X testm to test the denoising performance of the network;

[0071] Noisy seismic record:

[0072] X = Y + N

[0073] where Y represents the effective signal and N represents the random noise.

[0074] It should be noted that the purpose of this application is to improve the speed and quality of data processing, randomly extract fewer samples from the data collected in a single seismic exploration for training, and thus obtain a model with better generalization ability for this seismic data, rather than to obtain a model applicable to all seismic data.

[0075] The training data and the data to be processed all come from the same survey line, and there is a high degree of similarity between the data. By selecting a small number of samples, we first use a model trained with unsupervised learning to perform denoising and obtain labels. Then, we use the obtained labels to train a supervised learning model and perform denoising using the trained supervised model. The reason for adopting this cascaded approach is that traditional supervised models have problems with generalization ability. A model trained using seismic data with statistical features different from the current data to be processed is difficult to ensure good generalization ability for the current data. Therefore, we use unsupervised learning to process the data and obtain labels. At the same time, although using unsupervised learning for denoising can achieve relatively good results, since the process of processing seismic data is actually an optimization and iteration process, an explicit denoising function cannot be learned, resulting in relatively low denoising efficiency. Considering all the above points, adopting the cascaded approach of unsupervised and supervised learning for seismic signal processing can not only improve the processing efficiency but also have good generalization ability for seismic trace gathers with similar statistical features.

[0076] S2. Use the unsupervised pre-stack noise suppression method based on deep image prior to process the noisy seismic trace gathers X1, X2, X3, … X n , and obtain the corresponding labeled data Divide X1, X2, X3, … X n and into several corresponding data pairs of size 60 * 60 as the training set for the denoising task;

[0077] Specifically, use the unsupervised seismic random noise suppression algorithm to denoise the training set:

[0078] Use a deep generative convolutional network to learn a mapping f from a latent vector z to clean seismic data θ :

[0079] Y = f θ (z),

[0080] The noisy seismic data can be regarded as the superposition of the effective signal and random noise:

[0081] X = f θ (z) + N.

[0082] According to Bayes' theorem, assuming that the noise in the seismic record follows a Gaussian distribution, design the following loss function.

[0083] E(f θ (z), X) = ||f θ (z) - Y|| 2 ,

[0084] In a deep convolutional network, by reasonably setting the network, a certain prior can be imposed on the fitting process of the model, which is called the deep network prior. Two-dimensional seismic signals have local smoothness, and there is a relatively large correlation between adjacent data. Since the convolution operation acts on the entire field of view, it will impose a certain degree of stationarity and self-similarity on the output. By setting the network structure, the deep generative model tends to extract components with relatively large correlations in the signal.

[0085] Considering the prior attributes of the deep generative model, a regularization term constraint from the network structure is imposed after the loss function:

[0086] E(f θ (z),Y) = ||f θ (z) - Y|| 2 + R(f θ (z))

[0087] The useful signals in the seismic records have a large correlation, while the correlation between random noise data is small. During the iterative process, the deep generative model tends to extract useful signals and has a weak ability to extract noise in the data. Therefore, by using the early stop method to stop the iteration before the network fits to the noise, unsupervised denoising of seismic data can be achieved.

[0088] The specific method for dividing the seismic trace gather into training samples of size 60 * 60 is as follows:

[0089] First, determine the number of samples to be taken in the trace gather according to the number of traces in the seismic trace gather. The number of sampling points for the seismic trace gathers used for training in this patent is all 3001, and the sampling period is 2 ms. When the number of traces in the seismic trace gather is greater than 900, 2000 training samples are selected from it. When the number of traces is between 500 and 900, it is 1000 training samples. When the number of traces is less than 500, it is 500 samples. Denote the number of samples as m, the number of sampling points of the seismic trace gather as p, the number of traces as q, and the size of the sampled data block as s × s. The specific sampling process is as follows:

[0090] Randomly draw m integers in [0, q - s) to form a vector v vertical , and randomly draw m integers in [0, p - s) to form a vector v horizontal . Calculate the Cartesian product of v vertical and v horizontal :

[0091] v = v vertical × v horizontal

[0092] Then, using the positions corresponding to the elements in v as the positions of the upper left corners of the sampled data blocks, m seismic data blocks of size s×s are extracted from the seismic trace gather. It should be noted that the noisy seismic data and its corresponding unsupervised denoising results are sampled using the same set of random sampling methods to obtain training samples

[0093] S3. Construct a deep network model CABDCN for denoising, and use the data obtained in step S2 to train the deep network model CABDCN, and save the trained parameters; after the deep network model CABDCN is trained, load the saved network parameters, and use the test samples X test1 , X test2 , X test3 , … X testm input into the CABDCN network, and process the synthetic seismic data and actual seismic data in the test set through the 3D-SNACNN network to complete the suppression of random noise.

[0094] Send the training samples obtained in step S2 into the constructed CABDCN network for training, and use the mini-batch gradient descent method to adjust the parameters of the CABDCN network to minimize the empirical loss function, and optimize the CABDCN network by minimizing the empirical loss function.

[0095] After training, the network learns an explicit mapping h θ (·) from the noisy data to the clean data, and then outputs the data to be processed from the network to obtain the corresponding denoising result.

[0096] Use the mini-batch gradient descent method to adjust the parameters of the CABDCN network, and the specific objective function for optimization is:[[]]

[0097]

[0098] where Y is the effective signal in the noisy seismic data, represents the label data, N is the random noise and scattered noise, h represents the mapping from the noisy data to the effective signal, L represents the loss function adopted for network training, and in the present invention, the mean square error between h(x + n) and x is used as the loss function.

[0099] From the perspective of realizability, the distribution of finite samples is used to approximate the true distribution of seismic data. In the present invention, that is, the empirical loss function of the training samples is used to train the CABDCN network.

[0100] The objective function is as follows:

[0101]

[0102] Among them, h θ (·) is the mapping from noisy data to effective signals learned by the CABDCN network through training, θ is the parameter of the CABDCN network, and X i is the seismic data to be denoised, is the corresponding label data of X i ; is the sum of the mean square error losses between the denoising results obtained by passing N three-dimensional noisy seismic data through the CABDCN network and their corresponding label data.

[0103] Please refer to Figure 1 , the constructed CABDCN network is a fully convolutional network. Therefore, under the condition that the data size does not cause the computer memory to explode, the input of the network can be of any size, thus avoiding the use of the overlap operation that wastes computing power. The designed CABDCN network adopts the method of residual learning, enabling the network to fit the noise in the data and then subtract to obtain the effective signal, and then calculating the loss to perform the backpropagation operation. The CABDCN network contains 15 hidden layers, uses the ReLU function as the activation function, and adds the batch normalization (BatchNorm) layer in front of each activation function.

[0104] Please refer to Figure 2 , in order to improve the denoising effect, the channel attention mechanism is embedded in the designed network. After training, the CABDCN network learns a mapping h θ (X) = Y' ≈ Y from the original seismic data to extract the effective signal. The original three-dimensional seismic data is input into the CABDCN network to obtain the denoising result Y' in the seismic data.

[0105] Convolutional neural networks use different convolutional kernels to process the input two-dimensional data to extract various features of the data. There are differences between different output channels, with different structures, different scales, and many other differences. In the denoising task, due to the differences in the statistical characteristics of effective signals and random noise. When fitting the noise, different channels have different contributions. Based on this idea, the CABDCN network adopts the SE channel attention mechanism to emphasize some features and suppress some features, so that the features that contribute greatly to the output are propagated deeper into the network, and the influence of those with little contribution on the result is reduced.

[0106] The SE channel attention mechanism includes three steps:

[0107] Firstly, it is the Squeeze operation. Through a global average pooling, the h×w×c features are transformed into 1×1×c features, obtaining features with a global receptive field;

[0108] Next is the Excitation operation. The 1×1×c features obtained above are input into a multi-layer perceptron, and then the output is normalized to obtain the weights for each feature channel. It should be noted that the weights here can be dynamically learned;

[0109] Finally, there is a Scale operation. The weights obtained through Excitation represent the importance of the corresponding feature channels. The learned weights are multiplied by the previous features respectively to complete the recalibration of the original data in the channel dimension.

[0110] Batch normalization technology is used to accelerate the learning of the network. The forward propagation formula for batch normalization is as follows:

[0111]

[0112]

[0113]

[0114]

[0115] Among them, β (k) = E[x (k) , μ Β , are the mean and variance of the outputs of m data in a batch passing through a certain layer of neurons; is the result of normalizing x Β , using μ i . ε is a very small positive number; y i is the result of scaling the mean and variance of using γ and β. γ and β can be adjusted during the training process;

[0116] During the backpropagation process, the chain derivative of the batch normalization layer is as follows:

[0117]

[0118]

[0119]

[0120]

[0121]

[0122]

[0123] Among them, Denote the derivative of loss \(l\) with respect to the variable \((\cdot)\). After obtaining the derivatives of the loss with respect to each parameter, the gradient descent algorithm can be used to optimize the parameters to obtain a smaller empirical loss.

[0124] In another embodiment of the present invention, a prestack seismic data strong scattering noise suppression system based on a residual network is provided. This system can be used to implement the above-mentioned prestack seismic data strong scattering noise suppression method based on a residual network. Specifically, the prestack seismic data strong scattering noise suppression system based on a residual network includes an acquisition module, a division module, and a suppression module.

[0125] Among them, the acquisition module randomly extracts a number of noisy seismic trace sets \(X_1, X_2, X_3, \ldots, X\) from a seismic line of a seismic record n , and takes the seismic trace sets that are not extracted as test samples \(X\) test1 , \(X\) test2 , \(X\) test3 , \(\ldots, X\) testm ;

[0126] The division module processes the noisy seismic trace sets \(X_1, X_2, X_3, \ldots, X\) in the acquisition module using an unsupervised prestack noise suppression method based on deep image prior n , and obtains the corresponding label data The noisy seismic trace sets \(X_1, X_2, X_3, \ldots, X\) n and the label data are divided into a number of corresponding data pairs of size \(60\times60\)

[0127] The suppression module constructs a CABDCN network, inputs the data pairs obtained by the division module into the constructed CABDCN network for training, and inputs the test samples \(X\) obtained by the acquisition module test1 , \(X\) test2 , \(X\) test3 , \(\ldots, X\) testm into the trained CABDCN network. The CABDCN network processes the synthetic seismic data and actual seismic data in the test set to complete the suppression of random noise in the prestack seismic record.

[0128] In another embodiment of the present invention, a terminal device is provided. The terminal device includes a processor and a memory. The memory is used to store a computer program, and the computer program includes program instructions. The processor is used to execute the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding function. The processor described in the embodiment of the present invention can be used for the operation of the pre-stack seismic data strong scattering noise suppression method based on the residual network, including:

[0129] Randomly extract several noisy seismic trace sets X1, X2, X3, … X from a seismic line of a seismic record n , and use the seismic trace sets that are not extracted as test samples X test1 , X test2 , X test3 , … X testm ; Use the unsupervised pre-stack noise suppression method based on depth image prior to process the noisy seismic trace sets X1, X2, X3, … X n to obtain the corresponding label data Divide the noisy seismic trace sets X1, X2, X3, … X n and the label data into several corresponding data pairs of size 60*60 Construct a CABDCN network, and input the data pairs into the constructed CABDCN network for training. Input the test samples X test1 , X test2 , X test3 , … X testm into the trained CABDCN network, and process the synthetic seismic data and actual seismic data in the test set through the CABDCN network to complete the suppression of random noise in the pre-stack seismic record.

[0130] In another embodiment of the present invention, the present invention further provides a storage medium, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is a memory device in a terminal device and is used to store programs and data. It can be understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and, of course, the extended storage medium supported by the terminal device. The computer-readable storage medium provides a storage space, and the operating system of the terminal is stored in this storage space. And, one or more instructions suitable for being loaded and executed by the processor are also stored in this storage space. These instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory (Non-Volatile Memory), such as at least one disk memory.

[0131] One or more instructions stored in the computer-readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the method for suppressing strong scattering noise in pre-stack seismic data based on the residual network in the above embodiment; one or more instructions in the computer-readable storage medium are loaded and executed by the processor to perform the following steps:

[0132] Randomly extract several noisy seismic trace sets X1, X2, X3,... X from a seismic line of seismic records n , and use the seismic trace sets not extracted as test samples X test1 , X test2 , X test3 ,... X testm ; Use an unsupervised pre-stack noise suppression method based on deep image prior to process the noisy seismic trace sets X1, X2, X3,... X n , and obtain the corresponding label data Divide the noisy seismic trace sets X1, X2, X3,... X n and the label data into several corresponding data pairs of size 60*60 Construct a CABDCN network, input the data pairs into the constructed CABDCN network for training, and input the test samples X test1 , X test2 , X test3 ,... X testm into the trained CABDCN network, and process the synthetic seismic data and actual seismic data in the test set through the CABDCN network to complete the suppression of random noise in the pre-stack seismic records.

[0133] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components described and shown in the accompanying drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed present invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0134] Please refer to Figure 3 , (a) is the synthetic seismic data. Random noise is added to (a) to obtain the noisy data with a peak-to-average signal-to-noise ratio of 26.34 dB. The CABDCN model is used to process the data in (b) to obtain the results in (c) and (d). The peak-to-average signal-to-noise ratio of (c) is 41.88 dB, and the noise in (a) is effectively suppressed.

[0135] Please refer to Figure 4 , (a) is the original seismic data severely contaminated by noise, with a total of 1083 traces and 3001 sampling points. The U-SNANet is used to suppress the noise in (a) in an unsupervised learning manner to obtain the effective signal shown in (b), and (c) is the suppressed noise in (a). Observing (b), it can be seen that the event axis in the denoising result is very smooth, and the spikes caused by noise in the original signal are also eliminated. Observing (c), it can be seen that there is almost no residue of the effective signal in the suppressed noise in (a), proving that the result processed by the U-SNANet model has high fidelity and little damage to the effective signal in the original seismic record during the denoising process.

[0136] Please refer to Figure 5 , the U-SNANet model, the CABDCN models with and without residual structures are respectively used to process Figure 4 the noisy seismic data in (a) to obtain (a) in the figure, Figure 4 figure (b) and Figure 4 the results in figure (c), Figure 5 (d), Figure 5 (e) and Figure 5 (f) are respectively the corresponding Figure 4The noise data suppressed in (a). By comparing and observing Figures (d), (e), and (f), it can be seen that the errors among the three are extremely small. After careful comparison, it can be found that the result obtained using the residual network is smoother and has better fidelity. The peak signal-to-noise ratio (PSNR) of Figure (e), Figure (f), and the corresponding labeled data Figure (d) are calculated to be 52.8 dB and 52.2 dB respectively. This also shows from the figure and the PSNR index that the network using the residual structure can obtain better results.

[0137] Please refer to Figure 6 , the ablation experiment on the channel attention module in the CABDCN network. Figure (a) shows the effective signals processed by the CABDCN network with and without the channel attention module respectively Figure 4 (a). Figures (c) and (d) are the corresponding noise data of Figures (a) and (b) respectively. The labeled data corresponding to Figures (a) and (b) are as Figure 4 (a). By calculating the PSNR, the corresponding results of (a) and (b) are 55.2 and 52.8 respectively. It can be found that the network with the channel attention module can better promote training convergence and obtain better results.

[0138] In summary, the present invention provides a method and system for suppressing strong scattering noise in pre-stack seismic data based on a residual network, which solves the interference problem of random noise in post-stack three-dimensional seismic data, effectively suppresses the random noise in pre-stack seismic data, can obtain satisfactory results using less data for training, and at the same time adopts an end-to-end processing method, which can greatly improve the data processing efficiency and has good self-adaptability, meeting the large-scale computing requirements in the industry.

[0139] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the above division of each functional unit and module is used as an example. In practical applications, the above functions can be allocated to different functional units and modules according to needs, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated into a processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above integrated unit can be implemented in the form of hardware or in the form of a software functional unit. In addition, the specific names of each functional unit and module are only for the convenience of mutual distinction and do not limit the protection scope of the present application. The specific working processes of the units and modules in the above system can refer to the corresponding processes in the foregoing method embodiments and will not be repeated here.

[0140] In the above embodiments, the descriptions of the various embodiments each have their own emphasis. For parts not detailed or recorded in a certain embodiment, reference may be made to the relevant descriptions of other embodiments.

[0141] Those of ordinary skill in the art will recognize that the units and algorithm steps of the examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, or in a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. A professional technician can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0142] In the embodiments provided by the present invention, it should be understood that the disclosed device / terminal and method can be implemented in other ways. For example, the device / terminal embodiments described above are merely illustrative. For example, the division of the modules or units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of the devices or units can be in electrical, mechanical or other forms.

[0143] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they can be located in one place, or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0144] In addition, the functional units in the various embodiments of the present invention can be integrated in one processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above integrated units can be implemented in the form of hardware or in the form of software functional units.

[0145] When the integrated module / unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, to implement all or part of the processes in the above-described embodiment methods of the present invention, it can also be completed by a computer program instructing relevant hardware. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-described various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0146] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the specified function in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0147] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the specified function in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0148] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are executed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions for implementing the functions specified in one process or a plurality of processes and / or blocks Figure 1 in one block or a plurality of blocks Figure 1 in one process or a plurality of processes and / or blocks

[0149] The above is only to illustrate the technical idea of the present invention and should not be used to limit the protection scope of the present invention. Any modification made on the basis of the technical solution according to the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A pre-stack seismic data strong scattering noise suppression method based on a residual network, characterized in that Including the following steps: S1. Randomly extract several noisy seismic trace sets from a seismic line of the seismic record , and use the seismic trace sets not extracted as test samples ; S2. Process the noisy seismic trace gather in step S1 using an unsupervised pre-stack noise suppression method based on depth image prior to obtain the corresponding labeled data . Divide the noisy seismic trace gather and the labeled data into a number of corresponding data pairs of size 60*60 ; S3. Construct a CABDCN network, which is a fully convolutional network with 15 hidden layers. The ReLU function is used as the activation function, and a batch normalization layer is added in front of each activation function. The CABDCN network uses the SE channel attention mechanism to screen features. The specific SE channel attention mechanism is as follows: operation, through a global average pooling, features are transformed into a feature, obtaining features with a global receptive field; Operation, input the features obtained above into a multi-layer perceptron and then normalize the output to obtain the weights for each feature channel. It should be noted that the weights here can be dynamically learned; The operation, after The obtained weights represent the importance of the corresponding feature channels. Multiply the learned weights by the previous features respectively to complete the recalibration of the original data in the channel dimension. For the data obtained in step S2 Input it into the constructed CABDCN network for training. Input the test samples obtained in step S1 into the trained CABDCN network. Process the synthetic seismic data and actual seismic data in the test set through the CABDCN network to complete the suppression of random noise in the pre-stack seismic records.

2. The method for suppressing strong scattering noise of pre-stack seismic data based on the residual network according to claim 1, wherein In step S1, the seismic record is as follows: Among them, is a valid signal, is random noise.

3. The pre-stack seismic data strong scattering noise suppression method based on a residual network according to claim 1, characterized in that In step S2, the noisy seismic trace gather obtained in step S1 is processed using an unsupervised pre-stack noise suppression model based on depth image prior Specifically: Learning a mapping from a latent vector to clean seismic data using a deep generative convolutional network to clean seismic data ; The noisy seismic data is regarded as the superposition of effective signals and random noise. The loss function is determined based on the least mean square error criterion, and the earlystop method is used to stop the iteration before the network fits the noise, so as to realize unsupervised denoising of seismic data.

4. The method for suppressing strong scattering noise of pre-stack seismic data based on the residual network according to claim 3, wherein Usage Indicates the regularization term constraint of the network structure, and the loss function is: Among them, is a valid signal, is a latent vector mapping to clean seismic data, is a valid signal.

5. The method for suppressing strong scattering noise of pre-stack seismic data based on a residual network according to claim 1, wherein In step S2, the noisy seismic trace gather and the label data are divided into a number of corresponding data pairs of size 60*60 Specifically: Determine the number of samples taken in the seismic gather according to the number of traces in the seismic gather. When the number of traces in the seismic gather is greater than 900, 2000 training samples are selected from it. When the number of traces is between 500 and 900, 1000 training samples are selected. When the number of traces is less than 500, 500 samples are selected. Denote the number of samples as , the number of sampling points in the seismic gather is , the number of traces is , the size of the sampled data block is , and the sampling process is as follows: In randomly select integers to form a vector , randomly select integers to form a vector , calculate and Cartesian product ; then use the positions corresponding to the elements in as the positions of the upper left corners of the sampled data blocks, and extract seismic data blocks with a size of from the seismic trace gather. The noisy seismic data and its corresponding unsupervised denoising results are sampled using the same set of random sampling methods to obtain the training samples .

6. The method for suppressing strong scattering noise of pre-stack seismic data based on a residual network according to claim 5, wherein The number of sampling points for the seismic trace gather used for training is 3,001, and the sampling period is 2 .

7. The pre-stack seismic data strong scattering noise suppression method based on the residual network according to claim 1, characterized in that In step S3, the small batch gradient descent method is used to adjust the parameters of the CABDCN network to minimize the empirical loss function. The optimized objective function is: wherein, is the effective signal in the noisy seismic data, represents the label data, is the random noise and scattered noise, represents the mapping from the noisy data to the effective signal, represents the loss function adopted for network training; The CABDCN network is trained using the empirical loss function of the training samples. The objective function is as follows: Among them, is the mapping from noisy data to effective signals learned by the CABDCN network, are the parameters of the CABDCN network, is the seismic data to be denoised, is the corresponding labeled data, is the sum of the mean square error losses between the denoising results obtained by the 3D-SNACNN network for 8. A prestack seismic data strong scattering noise suppression system based on a residual network, characterized in that Including: Acquisition module, randomly extract several noisy seismic trace gathers from a seismic line of seismic records , and use the seismic trace gathers not extracted as test samples ; Partitioning module, using an unsupervised pre-stack noise suppression method based on depth image prior to process the noisy seismic trace gather in the acquisition module , to obtain the corresponding label data , the noisy seismic trace gather and the label data are partitioned into a number of corresponding data pairs of size 60*60 ; Suppression module. Construct a CABDCN network, which is a fully convolutional network with 15 hidden layers. The ReLU function is used as the activation function, and a batch normalization layer is added in front of each activation function. The CABDCN network uses the SE channel attention mechanism to screen features. The specific SE channel attention mechanism is as follows: operation, through a global average pooling to the features into a feature, obtaining features with a global receptive field; Operation, input the features obtained above into a multi-layer perceptron and then normalize the output to obtain the weights for each feature channel. It should be noted that the weights here can be dynamically learned; Operation, after The obtained weights represent the importance of the corresponding feature channels. Multiply the learned weights with the previous features respectively to complete the recalibration of the original data in the channel dimension. For the data pairs obtained by the partitioning module Input into the constructed CABDCN network for training. Input the test samples obtained by the acquisition module Into the trained CABDCN network. Process the synthetic seismic data and actual seismic data in the test set through the CABDCN network to complete the suppression of random noise in the pre-stack seismic records.

Images