CN119310628A - Seismic denoising method based on MLReal transform - Google Patents
Seismic denoising method based on MLReal transform Download PDFInfo
- Publication number
- CN119310628A CN119310628A CN202310855828.4A CN202310855828A CN119310628A CN 119310628 A CN119310628 A CN 119310628A CN 202310855828 A CN202310855828 A CN 202310855828A CN 119310628 A CN119310628 A CN 119310628A
- Authority
- CN
- China
- Prior art keywords
- data
- training
- seismic
- mlreal
- method based
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a seismic denoising method based on MLReal transformation, which comprises the steps of 1, selecting an objective function, 2, building a denoising neural network, 3, redefining training data, 4, redefining test data, 5, inputting the redefined training data and the redefined test data into the neural network, and performing training cycle iteration to obtain synthetic data with field data characteristics. The seismic denoising method based on MLReal transformation combines important information of field seismic data into a training process of synthetic data, and uses a training success network for denoising the field seismic data, so that the denoising effect is obviously improved.
Description
Technical Field
The invention relates to the technical field of oil and gas exploration seismic data processing, in particular to a MLReal transformation-based seismic denoising method.
Background
In the process of seismic data acquisition, various noise can appear in the acquired seismic data due to the interference of various environmental factors and equipment, so that the signal-to-noise ratio and the resolution of a seismic section are affected, and noise suppression is very necessary. With the continual efforts of students, the method of seismic denoising has a great breakthrough, specifically f-x deconvolution (Naghizadeh AND SACCHI, 2012), radon transformation (Foster and Mosher,1992;Hargreaves et al, 2001), curvelet transformation (Cao et al, 2015), dictionary learning (Beckouche and Ma, 2014) and singular value decomposition (Sang Yu, 2014), etc. These conventional methods have achieved some results in dealing with noise, but in cases where the subsurface structure is very complex and the seismic data acquired is complex, the result of the processing will be quite undesirable.
In recent years, with the tremendous development of deep learning theory and application, deep learning algorithms have been applied to various fields. In the aspect of seismic denoising, han Weixue achieves good effects by constructing a CNN network and removing random noise in seismic data, 2019, dong proposes to suppress low-frequency noise in a desert based on self-adaption DnCNN determined by high-order statistics, 2021, fang Wenqian and the like propose a dual residual network for random noise suppression of seismic signals to protect effective signals, 2022, mihai l-Antonio Chirtu divide the seismic signals into noise and signal parts in a time-frequency domain and denoise the signals by using a U-Net convolutional neural network structure, which proves the feasibility of the method, and 2022, zhang Hao propose to perform seismic denoising based on an imaging domain of an anti-neural network based on conditions, so that not only is the image structure undistorted, but also various types of noise are suppressed. With the continuous development of deep learning, the method can be seen to have great potential in the aspect of seismic data denoising, and becomes the mainstream of the current world.
The choice of training data in the deep learning process, however, determines the performance of the neural network and its suitability for further data sets. Particularly in seismic applications, selection is important for both methods of manually marking field seismic data or generating synthetic seismic data, both of which exhibit serious limitations. The advantage of manually tagging field seismic data tags is that the data of the training network may have the same characteristics as the future application network, but is often limited to the performance of conventional methods of denoising, and thus cannot exceed the conventional methods required to generate tags, and training using synthetic data sets overcomes this obstacle, allowing a completely noiseless tag to be generated for the training process. However, if the composite data is almost different from the actual data set, it can affect the quality of the decision made based on the data. Even though the composite data is indeed very good, it is a copy of the specific properties of the real dataset. The model looks for trends to replicate, so some random behavior may be ignored, greatly affecting the noise reduction effect.
The challenge in training our neural network model on synthetic data is to generalize the training model to real data because the process requires careful recognition of the training set and careful recognition of the real noise and other variables contained between the synthetic data and the real data. In other words, the synthetic data and the real data are not typically from the same distribution, which is critical to the success of training the neural network model (Kouw, 2018). Thus, many synthetically trained neural network models perform poorly on real data. On the other hand, training of real data provides models that are typically as good as the accuracy of labels determined by manual algorithms. Therefore, in this case, the data driving characteristics of the machine learning will be seriously impaired (Zhou, 2017). Because the synthesized data cannot usually capture the real situation of the field data, the training Neural Network (NN) of the reasoning stage is not good, so that the denoising effect is reduced. Therefore, how to inject as many field seismic data features as possible into the synthetic data training, so as to reduce the error between the synthetic training data and the real application data, and finally, the field seismic data features are used in the seismic denoising task, which is an urgent problem to be solved.
In the Chinese patent application with the application number of CN202210523847.2, a seismic data denoising method based on a convolutional neural network is related, an initial data set is firstly constructed, a circularly generated countermeasure convolutional neural network is constructed, the circularly generated countermeasure convolutional neural network takes a circularly generated countermeasure network structure as a main network, the circularly generated countermeasure network structure comprises a generator and a discriminator, a non-local neural network is used as a residual error to be connected between convolutional layers of the generator, and the discriminator replaces the original full-connection mode with a PatchGAN output mode. Then training the loop generation countermeasure convolutional neural network, and finally denoising the noisy seismic data. Experiments prove that the method has good denoising effect.
In the Chinese patent application with the application number of CN202211250180.X, a seismic denoising method based on a convolutional neural network and a visual transformation neural network is related to the technical field of information processing. The method comprises the steps of collecting noise-containing seismic data, carrying out denoising treatment on the noise-containing seismic data to obtain a plurality of denoised seismic data blocks, constructing a seismic data set comprising a training set, a verification set and a test set, constructing a seismic data denoising network based on a convolutional neural network and a vision transformation neural network, training the seismic data denoising network by using the training set, debugging the denoising effect of the trained seismic data denoising network by using the verification set to obtain the seismic data denoising network with the best denoising effect, and removing the noise of the noise-containing seismic data blocks in the test set by using the trained seismic data denoising network to obtain the denoised seismic data. The method effectively improves the denoising performance of the seismic data, reduces the training cost, and is favorable for quickly and accurately acquiring the underground medium construction information.
In the Chinese patent application with the application number of CN202111337078.9, an unsupervised seismic data denoising method based on a depth tensor neural network is applied to the field of seismic data processing, and aims at the problem of low denoising performance in the prior art; the invention firstly uses SURE to replace a cost function of MSE to convert data denoising into a corresponding unsupervised regression model, then establishes a SURE-TCNN network framework, follows a tNN framework and a t-product based on transformation, which is an expansion of the t-product of the traditional tNN framework, and utilizes the advantages in the t-product to split the SURE-TCNN based on tensor into independent SURE-CNN based on matrix for each front surface or time slice in a time-frequency domain, thus being easy to solve and attractive, and finally synthesizes and truly data experiments show that compared with the SOTA denoising method, the proposed method realizes superior performance.
In the Chinese patent application with the application number of CN201710334723.9, a seismic wave noise reduction method based on the combination of self-adaptive filtering and wavelet transformation is related to solving the problems of extracting effective signals which are beneficial to analysis and explanation from mixed signals, suppressing noise, namely highlighting the effective signals, removing interference signals as much as possible and realizing the purpose of separating signals from noise. The method comprises the steps of firstly, considering a time delay estimation method to quickly find a useful signal, wherein the useful signal still contains partial frequency band noise, then, utilizing a wavelet denoising method to realize partial noise overlapped with the useful signal frequency spectrum, realizing the purpose of fine denoising, finally obtaining clean and effective seismic wave data, reflecting the actual situation of an underground geological structure and carrying out correct and reasonable detection on the geological situation.
The prior art is greatly different from the invention, the technical problem which is needed to be solved by the user cannot be solved, and a novel seismic denoising method based on MLReal transformation is invented for the purpose.
Disclosure of Invention
The invention aims to provide a MLReal transform-based seismic denoising method with obviously improved denoising effect.
The invention aims at realizing the following technical measures that the seismic denoising method based on MLReal transformation comprises the following steps of:
step 1, selecting an objective function;
step2, constructing a denoising neural network;
Step 3, redefining training data;
Step 4, redefining test data;
And 5, inputting redefined training data and test data into a neural network, and performing training loop iteration to obtain synthesized data with field data characteristics.
The aim of the invention can be achieved by the following technical measures:
the MLReal transform-based seismic denoising method also includes, before step 1, defining an error bound to ensure a minimum value of a difference measure in order to solve the difference between the distribution of the training data and the test dataset.
In defining the error bound, the formula used is:
εt(NN)≤εs(NN)+d(Ps(xs),Pt(xt))+λ
Epsilon t (NN) is expressed as a margin of error at test, epsilon s (NN) is expressed as a margin of error at training, P s(xs) is a probability distribution of training data, P t(xt) is a probability distribution of test data, d (,) is the distance between the test set and the training set margin distribution, lambda is the optimal joint error of the neural network model between the source data set and the target data set, and thus the upper error limit epsilon t (NN) of the test is guided by epsilon s (NN), d (,) and lambda by these three parameters.
In step 1, the network model is optimized by using the test error limit as an objective function.
In step 1, the objective function is:
εt(NN)≤εs(NN)+d(Ps(xs),Pt(xt))+λ
where ε t (NN) is denoted as the margin of error at test, ε s (NN) is denoted as the margin of error at training, P s(xs) is the probability distribution of training data, P t(xt) is the probability distribution of test data, d (,) is the distance between the test set and the marginal distribution of the training set, λ is the optimal joint error of the neural network model between the source data set and the target data set, and therefore the upper error limit ε t (NN) of the test is guided by ε s (NN), d (,) and λ by these three parameters.
In step 2, a 4-layer residual U-Net neural network is constructed, using 32 filters as initial filters, with two-fold increase in each layer, resulting in 106M trainable parameters, using 8092 training samples and 1024 validation samples to train the network, number epochs of training 150, and image block size 64×64.
In step 3, in the training phase, the training data is subjected to T s transformation, so that the field data features are injected into the synthesized data.
In step 3, T s is transformed
Where k is the reference trace index from the synthetic input data, fixed for both synthetic and real data, j is the partial index from real data, whether the partial is from a shot set or seismic imaging, operatorRepresenting cross-correlation, operator represents convolution, in this equation, at the composite profile of the inputCross-correlating with a reference trace from the profile and then convolving with the reference trace and with a randomly extracted (j) autocorrelation portion of d ij from the real data, i.e. each synthetic data block is cross-correlated with a single reference trace extracted from the data block before inputting NN, and upon training iterations, for each iteration, a random field data block is selected, autocorrelation is performed on a trace-by-trace basis, and convolved with each cross-correlated synthetic data block, which becomes the input to the network.
In step 3, the randomly chosen j index for the autocorrelation real data is varied in each iterative training to allow for proper representation of the real data features on the training set.
In step 4, in the reasoning stage, the characteristics of the synthesized data are injected into the target data by performing T t transformation on the test data.
In step 4, the T t transform is:
Where k is the reference track index from the synthetic input data, fixed for both synthetic and real data, j is the partial index from real data, operator Representing cross-correlation, operator x represents convolution, d ij is the real data, N s is the number of parts synthesized in the training set, i.e. for the inference phase the real data undergoes a preprocessing process similar to the training phase synthesis data, however, at this stage, once the field data blocks are cross-correlated with their reference tracks, these blocks are convolved with the blocks of autocorrelation averages of all synthesized training data blocks.
At step 4, the randomly chosen j index for the autocorrelation real data is varied in each iterative training to allow for proper representation of the real data features on the training set.
And 5, inputting redefined training data and test data into a neural network, and performing training loop iteration, so that a test error objective function reaches a minimum value, namely, obtaining synthesized data with field data characteristics, finally taking the synthesized data as the input of the network, and finally, removing noise.
According to the seismic denoising method based on MLReal transformation, MLReal transformation is used for reducing errors between synthesized training data and real application data, and as many field seismic data features as possible are injected into the synthesized data training and finally used for seismic denoising tasks. Specifically, the invention provides an optimization algorithm based on seismic denoising, which aims to train synthesized data by using a neural network while training the synthesized data, so that important information of field seismic data is combined into a training process of the synthesized data by carrying out MLReal transformation on the training data and the real data, and finally, a training success network is used for denoising the field seismic data, so that the denoising effect is obviously improved.
Drawings
FIG. 1 is a flow chart of one embodiment of a MLReal transform-based seismic denoising method of the present invention;
FIG. 2 is a schematic diagram of an example application of seismic denoising of synthetic seismic data in accordance with an embodiment of the present invention;
FIG. 3 is a schematic diagram of an example of the application of the invention in the denoising of vibroseis field seismic data in western desert areas of China;
fig. 4 is a schematic diagram of an application example of field seismic data denoising of explosive source in eastern region of china according to an embodiment of the present invention.
Detailed Description
It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the invention. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of exemplary embodiments according to the present invention. As used herein, the singular forms also are intended to include the plural forms unless the context clearly indicates otherwise, and furthermore, it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of stated features, steps, operations, and/or combinations thereof.
Based on MLReal algorithm seismic denoising method, the MLReal transformation is used to reduce the error between the synthesized training data and the real application data, and in order to keep the characteristics of the original wave field, the neural network is used to train the synthesized data and the MLReal transformation is performed, so that the important information of the field seismic data is combined into the training process of the synthesized data, and a good denoising effect is achieved.
As shown in fig. 1, fig. 1 is a flowchart of the seismic denoising method based on MLReal transform according to the present invention. In the flow of fig. 1, before training the model, in order to resolve the differences between the training data and the distribution of the test data sets, it is necessary to define a margin of error to ensure a minimum value of the difference metric:
εt(NN)≤εs(NN)+d(Ps(xs),Pt(xt))+λ
Epsilon t (NN) is denoted as the margin of error at test, epsilon s (NN) is denoted as the margin of error at training, P s(xs) is the probability distribution of training data, P t(xt) is the probability distribution of test data, d (,) is the distance between the test set and the training set margin distribution, and lambda is the optimal joint error of the neural network model between the source data set and the target data set. Thus, the upper error limit ε t (NN) of the test is guided by ε s (NN), d (,) and λ by these three parameters.
For most artificial intelligence, machine learning projects, it is essential to have a large and processed data set, but acquiring such data is often a significant challenge. Not only must data be collected from the real world, but also must be manually cleaned and annotated, and the composite data is much easier than the actual data collected and annotated, but the composite data may lack the behavior and characteristics of the actual data. The training data and the real data are subjected to MLReal transformation, so that important information of the field seismic data is combined into the training process of the synthesized data, and finally, a training success network is used for denoising the field seismic data, so that the denoising effect is obviously improved.
The training phase and the reasoning phase of the neural network are realized based on MLReal transformation to obtain the probability distribution difference reduction of training data and test data, and the realization process is as follows:
1) And selecting an objective function, namely optimizing the network model by taking the test error limit as the objective function, wherein the objective function is as follows:
εt(NN)≤εs(NN)+d(Ps(xs),Pt(xt))+λ
Where ε t (NN) is represented as the margin of error at test, ε s (NN) is represented as the margin of error at training, P s(xs) is the probability distribution of training data, P t(xt) is the probability distribution of test data, d (,) is the distance between the test set and the marginal distribution of training set, and λ is the optimal joint error of the neural network model between the source data set and the target data set. Thus, the upper error limit ε t (NN) of the test is guided by ε s (NN), d (,) and λ by these three parameters.
2) Constructing a denoising neural network, namely constructing a 4-layer residual U-Net neural network, wherein the network uses 32 filters as initial filters, two times of each layer is increased, about 106M trainable parameters are generated, the network is trained by using 8092 training samples and 1024 verification samples, the training frequency epochs is 150, and the size of an image block is 64 multiplied by 64.
3) Redefining the test data, namely, in the training stage, carrying out T s transformation on the training data so as to enable field data characteristics to be injected into the synthesized data. T s transform
Where k is the reference trace index from the synthetic input data (fixed for both synthetic and real data), j is the partial index from the real data, whether the partial is from the shot set or seismic imaging, operatorRepresenting cross-correlation and operator representing convolution. In this equation we are in the composite profile of the inputCross-correlating with a reference trace from the profile and then convolving with the reference trace and the randomly extracted (j) autocorrelation portion of d ij from the real data. I.e. each composite data block is cross-correlated with a single reference track extracted from the data block before the input NN. In training iterations, for each iteration, a random field data block is selected, autocorrelation is performed on a channel-by-channel basis, and convolved with each cross-correlated composite data block, which becomes the input to the network.
The randomly chosen j index for the autocorrelation of the real data is varied in each iteration of the training to allow for proper representation of the real data features on the training set.
4) Redefining the test data, namely, in the reasoning stage, the characteristics of the synthesized data are injected into the target data by carrying out T t transformation on the test data. T t transforms are:
where k is the reference track index from the synthetic input data (fixed for both synthetic and real data), j is the partial index from real data, operator Representing cross-correlation, operator x represents convolution, d ij is the true data, and N s is the number of parts synthesized in the training set. I.e. for the reasoning phase, the real data undergoes a preprocessing process similar to the training phase synthesis data. However, at this stage, once the field data blocks are cross-correlated with their reference tracks, these blocks are convolved with blocks of the autocorrelation mean of all the synthetic training data blocks.
The randomly chosen j index for the autocorrelation of the real data is varied in each iteration of the training to allow for proper representation of the real data features on the training set.
5) And inputting redefined training data and test data into a neural network, and performing training loop iteration, so that a test error objective function reaches a minimum value, and obtaining synthesized data with field data characteristics, wherein the synthesized data is finally used as network input and is finally used for denoising.
The following are several embodiments of the invention
Example 1:
As shown in fig. 2, a successfully trained network was used for denoising synthetic seismic data for laboratory simulation using a MLReal transform-based seismic denoising method. On the basis of the simulated noiseless seismic data, strong noise with a signal to noise ratio of 1:1 is added, as shown in FIG. 2 a. It can be seen that the in-phase axis of the shallow layer is still distinguishable, but the reflected signal of the middle deep layer is not. After seismic denoising based on MLReal transformation, as shown in fig. 2b, the reflection phase axis of the full-layer section, particularly the middle deep layer, is well recovered, and the hyperbolic morphology is clearly visible. Meanwhile, the signals of the middle and shallow layers are improved to a certain extent.
Example 2:
As shown in fig. 3, the successfully trained network is used for denoising the vibroseis field seismic data in western desert areas of china by using a MLReal transform-based seismic denoising method. After cross-correlation processing is performed on the vibroseis data, noisy data is obtained, as shown in fig. 3 a. The data has low signal-to-noise ratio, deep reflection is submerged in noise, and is not easy to identify. FIG. 3b is the denoised data, and it can be seen that the MLReal transform based seismic denoising method works well for shallow, medium, and deep layers. In the denoising process, the effective signal is not damaged. The overall signal-to-noise ratio and the continuity of the reflection phase axis are significantly improved.
Example 3:
As shown in fig. 4, the network successfully trained is used for denoising explosive source field seismic data in eastern region of china by using a seismic denoising method based on MLReal transformation. By carrying out MLReal transformation on the training data and the test data, the important information of the field seismic data is combined to the training process of the actual data, and finally the trained network is used for denoising the field seismic data. Fig. 4a is noisy data and fig. 4b is denoised data. The data signal-to-noise ratio is relatively high, but the noise in the middle-short offset range is still strong, the strong noise is obviously removed through the seismic denoising application based on MLReal transformation, and the reflection phase axis originally submerged in the noise is clearly visible. It can be seen that the quality of single shot gather before and after denoising is obviously improved.
The seismic denoising method based on MLReal algorithm comprises the steps of firstly transforming and converting training data through T s in the training process, secondly transforming actual data through similar T t transformation in the reasoning stage in order to enable the distribution of a changed training set and a test set to be similar, enabling important information of field seismic data to be combined into the training process of synthetic data through T s transformation and T t transformation which are jointly called MLReal transformation, and finally enabling a training success network to be used for denoising of the field seismic data, so that denoising effect is improved obviously.
It should be noted that the above description is only a preferred embodiment of the present invention, and not intended to limit the present invention, but although the present invention has been described in detail with reference to the above embodiment, it will be apparent to those skilled in the art that modifications may be made to the technical solutions described in the above embodiment, or equivalents may be substituted for some of the technical features thereof. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Other than the technical features described in the specification, all are known to those skilled in the art.
Claims (13)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310855828.4A CN119310628A (en) | 2023-07-13 | 2023-07-13 | Seismic denoising method based on MLReal transform |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310855828.4A CN119310628A (en) | 2023-07-13 | 2023-07-13 | Seismic denoising method based on MLReal transform |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN119310628A true CN119310628A (en) | 2025-01-14 |
Family
ID=94180096
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310855828.4A Pending CN119310628A (en) | 2023-07-13 | 2023-07-13 | Seismic denoising method based on MLReal transform |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN119310628A (en) |
-
2023
- 2023-07-13 CN CN202310855828.4A patent/CN119310628A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Mandelli et al. | Interpolation and denoising of seismic data using convolutional neural networks | |
| Van den Ende et al. | A self-supervised deep learning approach for blind denoising and waveform coherence enhancement in distributed acoustic sensing data | |
| CN110806602B (en) | Intelligent seismic data random noise suppression method based on deep learning | |
| CN112946749A (en) | Method for suppressing seismic multiples based on data augmentation training deep neural network | |
| Birnie et al. | Self-supervised learning for random noise suppression in seismic data | |
| Liu et al. | Coherent noise suppression via a self-supervised blind-trace deep learning scheme | |
| CN114418886A (en) | Robustness denoising method based on deep convolution self-encoder | |
| Chen et al. | Efficient seismic data denoising via deep learning with improved MCA-SCUNet | |
| Yan et al. | A method for denoising seismic signals with a CNN based on an attention mechanism | |
| Dong et al. | Seismic data reconstruction based on a multicascade self-guided network | |
| Liu et al. | Seismic random noise suppression based on deep image prior and total variation | |
| Liu et al. | A self‐supervised scheme for ground roll suppression | |
| Wang et al. | Seismic data denoising using a self-supervised deep learning network | |
| Cheng et al. | An effective self-supervised learning method for various seismic noise attenuation | |
| Dong et al. | Multistage residual network for intense distributed acoustic sensing background noise attenuation | |
| Wang et al. | EFGW-UNet: a deep-learning-based approach for weak signal recovery in seismic data | |
| CN119310628A (en) | Seismic denoising method based on MLReal transform | |
| Mao et al. | DenoiseNet: Deep generator and discriminator learning network with self-attention applied to ocean data | |
| Durall et al. | An in-depth study of U-net for seismic data conditioning: Multiple removal by moveout discrimination | |
| CN115205534A (en) | Seismic acquisition footprint denoising method based on unsupervised model driving type deep learning | |
| Zhong et al. | Mutual-guided scale-aggregation denoising network for seismic noise attenuation | |
| CN116359982A (en) | Convolutional neural network-based method for enhancing small quantity of source virtual cannon signals | |
| Jayakumar et al. | Karhunen Loeve Transform with adaptive dictionary learning for coherent and random noise attenuation in seismic data | |
| Yang et al. | Interpretable unsupervised learning framework for multi-dimensional erratic and random noise attenuation | |
| Wang et al. | A Survey of Random Noise Suppression Methods for Seismic Data based on Deep Learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |