CN117250657A - Seismic data reconstruction denoising integrated method - Google Patents

Seismic data reconstruction denoising integrated method Download PDF

Info

Publication number
CN117250657A
CN117250657A CN202311531698.5A CN202311531698A CN117250657A CN 117250657 A CN117250657 A CN 117250657A CN 202311531698 A CN202311531698 A CN 202311531698A CN 117250657 A CN117250657 A CN 117250657A
Authority
CN
China
Prior art keywords
seismic data
layer
data
patch
training
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.)
Granted
Application number
CN202311531698.5A
Other languages
Chinese (zh)
Other versions
CN117250657B (en
Inventor
张岩
张一鸣
杨壮
宋利伟
王鹏
孙宇航
董宏丽
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Sanya Offshore Oil And Gas Research Institute Of Northeast Petroleum University
Original Assignee
Sanya Offshore Oil And Gas Research Institute Of Northeast Petroleum University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Sanya Offshore Oil And Gas Research Institute Of Northeast Petroleum University filed Critical Sanya Offshore Oil And Gas Research Institute Of Northeast Petroleum University
Priority to CN202311531698.5A priority Critical patent/CN117250657B/en
Publication of CN117250657A publication Critical patent/CN117250657A/en
Application granted granted Critical
Publication of CN117250657B publication Critical patent/CN117250657B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V1/00—Seismology; Seismic or acoustic prospecting or detecting
    • G01V1/28—Processing seismic data, e.g. for interpretation or for event detection
    • G01V1/282—Application of seismic models, synthetic seismograms
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V1/00—Seismology; Seismic or acoustic prospecting or detecting
    • G01V1/28—Processing seismic data, e.g. for interpretation or for event detection
    • G01V1/36—Effecting static or dynamic corrections on records, e.g. correcting spread; Correlating seismic signals; Eliminating effects of unwanted energy
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Remote Sensing (AREA)
  • Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Acoustics & Sound (AREA)
  • Environmental & Geological Engineering (AREA)
  • Geology (AREA)
  • General Life Sciences & Earth Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Geophysics (AREA)
  • Geophysics And Detection Of Objects (AREA)

Abstract

The invention relates to the field of the intersection technology of the earth science and technology and artificial intelligence, in particular to a seismic data reconstruction denoising integrated method, which comprises the following steps: acquiring seismic data; inputting the seismic data into a preset seismic data recovery model, and outputting the seismic data after reconstruction and denoising, wherein the data recovery model is obtained based on training set training, the training set comprises ideal seismic data and training seismic data, and the data recovery model is constructed by adopting a Swin transducer to generate an countermeasure network. According to the invention, by designing the generator, the discriminator and the loss function, a data recovery model is constructed, so that the characteristics of global information and relevance of the seismic data and the like can be better utilized, and the problem that the recovered seismic data is fuzzy due to the fact that the receptive field is limited and the global information cannot be introduced in the prior model is solved.

Description

Seismic data reconstruction denoising integrated method
Technical Field
The invention relates to the field of the intersection technology of the earth science technology and artificial intelligence, in particular to an integrated method for reconstructing and denoising seismic data.
Background
The seismic exploration is a geophysical exploration method which is used for measuring the response of the earth to manually excited seismic waves through a detector, acquiring underground structure information and further analyzing an underground medium layer and is beneficial to determining the position of an oil and gas ring, and is an important means of petroleum exploration. In an ideal case, the sampling of the seismic signals should be of high quality, regular and dense, and the seismic data is often severely corrupted by random noise due to environmental disturbances and acquisition conditions. In addition, due to economic and physical limitations, the collected seismic traces are usually insufficient in sampling rate, and irregular or missing phenomena can occur. The processing and interpretation of the seismic data have high requirements on the signal-to-noise ratio and regularity of the data, so that the recovery, reconstruction and denoising of the missing channels of the seismic data are very critical problems, and the research on efficient seismic data reconstruction and denoising algorithms has great significance in improving the accuracy of seismic data processing.
In recent years, students have proposed methods of seismic data reconstruction and denoising, which can be roughly classified into two types of methods driven by conventional theory and data driving from experimental mechanisms. The traditional theory driven seismic data reconstruction and denoising algorithm is mostly based on the propagation characteristics of seismic waves and digital signal processing technology. The main categories can be divided into four categories: prediction filters, wave equations, mathematical transformations and matrix reduction ranks. The traditional seismic data reconstruction and denoising algorithm has the advantages of being interpretable and obvious in defects. First, algorithm accuracy typically relies on a priori information. If the prior information estimation is inaccurate, a model cannot be accurately built, and the algorithm accuracy is low; second, conventional denoising and reconstruction algorithms are less versatile. Since algorithms are often designed for specific data, algorithms may be less adaptable after encountering new types of data; finally, conventional algorithms typically require a problem model to be designed according to a specific data format to solve the problems of reconstruction and denoising, with slow prediction speed.
In recent years, data-driven based methods are also widely used in geophysical fields such as seismic data processing, seismic waveform classification, velocity model building, seismic data interpretation, and the like. Current seismic data denoising and reconstruction algorithms are largely divided into two types of methods, automatic Auto-Encoder (VAE) and generation of countermeasure networks (Generator Adversarial Network, GAN). Wherein the method of the automatic encoder comprises: si et al and Zhao et al use a denoising convolutional neural network (Denoising Convolutional Neural Networks, dnCNN) model to achieve the denoising task of seismic data and exhibit good effects in the actual denoising process. Wang et al reconstruct the missing rule seismic data using a residual network, but do a cubic spline interpolation of the missing rule locations in advance during network prediction. Chai et al reconstruct seismic data using U-Net, reconstructing irregularly missing seismic data. The Zhong et al propose a U-Net based on a residual structure, and the purpose of denoising is achieved by learning multi-scale features of seismic data to distinguish signals from random noise. Wang and Li propose a multi-scale end-to-end network based on a residual error network, and a denoising network and an interpolation network are respectively used for processing the seismic data at the same time, so that the tasks of denoising and reconstructing the seismic data are completed. Jiang et al propose an improved convolutional self-encoder method to achieve simultaneous reconstruction and denoising of seismic data, training uses Adam optimization algorithm to optimize the loss function, and reconstruction and denoising of seismic data are achieved. Mandell et al propose a CNN-based simultaneous reconstruction and denoising method, which selects the mean square error between the network output and the tag data as a loss function, and adjusts the network structure to realize the reconstruction and denoising of the seismic data. The Haoshao et al constructs an end-to-end U-Net structure, proposes a seismic data simultaneous reconstruction and denoising algorithm based on a Huber loss function, and introduces a attention module to further enhance the feature extraction capability of the seismic signals.
In addition to this, there are some methods of reconstruction and denoising of GAN-based seismic data, but less research is underway. GAN adds a discriminator in the generator, introducing regional or global information, so it has better performance, especially for continuous and large-scale defects. Oliveira et al have realized interpolation and reconstruction of the Netherlands offshore seismic data via cGAN. Wei et al change the resistance loss to the Wasserstein loss based on cGAN and achieve two-dimensional interpolation of up to 35 consecutive seismic trace deletions. Kaur et al interpolate two-dimensional synthetic seismic data using GAN and CycleGAN. Most of the methods based on data driving use convolution kernels to learn the characteristics of the seismic data, and although the mapping relation between undersampled noisy seismic data and complete noiseless seismic data can be effectively learned, the long-distance dependence of the seismic data cannot be established by the method based on convolution data driving due to the limited convolution receptive field, and the problems of incomplete denoising, inaccurate texture details of the reconstructed seismic data and the like still exist.
Inspired by the superior performance of the transducer in the field of natural language processing (Natural Language Processing, NLP), researchers introduced the transducer into the field of Computer Vision (CV). Vision Transformer as a feature extractor for image recognition, which takes a 2D image block as a model input and pre-trains using a large dataset such as ImageNet, gives better results than CNN, however, these methods are computationally intensive and do not take into account the context information of the local features of the seismic data. The Swin transducer can solve the problems through a moving window mechanism, and achieves a good effect in the field of general image denoising.
Therefore, based on the Swin transducer generation countermeasure network, a seismic data reconstruction denoising integrated method is provided.
Disclosure of Invention
The invention aims to provide an integrated method for reconstructing and denoising seismic data, which combines a Swin transform and a cGAN network to be introduced into the reconstruction and denoising of the seismic data, and solves the problems that the convolutional network has limited receptive field and the recovered seismic data is fuzzy due to the fact that global information cannot be introduced.
In order to achieve the above object, the present invention provides the following solutions:
an integrated method for reconstructing and denoising seismic data, comprising the following steps:
acquiring seismic data;
inputting the seismic data into a preset seismic data recovery model, and outputting the seismic data after reconstruction and denoising, wherein the data recovery model is obtained based on training set training, the training set comprises ideal seismic data and training seismic data, and the data recovery model is constructed by adopting a Swin transducer to generate an countermeasure network.
Further, the training seismic data is obtained by sampling the ideal seismic data, the sampling the ideal seismic data comprising:
randomly dividing the ideal seismic data into three parts, uniformly sampling a first part, setting a first noise standard deviation coefficient, and adding random noise which is not more than the first noise standard deviation coefficient and has different intensities;
Randomly sampling the second part, setting a second noise standard deviation coefficient, and adding random noise which is not more than the second noise standard deviation coefficient and has different intensities;
and carrying out local sampling on the third part, setting a third noise standard deviation coefficient, and adding random noise which is not more than the third noise standard deviation coefficient and has different intensities.
Further, before training the data recovery model based on the training set, the method further comprises preprocessing data in the training set, wherein the preprocessing comprises the following steps:
and segmenting the training seismic data and the ideal seismic data in the training set into two-dimensional seismic data with the size of NT, and acquiring a training seismic data set and a theoretical seismic data set.
Further, the seismic data recovery model includes: the generator is used for processing the training set to acquire the seismic data and the content loss generated by the generator, and comprises a Patch Partition layer, an encoder, a bottleneck layer, a decoder and a Linear Projection layer which are connected in sequence;
the discriminator is used for judging the seismic data generated by the training set and the generator, acquiring errors, acquiring countermeasures by combining the content measures, optimizing the countermeasures counter propagation guide generator, and comprises a Patch Partition layer, a Patch Embedding layer, a Patch measuring layer and an MLP layer which are sequentially connected, and a first switch transform module which is connected after the Patch Embedding layer and after the Patch measuring layer.
Further, the encoder comprises a Patch Embedding layer, a second switch transformation module and a Patch Merging layer, wherein the second switch transformation module is placed between the Patch Embedding layer and the Patch Merging layer and between the Patch Merging layer and the Patch Merging layer;
the bottleneck layer comprises a third Swin Transformer module;
the decoder comprises a Patch expansion layer and a fourth Swin transform module, wherein the fourth Swin transform module is placed between the Patch expansion layer and the Patch expansion layer;
the second Swin transducer module and the fourth Swin transducer module are connected by a residual.
Further, the content loss is:
wherein,to evaluate the loss of denoising effect of non-missing parts, < >>To evaluate loss of effect of the partial reconstruction of the deletion +.>Is used for maintaining parameters on the same magnitude of the two;
wherein,x i in order to train the samples in the seismic dataset,y i for a sample in an ideal set of seismic data,i=1,2,3,…,N，G（x i ) Is thatx i The output of the sample down-generator,for Hadamard product, the operation rule is multiplication of matrix corresponding positions, ++>Mask matrix for missing seismic data, +.>Is->And taking the inverse.
Further, the countering loss is:
Wherein,for training the seismic dataset, < >>For an ideal seismic dataset, +.>Output of generator, ++>Data pairs input for the discriminator pairs +.>Is the discrimination result of->Data pairs input for the discriminator pairs +.>Is the discrimination result of->For the desired value->Is a probability distribution of real seismic data.
Further, the first Swin Transformer module, the second Swin Transformer module, the third Swin Transformer module and the fourth Swin Transformer module are all used for performing self-attention mechanism calculation, and the method for performing the self-attention mechanism calculation is as follows:
wherein,and->For seismic data in different windows, +.>Is->And->Degree of association between->Trainable variable greater than 0.01, < >>Encoding the relative positions of the different positions.
The beneficial effects of the invention are as follows:
according to the invention, the transducer is introduced into the field of seismic data reconstruction and denoising, and theoretical analysis and experiments prove that the transducer has feasibility in the seismic data reconstruction and denoising work, and has stronger competitiveness compared with a mainstream convolution method.
According to the invention, a network model is designed and improved, and Swin transform is introduced to replace convolution operation so as to improve the utilization efficiency of the model on the global information of the seismic data; the encoder is used for carrying out multi-scale feature extraction on the seismic data, the characteristics of the same-phase axis and texture details of the seismic data are considered by utilizing the characteristics of the U-Net structure, and the quality of the seismic data recovered by the generator is enhanced; compared with the existing similar reconstruction and denoising methods, the method has certain advantages in terms of result precision and model generalization.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the drawings that are needed in the embodiments will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present invention, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of a seismic data recovery model construction in accordance with an embodiment of the present invention;
FIG. 2 is a block diagram of a seismic data recovery model in accordance with an embodiment of the invention;
FIG. 3 is a block diagram of a generator according to an embodiment of the present invention;
FIG. 4 is a diagram of a arbiter in accordance with an embodiment of the present invention;
FIG. 5 is a block diagram of a Swin transducer module according to an embodiment of the present invention;
FIG. 6 is a schematic diagram of a Swin transducer module according to an embodiment of the present invention;
FIG. 7 is a graph of the comparative recovery effect of the different methods of the embodiments of the present invention on the random sampled BP2004 synthetic seismic data, wherein FIG. 7 (a) is an ideal seismic data partial area enlargement, FIG. 7 (b) is a missing noisy seismic data partial area enlargement, FIG. 7 (c) is a seismic data recovery model recovery seismic data partial area enlargement, SNR is 20.3dB, FIG. 7 (d) is a SUNet recovery seismic data partial area enlargement, SNR is 15.2dB, FIG. 7 (e) is a WGAN recovery seismic data partial area enlargement, SNR is 16.2dB;
FIG. 8 is a graph showing how the recovery effect of the different methods of the embodiment of the present invention on the random sampled Marmousi synthetic seismic data is compared, wherein FIG. 8 (a) is ideal seismic data, FIG. 8 (b) is noise-containing missing seismic data with a noise intensity of 0.1 and a sampling ratio of 0.5, FIG. 8 (c) is seismic data recovered by SUNet, the SNR is 7.5dB, FIG. 8 (d) is seismic data recovered by WGAN, the SNR is 15.3dB, FIG. 8 (e) is seismic data recovered by a seismic data recovery model, and the SNR is 17.6dB;
FIG. 9 is a graph comparing the recovery effect of locally sampled Marmousi synthetic seismic data by various methods in accordance with embodiments of the invention, wherein FIG. 9 (a) is an ideal seismic data; FIG. 9 (b) is seismic data with a noise intensity of 0.1 and a partial loss of 20 traces, with an SNR of-1.0 dB; FIG. 9 (c) is SUNet recovered seismic data with an SNR of 13.6dB; FIG. 9 (d) is the seismic data recovered by the WGAN with an SNR of 12.8dB; FIG. 9 (e) is the seismic data recovered by the seismic data recovery model, with an SNR of 20.9dB;
fig. 10 is a graph of recovery effects of a seismic data recovery model on randomly sampled real seismic data according to the third embodiment of the invention, in which fig. 10 (a) is ideal seismic data, fig. 10 (b) is real seismic data with added noise intensity of 0.1 and sampling ratio of 0.5, fig. 10 (c) is a graph of recovery effects of the seismic data recovery model, SNR is 16.8dB, fig. 10 (d) is a graph of residual profile waveforms of the recovery effects of the seismic data recovery model and the real data, and fig. 10 (e) is a spectrum diagram corresponding to residual errors of the real data.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
In order that the above-recited objects, features and advantages of the present invention will become more readily apparent, a more particular description of the invention will be rendered by reference to the appended drawings and appended detailed description.
Embodiment one:
the embodiment provides a seismic data reconstruction denoising integrated method, which comprises the following steps:
acquiring seismic data;
and inputting the seismic data into a preset seismic data recovery model, and outputting the seismic data after reconstruction and denoising.
As shown in fig. 1, the construction of the seismic data recovery model includes the steps of:
step 1, constructing a training data set;
according to the property of the simultaneous reconstruction and denoising tasks, a nonlinear mapping model from missing and noisy seismic data to ideal seismic data is established, and a large number of sample pairs of missing and noisy seismic data and ideal seismic data (labels) are required.
The ideal sample data set Sei _data_c sources for this embodiment include: (1) a synthesized seismic dataset, (2) a standard model forward data set, (3) an actually acquired and processed seismic dataset;
in order to simulate different conditions of missing channels and noise interference of the seismic data in the actual acquisition process, the following processing is carried out on the seismic data set Sei _data_c:
(1) Seismic data with the random extraction proportion of dataset_percentage_A is uniformly sampled with the sampling rate of dataset_sample_A, random noise with different intensities is added, and the maximum noise standard deviation coefficient is noise_maxdev_A;
(2) Extracting seismic data with the ratio of dataset_percentage_B, randomly sampling the sampling ratio dataset_sample_B, adding random noise with different intensities, wherein the maximum noise standard deviation coefficient is noise_maxdev_B;
(3) And (3) locally sampling the rest seismic data with the proportion of dataset_percentage_C, adding random noise with different intensities, and obtaining training seismic data Sei _data_init, wherein the noise standard deviation coefficient is noise_maxdev_C at the maximum.
The training seismic data Sei _data_init and samples in the ideal seismic data Sei _data_c are segmented into two-dimensional seismic data with the size of NT, so that a training seismic data set Sei _train_data and a theoretical seismic data set Sei _data are obtained, the training set not only can effectively describe possible missing and noise conditions of the seismic data in the acquisition process, but also can improve the calculation efficiency of a self-attention mechanism and improve the processing quality of recovered seismic data.
Step 2, network input and label setting;
(1) The input of the generator network is training seismic data set Sei _train_data, the label is ideal seismic data set Sei _data corresponding to the training seismic data set, the output of the generator is Sei _outputs, and the content loss L obtained through calculation of the input and the label C The training seismic data set Sei _train_data and the output Sei _output of the generator are spliced according to the channel and then are input into the discriminator together to obtain the resistance loss L of the generator cGAN_G ；
(2) The arbiter network (1) takes a training seismic data set Sei _train_data and a corresponding ideal seismic data set Sei _data as input after being spliced according to channel dimensions, and an error errD_real is obtained; (2) training seismic data set Sei _train_data and corresponding generator output Sei _output are spliced according to channels and then used as input, and the obtained error errD_fake, errD_fake and errD_real jointly form the antagonism loss L of the discriminator cGAN_D 。
Step 3, generating an countermeasure network based on the Swin Transformer, and constructing a seismic data recovery model;
as shown in fig. 2, the present embodiment uses a condition countermeasure network as a base network, builds a seismic data recovery model, and the model is divided into a generator and a discriminator section, specifically including:
(1) Generator M G
As shown in fig. 3, generator M G Is composed of three parts, including downsampled encoder, upsampled decoder and bottleneck layer, and is composed of the Patch Partition layer, encoder, bottleneck layer, decoder and LinAn ear project layer;
the coder comprises a Patch Embedding layer, a switch conversion former module and a Patch Messing layer, wherein the switch conversion former module is arranged between the Patch Embedding layer and the Patch Messing layer and between the Patch Messing layer and the Patch Messing layer;
the bottleneck layer comprises a Swin transducer module;
the decoder comprises a Patch expansion layer and a Swin transform module, wherein the Swin transform module is arranged between the Patch expansion layer and the Patch expansion layer;
the Swin transducer module in the encoder and the Swin transducer module in the decoder are connected by a residual.
The working process is as follows:
(1) The seismic data is input into a network, and H.times.W.times.1 two-dimensional seismic data are converted into H/4*W/4.times.16 slice seismic data by a Patch Partition layer.
(2) The method comprises the steps of (1) slicing seismic data into an encoder part, firstly converting the seismic data into one-dimensional sequence characteristics through a Patch Embedding layer, and setting the seismic data into a learnable parameter to learn to perform characteristic representation on the seismic data and improve the channel dimension to C because the seismic data is complex and the slices cannot be directly embedded E So that the subsequent Swin transducer can perform self-attention calculation on the Swin transducer; (2) the embedded slice seismic data are input into a Swin transducer module, relative position codes are added to each slice seismic data in order to better utilize the position information in the seismic data, indexes are carried out according to the corresponding relative position coding tables, and self-attention mechanisms are carried out on different areas in a window to obtain re-weighted slice seismic data; (3) seismic data is more correlated and content more complex than general-purpose data (general-purpose image, speech data). Thus adopt depth_G E The layer Swin transducer module processes slice seismic data to enhance the extraction effect of the self-attention mechanism; (4) downsampling the re-weighted slice seismic data through a Patch Merging layer to realize multi-scale feature extraction of the seismic data, and fully utilizing the position information in the seismic data; (5) through N E And (3) after the processing of the times (3) and (4), obtaining slice seismic data under different scales after the self-attention mechanism processing.
(3) Slice seismic data is transferred to bottleneck layer via N B Secondary depth_g B The Swin transducer module of the layer performs self-attention mechanism calculation on the layer and inputs the self-attention mechanism calculation into a decoder.
(4) The method comprises the steps of (1) inputting slice seismic data processed by a bottleneck layer into a Patch Expanding layer for up-sampling and Expanding the slice seismic data and reducing the channel number of the slice seismic data; (2) connecting the slice seismic data transferred by the residual connection with the slice seismic data corresponding to the size of the Patch expansion layer in the channel dimension, and then inputting depth_G D The calculation of the self-attention mechanism is carried out in a Swin transducer module of the layer; (3) through N D Obtaining the seismic data with the same size as the seismic data of the initial input slice after the processing of the times (1) and (2); (4) and carrying out channel dimension transformation on the seismic data through a Linear Projection layer to obtain the seismic data generated by the generator.
(2) Discriminator M D
As shown in FIG. 4, the arbiter M D And judging whether the input sample is ideal seismic data or the seismic data generated by the generator, outputting different values according to different judging results, and counter-propagating so as to guide the optimized direction of the generator. The discriminator comprises a Patch Partition layer, a Patch Embedding layer, an MLP layer and a switch transform module which are connected in sequence and behind the Patch Embedding layer.
The working process is as follows:
(1) Splicing the missing noisy seismic data and the corresponding ideal seismic data (or the missing noisy seismic data and the seismic data generated by the generator) according to the channel dimension, and converting the two-dimensional seismic data of H.times.W.times.2 into slice seismic data of H/4*W/4.times.32 through a Patch Partition layer; (2) slice seismic data is converted into one-dimensional sequence features through a Patch Embedding layer, learning is carried out by using learnable parameters to carry out feature representation on the seismic data, and channel dimension is increased to C D So that the subsequent Swin converters self-annotate the sameCalculating the meaning force; (3) through depth_D 1 The layer Swin transducer module adds relative position codes to each slice of seismic data, indexes the slice of seismic data according to a corresponding relative position code table, and calculates self-attentive mechanisms of different areas in a window to obtain re-weighted slice of seismic data; (4) downsampling the re-weighted slice seismic data through a Patch Merging layer to realize multi-scale feature extraction of the seismic data, and fully utilizing the position information in the seismic data; (5) repeating the steps (3) and (4) for 2 times to finish the feature extraction task of different scales of the slice seismic data, wherein the layer numbers of the Swin transducer module in (3) are respectively provided with depth_D 2 And depth_D 3 . (6) The characteristic passing layer number of the slice seismic data processed in the step (5) is depth_D 4 The Swin transducer module obtains the feature vector of the sliced seismic data, and inputs the feature vector into an MLP layer with an input layer of in_features dimension, a hidden layer of his_features dimension and an output layer of out_features dimension, and an activation function of Gelu to judge whether the input seismic data is the seismic data generated by the generator or ideal seismic data. Because the classification task is mainly executed in the discriminator, compared with the task of generating the seismic data in the generator, the method has the advantages that the training speed can be increased and the calculation efficiency can be improved by setting the layer number.
Step four: setting a loss function;
the loss function of the overall network is:
the loss function is mainly divided into two parts: (1) generating a resistance loss against a network(2) content loss of network-generated and ideal seismic data +.>。/>Is->Weight coefficient of (2), remain->And->On the same order of magnitude.
Loss of resistanceThe specific form of (2) is as follows:
wherein,for training the seismic dataset Sei _train_data +.>Sei _data, < ++for ideal seismic dataset>Output Sei _outputs of the generator, +.>Data pairs input for the discriminator pairs +. >Is used for judging the result of the judgment,data pairs input for the discriminator pairs +.>Is the discrimination result of->For the desired value->Is a probability distribution of real seismic data. According to the characteristics of simultaneous reconstruction and denoising of the seismic data, noise interference is not required to be considered when the seismic channel is required to be reconstructed, and only the reconstruction effect is required to be evaluated; the missing channels are not required to be considered in the seismic data needing to be denoised, and only the denoising effect needs to be evaluated. Content loss function based on the above-mentioned characteristics>The form is as follows:
wherein,to evaluate the loss of denoising effect of non-missing parts, < >>To evaluate loss of effect of the partial reconstruction of the deletion +.>The parameters are used for keeping the same magnitude of the two parameters, and the specific forms are as follows:
wherein,x i in order to train the samples in the seismic dataset,y i for a sample in an ideal set of seismic data,i=1,2,3,…,N，G（x i ) Is thatx i The output of the sample down-generator,the operation rule is multiplication of matrix corresponding positions for Hadamard product (Hadamard product)>For the mask matrix of the missing seismic data, the matrix shape is the same as the shape of the Sei _data slice of the seismic data, the missing channel value is 0, and the rest are 1./>Is->And taking the inverse.
Step five: model training and testing;
the training seismic data set Sei _train_data is divided into a training set Sei _train and a Test set Sei _test according to train_test_rate scale parameters.
(1) Because Sei _train is formed by combining multiple data sources (three parts), the data distribution difference is large, and the data is standardized when being input into a network, so that the subsequent training is facilitated. In the training process, model parameters are initialized by using a model pre-trained on a general data set, so that the training convergence speed is increased, and the recovery effect of seismic data is improved.
(2) In the training process, as shown in fig. 5 and 6, after the Layer Norm Layer in the Swin Transformer module is tuned to the W-MSA Layer, the MLP Layer and the SW-MSA Layer, the embodiment is used for solving the problem that the model cannot complete training due to the fact that the distribution mean value of the feature map output by the Swin Transformer after self-attention is calculated is too large in input distribution mean value difference.
(3) The self-attention mechanism is calculated by adopting a new calculation mode instead of the traditional dot product form. For seismic data with large data difference, the calculated feature map is often dominated by a few pixels with large values, and the recovery effect of the part with small values is poor. The new calculation mode formula is as follows:
when self-attention mechanism is calculated, slice seismic data is divided into different windows, whereinAnd->Are all seismic data within different windows. / >Is->And->Degree of association between->Is a trainable variable greater than 0.01,encoding the relative positions of the different positions.
(4) The generator and the discriminator need to train alternately, T is trained first D Wheel discriminator, retrain T G A wheel generator. Inputting the missing noisy seismic data and the corresponding ideal seismic data (or the missing noisy seismic data and the seismic data generated by the generator) into a discriminator according to the number of channels, and obtaining the antagonism loss by forward propagationUpdating parameters of the discriminator through back propagation; inputting the missing noisy seismic data into a generator, and forward propagating to obtain content loss +.>And resistance loss->The parameters of the generator are updated by back propagation until the arbiter is unable to distinguish between the seismic data generated by the generator and the ideal seismic data. After training is completed, the generator can realize the task of reconstructing and denoising the missing noisy seismic data. Saving the model and using the model to Test the Test seismic data Sei _test with SNAnd R is an index to evaluate the recovery effect of the seismic data, and the experimental result is recorded. Wherein the SNR calculation formula is as follows:
。
embodiment two:
the embodiment provides a seismic data reconstruction denoising integrated method, which comprises the following steps:
Acquiring seismic data;
and inputting the seismic data into a preset seismic data recovery model, and outputting the seismic data after reconstruction and denoising.
As shown in fig. 1 to 6, the construction of the seismic data recovery model specifically includes the following steps:
step 1, constructing a training data set;
according to the property of the simultaneous reconstruction and denoising tasks, a nonlinear mapping model from missing and noisy seismic data to ideal seismic data is established, and a large number of sample pairs of missing and noisy seismic data and ideal seismic data (labels) are required.
The ideal sample data set Sei _data_c sources of the present invention include: (1) a synthetic seismic dataset, (2) a standard model forward dataset, and (3) an actual acquired and processed seismic dataset.
In order to simulate different conditions of missing channels and noise interference of the seismic data in the actual acquisition process, the following processing is carried out on the seismic data set Sei _data_c:
(1) Uniformly sampling the seismic data of the first part with the sampling proportion of [0.3,0.8], adding random noise with different intensities, wherein the maximum standard deviation coefficient of the noise is 0.3;
(2) Carrying out random sampling of sampling proportion [0.5,0.8] on the seismic data of the second part, adding random noise with different intensities, wherein the maximum standard deviation coefficient of the noise is 0.2;
(3) And carrying out local sampling on the third part of seismic data, adding random noise with different intensities, wherein the maximum noise standard deviation coefficient is 0.3, and obtaining training seismic data Sei _data_init.
The training seismic data Sei _data_init and samples in the ideal seismic data Sei _data_c are segmented into two-dimensional seismic data with the size of 256 x 256, and a training seismic data set Sei _train_data and a theoretical seismic data set Sei _data are obtained.
Step 2, network input and label setting;
(1) The input of the generator network is training seismic data set Sei _train_data, the label is corresponding ideal seismic data set Sei _data, the output of the generator is Sei _output, the content loss LC obtained through calculation of the input and the label is input, and the training seismic data set Sei _train_data and the output Sei _output of the generator are spliced according to a channel and then are input into a discriminator together, so that the generator resistance loss LcGAN_G is obtained;
(2) The arbiter network (1) takes a training seismic data set Sei _train_data and a corresponding ideal seismic data set Sei _data as input after being spliced according to channel dimensions, and an error errD_real is obtained; (2) the training seismic data set Sei _train_data and the corresponding generator output Sei _output are spliced according to the channel and then are used as input, and the obtained error errD_fake, errD_fake and errD_real jointly form the antagonism loss LcGAN_D of the discriminator.
Step 3, generating an countermeasure network based on the Swin Transformer, and constructing a seismic data recovery model;
the invention adopts a condition countermeasure network as a basic network, a network model is divided into a generator and a discriminator part, and the network model is set as follows:
(1) Generator M G
Generator M G The system comprises three parts, namely a downsampled encoder, an upsampled decoder and a bottleneck layer, and specifically comprises a Patch Partition layer, an encoder, a bottleneck layer, a decoder and a Linear Projection layer which are connected in sequence;
the coder comprises a Patch Embedding layer, a switch conversion former module and a Patch Messing layer, wherein the switch conversion former module is arranged between the Patch Embedding layer and the Patch Messing layer and between the Patch Messing layer and the Patch Messing layer;
the bottleneck layer comprises a Swin transducer module;
the decoder comprises a Patch expansion layer and a Swin transform module, wherein the Swin transform module is arranged between the Patch expansion layer and the Patch expansion layer;
the Swin transducer module in the encoder and the Swin transducer module in the decoder are connected by a residual.
The working process is as follows:
(1) The seismic data is input into a network, and the Patch Partition layer converts 256×256×1 two-dimensional seismic data into 64×64×16 slice seismic data.
(2) The method comprises the steps of (1) slicing seismic data, entering an encoder part, firstly converting the seismic data into one-dimensional sequence characteristics through a Patch Embedding layer, and setting the seismic data into learnable parameters for learning to perform characteristic representation on the seismic data and raising the channel dimension to 96 so as to facilitate the follow-up self-attention calculation of a Swin converter because the seismic data is complex and the slices cannot be directly embedded; (2) the embedded slice seismic data are input into a Swin transducer module, relative position codes are added to each slice seismic data in order to better utilize the position information in the seismic data, indexes are carried out according to the corresponding relative position coding tables, and self-attention mechanisms are carried out on different areas in a window to obtain re-weighted slice seismic data; (3) seismic data is more correlated and content more complex than general-purpose data (general-purpose image, speech data). Therefore, 8 layers of Swin transducer modules are adopted to process slice seismic data so as to enhance the extraction effect of a self-attention mechanism; (4) downsampling the re-weighted slice seismic data through a Patch Merging layer to realize multi-scale feature extraction of the seismic data, and fully utilizing the position information in the seismic data; (5) and (3) processing for 3 times (3) (4), obtaining slice seismic data under different scales after self-attention mechanism processing.
(3) Slice seismic data are transmitted to a bottleneck layer, subjected to self-attention mechanism calculation by a Swin transducer module of 2 times 4 layers, and then input to a decoder.
(4) The method comprises the steps of (1) inputting slice seismic data processed by a bottleneck layer into a Patch Expanding layer for up-sampling and Expanding the slice seismic data and reducing the channel number of the slice seismic data; (2) connecting slice seismic data transmitted by residual connection with slice seismic data corresponding to the size of a Patch expansion layer in a channel dimension, and then inputting the slice seismic data into a Swin transducer module of 8 layers to calculate a self-attention mechanism; (3) obtaining the seismic data with the same size as the seismic data of the initial input slice after 3 times of (1) and (2) processing; (4) and carrying out channel dimension transformation on the seismic data through a Linear Projection layer to obtain the seismic data generated by the generator.
(2) Discriminator M D
Discriminator M D And judging whether the input sample is ideal seismic data or the seismic data generated by the generator, outputting different values according to different judging results, and counter-propagating so as to guide the optimized direction of the generator. The discriminator comprises a Patch Partition layer, a Patch Embedding layer, an MLP layer and a switch transform module which are connected in sequence and behind the Patch Embedding layer.
The working process is as follows:
(1) splicing the missing noisy seismic data and the corresponding ideal seismic data (or the missing noisy seismic data and the seismic data generated by the generator) according to the channel dimension, and converting the 256-by-2 two-dimensional seismic data into 64-by-32 slice seismic data through a Patch Partition layer; (2) the slice seismic data is converted into one-dimensional sequence features through a Patch Embedding layer, the learnable parameters are used for learning to perform feature representation on the seismic data, and the channel dimension is raised to 96 so that the subsequent Swin transducer can perform self-attention calculation on the seismic data; (3) adding a relative position code to each slice of seismic data through a 2-layer Swin transducer module, indexing according to a corresponding relative position code table, and calculating self-attentive mechanisms of different areas in a window to obtain re-weighted slice of seismic data; (4) downsampling the re-weighted slice seismic data through a Patch Merging layer to realize multi-scale feature extraction of the seismic data, and fully utilizing the position information in the seismic data; (5) and (3) repeating the steps (3) and (4) for 2 times to finish the feature extraction task of different scales of the slice seismic data, wherein the number of layers of the Swin transducer module in (3) is respectively set to 2 and 6. (6) The characteristics of the sliced seismic data processed in the step (5) are processed through a Swin transducer module with the layer number of 2 to obtain sliced seismic data characteristic vectors, and the sliced seismic data characteristic vectors are input into an input layer MLP layer to judge whether the input seismic data are generated by a generator or ideal seismic data. Because the classification task is mainly executed in the discriminator, compared with the task of generating the seismic data in the generator, the method has the advantages that the training speed can be increased and the calculation efficiency can be improved by setting the layer number.
Step four: setting a loss function;
the loss function of the overall network is:
the loss function is mainly divided into two parts: (1) generating a resistance loss against a network(2) content loss of network-generated and ideal seismic data +.>。/>Is->Weight coefficient of (2), remain->And->On the same order of magnitude.
Loss of resistanceThe specific form of (2) is as follows:
wherein,for training the seismic dataset Sei _train_data +.>Sei _data, < ++for ideal seismic dataset>Output Sei _outputs of the generator, +.>Data pairs input for the discriminator pairs +.>Is the discrimination result of->Data pairs input for the discriminator pairs +.>Is the discrimination result of->For the desired value->Is a probability distribution of real seismic data. According to the characteristics of simultaneous reconstruction and denoising of the seismic data, noise interference is not required to be considered when the seismic channel is required to be reconstructed, and only the reconstruction effect is required to be evaluated; the missing channels are not required to be considered in the seismic data needing to be denoised, and only the denoising effect needs to be evaluated. Content loss function based on the above-mentioned characteristics>The form is as follows:
wherein the method comprises the steps ofTo evaluate the loss of denoising effect of non-missing parts, < >>In order to evaluate the loss of the effect of the missing part reconstruction,the parameters are used for keeping the same magnitude of the two parameters, and the specific forms are as follows:
Wherein,x i in order to train the samples in the seismic dataset,y i for a sample in an ideal set of seismic data,i=1,2,3,…,N，G（x i ) Is thatx i The output of the sample down-generator,the operation rule is multiplication of corresponding positions of a matrix, the matrix is a mask matrix of missing seismic data, the shape of the matrix is identical to that of a Sei _data slice of the seismic data, the number of missing channels is 0, and the rest is 1./>Is->And taking the inverse.
Step five: training and testing a network;
the training seismic data set Sei _train_data is divided into a training set Sei _train and a Test set Sei _test according to a 8:2 scale parameter.
(1) Because Sei _train is formed by combining multiple data sources (three parts), the data distribution difference is large, and the data is standardized when being input into a network, so that the subsequent training is facilitated. In the training process, model parameters are initialized by using a model pre-trained on a general data set, so that the training convergence speed is increased, and the recovery effect of seismic data is improved.
(2) In the training process, after the Layer Norm Layer in the Swin converter module is adjusted to the W-MSA Layer, the MLP Layer and the SW-MSA Layer, the method is used for solving the problem that the model cannot finish training due to the fact that the distribution mean value of the feature map output by the Swin converter after self-attention is calculated is too large in input distribution mean value difference.
(3) The self-attention mechanism is calculated by adopting a new calculation mode instead of the traditional dot product form. For seismic data with large data difference, the calculated feature map is often dominated by a few pixels with large values, and the recovery effect of the part with small values is poor. The new calculation mode formula is as follows:
when self-attention mechanism is calculated, slice seismic data is divided into different windows, whereinAnd->Are all seismic data within different windows. />Is->And->Degree of association between->Is a trainable variable greater than 0.01,encoding the relative positions of the different positions.
(4) The generator and the discriminant need to be trained alternately, 10 rounds of discriminants are trained first, and then 1 round of generator is trained. Inputting the missing noisy seismic data and the corresponding ideal seismic data (or the missing noisy seismic data and the seismic data generated by the generator) into a discriminator according to the number of channels, and obtaining the antagonism loss by forward propagationUpdating parameters of the discriminator through back propagation; inputting the missing noisy seismic data into a generator, and forward propagating to obtain content loss +.>And resistance loss->The parameters of the generator are updated by back propagation until the arbiter is unable to distinguish between the seismic data generated by the generator and the ideal seismic data. After training is completed, the generator can realize the task of reconstructing and denoising the missing noisy seismic data. And (3) storing the model, testing the Test seismic data Sei _test by using the model, evaluating the recovery effect of the seismic data by taking the SNR as an index, and recording the experimental result. Wherein the SNR calculation formula is as follows:
。
Embodiment III:
experimental platform configuration of this embodiment: the computer operating system is Ubutu18.04, the GPU is NVIDA GTX-3080, the deep learning network model is built by using pytorch1.7.0 and python3.8, and the specific process is implemented as follows:
1. preprocessing a seismic data training set;
1.1 reading the mat file;
the mat-format File in the theoretical seismic dataset Sei _data is read by the File function of the toolkit h5 py. The calling mode is as follows:
Patch = h5py.File(file_name)['data']
and storing the read seismic data in a Patch, wherein file_name is a seismic data file, and data is a variable name of the file in matlab.
1.2 pre-processing seismic data;
in order to simulate different situations of the seismic data in the actual acquisition process, the theoretical seismic data set Sei _data is processed as follows: the seismic dataset was randomly divided into three parts by proportions of 30%,30% and 40%. (1) Uniformly sampling the seismic data of the first part with the sampling rate of [0.3,0.8], and adding random noise with different intensities, wherein the maximum standard deviation coefficient of the noise is 0.3; (2) Carrying out random sampling of the sampling proportion [0.5,0.8] on the seismic data of the second part, adding random noise with unequal intensities, wherein the maximum standard deviation coefficient of the noise is 0.2; (3) The seismic data of the third part is subjected to local sampling of which the missing channels are 10 channels, 20 channels and 30 channels respectively, random noise with unequal intensity is added, the maximum standard deviation coefficient of the noise is 0.3, and training seismic data Sei _data_init is obtained, wherein the key codes are as follows:
noise = np.random.normal(0, noise_rank*batch_x.max() , batch_x.shape)
out = noise + batch_x.numpy()
out = torch.from_numpy(out)
mask = irregular_mask(batch_x,self.rate)
batch_y = mask.mul(out)
The noise_rank is the seismic noise level, the irrgular_mask can generate mask files with different proportions and with random missing, and missing seismic data can be obtained after multiplication of the mask files with the seismic data. The partial and uniform deletions can be obtained by masking and multiplying by the same method. From this, a training seismic dataset Sei _train_data can be derived, along with a corresponding theoretical seismic dataset Sei _data.
2. Network input and label setting;
(1) The input of the generator network is training seismic data set Sei _train_data, the label is ideal seismic data set Sei _data corresponding to the training seismic data set, the output of the generator is Sei _outputs, and the content loss L obtained through calculation of the input and the label C The training seismic data set Sei _train_data and the output Sei _output of the generator are spliced according to the channel and then are input into the discriminator together, so that the resistance loss LcGAN_G of the generator is obtained; (2) The arbiter network (1) takes a training seismic data set Sei _train_data and a corresponding ideal seismic data set Sei _data as input after being spliced according to channel dimensions, and an error errD_real is obtained; (2) training seismic data set Sei _train_data and corresponding generator output Sei _output are spliced according to channels and then used as input, and the obtained error errD_fake, errD_fake and errD_real jointly form the antagonism loss L of the discriminator cGAN_D 。
3. Designing a seismic data recovery model based on a Swin transducer to generate an countermeasure network;
3.1 Generator M G Designing;
the generator network is combined with the switch transducer module and the U-Net network to strengthen the utilization of global relation in the seismic data, and mainly comprises a Patch Partition layer, a Patch Embedding layer, a switch transducer module, a Patch measurement layer, a Patch expansion layer and a Linear Projection layer, which relate to related standard functions in Pytorch, and the standard functions are described as follows:
standard functions defined in Pytorch are used in constructing the Patch Partition layer and the Patch Embedding layer:
conv2d (in_chams, emudbed_dim, kernel_size=patch_size, stride=patch_size), where in_chams represents the number of input channels, emudbed_dim represents the number of output channels, patch_size specifies the convolution kernel size, patch_size is the convolution step size.
To simplify the implementation, the Patch Messing layer and the Swin transform module are defined as one module, wherein standard functions defined in Pytorch are used when constructing the Patch Messing layer and the Swin transform module:
x0 = x[:, 0::2, 0::2, :]
x1 = x[:, 1::2, 0::2, :]
x2 = x[:, 0::2, 1::2, :]
x3 = x[:, 1::2, 1::2, :]
x = torch.cat([x0, x1, x2, x3], -1)
x = x.view(B, -1, 4 * C)
SwinTransformerBlock(dim=dim, input_resolution=input_resolution,
num_heads=num_heads, window_size=window_size,
shift_size=0 if (i % 2 == 0) else window_size // 2,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop, attn_drop=attn_drop,
drop_path=drop_path[i] if isinstance(drop_path, list)
else drop_path,
act_layer=nn.GELU, norm_layer=nn.LayerNorm)
wherein dim is a channel for inputting seismic data, input_resolution is the size of the input seismic data, num_heads is the number of heads of attention, window_size is the size of a window, mlp _ratio is the ratio of the hidden layer dimension to the embedded dimension, qkv _bias is whether a leachable bias value is added, and drop is the ratio of drop. act_layer is an activation function, a GELU activation function is adopted, and norm_layer adopts a layer normalization function LayerNorm (num_features);
Standard functions defined in Pytorch were used in constructing the Patch expansion layer and Linear Projection layer:
nn.Sequential(nn.Conv2d(in_channels, 2*in_channels, 1, 1, 0, bias=False),
nn.PReLU(),
nn.PixelShuffle(scale_factor),
nn.Conv2d(in_channels//2, in_channels//2, 1, stride=1, padding=0,
bias=False))
3.2 discriminator M D Designing;
the same generator code is designed, the drop_path parameter in the generator is modified to obtain the wanted depth [2,2,6,2] and the MLP module is added at last.
nn.Linear(in_features, hidden_features)
nn.GELU()
nn.Linear(hidden_features, out_features)
4. Setting a loss function;
the loss function of the overall network is:
the loss function is mainly divided into two parts: (1) generating a resistance loss against a network(2) content loss of network-generated and ideal seismic data +.>。/>Is->Weight coefficient of (2), remain->And->On the same order of magnitude.
Loss of resistanceThe specific form of (2) is as follows:
wherein,for training the seismic dataset Sei _train_data +.>Sei _data, < ++for ideal seismic dataset>Output Sei _outputs of the generator, +.>Data pairs input for the discriminator pairs +.>Is used for judging the result of the judgment,data pairs input for the discriminator pairs +.>Is the discrimination result of->For the desired value->Is a probability distribution of real seismic data. According to the characteristics of simultaneous reconstruction and denoising of the seismic data, noise interference is not required to be considered when the seismic channel is required to be reconstructed, and only the reconstruction effect is required to be evaluated; the missing channels are not required to be considered in the seismic data needing to be denoised, and only the denoising effect needs to be evaluated. Content loss function based on the above-mentioned characteristics >The form is as follows:
wherein,to evaluate the loss of denoising effect of non-missing parts, < >>To evaluate loss of effect of the partial reconstruction of the deletion +.>The parameters are used for keeping the same magnitude of the two parameters, and the specific forms are as follows: />
Wherein,x i in order to train the samples in the seismic dataset,y i for a sample in an ideal set of seismic data,i=1,2,3,…,N，G（x i ) Is thatx i The output of the sample down-generator,the operation rule is multiplication of corresponding positions of a matrix, the matrix is a mask matrix of missing seismic data, the shape of the matrix is identical to that of a Sei _data slice of the seismic data, the number of missing channels is 0, and the rest is 1./>Is->And taking the inverse. The key codes are as follows:
criterionadv = nn.BCELoss()
real=netD(target)
errD_real=criterionadv(real,target)
errD_real.backward(one)
fake = netD(inputv)
errD_fake=criterionadv(real,target)
errD_fake.backward(mone)
errD = errD_real - errD_fake
x_hat = netG(input)
err_MSER = criterionMSE((1-mask)*x_hat,(1-mask)*target)
err_MSED = criterionMSE(mask*x_hat,mask*target)
LossG=err_MSER+μerr_MSED+λerrG(x_hat,target)
wherein err_MSED isI.e. evaluate the loss of denoising effect of the non-missing part, err_MSER is +.>I.e. evaluating the loss of effect of the missing part reconstruction. errD is the arbiter penalty and LossG is the generator penalty.
5. Training a network and performing model test;
dividing a training set and a test set according to the ratio of 8:2, testing the model recovery effect by using the test set after training is completed on the training set, calculating SNR, wherein an SNR calculation formula is shown as follows, and recording an experimental result.
。
The implementation effect is as follows:
FIG. 7 is a graph of the comparative effects of recovery of random sampled BP2004 synthetic seismic data by various methods, wherein FIG. 7 (a) is an enlarged view of a portion of the area of the ideal seismic data; FIG. 7 (b) is an enlarged view of a portion of the area where noise-containing seismic data is missing; FIG. 7 (c) is an enlarged view of a portion of the recovered seismic data area of the seismic data recovery model, with an SNR of 20.3dB; FIG. 7 (d) is an enlarged view of a portion of the SUNet recovered seismic data with an SNR of 15.2dB; fig. 7 (e) is an enlarged view of a portion of the WGAN recovered seismic data area with an SNR of 16.2dB. Comparing the seismic data recovery model constructed by the implementation with the SUNet model recovery effect can obtain the generation of the countermeasure network, which is beneficial to improving the learning ability of the model to the seismic data, so that the detail part of the recovered seismic data is more approximate to the ideal seismic data. By comparing the recovery effect of the seismic data recovery model constructed by the embodiment with that of the WGAN model, the Swin transducer module can learn the relevance of the seismic data better than the convolution module, and the seismic data and noise can be distinguished better.
FIG. 8 is a graph of how a randomly sampled Marmousi synthetic seismic data is recovered versus a different method, where FIG. 8 (a) is an ideal seismic data; FIG. 8 (b) is a plot of noise-containing missing seismic data with a noise intensity of 0.1 and a sampling rate of 0.5; FIG. 8 (c) is SUNet recovered seismic data with an SNR of 7.5dB; FIG. 8 (d) is WGAN recovered seismic data with an SNR of 15.3dB; fig. 8 (e) shows the seismic data recovered by the seismic data recovery model, with an SNR of 17.6dB. The recovery graphs of different models are analyzed, the denoising effect of the seismic data recovery model constructed by the embodiment is obviously better than that of other two models in the part of 80-120 channels, the recovered phase axes are smooth and continuous in the part of 60-80 channels, and the SNR value is obviously higher than that of other two models, so that the seismic data recovery effect of the seismic data recovery model constructed by the embodiment is better than that of other two models.
FIG. 9 is a graph comparing the recovery effect of different methods on locally sampled Marmousi synthetic seismic data, wherein FIG. 9 (a) is ideal seismic data; FIG. 9 (b) is seismic data with a noise intensity of 0.1 and a partial loss of 20 traces, with an SNR of-1.0 dB; FIG. 9 (c) is SUNet recovered seismic data with an SNR of 13.6dB; FIG. 9 (d) is the seismic data recovered by the WGAN with an SNR of 12.8dB; fig. 9 (e) shows the seismic data recovered by the seismic data recovery model, with an SNR of 20.9dB. The recovery effect diagram of the seismic data recovery model constructed by observing SUNet, WGAN and the embodiment can be obtained, and compared with other two models, the seismic data recovery model constructed by the embodiment has smoother and more continuous phase axes of the seismic data, and is closer to ideal seismic data.
FIG. 10 is a graph of the recovery effect of the present embodiment on randomly sampled real seismic data, wherein FIG. 10 (a) is ideal seismic data; FIG. 10 (b) is a plot of real seismic data with an additive noise intensity of 0.1 and a sampling rate of 0.5; FIG. 10 (c) is a graph of recovery effects of a seismic data recovery model with an SNR of 16.8dB; FIG. 10 (d) is a residual profile waveform diagram of the recovery effect diagram of the seismic data recovery model and the real data; fig. 10 (e) is a spectrum diagram of the recovery effect diagram of the seismic data recovery model corresponding to the residual profile of the real data. And observing the recovery effect of the seismic data recovery model constructed by the embodiment, and recovering the seismic data clearly, enabling the same phase axis to be smooth and continuous, and enabling a small part of detail textures to be lost. The residual profile waveform diagram and the residual profile spectrogram are observed, the residual information intensity is still lower, the loss of effective information of the recovered seismic data is less, the recovery quality is higher, and the experiment of the method provided by the embodiment on the real seismic data set has better effect according to the past.
The above embodiments are merely illustrative of the preferred embodiments of the present invention, and the scope of the present invention is not limited thereto, but various modifications and improvements made by those skilled in the art to which the present invention pertains are made without departing from the spirit of the present invention, and all modifications and improvements fall within the scope of the present invention as defined in the appended claims.

Claims (7)

1. An integrated method for reconstructing and denoising seismic data, which is characterized by comprising the following steps:
acquiring seismic data;
inputting the seismic data into a preset seismic data recovery model, and outputting reconstructed and denoised seismic data, wherein the data recovery model is obtained based on training set training, the training set comprises ideal seismic data and training seismic data, and the data recovery model is constructed by adopting a Swin transducer to generate an countermeasure network;
the training seismic data is obtained by sampling the ideal seismic data, the sampling the ideal seismic data comprising:
randomly dividing the ideal seismic data into three parts, uniformly sampling a first part, setting a first noise standard deviation coefficient, and adding random noise which is not more than the first noise standard deviation coefficient and has different intensities;
randomly sampling the second part, setting a second noise standard deviation coefficient, and adding random noise which is not more than the second noise standard deviation coefficient and has different intensities;
and carrying out local sampling on the third part, setting a third noise standard deviation coefficient, and adding random noise which is not more than the third noise standard deviation coefficient and has different intensities.
2. The integrated seismic data reconstruction denoising method according to claim 1, further comprising preprocessing data in the training set before training the data recovery model based on the training set, the preprocessing process comprising:
and segmenting the training seismic data and the ideal seismic data in the training set into two-dimensional seismic data with the size of NT, and acquiring a training seismic data set and a theoretical seismic data set.
3. The integrated seismic data reconstruction denoising method according to claim 1, wherein the seismic data recovery model comprises: the generator is used for processing the training set to acquire the seismic data and the content loss generated by the generator, and comprises a Patch Partition layer, an encoder, a bottleneck layer, a decoder and a Linear Projection layer which are connected in sequence;
the discriminator is used for judging the seismic data generated by the training set and the generator, acquiring errors, acquiring countermeasures by combining the content measures, optimizing the countermeasures counter propagation guide generator, and comprises a Patch Partition layer, a Patch Embedding layer, a Patch measuring layer and an MLP layer which are sequentially connected, and a first switch transform module which is connected after the Patch Embedding layer and after the Patch measuring layer.
4. The integrated seismic data reconstruction and denoising method according to claim 3, wherein the encoder comprises a Patch Embedding layer, a second switch transformation module, a Patch Merging layer, the second switch transformation module being placed between the Patch Embedding layer and the Patch Merging layer, and between the Patch Merging layer and the Patch Merging layer;
the bottleneck layer comprises a third Swin Transformer module;
the decoder comprises a Patch expansion layer and a fourth Swin transform module, wherein the fourth Swin transform module is placed between the Patch expansion layer and the Patch expansion layer;
the second Swin transducer module and the fourth Swin transducer module are connected by a residual.
5. The integrated seismic data reconstruction denoising method of claim 3, wherein the content loss is:
wherein (1)>To evaluate the loss of denoising effect of non-missing parts, < >>To evaluate loss of effect of the partial reconstruction of the deletion +.>Is used for maintaining parameters on the same magnitude of the two;
wherein,x i in order to train the samples in the seismic dataset,y i for a sample in an ideal set of seismic data,i=1,2,3,…,N，G(x i ) Is thatx i Output of sample down generator, +. >For Hadamard product, the operation rule is multiplication of matrix corresponding positions, ++>Mask matrix for missing seismic data, +.>Is thatAnd taking the inverse.
6. The integrated seismic data reconstruction denoising method according to claim 3, wherein the countermeasures against loss are:
wherein (1)>For training the seismic dataset, < >>For an ideal seismic dataset, +.>Output of generator, ++>Data pairs input for the discriminator pairs +.>Is the discrimination result of->Data pairs input for the discriminator pairs +.>Is the discrimination result of->For the desired value->Is a probability distribution of real seismic data.
7. The integrated method for reconstructing and denoising seismic data according to claim 4, wherein the first Swin Transformer module, the second Swin Transformer module, the third Swin Transformer module, and the fourth Swin Transformer module are all used for performing self-attention mechanism calculation, and the method for performing self-attention mechanism calculation is as follows:
wherein (1)>And->For seismic data in different windows, +.>Is->And->Degree of association between->Trainable variable greater than 0.01, < >>Encoding the relative positions of the different positions.
CN202311531698.5A 2023-11-17 2023-11-17 Seismic data reconstruction denoising integrated method Active CN117250657B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311531698.5A CN117250657B (en) 2023-11-17 2023-11-17 Seismic data reconstruction denoising integrated method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311531698.5A CN117250657B (en) 2023-11-17 2023-11-17 Seismic data reconstruction denoising integrated method

Publications (2)

Publication Number Publication Date
CN117250657A true CN117250657A (en) 2023-12-19
CN117250657B CN117250657B (en) 2024-03-08

Family

ID=89128011

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311531698.5A Active CN117250657B (en) 2023-11-17 2023-11-17 Seismic data reconstruction denoising integrated method

Country Status (1)

Country Link
CN (1) CN117250657B (en)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115330643A (en) * 2022-10-13 2022-11-11 中国石油大学(华东) Earthquake denoising method based on convolution neural network and visual transformation neural network
CN115905805A (en) * 2022-11-02 2023-04-04 吉林大学 DAS data multi-scale noise reduction method based on global information judgment GAN
CN116009065A (en) * 2021-10-21 2023-04-25 中国石油化工股份有限公司 Seismic data noise removing method, electronic equipment and medium
CN116416156A (en) * 2023-03-10 2023-07-11 中加健康工程研究院(合肥)有限公司 Swin transducer-based medical image denoising method
CN116468083A (en) * 2023-04-27 2023-07-21 青岛理工大学 Transformer-based network generation countermeasure method
US20230252993A1 (en) * 2022-02-04 2023-08-10 Adobe Inc. Visual speech recognition for digital videos utilizing generative adversarial learning

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116009065A (en) * 2021-10-21 2023-04-25 中国石油化工股份有限公司 Seismic data noise removing method, electronic equipment and medium
US20230252993A1 (en) * 2022-02-04 2023-08-10 Adobe Inc. Visual speech recognition for digital videos utilizing generative adversarial learning
CN115330643A (en) * 2022-10-13 2022-11-11 中国石油大学(华东) Earthquake denoising method based on convolution neural network and visual transformation neural network
CN115905805A (en) * 2022-11-02 2023-04-04 吉林大学 DAS data multi-scale noise reduction method based on global information judgment GAN
CN116416156A (en) * 2023-03-10 2023-07-11 中加健康工程研究院(合肥)有限公司 Swin transducer-based medical image denoising method
CN116468083A (en) * 2023-04-27 2023-07-21 青岛理工大学 Transformer-based network generation countermeasure method

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
QIAN ZHANG ET AL.: "A Single-Stage Unsupervised Denoising Low-Illumination Enhancement Network Based on Swin-Transformer", 《IEEE ACCESS》, vol. 11, pages 75696 - 75706 *
SHUANG YAN ET AL.: "A Method for Denoising Seismic Signals With a CNN Based on an Attention Mechanism", 《IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING》, vol. 60, pages 1 - 15, XP011925711, DOI: 10.1109/TGRS.2022.3215819 *
XIN ZHOU ET AL.: "SID-TGAN: A Transformer-Based Generative Adversarial Network for Sonar Image Despeckling", 《REMOTE SENSING》, vol. 15, pages 1 - 19 *

Also Published As

Publication number Publication date
CN117250657B (en) 2024-03-08

Similar Documents

Publication Publication Date Title
CN110361778B (en) Seismic data reconstruction method based on generation countermeasure network
CN109086773B (en) Fault plane identification method based on full convolution neural network
CN107563433B (en) Infrared small target detection method based on convolutional neural network
CN113113030B (en) High-dimensional damaged data wireless transmission method based on noise reduction self-encoder
CN110490219B (en) Method for reconstructing seismic data by U-net network based on texture constraint
CN111931820B (en) Method for extracting line spectrum of target radiation noise LOFAR spectrogram in water based on convolution residual error network
CN114549925A (en) Sea wave effective wave height time sequence prediction method based on deep learning
CN112257741B (en) Method for detecting generative anti-false picture based on complex neural network
CN114966860A (en) Seismic data denoising method based on convolutional neural network
Gao et al. Deep learning vertical resolution enhancement considering features of seismic data
CN116626753B (en) Microseism event identification method and system based on multi-modal neural network
CN117250657B (en) Seismic data reconstruction denoising integrated method
CN117171514A (en) Seismic data denoising method based on multi-scale residual convolution
CN113570041B (en) Neural network and method for pressing seismic data noise of marine optical fiber towing cable by using same
CN116434074A (en) Target identification method based on adjacent branch complementation significance and multiple priori sparse representation
CN115639605A (en) Automatic high-resolution fault identification method and device based on deep learning
CN115147727A (en) Method and system for extracting impervious surface of remote sensing image
CN117975174B (en) Three-dimensional digital core reconstruction method based on improvement VQGAN
CN116405368B (en) Network fault diagnosis method and system under high-dimensional unbalanced data condition
CN115169407B (en) Weak transient signal detection method and system based on time domain depth feature learning
CN118051793A (en) Formation data real-time analysis and prediction method and system based on TRD construction method
CN117409316A (en) TransUNet-based seismic data karst characteristic intelligent identification positioning method
CN116859449A (en) Method for reconstructing seismic data by using convolutional self-encoder depth neural network
CN117150222A (en) Intelligent restoration method for underground vibration field measurement data based on diffusion model
CN116755145A (en) Three-dimensional seismic horizon intelligent tracking method based on multi-seismic attribute regression network

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
GR01 Patent grant
GR01 Patent grant