CN113484908B - Missing seismic data reconstruction method for deep learning network by combining partial convolution and attention mechanism - Google Patents
Missing seismic data reconstruction method for deep learning network by combining partial convolution and attention mechanism Download PDFInfo
- Publication number
- CN113484908B CN113484908B CN202110980357.0A CN202110980357A CN113484908B CN 113484908 B CN113484908 B CN 113484908B CN 202110980357 A CN202110980357 A CN 202110980357A CN 113484908 B CN113484908 B CN 113484908B
- Authority
- CN
- China
- Prior art keywords
- data
- network
- missing
- convolution
- seismic data
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 23
- 238000013135 deep learning Methods 0.000 title claims abstract description 12
- 230000007246 mechanism Effects 0.000 title claims abstract description 8
- 238000012545 processing Methods 0.000 claims description 20
- 230000006870 function Effects 0.000 claims description 10
- 238000012549 training Methods 0.000 claims description 10
- 238000005070 sampling Methods 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 5
- 230000008569 process Effects 0.000 claims description 5
- 239000011159 matrix material Substances 0.000 claims description 3
- 230000004927 fusion Effects 0.000 claims 1
- 230000001788 irregular Effects 0.000 abstract description 3
- 238000003672 processing method Methods 0.000 abstract description 2
- 230000000694 effects Effects 0.000 description 4
- 238000009499 grossing Methods 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000013136 deep learning model Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 239000003208 petroleum Substances 0.000 description 1
- 238000012876 topography Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
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
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 provides a data processing method for realizing the reconstruction of missing seismic data by fusing partial convolution and an attention mechanism into a deep learning network. The method can be used for adapting to irregular and large-space data missing situations in the seismic acquisition data, and realizing more accurate and reliable seismic data reconstruction.
Description
Technical Field
The invention belongs to the field of petroleum seismic exploration data processing, and relates to a seismic data processing method for reconstructing and recovering data when seismic data are missing or damaged.
Background
In the field acquisition of seismic exploration data, the acquired seismic data is often irregular, undersampled or has data missing due to the limitations of surface topography, geological conditions, data acquisition equipment, environmental factors and the like. In the process of seismic data processing, the deletion of seismic data can be caused by removing bad tracks, cutting off specific data segments or filtering and the like. The integrity of the original seismic data is a key foundation for subsequent seismic data processing and interpretation and reliable subsurface geologic structure and reservoir imaging, so that the reconstruction of the missing seismic data becomes one of key technologies for seismic data processing. The existing seismic data reconstruction method mainly comprises an artificial intelligence method based on wave equation, prediction filtering, transform domain reconstruction, compressed sensing theory and machine learning and deep learning.
In recent years, a method for introducing artificial intelligence into seismic data processing has become a trend, and a deep learning model represented by a GAN countermeasure generation network and a DNCNN deep convolution network has good effects on seismic data reconstruction processing. However, the output result of the ordinary GAN deep learning model often has problems such as deformation of the seismic reflection structure, blurring of effective signals, aliasing, and the like. In particular, in the case of seismic data with large-distance data channel missing, the seismic data channel reconstructed by the common deep learning method is often uncoordinated with the characteristics of the surrounding existing seismic data channels. An important reason for this is that the size of the convolution kernel is limited in such data reconstruction algorithms, so that the convolution result is mainly affected by the data in the convolution kernel, and the effect and influence of effective information at a far position outside the convolution kernel cannot be obtained.
The attention mechanism in the deep learning is used as an additional module in the deep learning network, and can be used for solving the problem of smaller network perception range caused by the limitation of the size of a convolution kernel.
Disclosure of Invention
The invention aims to overcome the defects of the existing seismic data reconstruction method and technology, and develops a missing seismic data reconstruction method of a deep learning network by fusing partial convolution and a attention mechanism, wherein the method comprises the following main steps:
the method comprises the steps of establishing seismic data slices with uniform sizes as a training set;
and secondly, generating an countermeasure network by utilizing the data input of the training set, so that masks M with the same number and size as the seismic data slices in the training set are randomly generated by the countermeasure network. The mask M randomly divides the rectangular missing regions marked with 0 and the known regions marked with 1 of different sizes, respectively. When the mask is multiplied by the number of the data slices, a missing region can be formed on the original data slice;
firstly, the input data I passes through a part of convolution layers in the generation network G, the part of convolution layers judges whether effective data exist or not according to the position of a convolution kernel, convolution operation is selectively executed, and then the input data I' is obtained after being processed by a downsampling layer in the generation network G and is used for further processing by an attention module;
the attention module in the generating network G calculates cosine distances between each point in the missing area and all points in the known area, and the calculation method is as follows:
wherein,,
representing cosine similarity at two points (x, y) and (x ', y') obtained through standardized inner product operation, mapping a calculation result into probability models through a softmax function, wherein the number of the probability models is the same as the number of points in a missing area;
fifthly, multiplying the known region data with the obtained different probability model numbers respectively, wherein the probability model numbers are as follows:
a new value of a point in the missing region corresponding to the probability model can be obtained, after the processing, the value in the missing region is fused with the data characteristics in the known region, and the obtained new data matrix is recorded as
Sixth step of comparing I' with
After being spliced together in the channel dimension, the channel dimension is subjected to pixel convolution operation and is conveyed to an up-sampling layer in a generating network G for subsequent processing;
and after the up-sampling layer in the generation network G processes the data, outputting a plurality of results, calculating the average value of the results, then transmitting the results to the discrimination network D, comparing the reconstructed result with the samples of the training set by the discrimination network D, and feeding back the comparison result to the generation network G, wherein the two networks are mutually influenced, so that the network performance is jointly improved, the constructed loss function is ensured to obtain the minimum value, and the reconstructed seismic data is output.
Drawings
FIG. 1 is a schematic diagram of a deep learning network model incorporating a partial convolution layer and an attention module in accordance with the present invention. The input data sequentially passes through a partial convolution layer in the generation network G, a downsampling layer, a attention module and an upsampling layer and then is input into the discrimination network D.
FIG. 2 is a plot of seismic profiles before and after processing, after random loss of trace data, for a fully acquired single shot trace set in accordance with an embodiment of the present invention. Wherein: FIG. 2 (a) is a cross section of a completely acquired, undelayed seismic trace data, with trace number on the abscissa and time on the ordinate, in seconds(s); FIG. 2 (b) is a cross section of FIG. 2 (a) with trace numbers randomly missing, time on the abscissa, and seconds(s) on the ordinate; FIG. 2 (c) is a cross section of the seismic data reconstruction method of the invention, obtained by processing FIG. 2 (b), with the abscissa being the trace number and the ordinate being time in seconds(s); fig. 2 (d) is a residual profile obtained by subtracting the data of fig. 2 (a) and 2 (c), the abscissa is the trace number, the ordinate is time, and the unit is seconds(s).
Detailed Description
The invention discloses a structure and a flow of a deep learning network for realizing the reconstruction of missing seismic data, which are shown in a figure 1, and the structure and the flow of the deep learning network are integrated with a part of convolution layers and a processing module of an attention mechanism, and the concrete implementation flow of the reconstruction processing of the missing seismic data is as follows:
firstly, using seismic data slices with uniform sizes as a training set;
and when a batch of training data is input to generate an countermeasure network, the network can randomly generate masks M with the same number and size as the data slices. Rectangular missing areas and known areas with different sizes are randomly divided on the mask, data in the known areas in the mask are marked as 1, and the missing areas are marked as 0. When the mask is multiplied by the number of the data slices, a missing region can be manufactured on the original data slice;
and thirdly, the input data I is processed by a partial convolution layer and a downsampling layer to obtain I'. I' is not directly input into the upsampling layer, but enters the attention module for further processing; wherein a portion of the convolution layers selectively convolve based on whether valid data is present in the location of the convolution kernel.
The attention module calculates cosine distances between each point in the missing region and all points in the known region by the following calculation method:
wherein the method comprises the steps of
Representing cosine similarity at two points (x, y), (x ', y') obtained by a normalized inner product operation.
Mapping the calculation result into probability models through a softmax function, and marking the probability models as score, wherein the number of the probability models is the same as the number of points in the missing area; however, with iterations of the neural network, score is not the only fixed. So define the final score after i iterations as score i ，score i And score i-1 The relationship of (2) is as follows:
score i ＝λsocre′ i +(1-λ)score i-1
in socre' i Representing the fraction generated by the softmax function during the ith iteration. Lambda is a parameter that can be learned during the iteration process.
And fifthly, multiplying the known region data with the obtained different probability model numbers respectively, wherein the formula is as follows:
obtaining the missing region corresponding to the probability modelNew values for points in the inner. After such processing, the values in the missing region fuse the data features in the known region, and the new data matrix is recorded as
Sixth step of comparing I' with
After being spliced together in the channel dimension, the pixel convolution is carried out and the pixel convolution is transmitted to the next up-sampling layer;
the data of the user is output a plurality of results through the up-sampling layer, and the average value of the results is taken and then is transmitted to a discrimination network; the loss function sampled by the model is designed as:
wherein the loss function in the generated network G is:
the loss function in the discrimination network D is:
where M represents the mask and ". Sur represents the dot product operation. Loss of deleted region L hole Representing the difference between the reconstructed result and the target sample. Known area loss is L valid The loss function may control the transformation of the known region before and after processing. Input sample is I input The network output result is I out 。L tv For the smoothing loss, P represents the hole area after expansion by 1 pixel. The smoothing loss represents the L1 loss for one pixel in the hole area and the right and lower pixels of that pixel. The smoothing loss measures the difference between adjacent data in the horizontal and vertical directions. L (L) perceptual For perceived loss, it is a pretrained VGG-16 network.
Characterization of the ith pooled layer map in VGG-16 network, H i W i C i The height, width and channel size of the ith feature map are represented. The perceived loss may be a deeper feature than the network output data and the real data. The output of the discrimination network D is 0 or 1. The output of the discrimination network D is 0 or 1.
When the complete single shot seismic trace data in fig. 2 (a) has serious data missing, as shown in the cross section in fig. 2 (b), the data missing of large space and small space simultaneously occurs, wherein the missing space of the largest missing region reaches 50 channels, the missing space of the smallest missing region is 6 channels, 150 channels of seismic data are missing in the cross section in fig. 2 (b), and the data missing rate is 58%. After the processing of the method of the invention is adopted, the data reconstruction section as shown in the figure 2 (c) is obtained, the original area with missing data is clear and continuous seismic reflection in-phase axis, compared with the figure 2 (a), the seismic reflection characteristics of the reconstructed seismic section are the same as the complete single shot seismic trace set, in order to illustrate the effect of data reconstruction, the section data in the figure 2 (a) and the figure 2 (c) are subtracted, the residual seismic sections of the two are obtained, as shown in the figure 2 (d), the effective seismic reflection signals with almost no obvious energy are visible from the residual seismic sections, and the effect of the invention on the reconstruction of the missing seismic data is quite reliable and superior.
The invention has the advantages that: (1) For a foreground data feature, the attention mechanism first calculates its similarity to all background data features, and then converts the similarity to an attention score by softmax. Finally, calculating an average data characteristic by using the background data characteristic and the attention score, wherein the average data characteristic contains useful information needed to be used according to the attention score; (2) After fusing the result with the original data characteristic pixels, carrying out convolution, so that the limitation of the convolution kernel size can be broken through in the convolution process, and the data outside the convolution kernel can also act on the convolution result, thereby reducing the defects of deformation, blurring, aliasing and the like of the stratum reflection structure in the reconstruction result; (3) The method can be suitable for effectively reconstructing irregular and random missing data in the seismic acquisition data.
The foregoing embodiments are merely illustrative of the present invention, and various implementation steps of the method may be changed, and all equivalent changes and modifications performed on the basis of the technical solutions of the present invention should not be excluded from the protection scope of the present invention.
Claims (1)
1. A partial convolution and attention mechanism fusion deep learning network missing seismic data reconstruction method comprises the following main steps:
(1) Establishing seismic data slices with uniform sizes as a training set;
(2) Generating an countermeasure network by using training set data input, wherein the countermeasure network consists of a generation network G and a discrimination network D; enabling the countermeasure network to randomly generate masks M with the same number and same size as the seismic data slices in the training set; the mask M randomly divides the rectangular missing area marked with 0 and the known area marked with 1 with different sizes respectively; when the mask is multiplied by the number of the data slices, a missing region can be formed on the original data slice;
(3) The input data I firstly passes through a part of convolution layers in the generation network G, the part of convolution layers judges whether effective data exist or not according to the position of a convolution kernel, then selectively carries out convolution operation, and then obtains I' after being processed by a downsampling layer in the generation network G for further processing by an attention module;
(4) The attention module in the generation network G calculates cosine distances between each point in the missing area and all points in the known area by the following method:
wherein,,
representing cosine similarity at two points (x, y), (x ', y') obtained by normalized inner product operation, will be calculatedThe calculation result is mapped into probability models through a softmax function, and the number of the probability models is the same as the number of points in the missing area;
(5) Multiplying the known region data with the obtained different probability model numbers respectively, wherein the following formula is as follows:
a new value of a point in the missing region corresponding to the probability model can be obtained, after the processing, the value in the missing region is fused with the data characteristics in the known region, and the obtained new data matrix is recorded as
(6) And I' is combined with
After being spliced together in the channel dimension, the channel dimension is subjected to pixel convolution operation and is conveyed to an up-sampling layer in a generating network G for subsequent processing;
(7) After the up-sampling layer in the network G is generated to process data, a plurality of results are output, the average value of the results is calculated and then is transmitted to the discrimination network D, the discrimination network D compares the reconstructed result with samples of the training set, the comparison result is fed back to the generation network G, the two networks are mutually influenced, the network performance is jointly improved, the constructed loss function is ensured to obtain the minimum value, and therefore the reconstructed seismic data is output.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110980357.0A CN113484908B (en) | 2021-08-25 | 2021-08-25 | Missing seismic data reconstruction method for deep learning network by combining partial convolution and attention mechanism |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110980357.0A CN113484908B (en) | 2021-08-25 | 2021-08-25 | Missing seismic data reconstruction method for deep learning network by combining partial convolution and attention mechanism |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113484908A CN113484908A (en) | 2021-10-08 |
| CN113484908B true CN113484908B (en) | 2023-07-14 |
Family
ID=77946207
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110980357.0A Active CN113484908B (en) | 2021-08-25 | 2021-08-25 | Missing seismic data reconstruction method for deep learning network by combining partial convolution and attention mechanism |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113484908B (en) |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103048690A (en) * | 2012-12-11 | 2013-04-17 | 成都理工大学 | Stratum reflection pickup technology of rapid matched projection decomposition extracted based on optimal seismic wavelet |
| CA2838711A1 (en) * | 2013-01-11 | 2014-07-11 | Cgg Services Sa | A system and method for the removal of shallow water multiples using a hybrid multi-channel prediction method |
| CA2838857A1 (en) * | 2013-01-15 | 2014-07-15 | Cgg Services Sa | Seismic data processing including true-azimuth three-dimensional internal multiple attentuation without subsurface information |
| CA2893718A1 (en) * | 2014-06-06 | 2015-12-06 | Schlumberger Canada Limited | Generalized spectral decomposition |
| WO2017155634A1 (en) * | 2016-03-11 | 2017-09-14 | Origin Wireless, Inc. | Methods, devices, servers, apparatus, and systems for wireless internet of things applications |
| GB2560991A (en) * | 2017-03-31 | 2018-10-03 | Seismic Apparition Gmbh | Method for seismic acquisition and processing |
| CN110197154A (en) * | 2019-05-30 | 2019-09-03 | 汇纳科技股份有限公司 | Pedestrian recognition methods, system, medium and the terminal again of fusion site texture three-dimensional mapping |
| CN110361778A (en) * | 2019-07-04 | 2019-10-22 | 华北电力大学 | A kind of Reconstruction of seismic data method based on generation confrontation network |
| CN111476717A (en) * | 2020-04-07 | 2020-07-31 | 西安电子科技大学 | Face image super-resolution reconstruction method based on self-attention generation countermeasure network |
| WO2020181685A1 (en) * | 2019-03-12 | 2020-09-17 | 南京邮电大学 | Vehicle-mounted video target detection method based on deep learning |
| CN113156513A (en) * | 2021-04-14 | 2021-07-23 | 吉林大学 | Convolutional neural network seismic signal denoising method based on attention guidance |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11947061B2 (en) * | 2019-10-18 | 2024-04-02 | Korea University Research And Business Foundation | Earthquake event classification method using attention-based convolutional neural network, recording medium and device for performing the method |
-
2021
- 2021-08-25 CN CN202110980357.0A patent/CN113484908B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103048690A (en) * | 2012-12-11 | 2013-04-17 | 成都理工大学 | Stratum reflection pickup technology of rapid matched projection decomposition extracted based on optimal seismic wavelet |
| CA2838711A1 (en) * | 2013-01-11 | 2014-07-11 | Cgg Services Sa | A system and method for the removal of shallow water multiples using a hybrid multi-channel prediction method |
| CA2838857A1 (en) * | 2013-01-15 | 2014-07-15 | Cgg Services Sa | Seismic data processing including true-azimuth three-dimensional internal multiple attentuation without subsurface information |
| CA2893718A1 (en) * | 2014-06-06 | 2015-12-06 | Schlumberger Canada Limited | Generalized spectral decomposition |
| WO2017155634A1 (en) * | 2016-03-11 | 2017-09-14 | Origin Wireless, Inc. | Methods, devices, servers, apparatus, and systems for wireless internet of things applications |
| GB2560991A (en) * | 2017-03-31 | 2018-10-03 | Seismic Apparition Gmbh | Method for seismic acquisition and processing |
| WO2020181685A1 (en) * | 2019-03-12 | 2020-09-17 | 南京邮电大学 | Vehicle-mounted video target detection method based on deep learning |
| CN110197154A (en) * | 2019-05-30 | 2019-09-03 | 汇纳科技股份有限公司 | Pedestrian recognition methods, system, medium and the terminal again of fusion site texture three-dimensional mapping |
| CN110361778A (en) * | 2019-07-04 | 2019-10-22 | 华北电力大学 | A kind of Reconstruction of seismic data method based on generation confrontation network |
| CN111476717A (en) * | 2020-04-07 | 2020-07-31 | 西安电子科技大学 | Face image super-resolution reconstruction method based on self-attention generation countermeasure network |
| CN113156513A (en) * | 2021-04-14 | 2021-07-23 | 吉林大学 | Convolutional neural network seismic signal denoising method based on attention guidance |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113484908A (en) | 2021-10-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111640159B (en) | Remote sensing image change detection method based on twin convolutional neural network | |
| CN106157232B (en) | A kind of general steganalysis method of digital picture characteristic perception | |
| CN111985381B (en) | Guidance area dense crowd counting method based on flexible convolution neural network | |
| CN113673590A (en) | Rain removing method, system and medium based on multi-scale hourglass dense connection network | |
| CN111179196B (en) | Multi-resolution depth network image highlight removing method based on divide-and-conquer | |
| CN113327231B (en) | Hyperspectral abnormal target detection method and system based on space-spectrum combination | |
| CN117611494B (en) | Panchromatic remote sensing image thin cloud removing method | |
| CN103871039A (en) | Generation method for difference chart in SAR (Synthetic Aperture Radar) image change detection | |
| CN112560719B (en) | High-resolution image water body extraction method based on multi-scale convolution-multi-core pooling | |
| CN117114984A (en) | Remote sensing image super-resolution reconstruction method based on generation countermeasure network | |
| CN111798359A (en) | Deep learning-based image watermark removing method | |
| CN108230365B (en) | SAR image change detection method based on multi-source difference image content fusion | |
| CN114460648A (en) | 3D convolutional neural network-based self-supervision 3D seismic data random noise suppression method | |
| Shit et al. | An encoder‐decoder based CNN architecture using end to end dehaze and detection network for proper image visualization and detection | |
| CN113484908B (en) | Missing seismic data reconstruction method for deep learning network by combining partial convolution and attention mechanism | |
| CN117746130A (en) | Weak supervision deep learning classification method based on remote sensing image punctiform semantic tags | |
| CN113409413A (en) | Time sequence image reconstruction method based on gated convolution-long and short memory network | |
| CN112734675A (en) | Image rain removing method based on pyramid model and non-local enhanced dense block | |
| Hepburn et al. | Enforcing perceptual consistency on generative adversarial networks by using the normalised laplacian pyramid distance | |
| CN115909084A (en) | Single remote sensing image cloud removing method based on edge information | |
| CN116468638A (en) | Face image restoration method and system based on generation and balance countermeasure identification | |
| CN115861108A (en) | Image restoration method based on wavelet self-attention generation countermeasure network | |
| CN112862729B (en) | Remote sensing image denoising method based on characteristic curve guidance | |
| CN114862803A (en) | Industrial image anomaly detection method based on fast Fourier convolution | |
| CN103002355A (en) | Production method and extracting method for video watermarking with robustness |
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 |