CN114509814A - Pre-stack seismic data random noise suppression method and system - Google Patents
Pre-stack seismic data random noise suppression method and system Download PDFInfo
- Publication number
- CN114509814A CN114509814A CN202011171331.3A CN202011171331A CN114509814A CN 114509814 A CN114509814 A CN 114509814A CN 202011171331 A CN202011171331 A CN 202011171331A CN 114509814 A CN114509814 A CN 114509814A
- Authority
- CN
- China
- Prior art keywords
- data
- network
- random noise
- seismic data
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 68
- 230000001629 suppression Effects 0.000 title claims abstract description 55
- 238000012549 training Methods 0.000 claims abstract description 80
- 238000012795 verification Methods 0.000 claims description 26
- 238000005070 sampling Methods 0.000 claims description 21
- 238000012545 processing Methods 0.000 claims description 16
- 238000004422 calculation algorithm Methods 0.000 claims description 13
- 230000000694 effects Effects 0.000 claims description 13
- 238000004088 simulation Methods 0.000 claims description 9
- 238000002360 preparation method Methods 0.000 claims description 8
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 6
- 238000013461 design Methods 0.000 claims description 6
- 238000012935 Averaging Methods 0.000 claims description 3
- 238000003860 storage Methods 0.000 claims description 3
- 238000005056 compaction Methods 0.000 claims description 2
- 230000008569 process Effects 0.000 abstract description 26
- 238000013135 deep learning Methods 0.000 abstract description 11
- 238000005516 engineering process Methods 0.000 abstract description 11
- 238000013528 artificial neural network Methods 0.000 abstract description 7
- 230000006870 function Effects 0.000 description 6
- 238000010586 diagram Methods 0.000 description 5
- 238000012360 testing method Methods 0.000 description 5
- 238000011161 development Methods 0.000 description 4
- 230000018109 developmental process Effects 0.000 description 4
- 238000009826 distribution Methods 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 230000009466 transformation Effects 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000005520 cutting process Methods 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 238000011176 pooling Methods 0.000 description 2
- 238000010200 validation analysis Methods 0.000 description 2
- 238000003491 array Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000013075 data extraction Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000003709 image segmentation Methods 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 239000003208 petroleum Substances 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000011218 segmentation 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
- G01V1/36—Effecting static or dynamic corrections on records, e.g. correcting spread; Correlating seismic signals; Eliminating effects of unwanted energy
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
- G01V2210/00—Details of seismic processing or analysis
- G01V2210/30—Noise handling
- G01V2210/32—Noise reduction
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)
- Image Analysis (AREA)
Abstract
The invention provides a method and a system for suppressing random noise of pre-stack seismic data, and belongs to the field of geophysical exploration technology and deep learning. The method for suppressing the random noise of the pre-stack seismic data obtains an intelligent random noise suppression network by utilizing actual seismic data, pre-stack shot set forward modeling data and a GPU parallel mode, and processes the actual pre-stack seismic data by utilizing the intelligent random noise suppression network to obtain the pre-stack seismic data after the random noise suppression. The method applies the U-NET network to the seismic data denoising, effectively avoids the high requirements of the conventional convolution network on the training data volume and the neural network scale, realizes the network training of the parallel multiple GPU data, improves the training efficiency, and realizes the effective suppression of the random noise of the pre-stack data.
Description
Technical Field
The invention belongs to the field of geophysical exploration technology and deep learning, and particularly relates to a method and a system for suppressing random noise of pre-stack seismic data, which are applied to random noise removal in denoising processing in petroleum geophysical exploration.
Background
The existence of seismic data noise seriously affects the imaging quality of seismic data, and how to effectively perform seismic data denoising is a key step for improving the signal-to-noise ratio of the seismic data in seismic data processing. The seismic data noise is mainly divided into regular noise and random noise, wherein the regular noise has deterministic characteristics and can be effectively suppressed according to a noise forming mechanism, the random noise only has statistical characteristics and does not have deterministic distribution form, and the noise suppression difficulty is relatively high. The conventional random noise suppression method mainly comprises a time-space domain denoising method and a transform domain denoising method, wherein the time-space domain denoising algorithm is directly processed based on time-space domain data, a representative method mainly comprises a denoising algorithm based on a non-local mean value of a redundant structure hypothesis, and the method can effectively suppress random noise without reducing effective signal resolution by mainly performing noise suppression on regional groups with similar local structures in seismic data, but has relatively high computational complexity and low processing efficiency. The denoising method based on the transform domain suppresses noise by performing data domain transformation on data to be processed to enhance the difference between the noise and an effective signal, and the current representative methods include f-k domain filtering, wavelet transformation denoising, curvelet transformation denoising and the like.
Generally speaking, various algorithms have been developed for suppressing random noise at present, and a wide range of practical applications have been obtained, but most of the existing methods rely on manual intervention and are influenced by the capability of processing personnel to a great extent, and a large amount of manual parameter adjustment may be required in the denoising process, so that the denoising efficiency is influenced to a great extent.
In recent years, with the rapid development of an automatic rapid image processing technology based on artificial intelligence, the appearance of a convolutional neural network algorithm promotes the application of deep learning in various fields, so that the seismic data processing based on artificial intelligence has a wide application prospect. Therefore, the seismic data noise removal technology based on the deep neural network is rapidly developed and applied and achieves certain effect. However, the existing intelligent denoising method is mainly based on the conventional convolutional network, and the development and practical popularization and application of the technology are greatly limited by the size of the available label data set and the size of the scale of the designable neural network. U-NET networks, originally used for medical cell segmentation in 2015, were an expanded and optimized full convolution network, consisting of a down-sampling path and a symmetric up-sampling path. A small number of data training models can be supported with higher resolution than other fully connected and conventional convolutional networks. The noise suppression precision can be effectively improved by developing the seismic data random noise suppression technology based on the U-net network.
Based on the problems, a seismic data random noise suppression technology based on a U-NET network is developed, and in addition, considering that the training of a model occupies the longest time in the intelligent processing process, the efficiency of network training is influenced because the conventional CPU-based training and single GPU training cannot fully utilize computing resources, so that the invention realizes the data parallel model training of multiple GPUs based on Tensiloflow, provides a high-precision and high-efficiency intelligent denoising process, realizes the intelligent processing of noise suppression and avoids the manual intervention of the denoising process. The formed technology can better serve seismic data processing, and lays a foundation for production, application and popularization of the seismic data random noise suppression technology, and corresponding research literature data are not found in the method.
Disclosure of Invention
The invention aims to solve the problems in the prior art, and provides a method and a system for suppressing random noise of pre-stack seismic data, so that multi-GPU parallel model training is realized, the efficiency of the model training is further enhanced, a set of intelligent denoising network with strong generalization capability and high denoising precision is established, manual intervention is reduced, the denoising efficiency is improved, data with high signal-to-noise ratio is provided for seismic data processing, a new technical thought is provided for the high-precision seismic data denoising technology in the future, and powerful technical support is provided for practical production and application.
The invention is realized by the following technical scheme:
the method comprises the steps of obtaining an intelligent random noise suppression network by utilizing actual seismic data, pre-stack shot set forward modeling data and a GPU parallel mode, and processing the actual pre-stack seismic data by utilizing the intelligent random noise suppression network to obtain pre-stack seismic data after random noise suppression.
A further development of the invention is that the method comprises:
(1) preparing a data set by using actual seismic data and forepart shot set forward modeling data, and dividing the data set into a training data set and a verification data set;
(2) designing a network, and performing parallel training on the network by using a GPU and a training data set to obtain a trained network;
(3) verifying the trained network by adopting a verification data set to obtain a verified intelligent random noise suppression network;
(4) and acquiring pre-stack seismic data, and inputting the pre-stack seismic data into the verified intelligent random noise suppression network to obtain pre-stack seismic data after random noise suppression.
The invention is further improved in that the actual seismic data in the step (1) is the actual seismic data with high-quality denoising effect selected from the denoising seismic data of a typical exploration area;
and (2) the forward modeling simulation data of the pre-stack shot set in the step (1) is simulation data obtained after forward modeling is carried out on the pre-stack shot set.
A further improvement of the present invention is that the actual seismic data is 40% and the simulated data is 60%.
In a further improvement of the present invention, the operation of preparing the data set by using the actual seismic data and the antecedent-stacked shot set forward modeling data in the step (1) comprises:
taking the analog data as tag data;
and carrying out random data matching on the noise data extracted from the actual seismic data and the simulation data, and then summing to obtain noise-containing data, or adding Gaussian random noise into the simulation data to obtain the noise-containing data.
Noisy data and tag data constitute the data set.
The operation of step (1) further comprises: writing the data set into the compaction file in the storage format of the MNIST data set.
The invention is further improved in that the network designed in the step (2) is a U-NET network;
and performing image up-sampling in an up-sampling part of the U-NET network by adopting a nearest neighbor interpolation algorithm.
In a further improvement of the present invention, the operation of training the network in parallel by using the GPU and the training data set in step (2) to obtain the trained network includes:
the CPU distributes the training data set to each GPU;
each GPU trains the U-NET network by using a training data set, obtains a gradient value according to a training result and then sends the gradient value to a CPU;
the CPU performs protocol averaging on the gradient values sent by all the GPUs to obtain a gradient mean value, and applies the gradient mean value to a loss function and an optimizer to obtain an updated model;
and obtaining the trained network after multiple updates.
A further refinement of the invention is that the loss function takes the L2 norm;
the optimizer adopts an Adam optimizer.
In a second aspect of the invention, there is provided a system for stochastic noise suppression of pre-stack seismic data, the system comprising:
the data set preparation unit is used for preparing a data set by utilizing actual seismic data and the ante-stack shot set forward modeling data and dividing the data set into a training data set and a verification data set;
the network design training unit is connected with the data set preparation unit and used for designing a network and training the network in parallel by using a GPU and a training data set to obtain a trained network;
the verification unit is connected with the network design training unit and used for verifying the trained network by adopting a verification data set to obtain a verified intelligent random noise suppression network;
and the noise suppression unit is connected with the verification unit and used for acquiring pre-stack seismic data and inputting the pre-stack seismic data into the verified intelligent random noise suppression network to obtain pre-stack seismic data after random noise suppression.
Compared with the prior art, the invention has the beneficial effects that:
the method applies the U-NET network to the seismic data denoising, effectively avoids the high requirements of the conventional convolutional network on the training data volume and the neural network scale, realizes the network training of the parallel multi-GPU data, improves the training efficiency, and realizes the effective suppression of the random noise of the pre-stack data.
Drawings
Fig. 1 is a schematic diagram of a U-NET network structure adopted by the present invention.
FIG. 2 is a schematic diagram illustrating the principle of implementing multi-GPU data parallelism according to the present invention.
FIG. 3 is a schematic diagram of a real-time operation state of a GPU card when multiple GPUs are in parallel during network training according to the present invention.
Fig. 4 shows the noise-free single shot data of the verification data.
Fig. 5 shows the single shot data after noise addition of the verification data.
Fig. 6 is a comparison of the single-pass results of the noise-free tag data and the noise-added data of the present verification data.
FIG. 7 is the single shot data after the trained network throttle recovery of this validation data.
Fig. 8 is a comparison of the single-channel results of the noise-free tag data and the noise-free data of the verification data.
Fig. 9 is actual noisy seismic data of this validation data.
FIG. 10 shows the data of this verification data after noise suppression of the actual data;
FIG. 11 is a block diagram of the steps of the method of the present invention;
FIG. 12 is a block diagram of the components of the system of the present invention.
Detailed Description
The invention is described in further detail below with reference to the accompanying drawings:
the method is oriented to the problem of random noise suppression of pre-stack seismic data, a U-NET network is established based on Tensorflow to perform multi-GPU data parallel training on the prepared data set, a computational efficient intelligent denoising process is developed, and the denoising effect of the generated algorithm process is verified through verification data.
The structure of the U-NET network adopted by the invention is shown in figure 1, the whole network is a full convolution neural network and is divided into a down-sampling part and an up-sampling part, a large number of channels are still kept in an up-sampling path of the network, and the high resolution of the grid is ensured. The basic structural unit of the downsampling part is three groups of two convolutional layers (the number of channels is the same, the size of a convolutional kernel is 3x3) which are connected with a maximum pooling layer (2x2, the step length is 2x2), the number of the convolutional layer channels is increased from the initial 64 to 256, the number of each group of channels is 2 times of that of the previous group, downsampling of the image is completed under the effect of the pooling layer, and the number of horizontal and vertical sampling points of the image is half of that of the previous group. The up-sampling part adopts a NEAREST NEIGHBOR interpolation algorithm (NEAREST _ NEIGHBOR) to perform image up-sampling, and further adopts a cutting and copying method to splice with a picture with higher resolution in the left down-sampling part, thereby ensuring the richness and resolution of information in the up-sampling process. The up-sampling convolution process reduces the number of channels continuously and keeps the symmetry with the down-sampling path, and finally obtains the final output result through convolution of 1x 1.
An important structure of the U-net network different from other deep learning networks is that: each time of upsampling, a network layer with the same number of channels as that of the corresponding downsampling part needs to be fused once, that is, when the two groups of images are different in size, the corresponding image in the downsampling process needs to be cut once to ensure that the size of the corresponding image is the same as that of the image in the upsampling process, and further, the two groups of image arrays are fused once in the dimension where the number of channels is located, that is, the images are copied and spliced, as shown in fig. 1. When the sizes of the images in the up-sampling and down-sampling processes are consistent, the process can be simplified into simple copy splicing, and the cutting process is omitted.
The deep learning network structure adopted by the invention is that on the basis of the existing U-net network (originally applied to image segmentation), the grid structure is basically kept consistent (being a full convolution neural network and divided into two parts of down sampling and up sampling), and further, according to the self-demand and the data characteristic of the invention, the number of layers of the whole network and the number of channels of each convolution layer are pertinently reduced, and algorithms such as an up sampling algorithm and the like are modified to integrally adapt to the image processing demand of the invention.
The sizes of the input and output images of the U-NET network are kept consistent, and the end-to-end training process is ensured. In order to ensure that the relative relation between the data energy and the amplitude before and after denoising is kept unchanged, the loss function is L2 norm (the existing L2 norm formula is adopted), and an Adam optimizer (the existing Adam optimizer) is adopted for iterative updating.
In addition, in order to improve training efficiency, the invention realizes network training based on GPU data parallel, and the implementation process is as shown in fig. 2, where the CPU is responsible for data distribution, gradient stipulation (after data distribution, gradient update is performed in each GPU card, then each update amount is averaged in the CPU to be used as the current gradient update amount), and model update, and the GPU is responsible for model training.
Specifically, the CPU distributes training sample data (for the present invention, data before denoising and corresponding denoised label data), the size of the training data taken by the CPU each time is batch _ size × GPU card number, and then the CPU distributes the data to each GPU card in an average manner, where the data size of each GPU card is batch _ size. And the CPU averagely distributes the training data taken each time to each GPU card, wherein the data volume of each GPU card is batch _ size.
The parallel mode adopted by the invention is data parallel, namely each parallel unit (each GPU card) adopts the same network model for training, the trained result (namely the gradient updating amount) is subjected to gradient calculation (the training process in each GPU card is the conventional deep learning network training process, the current gradient can be obtained by only calling the gradient calculation API of the deep learning network optimization iterator in the network training process by giving a corresponding loss function and an optimization iterator), the gradient value is transmitted to a CPU for specification averaging, the CPU is responsible for applying the gradient to the optimizer and updating the model (the updating of each network layer weight parameter (specifically, the convolution kernel parameter) of the network model updating index in the deep learning process), the Adam algorithm is adopted in the invention, the main idea is to continuously solve the gradient iteration for updating the neural network weight, eventually minimizing the target functional).
The loss in fig. 2 is a value of an objective function obtained based on a current network weight coefficient, that is, an L2 norm of a difference between an output result obtained by inputting current input data into the network and an expected result (a tag result). The gradient at this moment can be found based on the existing Adam optimization iterator in the deep learning tensoflow framework.
The embodiment of the method of the invention is as follows:
[ EXAMPLES one ]
As shown in FIG. 11, the pre-stack seismic data random noise suppression method based on the U-NET network comprises the following steps:
(1) preparing a data set by using actual seismic data and the ante-stack shot set forward modeling data: a key step of deep learning processing is the preparation of a label data set, and in order to ensure the generalization ability of a training network, a high-quality label data source for an intelligent noise suppression technology can consist of two parts: 1, carrying out data analysis and extraction on high-quality de-noising seismic data of a typical exploration area: the actual data label is an important component for improving the generalization capability of the deep network, so that the data selection should cover various actual typical exploration areas, and the data with high-quality denoising effect is selected from the actual typical exploration areas, so that the robustness of the data set is improved; 2, forward modeling data of a folded-in-front shot set: the actual data denoising effect is influenced by the denoising algorithm and the data quality of the data to a great extent, and the denoising result corresponding to the forward simulation data is definitely known, so that the network denoising effect can be improved. The actual seismic data accounts for 40%, and the simulation data of the prestack shot gather forward modeling accounts for 60%.
According to the invention, a large amount of noise-free pre-stack shot gather data (namely noise-free single shot data) can be obtained through forward modeling and used as the label data of expected output after denoising, and the preparation of the corresponding pre-denoising data is mainly realized through two ways: 1, only extracting noise data (data difference before and after denoising) from selected actual data, carrying out random data matching on the noise data and the prepared forward modeling data, and summing to obtain noise-containing data; and 2, adding Gaussian random noise to the noise-free simulation data to serve as noise-containing data. The noisy data and the tag data constitute a data set.
The data obtained in the above two ways each account for 50% of the model data. Finally, writing the data into a compressed file according to the magic number (magic-4 bytes), the image number (4 bytes), the single image row number (4 bytes), the single image column number (4 bytes) and the rule of the data in the storage format of the MNIST data set to prepare the data set; the data set is divided into a training data set and a verification data set, and the proportion of the training data set and the verification data set can be set according to actual needs.
(2) Designing a network, and utilizing a GPU and a training data set to carry out parallel training on the network to obtain a trained network: the network structure, the loss function, the iterative optimizer and the GPU parallel training algorithm designed based on the invention are used for network training: in the training process, noise-free single shot data in a training data set is used as label data, random noise addition is randomly carried out on the label data according to a set signal-to-noise ratio in the training process of each Batch (each GPU trains one Batch) to enhance the training data, two groups of data (the two groups of data are noise-containing data before and after the random noise is added and are used as data input, and the noise-free data is output as expected) are input into a network for training, and finally, an intelligent random noise suppression network is obtained (the whole training process is a process of continuously updating a deep learning network, network parameters are continuously optimized in the process, the network stops training when the expected training precision is reached, and the corresponding network weight parameters are the final intelligent noise suppression network) and the trained network is stored;
(3) verifying the trained network by adopting verification data to obtain a verified intelligent random noise suppression network: overloading the stored network, inputting actual seismic data to be denoised in the verification data set for data denoising, verifying the effect of the training network, and obtaining a verified intelligent random noise suppression network;
(4) and acquiring pre-stack seismic data, and inputting the pre-stack seismic data into the verified intelligent random noise suppression network to obtain pre-stack seismic data after random noise suppression.
The invention also provides a system for suppressing the random noise of the pre-stack seismic data, which comprises the following embodiments:
[ example two ]
As shown in fig. 12, the system includes:
the data set preparation unit 10 is used for preparing a data set by utilizing actual seismic data and the ante-stack shot set forward modeling data, and dividing the data set into a training data set and a verification data set;
the network design training unit 20 is connected with the data set preparation unit 10 and used for designing a network and training the network in parallel by using a GPU and a training data set to obtain a trained network;
the verification unit 30 is connected with the network design training unit 20 and is used for verifying the trained network by adopting a verification data set to obtain a verified intelligent random noise suppression network;
and the noise suppression unit 40 is connected with the verification unit 30 and used for acquiring pre-stack seismic data and inputting the pre-stack seismic data into the verified intelligent random noise suppression network to obtain pre-stack seismic data after random noise suppression.
Examples of applications of the invention are as follows:
[ EXAMPLE III ]
The training input data size is 128x1000, where a total dataset containing 10 thousand single shot results is prepared. The above-mentioned U-NET network is established, and the learning rate is set to 0.001, in order to further improve the generalization ability of the network during training, 5% gaussian noise is further added for data enhancement, and additionally, the batch _ size is set to 20, and the number of learning rounds is set to 15 rounds. In addition, the size of the video memory of the cluster GPU used in the test is RTX 2080Ti, the number of the GPU cards used for training is seven, the running state of the GPU cards in the model training process is shown in figure 3, and as can be seen from the figure, 7 GPU cards have high utilization rate, and the efficiency of network training is effectively improved.
And carrying out network training based on the parameters, and storing the trained model. In the testing process, firstly, model data is adopted for testing, 8% of noise is added to the live data for testing, the input noise-free single shot data is shown in figure 4, the single shot data added with 8% of noise is shown in figure 5, 20 th single-channel data pairs of figures 4 and 5 are extracted, for example, the 20 th single-channel data pairs are shown in figure 6, and compared with the data before and after noise addition, it can be obviously seen that effective signals are submerged to a great extent by the noise-added data noise at the moment, and the conventional noise removal has great difficulty. At this time, the noise-added data is denoised by adopting the trained network, the obtained denoised data is shown in fig. 7, as can be seen from fig. 7, the noise at this time is basically suppressed, the effective signal is recovered, in order to verify the denoising effect of the network, the 20 th channel of single-channel data of the noise-free data and the denoised data is extracted for comparison, as shown in fig. 8, the data can be obtained from fig. 8, the error between the denoised data and the noise-free data of the label is extremely small, the energy and the form of the position of the effective axis are well recovered, and the denoising effect of the established network is verified.
In order to further verify the denoising effect of the network, the actual data is further selected for testing, the input noisy data is shown in fig. 9, and as can be seen from the figure, the actual data has low signal-to-noise ratio and serious noise development, and the effective signals are submerged to a great extent. The data obtained by processing the data by using the network obtained by the invention is shown in fig. 10, and as can be seen from fig. 10, the random noise is well suppressed, the signal-to-noise ratio is greatly improved, and the processing effect of the intelligent denoising network obtained by the invention is further verified.
Finally, it should be noted that the above-mentioned technical solution is only one embodiment of the present invention, and it will be apparent to those skilled in the art that various modifications and variations can be easily made based on the application method and principle of the present invention disclosed, and the method is not limited to the above-mentioned specific embodiment of the present invention, so that the above-mentioned embodiment is only preferred, and not restrictive.
Claims (10)
1. A method for suppressing random noise of pre-stack seismic data is characterized by comprising the following steps: the method comprises the steps of obtaining an intelligent random noise suppression network by utilizing actual seismic data, pre-stack shot set forward modeling data and a GPU parallel mode, and processing the actual pre-stack seismic data by utilizing the intelligent random noise suppression network to obtain pre-stack seismic data after random noise suppression.
2. The method of random noise suppression for prestack seismic data as recited in claim 1, wherein: the method comprises the following steps:
(1) preparing a data set by using actual seismic data and forepart shot set forward modeling data, and dividing the data set into a training data set and a verification data set;
(2) designing a network, and performing parallel training on the network by using a GPU and a training data set to obtain a trained network;
(3) verifying the trained network by adopting a verification data set to obtain a verified intelligent random noise suppression network;
(4) and acquiring pre-stack seismic data, and inputting the pre-stack seismic data into the verified intelligent random noise suppression network to obtain pre-stack seismic data after random noise suppression.
3. The method of claim 2, wherein the method comprises: the actual seismic data in the step (1) are selected from de-noised seismic data of a typical exploration area and have a high-quality de-noising effect;
and (2) the forward modeling simulation data of the pre-stack shot set in the step (1) is simulation data obtained after forward modeling is carried out on the pre-stack shot set.
4. The method of claim 3, wherein the method comprises: the actual seismic data accounts for 40%, and the simulated data accounts for 60%.
5. The method of claim 4, wherein the method comprises: the operation of preparing the data set by using the actual seismic data and the folded-in-front shot set forward modeling data in the step (1) comprises the following steps:
taking the analog data as tag data;
carrying out random data matching on noise data extracted from actual seismic data and simulated data and then summing to obtain noise-containing data, or adding Gaussian random noise into the simulated data to obtain the noise-containing data;
noisy data and tag data constitute the data set.
6. The method of claim 5, wherein the method comprises: writing the data set into the compaction file in the storage format of the MNIST data set.
7. The method of random noise suppression for prestack seismic data as recited in claim 2, wherein: the network designed in the step (2) is a U-NET network;
and performing image up-sampling in an up-sampling part of the U-NET network by adopting a nearest neighbor interpolation algorithm.
8. The method of random noise suppression for prestack seismic data as recited in claim 2, wherein: the operation of performing parallel training on the network by using the GPU and the training data set in the step (2) to obtain the trained network includes:
the CPU distributes the training data set to each GPU;
each GPU trains the U-NET network by using a training data set, obtains a gradient value according to a training result and then sends the gradient value to a CPU;
the CPU performs protocol averaging on the gradient values sent by all the GPUs to obtain a gradient mean value, and applies the gradient mean value to a loss function and an optimizer to obtain an updated model;
and obtaining the trained network after multiple updates.
9. The method of random noise suppression for prestack seismic data as recited in claim 8, wherein: the loss function adopts an L2 norm;
the optimizer adopts an Adam optimizer.
10. The utility model provides a pre-stack seismic data random noise suppression system which characterized in that: the system comprises:
the data set preparation unit is used for preparing a data set by utilizing actual seismic data and the ante-stack shot set forward modeling data and dividing the data set into a training data set and a verification data set;
the network design training unit is connected with the data set preparation unit and used for designing a network and training the network in parallel by using a GPU and a training data set to obtain a trained network;
the verification unit is connected with the network design training unit and used for verifying the trained network by adopting a verification data set to obtain a verified intelligent random noise suppression network;
and the noise suppression unit is connected with the verification unit and used for acquiring pre-stack seismic data and inputting the pre-stack seismic data into the verified intelligent random noise suppression network to obtain pre-stack seismic data after random noise suppression.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011171331.3A CN114509814A (en) | 2020-10-28 | 2020-10-28 | Pre-stack seismic data random noise suppression method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011171331.3A CN114509814A (en) | 2020-10-28 | 2020-10-28 | Pre-stack seismic data random noise suppression method and system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114509814A true CN114509814A (en) | 2022-05-17 |
Family
ID=81546937
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011171331.3A Pending CN114509814A (en) | 2020-10-28 | 2020-10-28 | Pre-stack seismic data random noise suppression method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114509814A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115497573A (en) * | 2022-09-02 | 2022-12-20 | 广东省科学院生态环境与土壤研究所 | Method for predicting and preparing properties of carbon-based biological geological catalytic material |
-
2020
- 2020-10-28 CN CN202011171331.3A patent/CN114509814A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115497573A (en) * | 2022-09-02 | 2022-12-20 | 广东省科学院生态环境与土壤研究所 | Method for predicting and preparing properties of carbon-based biological geological catalytic material |
| CN115497573B (en) * | 2022-09-02 | 2023-05-19 | 广东省科学院生态环境与土壤研究所 | Carbon-based biological and geological catalytic material property prediction and preparation method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110119780B (en) | Hyper-spectral image super-resolution reconstruction method based on generation countermeasure network | |
| Liu et al. | Seismic data reconstruction via wavelet-based residual deep learning | |
| CN109035142B (en) | Satellite image super-resolution method combining countermeasure network with aerial image prior | |
| CN102156875B (en) | Image super-resolution reconstruction method based on multitask KSVD (K singular value decomposition) dictionary learning | |
| CN110599401A (en) | Remote sensing image super-resolution reconstruction method, processing device and readable storage medium | |
| Wu et al. | Incoherent dictionary learning for reducing crosstalk noise in least-squares reverse time migration | |
| CN110490219B (en) | Method for reconstructing seismic data by U-net network based on texture constraint | |
| CN109214989A (en) | Single image super resolution ratio reconstruction method based on Orientation Features prediction priori | |
| CN111368710A (en) | Seismic data random noise suppression method combined with deep learning | |
| CN113256508A (en) | Improved wavelet transform and convolution neural network image denoising method | |
| CN115330643B (en) | Earthquake denoising method based on convolutional neural network and visual transformation neural network | |
| CN113269818B (en) | Deep learning-based seismic data texture feature reconstruction method | |
| CN116029902A (en) | Knowledge distillation-based unsupervised real world image super-resolution method | |
| CN114418886A (en) | Robustness denoising method based on deep convolution self-encoder | |
| CN115526779A (en) | Infrared image super-resolution reconstruction method based on dynamic attention mechanism | |
| CN115205136A (en) | Image rain removing method based on Fourier prior | |
| CN114509814A (en) | Pre-stack seismic data random noise suppression method and system | |
| CN112862684A (en) | Data processing method for depth map super-resolution reconstruction and denoising neural network | |
| Yang et al. | RSAMSR: A deep neural network based on residual self-encoding and attention mechanism for image super-resolution | |
| CN116068644B (en) | Method for improving resolution and noise reduction of seismic data by using generation countermeasure network | |
| CN108037533B (en) | A kind of seismic prospecting noise drawing method recombinating expectation log-likelihood based on block | |
| Fu et al. | Double enhanced residual network for biological image denoising | |
| CN116205807A (en) | Deep learning-based frozen electron microscope image denoising method and system | |
| CN105761210A (en) | Low resolution image amplification method and device based on adaptive impact filtering | |
| CN112907456B (en) | Deep neural network image denoising method based on global smooth constraint prior model |
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 |