CN112213785A - Seismic data desert noise suppression method based on feature-enhanced denoising network - Google Patents
Seismic data desert noise suppression method based on feature-enhanced denoising network Download PDFInfo
- Publication number
- CN112213785A CN112213785A CN202011115922.9A CN202011115922A CN112213785A CN 112213785 A CN112213785 A CN 112213785A CN 202011115922 A CN202011115922 A CN 202011115922A CN 112213785 A CN112213785 A CN 112213785A
- Authority
- CN
- China
- Prior art keywords
- noise
- layer
- network
- desert
- feature
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 42
- 230000001629 suppression Effects 0.000 title claims abstract description 15
- 230000006870 function Effects 0.000 claims abstract description 15
- 230000004913 activation Effects 0.000 claims abstract description 6
- 238000012549 training Methods 0.000 claims description 26
- 230000010339 dilation Effects 0.000 claims description 15
- 238000005070 sampling Methods 0.000 claims description 8
- 238000013507 mapping Methods 0.000 claims description 7
- 238000010606 normalization Methods 0.000 claims description 5
- 230000015556 catabolic process Effects 0.000 claims description 4
- 238000005520 cutting process Methods 0.000 claims description 4
- 238000006731 degradation reaction Methods 0.000 claims description 4
- 150000001875 compounds Chemical class 0.000 claims description 3
- 230000008859 change Effects 0.000 claims description 2
- 238000010586 diagram Methods 0.000 description 22
- 230000000694 effects Effects 0.000 description 10
- 238000004422 calculation algorithm Methods 0.000 description 8
- 238000013527 convolutional neural network Methods 0.000 description 7
- 238000001228 spectrum Methods 0.000 description 7
- 238000012545 processing Methods 0.000 description 6
- 238000011084 recovery Methods 0.000 description 6
- 239000002131 composite material Substances 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 4
- 238000002474 experimental method Methods 0.000 description 4
- 238000001914 filtration Methods 0.000 description 4
- 230000000052 comparative effect Effects 0.000 description 3
- 230000009467 reduction Effects 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000009826 distribution Methods 0.000 description 2
- 239000000203 mixture Substances 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 108010003272 Hyaluronate lyase Proteins 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 229910052500 inorganic mineral Inorganic materials 0.000 description 1
- 239000011707 mineral Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000004445 quantitative analysis Methods 0.000 description 1
- 239000011435 rock Substances 0.000 description 1
- 239000013589 supplement Substances 0.000 description 1
- 230000002195 synergetic effect Effects 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
- 238000012360 testing method Methods 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
- 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
- G01V1/364—Seismic filtering
-
- 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
-
- 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
- G01V2210/324—Filtering
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)
- Filters That Use Time-Delay Elements (AREA)
Abstract
The invention discloses a seismic data desert noise suppression method based on a feature-enhanced denoising network, which comprises the following steps: step one, constructing an initial characteristic enhancement denoising network: the layers 1-3 and 9-12 are composed of convolution, batch standardization and linear rectification units, the layers 4-8 are composed of expansion convolution, batch standardization and linear rectification units, input noise records are mapped and mixed with the characteristic information of the layers 3, 6, 9 and 12 through cascade operation, the paths of the layers 3, 6, 9 and 12 in the cascade operation are provided with expansion convolution layers, and finally, the expansion convolution layers are output through convolution after passing through an activation function; the expansion factors of the 4 th layer and the 8 th layer are respectively 2, 3, 4, 3 and 2; step two, residual error learning is carried out on the initial feature enhancement denoising network to obtain an ideal feature enhancement denoising network; and thirdly, inputting the original noisy record into the ideal characteristic enhanced denoising network, and outputting a denoised seismic signal.
Description
Technical Field
The invention relates to the technical field of geophysical, in particular to a seismic data desert noise suppression method based on a feature enhancement denoising network.
Background
Seismic exploration technology is an effective resource exploration means. Seismic waves are obtained by a series of technologies, and are analyzed to obtain geological information, obtain an underground rock stratum structure and obtain the mineral resource distribution condition. With the continuous decrease of oil and gas resources, the environment of seismic exploration is more complex, and a great amount of noise is often introduced while seismic data are acquired. Therefore, the primary task in processing seismic signals is to suppress noise in seismic exploration data, thereby improving the quality of the seismic data and preparing for subsequent processing and interpretation. However, desert seismic recording noise has complex characteristics in desert regions due to special natural conditions and human activities, among other factors. The desert noise brings great difficulty to the denoising work: the desert noise has the characteristics of large amplitude, non-stability, nonlinearity, high energy and the like, and most algorithms seriously attenuate effective signals in the noise suppression process; desert noise has low frequency characteristics, which causes the noise and the effective signal to be heavily overlapped in the frequency domain. Therefore, many time-frequency domain filtering methods cannot fully recover the effective seismic signal while suppressing noise; besides gaussian noise, desert noise also contains a large amount of non-gaussian noise, but many methods for denoising desert seismic data can only remove the gaussian part, so that the final denoising effect is not ideal. Therefore, suppressing desert noise is very challenging.
Currently, researchers have proposed many effective seismic data noise suppression algorithms. These methods typically separate the noise and the desired signal in different transform domains. In particular, based on the predictability of the in-phase axes, Canales applied F-X deconvolution (Canales, 1984) to the task of denoising seismic signals, and a number of corresponding improved algorithms were proposed in succession (Gulunay, 1986; Hornbostel, 1991). F-X deconvolution has become a common algorithm in the field of seismic signal denoising. In 2014, Chen et al proposed an Empirical Mode Decomposition (EMD) based (Chen et al, 2014) seismic data noise removal method. EMD decomposes an input signal into multiple modes according to different frequency components to remove noise. In addition, there are other excellent classical seismic data denoising methods, such as band pass filters (Ma et al, 2019), time-frequency peak filtering (Liu et al, 2020), and wavelet transforms (mouswavi et al, 2016). Although the above method can improve the signal-to-noise ratio to some extent, it is often limited. For example, noise needs to meet certain characteristics (e.g., linearity, stationarity) and from a certain distribution (e.g., gaussian), the input signal-to-noise ratio cannot be too low, and many of these methods face the problems of manual parameter selection and complex optimization algorithms. Especially for complex desert noise, more effort and time are usually required to obtain the best noise reduction effect.
The traditional denoising algorithm achieves the optimal condition and is difficult to further improve. In recent years, with the rapid development of computational tools and deep learning, Convolutional Neural Networks (CNNs) have received much attention. Due to low computational complexity and good performance, CNNs have successfully handled different tasks such as object detection (Lee et al, 2020), image denoising (Zhang et al, 2016), image classification (Lee et al, 2017) and motion recognition (Cheron et al, 2015). In recent years, CNNs have also found wide application in seismic exploration, such as seismic data interpolation (Wang et al, 2019), seismic fault detection (Xiang et al, 2018), seismic inversion (Das et al, 2018), and the like. In addition, CNN also has many successful applications in seismic data denoising (Wu et al, 2019; ZHao et al, 2019), opens up a new way for solving the seismic data denoising task, but with the increase of depth, the influence of the shallow layer on the whole network, especially the deep layer, is gradually reduced, and the suppression of complex desert noise is not facilitated.
Disclosure of Invention
The invention provides a method for suppressing the desert noise of seismic data based on a feature-enhanced denoising network for solving the technical defects at present, and the reception field is improved by mixed expansion convolution; the method adopts multiple paths to fuse originally input strong noise information and characteristic information of different layers, thereby increasing the width of the network and obtaining better noise reduction effect.
The technical scheme provided by the invention is as follows:
a seismic data desert noise suppression method based on a feature-enhanced denoising network comprises the following steps:
step one, constructing an initial characteristic enhancement denoising network;
the 1 st to 3 rd layers and the 9 th to 12 th layers of the initial feature enhancement denoising network are composed of convolution, batch normalization and linear rectification units, the 4 th to 8 th layers are composed of expansion convolution, batch normalization and linear rectification units, the 13 th layer is cascade operation, the cascade operation is to map and mix input noise records and feature information of the 3 rd layer, the 6 th layer, the 9 th layer and the 12 th layer, expansion convolution layers are arranged in paths of the 3 rd layer, the 6 th layer, the 9 th layer and the 12 th layer in the cascade operation, the 14 th layer is an activation function, and the 15 th layer is composed of convolution;
the dilation factor of the 4 th layer is 2, the dilation factor of the 5 th layer is 3, the dilation factor of the 6 th layer is 4, the dilation factor of the 7 th layer is 3, the dilation factor of the 8 th layer is 2, the size of the convolution filter of the 1 st layer is set to be 3 × 3 × 1 × 128, the size of the convolution filter of the 2 nd to 12 th layers is set to be 3 × 3 × 128 × 128, the size of the convolution filter of the 15 th layer is set to be 3 × 3 × 5 × 1, and the size of the convolution filter of the extended convolutional layer is set to be 3 × 3 × 128 × 1;
step two, residual error learning is carried out on the initial feature enhancement denoising network to obtain an ideal feature enhancement denoising network;
and step three, inputting the original noisy record into the ideal characteristic enhancement denoising network, and outputting the denoised seismic signal.
Preferably, the residual learning includes the steps of:
n≈R(y；Θ)；
in the formula, y is an original noisy seismic signal, x is a denoised seismic signal, v is desert noise, n is residual mapping, and theta is a network parameter;
in the formula (I), the compound is shown in the specification,
representing S pairs of the noisy-clean training set,
represents the Frobenius norm;
and 3, substituting the optimal network parameters into the initial characteristic enhancement denoising network to obtain an ideal characteristic enhancement denoising network.
Preferably, the training set comprises a signal set and a noise set.
Preferably, the signal set is Ricker waves with different main frequencies and curvatures, 34 records with the size of 1250 × 100 are generated, and then the Ricker waves are cut to obtain training blocks with the size of 50 × 50, and the cutting step size is 10.
Preferably, the noise set comprises desert noise and surface waves, the desert noise record comprises 2300 tracks, and each track comprises 30000 sampling points; the surface wave record comprises 800 channels, and each channel comprises 1200 sampling points; the desert noise record or surface wave record is then cut into 50 x 50 noise blocks with a cutting step size of 25.
Preferably, the change in signal-to-noise ratio of the training set is achieved by randomly selecting a random coefficient within the range of [0.1, 1] to multiply the noise block.
Preferably, the activation function is a hyperbolic tangent function.
Compared with the prior art, the invention has the following beneficial effects:
(1) the invention adopts a plurality of paths to fuse the originally input strong noise information and the characteristic information of different layers, increases the width of the network, fully utilizes the influence of the whole network, and is beneficial to extracting more desert noise characteristics under a complex background.
(2) The method applies the hybrid expansion convolution technology to the model to improve the receptive field, which plays an important role in obtaining more context information in the denoising task.
(3) The method also adopts residual learning to promote network training, and compared with the existing denoising method, the method has better performance in desert noise, surface wave suppression and seismic event recovery than the existing denoising method, and has better performance in seismic data denoising.
Drawings
FIG. 1 is a schematic diagram of a feature-enhanced denoising network according to the present invention.
Fig. 2 is a schematic diagram of the dilated convolutional network of the present invention.
Fig. 3 is a schematic diagram of a dilated convolutional network with symmetric hopping connections according to the present invention.
FIG. 4 is a schematic diagram of a dilated convolution network with a single feature enhancement path according to the present invention.
FIG. 5 is a schematic diagram of a clean record of a composite record according to the present invention.
Fig. 6 is a schematic diagram of a desert noise recording of the composite recording of the present invention.
Fig. 7 is a schematic illustration of a noisy recording of a composite recording according to the present invention.
FIG. 8 is a diagram illustrating the FEDnet denoising result of the synthetic record according to the present invention.
FIG. 9 is a schematic representation of the residual results of the synthetic recordings of the present invention.
FIG. 10 is a schematic of FK spectra of clean records of synthetic records of the invention.
Fig. 11 is a schematic diagram of FK spectra of desert noise for the synthetic recordings described in this invention.
Fig. 12 is a schematic of the FK spectra of noisy recordings of synthetic recordings according to the invention.
FIG. 13 is a schematic diagram of FK spectra of FEDnet denoising results of synthetic recordings according to the present invention.
Fig. 14 is a schematic of the FK spectrum of the residual results of the synthetic recordings of the present invention.
FIG. 15 is a schematic diagram of the denoising result of the bandpass filter for synthetic recording according to the present invention.
FIG. 16 is a diagram of the DnCNN denoising result of the synthetic record of the present invention.
FIG. 17 is a diagram of the FEDnet denoising result of the synthetic record of the present invention.
FIG. 18 is a diagram illustrating comparison of 28 th-channel signals of the denoising result of the band-pass filter for synthesis recording according to the present invention.
FIG. 19 is a comparative schematic diagram of the 28 th signal of the DnCNN de-noising result of the synthetic record of the present invention.
FIG. 20 is a schematic diagram comparing the 28 th signal of the FEDnet denoising result of the synthetic record according to the present invention.
Fig. 21 is a schematic diagram of an actual desert earthquake according to the present invention.
FIG. 22 is a schematic diagram of the denoising result of the band-pass filter of the actual desert earthquake according to the present invention.
FIG. 23 is a schematic diagram of a DnCNN denoising result of an actual desert earthquake according to the present invention.
FIG. 24 is a diagram of the FEDnet denoising result of an actual desert earthquake according to the present invention.
FIG. 25 is a schematic diagram showing comparison of the band-pass filter residual results of an actual desert earthquake according to the present invention.
FIG. 26 is a schematic diagram showing a comparison of DnCNN residual results of actual desert earthquakes in accordance with the present invention.
FIG. 27 is a schematic diagram showing comparison of FEDnet residual results of actual desert earthquake according to the present invention.
Detailed Description
The present invention is described in further detail below to enable those skilled in the art to practice the invention with reference to the description.
The invention provides a seismic data desert noise suppression method based on a feature-enhanced denoising network, which utilizes background information to reconstruct signal points for denoising and comprises the following steps:
step one, constructing an initial characteristic enhancement denoising network;
the expansion convolution is adopted to improve the receptive field and simultaneously help to reduce the depth of the network and the complexity of training, and the expansion convolution with the expansion factor is to spread convolution kernels by 0, namely the expansion factor represents the space between kernels. In the dilated convolutional network, when the dilation factor of each layer is 2, the receptive field size is (4n +1) × (4n +1), where n represents the depth. For example, when n is 8, the receptive field size is 33 × 33, which corresponds to the receptive field size of a 16-layer conventional CNN with a filter size of 3 × 3.
Although the dilation convolution can increase the receptive field size and make better use of the context information than conventional convolution, it can lead to grid artifacts, with zeros participating in the computation in the filter of the dilation convolution network. When the expansion factor is 2, the convolution kernel is expanded from 3 × 3 to 5 × 5, and although the region has 25 signal points, only 9 signal points are used for calculation, and the problem of grid artifacts is solved by adopting an almost symmetrical mixed expansion factor setting so that each layer has a different expansion factor.
As shown in FIG. 1, Conv, scaled Conv, BN and ReLU in the graph are convolution, expanded convolution, batch normalization and linear rectification units, respectively, the 1-3 layers and 9-12 layers of the initial feature enhanced denoising network are Conv + BN + ReLU, 4-8 layers of scaled Conv + BN + ReLU, the expansion factor of the 4 layer is 2, the expansion factor of the 5 layer is 3, the expansion factor of the 6 layer is 4, the expansion factor of the 7 layer is 3, the expansion factor of the 8 layer is 2, the convolution filter of the 1 layer is set to be 3 × 3 × 1 × 128, the convolution filter of the 2-12 layers is set to be 3 × 3 × 128 × 128, therefore, the receptive field of the first 12 layers of the initial feature enhanced denoising network has the same receptive field size as that of the 21-layer conventional network (the receptive field size is 43), the depth of the initial characteristic enhanced denoising network is reduced, and therefore the efficiency of network training is improved.
An increase in depth may lead to a decrease in network performance, one of the main reasons being that relatively deep networks may face problems with shallow layers having a reduced effect on deep layers. Therefore, in order to suppress desert noise and fully utilize the influence of the whole network, the initial feature enhancement denoising network adopts a plurality of paths to design a feature enhancement technology, not only selects original noisy input, but also selects feature information of different layers as the supplement of the network, the 13 th layer is a cascade operation, the cascade operation is to map and mix input noise records with the feature information of the 3 rd layer, the 6 th layer, the 9 th layer and the 12 th layer, so that the width of the network is improved, the network is facilitated to obtain more useful noise features, in order to keep the same size as the input noise records, the paths of the 3 rd layer, the 6 th layer, the 9 th layer and the 12 th layer in the cascade operation are provided with extension convolution layers, and the size of convolution filters of the extension convolution layers is set to be 3 × 3 × 128 × 1; "Tanh" at layer 14 represents the hyperbolic tangent function, which is an activation function; the 15 th layer is composed of convolution, the size of the convolution filter of the 15 th layer is set to be 3 multiplied by 5 multiplied by 1, the residual record is reconstructed, and finally the output is a denoised signal.
Step two, residual error learning is carried out on the initial feature enhancement denoising network to obtain an ideal feature enhancement denoising network;
wherein the residual learning comprises:
y＝x+v；
wherein y is an original noisy seismic signal, x is a denoised seismic signal, and v is desert noise;
the purpose of residual learning is to train a model to obtain residual mapping (i.e., desert noise) v, and finally indirectly obtain a target pure record x as y-v, so that training is easier.
Training the initial feature enhancement denoising network based on a degradation equation y which is x + v, and outputting a residual mapping:
n≈R(y；Θ)；
in the formula, y is an original noisy seismic signal, x is a denoised seismic signal, v is desert noise, n is residual mapping, and theta is a network parameter;
in the formula (I), the compound is shown in the specification,
representing S pairs of the noisy-clean training set,
represents the Frobenius norm;
since desert noise has non-gaussian and low frequency characteristics, it is difficult to effectively recover valid seismic events. In order to effectively remove desert noise from seismic data, it is critical to establish a high quality training set that includes a signal set and a noise set.
And selecting Ricker waves simulating the seismic event as a signal set. In order to obtain better denoising effect, the parameters of the Ricker wave should be as close to the real signal as possible. Different dominant frequencies and curvatures were set to generate 34 recordings of 1250 × 100 size, which were then sliced to obtain training blocks of 50 × 50 size with a slicing step size of 10.
The noise set is a main factor influencing the network performance and directly influences the denoising performance of the network. The desert noise and the surface wave are main factors causing the low quality of the desert earthquake record, and the invention takes the two types of noise as a training noise set. In this embodiment, the desert noise data used is from the desert area of the Chinese Tarim basin, the number of tracks in the noise record is 2300, and each track has 30000 sampling points; the surface wave record is collected from the actual seismic data of desert areas in western China, the noise record comprises 800 channels and 1200 sampling points in each channel, and the record is cut into 50 x 50 training blocks by taking 25 as step length.
Randomly taking noise blocks v from a noise setiTo obtain training sets with different signal-to-noise ratios, we are at [0.1, 1]]Within range ofThe machine selects a random coefficient c and then multiplies it by a noise block viAnd obtaining a single network model to finish the denoising tasks with different noise levels.
And 3, substituting the optimal network parameters into the initial characteristic enhancement denoising network to obtain an ideal characteristic enhancement denoising network.
And step three, inputting the original noisy record into the ideal characteristic enhancement denoising network, and outputting the denoised seismic signal.
In order to verify the effectiveness of the method provided by the invention, all networks are trained under the same network training conditions (including training set, patch size, batch size, iteration number and the like), and the network parameters are set according to table 1.
TABLE 1 network training parameters
| Training environment | Is provided with |
| |
50×50 |
| Number of batches | 128 |
| Number of |
50 |
| Range of learning rate | [10-3-10-5] |
(1) In order to verify the denoising capability of different expansion factor setting modes, experiments prove that different expansion convolution network models are used.
As shown in fig. 2, the present invention compares a network without feature enhancement techniques with a conventional convolutional network and a convolutional network with a dilation factor of 2.
In three different networks, the filter size of 2-12 layers, except the first and last layer, is set to 3 × 3 × 128 × 128. The filter sizes of the first layer and the last layer are set to 3 × 3 × 1 × 128 and 3 × 3 × 128 × 1, respectively. For a 13-layer convolutional network, the field of view (field of 45) of the model in fig. 2 is 1.6 times that of the conventional convolutional network (field of 27), which is smaller than the dilated convolutional network (field of 53) with a dilation factor of 2.
In the embodiment, a synthetic seismic record with a signal-to-noise ratio of-6.8420 db and polluted by desert noise is processed through a common convolution network, an expansion convolution network with an expansion factor of 2 and a denoising network without feature enhancement (FEDnet without peak enh) to perform a comparison experiment. Two widely used indicators are used to quantitatively evaluate denoising quality: signal-to-noise ratio (SNR) and Mean Square Error (MSE). As can be seen from table 2, when the network has a spreading factor of 2, the signal-to-noise ratio is higher than that of the ordinary convolutional network since the receptive field is larger, which confirms the importance of the receptive field. Also, although the receptive field of "FEDnet with out eat enh" is not the largest, the SNR is the highest, where "FEDnet with out eat enh" represents the structure of fig. 2, which means that in the dilated convolutional network, the grid artifact is not negligible. Therefore, the feature enhanced denoising network provided by the invention is provided.
TABLE 2 SNR and MSE comparison results for different network architectures
| Method of producing a composite material | SNR(dB) | MSE |
| Expansion factor of 1 | 9.2279 | 0.0061 |
| Expansion factor of 2 | 10.1423 | 0.0050 |
| FEDnet without Feat Enh | 11.5041 | 0.0036 |
(2) The superiority of the method of the invention is illustrated by comparing the performance of different connection modes.
The invention compares the network of three different connection modes and compares the SNR and MSE results. Hopping connections help to train and help the network carry more detailed information. Fig. 3 shows a symmetric jumper scheme, which is widely used in many network architectures such as RED and DSNet. The filter sizes of the first layer and the last layer are set to 3 × 3 × 1 × 128 and 3 × 3 × 128 × 1, respectively. For the remaining layers, the filter size is set to 3 × 3 × 128 × 128. Fig. 4 uses a cascade operation to connect the input original record (global feature) and the feature information map (local feature) of the twelfth layer to enhance the influence of the shallow layer, and in fig. 4, only one extension layer is used, and thus, the filter size of the last layer is set to 3 × 3 × 2 × 1. The filter parameters of the first 12 layers are the same as in fig. 3. However, a single feature enhancement path does not fully exploit the network's impact. Therefore, as shown in fig. 1, the present invention adopts a new feature enhancement network structure to mine more features for denoising task. When the desert seismic data are denoised, the originally input strong noise information is selected, and the characteristic information of a plurality of network layers is selected to be complementary with the characteristic information of a deep layer. The same records as in the previous section were processed in this example and the results of the comparative experiments are shown in table 3, which demonstrates that FEDnet achieves superior performance. Specifically, "Sym Skip Connect" and "Single flat Enh" in table 3 represent the models of fig. 3 and 4, respectively.
TABLE 3 SNR and MSE comparison results for various network architectures
| Method of producing a composite material | SNR(dB) | MSE |
| Expansion factor of 1 | 9.2279 | 0.0061 |
| Expansion factor of 2 | 10.1423 | 0.0050 |
| FEDnet without Feat Enh | 11.5041 | 0.0036 |
| Sym Skip Connect | 12.0989 | 0.0032 |
| Single Feat Enh | 12.1882 | 0.0031 |
| FEDnet | 13.5690 | 0.0023 |
(3) Simulating the results of the recording process
In order to verify the denoising capability of the feature-enhanced denoising network, a simulation record is constructed in the embodiment for simulation experiments. As shown in fig. 5, the seismic event is generated by Ricker waves, the sampling frequency is 500hz, different main frequencies are set to approximate to real noisy seismic signals in the embodiment, and the recording size is 200 × 1400; fig. 6 shows desert noise, and we finally obtain the noise-containing record shown in fig. 7 according to the equation y ═ x + v, and as shown in fig. 8, it is a denoising result of the FEDnet, which illustrates that the FEDnet can completely suppress noise and effectively recover the effective signal in the noise-containing record, and fig. 9 is a residual result of the synthesized record obtained by subtracting the denoising result from the input record, and the result shows that there is no signal energy loss after the FEDnet is denoised. As shown in fig. 10 to 14, the present invention plots a frequency domain wave number spectrum (FK spectrum). As shown in fig. 12, it is apparent that the effective signal is almost submerged in strong low-frequency noise. As can be seen from fig. 13, there is almost no desert noise residual in the denoising result.
As shown in fig. 15-17, to prove the advantage of the feature enhanced denoising network of the present invention in denoising performance, FEDnet is compared with a band pass filter and DnCNN in this embodiment. The filter size of the DnCNN algorithm is set to be the same as the network in fig. 2 to handle both analog recordings and actual recordings: the filter sizes of the first and last layers are set to 3 × 3 × 1 × 128 and 3 × 3 × 128 × 1, respectively, and the filter sizes of the remaining layers are set to 3 × 3 × 128 × 128. The denoising results of the three methods are shown in fig. 15-17, the single-channel comparison results of the three methods are shown in fig. 18-20, the band-pass filtering effect is poor, a large amount of noise is left, and effective signals are still submerged in the noise. As shown in fig. 16, the noise reduction effect of DnCNN seems to be good, but the one-pass results in fig. 19 indicate that DnCNN cannot completely suppress desert noise. In contrast, the single-pass results in fig. 20 indicate that the FEDnet can fully recover the valid signal without noise residue.
Quantitative analysis can more objectively evaluate the quality of the denoising result. Synthetic seismic records of 4 different noise levels were processed in the three methods described above. The comparative results are shown in Table 4. FEDnet demonstrates the best performance.
TABLE 4 comparison of SNR and MSE results for different methods
(4) Actually recording the processing result
The FEDnet is further applied to actual recording of desert earthquake to test the denoising performance of the feature enhanced denoising network. An actual seismic record is shown in FIG. 21, where the record contains 140 traces and 1101 sample points per trace. The sampling frequency was recorded as 500 hz. Likewise, we also apply a band pass filter and DnCNN to the recording. Fig. 22-24 show the denoising effect of the three methods.
Observing the original recording, the seismic events in FIG. 21 are overwhelmed by noise. Particularly, the effective seismic signals marked by the boxes 1 and 2 can not be identified almost, and the seismic event of the box 3 is seriously influenced by surface waves and has poor continuity. The denoising effect of the band-pass filter is poor: although some noise is suppressed, there is still a lot of noise in the recordings and seismic events are still difficult to identify. After the DnCN processing, the continuity of the seismic reflection waves of the frame 1 and the frame 3 is good. But the recovery of the seismic reflections of box 2 is not good enough. In addition, the noise in box 4 is not completely suppressed. In contrast, FEDnet is superior to the above method in terms of noise suppression and signal recovery. In particular, after FEDnet processing, the seismic effective signal recovery in box 2 is better, and the overall de-noised record appears cleaner than DnCNN.
Meanwhile, as shown in fig. 25 to 27, residual maps of the three methods are compared. The DnCNN and the residual image of the feature-enhanced denoising network are found to be darker than a band-pass filter, which means that the DnCNN and the method can more thoroughly suppress desert noise.
For the suppression of the low-frequency desert noise, the processing result of the traditional filtering algorithm or the traditional CNN often has the problems of incomplete signal recovery, serious noise residue and the like. The invention provides a new denoising CNN-FEDnet aiming at the denoising problem of seismic data. It has the following characteristics:
(a) the FEDnet not only adopts the expansion convolution, but also makes full use of the synergistic effect of different layers of characteristic information, and makes a balance between the denoising performance and the denoising efficiency. Particularly, through comparison of different connection modes, the characteristic enhancement denoising network achieves the best denoising performance.
(b) The noise suppression can be automatically and intelligently finished without debugging parameters. For different noise level records, only one model needs to be trained to complete the denoising task, and better performance than that of the traditional denoiser is obtained.
(c) Experiments show that the FEDnet has good desert noise suppression capability and effective signal recovery capability on seismic data with low signal-to-noise ratio. The high-quality and rich training set plays an important role in the denoising performance of the network.
While embodiments of the invention have been described above, it is not limited to the applications set out in the description and the embodiments, which are fully applicable in all kinds of fields suitable for the invention, and further modifications may readily be effected by those skilled in the art, without departing from the general concept defined by the claims and the equivalents thereof, and the invention is therefore not limited to the specific details and embodiments shown and described herein.
Claims (7)
1. A seismic data desert noise suppression method based on a feature-enhanced denoising network is characterized by comprising the following steps:
step one, constructing an initial characteristic enhancement denoising network;
the 1 st to 3 rd layers and the 9 th to 12 th layers of the initial feature enhancement denoising network are composed of convolution, batch normalization and linear rectification units, the 4 th to 8 th layers are composed of expansion convolution, batch normalization and linear rectification units, the 13 th layer is cascade operation, the cascade operation is to map and mix input noise records and feature information of the 3 rd layer, the 6 th layer, the 9 th layer and the 12 th layer, expansion convolution layers are arranged in paths of the 3 rd layer, the 6 th layer, the 9 th layer and the 12 th layer in the cascade operation, the 14 th layer is an activation function, and the 15 th layer is composed of convolution;
the dilation factor of the 4 th layer is 2, the dilation factor of the 5 th layer is 3, the dilation factor of the 6 th layer is 4, the dilation factor of the 7 th layer is 3, the dilation factor of the 8 th layer is 2, the convolution filter of the 1 st layer is set to be 3 × 3 × 1 × 128 in size, the convolution filter of the 2 nd to 12 th layers is set to be 3 × 3 × 128 × 128, the convolution filter of the 15 th layer is set to be 3 × 3 × 5 × 1 in size, and the convolution filter of the extended convolutional layer is set to be 3 × 3 × 128 × 1 in size;
step two, residual error learning is carried out on the initial feature enhancement denoising network to obtain an ideal feature enhancement denoising network;
and thirdly, inputting the original noisy record into the ideal characteristic enhanced denoising network, and outputting a denoised seismic signal.
2. The method for suppressing noise in a seismic data desert based on the feature enhanced denoising network of claim 1, wherein the residual learning comprises the following steps:
step 1, training the initial feature enhancement denoising network based on a degradation equation y which is x + v, and outputting a residual mapping:
n≈R(y；Θ)；
in the formula, y is an original noisy seismic signal, x is a denoised seismic signal, v is desert noise, n is residual mapping, and theta is a network parameter;
step 2, obtaining the optimal network parameters by minimizing a loss function, wherein the loss function satisfies the following conditions:
in the formula (I), the compound is shown in the specification,
representing S pairs of the noisy-clean training set,
represents the Frobenius norm;
and 3, substituting the optimal network parameters into the initial characteristic enhancement denoising network to obtain an ideal characteristic enhancement denoising network.
3. The method of suppressing noise in a seismic data desert based on a feature enhanced denoising network of claim 2, wherein the training set comprises a signal set and a noise set.
4. The method for suppressing the noise of the seismic data desert based on the feature-enhanced denoising network of claim 3, wherein the signal set is Ricker waves with different main frequencies and curvatures, 34 records with the size of 1250 x 100 are generated, then the Ricker waves are cut to obtain training blocks with the size of 50 x 50, and the cutting step size is 10.
5. The method for suppressing the noise in the desert of the seismic data based on the feature-enhanced denoising network of claim 3, wherein the noise set comprises desert noise and surface waves, the desert noise record comprises 2300 tracks, and each track has 30000 sampling points; the surface wave record comprises 800 channels, and each channel comprises 1200 sampling points; the desert noise record or surface wave record is then cut into 50 x 50 noise blocks with a cutting step size of 25.
6. The method as claimed in claim 5, wherein the change of the signal-to-noise ratio of the training set is obtained by randomly selecting a random coefficient to be multiplied by the noise block in the range of [0.1,1 ].
7. The method for suppressing noise in a seismic data desert based on a feature-enhanced denoising network of claim 6, wherein the activation function is a hyperbolic tangent function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011115922.9A CN112213785B (en) | 2020-10-19 | 2020-10-19 | Seismic data desert noise suppression method based on feature-enhanced denoising network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011115922.9A CN112213785B (en) | 2020-10-19 | 2020-10-19 | Seismic data desert noise suppression method based on feature-enhanced denoising network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112213785A true CN112213785A (en) | 2021-01-12 |
| CN112213785B CN112213785B (en) | 2021-08-13 |
Family
ID=74055719
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011115922.9A Expired - Fee Related CN112213785B (en) | 2020-10-19 | 2020-10-19 | Seismic data desert noise suppression method based on feature-enhanced denoising network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112213785B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116594061A (en) * | 2023-07-18 | 2023-08-15 | 吉林大学 | Seismic data denoising method based on multi-scale U-shaped attention network |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP2196827B1 (en) * | 2008-12-03 | 2013-04-24 | PGS Geophysical AS | Method for Determining Signal Quality in Dual Sensor Seismic Streamer Signals |
| CN103218783A (en) * | 2013-04-17 | 2013-07-24 | 国家测绘地理信息局卫星测绘应用中心 | Fast geometric correction method for satellite remote sensing image and based on control point image database |
| CN106569275A (en) * | 2015-10-10 | 2017-04-19 | 中国石油化工股份有限公司 | Wavelet zero-phasing treatment method and device |
| CN107942390A (en) * | 2018-01-16 | 2018-04-20 | 吉林大学 | A kind of desert area seismic prospecting denoising method based on adaptive multiple sharpening diffusion |
| CN111667435A (en) * | 2020-06-19 | 2020-09-15 | 山东工商学院 | GPR noise suppression method and system based on Bayes nonnegative matrix factorization |
-
2020
- 2020-10-19 CN CN202011115922.9A patent/CN112213785B/en not_active Expired - Fee Related
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP2196827B1 (en) * | 2008-12-03 | 2013-04-24 | PGS Geophysical AS | Method for Determining Signal Quality in Dual Sensor Seismic Streamer Signals |
| CN103218783A (en) * | 2013-04-17 | 2013-07-24 | 国家测绘地理信息局卫星测绘应用中心 | Fast geometric correction method for satellite remote sensing image and based on control point image database |
| CN106569275A (en) * | 2015-10-10 | 2017-04-19 | 中国石油化工股份有限公司 | Wavelet zero-phasing treatment method and device |
| CN107942390A (en) * | 2018-01-16 | 2018-04-20 | 吉林大学 | A kind of desert area seismic prospecting denoising method based on adaptive multiple sharpening diffusion |
| CN111667435A (en) * | 2020-06-19 | 2020-09-15 | 山东工商学院 | GPR noise suppression method and system based on Bayes nonnegative matrix factorization |
Non-Patent Citations (2)
| Title |
|---|
| 练秋生，等: "基于多尺度残差网络的压缩感知重构算法", 《自动化学报》 * |
| 郑升: "基于RCSST和CNN的微弱地震信号噪声压制方法研究", 《中国优秀硕士学位论文全文数据库 基础科学辑》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116594061A (en) * | 2023-07-18 | 2023-08-15 | 吉林大学 | Seismic data denoising method based on multi-scale U-shaped attention network |
| CN116594061B (en) * | 2023-07-18 | 2023-09-22 | 吉林大学 | Seismic data denoising method based on multi-scale U-shaped attention network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112213785B (en) | 2021-08-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108107475B (en) | Borehole microseismic denoising method based on experience wavelet transformation and multi-threshold function | |
| CN108845352B (en) | Desert Denoising of Seismic Data method based on VMD approximate entropy and multi-layer perception (MLP) | |
| CN102998706B (en) | Method and system for attenuating seismic data random noise | |
| CN111190227B (en) | Low signal-to-noise ratio seismic data denoising method based on residual convolution generation countermeasure model | |
| CN109031422A (en) | A kind of seismic signal noise suppressing method based on CEEMDAN and Savitzky-Golay filtering | |
| CN106443787B (en) | Prestack seismic gather presses method for de-noising and its device | |
| CN117111155B (en) | Microseism data denoising method based on integrated framework | |
| CN112213785B (en) | Seismic data desert noise suppression method based on feature-enhanced denoising network | |
| CN104133248A (en) | High-fidelity sound wave interference suppression method | |
| CN109212608B (en) | Borehole microseismic signal antinoise method based on 3D shearlet transformation | |
| CN110780349A (en) | Weighted kernel norm minimization algorithm based on enhanced block matching precision, low-frequency noise suppression method in desert earthquake and application | |
| CN114325821A (en) | Method and system for suppressing strong scattering noise in pre-stack seismic data based on 3D-SNACNN network | |
| CN111308555B (en) | Surface wave interference elimination method based on inter-frequency similarity low-frequency reconstruction | |
| Jia et al. | Separating ground-roll from land seismic record via convolutional neural network | |
| CN108037533A (en) | It is a kind of that the seismic prospecting noise drawing method for it is expected log-likelihood is recombinated based on block | |
| CN113093282A (en) | Desert data denoising method based on geometric modal characteristic parallel network | |
| CN102608658A (en) | Strong energy interference suppression method | |
| CN112764108B (en) | Novel seismic data noise suppression algorithm based on improved empirical wavelet transformation | |
| CN115561817A (en) | Desert earthquake denoising method based on multiple attention mechanism | |
| CN110764147B (en) | Desert exploration weak signal recovery method based on VMD local F-X spectrum decomposition | |
| CN113567129A (en) | CEEMD-based noise reduction method for train bearing vibration signal | |
| Lin et al. | Self-similarity convolution neural network for seismic noise suppression in desert environment | |
| Wu et al. | A SNR enhancement method for desert seismic data: Simplified low-rank selection in time–frequency decomposition domain | |
| Ni et al. | Seismic random noise attenuation via noise assisted-multivariate EMD based MSSA | |
| CN118311669B (en) | Distributed optical fiber sensing coupling noise forward modeling and suppressing method |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20210813 |