CN115346112A - Seismic data oil pumping unit noise suppression method based on multilayer feature fusion - Google Patents
Seismic data oil pumping unit noise suppression method based on multilayer feature fusion Download PDFInfo
- Publication number
- CN115346112A CN115346112A CN202210506583.XA CN202210506583A CN115346112A CN 115346112 A CN115346112 A CN 115346112A CN 202210506583 A CN202210506583 A CN 202210506583A CN 115346112 A CN115346112 A CN 115346112A
- Authority
- CN
- China
- Prior art keywords
- noise
- seismic
- pumping unit
- seismic data
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/30—Noise filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
- G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Multimedia (AREA)
- Computing Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to a method for suppressing noise of a seismic data pumping unit based on multilayer feature fusion, which comprises the steps of constructing an end-to-end denoising network of multilayer generators, extracting seismic data features on different scales by a coder of each layer of generator, fusing the features extracted by the generator on the upper layer into a feature map extracted by the generator on the lower layer, and realizing the denoising performance from coarse to fine by extracting the features on different scales layer by layer. Meanwhile, the input of each layer of generator is divided into data blocks with different sizes, so that the network can feel more seismic data areas, the receptive field of the de-noising network is improved, and the generators can extract more useful seismic data characteristics. The structure of the network is benefited, and the seismic data are divided into larger data blocks, so that the network is favorable for extracting more global noise characteristics, and the recognition performance of the noise of the pumping unit is improved.
Description
Technical Field
The invention relates to the field of seismic exploration, in particular to a method for suppressing noise of a seismic data oil pumping unit based on multi-layer feature fusion.
Background
In the actual seismic data acquisition, the seismic data are inevitably polluted by noise, and the introduction of the noise can greatly reduce the quality and the precision of the seismic data, and directly influence the subsequent analysis, explanation and application of the seismic data. In the field of seismic exploration, effective separation of signals from noise has been a hotspot and difficulty of research. The seismic noise mainly comprises coherent noise and incoherent noise, wherein the incoherent noise is irregularly represented in the seismic data and has no fixed frequency and apparent velocity, such as random noise. Coherent noise exhibits relatively obvious and regular characteristics in seismic recording, and its main frequency and apparent velocity are relatively fixed, such as surface waves and multiples. This document focuses on coherent noise generated by pumping unit vibration.
Existing seismic data noise suppression methods can be divided into two categories. The first category is the traditional classification, including filtering and transform-based methods. The filtering method is to separate the signal and the noise by designing a suitable filter, such as an fk filter, an adaptive time-frequency peak filter, a kalman filter, and a particle filter. Considering that seismic data and noise have significant feature differences in some sparse domains, researchers have proposed many successful methods based on sparse transforms, such as: singular value decomposition, principal component analysis, curvelet, wavelet, and shearlet. The traditional method mainly depends on prior information, needs manual parameter adjustment, is time-consuming and labor-consuming, and is not beneficial to efficient processing of massive seismic data. Therefore, a more intelligent and better seismic denoising method is urgently needed to be explored.
The second category is data-driven based. With the success of the deep learning method in the aspects of image processing, target detection, NLP problems and the like, the deep learning method is gradually introduced into the field of seismic exploration and is mainly used for automatic fault detection, phase classification, lithology prediction, noise attenuation and the like of random noise, and Kimiaefar et al combines an artificial neural network with a wavelet packet analysis method to attenuate seismic random noise. Ambigua et al describe deep CNN and transfer learning to suppress not only random noise, but also linear noise and surface wave noise. Xu et al employs a denoised convolutional neural network (DnCNN) to attenuate random noise. The original DnCNN is improved in aspects of patch size, convolution kernel size, network depth and the like, so that the method is suitable for low-frequency desert noise suppression. Poplar et al added residual learning and batch normalization methods to DnCN to reconstruct noise-free seismic data. Residual learning is also introduced into the period GAN to improve the training efficiency of seismic data denoising. For the surface wave noise, the existing surface wave noise attenuation research is mainly a traditional method, the research on a surface wave deep learning method is relatively less, and the method is mainly based on a deep neural network, a generation countermeasure network and a condition generation countermeasure network.
In the existing research on seismic data noise suppression, coherent noise generated by an oil pumping unit is rarely researched. In actual seismic exploration, along with exploration integration, the condition that seismic data are acquired in a work area of an oil pumping unit at the same time cannot be avoided, and under the condition, the acquired seismic data can be interfered by coherent noise of the oil pumping unit, so that the quality of the seismic data is seriously influenced.
According to the spectrum analysis of actual pumping unit data, the fact that the spectrum ranges of pumping unit noise and effective signals are almost completely overlapped is found, so that the traditional denoising method cannot effectively separate seismic signals from pumping unit noise. Furthermore, the coherent noise of the pumping unit is different from the principles of the generation of seismic noise such as random noise, surface wave and the like, and the expression characteristics of the coherent noise of the pumping unit are different on seismic data, so that the conventional deep learning denoising method aiming at the random noise and the surface wave is not suitable for noise suppression of the pumping unit.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a multilayer feature fusion seismic data pumping unit noise suppression method, which obtains simulation data through a Marmousi-II model, realizes coarse-to-fine noise suppression through fusing features extracted by different generators based on a multilayer generator, divides the input of each layer of generator into seismic data blocks with different sizes, effectively enlarges the receptive field of a network, and extracts more useful pumping unit noise features, and comprises the following specific steps:
step 1: producing a simulated data set comprising:
step 11: acquiring a simulated noisy seismic record, selecting a part of a Marmousi-II longitudinal wave velocity model as a forward model, wherein the size of the model is 271 multiplied 351, the space step length is h =5m, the time sampling interval is delta t =0.5ms, the total receiving time length is 3.6s, the Ricker wavelet is used as an explosive source and an oil pumping unit source, the explosive source and the oil pumping unit source work simultaneously, numerical simulation is carried out by adopting a first-order stress-velocity sound wave equation, and the 31-shot seismic record with the oil pumping unit noise is obtained in total;
step 12: setting the seismic source of the oil pumping unit not to vibrate and only the seismic source to vibrate based on the forward modeling of the step 11 to obtain a clean seismic record;
step 13: making a simulation training set, and making noise-containing and noise-free simulation data pairs according to the noise-containing seismic records obtained in the step 11 and the clean seismic records obtained in the step 12, wherein the size of each data pair is 256 × 256, and 1541 data pairs are obtained;
step 2: producing an actual data set comprising: firstly, pumping unit noise which is not mixed with effective signals is extracted, namely pumping unit signals of first arrival wavefront collected by a detector are superposed on seismic data without pumping unit noise to form actual noise-containing seismic data, actual seismic pairs without pumping unit noise and with pumping unit noise are obtained through the matching mode, and finally 45678 data blocks with the size of 256 × 256 are obtained.
And 3, step 3: the method comprises the steps of constructing a pumping unit noise suppression network based on multiple layers of generators, wherein the network comprises four generators which are sequentially connected, dividing a noise-containing seismic image into data blocks with different sizes and respectively sending the data blocks into the generators of corresponding layers, sequentially fusing extracted high-level semantic features into features extracted by the generator of the next layer from the first layer to the fourth layer, superposing a primarily processed denoising result to the input of the next layer, and finally obtaining a final denoising seismic image, wherein the training process of the network specifically comprises the following steps:
step 31: inputting original noisy seismic data into a first layer generator in an eighth division mode, outputting corresponding eight characteristic graphs after passing through a first decoder, splicing the characteristic graphs into four first characteristic graphs according to wide dimensions, and outputting four blocks of coarse denoising seismic data through the first decoder;
step 32: inputting the original noise-containing seismic image into a second generator in four equal parts, fusing the original noise-containing seismic image with four blocks of coarse de-noising seismic images output by a first layer decoder, and inputting the fused image into a second encoder, outputting four second characteristic graphs by the second encoder, fusing the second characteristic graphs with a first characteristic graph, inputting the fused image into a second decoder, and outputting two blocks of coarse de-noising seismic data by the second decoder;
step 33: halving the original noise-containing seismic image and inputting the halved noise-containing seismic image into a third generator, fusing the halved noise-containing seismic image with the coarse noise-removing seismic data of the two blocks output by the second layer decoder and inputting the halved noise-removing seismic data into a third encoder, outputting a third characteristic diagram of the two blocks by the third encoder, fusing the third characteristic diagram with the second characteristic diagram and inputting the fused third characteristic diagram into a third decoder, and outputting complete coarse noise-removing seismic data by the third decoder;
step 34: fusing the original noise-containing seismic image with the de-noised seismic data output by the third layer and inputting the fused data into a fourth generator, outputting a whole block of fourth characteristic diagram by a fourth encoder, fusing the fourth characteristic diagram with the third characteristic diagram and inputting the fused graph into a fourth decoder, and outputting final fine de-noised seismic data by the fourth decoder;
and step 36: judging whether the set verification iteration times are reached, if so, saving the model, and if not, executing a step 37;
step 37: and judging whether the set total iteration times are reached, if so, ending the training, and otherwise, repeating the step 31 and the step 36.
According to a preferred embodiment, the generator comprises an encoder and a decoder, the encoder extracts the characteristics of the seismic data through convolution, the decoder uses the characteristics to reconstruct the denoised seismic data, and the encoder consists of 14 convolution layers and 6 residual connection; the layers of the decoder are the same as the encoder, except that there are two deconvolution layers in the encoder instead of the convolution layers to generate denoised seismic data.
According to a preferred embodiment, the loss function of the noise suppression network is expressed using a reconstruction error, the mathematical expression being as follows:
wherein G represents true noise-free seismic data,
representing the generated de-noised seismic data.
Compared with the prior art, the invention has the beneficial effects that:
1. the denoising network of the end-to-end multilayer generators is provided, the encoder of each layer of generators extracts seismic data features on different scales, the features extracted by the generators in the previous layer are fused into the feature map extracted by the generators in the next layer, and the denoising performance from coarse to fine is realized by extracting the features of different scales layer by layer.
2. The input of each layer of generator is divided into data blocks with different sizes, so that the network can feel more seismic data areas, the receptive field of the denoising network is improved, and the generators can extract more useful seismic data characteristics.
3. The method provided by the text does not comprise a discriminator, only restrains the denoising result through a reconstruction function, guides a generator to generate a high-quality denoising image, reduces the complexity of model training and shortens the training time while ensuring the denoising performance.
4. The existing denoising method generally divides seismic data into 40 × 40 or 50 × 50 data blocks, while the noise characteristics of the pumping unit are relatively large, and the division of the data blocks which are too small enables an encoder to be incapable of extracting the global characteristics of noise, so that the denoising performance is reduced.
Drawings
FIG. 1 is a flow chart of a method of the present invention for noise suppression in a pumping unit;
FIG. 2 is a block diagram of a noise suppression network of the present invention;
FIG. 3 is a forward velocity model used by the present invention to obtain simulation data;
FIG. 4 is a block diagram of an encoder and decoder of the present invention;
FIG. 5 is a graph comparing experimental results of simulation data according to the present invention;
FIG. 6 is a graph of F-K frequency spectrum comparison of simulation data of the present invention;
FIG. 7 is a graph comparing experimental results of actual data of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the accompanying drawings in conjunction with the following detailed description. It should be understood that the description is intended to be exemplary only, and is not intended to limit the scope of the present invention. Moreover, in the following description, descriptions of well-known structures and techniques are omitted so as to not unnecessarily obscure the concepts of the present invention.
Along with the exploration integration, the economic loss is brought when the oil extraction is stopped, the acquisition and the seismic exploration are inevitably carried out simultaneously in an old oil area, so the acquired seismic data are inevitably interfered by ground noise of an oil pumping unit and the like, and because the noise energy of the oil pumping unit is strong, the frequency band is completely overlapped with the frequency band of an effective signal, and the characteristics expressed on the seismic data are completely different from random noise and surface wave, the two denoising methods based on the prior method are not suitable for denoising with the noise of the oil pumping unit. Aiming at the defects of the prior art and the urgent need of oil field exploitation, the invention provides a seismic data pumping unit noise suppression method research based on a multilayer generator.
The following detailed description is made with reference to the accompanying drawings.
FIG. 1 is a flow chart of a method of the present invention for noise suppression in a pumping unit; the method is based on a plurality of layers of generators, achieves coarse-to-fine noise suppression by fusing the features extracted by the blind generators, divides the input of each layer of generators into data blocks with the blind sizes, effectively enlarges the receptive field of the network, and extracts more useful noise features of the pumping unit, and comprises the following specific steps:
step 1: producing a simulated data set comprising:
the principle of the noise generation of the oil pumping unit is that vibration generated when the oil pumping unit works generates a vibration signal to the ground, and the vibration signal is transmitted to the ground and received by the wave detector. The principle of pumping unit noise is similar to seismic data, so we can use the wave equation to obtain the simulated pumping unit noise.
The existing denoising method usually adopts a simpler horizontal layered model when generating simulation data, so that the synthesized data lacks rich characteristics of field data, and the denoising performance of a neural network is influenced. In order to make the characteristics of the synthetic data closer to the field seismic data and improve the denoising performance, a more complex Marmousi-II P wave velocity model which more conforms to the actual geological condition is selected as a forward model, as shown in FIG. 2.
Step 11: acquiring a simulated noisy seismic record, selecting a part of a Marmousi-II longitudinal wave velocity model as a forward model, wherein the size of the model is 271 multiplied 351, the space step length is h =5m, the time sampling interval is delta t =0.5ms, the total receiving time is 3.6s, the Ricker wavelet is used as an explosive source and an oil pumping unit source and is simultaneously used as an explosive source, and a first-order stress-velocity sound wave equation is adopted for carrying out numerical simulation to obtain 31-cannon seismic records with oil pumping unit noise;
when data is collected in the field, the oil pumping unit in the work area can also vibrate the stratum. According to the working characteristics of the pumping unit, the seismic source S generated by the pumping unit can be known d It can be represented by a fixed frequency point source:
S d ＝A*sin(2πωt) (1)
wherein A is amplitude, and omega is vibration angular frequency of the oil pumping unit.
The first order stress-velocity constant density acoustic wave equation is:
in the formula: p is stress; v x And V z Representing the vibration velocity components of the mass point in the x and z directions respectively; rho is the density of the medium; v is the seismic wave velocity.
In order to better accord with the noise data of the actual oil pumping unit, the invention reduces the number of the vibration sources of the noise of the oil pumping unit and simultaneously attenuates and attenuates the energy transmitted by the noise of the oil pumping unit to a certain degree. This is due to the real pumping noise, and the wave energy is absorbed by the formation as it passes through it, thus there is a loss of energy transmitted to the surface.
Step 12: and (4) setting the seismic source of the oil pumping unit not to vibrate and only the seismic source to vibrate based on the forward modeling in the step (11), and obtaining a clean seismic record.
Step 13: and (3) making a simulation training set, and making noise-containing and noise-free simulation data pairs according to the noise-containing seismic records obtained in the step (11) and the clean seismic records obtained in the step (12), wherein the size of each data pair is 256 × 256, and 1541 data pairs are obtained.
Step 2: producing an actual data set comprising: firstly, pumping unit noise which is not mixed with effective signals is extracted and superposed on seismic data without pumping unit noise to form actual noise-containing seismic data, actual seismic pairs without pumping unit noise and with pumping unit noise are obtained through the matching mode, and finally 45678 pairs with the size of 256 × 256 are obtained in total. The actual data set manufacturing method is adopted, because seismic data without noise of the oil pumping unit does not exist in practice, and the method is a supervised method, the actual seismic data without noise of the oil pumping unit is used as constraint by adopting the method, and the denoising performance and precision are improved. The influence distance of the noise of the pumping unit is in the range of 0-100m, and the energy is continuously weakened along with the increase of the distance. During simulation, the simulation method is researched according to the vibration characteristics of the pumping unit, so that the simulated pumping unit noise is closer to the real pumping unit noise.
And 3, step 3: and constructing a pumping unit noise suppression network based on a plurality of generators, wherein the network comprises four generators which are sequentially connected, as shown in fig. 3, dividing a noise-containing seismic image into data blocks with different sizes and respectively sending the data blocks into the generators of corresponding layers, sequentially fusing the extracted high-level semantic features into the features extracted by the generator of the next layer from the first layer to the fourth layer, superposing the primarily processed denoising result to the input of the next layer, and finally obtaining a final denoising seismic image.
The generator comprises an encoder and a decoder, the structural characteristics of the encoder are shown in figure 4, the encoder extracts the characteristics of the seismic data through convolution, the decoder reconstructs the denoised seismic data by using the characteristics, and the encoder adopts the thought of a residual error network ResNet and consists of 14 convolutional layers and 6 residual error networks. The method is a multilayer generator structure and is an obvious deep neural network, the deep neural network has stronger learning capability and more extracted high-level semantic features, but the deep residual error network has the problem that the performance of the model is reduced along with the increase of the depth of the model during training, namely the degradation of the neural network, the performance of the deep network is not as good as that of the shallow network at some times, and the problems of gradient disappearance and gradient explosion exist as another reason. Therefore, the invention uses the thought of the residual error network for reference, adopts a plurality of jump connections, and has more continuous loss after the residual error network is introduced, and the problem of extremely large fluctuation of the loss value can not occur. The shallow feature is directly transmitted to the deep layer through a residual error network, reusability of the shallow feature is improved, low-dimensional features are fused into high dimensions, and feature learning capability is improved.
The layers of the decoder are the same as the encoder, except that there are two deconvolution layers in the encoder instead of the convolution layers to generate denoised seismic data, while a residual network is also employed.
The training process of the seismic data pumping unit noise suppression network based on the multilayer generator provided by the invention is as follows:
step 31: the method comprises the steps of inputting original noisy seismic data into a first layer generator in an eighth division mode, outputting corresponding eight characteristic graphs after passing through a first decoder, splicing the characteristic graphs into four first characteristic graphs according to wide dimensions, and outputting four blocks of coarse denoising seismic data through the first decoder.
Step 32: and the original noise-containing seismic image is input into a second generator in four equal parts, and is input into a second encoder after being fused with the four blocks of coarse de-noising seismic images output by the first layer decoder, the second encoder outputs four second characteristic maps, the second characteristic map is input into a second decoder after being fused with the first characteristic map, and the second decoder outputs coarse de-noising seismic data of the two blocks.
Step 33: and (3) halving the original noise-containing seismic image and inputting the halved noise-containing seismic image into a third generator, fusing the halved noise-containing seismic image with the coarse noise-removing seismic data of the two blocks output by the second layer decoder and inputting the fused coarse noise-removing seismic data into a third encoder, outputting a third characteristic diagram of the two blocks by the third encoder, fusing the third characteristic diagram with the second characteristic diagram and inputting the fused third characteristic diagram into a third decoder, and outputting complete coarse noise-removing seismic data by the third decoder.
Step 34: and fusing the original noise-containing seismic image with the de-noised seismic data output by the third layer and inputting the fused data into a fourth generator, outputting a whole block of fourth characteristic diagram by a fourth encoder, fusing the fourth characteristic diagram with the third characteristic diagram and inputting the fused image into a fourth decoder, and outputting final fine de-noised seismic data by the fourth decoder.
The loss function of the noise suppression network is expressed using the reconstruction error, and the mathematical expression is as follows:
wherein G represents true noise-free seismic data,
representing the generated de-noised seismic data. Because the network adopts a multi-layer and multi-chip strategy and follows the residual learning principle, the generator at the last layer captures the characteristics of different scales, and therefore, the total loss of the network can be obtained only by calculating the loss function at the fourth layer.
And step 36: and judging whether the set verification iteration number is reached, if so, saving the model, and if not, executing the step 37.
Step 37: and judging whether the set total iteration times are reached, if so, finishing the training, and otherwise, repeating the step 31 and the step 36.
In order to show the noise suppression performance of the pumping unit, the method provided by the invention takes two most common denoising methods in the seismic denoising field, namely denoising convolutional neural network DnCNN and generating antagonistic network GAN as comparison methods. FIG. 5 is a comparison of the results of the present invention simulation data experiments, with FIG. 5 (a) being clean seismic data, FIG. 5 (b) being the effect after DnCN treatment, FIG. 5 (c) being the effect after GAN treatment, and FIG. 5 (d) being the effect after MLGN treatment of the present invention. It can be seen from the processing effect graph that the method of the present invention better suppresses the noise of the pumping unit and better noise residue, while the DnCNN method still has significant pumping unit noise residue, and the GAN method can also better suppress the noise, but to a certain extent, some additional noise will be added. Fig. 5 (e) is seismic data containing pumping unit noise. Fig. 5 (f) shows the residual between fig. 5 (b) and fig. 5 (a), fig. 5 (g) shows the residual between fig. 5 (c) and fig. 5 (a), and fig. 5 (h) shows the residual between fig. 5 (d) and fig. 5 (a). It can be seen from the figure that the method of the present invention has the least loss of effective signals and the best performance.
FIG. 6 is a comparison of F-K spectra. FIG. 6 (a) is the F-K spectrum of clean data, FIG. 6 (b) is the F-K spectrum of seismic data containing pumping unit noise, FIG. 6 (c) is the F-K spectrum of seismic data processed by the DnCNN method, FIG. 6 (d) is the F-K spectrum of seismic data processed by the GAN method, and FIG. 6 (e) is the F-K spectrum of seismic data processed by the method of the present invention. It can be seen from the spectrum comparison that the frequency spectrum processed by the method of the invention is closer to the real frequency spectrum. The method has the advantage of optimal denoising performance.
Fig. 7 is a graph comparing the effects of actual data. Fig. 7 (a) and fig. 7 (e) are real noisy data, fig. 7 (b) and fig. 7 (f) are effect diagrams after DnCNN denoising, it can be seen from the diagrams that there is a significant loss of effective signals, and noise cannot be removed cleanly, and still both noise remains, fig. 7 (c) and fig. 7 (g) are effect diagrams after GAN denoising, and fig. 7 (d) and fig. 7 (h) are effect diagrams after denoising by the method of the present invention. From the processing effect of actual data, it can be seen that neither the DnCNN method nor GAN can suppress noise, and the GAN method may additionally introduce some noise. Only the method of the invention has the best performance of inhibiting the noise of the oil pumping unit.
It should be noted that the above-mentioned embodiments are exemplary, and that those skilled in the art, having benefit of the present disclosure, may devise various arrangements that are within the scope of the present disclosure and that fall within the scope of the invention. It should be understood by those skilled in the art that the present specification and figures are illustrative only and are not limiting upon the claims. The scope of the invention is defined by the claims and their equivalents.
Claims (3)
1. The method for suppressing the noise of the seismic data pumping unit based on the multilayer feature fusion is characterized in that simulation data are obtained through a Marmousi-II model, coarse-to-fine noise suppression is achieved through fusing features extracted by different generators based on the multilayer generators, the input of each layer of generators is divided into seismic data blocks with different sizes, the receptive field of a network is effectively enlarged, and more useful noise features of the pumping unit are extracted, and the method specifically comprises the following steps:
step 1: producing a simulation data set, comprising:
step 11: obtaining a simulated noisy seismic record, selecting a part of a Marmousi-II longitudinal wave velocity model as a forward model, wherein the size of the model is 271 multiplied by 351, the space step length is h =5m, and the time sampling interval is
The total receiving time is 3.6s, the Rake wavelets are used as an explosive source and a pumping unit source, the explosive source and the pumping unit source work simultaneously, numerical simulation is carried out by adopting a first-order stress-velocity sound wave equation, and seismic records with 31 guns and pumping unit noise are obtained in total;
step 12: setting the seismic source of the oil pumping unit not to vibrate and only the seismic source to vibrate based on the forward modeling of the step 11 to obtain a clean seismic record;
step 13: making a simulation training set, and making noise-containing and noise-free simulation data pairs according to the noise-containing seismic records obtained in the step 11 and the clean seismic records obtained in the step 12, wherein the size of the data pairs is 256 × 256, so that 1541 data pairs are obtained;
and 2, step: producing an actual data set comprising: firstly, pumping unit noise without aliasing with effective signals, namely pumping unit signals of first arrival wavefront collected by a detector, are extracted and are superposed on seismic data without pumping unit noise to form actual noisy seismic data, actual seismic pairs without pumping unit noise and with pumping unit noise are obtained through the matching mode, and 45678 data blocks with the size of 256 x 256 are finally obtained;
and step 3: the method comprises the steps of constructing a pumping unit noise suppression network based on multiple layers of generators, wherein the network comprises four generators which are sequentially connected, dividing a noise-containing seismic image into data blocks with different sizes and respectively sending the data blocks into the generators of corresponding layers, sequentially fusing extracted high-level semantic features into features extracted by the generator of the next layer from the first layer to the fourth layer, superposing a primarily processed denoising result to the input of the next layer, and finally obtaining a final denoising seismic image, wherein the training process of the network specifically comprises the following steps:
step 31: inputting original noisy seismic data into a first layer generator in an eighth division mode, outputting corresponding eight characteristic graphs after passing through a first decoder, splicing the characteristic graphs into four first characteristic graphs according to wide dimensions, and outputting four blocks of coarse denoising seismic data through the first decoder;
step 32: inputting the original noise-containing seismic image into a second generator in four equal parts, fusing the original noise-containing seismic image with four blocks of coarse de-noising seismic images output by a first layer decoder, and inputting the fused image into a second encoder, outputting four second characteristic graphs by the second encoder, fusing the second characteristic graphs with a first characteristic graph, inputting the fused image into a second decoder, and outputting two blocks of coarse de-noising seismic data by the second decoder;
step 33: halving and inputting the original noise-containing seismic image into a third generator, fusing the halved noise-containing seismic image with the coarse noise-removing seismic data of the two blocks output by the second layer decoder, and inputting the fused noise-removing seismic data into a third encoder, outputting a third feature map of the two blocks by the third encoder, fusing the third feature map with the second feature map, and inputting the fused feature map into a third decoder, and outputting complete coarse noise-removing seismic data by the third decoder;
step 34: fusing the original noise-containing seismic image with the de-noised seismic data output by the third layer and inputting the fused data into a fourth generator, outputting a whole block of fourth characteristic diagram by a fourth encoder, fusing the fourth characteristic diagram with the third characteristic diagram and inputting the fused graph into a fourth decoder, and outputting final fine de-noised seismic data by the fourth decoder;
and step 36: judging whether the set verification iteration times are reached, if so, saving the model, and if not, executing the step 37;
step 37: and judging whether the set total iteration times are reached, if so, finishing the training, and otherwise, repeating the step 31 and the step 36.
2. The method as claimed in claim 1, wherein the generator comprises an encoder and a decoder, the encoder extracts the features of the seismic data by convolution, the decoder reconstructs the de-noised seismic data by using the features, the encoder is composed of 14 convolution layers and 6 residual connections; the layers of the decoder are the same as the encoder, except that there are two deconvolution layers in the encoder instead of the convolution layers to generate denoised seismic data.
3. The seismic data pumping unit noise suppression method based on multi-layer feature fusion of claim 2, wherein the loss function of the noise suppression network is expressed by using a reconstruction error, and the mathematical expression is as follows:
where G represents true noise-free seismic data,
representing the generated de-noised seismic data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210506583.XA CN115346112A (en) | 2022-05-11 | 2022-05-11 | Seismic data oil pumping unit noise suppression method based on multilayer feature fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210506583.XA CN115346112A (en) | 2022-05-11 | 2022-05-11 | Seismic data oil pumping unit noise suppression method based on multilayer feature fusion |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115346112A true CN115346112A (en) | 2022-11-15 |
Family
ID=83947604
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210506583.XA Pending CN115346112A (en) | 2022-05-11 | 2022-05-11 | Seismic data oil pumping unit noise suppression method based on multilayer feature fusion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115346112A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115577247A (en) * | 2022-12-09 | 2023-01-06 | 中海油田服务股份有限公司 | Seismic noise removing method and device based on stack feedback residual error network |
| CN116736372A (en) * | 2023-06-05 | 2023-09-12 | 成都理工大学 | Seismic interpolation method and system for generating countermeasure network based on spectrum normalization |
-
2022
- 2022-05-11 CN CN202210506583.XA patent/CN115346112A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115577247A (en) * | 2022-12-09 | 2023-01-06 | 中海油田服务股份有限公司 | Seismic noise removing method and device based on stack feedback residual error network |
| CN116736372A (en) * | 2023-06-05 | 2023-09-12 | 成都理工大学 | Seismic interpolation method and system for generating countermeasure network based on spectrum normalization |
| CN116736372B (en) * | 2023-06-05 | 2024-01-26 | 成都理工大学 | Seismic interpolation method and system for generating countermeasure network based on spectrum normalization |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115346112A (en) | Seismic data oil pumping unit noise suppression method based on multilayer feature fusion | |
| CN102998706B (en) | Method and system for attenuating seismic data random noise | |
| US20100161235A1 (en) | Imaging of multishot seismic data | |
| CN111190227B (en) | Low signal-to-noise ratio seismic data denoising method based on residual convolution generation countermeasure model | |
| CN110658557B (en) | Seismic data surface wave suppression method based on generation of countermeasure network | |
| CN108267784A (en) | A kind of seismic signal random noise compression process method | |
| CN112285768B (en) | High-frequency marine acoustic guided wave frequency dispersion analysis device and method | |
| CN107144879A (en) | A kind of seismic wave noise-reduction method combined based on adaptive-filtering with wavelet transformation | |
| CN113935467A (en) | DAS (data acquisition System) borehole exploration data noise suppression method based on iterative multi-scale attention network | |
| CN104077749A (en) | Seismic data denoising method based on contourlet transformation | |
| CN112598593B (en) | Seismic noise suppression method based on non-equilibrium depth expectation block log-likelihood network | |
| CN106873036A (en) | Denoising method based on well-seismic combination | |
| CN111522062A (en) | Underburden amplitude compensation method based on volcanic shielding quantitative analysis | |
| CN113077386A (en) | Seismic data high-resolution processing method based on dictionary learning and sparse representation | |
| CN115905805A (en) | DAS data multi-scale noise reduction method based on global information judgment GAN | |
| CN111948711A (en) | Method and system for extracting low-frequency part of seismic data by using deep learning method | |
| CN112926232B (en) | Seismic low-frequency component recovery method based on layered fusion | |
| Li et al. | Removing abnormal environmental noise in nodal land seismic data using deep learning | |
| CN107703548A (en) | Shallow stratum stratum boundary division methods based on the deposit qualities factor and RL return loss level curve peak valley | |
| Yang et al. | Pyramid residual neural network with attention for seismic data denoising | |
| Ma et al. | A Global and Multi-Scale Denoising Method Based on Generative Adversarial Network for DAS VSP Data | |
| CN107102358B (en) | The method and apparatus of compacting bubble interference based on seismic data | |
| CN112946742B (en) | Method for picking up accurate superposition velocity spectrum | |
| CN112213785B (en) | Seismic data desert noise suppression method based on feature-enhanced denoising 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 |
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 |