Seismic Data Pumping Unit Noise Suppression Method Based on Multi-Layer Feature Fusion

Abstract

The present invention relates to a method for suppressing the noise of a pumping unit in seismic data based on multi-layer feature fusion. A denoising network with an end-to-end multi-layer generator is constructed. The encoder of each layer of the generator extracts seismic data features at different scales, and fuses the features extracted by the previous layer of the generator into the feature map extracted by the next layer of the generator. By extracting features at different scales layer by layer, the denoising performance from coarse to fine is achieved. At the same time, the input of each layer of the generator is divided into data blocks of different sizes, enabling our network to sense more regions of seismic data, increasing the receptive field of the denoising network, and facilitating the generator to extract more useful seismic data features. Thanks to the structure of the network, the seismic data is divided into larger data blocks, which is conducive to the network to extract more global noise features and improve the recognition performance of the pumping unit noise.

Description

Technical Field

[0001] The present invention relates to the field of seismic exploration, and in particular to a method for suppressing the noise of pumping units in seismic data based on multi-layer feature fusion. Background Art

[0002] In actual seismic data acquisition, it is inevitably contaminated by noise. The introduction of noise will greatly reduce the quality and accuracy of seismic data, directly affecting the subsequent analysis, interpretation and application of seismic data. In the field of seismic exploration, the effective separation of signals and noise has always been a research hotspot and difficulty. Seismic noise mainly includes coherent noise and incoherent noise. Incoherent noise has no regular pattern in seismic data, no fixed frequency and apparent velocity, such as random noise. Coherent noise shows more obvious and regular characteristics in seismic records, with relatively fixed dominant frequency and apparent velocity, such as surface waves and multiple waves. This article mainly studies the coherent noise generated by the vibration of pumping units.

[0003] Existing methods for suppressing seismic data noise can be divided into two categories. The first category is traditional classification, including filtering and transform-based methods. The filtering method separates signals and noise by designing appropriate filters, such as fk filters, adaptive time-frequency peak filters, Kalman filters, and particle filters. Considering the obvious feature differences between seismic data and noise in some sparse domains, many successful methods based on sparse transforms have been proposed by researchers, such as: singular value decomposition, principal component analysis, curvelet, wavelet, and shearlet. Traditional methods mainly rely on prior information and require manual parameter adjustment, which is time-consuming and laborious and is not conducive to the efficient processing of massive seismic data. Therefore, there is an urgent need to explore a more intelligent and better seismic denoising method.

[0004] The second category is data-driven. With the success of deep learning methods in image processing, object detection, NLP problems, etc., deep learning methods have gradually been introduced into the field of seismic exploration, mainly used for automatic fault detection of random noise, phase classification, lithology prediction, noise attenuation, etc. Kimiaefar et al. combined artificial neural networks with wavelet packet analysis methods to attenuate seismic random noise. Yu Siwei et al. introduced that deep CNN and transfer learning can not only suppress random noise, but also suppress linear noise and surface wave noise. Xu et al. used a denoising convolutional neural network (DnCNN) to attenuate random noise. Zhao Yuxing et al. improved the original DnCNN in terms of patch size, convolution kernel size, network depth, etc. to make it suitable for low-frequency desert noise suppression. Yang et al. added residual learning and batch normalization methods to DnCNN to reconstruct noise-free seismic data. There is also the introduction of residual learning into the cycle GAN to improve the training efficiency of seismic data denoising. For surface wave noise, the existing research on surface wave noise attenuation is mainly traditional methods, and the research on deep learning methods for surface waves is relatively less, mainly based on deep neural networks, generative adversarial networks, and conditional generative adversarial networks.

[0005] In the existing research on seismic data noise suppression, there is little research on the coherent noise generated by pumping units. In actual seismic exploration, with the integration of exploration, it is inevitable to collect seismic data in the work area of pumping units at the same time. In this case, the collected seismic data will be interfered by the coherent noise of the pumping unit, seriously affecting the quality of seismic data.

[0006] According to the spectral analysis of actual pumping unit data, it is found that the spectral ranges of the pumping unit noise and the effective signal almost completely overlap, which makes it impossible for traditional denoising methods to effectively separate the seismic signal and the pumping unit noise. Furthermore, due to the different generation principles of the coherent noise of the pumping unit and seismic noises such as random noise and surface waves, and their different manifestation characteristics on seismic data, the existing deep learning denoising methods for random noise and surface waves are not applicable to the suppression of pumping unit noise. Summary of the Invention

[0007] Aiming at the deficiencies of the existing technology, the present invention proposes a method for suppressing pumping unit noise in seismic data with multi-layer feature fusion. The method obtains simulated data through the Marmousi-II model. Based on a multi-layer generator, by fusing the features extracted by different generators, it realizes noise suppression from coarse to fine. The input of each layer of the generator is segmented into seismic data blocks of different sizes, effectively expanding the receptive field of the network and extracting more useful pumping unit noise features. The specific steps include:

[0008] Step 1: Make a simulated data set, including:

[0009] Step 11: Obtain the simulated noisy seismic records. Select a part of the Marmousi-II P-wave velocity model as the forward model. The model size is 271×351, the spatial step size is h = 5 m, the time sampling interval is Δt = 0.5 ms, the total receiving duration is 3.6 s. The Ricker wavelet is used as the explosive source and the pumping unit source. The explosive source and the pumping unit source work simultaneously. The first-order stress-velocity acoustic wave equation is used for numerical simulation, and a total of 31 shots of seismic records with pumping unit noise are obtained;

[0010] Step 12: Based on the forward model in Step 11, set the pumping unit source not to vibrate and only the seismic source vibrates to obtain clean seismic records;

[0011] Step 13: Make the simulated training set. According to the noisy seismic records obtained in Step 11 and the clean seismic records obtained in Step 12, make the simulated data pairs of noisy and noise-free. The size of the data pair is 256*256, and a total of 1541 data pairs are obtained;

[0012] Step 2: Make the actual data set, including: First, extract the pumping unit noise that is not aliased with the effective signal, that is, the pumping unit signal in front of the first arrival wave collected by the geophone, and superimpose it on the seismic data without pumping unit noise to form the actual noisy seismic data. Through this matching method, the actual seismic pairs without pumping unit noise and with pumping unit noise are obtained. Finally, a total of 45,678 data blocks of size 256*256 are obtained.

[0013] Step 3: Construct a pumping unit noise suppression network based on a multi-layer generator. The network includes four generators connected in sequence. The noisy seismic images are divided into data blocks of different sizes and sent into the generators of the corresponding layers respectively. From the first layer to the fourth layer, the extracted high-level semantic features are integrated into the features extracted by the next-layer generator in turn, and the preliminarily processed denoising results are superimposed on the input of the next layer. Finally, the final denoised seismic image is obtained. The training process of the network is as follows:

[0014] Step 31: Input the original noisy seismic data divided into eight equal parts into the first-layer generator. After passing through the first decoder, the corresponding eight feature maps are output respectively. The feature maps are spliced along the width dimension into four first feature maps and then output four blocks of roughly denoised seismic data through the first decoder;

[0015] Step 32: Input the original noisy seismic image divided into four equal parts into the second generator, and fuse it with the four blocks of roughly denoised seismic images output by the first-layer decoder and then input it into the second encoder. The second encoder outputs four second feature maps. The second feature maps are fused with the first feature maps and then input into the second decoder, and the second decoder outputs two blocks of roughly denoised seismic data;

[0016] Step 33: Bisect the original noisy seismic image and input it into the third generator. After fusing it with the coarsely denoised seismic data of the two blocks output by the second-layer decoder, input it into the third encoder. The third encoder outputs the third feature maps of the two blocks. After fusing the third feature maps with the second feature maps, input them into the third decoder, and the third decoder outputs the complete coarsely denoised seismic data;

[0017] Step 34: Input the original noisy seismic image after fusing the denoised seismic data output by the third layer into the fourth generator. The fourth encoder outputs a whole block of fourth feature maps. After fusing the fourth feature maps with the third feature maps, input them into the fourth decoder, and the fourth decoder outputs the final finely denoised seismic data;

[0018] Step 36: Determine whether the set number of validation iterations is reached. If it is reached, save the model. If it is not reached, execute Step 37;

[0019] Step 37: Determine whether the total number of iterations is reached. If it is reached, end the training. Otherwise, repeat Step 31 and Step 36.

[0020] According to a preferred embodiment, the generator includes an encoder and a decoder. The encoder extracts the features of the seismic data through convolution, and the decoder uses these features to reconstruct the denoised seismic data. The encoder consists of 14 convolutional layers and 6 residual connections; the layers of the decoder are the same as those of the encoder, except that there are two transposed convolutional layers instead of convolutional layers in the encoder to generate the denoised seismic data.

[0021] According to a preferred embodiment, the loss function of the noise suppression network is represented by the reconstruction error, and the mathematical expression is as follows:

[0022]

[0023] where G represents the true noise-free seismic data, represents the generated denoised seismic data.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] 1. We propose an end-to-end denoising network with a multi-layer generator. The encoder of each layer of the generator extracts the features of the seismic data at different scales, and fuses the features extracted by the previous layer of the generator into the feature maps extracted by the next layer of the generator. By extracting the features at different scales layer by layer, the denoising performance from coarse to fine is achieved.

[0026] 2. The input of each layer of the generator is divided into data blocks of different sizes, enabling our network to sense more regions of the seismic data, improving the receptive field of the denoising network, and facilitating the generator to extract more useful features of the seismic data.

[0027] 3. The method proposed in this paper does not include a discriminator. Instead, it only constrains the denoising result through a reconstruction function to guide the generator to generate high-quality denoised images. While ensuring the denoising performance, it reduces the model training complexity and shortens the training time.

[0028] 4. Existing denoising methods generally divide seismic data into data blocks of 40*40 or 50*50. However, the noise characteristics of pumping units are relatively large. Dividing the data into too small blocks makes the encoder unable to extract the global features of the noise, reducing the denoising performance. Therefore, thanks to the multi-layer network structure of the present invention, the seismic data is divided into data blocks of 256*256, which is conducive to the network extracting more global noise features and improving the recognition performance of pumping unit noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is the method flow chart of the pumping unit noise suppression method of the present invention;

[0030] Figure 2 is the framework schematic diagram of the noise suppression network of the present invention;

[0031] Figure 3 is the forward velocity model used to obtain the simulated data of the present invention;

[0032] Figure 4 The structural diagrams of the encoder and decoder of the present invention;

[0033] Figure 5 is the experimental effect comparison diagram of the simulated data of the present invention;

[0034] Figure 6 is the F-K spectrum comparison diagram of the simulated data of the present invention;

[0035] Figure 7 is the experimental effect comparison diagram of the actual data of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0036] To make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with the specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the descriptions of well-known structures and technologies are omitted to avoid unnecessarily confusing the concepts of the present invention.

[0037] With the integration of exploration, stopping oil production will result in economic losses. In old oil fields, seismic exploration and data acquisition inevitably occur simultaneously. Therefore, the collected seismic data will inevitably be interfered by surface noise such as pumping units. Due to the strong noise energy of pumping units and the complete overlap of their frequency bands with those of effective signals, the characteristics manifested in seismic data are completely different from those of random noise and surface waves. Therefore, the existing denoising methods based on these two types are not applicable to denoising pumping unit noise. In view of the deficiencies of the existing technology and the urgent needs of oilfield exploitation, the present invention proposes a research on a method for suppressing pumping unit noise in seismic data based on a multi-layer generator.

[0038] The following is a detailed description with reference to the accompanying drawings.

[0039] Figure 1 It is the flowchart of the method for suppressing pumping unit noise in the present invention; the method is based on a multi-layer generator, and by fusing the features extracted by different generators, it realizes noise suppression from coarse to fine. The input of each layer of the generator is segmented into data blocks of different sizes, effectively expanding the receptive field of the network and extracting more useful features of pumping unit noise. The specific steps include:

[0040] Step 1: Produce a simulated data set, including:

[0041] The principle of pumping unit noise generation is the vibration generated when the pumping unit works, which generates vibration signals on the ground and transmits the vibration signals to the ground, where they are received by the geophones. The principle of pumping unit noise is similar to that of seismic data. Therefore, we can use the wave equation to obtain simulated pumping unit noise.

[0042] When generating simulated data, existing denoising methods usually adopt a relatively simple horizontal layered model, which makes the synthetic data lack the rich features of field data, thus affecting the denoising performance of the neural network. To make the features of the synthetic data closer to field seismic data and improve the denoising performance, a more complex and more realistic geological situation Marmousi-II P-wave velocity model is selected as the forward model, as Figure 2 shown.

[0043] Step 11: Obtain simulated noisy seismic records. Select a part of the Marmousi-II P-wave velocity model as the forward model, with a model size of 271×351, a spatial step size of h = 5m, a time sampling interval of Δt = 0.5ms, a total receiving duration of 3.6s, and a Ricker wavelet as the explosive source and the pumping unit source as the source at the same time. Use the first-order stress-velocity acoustic wave equation for numerical simulation, and a total of 31 shots of seismic records with pumping unit noise are obtained;

[0044] During field data acquisition, the pumping units in the work area will also cause the formation to vibrate. According to the working characteristics of the pumping unit, it can be known that the source S generated by the pumping unit dIt can be represented by a point source with a fixed frequency:

[0045] S d = A*sin(2πωt) (1)

[0046] Where A is the amplitude and ω is the vibration angular frequency of the pumping unit.

[0047] The first-order stress-velocity constant density acoustic wave equation is:

[0048]

[0049]

[0050] Where: P is the stress; V x and V z respectively represent the particle vibration velocity components in the x and z directions; ρ is the medium density; v is the seismic wave velocity.

[0051] In order to be more in line with the actual pumping unit noise data, the present invention reduces the number of sources of the pumping unit noise. At the same time, the energy of the propagation of the pumping unit noise is attenuated to a certain extent. This is because for the real pumping unit noise, when passing through the formation, the energy of its wave will be absorbed by the formation, so that the energy reaching the ground will be lost.

[0052] Step 12: Based on the forward model in Step 11, set the pumping unit source not to vibrate and only the seismic source vibrates to obtain a clean seismic record.

[0053] Step 13: Make a simulation training set. According to the noisy seismic record obtained in Step 11 and the clean seismic record obtained in Step 12, make a pair of simulated data with noise and without noise. The size of the data pair is 256*256, and a total of 1541 data pairs are obtained.

[0054] Step 2: Make an actual data set, including: First, extract the pumping unit noise that is not mixed with the effective signal, and superimpose it on the seismic data without the pumping unit noise to form the actual noisy seismic data. Through this matching method, an actual seismic pair without the pumping unit noise and with the pumping unit noise is obtained. Finally, a total of 45678 image pairs with a size of 256*256 are obtained. The method of making this actual data set is adopted because in reality, there is no seismic data without the pumping unit noise, and the method of the present invention is a supervised method. Therefore, by adopting this method, there is an actual seismic data without the pumping unit noise as a constraint, which improves the denoising performance and accuracy. The influence distance of the pumping unit noise is in the range of 0-100m, and as the distance increases, the energy is continuously weakened. During the simulation, the present invention studies the simulation method according to the vibration characteristics of the pumping unit, so that the simulated pumping unit noise is closer to the real pumping unit noise.

[0055] Step 3: Construct a pumping unit noise suppression network based on a multi-layer generator. The network includes four generators connected in sequence. As shown in Figure 3 , the noisy seismic image is divided into data blocks of different sizes and fed into the generators of the corresponding layers respectively. From the first layer to the fourth layer, the high-level semantic features extracted are successively integrated into the features extracted by the next-layer generator, and the preliminarily processed denoising results are superimposed on the input of the next layer. Finally, the final denoised seismic image is obtained.

[0056] The generator includes an encoder and a decoder, and its structural features are as shown in Figure 4 . The encoder extracts the features of seismic data through convolution, and the decoder uses these features to reconstruct the denoised seismic data. The encoder adopts the idea of the residual network ResNets and consists of 14 convolutional layers and 6 residual networks. Since the method of the present invention is a multi-layer generator structure, which is an obvious deep neural network, the deep neural network has stronger learning ability and extracts more high-level semantic features. However, there is a problem with the deep residual network that as the depth of the model increases, the performance of the model will decline during training, which is the degradation of the neural network. Sometimes the performance of the deep network is inferior to that of the shallow network. Another reason is that there will be problems of gradient disappearance and gradient explosion. Therefore, the present invention draws on the idea of the residual network and adopts multiple skip connections. After introducing the residual network, the loss is more continuous and there will be no problem of particularly large fluctuations in the loss value. Through the residual network, the shallow features are directly transmitted to the deep layer, increasing the reusability of the shallow features, fusing the low-dimensional features into the high-dimensional ones, and improving the feature learning ability.

[0057] The layers of the decoder are the same as those of the encoder, except that there are two deconvolutional layers in the encoder instead of convolutional layers to generate the denoised seismic data, and the residual network is also adopted.

[0058] The training process of the seismic data pumping unit noise suppression network based on a multi-layer generator proposed by the present invention is specifically as follows:

[0059] Step 31: Divide the original noisy seismic data into eight equal parts and input them into the first-layer generator. After passing through the first decoder, the corresponding eight feature maps are output respectively. The feature maps are spliced according to the width dimension into four first feature maps and then output four blocks of roughly denoised seismic data through the first decoder.

[0060] Step 32: Divide the original noisy seismic image into four equal parts and input it into the second generator, and fuse it with the four blocks of roughly denoised seismic images output by the first-layer decoder and then input it into the second encoder. The second encoder outputs four second feature maps. The second feature maps are fused with the first feature maps and then input into the second decoder, and the second decoder outputs two blocks of roughly denoised seismic data.

[0061] Step 33: Bisect the original noisy seismic image and input it into the third generator. After fusing it with the coarsely denoised seismic data of the two blocks output by the second-layer decoder, input it into the third encoder. The third encoder outputs the third feature maps of the two blocks. After fusing the third feature maps with the second feature maps, input them into the third decoder, and the third decoder outputs the complete coarsely denoised seismic data.

[0062] Step 34: Input the original noisy seismic image after fusing the denoised seismic data output by the third layer into the fourth generator. The fourth encoder outputs a whole block of fourth feature maps. After fusing the fourth feature maps with the third feature maps, input them into the fourth decoder, and the fourth decoder outputs the final finely denoised seismic data.

[0063] The loss function of the noise suppression network is represented by the reconstruction error, and the mathematical expression is as follows:

[0064]

[0065] Among them, G represents the true noise-free seismic data, represents the generated denoised seismic data. Since our network adopts a multi-layer and multi-slice strategy and follows the principle of residual learning, the generator of the last layer captures features of different scales. Therefore, we only need to calculate the loss function of the fourth layer to obtain the total loss of the network.

[0066] Step 36: Determine whether the set validation iteration times are reached. If so, save the model. If not, execute Step 37.

[0067] Step 37: Determine whether the set total iteration times are reached. If so, end the training. Otherwise, repeat Step 31 and Step 36.

[0068] To illustrate the performance of the method proposed in the present invention in suppressing the noise of pumping units, two of the most common denoising methods in the field of seismic denoising, the denoising convolutional neural network DnCNN and the generative adversarial network GAN, are used as comparison methods. Figure 5 is a comparison chart of the experimental results of the simulated data of the present invention. Figure 5 (a) is the clean seismic data. Figure 5 (b) is the effect after being processed by DnCNN. Figure 5 (c) is the effect after being processed by GAN. Figure 5 (d) is the effect after being processed by the MLGN of the present invention. It can be seen from the processing effect diagrams that the method of the present invention better suppresses the noise of the pumping unit and has less noise residue. While the DnCNN method still has obvious pumping unit noise residue, although the GAN method can also better suppress the noise, to a certain extent, it will additionally increase some noise. Figure 5 (e) is the seismic data containing the noise of the pumping unit. Figure 5 (f) isFigure 5 (b) and Figure 5 the residual result of (a), Figure 5 (g) is Figure 5 (c) and Figure 5 the residual result of (a), Figure 5 (h) is Figure 5 (d) and Figure 5 the residual result of (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.

[0069] Figure 6 is the F-K spectrum comparison. Figure 6 (a) is the F-K spectrum of clean data, Figure 6 (b) is the F-K spectrum of seismic data containing pumping unit noise, Figure 6 (c) is the F-K spectrum of seismic data processed by the DnCNN method, Figure 6 (d) is the F-K spectrum of seismic data processed by the GAN method, Figure 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 spectrum processed by the method of the present invention is closer to the real spectrum. It shows that the denoising performance of the method of the present invention is optimal.

[0070] Figure 7 is the effect comparison diagram of actual data. Figure 7 (a) and Figure 7 (e) are real noisy data, Figure 7 (b) and Figure 7 (f) are the effect diagrams after denoising by DnCNN. It can be clearly seen from the figure that there is an obvious loss of effective signals, and the noise cannot be removed cleanly, and there is still noise residue. 7(c) and Figure 7 (g) are the effect diagrams after denoising by GAN, Figure 7 (d) and Figure 7 (h) are the effect diagrams after denoising by the method of the present invention. It can be seen from the processing effect of actual data that neither the DnCNN method nor the GAN can suppress noise, and the GAN method will additionally introduce some noise. Only the method of the present invention has the best performance in suppressing pumping unit noise.

[0071] It should be noted that the above specific embodiments are exemplary. Those skilled in the art can come up with various solutions inspired by the disclosed content of the present invention, and these solutions also belong to the disclosed scope of the present invention and fall within the protection scope of the present invention. Those skilled in the art should understand that the description and drawings of the present invention are illustrative and do not constitute a limitation on the claims. The protection scope of the present invention is defined by the claims and their equivalents.

Claims

1. A method for suppressing the noise of a pumping unit in seismic data based on multi-layer feature fusion, characterized in that, The method obtains simulated data through the Marmousi-II model. Based on a multi-layer generator, by fusing the features extracted by different generators, it realizes noise suppression from coarse to fine. The input of each layer of the generator is segmented into seismic data blocks of different sizes, effectively expanding the receptive field of the network and extracting more useful pumping unit noise features. The specific steps are as follows: Step 1: Make a simulated data set, including: Step 11: Obtain simulated noisy seismic records. Select a part of the Marmousi-II P-wave velocity model as the forward model. The model size is 271×351, the spatial step size is h = 5 m, and the time sampling interval is , the total receiving duration is 3.6 s. The Ricker wavelet is used as the explosive source and the pumping unit source. The explosive source and the pumping unit source work simultaneously. The first-order stress-velocity acoustic wave equation is used for numerical simulation. A total of 31 shots of seismic records with pumping unit noise are obtained; Step 12: Based on the forward model in Step 11, set the pumping unit source not to vibrate and only the seismic source vibrates to obtain a clean seismic record; Step 13: Make a simulated training set. According to the noisy seismic record obtained in Step 11 and the clean seismic record obtained in Step 12, make pairs of noisy and noise-free simulated data. The size of the data pair is 256*256, and a total of 1541 data pairs are obtained; Step 2: Make an actual data set, including: First, extract the pumping unit noise that is not mixed with the effective signal, that is, the pumping unit signal in front of the first arrival wave collected by the geophone, and superimpose it on the seismic data without pumping unit noise to form the actual noisy seismic data. Through this matching method, pairs of actual seismic data without pumping unit noise and with pumping unit noise are obtained. Finally, a total of 45,678 data blocks of size 256*256 are obtained; Step 3: Construct a pumping unit noise suppression network based on a multi-layer generator. The network includes four generators connected in sequence. The noisy seismic image is divided into data blocks of different sizes and sent into the corresponding layer of the generator respectively. From the first layer to the fourth layer, the advanced semantic features extracted are integrated into the features extracted by the next layer of the generator in turn, and the preliminarily processed denoising result is superimposed on the input of the next layer. Finally, the final denoised seismic image is obtained. The training process of the network is as follows: Step 31: Divide the original noisy seismic data into eight equal parts and input it into the first layer of the generator. After passing through the first decoder, eight corresponding feature maps are output respectively. The feature maps are spliced according to the width dimension into four first feature maps and then output four blocks of coarsely denoised seismic data through the first decoder; Step 32: Divide the original noisy seismic image into four equal parts and input it into the second generator, and fuse it with the four blocks of coarsely denoised seismic images output by the first layer decoder and then input it into the second encoder. The second encoder outputs four second feature maps. The second feature maps are fused with the first feature maps and then input into the second decoder. The second decoder outputs two blocks of coarsely denoised seismic data; Step 33: Divide the original noisy seismic image into two equal parts and input it into the third generator, and fuse it with the two blocks of coarsely denoised seismic data output by the second layer decoder and then input it into the third encoder. The third encoder outputs two blocks of third feature maps. The third feature maps are fused with the second feature maps and then input into the third decoder. The third decoder outputs the complete coarsely denoised seismic data; Step 34: Input the original noisy seismic image after fusing the denoised seismic data output from the third layer into the fourth generator. The fourth encoder outputs a whole block of fourth feature maps. After the fourth feature maps are fused with the third feature maps, they are input into the fourth decoder, and the fourth decoder outputs the final precisely denoised seismic data; Step 36: Determine whether the set verification iteration times are reached. If so, save the model. If not, execute Step 37; Step 37: Determine whether the total set iteration times are reached. If so, end the training. Otherwise, repeat Step 31 and Step 36.

2. The method for suppressing the noise of a pumping unit in seismic data based on multi-layer feature fusion according to claim 1, wherein, The generator includes an encoder and a decoder. The encoder extracts the features of the seismic data through convolution, and the decoder uses these features to reconstruct the denoised seismic data. The encoder consists of 14 convolutional layers and 6 residual connections; the layers of the decoder are the same as those of the encoder, except that there are two transposed convolutional layers in the encoder instead of convolutional layers to generate the denoised seismic data.

3. The method for suppressing the noise of a pumping unit in seismic data based on multi-layer feature fusion according to claim 2, characterized in that, The loss function of the noise suppression network is represented by the reconstruction error, and the mathematical expression is as follows: where G represents the true noise-free seismic data, represents the generated denoised seismic data.

Images