DAS-VSP data background noise suppression method based on deep learning

Abstract

The application discloses a DAS-VSP data background noise suppression method based on deep learning, four random inactivation layers, residual units, global context modules and attention mechanism modules are added in the U-shaped network. Random inactivation prevents overfitting and improves the generalization ability of the model; in order to improve the training efficiency and convergence speed of the network, the output of the network is changed from denoising data to residual units; the global context module is introduced in the middle of the network, which can not only pay attention to local information, but also extract global context information; the attention mechanism module is added at the end of the network, which can not only capture the key features of the seismic signal, but also extract complex noise information. The GC-AB-Unet network of the application keeps a good balance between removing background noise and retaining effective signals, avoids the disadvantages of traditional methods, improves the defects of existing deep learning methods, and greatly improves the signal-to-noise ratio of seismic data.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of seismic exploration data denoising, and particularly relates to a DAS-VSP data background noise suppression method based on deep learning. BACKGROUND

[0002] Seismic exploration is one of important means for oil and gas resource development, and natural energy is explored and developed and produced mainly through oil and gas geophysical exploration technology taking seismic exploration as the main part. Generally, seismic exploration adopts artificial seismic source to excite elastic waves, and the vibration of the earth is detected at different positions along the survey line by using seismic instruments, and the obtained data is recorded in the form of numbers. Due to the interference of various noises, the effective seismic signal phase axis on the seismic record is often difficult to distinguish, which affects the subsequent seismic data processing and interpretation. Therefore, improving the signal-to-noise ratio of the seismic signal is an important content in the work of seismic data processing.

[0003] With the increasing universality of these oil wells, new exploration and production technologies are urgently needed to provide high-precision reservoir description around the oil wells in order to guide further development of oil and gas fields. Traditional seismic detectors have low spatial sampling density, high cost and difficult instrument layout, so it is extremely important to find a new sensing method to promote the detection process. Compared with traditional detectors, the distributed optical fiber acoustic sensing system (Distributed Acoustic Sensing, DAS) has the advantages of full coverage, wide screen band, high sensitivity, high precision, high density, resistance to high temperature and high pressure, and anti-electromagnetic interference. The optical fiber has stability, and it uses short light pulses to realize high spatial resolution monitoring. However, distributed optical fibers also have their drawbacks, and they will be greatly disturbed by noise during collection, and the collected data has low signal-to-noise ratio. DAS technology has been successfully used in vertical seismic (VSP) data acquisition, and distributed optical fiber vertical seismic (DAS-VSP) data contains various noises, such as random noise, instrument interference, coupling noise, chessboard noise and horizontal noise. Therefore, improving the signal-to-noise ratio of DAS-VSP seismic data and removing various noises are a very challenging task.

[0004] Traditional denoising methods often have certain limitations, such as low denoising efficiency and serious damage to effective signals. Therefore, designing a method that can both denoise efficiently and protect effective signals has always been an important research topic in denoising processing. Deep learning method has been successfully applied in seismic data denoising, and the present application uses deep learning method to suppress background noise containing random noise and instrument noise for distributed optical fiber vertical seismic (DAS-VSP) data, providing basic data for subsequent seismic data processing. SUMMARY

[0005] The technical problems to be solved by the present application are to avoid the limitations of traditional methods, improve existing deep learning methods, and then propose a distributed optical fiber vertical seismic (DAS-VSP) data background noise suppression method based on deep learning, extract the noise in the actual data and inject it into the training data, then propose a U-shaped network (U-net with Global Context Block and Attention Block, GC-AB-Unet) with a global context module and an attention mechanism, that is, based on the U-shaped network (U-net), a random dropout layer is added to avoid overfitting, a global context module (GC-Block) is added to extract global context information, and an attention mechanism (Attention Block) is added to the network to capture key features in seismic data, the network output is changed from denoised data to residual data to further improve the computational efficiency of the network, and finally the proposed method is applied to synthetic seismic data and actual DAS-VSP seismic data, and the results of the synthetic data and the actual data both reflect the good denoising performance of the GC-AB-Unet network.

[0006] The present application adopts the following technical solutions:

[0007] A distributed optical fiber vertical seismic (DAS-VSP) data background noise suppression method based on deep learning, comprising the following steps:

[0008] S1, obtain synthetic seismic data as clean training data set s by three different ways, extract noise n in actual DAS-VSP seismic data and randomly inject it into the clean training data to construct the corresponding noisy training data set y;

[0009] S2, based on the U-shaped network (U-net), change the original activation function from the linear rectifier function (ReLU) to the leaky linear rectifier function (LeakyReLU), so as to retain the negative value in the amplitude of the seismic signal;

[0010] S3, a random dropout layer is added before each downsampling layer, that is, randomly discard parameters that are not important to the network to prevent overfitting;

[0011] S4, introduce a global context module (GC-Block) to connect the contraction network and the expansion network, which is used to extract global context information and improve the feature extraction capability of the network;

[0012] S5, an attention mechanism module is added before the network output to capture the key features of the complex geological structure in the seismic data;

[0013] S6, the network output is changed from the denoised data to residual data, which greatly improves the calculation efficiency of the network;

[0014] S7, the trained network is applied to the synthetic seismic data and the actual DAS-VSP seismic data, and the denoising effect is analyzed through the quantitative index SNR and the visual effect, so as to reflect the denoising performance of the network.

[0015] Specifically, in step S1, the clean training data set is obtained by sliding window method from the complete seismic record, and the window size is 240*240. The data is extracted from left to right and from top to bottom, and the sliding window step is 120. In order to ensure the diversity of the training data, the synthetic VSP data is obtained by three different ways (SEG public data: http: / / s3.amazonaws.com / open.source.geoscience / open_data / SModels / SModels.html, reflectivity method and finite difference method in time domain), and the noise is directly extracted from the actual DAS-VSP seismic data. The extracted noise is randomly injected into the clean VSP data set to obtain the corresponding noise VSP data set.

[0016] Specifically, in step S2, the activation function with leaky rectified linear function (LeakyReLU) is defined as follows:

[0017]

[0018] Wherein, x is the input seismic signal, and a is a constant multiplied by the signal when it is negative. The results of multiple experiments show that when a = 0.3, the network denoising performance is the best.

[0019] Further, the shrinkage network contains four down-sampling layers, and the expansion network contains four up-sampling layers, which are realized by maximum pooling and bilinear interpolation respectively.

[0020] Specifically, in step S3, each random inactivation (dropout) layer randomly retains 90% of the hyperparameters, and discards the unimportant parameters in the network to prevent overfitting.

[0021] Specifically, in step S4, a global context block (GC-Block) is introduced between the 7th layer and the 8th layer of the network to enhance the ability of the network to obtain global context information. First, the input data is processed by 1x1 convolution and Softmax function, then the obtained result is multiplied with the feature map initially input into the module, and then 1x1 convolution, LayerNorm, ReLU and 1x1 convolution are used in sequence for feature transformation. LayerNorm represents the normalization processing of the input seismic data in the channel direction. Finally, the feature fusion is performed by adding the initially input feature map and the feature-transformed feature map, that is, the global context features are aggregated to the features at each position.

[0022] Specifically, in step S5, the attention mechanism uses cascading operation to fuse the noisy seismic data input into the network with the feature map output by the last layer of the network. First, feature compression is performed, that is, the channel number is converted to 1 by using 1x1 convolution; second, the obtained data is fused with the network input data through feature channels; third, the obtained features are converted into nonlinear features by using Tanh layer, and then normalized; finally, 1x1 convolution operation is performed on the data obtained in the third step, and then the data is multiplied with the data output in the first step, so as to enhance the ability of the network to capture key information in the seismic signal.

[0023] Specifically, in step S6, the network output is residual data, and the de-noised seismic data is obtained by subtracting the network output from the network input.

[0024] Further, the network is established to remove the background noise in the noisy data and retain the effective signal, and the de-noising process can be specifically described as follows:

[0025]

[0026] wherein y is the seismic signal contaminated by noise, denotes the signal after de-noising by the network, G represents the proposed U-shaped network with global context block and attention mechanism (GC-AB-Unet), and θ denotes the hyperparameters in the network, including weights and biases, which are adjusted by minimizing the loss function. The loss function is expressed as follows:

[0027]

[0028] wherein {y i , s i} represents N pairs of noisy and corresponding clean training data sets, both from synthetic seismic data. iFor clean seismic signals, G represents the proposed U-shaped network with global context module and attention mechanism (GC-AB-Unet), and theta refers to the hyperparameters in the network, including weights and biases. N refers to the batch size, i.e. the number of input data required to calculate the loss value once. The network training adopts the adaptive gradient algorithm (Adam) for gradient update.

[0029] Specifically, in step S7, in order to evaluate the denoising performance of the proposed network, the signal-to-noise ratio (SNR) is selected as a quantitative indicator to measure the denoising performance of the network, and the visual effect is presented through the denoised data and the difference data (the difference between the noisy data and the denoised data). The time domain graph and the frequency domain graph of the single channel signal are also used to analyze the denoising performance. The two are combined to analyze whether the network can remove the background noise while retaining the seismic signal well.

[0030] Compared with the existing deep learning method, the present application has at least the following improvements and achievements:

[0031] The present application proposes a DAS-VSP data background noise suppression method based on deep learning, and proposes a U-shaped network with global context module and attention mechanism (GC-AB-Unet), which is used for DAS-VSP data background noise suppression. Specifically, four dropout layers, residual units, global context modules (GC-Block) and attention mechanism modules (Attention Block) are added to the U-net. The dropout layer is designed to prevent overfitting and improve the generalization ability of the model; in order to improve the training efficiency and convergence speed of the network, the output of the network is modified to a residual unit; the global context module is introduced in the middle of the network, which can not only pay attention to local information, but also extract global context information; the attention mechanism module is added at the end of the network, which can not only capture the key features of the seismic signal, but also extract complex noise information. The denoising results of synthetic seismic data and actual DAS-VSP seismic data show that the GC-AB-Unet network maintains a good balance between removing background noise and retaining effective signals, avoids the drawbacks of traditional methods, improves the defects of existing deep learning methods, and greatly improves the signal-to-noise ratio of seismic data.

[0032] In summary, the present application can effectively and quickly suppress the background noise of DAS-VSP seismic data, and well retain the seismic signal, so that the extracted DAS-VSP seismic signal structure is clearer, the lateral continuity is better, and the signal-to-noise ratio is higher.

[0033] The technical solutions of the present application will be further described in detail below with the help of the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 Figure 1 is a network structure diagram of the present application;

[0035] Figure 2 Figure 2 is a global context (GC-Block) network structure diagram of the present application;

[0036] Figure 3(a) is the clean data of the first synthetic seismic test data of the present application;

[0037] Figure 3(b) is the noisy data of the first synthetic seismic test data of the present application;

[0038] Figure 3(c) is the data after denoising of the first synthetic seismic test data of the present application by the U-shaped network with global context module and attention mechanism (GC-AB-Unet);

[0039] Figure 3(d) is the difference data between the noisy data of the first synthetic seismic test data of the present application Figure 3(b) and Figure 3(c);

[0040] Figure 4(a) is the clean data of the second synthetic seismic test data of the present application;

[0041] Figure 4(b) is the noisy data of the second synthetic seismic test data of the present application;

[0042] Figure 4(c) is the data after denoising of the second synthetic seismic test data of the present application by the U-shaped network with global context module and attention mechanism (GC-AB-Unet);

[0043] Figure 4(d) is the difference data between the noisy data of the second synthetic seismic test data of the present application Figure 4(b) and Figure 4(c);

[0044] Figure 5(a) is the noisy data of the first actual DAS-VSP seismic data of the present application;

[0045] Figure 5(b) is the data after denoising of the first actual DAS-VSP seismic data of the present application by the U-shaped network with global context module and attention mechanism (GC-AB-Unet);

[0046] Figure 5(c) is the difference data between the noisy data of the first actual DAS-VSP seismic data of the present application Figure 5(a) and Figure 5(b);

[0047] Figure 5(d) is the comparison of time domain graph before denoising (black thin line) and after denoising (black thick line) of the 109th data of the first actual DAS-VSP seismic data of the present application;

[0048] Figure 5(e) is a comparison of the frequency domain plot of the 167th trace of the first actual DAS-VSP seismic data before denoising (thin black line) and after denoising (thick black line) of the present application.

[0049] Figure 6(a) is the noisy data of the second actual DAS-VSP seismic data of the present application.

[0050] Figure 6(b) is the denoised data of the second actual DAS-VSP seismic data of the present application by the U-shaped network with global context module and attention mechanism (GC-AB-Unet).

[0051] Figure 6(c) is the difference data between the noisy data of the second actual DAS-VSP seismic data of Figure 6(a) and the denoised data of Figure 6(b) of the present application.

[0052] Figure 6(d) is a comparison of the time domain plot of the 167th trace of the second actual DAS-VSP seismic data before denoising (thin black line) and after denoising (thick black line) of the present application.

[0053] Figure 6(e) is a comparison of the frequency domain plot of the 167th trace of the second actual DAS-VSP seismic data before denoising (thin black line) and after denoising (thick black line) of the present application.

[0054] Figure 7(a) is the noisy data of the third actual DAS-VSP seismic data of the present application.

[0055] Figure 7(b) is the denoised data of the third actual DAS-VSP seismic data of the present application by the U-shaped network with global context module and attention mechanism (GC-AB-Unet).

[0056] Figure 7(c) is the difference data between the noisy data of the third actual DAS-VSP seismic data of Figure 7(a) and the denoised data of Figure 7(b) of the present application.

[0057] Figure 7(d) is a comparison of the time domain plot of the 99th trace of the third actual DAS-VSP seismic data before denoising (thin black line) and after denoising (thick black line) of the present application.

[0058] Figure 7(e) is a comparison of the frequency domain plot of the 167th trace of the third actual DAS-VSP seismic data before denoising (thin black line) and after denoising (thick black line) of the present application. DETAILED DESCRIPTION

[0059] The application provides a DAS-VSP data background noise suppression method based on deep learning, proposes a U-shaped network (GC-AB-Unet) with a global context module and an attention mechanism, and applies the U-shaped network to DAS-VSP seismic data background noise suppression. Four random dropout layers, residual units, a global context module (GC-Block) and an attention mechanism module (Attention Block) are added to the U-net. First, the random dropout layer is used to prevent overfitting and improve the generalization ability of the model; second, in order to improve the training efficiency and convergence speed of the network, the output of the network is changed from denoised data to a residual unit; in addition, the global context module is introduced in the middle of the network, which can not only pay attention to local information, but also extract global context information; finally, the attention mechanism module is added at the end of the network, which can not only capture key features in the seismic signal, but also extract complex noise information. The denoising results of the synthetic seismic data and the actual DAS-VSP seismic data show that the GC-AB-Unet network can keep a good balance between removing background noise and preserving effective signals, avoid the disadvantages of traditional methods, improve the defects of existing deep learning methods, and greatly improve the signal-to-noise ratio of seismic data.

[0060] Referring to Figure 1 and Figure 2 The application provides a DAS-VSP data background noise suppression method based on deep learning, which comprises the following steps:

[0061] S1, obtain synthetic seismic data as clean training data set s by three different ways, extract noise n in actual DAS-VSP seismic data and randomly inject it into the clean training data to construct the corresponding noisy training data set y, the method is as follows:

[0062] The clean training data set contains 15000 synthetic seismic data with a size of 240x240, which are obtained by cutting data from complete seismic records by the sliding window method. In order to ensure the diversity of the training data, the synthetic VSP data is obtained by three different ways.

[0063] Among them, 8035 synthetic VSP data come from the website (http: / / s3.amazonaws.com / open.source.geoscience / open_data / SModels / SModels.html). Then, 1965 synthetic VSP data are synthesized by using the reflectivity method, which contains 480 channels, and each channel is composed of 2400 samples. The interval of the seismic detector is set to 5 meters, the main frequency is randomly changed in the range of 10-60 Hz, and the time domain sampling interval is 0.001 s. In addition, 5000 synthetic VSP data are simulated in different layered medium models based on the two-dimensional acoustic wave equation using the Finite Difference Time Domain (FDTD) method. The source is a Ricker wavelet with a main frequency of 25 Hz, the sampling interval in the time domain is 0.001 s, and the apparent velocity varies in the range of 1500-2900 m / s.

[0064] Noise is directly extracted from actual DAS-VSP seismic data, and the extracted part contains almost no valid signal, mainly containing a large amount of random noise and optical system noise. At the same time, the extracted noise is randomly injected into the clean VSP data set to obtain the corresponding noise VSP data set.

[0065] S2, on the basis of U-net, the original activation function is changed from linear rectifier function (ReLU) to leaky rectifier function (LeakyReLU), so as to retain the negative value in the amplitude of the seismic signal, wherein through multiple experiments, the parameter α of the activation function leaky rectifier function (LeakyReLU) is set to 0.3, at this time the network can achieve the best denoising performance.

[0066] Further, the network depth is 15 layers, the number of convolution kernels of the original convolution layer is 32, which is doubled after down-sampling operation, and is halved after up-sampling operation. The pooling window size, the up-sampling factor and the step size are all set to 2.

[0067] S3, a random inactivation (dropout) layer is added before each down-sampling layer, that is, the parameters unimportant to the network are randomly discarded to prevent overfitting, that is, each random inactivation (dropout) layer randomly retains 90% of the hyperparameters.

[0068] S4, a global context module (GC-Block) is introduced to connect the shrinkage network and the expansion network to extract global context information and improve the feature extraction capability of the network, and the method is as follows:

[0069] A global context block (GC-Block) is introduced between the 7th layer and the 8th layer of the network to enhance the ability of the network to obtain global context information. First, the input data is processed by 1x1 convolution and Softmax function, then the obtained result is multiplied with the feature map initially input into the module, and then 1x1 convolution, LayerNorm, ReLU and 1x1 convolution are used in sequence for feature transformation. LayerNorm represents the normalization processing of the input seismic data in the channel direction. Finally, the feature fusion is performed by adding the initially input feature map and the feature-transformed feature map, that is, the global context features are aggregated to the features at each position.

[0070] S5, an attention mechanism block (AB-Block) is added before the network output to capture key features in the seismic data, which is implemented in the following way:

[0071] First, feature compression is performed by using 1x1 convolution to convert the channel number to 1; second, the obtained data is fused with the network input data through the feature channel; third, the obtained features are converted to nonlinear features by Tanh layer, and then normalized; finally, 1x1 convolution is performed on the data obtained in the third step, and then multiplied with the data output in the first step, so as to enhance the ability of the network to capture key information in the seismic signal.

[0072] S6, the network output is changed from the denoised data to the residual data, which greatly improves the calculation efficiency of the network, that is, the network output is residual data, and the denoised seismic data is obtained by subtracting the network output from the network input. Further, in the experiment, the number N of input data required for calculating the loss value in the loss function is set to 32. The network training adopts adaptive gradient algorithm (Adam) for gradient update. In addition, the network is trained by NVIDIA GeForce RTX 2080Ti GPU, and the training time is about 1.28 hours.

[0073] S7, the trained network is applied to synthetic seismic data and actual DAS-VSP seismic data, and the denoising effect is analyzed by quantitative indicators and visual effects, so as to reflect the denoising performance of the network, and the specific method is as follows:

[0074] Both synthetic data are from the website (http: / / s3.amazonaws.com / open.source.geoscience / open_data / SModels / SModels.html). The SNR is selected as the quantitative indicator to measure the denoising performance of the network, and the visualization effect is presented by the denoised data and the difference data (the difference between the noisy data and the denoised data). The time domain graph and the frequency domain graph of the single channel signal are also used to analyze the denoising performance. The combination of the two can analyze whether the network can remove the background noise while retaining the seismic signal well.

[0075] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0076] Referring to Figures 3(a) to 4(d) , the denoising effects of two synthetic seismic test data are presented, FIG. 3(a) and FIG. 4(a) show the clean seismic data, and the noise data is shown in FIG. 3(b) and FIG. 4(b), it is obvious that the effective signal is contaminated by strong energy background noise; FIG. 3(c) and FIG. 4(c) are the results after denoising by the U-shaped network with global context module and attention mechanism (GC-AB-Unet), it can be seen that the noise is largely removed, and the seismic signal is clearer; FIG. 3(d) shows the difference data of the noisy data FIG. 3(b) and the denoised data FIG. 3(c), and FIG. 4(d) is the difference data of the noisy data FIG. 4(b) and the denoised data FIG. 4(c), it can be seen that the seismic signal is basically not damaged, and a large amount of noise is suppressed; from the quantitative indicator analysis, the signal-to-noise ratio (SNR) of the first synthetic data is increased from 5.89 to 13.14, and the signal-to-noise ratio of the second synthetic data is increased from 5.39 to 16.83, the signal-to-noise ratio is greatly improved; the above analysis shows that the noise in the synthetic seismic data is well suppressed, and the signal is well preserved.

[0077] Referring to Figures 5(a) to 7(e), further test the network de-noising performance by using three actual DAS-VSP seismic data. Fig. 5(a), Fig. 6(a) and Fig. 7(a) are noisy actual DAS-VSP seismic data, which contain 225 channels, each channel data has 7000 samples, the sampling rate is 0.001s, and it can be seen that a large amount of strong energy background noise will cover the seismic effective signal, which cannot be clearly identified; Fig. 5(b), Fig. 6(b) and Fig. 7(b) are data after de-noising by the application, the quality of seismic data is significantly improved, and the effective signal which is interfered by noise and relatively weak becomes more continuous and clear; correspondingly, Fig. 5(c), Fig. 6(c) and Fig. 7(c) are the difference data between the noisy data and the de-noised data, and there is a small amount of signal leakage in the removed noise part, but the noise is greatly suppressed; further, the 109th, 167th and 99th channels of the three actual DAS-VSP seismic data are extracted respectively for single-channel signal analysis, Fig. 5(d), Fig. 6(d) and Fig. 7(d) are the results of time domain analysis of single-channel signal, it can be seen that the noise is greatly suppressed, and correspondingly, Fig. 5(e), Fig. 6(e) and Fig. 7(e) are the results of frequency domain analysis of single-channel signal, it can be found that the high-frequency noise and the fixed frequency noise are well suppressed in the frequency domain graph; the actual DAS-VSP seismic data shows that the application is suitable for effectively removing complex background noise while preserving seismic effective signal.

[0078] The above is only to illustrate the technical idea of the application, and cannot limit the protection scope of the application, and any modification made according to the technical idea of the application on the basis of the technical scheme falls within the protection scope of the claims of the application.

Claims

1. A deep learning-based DAS-VSP data background noise suppression method, characterized in that, The method comprises the following steps: S1, obtaining synthetic seismic data as clean training data set s by three different ways, extracting noise n in actual DAS-VSP seismic data and randomly injecting it into the clean training data to construct the corresponding noisy training data set y; Wherein, the three different ways are SEG public data: reflectivity method and finite difference method in time domain; S2, on the basis of the U-shaped network, the original activation function is changed from linear rectifier function to leaky linear rectifier function, so as to retain the negative values existing in the amplitude of the seismic signal; S3, a random inactivation layer is added before each down-sampling layer, that is, the parameters unimportant to the network are randomly discarded to prevent overfitting phenomenon; S4, a global context module is introduced to connect the shrinkage network and the expansion network, which is used to extract global context information and improve the feature extraction capability of the network; S5, an attention mechanism module is added before the network output to capture the key features of the complex geological structure in the seismic data; S6, the network output is changed from the denoised data to the residual data; S7, the trained network is applied to synthetic seismic data and actual DAS-VSP seismic data, and the denoising effect is analyzed by quantitative index signal-to-noise ratio and visual effect.

2. The deep learning based DAS-VSP data background noise suppression method of claim 1, wherein, In step S1, the clean training data set is obtained by windowing method from the complete seismic record, the window size is 240*240, the data is sequentially cut from left to right and from top to bottom, and the window step is 120; synthetic VSP data is obtained by three different ways, noise is directly extracted from actual DAS-VSP seismic data, and the extracted noise is randomly injected into the clean VSP data set to obtain the corresponding noise VSP data set.

3. The deep learning based DAS-VSP data background noise suppression method of claim 1, wherein, In step S2, the activation function with leaky linear rectifier function is defined as follows: Wherein, x is the input seismic signal, and a is a constant multiplied by the signal when it is negative. The results of multiple experiments show that when a = 0.3, the signal-to-noise ratio of the network denoising performance reaches the highest value.

4. The deep learning based DAS-VSP data background noise suppression method of claim 3, wherein, The shrinkage network contains four down-sampling layers, and the expansion network contains four up-sampling layers, which are realized by maximum pooling and bilinear interpolation respectively.

5. The deep learning based DAS-VSP data background noise suppression method of claim 1, wherein, In step S3, each random inactivation layer randomly retains 90% of the hyperparameters, and discards the unimportant parameters in the network to prevent overfitting.

6. The deep learning based DAS-VSP data background noise suppression method of claim 1, wherein, In step S4, the global context module is introduced. First, the input data is processed by 1*1 convolution and normalization exponential function, then the obtained results are multiplied with the feature maps input into the module, and then 1*1 convolution, normalization layer, linear rectifier function and 1*1 convolution are used for feature transformation; finally, the initial input and the feature maps after feature transformation are added for feature fusion, that is, the global context features are aggregated to the features at each position.

7. The deep learning based DAS-VSP data background noise suppression method of claim 1, wherein, In step S5, the attention mechanism adopts cascade operation to fuse the noisy seismic data input into the network with the feature maps output by the last layer of the network. The specific steps are as follows: first, feature compression, that is, using 1*1 convolution to convert the channel number to 1; second, the obtained data is fused with the network input data through feature channel; In the third step, the obtained features are converted into nonlinear features through a hyperbolic tangent layer, and then normalized; finally, the data obtained in the third step are subjected to a 1*1 convolution operation, and then multiplied by the data output in the first step.

8. The deep learning based DAS-VSP data background noise suppression method of claim 1, wherein, In step S6, the network output is residual data, and the denoised seismic data is obtained by subtracting the network output from the network input.

9. The deep learning based DAS-VSP data background noise suppression method of claim 8, wherein, The denoising process is described as follows: where y is the seismic signal contaminated by noise, where y is the seismic signal contaminated by noise, where y is the seismic signal contaminated by noise, where y is the seismic signal contaminated by noise, where y is the seismic signal contaminated by noise, where y is the seismic signal contaminated by noise, where y is the seismic signal contaminated by noise, where y is the seismic signal contaminated by noise, where y is the seismic signal contaminated by noise, where y is the seismic signal contaminated by noise, where y is the seismic signal contaminated by noise, where y is the seismic signal contaminated where {y i , s i} represents N pairs of noisy and corresponding clean training data sets, both from synthetic seismic data; N refers to batch size, i.e. the number of input data required to calculate a loss value; network training uses adaptive gradient algorithm.

10. The deep learning based DAS-VSP data background noise suppression method of claim 1, wherein, In step S7, the signal-to-noise ratio is selected as a quantitative index to measure the denoising performance of the network, and the visual effect is presented through the denoised data and the difference data, and the denoising performance is analyzed by using the time-domain graph and the frequency-domain graph of the single-channel signal.

Images