DAS VSP data processing method and system based on multi-scale residual attention network

Abstract

The embodiment of the invention provides a DAS VSP data processing method and system based on a multi-scale residual attention network, and relates to the technical field of data processing, and the method comprises the following steps: obtaining first seismic data; wherein the first seismic data is continuous and noiseless seismic data; acquiring second seismic data; wherein the second seismic data is step data with fading noise obtained by processing on the basis of the first seismic data; building a multi-scale residual attention network model; wherein the multi-scale residual attention network model comprises an encoder, a multi-scale pyramid extraction module, a decoder and an output layer which are connected in sequence; inputting the first seismic data and the second seismic data into a multi-scale residual attention network model for training, and storing the trained multi-scale residual attention network model; in the training process, the signal-to-noise ratio and the structural similarity are used as evaluation indexes to evaluate the performance of the multi-scale residual attention network model.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular to a DAS VSP data processing method and system based on a multi-scale residual attention network. Background Art

[0002] Energy is a key area that is related to national security and development. In recent years, energy exploration technology based on seismic exploration data has developed rapidly. Energy exploration is mainly carried out to obtain underground structural information and distribution characteristics, etc., to provide a basis for further oil and gas energy development.

[0003] Among the related technologies, the combination of Distributed Acoustic Sensing (DAS) system and Weak Fiber Bragg Grating (WFBG) is widely used in the field of oil and gas energy exploration due to its high sensitivity, low cost, strong shape adaptability, one-time excitation and full-cable reception.

[0004] However, there are problems of fading noise and step-shaped signals in the vertical seismic profile (VSP) data collected by the DAS system, which affects the validity of the seismic data. Summary of the invention

[0005] The embodiments of the present application provide a DAS VSP data processing method and system based on a multi-scale residual attention network to solve the problem of fading noise and step-shaped signals in the VSP data collected by the DAS system in the related art, which affects the technical problem of the validity of the seismic data.

[0006] The embodiment of the present application provides a DAS VSP data processing method based on a multi-scale residual attention network, comprising the following steps: Acquire first seismic data; wherein the first seismic data is continuous and noise-free seismic data; Acquire second seismic data; wherein the second seismic data is step data with fading noise obtained by processing the first seismic data; Build a multi-scale residual attention network model; wherein the multi-scale residual attention network model includes an encoder, a multi-scale pyramid extraction module, a decoder and an output layer connected in sequence; Input the first seismic data and the second seismic data into the multi-scale residual attention network model for training, and save the trained multi-scale residual attention network model; wherein the first seismic data is the actual target value, and the second seismic data is the input value; the encoder is configured to extract the first feature of the second seismic data using different receptive fields; the multi-scale pyramid extraction module is configured to capture global information from multiple receptive fields for the first feature input by the encoder to obtain the second feature; the decoder is configured to stack and fuse the first feature and the second feature to obtain the third feature; the output layer is configured to convolve the third feature and then output the training result; During the training process, signal-to-noise ratio and structural similarity are used as evaluation indicators to evaluate the performance of the multi-scale residual attention network model.

[0007] In one possible implementation, the multi-scale pyramid extraction module and the decoder both include an attention mechanism, and the attention mechanism is configured to assign different weights to features in the multi-scale residual attention network model.

[0008] In a feasible implementation, the multi-scale residual attention network model also includes a residual structure, and the encoder, decoder and multi-scale pyramid extraction module are connected through the residual structure.

[0009] In one possible implementation, the encoder includes two first convolutional layers and a downsampling layer; The encoder is configured to perform first feature extraction on the second seismic data, specifically comprising: The first feature of the second seismic data is extracted through the convolution operations of the two first convolutional layers in sequence and then output to the downsampling layer.

[0010] In a feasible implementation, three encoders are provided; The three encoders are connected via a downsampling layer. The second seismic data is sequentially subjected to first feature extraction using different receptive fields by the three encoders and then output to a multi-scale pyramid extraction module via the downsampling layer.

[0011] In one feasible implementation, the multi-scale pyramid extraction module includes: four dilated convolutional layers and a global pooling layer; The multi-scale pyramid extraction module is configured to capture global information from multiple receptive fields for the first feature input by the encoder to obtain a second feature, specifically including: The multi-scale pyramid extraction module is configured to capture global information from multiple receptive fields on the first feature of the encoder input through four dilated convolutional layers and a global pooling layer to obtain a second feature.

[0012] In one possible implementation, the decoder includes two second convolutional layers and an upsampling layer; The decoder is configured to stack and fuse the first feature and the second feature to obtain a third feature, specifically including: The decoder is configured to use two second convolutional layers to perform convolution to achieve stacking and fusion of the first features and the second features of each scale in the channel direction to obtain the third feature, which is output to the output layer by the upsampling layer.

[0013] In a feasible implementation, there are three decoders; The three decoders are connected via an upsampling layer, and the first feature and the second feature are stacked and fused in turn via the three decoders to obtain the third feature, which is then output to the output layer via the upsampling layer.

[0014] In one possible implementation, the encoder also includes batch normalization and ReLU activation function; Batch Normalization is configured to normalize the feature data; The ReLU activation function is configured to increase the nonlinear modeling capability of the multi-scale residual attention network model.

[0015] In a feasible implementation, after the step of inputting the first seismic data and the second seismic data into the multi-scale residual attention network model for training, and saving the trained multi-scale residual attention network model, the step further includes: The parameters of the trained multi-scale residual attention network model are adjusted according to the loss function; the calculation formula of the loss function is as follows: ; in, A loss function that represents structural information and signal amplitude; A loss function that represents structural information; A loss function representing the signal amplitude; represents the first earthquake data; represents the data estimated by the model; and Respectively represent the number of seismic traces and the number of samples per trace; and express and The average value of and yes and The standard deviation of yes and The covariance of , is a small constant.

[0016] In a feasible implementation, during the training process, the signal-to-noise ratio and structural similarity are used to evaluate the performance of the multi-scale residual attention network model, where the calculation formula of the evaluation index is as follows: ; ; In the above formula, represents the signal-to-noise ratio; Indicates structural similarity; and Represent the energy of the whole data and the energy of the noise respectively; and Respectively represent the total number of seismic traces and time sampling points of the data; represents the first earthquake data; express The average value of Indicates data to be processed or processed; express The average value of and yes and The standard deviation of yes and The covariance of , is a small constant.

[0017] In a second aspect, a processing system is applied to a DAS data processing method based on a multi-scale residual attention network of any technical solution in the first aspect, and the processing system includes: Encoder, multi-scale pyramid module, decoder, and output layer; Among them, the encoder is configured to use different receptive fields for feature extraction; the multi-scale pyramid extraction module is configured to capture global information from multiple receptive fields; the decoder is configured to use a splicing layer to achieve stacking and fusion of features of each scale in the channel direction; and the output layer is configured to output the processing results.

[0018] In the first aspect, the embodiment of the present application provides a DAS VSP data processing method based on a multi-scale residual attention network. In the embodiment of the present application, a multi-scale residual attention network model is constructed, and the continuous and noise-free first seismic data is used as the actual target value, and the noisy second seismic data collected by the DAS is used as the input value to enter the multi-scale residual attention model for training, and the trained multi-scale residual attention network model is obtained. In the trained multi-scale residual attention network model, the encoder can use different receptive fields for feature extraction, thereby realizing a wide capture of details and global features; further, the embodiment of the present application can use different receptive fields to extract spatial information through the setting of a multi-scale pyramid extraction module, thereby helping the network to maintain continuity when reconstructing the signal, so as to enhance the network denoising and smoothing capabilities. Furthermore, the embodiment of the present application realizes the stacking and fusion of features of each scale in the channel direction by the splicing layer through the setting of the decoder. This strategy helps to capture more detail information and reduce the loss of detail features caused by early pooling operations. That is to say, the embodiment of the present application can denoise the noisy step data collected by DAS and smooth the step data in the seismic data by building and training the multi-scale residual attention network model. That is, through the setting of the embodiment of the present application, a DAS VSP data processing method of a multi-scale residual attention network that can denoise and smooth the step data of the VSP data collected by DAS is provided.

[0019] In the second aspect, an embodiment of the present application provides a DAS VSP data processing system based on a multi-scale residual attention network. Since the system is applied to the DAS VSP data processing method based on a multi-scale residual attention network of the first aspect, the system has all the technical effects of the technical solution of the first aspect and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 is an earthquake forward model provided by an embodiment of the present application; FIG. 2( a ) is a partial characteristic diagram of first seismic data provided in an embodiment of the present application; FIG2( b ) is a partial characteristic diagram of the second seismic data provided in an embodiment of the present application; Figure 3 It is a structural diagram of a DAS VSP data processing system based on a multi-scale residual attention network provided in an embodiment of the present application; Figure 4It is a method flow chart of a DAS VSP data processing method based on a multi-scale residual attention network provided in one embodiment of the present application; Figure 5 This is a curve chart of the loss changes of the training set and the test set during the training process; FIG6 (a) is a curve diagram of the signal-to-noise ratio variation of a test set during the training process of a multi-scale residual attention network model provided in an embodiment of the present application; FIG6 (b) is a curve diagram of the change in structural similarity of a test set during the training process of a multi-scale residual attention network model provided in an embodiment of the present application; FIG. 7 ( a ) is a smoothed third seismic data without fading noise provided by an embodiment of the present application; FIG7( b ) is a fourth step-shaped seismic data obtained by adding fading noise to the third seismic data in FIG7( a ); FIG. 7 ( c ) is fifth seismic data after the fourth seismic data in FIG. 7 ( b ) is processed by the DAS VSP data processing method based on the multi-scale residual attention network provided in an embodiment of the present application; FIG7( d ) is sixth seismic data obtained by processing the fourth seismic data in FIG7( b ) using a denoising convolutional neural network; FIG7( e ) is the seventh seismic data obtained by processing the fourth seismic data in FIG7( b ) using UNet3; FIG7( f ) is the eighth seismic data obtained by processing the fourth seismic data in FIG7( b ) using UNet+3.

[0021] Description of reference numerals: 110-first encoder; 120-second encoder; 130-third encoder; 140-first convolutional layer; 150-downsampling layer; 200-multi-scale pyramid extraction module; 210-first dilated convolution layer; 220-second dilated convolution layer; 230-third dilated convolution layer; 240-fourth dilated convolution layer; 250-global pooling layer; 310 - first decoder; 320 - second decoder; 330 - third decoder; 340 - second convolutional layer; 350 - upsampling layer; 400-Output layer. DETAILED DESCRIPTION

[0022] In order to enable those skilled in the art to better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present application.

[0023] It should be noted that many specific details are set forth in the following description to facilitate a full understanding of the present application. However, the present application may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present application is not limited to the specific implementation methods disclosed below.

[0024] In the description of the present application, it should be understood that the terms "center", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "axial", "radial", "circumferential" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application. In the present application, unless otherwise expressly specified and limited, a first feature being "up" or "down" a second feature may be the first and second features directly contacting, or the first and second features indirectly contacting through an intermediate medium.

[0025] In this application, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral one; it can be directly connected or indirectly connected through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. However, if it is indicated as a direct connection, it means that the two connected entities are not connected through an intermediate structure, but are only connected to form a whole through a connecting structure. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to the specific circumstances.

[0026] In this application, the descriptions of "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" or "second" may explicitly or implicitly include at least one of the features.

[0027] Energy is a key area related to national security and development. In recent years, energy exploration technology based on seismic exploration data has developed rapidly. Energy exploration is mainly to obtain underground structural information and distribution characteristics, etc., to provide a basis for further oil and gas energy development. At present, traditional seismic detectors used for seismic exploration data collection have many limitations and can no longer meet the needs of seismic exploration in complex geological environments.

[0028] Among the related technologies, the combination of Distributed Acoustic Sensing (DAS) system and Weak Fiber Bragg Grating (WFBG) is widely used in the field of oil and gas energy exploration due to its high sensitivity, low cost, strong shape adaptability, one-time excitation and full-cable reception.

[0029] It should be noted that DAS is a new type of high-density seismic acquisition system technology based on fiber optic sensing. It converts optical fiber into an array tool for measuring strain. It can perform continuous and real-time measurements on the entire optical fiber and is usually used to obtain temperature, strain and vibration data.

[0030] It should be further explained that WFBG is a low-reflectivity fiber Bragg Grating (FBG), and the reflectivity of WFBG is usually less than 0.1%. Weakly reflective fiber grating arrays are densely engraved with thousands of WFBGs on an optical fiber, and the grating pitch (the spacing from center to center) can be as small as 10mm, and the spacing between two adjacent gratings can be as small as 1mm.

[0031] It is necessary to further explain that the Vertical Seismic Profile (VSP) is a seismic observation method. The vertical profile is equivalent to the ground seismic profile. This method is to observe the seismic wave field in the well, and place seismic detectors at different depths in the well to record the seismic signals generated by the ground source.

[0032] However, the data collected by the DAS system using WFBG as the sensing unit has fading noise, and the collected near-shot point seismic data shows step-like characteristics due to the influence of pulse width and grid spacing. That is, there are fading noise and step-like signal problems in the seismic data collected by the DAS system, which affects the validity of the seismic data and limits the application and promotion of DAS in the field of seismic exploration.

[0033] Therefore, the embodiments of the present application provide a DAS VSP data processing method and system based on a multi-scale residual attention network to solve the problem of fading noise and step-shaped signals in the VSP data collected by the DAS system in the related art, which is a technical problem that affects the validity of seismic data.

[0034] Figure 1 is a seismic forward model provided by an embodiment of the present application; FIG. 2 (a) is a partial characteristic diagram of the first seismic data provided by an embodiment of the present application; FIG. 2 (b) is a partial characteristic diagram of the second seismic data provided by an embodiment of the present application; Figure 3 It is a structural diagram of a DAS VSP data processing system based on a multi-scale residual attention network provided in an embodiment of the present application; Figure 4 This is a method flow chart of a DAS VSP data processing method based on a multi-scale residual attention network provided in one embodiment of the present application.

[0035] First, refer to Figures 1 to 4 , the embodiment of the present application provides a DASVSP data processing method based on a multi-scale residual attention network, comprising the following steps: S1: Acquire the first seismic data.

[0036] A forward model is established through simulation software to perform simulation, and a number of continuous and noise-free first seismic data are obtained and saved.

[0037] It should be noted that earthquake forward modeling Figure 1 As shown, in Figure 1 In the figure, the horizontal axis represents the width of the formation, the vertical axis represents the depth of the formation, and different areas represent the sound speed of different formations. Figure 1 In the figure, area A indicates that the sound speed of the stratum is 4000m / s, area B indicates that the sound speed of the stratum is 3200m / s, area C indicates that the sound speed of the stratum is 2400m / s, area D indicates that the sound speed of the stratum is 1800m / s, and area E indicates that the sound speed of the stratum is 1200m / s.

[0038] The first seismic data is continuous and noise-free seismic data.

[0039] In the subsequent processing process, since the first seismic data is continuous and noise-free, the first seismic data is used as the target data in the processing process.

[0040] S2: Acquire the second seismic data.

[0041] The second seismic data is step-shaped data with fading noise obtained after further simulation processing based on the first seismic data.

[0042] In the specific implementation, in S2, the first seismic data is processed by column averaging according to the characteristics of the weak grating to obtain step data, and then the step data is normalized, and a 128×128 sliding window is used to segment the first seismic data and the step data with the same moving direction and moving step size, and the corresponding data blocks are named with the same serial number and saved separately.

[0043] Exemplarily, the second seismic data is averaged by region, with 10 columns of data as a step region, and the average of ten data in each row is used to replace the original ten data.

[0044] DAS is used to collect the first seismic data without adding a signal source, and the fading noise part is extracted, and several noise data are normalized. The noise block is cut into data segments with a length of 128, and then randomly added to the step-shaped data block and saved to obtain the second seismic data. A database is constructed with the step-shaped second seismic data containing fading noise as the feature and the continuous and noise-free first seismic data as the label. Figure 2 (a) shows the continuous and noise-free first seismic data, and Figure 2 (b) shows the step-shaped second seismic data containing fading noise.

[0045] S3: Build a multi-scale residual attention network model.

[0046] Reference Figure 3 The multi-scale residual attention network model includes an encoder (Encode), a multi-scale pyramid extraction module 200, a decoder (Decode) and an output layer 400 connected in sequence.

[0047] The multi-scale residual attention network includes an encoder, a multi-scale pyramid extraction module 200 and a decoder; the encoder is configured to perform feature extraction; the multi-scale pyramid extraction module 200 is configured to capture global information from multiple receptive fields; the decoder is configured to use a splicing layer to achieve stacking and fusion of features of each scale in the channel direction; the output layer is configured to output the processing results.

[0048] Among them, when training the multi-scale residual attention network model, the network training parameters used are shown in Table 1.

[0049] Table 1 Network training parameters

[0050] S4: Input the first earthquake data and the second earthquake data into the multi-scale residual attention network model for training, and save the trained multi-scale residual attention network model.

[0051] Among them, the first seismic data is the actual target value, and the second seismic data is the input value to establish a database; the encoder is configured to extract the first feature of the second seismic data; the multi-scale pyramid extraction module 200 is configured to capture global information from multiple receptive fields for the first feature input by the encoder to obtain the second feature; the decoder is configured to stack and fuse the second feature to obtain the third feature; the output layer 400 is configured to convolve the third feature and output the training result.

[0052] Reference Figure 3 As shown, there are three encoders.

[0053] In the embodiment of the present application, by setting the number of encoders to three, it is not only possible to use different receptive fields to extract features from the second seismic data, but also possible to reduce the amount of data calculation.

[0054] In some examples, the encoder includes two first convolutional layers 140 and a downsampling layer 150; The encoder is configured to use different receptive fields to extract the first feature of the second seismic data, specifically including: The first feature of the second seismic data is extracted by performing convolution operations of two first convolutional layers 140 in sequence and then output to the downsampling layer 150 for compression.

[0055] The first convolutional layer 140 is configured to perform a first feature extraction on the input information through a convolution operation. The size of the area mapped by a point in the obtained feature map on the input is the receptive field. The larger the receptive field, the more context information it can contain.

[0056] The downsampling layer 150 is located at the output end of the two first convolutional layers 140 to compress the first features output by the two first convolutional layers 140, thereby reducing the amount of calculation, increasing the receptive field, and extracting more abstract high-level features by reducing the resolution of the feature map.

[0057] Reference Figure 3As shown, the encoder includes a first encoder 110, a second encoder 120 and a third encoder 130, each of which includes two first convolutional layers 140 and a downsampling layer 150. In the first encoder 110, the two first convolutional layers 140 each use 32 3×3 convolution kernels to process the input second seismic data, and the downsampling layer 150 compresses the first feature output by the first convolutional layer 140 in the first encoder 110; in the second encoder 120, the two first convolutional layers 140 each use 64 3×3 convolution kernels to process the input second seismic data, and the downsampling layer 150 compresses the first feature output by the second convolutional layer 340 in the second encoder 120; in the third encoder 130, the two first convolutional layers 140 each use 128 3×3 convolution kernels to process the input second seismic data, and the downsampling layer 150 compresses the first feature output by the first convolutional layer 140 in the first encoder 110.

[0058] In some examples, the encoder further includes batch normalization and ReLU activation functions, which are arranged after the first convolutional layer 140; Batch normalization is configured to normalize the data input to the first convolutional layer 140; The ReLU activation function is configured to increase the nonlinear modeling capability of the multi-scale residual attention network model.

[0059] In the embodiment of the present application, batch normalization (Batch Normalization, BN) is to normalize each batch of second seismic data and each layer of input data to the multi-scale residual attention network model (mean is 0, standard deviation is 1) to accelerate network training. As the number of network layers increases, the operation of the first convolutional layer 140 makes the distribution difference of each layer of training data increasingly larger, and the parameters need to be continuously adjusted to adapt to the new data distribution, which will make network learning difficult and the convergence speed of the loss function slow down. Therefore, the convergence speed of the network is accelerated by batch normalization, so that the multi-scale residual attention network model can also converge smoothly.

[0060] Furthermore, in the embodiment of the present application, a linear rectified linear unit (RLU) activation function is introduced between layers to increase the nonlinear modeling capability of the network, so that the network can better handle more complex problems. The embodiment of the present application can improve the convergence speed of the multi-scale residual attention network model by setting the RLU activation function.

[0061] Exemplarily, the encoder also integrates residual connections.

[0062] In an embodiment of the present application, residual connections are used to prevent the deep neural network from degenerating, thereby improving the model stability.

[0063] In some examples, the multi-scale pyramid extraction module 200 includes: four dilated convolutional layers and a global pooling layer 250; The multi-scale pyramid extraction module 200 is configured to obtain a second feature by capturing global information from multiple receptive fields for a first feature input by the encoder, specifically including: The multi-scale pyramid extraction module 200 is configured to obtain a second feature by convolving the first feature of the encoder input through four dilated convolution layers and a global pooling layer 250 to capture global information from multiple receptive fields.

[0064] It should be noted that dilated convolution is to inject holes into the standard convolution map, that is, fill it with zeros, and expand the convolution kernel to a specified size to achieve the purpose of expanding the receptive field. Compared with ordinary convolution, the similarity between dilated convolution and ordinary convolution is that the size of the actual convolution kernel is consistent, which means that the number of parameters in the network remains unchanged. The difference is that the receptive field of dilated convolution is larger, that is, dilated convolution can increase the receptive field of the convolution kernel while keeping the number of parameters unchanged, so that the output of each convolution contains a larger range of information. Dilated convolution can have a larger range of receptive fields and extract richer features under the condition of the same number of parameters. Therefore, we use dilated convolutions of different sizes in the network to obtain multi-scale information.

[0065] Reference Figure 3 As shown, the four dilated convolution layers include: a first dilated convolution layer 210, a second dilated convolution layer 220, and a third dilated convolution layer 230. In the first dilated convolution layer 210, the first feature of the encoder input is processed by the first dilated convolution layer 210 using 128 3×3 convolution kernels and a dilation rate of 1; in the second dilated convolution layer 220, the first feature of the encoder input is processed by the second dilated convolution layer 220 using 128 3×3 convolution kernels and a dilation rate of 2; in the third dilated convolution layer 230, the first feature of the encoder input is processed by the third dilated convolution layer 230 using 128 3×3 convolution kernels and a dilation rate of 4; in the fourth dilated convolution layer 240, the first feature of the encoder input is processed by the fourth dilated convolution layer 240 using 128 3×3 convolution kernels and a dilation rate of 6.

[0066] Furthermore, the first feature is also globally averaged pooled in the global pooling layer 250 to compress the spatial information on each channel of the feature map into a single numerical value, which represents the average activation value of the entire feature map, thereby compressing the image data, reducing network parameters, and improving the generalization ability of the network.

[0067] In some examples, the multi-scale pyramid extraction module 200 also includes an attention mechanism, which is configured to assign different weights to features in the multi-scale residual attention network model.

[0068] It should be noted that the attention mechanism is a mechanism for learning the importance of different information. The "attention" in it is similar to the visual attention mechanism possessed by humans. Human vision can quickly scan the things seen and obtain the area that needs to be focused on, which is what we call the focus of attention. Then, by paying more attention to this area, more detailed information about the target of attention can be obtained and other irrelevant information can be suppressed. By using a simple attention mechanism, useful information can be quickly obtained from massive information, which greatly improves the efficiency and accuracy of information processing.

[0069] In the specific implementation, the embodiment of the present application adopts SimAM (Simple Attention Module).

[0070] It should be noted that SimAM is a lightweight attention mechanism designed to enhance the representation ability of neural networks without significantly increasing computational overhead. It improves the network's attention to key features by adaptively weighting feature maps.

[0071] In the specific implementation, the basic idea of ​​SimAM is to use a simple module to capture the important information in the feature map, that is, SimAM achieves efficient feature enhancement by introducing an adaptive weighting mechanism.

[0072] The embodiment of the present application adopts SimAM, and since SimAM does not introduce additional convolutional layers or fully connected layers, the amount of computation is relatively low. Furthermore, the implementation of SimAM is relatively simple and can be easily embedded in existing convolutional neural networks. Furthermore, SimAM performs well in improving network performance and can significantly improve the accuracy of tasks such as image classification and target detection.

[0073] For example, each dilated convolutional layer is connected to a SimAM, and different weights can be assigned to the first feature through SimAM, so that the network can focus on the effective signal while quickly reducing the influence of noise and step edge features. That is, SimAM enhances the accuracy and effectiveness of network learning.

[0074] In an embodiment of the present application, the multi-scale pyramid extraction module 200 uses four different dilated convolutional layers and a global pooling layer 250 to capture global information from multiple receptive fields, helping the network to maintain continuity when reconstructing signals, thereby enhancing its denoising and smoothing capabilities.

[0075] In some examples, the encoder includes two second convolutional layers 340 and an upsampling layer 350 .

[0076] The decoder is configured to use two second convolutional layers 340 to perform convolution to achieve stacking and fusion of the first features and the second features of each scale in the channel direction to obtain the third feature, which is output to the output layer 400 by the upsampling layer 350.

[0077] Continue to refer to Figure 3 , there are three decoders, including a first decoder 310, a second decoder 320 and a third decoder 330, each of which includes two second convolution layers 340 and an upsampling layer 350. In the first decoder 310, upsampling firstly enlarges the first feature output by the first encoder and the second feature output by the multi-scale pyramid feature extraction module, and then the two first convolution layers 140 use 128 3×3 convolution kernels to process the first and second features of the input to obtain the third feature; in the second decoder 320, upsampling firstly enlarges the first feature output by the second encoder and the second feature output by the multi-scale pyramid feature extraction module, and then the two first convolution layers 140 use 64 3×3 convolution kernels to process the first and second features of the input to obtain the third feature; upsampling firstly enlarges the first feature output by the third encoder and the second feature output by the multi-scale pyramid feature extraction module, and then the two first convolution layers 140 use 32 3×3 convolution kernels to process the first and second features of the input to obtain the third feature.

[0078] In some examples, the decoder also includes SimAM, and each second convolution layer 340 is connected to a SimAM, and different weights can be assigned to the second features through SimAM, so that the network can focus on valid signals while quickly reducing the influence of noise and step edge features. That is, SimAM enhances the accuracy and effectiveness of network learning.

[0079] In some examples, the decoder also includes batch normalization and ReLU activation functions, which are set after the second convolutional layer 340; batch normalization is configured to normalize the data input to the second convolutional layer 340; the ReLU activation function is configured to increase the nonlinear modeling capability of the multi-scale residual attention network model.

[0080] In the embodiment of the present application, batch normalization (Batch Normalization, BN) is to normalize each batch of second seismic data and each layer of input data to the multi-scale residual attention network model (mean is 0, standard deviation is 1) to accelerate network training. As the number of network layers increases, the second convolutional layer 340 operation makes the distribution difference of each layer of training data become larger and larger, and the parameters need to be continuously adjusted to adapt to the new data distribution, which will make network learning difficult and the convergence speed of the loss function slow down. Therefore, the convergence speed of the network is accelerated by batch normalization, so that the multi-scale residual attention network model can also converge smoothly.

[0081] Furthermore, in the embodiment of the present application, a linear rectified linear unit (RLU) activation function is also introduced between layers to increase the nonlinear modeling capability of the network, so that the network can better handle more complex problems. The embodiment of the present application can improve the convergence speed of the multi-scale residual attention network model by setting the RLU activation function.

[0082] In some other examples, the multi-scale residual attention network model also includes a residual structure, and the encoder, decoder and multi-scale pyramid extraction module 200 are connected through the residual structure.

[0083] It should be noted that the residual structure connection is used to connect two or more tensors together in a certain dimension to generate a larger tensor. This can be used to increase the number of channels or feature dimensions in a deep learning model so that the relationship between different features can be better captured in subsequent layers.

[0084] In an embodiment of the present application, the multi-scale residual attention network model widely integrates residual connections to prevent the degradation of the deep neural network and improve the stability of the model.

[0085] Continue to refer to Figure 3 ,The feature map decoded by the decoder is finally processed by an output layer 400 of a 3×3 convolution kernel, outputting 1×128×128 data and the final training result.

[0086] Figure 5 This is a curve chart of the loss changes of the training set and the test set during the training process.

[0087] S6: During the training process, signal-to-noise ratio and structural similarity are used as evaluation indicators to evaluate the performance of the multi-scale residual attention network model.

[0088] During the training process, refer to Figure 5Curve F is the curve diagram of the training set changing with the number of iterations, and curve G is the curve diagram of the test set changing with the number of iterations. Figure 5 It can be seen that during the training process of the multi-scale residual attention network model, the loss of the training set and the test set continues to decrease.

[0089] It should be noted that the training set is the data set used to train the model. On this data set, the model learns to identify patterns and relationships in the data and adjusts its parameters to minimize the prediction error. The test set is the data set used to evaluate the performance of the model. It contains data that the model has not seen during the training process and is used to test the generalization ability of the model.

[0090] The signal-to-noise ratio and structural similarity are used as evaluation indicators to evaluate the performance of the multi-scale residual attention network model. The calculation formula of the evaluation indicator is as follows: (1) (2) In formula (1) and formula (2), represents the signal-to-noise ratio; Indicates structural similarity; and Represent the energy of the whole data and the energy of the noise respectively; and Respectively represent the total number of seismic traces and time sampling points of the data; represents the first earthquake data; express The average value of Indicates data to be processed or processed; express The average value of and yes and 's mark; yes and The covariance of , is a small constant.

[0091] S6: Adjust the parameters of the trained multi-scale residual attention network model according to the loss function.

[0092] The loss considers both structural information and signal amplitude.

[0093] It should be noted that Represents structural information, represents the signal amplitude, and and The loss function is used as the evaluation indicator.

[0094] Loss Function The formula is as follows: (3) In formula (3), A loss function that represents structural information and signal amplitude; A loss function that represents structural information; A loss function representing the signal amplitude; represents the first earthquake data; represents the data estimated by the model; and represent the number of seismic channels and the number of samples per channel respectively; and express and The average value of and yes and The standard deviation of yes and The covariance of , is a small constant.

[0095] Figure 6 (a) is a curve diagram of the change in signal-to-noise ratio of the test set during the training process of the multi-scale residual attention network model provided in an embodiment of the present application; Figure 6 (b) is a curve diagram of the change in structural similarity of the test set during the training process of the multi-scale residual attention network model provided in an embodiment of the present application.

[0096] The evaluation results are shown in Figure 6 (a) and Figure 6 (b). In Figure 6 (a), curve H represents the curve of the signal-to-noise ratio of the input noisy data during the training of the multi-scale residual attention network model as the number of iterations changes; the dotted line I represents the noisy data. It can be seen from curve H that the signal-to-noise ratio is constantly improving.

[0097] In Figure 6 (b), curve J represents the curve of the structural similarity of the input noisy data as the number of iterations changes during the training of the multi-scale residual attention network model; the dotted line K represents the noisy data. It can be seen from curve H that the structural similarity is constantly improving.

[0098] Figure 7 (a) is the smoothed third seismic data without fading noise provided by an embodiment of the present application; Figure 7 (b) is the stepped fourth seismic data after adding fading noise to the third seismic data in Figure 7 (a); Figure 7 (c) is the fifth seismic data after the fourth seismic data in Figure 7 (b) is processed by the DAS VSP data processing method based on the multi-scale residual attention network provided in an embodiment of the present application; Figure 7 (d) is the sixth seismic data obtained by processing the fourth seismic data in Figure 7 (b) using a denoising convolutional neural network; Figure 7 (e) is the seventh seismic data obtained by processing the fourth seismic data in Figure 7 (b) using UNet; Figure 7 (f) is the eighth seismic data obtained by processing the fourth seismic data in Figure 7 (b) using UNet+3.

[0099] In FIG. 7 ( a ), block l is a partially enlarged schematic diagram of block L. The data in FIG. 7 ( a ) does not contain noise and step data.

[0100] In FIG7( b ), block m is a local enlarged schematic diagram of block M. In block m, a step-like signal can be clearly seen. The arrow N in FIG7( b ) points to coherent fading noise.

[0101] FIG7 (c) is the result of processing FIG7 (b) using the multi-scale residual attention network based on the embodiment of the present application. No fading noise or step signal is observed in the data after processing using the DAS VSP data processing method based on the multi-scale residual attention network. In FIG7 (c), block q is a partial enlarged schematic diagram of block Q.

[0102] FIG7(d) is the sixth seismic data obtained by processing the fourth seismic data in FIG7(b) using a denoising convolutional neural network (Dncnn). In FIG7(d), block r is a partial enlarged schematic diagram of block R. In block R, it can be seen that the signal is still in a step shape. Furthermore, in FIG7(d), it can be seen that there is fading noise at the arrow S, and the signal structure is indeed and discontinuous.

[0103] FIG7(e) is the seventh seismic data obtained by processing the fourth seismic data in FIG7(b) by UNet; in FIG7(e), block t is a partial enlarged schematic diagram of block T. No step-like signal is seen in the figure. The network can achieve signal smoothing, but there is fading noise at the arrow W.

[0104] Figure 7 (f) is the eighth seismic data obtained by processing the fourth seismic data in Figure 7 (b) using UNet+3; in Figure 7 (f), block y is a local enlarged schematic diagram of block Y. No step-like signal is seen in the figure. The network can achieve signal smoothing, but there is fading noise at the arrow Z, the signal structure is discontinuous, and UNet3+ is insufficient to recover the signal.

[0105] The noise-free smooth data shown in Figure 7 (a) is used as the standard image to calculate the evaluation index. The results are shown in Table 2. The changes in the evaluation index of the synthetic seismic record. After the multi-scale DAS VSP data processing method based on the residual attention network is processed, the data signal-to-noise ratio is improved by 13.125dB, and the structural similarity is increased from 0.933 to 0.995. Among the four networks of multi-scale residual attention network, Dncnn, UNet and UNet+3, the DAS data processing method based on the multi-scale residual attention network has the highest signal-to-noise ratio improvement and the highest structural similarity.

[0106] Table 2 Changes in evaluation indicators of synthetic seismic records

[0107] In a second aspect, an embodiment of the present application further provides a DAS VSP data processing system based on a multi-scale residual attention network, the processing system is applied to the DASVSP data processing method based on a multi-scale residual attention network in any technical solution of the first aspect, and the processing system includes: Encoder, multi-scale pyramid module and decoder; Among them, the encoder is configured to use different receptive fields for feature extraction; the multi-scale pyramid extraction module 200 is configured to capture global information from multiple receptive fields; and the decoder is configured to use a splicing layer to achieve stacking of features of each scale in the channel direction.

[0108] It is easy to understand that those skilled in the art can combine, split, reorganize, etc. the embodiments of the present application to obtain other embodiments based on the several embodiments provided in the present application, and these embodiments do not exceed the protection scope of the present application.

[0109] The above specific implementation methods further explain in detail the purpose, technical solutions and beneficial effects of the embodiments of the present application. It should be understood that the above are only specific implementation methods of the embodiments of the present application and are not used to limit the protection scope of the embodiments of the present application. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solutions of the embodiments of the present application should be included in the protection scope of the embodiments of the present application.

Claims

1. A DAS VSP data processing method based on a multi-scale residual attention network, characterized in that: The following steps are involved: Acquire first seismic data; wherein the first seismic data is continuous and noise-free seismic data; Acquire second seismic data; wherein the second seismic data is step data with fading noise obtained by processing the first seismic data; Building a multi-scale residual attention network model; wherein the multi-scale residual attention network model includes an encoder, a multi-scale pyramid extraction module, a decoder and an output layer connected in sequence; The first seismic data and the second seismic data are input into the multi-scale residual attention network model for training, and the trained multi-scale residual attention network model is saved; wherein the first seismic data is the actual target value, and the second seismic data is the input value; the encoder is configured to extract the first feature of the second seismic data; the multi-scale pyramid extraction module is configured to capture global information from multiple receptive fields for the first feature input by the encoder to obtain the second feature; the decoder is configured to stack and fuse the first feature and the second feature to obtain the third feature; the output layer is configured to convolve the third feature and output the training result; During the training process, signal-to-noise ratio and structural similarity are used as evaluation indicators to evaluate the performance of the multi-scale residual attention network model.

2. The DAS VSP data processing method based on multi-scale residual attention network according to claim 1 is characterized in that: The multi-scale pyramid extraction module and the decoder both include an attention mechanism configured to assign different weights to features in the multi-scale residual attention network model.

3. The DAS VSP data processing method based on multi-scale residual attention network according to claim 1 is characterized in that: The multi-scale residual attention network model also includes a residual structure, and the encoder, the decoder and the multi-scale pyramid extraction module are connected through the residual structure.

4. The DAS VSP data processing method based on multi-scale residual attention network according to claim 1 is characterized in that: The encoder comprises two first convolutional layers and a downsampling layer; The encoder is configured to extract the first feature from the second seismic data, specifically comprising: The first feature of the second seismic data is extracted by sequentially performing convolution operations of two first convolutional layers and then outputted to the downsampling layer.

5. The DAS VSP data processing method based on multi-scale residual attention network according to claim 4 is characterized in that: The encoder is provided with three; The three encoders are connected via the downsampling layer, and the second seismic data is sequentially subjected to first feature extraction using different receptive fields by the three encoders, and then output to the multi-scale pyramid extraction module via the downsampling layer.

6. The DAS VSP data processing method based on a multi-scale residual attention network according to any one of claims 1 to 5, characterized in that: The multi-scale pyramid extraction module includes: four dilated convolutional layers and a global pooling layer; The multi-scale pyramid extraction module is configured to capture global information from multiple receptive fields for the first feature input by the encoder to obtain a second feature, specifically comprising: The multi-scale pyramid extraction module is configured to capture global information from multiple receptive fields on the first feature input by the encoder through four dilated convolution layers and a global pooling layer to obtain a second feature.

7. The DAS VSP data processing method based on a multi-scale residual attention network according to any one of claims 1 to 5, characterized in that: The decoder comprises two second convolutional layers and an upsampling layer; The decoder is configured to obtain a third feature by stacking and fusing the first feature and the second feature, specifically including: The decoder is configured to use two of the second convolutional layers to perform convolution to achieve stacking and fusion of the first features and the second features at each scale in the channel direction to obtain a third feature, which is output to the output layer by the upsampling layer.

8. The DAS VSP data processing method based on multi-scale residual attention network according to claim 7 is characterized in that: The decoder is provided with three; The three decoders are connected via the upsampling layer, the first feature and the second feature are stacked and fused in sequence via the three decoders to obtain the third feature, and the third feature is output to the output layer via the upsampling layer.

9. The DAS VSP data processing method based on a multi-scale residual attention network according to any one of claims 1 to 5, characterized in that: The encoder and the decoder both include batch normalization and ReLU activation functions; The batch normalization is configured to normalize the feature data; The ReLU activation function is configured to increase the nonlinear modeling capability of the multi-scale residual attention network model.

10. The DAS VSP data processing method based on a multi-scale residual attention network according to any one of claims 1 to 3, characterized in that: After the step of inputting the first seismic data and the second seismic data into the multi-scale residual attention network model for training, and saving the trained multi-scale residual attention network model, the step further includes: The parameters of the trained multi-scale residual attention network model are adjusted according to the loss function; wherein the calculation formula of the loss function is as follows: ； in, A loss function that represents structural information and signal amplitude; A loss function that represents structural information; A loss function representing the signal amplitude; representing said first seismic data; represents the data estimated by the model; and Respectively represent the number of seismic traces and the number of samples per trace; and express and The average value of and yes and The standard deviation of yes and The covariance of , is a small constant.

11. The DAS VSP data processing method based on multi-scale residual attention network according to claim 1 is characterized in that: During the training process, the signal-to-noise ratio and structural similarity are used as evaluation indicators to evaluate the performance of the multi-scale residual attention network model, wherein the calculation formula of the evaluation indicator is as follows: ； ； In the above formula, represents the signal-to-noise ratio; Indicates structural similarity; and Represent the energy of the whole data and the energy of the noise respectively; and Respectively represent the total number of seismic traces and time sampling points of the data; represents the first earthquake data; express The average value of Indicates data to be processed or processed; express The average value of and yes and The standard deviation of yes and The covariance of , is a small constant.

12. DAS VSP data processing system based on multi-scale residual attention network, characterized by: The processing system is applied to the DAS VSP data processing method based on the multi-scale residual attention network according to any one of claims 1 to 11, and the processing system comprises: an encoder, a multi-scale pyramid module, a decoder and an output layer; The encoder is configured to perform feature extraction; the multi-scale pyramid extraction module is configured to capture global information from multiple receptive fields; the decoder is configured to use a splicing layer to achieve stacking and fusion of features of each scale in the channel direction; and the output layer is configured to output processing results.

Images