DAS seismic noise suppression method based on multi-scale feature fusion network

Abstract

A DAS seismic noise suppression method based on a multi-scale feature fusion network belongs to the technical field of machine learning and seismic image processing, and aims to solve the problem of non-uniform noise distribution in seismic data, a multi-scale feature interaction module formed by instance normalization is introduced, features of adjacent scales are used as module input, and a multi-scale feature fusion network is constructed; context semantic information contained in the low-scale features and detail information contained in the current-scale features are used for complementation, and therefore self-adaptive processing is carried out on noise which is not evenly distributed. In order to solve the problem that noise is not stable, a self-adaptive feature fusion module is introduced, useful features are dynamically selected from each scale, the capability of removing complex noise can be enhanced, weak reflection signals can be recovered, self-adaptive fusion of multi-scale features is achieved, and meanwhile correlation between different scales can be enhanced. According to the method, complex noise in DAS seismic data can be effectively suppressed, and the structure and continuity of effective signals can be well recovered.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning and seismic exploration data denoising, and specifically relates to a DAS seismic noise suppression method based on a multi-scale feature fusion network. Background Art

[0002] Seismic exploration is one of the important means of oil and gas resource development. It mainly uses oil and gas geophysical exploration technology based on seismic exploration to explore and develop natural energy. Generally speaking, seismic exploration uses artificial sources to excite elastic waves, uses seismic instruments to detect the vibration of the earth at different positions along the survey line, and records the obtained data in digital form. Due to the interference of various noises, effective seismic signals on seismic records are often difficult to distinguish, which affects the subsequent seismic data processing and interpretation. Therefore, improving the signal-to-noise ratio of seismic signals is an important part of seismic data processing.

[0003] In previous seismic data processing, geophysical researchers have proposed many seismic noise denoising methods. They are not very effective when processing DAS seismic data with complex and diverse noise. For example, bandpass filters and fx deconvolution cannot effectively separate the reflection signal from the noise in the shared frequency band; the spectrum analysis-based methods can perform the denoising task well under the appropriate threshold function, however, it is difficult to handle aliased noise because it shares similar frequency bands with the effective signal. In addition, the determination of the threshold function is often done manually, and inappropriate filtering parameters will reduce the denoising ability; variational mode decomposition uses iterative search for the optimal solution to determine the optimal center frequency and bandwidth of each mode, but it does not work well under low signal-to-noise ratio (S / N); sparse transform-based methods are effective in seismic data denoising because seismic data are usually sparse. However, it is often challenging to determine the optimal threshold for sparsity-based methods, and inappropriate parameters will have a negative impact on the denoising ability and may restore false information to valid events. The above methods involve manual selection of parameters during data processing, which increases processing time and relies too much on manual experience.

[0004] In recent years, the rapid development of convolutional neural networks (CNNs) has stimulated the emergence of many effective seismic data denoising methods. Networks such as the feedforward denoising convolutional neural network (DnCNN) and the residual encoding and decoding network (REDNet) adaptively learn the structural features and detail features of signals in seismic images to better suppress noise in seismic images. Although these CNN-based methods have better performance than traditional methods, they usually only use features at a single scale, which limits the denoising ability of the network. The present invention uses a multi-scale feature fusion method for distributed optical fiber vertical seismic (DAS-VSP) data to suppress the noise in DAS-VSP seismic data, providing basic data for subsequent seismic data processing. 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 efficiently denoise and protect effective signals has always been an important research topic in denoising processing. Summary of the invention

[0005] The purpose of the present invention is to provide a DAS seismic noise suppression method based on a multi-scale feature fusion network that can overcome the problems that complex noise in DAS seismic data reduces the signal-to-noise ratio of seismic data, seriously damages effective signals, and makes the identification and recovery of seismic signals difficult.

[0006] The DAS seismic noise suppression method based on a multi-scale feature fusion network of the present invention comprises the following steps:

[0007] 1) Build a multi-scale feature fusion network framework, including the following steps:

[0008] 1.1 The network consists of three branches. The multi-scale feature fusion network consists of five parts: initial layer module, coarse fusion module, feature enhancement module, fine fusion module and reconstruction module;

[0009] 1.2 Build a shallow feature extraction module. The shallow feature extraction module is embedded in the upper, middle and lower branches to form the initial layer module. The noisy seismic data is taken as input and downsampled through a 3×3 convolution layer with a step size of 2. The feature map output by each convolution layer is passed through two convolution layers to obtain features of different scales X i , i∈{1,2,3}, where i is the branch number, and {1,2,3} correspond to the upper, middle, and lower branches, respectively, corresponding to large-scale features, medium-scale features, and small-scale features;

[0010] 1.3 Build a multi-scale feature interaction module, embed the multi-scale feature interaction module into the upper and middle branches respectively, the upper and middle branches include two multi-scale feature interaction modules, and the lower branch includes two convolutional layers to form a coarse fusion module; the multi-scale feature interaction module is used to input the i-th branch feature X i, i∈{1,2} performs convolution and instance normalization to extract the current scale feature M i , the lower-scale features of the input are transposed convolution to obtain the transposed features M that are consistent with the dimensions of the current scale features i+1 , and with the current scale feature M i Multiply element by element to get the aggregate feature, the aggregate feature and the transposed feature M i+1 After adding element by element, it is combined with the current scale feature X of the input i Add element by element to get the interaction feature I i ; The small-scale features are input into the third branch, and after two convolutional layers, the interactive features I3 are obtained;

[0011] 1.4 Build a hybrid attention module and embed the hybrid attention module into the upper, middle and lower branches to form a feature enhancement module; i , i∈{1,2,3} is used as the input of the hybrid attention module. It is first normalized through the layer, and then passes through the channel attention and spatial attention respectively. The channel attention includes global average pooling, fully connected layer and Sigmod function to obtain the channel attention feature; the spatial attention includes mean operation, convolution and Sigmod function to obtain the spatial attention feature. The channel attention feature and the spatial attention feature are added element by element, and then passed through 3 convolution layers and combined with the interaction feature I i , i∈{1,2,3} are added to obtain the enhanced feature E i , i∈{1,2,3};

[0012] 1.5 Build an adaptive feature fusion module and embed the adaptive feature fusion module into the upper, middle and lower branches respectively. The upper branch includes three adaptive feature fusion modules, and the middle and lower branches include two adaptive feature fusion modules to form a detailed fusion module. i , i∈{1,2,3} is used as the input of the adaptive feature fusion module. First, the enhanced feature E i The dimension is adjusted to the dimension of the branch where the current adaptive feature fusion module is located. The adjusted feature E i ', i∈{1,2,3} is concatenated, and the channel statistics are obtained by adaptive average pooling, the compact channel statistics are obtained by linear layer downsampling, and three parallel linear layers are upsampled and the Softmax function is used to obtain three weights, which are respectively compared with the adjusted features E of the three branches. i 'Weighted, the weighted features are added element by element to obtain the fusion feature F i r , i,r∈{1,2,3}, where: r is the number of adaptive feature fusion modules; the final output of the fine fusion module is the fusion feature F1 3 ;

[0013] 1.6 Convolutional layers constitute the reconstruction module, which combines the fusion feature F1 3 As the input of the convolutional layer, the predicted noise data is obtained;

[0014] 2) Optimizing the multi-scale feature fusion network, including the following steps:

[0015] 2.1 Construct a training data set, use the finite difference method to solve the elastic wave equation to construct a clean signal set, and extract noise from the actual data to construct a noise set;

[0016] 2.2 Set the hyperparameters of the multi-scale feature fusion network framework, and use the mean square error between the output denoised data and the input noisy data as the loss function;

[0017] 2.3 Adopting the adaptive moment estimation optimization algorithm, the constructed multi-scale feature fusion network is trained to obtain the multi-scale feature fusion network model;

[0018] 3) Test the multi-scale feature fusion network:

[0019] The noisy seismic data is input into the trained multi-scale feature fusion network model, and the noise data predicted by the network is subtracted from the input data to obtain the denoised data.

[0020] The technical problem solved by the present invention is that in DAS seismic data, complex noise reduces the signal-to-noise ratio of seismic data, seriously destroys the structure and continuity of effective signals, and makes it difficult to identify and recover seismic signals. In order to obtain high-quality seismic data, the present invention combines multi-scale strategy and attention mechanism to propose a DAS seismic noise suppression method based on multi-scale feature fusion network.

[0021] A multi-scale strategy is introduced to obtain the multi-scale features of seismic data and increase the receptive field of the network. To address the problem of uneven noise distribution in seismic data, a multi-scale feature interaction module is introduced. Features of adjacent scales are used as module inputs, and the contextual semantic information contained in lower-scale features is used to achieve adaptive processing of noise with spatial variations. To address the problem of non-stationary noise, an adaptive feature fusion module is introduced to concatenate the scale-adjusted features, use adaptive average pooling to obtain channel statistics, and use linear layers to perform linear transformation on channel statistics to obtain compact channel statistics. Each output is a weighted combination of the inputs, and useful features are dynamically selected from each scale, which enhances the ability to remove complex noise, thereby achieving adaptive fusion of multi-scale features and enhancing the correlation between different scales.

[0022] In view of the limitations of CNN single-scale feature extraction, incomplete DAS seismic noise suppression, incomplete effective signal recovery and destroyed signal structure, the present invention proposes a DAS seismic data noise suppression method based on a multi-scale feature fusion network, which can effectively retain the seismic signal and make the extracted DAS seismic signal structure clearer, with better continuity, higher signal-to-noise ratio and more thorough noise suppression, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flowchart of seismic denoising based on multi-scale feature fusion network;

[0024] Figure 2 It is a schematic diagram of the structure of the multi-scale feature fusion network;

[0025] Figure 3 It is a schematic diagram of the structure of the multi-scale feature interaction module;

[0026] Figure 4 Schematic diagram of the adaptive feature fusion module;

[0027] Figure 5 This is a schematic diagram of actual DAS-VSP noisy seismic data;

[0028] Figure 6 A schematic diagram of denoising results processed by the method of the present invention;

[0029] Figure 7 Schematic diagram of the difference between actual DAS-VSP noisy data and the denoising result of the method of the present invention. DETAILED DESCRIPTION

[0030] The present invention is described below with reference to the accompanying drawings.

[0031] The DAS seismic noise suppression method based on a multi-scale feature fusion network of the present invention comprises the following steps:

[0032] 1) Build a multi-scale feature fusion network framework, including the following steps:

[0033] 1.1 The network consists of three branches. The multi-scale feature fusion network consists of five parts: initial layer module, coarse fusion module, feature enhancement module, fine fusion module and reconstruction module;

[0034] 1.2 Build a shallow feature extraction module. The shallow feature extraction module is embedded in the upper, middle and lower branches to form the initial layer module. The noisy seismic data is taken as input and downsampled through a 3×3 convolution layer with a step size of 2. The feature map output by each convolution layer is passed through two convolution layers to obtain features of different scales X i, i∈{1,2,3}, where i is the branch number, and {1,2,3} correspond to the upper, middle, and lower branches, respectively, corresponding to large-scale features, medium-scale features, and small-scale features;

[0035] 1.3 Build a multi-scale feature interaction module, embed the multi-scale feature interaction module into the upper and middle branches respectively, the upper and middle branches include two multi-scale feature interaction modules, and the lower branch includes two convolutional layers to form a coarse fusion module; the multi-scale feature interaction module is used to input the i-th branch feature X i , i∈{1,2} performs convolution and instance normalization to extract the current scale feature M i , the lower-scale features of the input are transposed convolution to obtain the transposed features M that are consistent with the dimensions of the current scale features i+1 , and with the current scale feature M i Multiply element by element to get the aggregate feature, the aggregate feature and the transposed feature M i+1 After adding element by element, it is combined with the current scale feature X of the input i Add element by element to get the interaction feature I i ; The small-scale features are input into the third branch, and after two convolutional layers, the interactive features I3 are obtained;

[0036] 1.4 Build a hybrid attention module and embed the hybrid attention module into the upper, middle and lower branches to form a feature enhancement module; i , i∈{1,2,3} is used as the input of the hybrid attention module. It is first normalized through the layer, and then passes through the channel attention and spatial attention respectively. The channel attention includes global average pooling, fully connected layer and Sigmod function to obtain the channel attention feature; the spatial attention includes mean operation, convolution and Sigmod function to obtain the spatial attention feature. The channel attention feature and the spatial attention feature are added element by element, and then passed through 3 convolution layers and combined with the interaction feature I i , i∈{1,2,3} are added to obtain the enhanced feature E i , i∈{1,2,3};

[0037] 1.5 Build an adaptive feature fusion module and embed the adaptive feature fusion module into the upper, middle and lower branches respectively. The upper branch includes three adaptive feature fusion modules, and the middle and lower branches include two adaptive feature fusion modules to form a detailed fusion module. i , i∈{1,2,3} is used as the input of the adaptive feature fusion module. First, the enhanced feature E i The dimension is adjusted to the dimension of the branch where the current adaptive feature fusion module is located. The adjusted feature E i', i∈{1,2,3} is concatenated, and the channel statistics are obtained by adaptive average pooling, the compact channel statistics are obtained by linear layer downsampling, and three parallel linear layers are upsampled and the Softmax function is used to obtain three weights, which are respectively compared with the adjusted features E of the three branches. i 'Weighted, the weighted features are added element by element to obtain the fusion feature F i r , i,r∈{1,2,3}, where: r is the number of adaptive feature fusion modules; the final output of the fine fusion module is the fusion feature F1 3 ;

[0038] 1.6 Convolutional layers constitute the reconstruction module, which combines the fusion feature F1 3 As the input of the convolutional layer, the predicted noise data is obtained;

[0039] 2) Optimizing the multi-scale feature fusion network, including the following steps:

[0040] 2.1 Construct a training data set, use the finite difference method to solve the elastic wave equation to construct a clean signal set, and extract noise from the actual data to construct a noise set;

[0041] 2.2 Set the hyperparameters of the multi-scale feature fusion network framework, and use the mean square error between the output denoised data and the input noisy data as the loss function;

[0042] 2.3 Adopting the adaptive moment estimation optimization algorithm, the constructed multi-scale feature fusion network is trained to obtain the multi-scale feature fusion network model;

[0043] 3) Test the multi-scale feature fusion network:

[0044] The noisy seismic data is input into the trained multi-scale feature fusion network model, and the noise data predicted by the network is subtracted from the input data to obtain the denoised data.

[0045] Example

[0046] The effect of the present invention is described below through experiments on synthetic DAS-VSP noisy seismic data and actual DAS-VSP noisy seismic data.

[0047] The experimental platform of the present invention adopts Intel(R) Xeon(R) CPU E5-2678 v3@2.50GHz and anNVIDIA GeForce GTX 1080Ti 411GPU, and the programming language is Python.

[0048] Figure 4 is the actual DAS-VSP noisy seismic data collected, Figure 5The processing results of the actual seismic data using this method show that when the noise is removed to a certain extent, the continuous signal structure will be restored. Figure 6 The difference diagram between the denoising result of the method of the present invention and the actual DAS-VSP noisy seismic data shows that the effective signal is leaked less and the noise is removed more thoroughly. It can be seen that the present invention can better restore the signal structure of DAS-VSP seismic data and remove noise.

Claims

1. A DAS seismic noise suppression method based on a multi-scale feature fusion network, characterized in that: The following steps are involved: 1) Build a multi-scale feature fusion network framework, including the following steps: 1.1 The network consists of three branches. The multi-scale feature fusion network consists of five parts: initial layer module, coarse fusion module, feature enhancement module, fine fusion module and reconstruction module; 1.2 Build a shallow feature extraction module. The shallow feature extraction module is embedded in the upper, middle and lower branches to form the initial layer module. The noisy seismic data is taken as input and downsampled through a 3×3 convolution layer with a step size of 2. The feature map output by each convolution layer is passed through two convolution layers to obtain features of different scales X i , i∈{1,2,3}, where i is the branch number, and {1,2,3} correspond to the upper, middle, and lower branches, respectively, corresponding to large-scale features, medium-scale features, and small-scale features; 1.3 Build a multi-scale feature interaction module, embed the multi-scale feature interaction module into the upper and middle branches respectively, the upper and middle branches include two multi-scale feature interaction modules, and the lower branch includes two convolutional layers to form a coarse fusion module; The multi-scale feature interaction module is used to input the i-th branch feature X i , i∈{1,2} performs convolution and instance normalization to extract the current scale feature M i , the lower-scale input features are transposed and convolved to obtain the transposed features M that are consistent with the current scale feature dimensions i+1 , and with the current scale feature M i Multiply element by element to get the aggregate feature, the aggregate feature and the transposed feature M i+1 After adding element by element, it is combined with the current scale feature X of the input i Add element by element to get the interaction feature I i ; The small-scale features are input into the third branch and after two convolutional layers, the interactive features I3 are obtained; 1.4 Build a hybrid attention module and embed the hybrid attention module into the upper, middle and lower branches to form a feature enhancement module; i , i∈{1,2,3} is used as the input of the hybrid attention module. It is first normalized through the layer, and then passes through the channel attention and spatial attention respectively. The channel attention includes global average pooling, fully connected layer and Sigmod function to obtain the channel attention feature; the spatial attention includes mean operation, convolution and Sigmod function to obtain the spatial attention feature. The channel attention feature and the spatial attention feature are added element by element, and then passed through 3 convolution layers and combined with the interaction feature I i , i∈{1,2,3} are added to obtain the enhanced feature E i , i∈{1,2,3}; 1.5 Build an adaptive feature fusion module and embed the adaptive feature fusion module into the upper, middle and lower branches respectively. The upper branch includes three adaptive feature fusion modules, and the middle and lower branches include two adaptive feature fusion modules to form a detailed fusion module. i , i∈{1,2,3} is used as the input of the adaptive feature fusion module. First, the enhanced feature E i The dimension is adjusted to the dimension of the branch where the current adaptive feature fusion module is located. The adjusted feature E i ', i∈{1,2,3} is concatenated, and the channel statistics are obtained by adaptive average pooling, the compact channel statistics are obtained by linear layer downsampling, and three parallel linear layers are upsampled and the Softmax function is used to obtain three weights, which are respectively compared with the adjusted features E of the three branches. i 'Weighted, the weighted features are added element by element to obtain the fusion feature F i r , i,r∈{1,2,3}, where: r is the number of adaptive feature fusion modules; the final output of the fine fusion module is the fusion feature F1 3 ; 1.6 Convolutional layers constitute the reconstruction module, which combines the fusion feature F1 3 As the input of the convolutional layer, the predicted noise data is obtained; 2) Optimizing the multi-scale feature fusion network, including the following steps: 2.1 Construct a training data set, use the finite difference method to solve the elastic wave equation to construct a clean signal set, and extract noise from the actual data to construct a noise set; 2.2 Set the hyperparameters of the multi-scale feature fusion network framework, and use the mean square error between the output denoised data and the input noisy data as the loss function; 2.3 Adopting the adaptive moment estimation optimization algorithm, the constructed multi-scale feature fusion network is trained to obtain the multi-scale feature fusion network model; 3) Test the multi-scale feature fusion network: The noisy seismic data is input into the trained multi-scale feature fusion network model, and the noise data predicted by the network is subtracted from the input data to obtain the denoised data.

Images