DAS seismic noise suppression method based on semantic information feature fusion network

Abstract

The invention discloses a DAS seismic noise suppression method based on a semantic information feature fusion network, and belongs to the technical field of machine learning and seismic image processing. The method comprises the steps that a lightweight attention module is introduced, self-attention lightweight calculation is achieved by reducing the size of a feature map, and non-local self-similar features of DAS data are effectively captured; semantic information of noisy data is extracted by using a semantic information extraction sub-network, and is fused with self-similar features to better extract seismic signal features, and a subsequent noise reduction network is guided to recover an effective signal form; the local feature reconstruction module and the global feature reconstruction module perform feature reconstruction from two scales to obtain predicted noise; and reconstructing clean DAS VSP data by using a residual learning strategy. According to the method, the suppression effect on various types of noise in the DAS VSP data is remarkable, and effective signals can be better recovered.

Description

Technical Field

[0001] The invention belongs to the technical field of machine learning and seismic image processing, and in particular relates to a DAS seismic noise suppression method based on a semantic information feature fusion network. Background Art

[0002] In recent years, distributed acoustic sensing (DAS) technology has made rapid progress. This method has been successfully applied to various seismic exploration tasks due to its advantages such as full well coverage, high temporal and spatial resolution, and low cost. With the successful application of this method in downhole vertical seismic profiles (VSP), it has greatly promoted the development of land oil and gas exploration towards deep exploration and fine structural analysis. However, limited by environmental conditions and exploration equipment, the acquired DAS VSP data are usually contaminated by various types of complex noise, such as random noise, fading noise, horizontal noise, checkerboard noise, and low-frequency noise. The interference of these noises makes the collected DAS VSP data present a low signal-to-noise ratio, which seriously affects the subsequent imaging and interpretation process. At the same time, the effective DAS VSP signal obtained under complex wave field conditions becomes very weak after multiple reflections and refractions during the transmission process, which increases the difficulty of signal recovery. Therefore, how to recover the weak reflection signal submerged in various types of strong noise has become a key problem in the field of DAS VSP data processing today.

[0003] In order to improve the quality of seismic data, domestic and foreign researchers have done a lot of research in the field of seismic exploration and have proposed many seismic noise suppression methods, such as complex diffusion filtering, dictionary learning, non-local Bayesian filtering, sparse low-rank estimation and non-local mean algorithm. These methods have promoted the development of seismic data processing to a certain extent, but because these methods are heavily dependent on prior information, manual parameter adjustment based on human expertise is required to achieve the best noise reduction performance. The generalization performance for low signal-to-noise ratio data is poor and the computational cost is huge, which limits their scope of application.

[0004] The rapid development of deep learning technology in recent years has inspired many seismic data denoising methods based on convolutional neural networks. Networks such as the feedforward denoising convolutional neural network (DnCNN) and the feature-attention-based image denoising network (RIDNet) can adaptively learn the structural and detail features of signals in seismic data to better suppress noise in seismic data. However, these networks usually only use features on a single scale, and their denoising capabilities are greatly reduced when faced with complex types of noise in DAS VSP data. The present invention suppresses complex types of noise in DAS VSP data by extracting and aggregating different types of features in DAS VSP data. Traditional denoising methods have limitations such as incomplete noise suppression and poor signal amplitude preservation when processing DAS VSP data. Therefore, designing a method that can both efficiently suppress noise and protect effective signals is an important research topic for seismic data denoising. Summary of the invention

[0005] The purpose of the present invention is to provide a method for suppressing DAS VSP seismic noise using a semantic information feature fusion network, so as to overcome the problems that complex types of noise in DAS VSP data destroy effective signals, the signal-to-noise ratio of seismic data is low, and the identification and recovery of seismic signals are very difficult.

[0006] The DAS seismic noise suppression method based on the semantic information feature fusion network of the present invention comprises the following steps:

[0007] 1) Constructing a semantic information feature fusion network, including the following steps:

[0008] 1.1 The semantic information feature fusion network is composed of a cascade of a semantic information extraction subnetwork and a dual-branch denoising subnetwork;

[0009] 1.2 The semantic information extraction subnetwork consists of a semantic information extraction module, a lightweight self-similar attention module, and a feature aggregation module, including the following steps:

[0010] 1.2.1 Constructing semantic information extraction module:

[0011] The noisy seismic data Y∈R H×W Input semantic information extraction module, where H is the height of noisy seismic data and W is the width of noisy seismic data; after Y passes through a U-type encoder-decoder with skip connection, and then passes through the Softmax function, the semantic information of DAS VSP seismic data is obtained. The U-type codec is composed of an encoder and a decoder cascaded. The encoder includes four layers of convolution blocks, each of which is composed of a 3×3 convolution, batch normalization, linear rectification activation function, and a downsampling layer cascaded; the decoder includes four layers of convolution blocks, each of which is composed of a 3×3 convolution, batch normalization, linear rectification activation function, and an upsampling layer cascaded;

[0012] 1.2.2 Constructing a lightweight self-similar attention module:

[0013] The noisy seismic data Y∈R H×W Input the lightweight self-similar attention module, and after 1 C@3×3 convolution, where C is the number of convolution kernels, the shallow feature F∈R is obtained. C×H×W , F p is the feature map on the pth channel of F, where p∈{1,2,...,C}, and F p Divided into d 2 Group q∈{1,2,...,d 2};right Sampling i, j∈{0,1,...,d-1}, m∈{1,2,...,H / d}, n∈{1,2,...,W / d}, change F r Rearrange to 1×HW / d 2 The vector is put into the qth row of the pth channel of M, and then layer normalization is performed to obtain Set a learnable parameter E, obtained by Q = ME and K = ME and The two-dimensional feature map on the μth channel of Q and the two-dimensional feature map on the Kth μ channel Transpose Perform matrix multiplication, pass through the Softmax function and Perform matrix multiplication to obtain weighted features The δth channel z' δ Divide into H / d groups along the column direction, and divide z' δ Put the q'th row of the p'th group in F ss The p'th row of the q'th group in the above example obtains the self-similar attention feature F ss ∈R C×H×W ,Will and F ss Concatenate in the channel dimension to obtain the fusion feature F that combines semantic information and self-similar attention features a ;

[0014] 1.2.3 Construct feature aggregation module:

[0015] F a After 1 C@3×3 convolution, we get F b ∈R C×H×W ; F b Add element-by-element to F in the corresponding channel to get F k ∈R C×H×W , Fk As the input of the feature aggregation module, it goes through global average pooling, C@3×3 convolution, linear rectification activation function, C@3×3 convolution, and Sigmoid function in sequence to obtain the feature F h ∈R C×H×W Finally, F h With F k The elements on the corresponding channels are multiplied point by point and then added to F to obtain the aggregate feature F. c ∈R C×H×W ;

[0016] 1.3 The dual-branch denoising sub-network consists of a local feature reconstruction module and a global feature reconstruction module, including the following steps:

[0017] 1.3.1 Construct local feature reconstruction module:

[0018] F c As the input of the local feature reconstruction module, its output is added element by element to Y to obtain the predicted noise The 4th, 7th, 10th and 13th layers of the local feature reconstruction module are convolutional attention fusion blocks, and the remaining 12 layers are residual blocks; the residual block consists of two C@3×3 convolutions and a linear rectification activation function; the convolutional attention fusion block first normalizes the input layer, then passes through a hybrid attention module consisting of channel attention and spatial attention in parallel, and finally normalizes the input layer again and passes through a multi-scale feedforward network;

[0019] 1.3.2 Constructing global feature reconstruction module:

[0020] F c As the input of the local feature reconstruction module, its output is added element by element to Y to obtain the predicted noise The global feature reconstruction module is composed of 16 cascaded dilated convolutional layers, each of which includes a C@3×3 convolution, batch normalization, and linear rectification activation function;

[0021] 1.3.3 and After adding and passing through a 3×3 convolution, the predicted noise f(Y;θ) is obtained, and the final denoising result is:

[0022]

[0023] Where: Y is the noisy DAS VSP data; θ is the trainable parameter of the network; f represents the proposed semantic information feature fusion network;

[0024] 2) Optimizing the semantic information feature fusion network, including the following steps:

[0025] 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;

[0026] 2.2 Set the loss function L of the semantic information feature fusion network:

[0027] The loss function L of the semantic information feature fusion network is the mean square error loss function L MSE , similarity constraint loss function L SSIM and the semantic constraint loss function L SIE The weighted sum of is expressed as:

[0028] L=L MSE +λ SSIM L SSIM +λ SIE L SIE ;

[0029] 2.2.1 Mean Square Error Loss Function L of Semantic Information Feature Fusion Network MSE , its mathematical expression is:

[0031]

[0032] Where: N i is the noise in the training set data; is the prediction noise of the network; I is the number of noisy-clean training pairs;

[0033] 2.2.2 Similarity Constrained Loss Function L of Semantic Information Feature Fusion Network SSIM , its mathematical expression is:

[0034]

[0035] in: It is expressed as:

[0036]

[0037] Among them: μ represents the average value; σ 2 is the variance; express and N i The covariance coefficient of ; ξ1 and ξ2 are constants;

[0038] 2.2.3 Semantic Constraint Loss Function L of Semantic Information Feature Fusion Network SIE The mathematical expression is:

[0039]

[0040] Where: G i Represents a semantic tag; represents the semantic information predicted by the semantic information extraction subnetwork; ln represents the natural logarithm.

[0041] Technical Problems Solved by the Invention

[0042] In the actual collected DAS VSP data, weak effective signals are often submerged in the DAS noise with strong energy, wide distribution and complex types, making it more difficult to suppress DAS noise, identify and recover weak effective signals. In order to obtain high-quality DAS VSP seismic data, this paper proposes a DAS VSP seismic noise suppression method based on semantic information feature fusion network by extracting the semantic information of the signal and combining it with multi-scale attention.

[0043] The present invention uses a semantic information extraction module to extract the semantic information of effective signals, uses a lightweight attention module to achieve self-attention lightweight calculation by reducing the size of the feature map while capturing the non-local similarity features of the data, and uses a feature aggregation module to fuse the semantic information with the non-local similarity features to guide subsequent denoising; then a dual-branch denoising subnetwork is used for noise prediction, and finally a residual learning strategy is used to reconstruct clean data. The present invention can effectively solve the problem that multiple types of noise in actual DAS VSP data are difficult to suppress and weak reflection signals are difficult to recover.

[0044] In view of the problems that some existing denoising methods based on convolutional neural networks only consider single-scale features, resulting in incomplete noise suppression, poor amplitude preservation of restored effective signals, and discontinuous structure, the present invention proposes a DAS VSP seismic noise suppression method based on a semantic information feature fusion network. The method of the present invention combines semantic information, multi-scale strategy and attention mechanism, can effectively suppress multiple types of noise in DAS VSP data, and make the extracted DAS VSP seismic signal structure clearer, with better continuity and higher signal-to-noise ratio, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a denoising flow chart based on semantic information feature fusion network;

[0046] Figure 2 This is the structural diagram of the semantic information feature fusion network;

[0047] Figure 3 Structural diagram of the subnetwork for semantic information extraction;

[0048] Figure 4 This is a schematic diagram of actual noisy DAS VSP seismic data;

[0049] Figure 5 It is a schematic diagram of the results of processing by the method of the present invention.

[0050] Figure 6 Schematic diagram of the difference between actual DAS VSP data and the denoising result of the method of the present invention. DETAILED DESCRIPTION

[0051] The DAS seismic noise suppression method based on the semantic information feature fusion network of the present invention comprises the following steps:

[0052] 1) Constructing a semantic information feature fusion network, including the following steps:

[0053] 1.1 The semantic information feature fusion network is composed of a cascade of a semantic information extraction subnetwork and a dual-branch denoising subnetwork;

[0054] 1.2 The semantic information extraction subnetwork consists of a semantic information extraction module, a lightweight self-similar attention module, and a feature aggregation module, including the following steps:

[0055] 1.2.1 Constructing semantic information extraction module:

[0056] The noisy seismic data Y∈R H×W Input semantic information extraction module, where H is the height of noisy seismic data and W is the width of noisy seismic data; after Y passes through a U-type encoder-decoder with skip connection, and then passes through the Softmax function, the semantic information of DAS VSP seismic data is obtained. The U-type codec is composed of an encoder and a decoder cascaded. The encoder includes four layers of convolution blocks, each of which is composed of a 3×3 convolution, batch normalization, linear rectification activation function, and a downsampling layer cascaded; the decoder includes four layers of convolution blocks, each of which is composed of a 3×3 convolution, batch normalization, linear rectification activation function, and an upsampling layer cascaded;

[0057] 1.2.2 Constructing a lightweight self-similar attention module:

[0058] The noisy seismic data Y∈R H×W Input the lightweight self-similar attention module, and after 1 C@3×3 convolution, where C is the number of convolution kernels, the shallow feature F∈R is obtained. C×H×W , F p is the feature map on the pth channel of F, where p∈{1,2,...,C}, and F p Divided into d 2 Group q∈{1,2,...,d 2};right Sampling i, j∈{0,1,...,d-1}, m∈{1,2,...,H / d}, n∈{1,2,...,W / d}, change F r Rearrange to 1×HW / d2 The vector is put into the qth row of the pth channel of M, and then layer normalization is performed to obtain Set a learnable parameter E, obtained by Q = ME and K = ME and The two-dimensional feature map on the μth channel of Q and the two-dimensional feature map on the Kth μ channel Transpose Perform matrix multiplication, pass through the Softmax function and Perform matrix multiplication to obtain weighted features The δth channel z' δ Divide into H / d groups along the column direction, and divide z' δ Put the q'th row of the p'th group in F ss The p'th row of the q'th group in the above example obtains the self-similar attention feature F ss ∈R C×H×W ,Will and F ss Concatenate in the channel dimension to obtain the fusion feature F that combines semantic information and self-similar attention features a ;

[0059] 1.2.3 Construct feature aggregation module:

[0060] F a After 1 C@3×3 convolution, we get F b ∈R C×H×W ; F b Add element-by-element to F in the corresponding channel to get F k ∈R C×H×W , F k As the input of the feature aggregation module, it goes through global average pooling, C@3×3 convolution, linear rectification activation function, C@3×3 convolution, and Sigmoid function in sequence to obtain the feature F h ∈R C×H×W Finally, F h With F k The elements on the corresponding channels are multiplied point by point and then added to F to obtain the aggregate feature F. c ∈R C×H×W ;

[0061] 1.3 The dual-branch denoising sub-network consists of a local feature reconstruction module and a global feature reconstruction module, including the following steps:

[0062] 1.3.1 Construct local feature reconstruction module:

[0063] F c As the input of the local feature reconstruction module, its output is added element by element to Y to obtain the predicted noise The 4th, 7th, 10th and 13th layers of the local feature reconstruction module are convolutional attention fusion blocks, and the remaining 12 layers are residual blocks; the residual block consists of two C@3×3 convolutions and a linear rectification activation function; the convolutional attention fusion block first normalizes the input layer, then passes through a hybrid attention module consisting of channel attention and spatial attention in parallel, and finally normalizes the input layer again and passes through a multi-scale feedforward network;

[0064] 1.3.2 Constructing global feature reconstruction module:

[0065] F c As the input of the local feature reconstruction module, its output is added element by element to Y to obtain the predicted noise The global feature reconstruction module is composed of 16 cascaded dilated convolutional layers, each of which includes a C@3×3 convolution, batch normalization, and linear rectification activation function;

[0066] 1.3.3 and After adding and passing through a 3×3 convolution, the predicted noise f(Y;θ) is obtained, and the final denoising result is:

[0067]

[0068] Where: Y is the noisy DAS VSP data; θ is the trainable parameter of the network; f represents the proposed semantic information feature fusion network;

[0069] 2) Optimizing the semantic information feature fusion network, including the following steps:

[0070] 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;

[0071] 2.2 Set the loss function L of the semantic information feature fusion network:

[0072] The loss function L of the semantic information feature fusion network is the mean square error loss function L MSE , similarity constraint loss function L SSIM and the semantic constraint loss function L SIE The weighted sum of is expressed as:

[0073] L=L MSE +λ SSIM L SSIM +λ SIE L SIE ;

[0074] 2.2.1 Mean Square Error Loss Function L of Semantic Information Feature Fusion Network MSE, its mathematical expression is:

[0076]

[0077] Where: N i is the noise in the training set data; is the prediction noise of the network; I is the number of noisy-clean training pairs;

[0078] 2.2.2 Similarity Constrained Loss Function L of Semantic Information Feature Fusion Network SSIM , its mathematical expression is:

[0079]

[0080] in: It is expressed as:

[0081]

[0082] Among them: μ represents the average value; σ 2 is the variance; express and N i The covariance coefficient of ; ξ1 and ξ2 are constants;

[0083] 2.2.3 Semantic Constraint Loss Function L of Semantic Information Feature Fusion Network SIE The mathematical expression is:

[0084]

[0085] Where: G i Represents a semantic tag; represents the semantic information predicted by the semantic information extraction subnetwork; ln represents the natural logarithm.

[0086] Example

[0087] The effect of the present invention is described below through experiments on actual DAS VSP seismic data.

[0088] The experimental platform of the present invention adopts the 13th generation Intel Core i9-13900HX CPU and NVIDIA GeForceRTX 4060GPU, and the language is Python.

[0089] Figure 4 is the actual DAS VSP seismic data collected, Figure 5 The denoising results of the actual seismic data using the method of the present invention are shown in Figure 2. It can be seen that the method of the present invention can effectively remove various types of noise in DAS VSP data and can well restore the weak reflection signal in DAS VSP data. Figure 4 Actual DAS VSP data and Figure 5 The denoised data shown in the figure is subtracted to obtain the corresponding difference image, such as Figure 6 As shown, it can be seen that there is very little effective signal remaining and the signal has a good amplitude preservation performance.

Claims

1. A DAS seismic noise suppression method based on a semantic information feature fusion network, characterized in that: The following steps are involved: 1) Constructing a semantic information feature fusion network, including the following steps: 1.1 The semantic information feature fusion network is composed of a cascade of a semantic information extraction subnetwork and a dual-branch denoising subnetwork; 1.2 The semantic information extraction subnetwork consists of a semantic information extraction module, a lightweight self-similar attention module, and a feature aggregation module, including the following steps: 1.2.1 Constructing semantic information extraction module: The noisy seismic data Y∈R H×W Input semantic information extraction module, where H is the height of noisy seismic data and W is the width of noisy seismic data; after Y passes through a U-type encoder-decoder with skip connection, and then passes through the Softmax function, the semantic information of DAS VSP seismic data is obtained. The U-type codec is composed of an encoder and a decoder cascaded. The encoder includes four layers of convolution blocks, each of which is composed of a 3×3 convolution, batch normalization, linear rectification activation function, and a downsampling layer cascaded; the decoder includes four layers of convolution blocks, each of which is composed of a 3×3 convolution, batch normalization, linear rectification activation function, and an upsampling layer cascaded; 1.2.2 Constructing a lightweight self-similar attention module: The noisy seismic data Y∈R H×W Input the lightweight self-similar attention module, and after 1 C@3×3 convolution, where C is the number of convolution kernels, the shallow feature F∈R is obtained. C×H×W , F p is the feature map on the pth channel of F, where p∈{1,2,...,C}, and F p Divided into d 2 Group right Sampling F r Rearrange to 1×HW / d 2 The vector is put into the qth row of the pth channel of M, and then layer normalization is performed to obtain Set a learnable parameter E, obtained by Q = ME and K = ME and The two-dimensional feature map on the μth channel of Q and the two-dimensional feature map on the Kth μ channel Transpose Perform matrix multiplication, pass through the Softmax function and Perform matrix multiplication to obtain weighted features The δth channel z' δ Divide into H / d groups along the column direction, and divide z' δ Put the q'th row of the p'th group in F ss The p'th row of the q'th group in the above example obtains the self-similar attention feature F ss ∈R C×H×W ,Will and F ss Concatenate in the channel dimension to obtain the fusion feature F that combines semantic information and self-similar attention features a ; 1.2.3 Construct feature aggregation module: F a After 1 C@3×3 convolution, we get F b ∈R C×H×W ; F b Add element-by-element to F in the corresponding channel to get F k ∈R C×H×W , F k As the input of the feature aggregation module, it goes through global average pooling, C@3×3 convolution, linear rectification activation function, C@3×3 convolution, and Sigmoid function in sequence to obtain the feature F h ∈R C×H×W Finally, F h With F k The elements on the corresponding channels are multiplied point by point and then added to F to obtain the aggregate feature F. c ∈R C×H×W ; 1.3 The dual-branch denoising sub-network consists of a local feature reconstruction module and a global feature reconstruction module, including the following steps: 1.3.1 Construct local feature reconstruction module: F c As the input of the local feature reconstruction module, its output is added element by element to Y to obtain the predicted noise The 4th, 7th, 10th and 13th layers of the local feature reconstruction module are convolutional attention fusion blocks, and the remaining 12 layers are residual blocks; the residual block consists of two C@3×3 convolutions and a linear rectification activation function; the convolutional attention fusion block first normalizes the input layer, then passes through a hybrid attention module consisting of channel attention and spatial attention in parallel, and finally normalizes the input layer again and passes through a multi-scale feedforward network; 1.3.2 Constructing global feature reconstruction module: F c As the input of the local feature reconstruction module, its output is added element by element to Y to obtain the predicted noise The global feature reconstruction module is composed of 16 cascaded dilated convolutional layers, each of which includes a C@3×3 convolution, batch normalization, and linear rectification activation function; 1.3.3 and After adding and passing through a 3×3 convolution, the predicted noise f(Y;θ) is obtained, and the final denoising result is: Where: Y is the noisy DAS VSP data; θ is the trainable parameter of the network; f represents the proposed semantic information feature fusion network; 2) Optimizing the semantic information 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 loss function L of the semantic information feature fusion network: The loss function L of the semantic information feature fusion network is the mean square error loss function L MSE , similarity constraint loss function L SSIM and the semantic constraint loss function L SIE The weighted sum of is expressed as: L=L MSE +λ SSIM L SSIM +λ SIE L SIE ; 2.2.1 Mean Square Error Loss Function L of Semantic Information Feature Fusion Network MSE , its mathematical expression is: in: N i is the noise in the training set data; is the prediction noise of the network; I is the number of noisy-clean training pairs; 2.2.2 Similarity Constrained Loss Function L of Semantic Information Feature Fusion Network SSIM , its mathematical expression is: in: It is expressed as: Among them: μ represents the average value; σ 2 is the variance; express and N i The covariance coefficient of ; ξ1 and ξ2 are constants; 2.2.3 Semantic Constraint Loss Function L of Semantic Information Feature Fusion Network SIE The mathematical expression is: Where: G i Represents a semantic tag; represents the semantic information predicted by the semantic information extraction subnetwork; ln represents the natural logarithm.

Images