A desert seismic noise suppression method based on a multi-scale attention interaction network

Abstract

This invention relates to a desert earthquake noise suppression method based on a multi-scale attention interaction network, belonging to the fields of machine learning and seismic image processing. Addressing the problem of low-frequency noise suppression in desert earthquake images, this invention proposes a multi-scale attention interaction network. This network first downsamples the input seismic data, then adaptively learns the complex features of the multi-scale seismic data using dual-branch convolutional layers with different kernel sizes. It then employs multiple attention mechanisms, including permutation attention and coordinate attention, to integrate and interact with multi-scale features across channel, spatial, and coordinate dimensions to obtain continuous seismic signals. This method uses a hybrid loss function of mean squared error and mean cosine similarity, combining the multi-scale strategy with the attention mechanism to improve the denoising effect of desert earthquake images. Compared with single-scale denoising convolutional neural networks, this invention significantly suppresses desert noise and outperforms single-scale denoising convolutional neural networks in restoring the continuity of the phase axis.

Description

Technical Field

[0001] This invention belongs to the field of machine learning and earthquake image processing technology, specifically relating to a desert earthquake noise suppression method based on a multi-scale attention interaction network. Background Technology

[0002] Seismic exploration is a crucial method for finding oil and gas resources. Currently, due to the complexity of geological structures and limitations in acquisition technology, the obtained seismic data often contains strong and complex random noise. This irregular and unpredictable random noise severely damages the effective signal. The key and challenge of seismic signal processing is to distinguish noise from the effective signal as much as possible from the acquired noisy seismic images, thereby accurately extracting the effective seismic signal and improving the signal-to-noise ratio and resolution of the seismic data. Especially in desert regions, desert random noise has similar waveforms and overlapping spectra with the effective signal, greatly increasing the difficulty of seismic signal identification and extraction. Therefore, noise suppression is of great significance for seismic exploration.

[0003] To improve the quality of seismic data, researchers both domestically and internationally have conducted extensive studies in the field of seismic exploration, continuously proposing a series of seismic denoising methods and theories, such as complex diffusion filtering, dictionary learning, nonlocal Bayesian filtering, sparse low-rank estimation, and nonlocal mean algorithms. These methods have improved the quality of seismic images to some extent. However, under conditions of low signal-to-noise ratio and spatiotemporally varying random seismic noise, the denoising effects of different seismic records still vary significantly. Moreover, obtaining the best denoising effect for processing large amounts of seismic data requires a considerable amount of time. These denoising methods have failed to fully exploit the complex structural and detailed features of seismic signals, resulting in low continuity of the effective signal and unclear details while suppressing noise. In recent years, the rapid development of convolutional neural networks (CNNs) has spurred the emergence of many effective seismic data denoising methods. Networks such as feedforward denoising convolutional neural networks (DnCNN) and residual encoding / decoding networks (REDNet) adaptively learn the structural and detailed features of signals in seismic images, better suppressing low-frequency random noise in seismic images. Although these CNN-based methods have superior performance compared to traditional methods, they typically only utilize features at a single scale, limiting the network's denoising capabilities. Summary of the Invention

[0004] Seismic data contains a large amount of random noise. The random noise studied in this invention is a non-stationary colored noise, whose energy is mainly concentrated in the low-frequency band of 0-20Hz and overlaps with the effective signal in the spectrum. Random noise interferes with the effective signal, making it difficult to identify. Some supervised networks often only consider single-scale features and ignore multi-scale features, resulting in low signal continuity during the denoising process.

[0005] The purpose of this invention is to provide a desert earthquake noise suppression method based on a multi-scale attention interaction network. First, dual-branch convolutional layers with different kernel sizes are used to adaptively capture multi-scale features of seismic data. Then, a multi-attention mechanism including permutation attention and coordinate attention is employed. During the iteration process, the network gradually interacts and integrates multi-scale features. By combining multi-scale strategies and attention mechanisms, low-frequency noise in desert earthquake images can be effectively suppressed, and the recovered signal can be made more continuous. This invention can achieve excellent denoising performance while recovering seismic signals with complex morphology.

[0006] The desert earthquake noise suppression method based on a multi-scale attention interaction network of the present invention includes the following steps:

[0007] 1) Constructing a multi-scale attention interaction network: The constructed multi-scale attention interaction network consists of a multi-scale feature extraction module, a multi-scale feature fusion module, and a reconstruction module;

[0008] 1.1) The multi-scale feature extraction module consists of a downsampling block and a residual multi-scale block. The downsampling block represents sampling at intervals in rows and columns, transforming the input data into sub-data of half the size. The residual multi-scale block consists of convolutional layers of different sizes with two branches. Specifically, 1×3 and 3×1 convolutions and ReLU operations are performed respectively. The information from the two branches is cross-fused and sent to the 3×1 convolution and 1×3 convolution respectively. Then, the ReLU operation is performed. The outputs of the two branches are then connected and sent to a 1×1 convolutional layer. Finally, the input data is added element-wise.

[0009] 1.2) The multi-scale feature fusion module consists of permutation attention and coordinate attention; the first layer of coordinate attention is a pooling kernel in the horizontal and vertical directions, the second layer is a connection operation and a convolutional layer, the third layer is a batch normalization (BN) layer and a nonlinear layer, the fourth layer is a convolutional layer, and the fifth layer is a sigmoid function.

[0010] 1.3) The reconstruction module consists of convolutional layers and inverse downsampling blocks; there are 7 convolutional layers in total. Layers 1 to 6 include convolution, BN and ReLU operations, while layer 7 only performs convolution operations.

[0011] 2) Given a noisy seismic exploration image Y = X + V, where: X represents a clean seismic image; V represents low-frequency colored noise;

[0012] Suppressing low-frequency color noise in noisy seismic data using a multi-scale attention interaction network includes the following steps:

[0013] 2.1) Use the noisy seismic exploration image Y as the input to the network;

[0014] 2.2) The input data is decomposed into four sub-data points of half their original size using downsampling blocks;

[0015] 2.3) Input the sub-data into a parallel residual multi-scale block to obtain the output M;

[0016] 2.4) Divide M into G groups along the channel dimension: M = [M1, ..., M2] G For the Kth feature M in M k First, we divide it into two smaller sub-features along the channel dimension, namely M. k1 M k2 The signal is processed through the channel attention module and the spatial attention module respectively, resulting in the output M. k1 'and M k2 Finally, connect M in the channel direction. k1 'and M k2 Perform a channel shuffle operation to obtain the output N; in the channel attention module, first use global average pooling to generate channel statistics, and then use the sigmoid function to multiply them by the input feature M. k1 Obtain the final output M of the channel attention module. k1 In the spatial attention module, group normalization is used for M. k2 The data is processed to obtain spatial-level statistical information, then enhanced using Fc(·) = Wx + b, where W and b are the scaling and offset parameters, respectively. Finally, it is multiplied by M. k2 Obtain the final output M of the spatial attention module. k2 ';

[0017] 2.5) Using pooling kernels of size (H, 1) and (1, W) along the horizontal and vertical directions, information from different channels of N is encoded to obtain N. x and N y Then connect N x and N y The data is fed into a shared 1×1 convolutional layer, then processed using a non-linear activation function and a sigmoid function. The resulting feature maps are used as attention weights for the horizontal and vertical coordinates, respectively, and finally reweighted by N.

[0018] 2.6) In the reconstruction module, convolutional layers and inverse downsampling blocks are used to obtain a denoised, clean desert earthquake image.

[0019] The denoising process in steps 2.1) to 2.6) is mapped by the network. express;

[0020] 3) Constructing signal and noise sets: Analyzing the structural characteristics of actual desert seismic signals, 64 clean desert seismic images of 2000×200 pixels were generated using the Ricker wavelet simulation, resulting in a clean signal training set {X}. i}; Prepare noise set {V i This refers to actual desert noise, which includes 480 channels, each with 30,000 sampling points;

[0021] 4) Train the network using the signal set and noise set to learn the parameters Θ of network D, including the following steps:

[0022] 4.1) During training, first {X} i Divide the signal into 80×80 blocks with a step size of 20, and then perform amplitude normalization on the signal blocks to obtain the training signal block set; similarly, the actual desert noise {V} is divided into blocks. i The noise data is cut into 80×80 blocks, then normalized based on the maximum amplitude of each noise block. A random number in the range (0, 2) is then selected as a weight and multiplied with the noise block to simulate different noise levels, thus obtaining a noise data block. Finally, the signal block and the noise block are superimposed to obtain the noisy data block {Y}. i};

[0023] 4.2) Train a multi-scale attention interaction network to establish a network that takes a noisy seismic image Y as input. i to network output mapping Where: Θ={w,b} are the network parameters, and w and b are the network weights and biases, respectively;

[0024] 4.3) Using a hybrid loss function of mean squared error and mean cosine similarity, the network parameters Θ are learned, and the network output is... The formula for calculating the mean cosine similarity (ACS) is:

[0025]

[0026] The formula for calculating the network loss function is:

[0027]

[0028] Where: α is the weighting parameter; M represents the noisy-clean training set, and the Θ value is the final network parameter when L(Θ) reaches its minimum.

[0029] The beneficial effects of this invention are as follows: In seismic data, random noise reduces the signal-to-noise ratio, severely contaminates the effective signal, and makes the identification and recovery of seismic signals difficult. To obtain high-quality seismic data, this invention proposes a desert seismic noise suppression method based on a multi-scale attention interaction network, combining multi-scale strategies and attention mechanisms. Downsampled seismic data is passed through a dual-branch residual multi-scale block, with different branches using convolutional kernels of different sizes to capture the multi-scale features of the seismic data. Then, a multiple attention mechanism, including permutation attention and coordinate attention, is used to gradually interact with multi-scale features from shallow to deep layers of the network at the channel, spatial, and coordinate levels, thereby identifying reflected signals from seismic noise. This invention employs a hybrid loss function of mean square error and mean cosine similarity to adjust the learning of network parameters, thereby better recovering the continuous signal structure under noise suppression. Attached Figure Description

[0030] Figure 1 This is a structural diagram of a multi-scale attention interaction network;

[0031] Figure 2 This is a structural diagram of the residual multi-scale block;

[0032] Figure 3 A structural diagram for attention displacement;

[0033] Figure 4 A structural diagram of coordinate attention;

[0034] Figure 5 A schematic diagram of a clean seismic record;

[0035] Figure 6 A schematic diagram of random noise in the desert;

[0036] Figure 7 To be Figure 3 The desert random noise shown Figure 2 A schematic diagram of noise-contaminated earthquake records obtained from [the source].

[0037] Figure 8 This is a schematic diagram of the denoising results using DnCNN;

[0038] Figure 9 This is a schematic diagram of the denoising result achieved using the method of the present invention;

[0039] Figure 10 A schematic diagram showing the difference between clean seismic data and the denoised result of DnCNN;

[0040] Figure 11 A schematic diagram showing the difference between clean seismic data and the denoised result of the method of this invention;

[0041] Figure 12 This is a schematic diagram of actual noisy seismic data;

[0042] Figure 13 This is a schematic diagram showing the result of the method of the present invention. Detailed Implementation

[0043] The present invention will now be described in conjunction with the accompanying drawings.

[0044] The desert earthquake noise suppression method based on a multi-scale attention interaction network of the present invention includes the following steps:

[0045] 1) Constructing a multi-scale attention interaction network: The constructed multi-scale attention interaction network consists of a multi-scale feature extraction module, a multi-scale feature fusion module, and a reconstruction module;

[0046] 1.1) The multi-scale feature extraction module consists of a downsampling block (DB) and a residual multi-scale block (RMB). The downsampling block represents sampling at intervals in rows and columns, transforming the input data into sub-data of half the size. The residual multi-scale block consists of convolutional layers of different sizes with two branches. Specifically, 1×3 and 3×1 convolutions and ReLU operations are performed respectively. The information from the two branches is cross-fused and sent to the 3×1 convolution and 1×3 convolution respectively. Then, the ReLU operation is performed. The outputs of the two branches are then connected and sent to a 1×1 convolutional layer. Finally, the input data is added element-wise.

[0047] 1.2) The multi-scale feature fusion module consists of permutation attention (SAB) and coordinate attention (CAB); the first layer of coordinate attention is a pooling kernel in the horizontal and vertical directions, the second layer is a connection operation and a convolutional layer, the third layer is a batch normalization (BN) layer and a nonlinear layer, the fourth layer is a convolutional layer, and the fifth layer is a sigmoid function.

[0048] 1.3) The reconstruction module consists of convolutional layers and inverse downsampling blocks; there are 7 convolutional layers in total. Layers 1 to 6 include convolution, BN and ReLU operations, while layer 7 only performs convolution operations.

[0049] 2) Given a noisy seismic exploration image Y = X + V, where: X represents a clean seismic image; V represents low-frequency colored noise;

[0050] Suppressing low-frequency color noise in noisy seismic data using a multi-scale attention interaction network includes the following steps:

[0051] 2.1) Use the noisy seismic exploration image Y as the input to the network;

[0052] 2.2) The input data is decomposed into four sub-data points of half their original size using downsampling blocks;

[0053] 2.3) Input the sub-data into a parallel residual multi-scale block to obtain the output M;

[0054] 2.4) Divide M into G groups along the channel dimension: M = [M1, ..., M2] G For the Kth feature M in M k First, we divide it into two smaller sub-features along the channel dimension, namely M. k1 M k2 The signal is processed through the channel attention module and the spatial attention module respectively, resulting in the output M. k1 'and M k2 Finally, connect M in the channel direction. k1 'and M k2 Perform a channel shuffle operation to obtain the output N; in the channel attention module, first use global average pooling to generate channel statistics, and then use the sigmoid function to multiply them by the input feature M. k1 Obtain the final output M of the channel attention module. k1 In the spatial attention module, group normalization is used for M. k2 The data is processed to obtain spatial-level statistical information, then enhanced using Fc(·) = Wx + b, where W and b are the scaling and offset parameters, respectively. Finally, it is multiplied by M. k2 Obtain the final output M of the spatial attention module. k2 ';

[0055] 2.5) Using pooling kernels of size (H, 1) and (1, W) along the horizontal and vertical directions, information from different channels of N is encoded to obtain N. x and N y Then connect N x and N y The data is fed into a shared 1×1 convolutional layer, then processed using a non-linear activation function and a sigmoid function. The resulting feature maps are used as attention weights for the horizontal and vertical coordinates, respectively, and finally reweighted by N.

[0056] 2.6) In the reconstruction module, convolutional layers and inverse downsampling blocks (IDRB) are used to obtain a denoised, clean desert seismic image.

[0057] The denoising process in steps 2.1) to 2.6) is mapped by the network. express;

[0058] 3) Constructing signal and noise sets: Analyzing the structural characteristics of actual desert seismic signals, 64 clean desert seismic images of 2000×200 pixels were generated using the Ricker wavelet simulation, resulting in a clean signal training set {X}. i}; Prepare noise set {Vi This refers to actual desert noise, which includes 480 channels, each with 30,000 sampling points;

[0059] 4) Train the network using the signal set and noise set to learn the parameters Θ of network D, including the following steps:

[0060] 4.1) During training, first {X} i Divide the signal into 80×80 blocks with a step size of 20, and then perform amplitude normalization on the signal blocks to obtain the training signal block set; similarly, the actual desert noise {V} is divided into blocks. i The noise data is cut into 80×80 blocks, then normalized based on the maximum amplitude of each noise block. A random number in the range (0, 2) is then selected as a weight and multiplied with the noise block to simulate different noise levels, thus obtaining a noise data block. Finally, the signal block and the noise block are superimposed to obtain the noisy data block {Y}. i};

[0061] 4.2) Train a multi-scale attention interaction network to establish a network that takes a noisy seismic image Y as input. i to network output mapping Where: Θ={w,b} are the network parameters, and w and b are the network weights and biases, respectively;

[0062] 4.3) Using a hybrid loss function of mean squared error and mean cosine similarity, the network parameters Θ are learned, and the network output is... The formula for calculating the mean cosine similarity (ACS) is:

[0063]

[0064] The formula for calculating the network loss function is:

[0065]

[0066] Where: α is the weighting parameter; M represents the noisy-clean training set, and the Θ value is the final network parameter when L(Θ) reaches its minimum.

[0067] Example

[0068] The effects of this invention will be illustrated below through experiments using simulated earthquake data and actual earthquake data.

[0069] The experimental platform of this invention uses an Intel(R) Xeon(R) CPU E5-2678 v3@2.50GHz and an NVIDIA GeForce GTX 1080Ti 411 GPU, and the programming language is Python.

[0070] Synthetic clean seismic data, such as Figure 5 As shown, there are a total of 100 channels, each containing 800 sampling points, consisting of 8 seismic hyperbolic events with dominant frequencies of 25, 24, 22, 20, 18, 16, and 15 Hz. Desert random noise is as follows: Figure 6 As shown. Figure 7 To be Figure 6 Join Figure 5 The desert earthquake data obtained was contaminated by desert noise (signal-to-noise ratio of -9dB). In this embodiment, the denoising result of the method of the present invention was tested and compared with the denoising result of DnCNN. The denoising result is as follows: Figure 8 and Figure 9 As shown. Comparing the denoising results, it can be found that the method of the present invention makes the background clearer and the recovered seismic events clearer and more continuous. Furthermore, Figure 8 and Figure 9 The denoising results and Figure 5 The difference image is obtained by subtracting the clean image shown, as follows: Figure 10 and Figure 11 As shown. The method of the present invention has minimal signal residue. Figure 12 For the noisy seismic data collected, Figure 13 The results of processing actual seismic data using this method show that noise is removed to a certain extent and the continuous signal structure is recovered.

Claims

1. A desert earthquake noise suppression method based on a multi-scale attention interaction network, comprising the following steps: 1) Constructing a multi-scale attention interaction network: The constructed multi-scale attention interaction network consists of a multi-scale feature extraction module, a multi-scale feature fusion module, and a reconstruction module; 1.1) The multi-scale feature extraction module consists of a downsampling block and a residual multi-scale block. The downsampling block represents sampling at intervals in rows and columns, transforming the input data into sub-data of half the size. The residual multi-scale block consists of convolutional layers of different sizes with two branches. Specifically, 1×3 and 3×1 convolutions and ReLU operations are performed respectively. The information from the two branches is cross-fused and sent to the 3×1 convolution and 1×3 convolution respectively. Then, the ReLU operation is performed. The outputs of the two branches are then connected and sent to a 1×1 convolutional layer. Finally, the input data is added element-wise. 1.2) The multi-scale feature fusion module consists of permutation attention and coordinate attention; the first layer of coordinate attention is a pooling kernel in the horizontal and vertical directions, the second layer is a connection operation and a convolutional layer, the third layer is a batch normalization (BN) layer and a nonlinear layer, the fourth layer is a convolutional layer, and the fifth layer is a sigmoid function. 1.3) The reconstruction module consists of convolutional layers and inverse downsampling blocks; there are 7 convolutional layers in total. Layers 1 to 6 include convolution, BN and ReLU operations, while layer 7 only performs convolution operations. 2) Given a noisy seismic exploration image Y = X + V, where: X represents a clean seismic image; V represents low-frequency color noise; Suppressing low-frequency color noise in noisy seismic data using a multi-scale attention interaction network includes the following steps: 2.1) Use the noisy seismic exploration image Y as the input to the network; 2.2) The input data is decomposed into four sub-data points of half their original size using downsampling blocks; 2.3) Input the sub-data into a parallel residual multi-scale block to obtain the output M; 2.4) Divide M into G groups along the channel dimension: M = [M1, ..., M2] G For the Kth feature M in M k First, we divide it into two smaller sub-features along the channel dimension, namely M. k1 M k2 The signal is processed through the channel attention module and the spatial attention module respectively, resulting in the output M. k1 'and M k2 Finally, connect M in the channel direction. k1 'and M k2 Perform a channel shuffle operation to obtain the output N; in the channel attention module, first use global average pooling to generate channel statistics, and then use the sigmoid function to multiply them by the input feature M. k1 Obtain the final output M of the channel attention module. k1 In the spatial attention module, group normalization is used for M. k2 The data is processed to obtain spatial-level statistical information, then enhanced using Fc(·) = Wx + b, where W and b are the scaling and offset parameters, respectively. Finally, it is multiplied by M. k2 Obtain the final output M of the spatial attention module. k2 '; 2.5) Using pooling kernels of size (H, 1) and (1, W) along the horizontal and vertical directions, information from different channels of N is encoded to obtain N. x and N y Then connect N x and N y The data is fed into a shared 1×1 convolutional layer, then processed using a non-linear activation function and a sigmoid function. The resulting feature maps are used as attention weights for the horizontal and vertical coordinates, respectively, and finally reweighted by N. 2.6) In the reconstruction module, convolutional layers and inverse downsampling blocks are used to obtain a denoised, clean desert earthquake image. The denoising process in steps 2.1) to 2.6) is mapped by the network. express; 3) Constructing signal and noise sets: Analyzing the structural characteristics of actual desert seismic signals, 64 clean desert seismic images of 2000×200 pixels were generated using the Ricker wavelet simulation, resulting in a clean signal training set {X}. i }; Prepare noise set {V i This refers to actual desert noise, which includes 480 channels, each with 30,000 sampling points; 4) Train the network using the signal set and noise set to learn the parameters Θ of network D, including the following steps: 4.1) During training, first {X} i Divide the signal into 80×80 blocks with a step size of 20, and then perform amplitude normalization on the signal blocks to obtain the training signal block set; similarly, the actual desert noise {V} is divided into blocks. i The noise data is cut into 80×80 blocks, then normalized based on the maximum amplitude of each noise block. A random number in the range (0, 2) is then selected as a weight and multiplied with the noise block to simulate different noise levels, thus obtaining a noise data block. Finally, the signal block and the noise block are superimposed to obtain the noisy data block {Y}. i }; 4.2) Train a multi-scale attention interaction network to establish a network that takes a noisy seismic image Y as input. i to network output mapping Where: Θ={w,b} are the network parameters, and w and b are the network weights and biases, respectively; 4.3) Using a hybrid loss function of mean squared error and mean cosine similarity, the network parameters Θ are learned, and the network output is... The formula for calculating the mean cosine similarity (ACS) is: The formula for calculating the network loss function is: Where: α is the weighting parameter; M represents the noisy-clean training set, and the Θ value is the final network parameter when L(Θ) reaches its minimum.

Images