Earthquake intelligent filtering method based on multi-scale information fusion and structural feature enhancement

Abstract

The invention discloses an earthquake intelligent filtering method based on multi-scale information fusion and structural feature enhancement, and belongs to the technical field of geophysical exploration data processing, and the method comprises the following steps: obtaining original earthquake data, carrying out the normalization processing of the data, and obtaining the preprocessing earthquake data; performing anisotropic diffusion filtering processing on the original seismic data based on an anisotropic diffusion equation to obtain structure enhanced seismic data; constructing a parallel convolutional neural network, and performing feature extraction on the preprocessed seismic data and the structure enhanced seismic data to obtain a multi-scale seismic feature image; based on a feature fusion neural network, performing cross-scale fusion on the multi-scale feature image to generate a fused feature image; and constructing and training a residual neural network MA-ResCNN based on multi-scale fusion and an attention mechanism. The method can effectively remove the noise interference of the exploration seismic data in a complex structure region, and improves the resolution and the signal-to-noise ratio of the seismic data.

Description

Technical Field

[0001] The present invention relates to the technical field of geophysical exploration data processing, and in particular to a seismic intelligent filtering method based on multi-scale information fusion and structural feature enhancement, which is suitable for noise suppression and geological feature recovery of seismic data in complex structural areas. Background Art

[0002] Seismic data processing is a core component of oil and gas exploration and development, where noise suppression and effective signal enhancement directly impact the accuracy of subsequent structural interpretation and reservoir inversion. As exploration targets shift toward deep, complex structural zones, seismic data generally exhibit low signal-to-noise ratios, multi-scale noise interference, and non-stationary signal characteristics. This is particularly true in complex geological areas such as fault zones and thin interbedded areas. Traditional seismic filtering methods (such as singular value decomposition and fx-domain filtering) are prone to effective signal distortion or structural ambiguity in complex structural zones, severely restricting the reliability of geological interpretation of seismic data. While existing deep learning models (such as convolutional neural networks) have made progress, they fail to fully exploit the multi-scale correlations of seismic data and prior knowledge of geological structures during training, leading to secondary problems such as weakened thin-layer reflection characteristics and blurred fault edges. Summary of the Invention

[0003] Technical problems solved: In view of the problems that existing seismic filtering methods lack the ability to model targeted multi-scale noise interference, insufficient recovery of small-scale structures, and lack of effective recovery of discontinuous geological features such as faults and pinch-out points during the denoising process, the present invention provides a seismic intelligent filtering method based on multi-scale information fusion and structural feature enhancement. Through multi-scale convolution feature extraction, structural guided filtering and attention mechanism fusion, the signal-to-noise ratio and geological interpretation reliability of seismic data in complex structural areas are improved, and the edge retention ability of key structural features such as fault systems and stratigraphic unconformity surfaces is enhanced, providing high-quality seismic data for the fine characterization of oil and gas reservoirs.

[0004] Technical solution: The seismic intelligent filtering method based on multi-scale information fusion and structural feature enhancement described in the present invention includes the following steps: Step 1: Obtain the original seismic data, separate the effective frequency band signal through bandpass filtering, retain the seismic data in the main frequency range, and normalize the data to obtain preprocessed seismic data; Step 2: Based on the anisotropic diffusion equation, the diffusion coefficient is calculated according to the direction of the seismic phase axis, and the original seismic data is processed by anisotropic diffusion filtering to obtain structural enhancement seismic data; Step 3: Construct a parallel convolutional neural network and use three different sizes of convolution kernels to extract features from the preprocessed seismic data and the structurally enhanced seismic data to obtain multi-scale seismic feature images; Step 4: Based on the feature fusion neural network, the channel attention mechanism is used to perform cross-scale fusion of multi-scale feature images to generate a fused feature image; Step 5: Construct a residual neural network, input the fused feature image and combine it with the residual block to associate the preprocessed seismic data and the structural enhanced seismic data to form a residual neural network MA-ResCNN based on multi-scale fusion and attention mechanism; Step 6: Use earthquake label data to train the MA-ResCNN model end-to-end, input the observed earthquake data into the model, and output the final filtered earthquake data.

[0005] Preferably, the bandpass filtering in step 1 adopts a Butterworth filter, the cutoff frequency of which is set according to the main frequency range of the effective signal, and the time domain seismic data of the main frequency band is reconstructed by inverse Fourier transform; the normalization processing adopts the maximum minimum normalization algorithm, and the calculation formula for the preprocessed seismic data is: ; Where: d norm To pre-process seismic data; d rec is the main frequency band time domain seismic data; d min and d max are the minimum and maximum values of the seismic data amplitude, respectively.

[0006] Preferably, the partial differential equation of the anisotropic diffusion filtering in step 2 is: ; Where f is the diffusion function used to calculate the diffusion coefficient related to structural features; t is the number of iterations, Ñ is the gradient, div is the divergence, and θ is the dip of the seismic image event.

[0007] Preferably, the diffusion coefficient f in step 2 is calculated by the seismic data gradient and the event dip, and the structural enhancement seismic data is obtained after iterative solution. The calculation formula is: .

[0008] Preferably, the three convolution kernel sizes of the parallel convolutional neural network in step 3 are 3×3, 5×5, and 7×7, respectively, the number of output channels of each branch is the same, and the number of output channels is C, and each convolution layer is followed by batch normalization and ReLU activation function; The preprocessed seismic data and structurally enhanced seismic data are used as input data. The input data retains the original seismic data spatial size H×W and the number of channels is 2. The input data is passed through a parallel convolutional neural network to form a seismic feature image output by three branches, and a multi-scale seismic feature image is obtained.

[0009] Preferably, in step 4, for the multi-scale seismic characteristic image, the seismic characteristic images output by the three branches are spliced along the channel dimension, and the calculation formula is: ; Where: F1, F2 and F3 are the seismic feature images output by the three branches; F is the spliced seismic feature image; g is the splicing function; the dimension of the spliced seismic feature image is H×W×3C, where H is the height of the image, W is the width of the image, and C is the number of channels of the image.

[0010] Preferably, the channel attention mechanism in step 4 generates a channel statistics vector by global average pooling , use a single fully connected layer to generate the attention weight matrix: ; Where: W is the attention weight matrix, W i,j is the influence weight of the input j-th channel on the output i-th channel.

[0011] Preferably, in step 4, a normalized attention weight matrix is generated by a Sigmoid function, the normalized attention weight matrix is split into three sub-weight vectors, and channel weighting is performed on the original seismic feature image. The weighted feature map is element-by-element added and fused to achieve cross-channel fusion. The calculation formula is: ; Where: Indicates channel-by-channel multiplication; k indicates the number of channels; W k is the sub-weight vector; F fs is the fused feature image.

[0012] Preferably, the residual neural network in step 5 includes two cascaded residual blocks, each residual block is composed of a 3×3 convolutional layer and a jump connection layer, and the number of channels is compressed by a 1×1 convolutional layer at the end of the network.

[0013] Preferably, the earthquake label data in step 6 includes denoised earthquake label data and structurally enhanced earthquake label data; the loss function L used in the end-to-end training is: ; Where: d predt is the seismic data filtered by MA-ResCNN, d label_1 is the denoised earthquake label data, d label_2 To enhance earthquake label data for construction; L err is the mean square error between filtered data and denoised labels, L simis the mean square error between filtered data and constructed enhanced labels, α and β are the weight coefficients of the loss function; the gradient descent algorithm is used to update the network parameters, and the MA-ResCNN model is obtained after training.

[0014] Compared with the prior art, the present invention has at least the following beneficial effects: The present invention improves the signal-to-noise ratio and geological interpretation reliability of seismic data in complex tectonic areas through the fusion of multi-scale convolution feature extraction, structure-guided filtering and attention mechanism, effectively enhances the edge retention capability of key structural features such as fault systems and stratigraphic unconformity surfaces, and is suitable for seismic data processing in deep complex tectonic areas, providing high-quality seismic data for the detailed characterization of oil and gas reservoirs. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of earthquake intelligent filtering of the present invention; Figure 2 The original seismic data in the embodiment of the present invention; Figure 3 This is the noisy seismic data in the embodiment of this application; Figure 4 The seismic noise predicted by the method of the present invention; Figure 5 For filtering seismic noise using the method of the present invention; Figure 6 The locally filtered seismic data in the embodiment of the present invention; Figure 7 This is the noisy seismic data corresponding to the locally filtered seismic data in the embodiment of the present invention. DETAILED DESCRIPTION

[0016] To make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the following Figures 1 to 7 The technical solutions of the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field are within the scope of protection of the present invention.

[0017] like Figure 1 As shown, the present invention discloses a seismic intelligent filtering method based on multi-scale information fusion and structural feature enhancement, comprising the following steps: 1. Obtain the original seismic data, separate the effective frequency band signal through bandpass filtering, retain the seismic data in the main frequency range, and normalize the data to obtain preprocessed seismic data.

[0018] Raw seismic data (SEG-Y format file) is read from a seismic exploration device or storage system. The bandpass filter uses a Butterworth filter, whose cutoff frequency is set according to the main frequency range of the effective signal, and reconstructs the time domain seismic data of the main frequency band through inverse Fourier transform. This embodiment uses post-stack seismic data based on the Marmousi model, source Ricker wavelet, and a main frequency of 60 Hz.

[0019] Normalization processing uses the maximum and minimum normalization algorithm, and the calculation formula for preprocessing seismic data is: ; Where: d norm To pre-process seismic data; d rec is the main frequency band time domain seismic data; d min and d max are the minimum and maximum values of the seismic data amplitude, respectively.

[0020] 2. Based on the anisotropic diffusion equation, the diffusion coefficient is calculated according to the direction of the seismic phase axis, and the original seismic data is processed by anisotropic diffusion filtering to obtain structural enhancement seismic data.

[0021] (1) The partial differential equation of the anisotropic diffusion filter is: ; Where f is the diffusion function used to calculate the diffusion coefficient related to structural features; t is the number of iterations, Ñ is the gradient, div is the divergence, and θ is the dip of the seismic image event.

[0022] (2) The diffusion coefficient f is calculated from the seismic data gradient and the phase axis dip, and the structural enhancement seismic data is obtained after iterative solution. The calculation formula is: .

[0023] 3. Construct a parallel convolutional neural network and use three different sizes of convolution kernels to extract features from preprocessed seismic data and structurally enhanced seismic data to obtain multi-scale seismic feature images.

[0024] (1) The three convolution kernel sizes of the parallel convolutional neural network are 3×3, 5×5 and 7×7 respectively. The number of output channels of each branch is the same, and the number of output channels is C. Each convolution layer is followed by batch normalization and ReLU activation function; among them, the step size of the 3×3 convolution kernel is 1 and the padding is 1, the step size of the 5×5 convolution kernel is 1 and the padding is 2, and the step size of the 7×7 convolution kernel is 1 and the padding is 3.

[0025] (2) The preprocessed seismic data and structurally enhanced seismic data are used as input data. The input data retains the original seismic data spatial size H×W and the number of channels is 2. The input data is passed through a parallel convolutional neural network to form a seismic feature image output by three branches, and a multi-scale seismic feature image is obtained.

[0026] 4. Based on the feature fusion neural network, the channel attention mechanism is used to perform cross-scale fusion of multi-scale feature images to generate a fused feature image.

[0027] (1) For the multi-scale seismic characteristic image, the seismic characteristic images output by the three branches are spliced along the channel dimension, and the calculation formula is: ; Where: F1, F2, and F3 are the seismic feature images output by the three branches; F is the spliced seismic feature image; g is the splicing function; the dimensions of the spliced seismic feature image are H × W × 3C, where H is the image height, W is the image width, and C is the number of image channels. In this embodiment of the present invention, the image height H = 64, the image width W = 64, and the number of output channels C = 2.

[0028] (2) The channel attention mechanism generates a channel statistics vector through global average pooling , use a single fully connected layer to generate the attention weight matrix: ; Where: W is the attention weight matrix, W i,j is the influence weight of the input j-th channel on the output i-th channel.

[0029] (3) Generate a normalized attention weight matrix through the Sigmoid function, split the normalized attention weight matrix into three sub-weight vectors, and implement channel weighting on the original seismic feature image. Perform element-by-element addition fusion on the weighted feature map to achieve cross-channel fusion. The calculation formula is: ; Where: Indicates channel-by-channel multiplication; k indicates the number of channels; W k is the sub-weight vector; F fs is the fused feature image.

[0030] 5. Construct a residual neural network, input the fused feature image and combine the residual block association preprocessing seismic data and structural enhanced seismic data to form a residual neural network MA-ResCNN based on multi-scale fusion and attention mechanism.

[0031] The residual neural network consists of two cascaded residual blocks. Each residual block consists of a 3×3 convolutional layer and a skip connection layer. The input and output channels are the same, with a stride of 1 and padding of 1, followed by a ReLU activation function. The skip connection layer directly adds the feature image to the convolutional layer output, forming a residual learning mechanism. At the end of the network, a 1×1 convolutional layer compresses the feature image channels to the target dimension. In conjunction with a parallel convolutional neural network, a feature fusion neural network, and a residual neural network, a residual convolutional neural network (MA-ResCNN) model based on multi-scale fusion and attention mechanisms is constructed.

[0032] 6. Use earthquake label data to train the MA-ResCNN model end-to-end, input the observed earthquake data into the model, and output the final filtered earthquake data.

[0033] The earthquake label data includes denoised earthquake label data and structurally enhanced earthquake label data; the loss function L used in the end-to-end training is: ; Where: d predt is the seismic data filtered by MA-ResCNN, d label_1 is the denoised earthquake label data, d label_2 To enhance earthquake label data for construction; L err is the mean square error between filtered data and denoised labels, L sim is the mean square error between the filtered data and the constructed enhanced labels, α and β are the weight coefficients of the loss function; in the embodiment of the present invention, α=0.6 and β=0.4.

[0034] The gradient descent algorithm is used to update the network parameters, and the MA-ResCNN model is obtained after training. The embodiment of the present invention uses the gradient descent algorithm Adam optimizer to update the network parameters, and the initial learning rate is set to 1×10 −4 , the batch size (batchsize) is set to 32, and the training round (epoch) is set to 30; the observed or processed seismic data are input into the MA-ResCNN model to obtain the final filtered seismic data.

[0035] Figure 2 The original seismic data in the embodiment of the present invention, that is, the noise-free stacked seismic data based on the Marmousi model, has a total of 737 seismic traces, each containing 750 sampling points. Figure 3 This is the noisy seismic data in this embodiment, and its data signal-to-noise ratio is 3.56dB. Figure 4 For the seismic noise predicted by the MA-ResCNN model of the present invention, Figure 5 To filter seismic data using the MA-ResCNN model of the present invention; Figure 4It can be seen that the method of the present invention can effectively predict data residuals (i.e. seismic noise); Figure 5 It can be seen that the method of the present invention can effectively restore the information and geological structure characteristics of seismic data, wherein the signal-to-noise ratio of the filtered seismic data is 17.81 dB, which is significantly improved compared with the data before filtering (containing noise). Figure 6 、 Figure 7 The local filtered seismic data (257th to 318th seismic traces, 322nd to 383rd sampling points) of the embodiment of the present invention are compared with the corresponding noisy seismic data. Figure 6 and Figure 7 By comparison, it can be seen that the method of the present invention not only effectively removes the noise interference of the seismic data, but also makes the boundary information of the seismic phase axis in the filtered data clear. Therefore, it can be seen that the method of the present invention can effectively improve the signal-to-noise ratio of seismic data and protect geological structural features.

[0036] The present invention improves the signal-to-noise ratio and geological interpretation reliability of seismic data in complex tectonic areas through the fusion of multi-scale convolution feature extraction, structure-guided filtering and attention mechanism, effectively enhances the edge retention capability of key structural features such as fault systems and stratigraphic unconformity surfaces, and is suitable for seismic data processing in deep complex tectonic areas, providing high-quality seismic data for the detailed characterization of oil and gas reservoirs.

[0037] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A seismic intelligent filtering method based on multi-scale information fusion and structural feature enhancement, characterized in that: The following steps are involved: Step 1: Obtain the original seismic data, separate the effective frequency band signal through bandpass filtering, retain the seismic data in the main frequency range, and normalize the data to obtain preprocessed seismic data; Step 2: Based on the anisotropic diffusion equation, the diffusion coefficient is calculated according to the direction of the seismic phase axis, and the original seismic data is processed by anisotropic diffusion filtering to obtain structural enhancement seismic data; Step 3: Construct a parallel convolutional neural network and use three different sizes of convolution kernels to extract features from the preprocessed seismic data and the structurally enhanced seismic data to obtain multi-scale seismic feature images; Step 4: Based on the feature fusion neural network, the channel attention mechanism is used to perform cross-scale fusion of multi-scale feature images to generate a fused feature image; Step 5: Construct a residual neural network, input the fused feature image and combine it with the residual block to associate the preprocessed seismic data and the structural enhanced seismic data to form a residual neural network MA-ResCNN based on multi-scale fusion and attention mechanism; Step 6: Use earthquake label data to train the MA-ResCNN model end-to-end, input the observed earthquake data into the model, and output the final filtered earthquake data.

2. The earthquake intelligent filtering method based on multi-scale information fusion and structural feature enhancement according to claim 1 is characterized in that: The bandpass filtering in step 1 adopts a Butterworth filter, whose cutoff frequency is set according to the main frequency range of the effective signal, and reconstructs the time domain seismic data of the main frequency band through inverse Fourier transform; the normalization processing adopts the maximum minimum normalization algorithm, and the calculation formula of the preprocessed seismic data is: ； Where: d norm To pre-process seismic data; d rec It is the main frequency band time domain seismic data; d min and d max are the minimum and maximum values of the seismic data amplitude, respectively.

3. The earthquake intelligent filtering method based on multi-scale information fusion and structural feature enhancement according to claim 2 is characterized in that: The partial differential equation of the anisotropic diffusion filter described in step 2 is: ； Where f is the diffusion function used to calculate the diffusion coefficient related to structural features; t is the number of iterations, Ñ is the gradient, div is the divergence, and θ is the dip of the seismic image event.

4. The earthquake intelligent filtering method based on multi-scale information fusion and structural feature enhancement according to claim 3 is characterized in that: The diffusion coefficient f in step 2 is calculated from the seismic data gradient and the event dip. After iterative solution, the structural enhancement seismic data is obtained. The calculation formula is: 。 5. The earthquake intelligent filtering method based on multi-scale information fusion and structural feature enhancement according to claim 4 is characterized in that: The three convolution kernel sizes of the parallel convolutional neural network described in step 3 are 3×3, 5×5, and 7×7 respectively. The number of output channels of each branch is the same, and the number of output channels is C. Each convolution layer is followed by batch normalization and ReLU activation function; The preprocessed seismic data and structurally enhanced seismic data are used as input data. The input data retains the original seismic data spatial size H×W and the number of channels is 2. The input data is passed through a parallel convolutional neural network to form a seismic feature image output by three branches, and a multi-scale seismic feature image is obtained.

6. The earthquake intelligent filtering method based on multi-scale information fusion and structural feature enhancement according to claim 5 is characterized in that: In step 4, for the multi-scale seismic characteristic image, the seismic characteristic images output by the three branches are spliced along the channel dimension. The calculation formula is: ； Where: F1, F2 and F3 are the seismic feature images output by the three branches; F is the spliced seismic feature image; g is the splicing function; the dimension of the spliced seismic feature image is H×W×3C, where H is the height of the image, W is the width of the image, and C is the number of channels of the image.

7. The earthquake intelligent filtering method based on multi-scale information fusion and structural feature enhancement according to claim 6 is characterized in that: The channel attention mechanism described in step 4 generates a channel statistics vector through global average pooling , use a single fully connected layer to generate the attention weight matrix: ； Where: W is the attention weight matrix, W i,j is the influence weight of the input j-th channel on the output i-th channel.

8. The earthquake intelligent filtering method based on multi-scale information fusion and structural feature enhancement according to claim 7 is characterized in that: In step 4, the normalized attention weight matrix is generated by the Sigmoid function, which is then split into three sub-weight vectors. Channel weighting is then applied to the original seismic feature image, and cross-channel fusion is achieved by element-by-element addition of the weighted feature map. The calculation formula is: ； Where: Indicates channel-by-channel multiplication; k indicates the number of channels; W k is the sub-weight vector; F fs is the fused feature image.

9. The earthquake intelligent filtering method based on multi-scale information fusion and structural feature enhancement according to claim 8 is characterized in that: The residual neural network described in step 5 contains two cascaded residual blocks. Each residual block consists of a 3×3 convolutional layer and a skip connection layer. The number of channels is compressed by a 1×1 convolutional layer at the end of the network.

10. The earthquake intelligent filtering method based on multi-scale information fusion and structural feature enhancement according to claim 9 is characterized in that: The earthquake label data in step 6 includes denoised earthquake label data and structurally enhanced earthquake label data; the loss function L used in the end-to-end training is: ； Where: d predt is the seismic data filtered by MA-ResCNN, d label_1 is the denoised earthquake label data, d label_2 To enhance earthquake label data for construction; L err is the mean square error between filtered data and denoised labels, L sim is the mean square error between filtered data and constructed enhanced labels, α and β are the weight coefficients of the loss function; the gradient descent algorithm is used to update the network parameters, and the MA-ResCNN model is obtained after training.

Images