A 3D fault recognition method integrating multi-layer dilated convolution and attention mechanism

Abstract

A three-dimensional fault recognition method integrating multi-layer dilated convolution and attention mechanism, outputs fault recognition results for the input image to be predicted through a recognition network, the recognition network includes an H-Unet network based on multi-layer dilated convolution and an RA-Unet network based on an attention mechanism, and uses a three-dimensional synthetic seismic data body and its label as input to train the H-Unet network and the RA-Unet network respectively, fuses the prediction results of the two networks, and performs weighted averaging on the prediction results of the two networks at each pixel point, and outputs the final predicted image as the recognition result. Aiming at the characteristics of three-dimensional seismic fault recognition, the present invention designs a recognition network that integrates the H-Unet network based on multi-layer dilated convolution and the RA-Unet network based on the attention mechanism, which can effectively extract the information of the three-dimensional seismic data body and identify the fault and its width and local details.

Description

Technical Field

[0001] The present invention belongs to the field of computer data processing technology, and relates to seismic exploration data interpretation and computer deep learning technology, and specifically is a three-dimensional fault recognition method integrating multi-layer hole convolution and attention mechanism. Background Art

[0002] Faults are prominent geological features formed by brittle deformation in the upper part of the earth's crust. Fault identification is of great significance to oil and gas exploration, so fault identification is an important part of seismic data interpretation. Traditional fault interpretation of seismic reflection data is a manual process that depends on the quality of seismic reflection data and the experience of the interpreter. This process is very time-consuming and requires a lot of manual repetitive work. The interpretation of a single data set usually takes weeks to months to complete. Researchers have tried to use seismic attributes to assist in interpretation, but the results are uncertain due to the need to manually select features, and the interpretation results are difficult to quantify. The main advantage of the application of deep convolutional neural networks in fault identification is that they can automatically identify and describe geological features, thereby automating the fault identification workflow. Due to the rapid development of computer performance, deep convolutional neural networks have shortened the life cycle of fault generation and fault interpretation from months to hours or minutes. There are two main methods for using DCNN for fault identification. One is to convert the fault identification problem into a binary classification problem, such as Di et al. [1] Use a 31\times31 window to slide and use the sliding window to identify whether the central pixel belongs to a fault. This method introduces redundant parameters and requires reconstruction of the processing results. Another method is to regard the fault identification problem as an image segmentation problem, such as Wu et al. [2] The 3D U-Net was trained using synthetic data and then tested on synthetic and actual field data. However, the fault refinement of this method was not high enough, the fault edges were blurred, and isolated small connected fault areas appeared, so there is still room for improvement in this method.

[0003] For fault identification of three-dimensional seismic data, the existing technology has also studied various improvements based on the U-Net network, such as the following improved scheme.

[0004] 1. Double Unet [3] The input image passes through the encoder-decoder structure of two UNets in turn, and based on the VGG-19 network pre-trained on ImageNet, it has achieved good results in medical image segmentation. This framework uses the pre-trained VGG, and the decoder part of the two networks is not much changed from the original UNet. However, due to the large difference in the recognition features of the ImageNet and seismic image datasets corresponding to the network, this method cannot achieve good results on the seismic dataset.

[0005] 2. MultiResUNet [4] The module in the figure is a combination of three 3x3 convolution results, which are concatenated as a combined feature map and then added to the result of the 1x1 convolution of the input feature map. For fault recognition, the unet has the limitation that the receptive field is not large enough.

[0006] References

[0007] [1]DI H,WANG Z,ALREGIB G.Seismic fault detection from post-stackamplitude by conv olutional neural networks[C] / / 80th EAGE Conference andExhibition 2018:Vol 2018.2018:1–5.

[0008] [2]WU

[0009] [3]D.Jha, MARiegler, D.Johansen, P.Halvorsen and HDJohansen, "DoubleU-Net: ADeep Convolutional Neural Network for Medical ImageSegmentation," 2020IEEE 33rd International Symposium on Computer-BasedMedical Systems(CBMS), 2020, pp.558-564, doi:10.1109 / CBMS49503.2020.00111.

[0010] [4]Ibtehaz N,Rahman M S.MultiResUNet:Rethinking the U-Netarchitecture for multimodal biomedical image segmentation[J].Neural Networks,2020,121:74-87. Summary of the invention

[0011] The problem to be solved by the present invention is: the problems existing in the prior art in three-dimensional fault recognition are: first, fault recognition needs to focus on local fault information, and does not need to consider irrelevant noise information, such as the false shadow area in the processing of seismic reflection signals. Isolated small connected areas often appear in the current recognition results. Secondly, there is a problem of discontinuity of faults in the recognition results.

[0012] The technical solution of the present invention is: a three-dimensional fault recognition method integrating multi-layer dilated convolution and attention mechanism, outputting fault recognition results for an input image to be predicted through a recognition network, wherein the recognition network includes an H-Unet network based on multi-layer dilated convolution and an RA-Unet network based on an attention mechanism, respectively training the H-Unet network and the RA-Unet network with a three-dimensional synthetic seismic data volume and its label as input, integrating the prediction results of the two networks, performing weighted averaging on the prediction results of the two networks at each pixel point, and outputting the final prediction image as the recognition result,

[0013] The H-Unet network improves the contraction path based on the Unet network. The convolution block in the Unet encoder is replaced by the HD module in the contraction path. The HD module is composed of a 3×3×3 convolution layer and a dilated convolution layer with a dilation rate of 3 and 5. At the same time, mixed dilated convolution and normal convolution are used to extract multi-scale features of seismic images.

[0014] The RA-Unet network improves the extension path based on the Unet network. The extension path includes a three-level residual module and an attention gate module. The attention gate module is added to the jump connection in the U-Net network. The low-level feature map is output through upsampling and residual modules. At the same time, the feature map from the jump connection is output through the attention gate module as a gated signal vector. The two outputs are spliced ​​and input to the next level to finally obtain the network output.

[0015] The training inputs two models separately and performs joint training. The loss function is:

[0016]

[0017] Where N is the total number of labels, g i is the true value, i.e. the label value, p iis the predicted value of the network output, w is the proportion of non-fault points, λ is the adjustment parameter, the parameter gradient of each network layer in the network is calculated by the loss function, and the Adam optimizer is used for supervised training.

[0018] Furthermore, the attention gate module receives the upper left feature map x l And the lower right feature map g l The upper left layer refers to the feature map of the encoder end connected by the jump, and the lower right layer refers to the feature map of the previous level output of the decoder. The feature map of the lower right layer g l As the gate signal vector, we get the output αx of the attention gate module l , α is the attention coefficient:

[0019]

[0020] Where σ is the sigmoid activation function, σ l is the Relu activation function, w x 、w g and ψ is a linear transformation, b g and b ψ is the offset term.

[0021] Furthermore, a decay strategy is set for the adjustment parameter λ, and λ is set to change with the size of the current epoch.

[0022]

[0023] j is the current epoch number, and n is the total number of epochs.

[0024] Compared with the existing methods, the fault recognition method based on the attention mechanism of the present invention has the following advantages:

[0025] 1. Compared with marking faults on a two-dimensional seismic profile, the recognition network of the present invention can maintain the spatial morphological characteristics of the faults for recognition in three-dimensional space.

[0026] 2. The present invention designs an H-Unet network based on multi-layer dilated convolution and an RA-Unet network based on an attention mechanism. The H-Unet network has mixed dilated convolutions with different field of view scales, which can view contextual information in seismic images more widely without losing resolution, thereby enhancing the continuity of fault identification. The RA-Unet network's emphasis on local features enables it to focus on geological features, suppress the feature expression of geological noise and pseudo-shadow areas in seismic processing, and solve the problem of insufficient accuracy of U-Net segmentation of fault boundaries. Moreover, faults appear as sudden changes in seismic signals on images, and fault identification requires that the broken lines of fault interpretation be as clear as possible. The strong inhibition brought by the attention mechanism designed in the present invention can help achieve this goal.

[0027] 3. The present invention utilizes the idea of ​​model fusion, makes full use of the differences between networks, and fuses the prediction results of the H-Unet and RA-Unet models designed by the present invention. The integrated network obtained by model fusion has better learning ability, can better extract generalization features in seismic data, and effectively improve the accuracy of model classification; at the same time, it avoids the problem that a single model is prone to fall into the local minimum during calculation.

[0028] 4. Due to the serious imbalance between positive and negative samples of seismic data in fault identification tasks, the present invention also proposes a new dynamic combination loss function. The first half of the expression in the loss function, the true fault label value and the predicted fault pixel value are trained to overlap gradually. The denominator considers the total number of fault pixels in the global range, while the numerator considers the overlap between the true fault label value and the predicted fault pixel value in the local range. Therefore, the loss function of the present invention considers local and global loss information, and the consideration of the pixel weight can solve the problem of class imbalance. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 HD module designed for the present invention.

[0030] Figure 2 It is the attention gate module involved in the RA-Unet model of the present invention.

[0031] Figure 3 Schematic diagram of the attention coefficient of the attention gate module of the RA-Unet model of the present invention.

[0032] Figure 4 It is the residual module involved in the RA-Unet model of the present invention.

[0033] Figure 5 Schematic diagram of the attention mechanism of the RA-Unet model of the present invention.

[0034] Figure 6 It is the RA-Unet network of the present invention.

[0035] Figure 7 Schematic diagram of the training process of the present invention.

[0036] Figure 8 These are three identification embodiments of the present invention. (a), (b), and (c) respectively show three seismic 3D data volumes and their fault identification results. DETAILED DESCRIPTION

[0037] The present invention uses the three-dimensional synthetic seismic data volume and its label as input, and uses the new dynamic loss function designed by the present invention as the loss function to train the H-UNet and RA-Unet networks proposed by the present invention. The image to be predicted is input into the trained model, the prediction results of the two models are merged, the prediction results of the two models at each pixel are weighted and averaged, and the final predicted image is output.

[0038] The traditional U-Net network includes an encoder and a decoder, which are also called a contraction path and an expansion path respectively. The two networks designed in the present invention have made improvements on the contraction path and the expansion path respectively.

[0039] Since fault identification is processed as an image segmentation problem into a binary classification problem of pixels, there will be discontinuities and breakpoints in the fault line. To solve this problem, the present invention designs a new HD module based on dilated convolution and residual connection, replaces the convolution block in the Unet encoder, improves the contraction path, and forms a new H-Unet network. The HD module proposed in the present invention uses mixed dilated convolution and normal convolution to extract multi-scale features of seismic images. The HD module designed in the present invention is composed of a 3×3×3 convolution layer and a dilated convolution layer with a dilation rate of 3.5. The specific design is as follows: Figure 1 The hybrid atrous convolution with different field of view scales in the present invention can more extensively view the contextual information in the seismic image without losing resolution, thereby enhancing the continuity of fault identification.

[0040] Because the segmentation result of geological faults is linear, the fault lines should be narrowed as much as possible during fault identification, and the existence of isolated small connected areas should be reduced. There are noise and false shadow areas in seismic images, and it is necessary to distinguish these irrelevant features from fault features. Based on the above considerations. The present invention designs a RA-Unet network, which improves the Unet network extension path. The extension path consists of three residual blocks and three attention gate modules. The traditional U-Net network uses a jump connection to connect the upsampling result with the output of the sub-module with the same resolution in the encoder as the input of the next sub-module in the decoder. The present invention adds an attention gate module to the jump connection, such as Figure 2 As shown, the attention gate module receives the upper left feature map x l And the lower right feature map g l The upper left layer refers to the feature map of the encoder end connected by the jump, and the lower right layer refers to the feature map of the previous level output of the decoder. The feature map of the lower right layer g l As the gate signal vector, we get the output αx of the attention gate module l , α is the attention coefficient. Figure 3 As shown, the attention module uses a 1×1×1 convolution w x 、w gand ψ for the gated signal vector g l and feature map x l Perform a linear transformation, then apply the Relu and Sigmoid activation functions, upsample the result, and obtain the multidimensional attention coefficient α:

[0041]

[0042] Where σ is the sigmoid activation function, σ l is the Relu activation function, w x 、w g and ψ is a linear transformation, b g and b ψ is the offset term.

[0043] Figure 4 is the residual block of the RA-Unet network. The three residual blocks and three attention gate modules of the present invention constitute a three-level decoding, such as Figure 5 As shown in Figure 1, in each level, the attention gate module receives the feature map from the upper left layer and the feature map from the lower right layer, i.e., the low-level feature map, and obtains the attention output. At the same time, the feature map from the lower right layer is upsampled and the residual module is used to obtain the feature map from the upper right layer. The feature map from the upper right layer is concatenated with the attention output to obtain the decoding output of this level. Figure 6 As shown, the RA-Unet network of the present invention highlights significant features and eliminates irrelevant and noisy features through an attention mechanism, and finally uses a softmax activation function to map the feature map to two categories.

[0044] Due to the serious class imbalance in earthquake fault identification tasks, the proportion of fault points to the total number of points is 0.077. The present invention considers designing a new loss function by combining the dice loss and the weighted cross entropy loss function to solve the above problem.

[0045]

[0046] Where N is the total number of labels (samples), g i The true value is the label value, p i is the predicted value of the network output, w is the proportion of non-fault points, is the known parameter of the training sample, and λ is the adjustment parameter. The parameter gradient of each network layer in the model is calculated through the loss function, and the Adam optimizer is used for supervised training. In the training process of deep learning, one pass through the training set is usually called an epoch.

[0047] The present invention proposes a decay strategy for λ, setting λ to change with the size of the current epoch:

[0048]

[0049] j is the current epoch number, and n is the total number of epochs.

[0050] The implementation of the present invention is further illustrated by a specific example. This example uses a public data set provided by Wu, which is randomly generated using a toolkit developed by Hale et al. The data set contains 220 pairs of 3D seismic data volumes, 200 of which are used as training sets and 20 as test sets. Each pair contains a 128×128×128 seismic data volume and a seismic fault data volume labeled as 128×128×128. The 3D seismic data volume has three dimensions: Xline, Inline, and Time.

[0051] The present invention designs H-Unet network and RA-Unet network, and uses synthetic data and the loss function proposed by the present invention to train the two networks. The idea of ​​model fusion is used to perform weighted averaging on the prediction results of the two models at each pixel point. Since the designed base models are diverse and different, the fused model can improve the final accuracy.

[0052] Step 1: Take the three-dimensional synthetic seismic data volume and its label as input and perform data preprocessing. The image enhancement method can keep the image features unchanged and generate derived data from the original data to increase the sample size, improve the problem of limited sample size, and effectively prevent model overfitting. The present invention adopts the data enhancement method of horizontal flipping and vertical flipping.

[0053] Step 2: Build the model and train. The HD module uses multi-layer hole convolution and residual connection to cover the seismic image with a sawtooth waveform receptive field. The HD module replaces the convolution block in the original Unet to form a new H-Unet. The RA-Unet network combines the residual block and attention mechanism with the three-dimensional U-Net architecture, in which a series of attention gate units are added to replace the jump connection to highlight important feature information and eliminate the redundant information of irrelevant and noisy feature responses. Combined with the dynamic loss function designed in the present invention, the parameter gradients of each network layer in the model are calculated by the loss function, and the Adam optimizer is used to supervise the training. In the training process of deep learning, one traversal of the training set is usually called an epoch. In this embodiment, the initial learning rate during training is set to 0.001. Both networks are trained for 50 epochs, and the training set loss values ​​converge to about 0.03, obtaining two pre-trained models.

[0054] Step 3: Input the image to be predicted into the two trained models to obtain two predicted probability 3D volumes. Sum and average each pixel of the two 3D volumes to obtain a new probability 3D volume, such as Figure 7 shown.

[0055] Based on the Unet model, this paper proposes a three-dimensional fault recognition method using the idea of ​​model fusion. This method treats the fault recognition problem as a binary image segmentation problem, and uses the ideas of multi-layer dilated convolution, residual connection and attention mechanism to propose a new dynamic loss function to complete the fault recognition of the three-dimensional body. Figure 8 (a)(b)(c) are the identification results of three groups of earthquake three-dimensional faults. It can be seen that the present invention can effectively identify the location and distribution of faults, and has obvious advantages in terms of fault continuity and background noise resistance. There are almost no burrs around the fault, the edges are clear, and the background also has good noise resistance and almost no noise points.

Claims

1. A three-dimensional fault recognition method integrating multi-layer dilated convolution and attention mechanism, characterized by The recognition network outputs the fault recognition result for the input image to be predicted. The recognition network includes the H-Unet network based on multi-layer void convolution and the RA-Unet network based on the attention mechanism. The H-Unet network and the RA-Unet network are trained respectively with the three-dimensional synthetic seismic data volume and its label as input. The prediction results of the two networks are fused, and the prediction results of the two networks at each pixel are weighted and averaged. The final prediction image is output as the recognition result. The H-Unet network improves the contraction path based on the Unet network. The convolution block in the Unet encoder is replaced by the HD module in the contraction path. The HD module is composed of a 3×3×3 convolution layer and a dilated convolution layer with a dilation rate of 3 and 5. At the same time, mixed dilated convolution and normal convolution are used to extract multi-scale features of seismic images. The RA-Unet network improves the extension path based on the Unet network. The extension path includes a three-level residual module and an attention gate module. The attention gate module is added to the jump connection in the U-Net network. The low-level feature map is output by upsampling and residual modules. At the same time, the feature map from the jump connection is output by the attention gate module as a gated signal vector. The two outputs are spliced ​​and input to the next level to finally obtain the network output. The attention gate module receives the upper left feature map x l And the lower right feature map g l The upper left layer refers to the feature map of the encoder end connected by the jump, and the lower right layer refers to the feature map of the previous level output of the decoder. The feature map of the lower right layer g l As the gate signal vector, we get the output αx of the attention gate module l , α is the attention coefficient: Where σ is the sigmoid activation function, σ l is the Relu activation function, w x 、w g and ψ is a linear transformation, b g and b ψ is the offset term; The training inputs two models separately and performs joint training. The loss function is: Where N is the total number of labels, g i is the true value, i.e. the label value, p i is the predicted value of the network output, w is the proportion of non-fault points, λ is the adjustment parameter, the parameter gradient of each network layer in the network is calculated through the loss function, and the Adam optimizer is used for supervised training; the attenuation strategy is set for the adjustment parameter λ, and λ is set to change with the size of the current epoch. j is the current epoch number, and n is the total number of epochs.

Images