Automatic identification method of earthquake faults based on integration of Res-50 and CBAM

Abstract

The present invention discloses an automatic seismic fault identification method integrating Res-50 and CBAM. Based on the U-Net network, the Res-50 residual block is introduced in the coding layer of the network. When the seismic data image is input into the module, the training parameters can be greatly reduced, the network is deepened and the learning ability is improved. In the decoding layer of the network, the CBAM convolutional attention module is introduced, which assigns a higher weight to the fault area in the seismic data feature map and a lower weight to the background, so as to collaboratively learn the fault detail information in the seismic data image, improve the neural network's attention to the fault in the image, and thus improve the accuracy of the network in identifying faults.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and more particularly to an automatic earthquake fault identification method integrating Res-50 and CBAM. Background Art

[0002] One of the important tasks in seismic data processing is fault interpretation. As a channel for the storage and migration of oil and gas, how to accurately and efficiently identify faults is of great significance to the exploration and exploitation of oil and gas reservoirs. The traditional method of fault identification is to manually interpret faults using the discontinuity of event axes in two-dimensional seismic profiles. The results are subjective, with large errors and low efficiency.

[0003] As one of the main methods of machine learning, deep learning is widely used in various fields. In the early days of seismic exploration, machine learning has been applied to seismic data interpretation. With the continuous development of machine learning, research on seismic data processing based on deep learning has gradually emerged. Dong Shouhua et al. (1997) completed the accurate identification of small faults by building a BP neural network. Teng Chao et al. (2016) used scatter plot method and grey correlation analysis method to calculate the correlation between various fault attributes and their extension length, so as to train the support vector machine model (SVM) to achieve the purpose of identifying small faults. Di et al. (2018) constructed a multi-layer perceptron (MLP) network structure model and trained the model with multiple seismic attributes for fault identification. Huang et al. (2017) used convolutional neural network for fault identification, which effectively improved the accuracy of fault identification. Wu et al. (2018) used the network model U-Net based on the fully convolutional network (FCN) for automatic fault identification, which greatly improved the accuracy of fault identification. Liu et al. (2020) introduced the residual module in ResNet based on the U-Net network, which greatly reduced the parameters required to learn the network, not only improving the accuracy of fault identification but also improving the efficiency.

[0004] Although great progress has been made in fault identification using deep learning, there are still problems in practice such as slow convergence of network training and unsatisfactory actual test results.

[0005] Therefore, how to provide an automatic earthquake fault identification method with high recognition accuracy, good fault continuity and clear morphology is a technical problem that technical personnel in this field urgently need to solve. Summary of the invention

[0006] In view of this, the present invention provides an automatic earthquake fault identification method integrating Res-50 and CBAM, which can reduce the amount of parameter calculation, learn the detailed information of faults in seismic data images, and improve the recognition accuracy of neural networks for faults.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] An automatic earthquake fault identification method integrating Res-50 and CBAM, comprising:

[0009] S1. Based on the U-Net network, multiple groups of Res-50 residual blocks are introduced in the encoding part of the U-Net network, and multiple CBAM convolutional attention modules are introduced in the decoding part to build a fault recognition model;

[0010] S2, in the encoding part, the original seismic image is converted into the original feature map;

[0011] S3, after downsampling, the original feature map enters the first group of Res-50 residual blocks for convolution processing, and the convolution result is linearly superimposed with the input and activated by the ReLU function to obtain a new feature map;

[0012] S4, repeatedly down-sample the feature map obtained in S3 as shown in S3, input the next set of Res-50 residual blocks, and repeat the convolution process as shown in S3 until all Res-50 residual blocks are passed to obtain the feature map F of the encoding part;

[0013] S5, after deconvolution of the feature map F through the upsampling layer, enter the first CBAM convolution attention module to perform CBAM operation, assign different weights to the fault area and background in the feature map F, and obtain a feature map F″ that significantly contains channel position information and spatial position information;

[0014] S6, performing feature mapping cascade on the feature map F″ and the feature map F, and performing a convolution operation on the cascaded feature map;

[0015] S7, repeatedly performing upsampling, CBAM operation, feature map cascade and convolution operation as shown in S5-S6 on the feature map obtained in S6, until all CBAM convolution attention modules are passed to obtain a feature image containing key information of the fault;

[0016] S8, restoring the feature image obtained in S7 to the original image size to obtain a fault recognition result image.

[0017] Furthermore, each downsampling layer of the encoding part corresponds to a group of Res-50 residual blocks, and the number of groups of downsampling layers and Res-50 residual blocks is 4; the encoding part sets 3, 4, 6, and 3 groups of Res-50 residual blocks from shallow to deep layers.

[0018] Furthermore, each group of Res-50 residual blocks contains three consecutive Res-50 residual blocks, and each Res-50 residual block contains three consecutive convolution kernels of 1×1, 3×3 and 1×1.

[0019] Furthermore, S2 includes: performing 3×3 convolution operation, normalization processing and maximum pooling processing on the original seismic image in sequence to obtain the original feature map; wherein, the 3×3 convolution operation is used to extract feature information in the input image; the normalization processing is used to adjust the size of each eigenvalue to a similar range to make its distribution consistent; the maximum pooling processing is used to reduce the dimension of the eigenvalue.

[0020] Furthermore, in S5, the CBAM convolutional attention module includes a channel attention module and a spatial attention module, and the CBAM operations performed by the module include:

[0021] S51, input the feature map F of the encoded part into the channel attention module, and obtain the feature map F′ reflecting the fault feature channel information by recalibrating the feature of F;

[0022] S52. The feature map F′ enters the spatial attention module and multiplies it with the spatial weight coefficient matrix to obtain a feature map F″ that significantly contains channel position information and spatial position information.

[0023] Furthermore, the feature map cascade operation in S6 is: concatenating the feature map F″ of the decoding part with the feature map F at the corresponding position of the encoding part.

[0024] Furthermore, in S8, after performing a 3×3 convolution on the feature image obtained in S7, the feature image size is restored to the original image size through a 1×1 convolution kernel.

[0025] Furthermore, in the training phase of the fault identification model, a weight factor is added to the loss function, and the Adam algorithm is used to iteratively update the weight factor, where the loss function expression is:

[0026]

[0027]

[0028] Where N is the number of input earthquake data points, and m+n=N; m is the positive sample, that is, the number of fault points; n is the negative sample, that is, the number of non-fault points; y i is the fault label data of the i-th feature image, where 1 represents a fault point and 0 represents a non-fault point; α∈[0,1] is a weight factor, which assigns a high weight value to the fault point sample and a low weight value to the non-fault point sample; x i ∈(0,1) is the fault prediction probability of the i-th feature image.

[0029] Furthermore, before S2, it also includes: after enhancing, trimming and normalizing the data set in sequence, combining the label data together as training samples of the fault recognition model, and taking each image of the training sample as the original seismic image.

[0030] Through the above technical solutions, it can be known that compared with the prior art, the present invention discloses a method for automatic identification of earthquake faults integrating Res-50 and CBAM, which is based on the U-Net network and combined with the Res-50 residual block in the encoding area to improve the network's ability to extract fault features and reduce the difficulty of model training; it is integrated with the Convolutional Block Attention Module (CBAM) in the decoding area to suppress information irrelevant to the fault and enhance the ability to focus on fault target extraction. In terms of loss function selection, in order to solve the problem of imbalance in the number of fault points and non-fault points in fault identification, this paper uses the Balanced Cross Entropy function. At the same time, in order to enable the network to better adapt to the fault identification of actual seismic data, the present invention performs data enhancement on the synthetic seismic data to generate a training data set for model training. Finally, the trained network model is used to perform fault identification test on actual seismic data. The fault identification model of the present invention not only has high recognition accuracy, but also has good fault continuity and clear morphology. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0032] Figure 1 A schematic diagram of the structure of the fault identification model provided by the present invention;

[0033] Figure 2 Schematic diagram of the structure of Res-34 and Res-50 residual blocks;

[0034] Figure 3 is a training data set; wherein (a) is a schematic diagram of the original seismic profile in the training data set, (b) is the labeled data in the training data set, and (c) is a superimposed display diagram of the original seismic profile and the labeled data in the training data set;

[0035] Figure 4 is the loss change curve;

[0036] Figure 5is the accuracy change curve;

[0037] Figure 6 The fault identification results of the fault identification model for the validation set; where (a) is the original seismic profile in the validation set, (b) is the fault label in the validation set, and (c) is the fault identification result.

[0038] Figure 7 The accuracy change curves of different networks;

[0039] Figure 8 It is the change curve of different network losses;

[0040] Fig. 9 The recognition results of U-Net, Res50-UNet and Res-CBAM-UNet applied to the seismic data in the actual work area; (a) is the original seismic profile; (b) is the U-Net recognition result; (c) is the Res50-UNet recognition result; (d) is the Res-CBAM-UNet recognition result;

[0041] Fig.10 The recognition results of normal faults in a certain area of ​​seismic data in the actual work area using U-Net, Res50-UNet and Res-CBAM-UNet; (a) is the original seismic profile; (b) is the U-Net recognition result; (c) is the Res50-UNet recognition result; (d) is the Res-CBAM-UNet recognition result. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] like Figure 1 As shown, the embodiment of the present invention discloses an automatic identification method of earthquake faults integrating Res-50 and CBAM, comprising:

[0044] S1. Based on the U-Net network, multiple groups of Res-50 residual blocks are introduced in the encoding part of the U-Net network, and multiple CBAM convolutional attention modules are introduced in the decoding part to build a fault recognition model (Res-CBAM-UNet);

[0045] S2. In the encoding part, the original seismic image is converted into the original feature map; specifically, the original seismic image is subjected to 3×3 convolution operation, normalization processing and maximum pooling processing in sequence to obtain the original feature map; among them, the 3×3 convolution operation is used to extract the feature information in the input image; the normalization processing is used to adjust the size of each eigenvalue to a similar range to make its distribution consistent, thereby speeding up the speed of gradient descent to find the optimal solution, that is, speeding up the convergence speed when training the network; the maximum pooling processing is used to reduce the dimension of the eigenvalue, thereby simplifying the network complexity and improving the receptive field of subsequent features.

[0046] S3, after downsampling, the original feature map enters the first group of Res-50 residual blocks for convolution processing, and the convolution result is linearly superimposed with the input and activated by the ReLU function to obtain a new feature map; among them, the pooling layer is also called the down-sampling layer (Down-Sampling), which can retain the effective information in the image and avoid overfitting to a certain extent. There is a ReLU activation function after each convolution. The ReLU activation function is a nonlinear activation function responsible for mapping the input of the neuron to the output.

[0047] Each group of Res-50 residual blocks contains three consecutive Res-50 residual blocks, and each Res-50 residual block contains three consecutive convolution kernels of 1×1, 3×3 and 1×1.

[0048] S4. Repeat the downsampling of the feature map obtained in S3 as shown in S3, and input the next group of Res-50 residual blocks, and repeat the convolution process as shown in S3 until all Res-50 residual blocks are passed to obtain the feature map F of the encoding part; this embodiment needs to repeat S3 three times.

[0049] Among them, the number of groups of Res-50 residual blocks corresponds to the downsampling layer, and each downsampling layer of the encoding part corresponds to a group of Res-50 residual blocks. The number of downsampling needs to be determined in combination with the actual task. In this embodiment, the number of groups of downsampling layers and Res-50 residual blocks is 4; the encoding part sets 3, 4, 6, and 3 groups of Res-50 residual blocks from shallow to deep layers.

[0050] S5. Deconvolve the feature map F through a 2×2 upsampling layer to map the image from a small resolution to a large resolution. After deconvolution, enter the first CBAM convolution attention module to perform CBAM operation, assign different weights to the fault area and background in the feature map F, and obtain a feature map F″ that significantly contains channel position information and spatial position information; pass the obtained feature map F through the upsampling layer.

[0051] Among them, the CBAM convolutional attention module includes a channel attention module and a spatial attention module, and the CBAM operations it performs include:

[0052] S51, input the feature map F of the encoded part into the channel attention module, and obtain the feature map F′ reflecting the fault feature channel information by recalibrating the feature of F;

[0053] S52. The feature map F′ enters the spatial attention module and multiplies it with the spatial weight coefficient matrix to obtain a feature map F″ that significantly contains channel position information and spatial position information.

[0054] After being processed by the CBAM convolutional attention module, the fault information of the image is paid more attention and the fault features are significantly amplified, while CBAM does not change the size of the input feature image.

[0055] S6, perform feature mapping cascade on the feature map F″ and the feature map F, and perform a convolution operation on the cascaded feature map; the feature mapping cascade operation in S6 is: splicing the feature map F″ of the decoding part with the feature map F at the corresponding position of the encoding part, thereby fusing the feature maps at the corresponding positions in the encoding part and the decoding part, while obtaining more high-resolution information and more completely restoring the detail information in the original image.

[0056] S7. Repeat the upsampling, CBAM operation, feature map cascade and convolution operation shown in S5-S6 for the feature map obtained in S6 until it passes through all the CBAM convolution attention modules to obtain a feature image containing key information of the fault; this embodiment repeats S5-S6 three times.

[0057] S8, restore the feature image obtained in S7 to the original image size to obtain a fault recognition result image. Specifically, after performing a 3×3 convolution on the feature image obtained in S7, the feature image size is restored to the original image size through a 1×1 convolution kernel.

[0058] Specifically, the U-Net network consists of an encoding part and a decoding part, namely, an encoding-decoding network structure. The encoding part is used to extract the features of the image layer by layer. The input image passes through two convolution kernels of size 3×3 and is activated by the nonlinear activation function ReLU after each convolution. In downsampling, the feature image is max-pooled through a pooling layer with a step size of 2×2. This operation is used to extract the part with the strongest features in the image to reduce the size of the output features, thereby reducing the amount of calculation. The decoding part is used to perform feature fusion and restore the position information of the image. The feature image is deconvolved through upsampling. The deconvolution operation can halve the number of feature channels and double the size of the feature map, thereby reconstructing the fault features more accurately. Through jump connections, the feature image corresponding to the encoding part is delivered to the decoding part. After splicing, the deep and shallow feature information can be fused, thereby improving the network performance. The spliced ​​image is convolved twice with 3×3 and activated by the ReLU function. Finally, the image is restored to the original image size through a 1×1 convolution kernel.

[0059] The core idea of ​​the ResNet residual network is that each additional layer should more easily contain the original function as one of its elements.

[0060] The ResNet residual network consists of multiple residual blocks, among which the residual block is the core of solving the degradation problem. It converts several layers behind the deep network from learning the identity mapping function to learning the residual function, so that the model degenerates into a shallow network. The conversion formula is:

[0061]

[0062] In the formula, x is the input, F(x) is the residual function, and H(x) is the function to be fitted. Therefore, as long as F(x) = 0, an identity mapping H(x) = x is formed, which helps the neural network to fit efficiently.

[0063] There are two forms of residual blocks, Res-34 and Res-50. The specific structures of the two are as follows Figure 2 As shown. Figure 2 As shown in (a), in the Res-34 residual block, it is mainly composed of two 3×3 convolution kernels and two ReLU activation functions. When the input x passes through the module, x will pass through a 3×3 convolution kernel and then be activated by the ReLU function. Finally, it will pass through a 3×3 convolution kernel to obtain the residual function F(x). After obtaining F(x), the input x is linearly superimposed with it through the identity mapping. It can speed up the training speed of the model and improve the training effect of the model. After being activated by the ReLU function, the function H(x) to be fitted is finally obtained. Compared with the Res-34 module, in the Res-50 residual block (such as Figure 2As shown in (b), the Relu function is increased by one, and the number of convolution kernels becomes three, namely 1×1 convolution kernel, 3×3 convolution kernel, and 1×1 convolution kernel, where the 1×1 convolution kernel is used to adjust the number of channels to increase or decrease.

[0064] In terms of the size of the computational parameters, Res-50 is 68K, while Res-34 is 72K. Therefore, when the amount of computation is similar, Res-50 not only deepens the network layer and makes the network have stronger learning ability, but also does not reduce efficiency.

[0065] The convolutional attention module (CBAM) is a lightweight module that combines channel attention and spatial attention, which can pay more attention to the feature information in the image. After the feature map F is input into the CBAM module, attention feature information can be generated in two dimensions, channel and space, in a serialized manner. The two feature information are multiplied with the initial input features for adaptive feature correction, and finally a feature map F″ containing both channel key information and spatial key information is generated. The overall process can be expressed as:

[0066]

[0067] Where F′ is the channel weighted output; represents matrix multiplication; F″ is the final refined output.

[0068] The channel attention module and the spatial attention module can be placed sequentially or in parallel, but the results of sequential placement are better than those of parallel placement.

[0069] In a specific embodiment, the training process of the fault recognition model is as follows:

[0070] ①Dataset processing:

[0071] After the data set is enhanced, trimmed and normalized in turn, it is combined with the label data and used as training samples for the fault recognition model. Each image of the training sample is used as the original seismic image.

[0072] In the training process of the fault recognition model, the quality of the training data determines the training effect to a large extent. When the amount of training data is too small or the data label is wrong or the data noise interference is too large, the network is prone to overfitting during the training process, resulting in poor robustness of the model. Therefore, a certain amount of data sets must be prepared before training the network. At the same time, in order to improve the diversity of data and the generalization ability of the training model, the present invention enhances the data set and rotates the training data set at different angles. The coordinate transformation expression of the angle rotation is as follows:

[0073]

[0074] Where (v, w) is the coordinate of the pixel in the original image, (x, y) is the new coordinate of the pixel after rotation, and θ is the rotation angle.

[0075] After trimming and normalizing the enhanced data, 500 data blocks of size 512×512 are obtained. 60% of the data are used as training sets, 20% of the data are used as validation sets, and the rest of the data are used as test sets. The original seismic profiles, label data, and their superposition in the training data set are shown as follows: Figure 3 shown.

[0076] ②Software and hardware environment and parameter settings

[0077] The hardware and software configurations used are shown in Table 1.

[0078] Table 1 Hardware and software environment

[0079]

[0080] The parameters set in network model training are shown in Table 2.

[0081] Table 2 Training parameter settings

[0082] parameter Numeric BatchSize 1 Initial learning rate 0.0001 Optimizer Adam Epoch 50 Loss Function Balanced Cross Entropy

[0083] ③Weight initialization

[0084] In deep learning, the weight initialization method of the neural network model has a great impact on the convergence speed and performance of the model. In order to ensure that the model can learn effectively during the training process, the weight initialization should meet two conditions: the activation values ​​of each layer will not be saturated, and the activation values ​​of each layer are not 0. In addition, in order to prevent the occurrence of gradient disappearance and gradient explosion, the weight should not be too large or too small, so the variance (Var) of the weight should be controlled.

[0085] The ReLU activation function has an unsaturated characteristic, so the superposition of the ReLU activation function is very likely to cause problems such as the disappearance of neuron activity. For the initialization weight method of the ReLU activation function, this embodiment adopts the He initialization method to turn off half of the neurons in each layer of the network to reduce the variance of its distribution. Therefore, He initialization can be considered as the result of Xavier initialization divided by 2. He initialization is essentially a Gaussian distribution, which can be expressed by the formula:

[0086]

[0087] In the formula, w i is the initialization weight; n i is the number of neurons.

[0088] ④Model training

[0089] To prevent overfitting, this embodiment adds a Dropout layer to the fault recognition model. During each training, a certain number of convolution kernels are randomly stopped. This can improve the generalization ability of the network and effectively alleviate the occurrence of overfitting.

[0090] The present invention uses the Adam algorithm to train the Res-CBAM-UNet network model in the GTX1650Ti environment for 50 cycles, and the change curves of its loss and accuracy are as follows: Figure 4 and Figure 5 shown.

[0091] Specifically, a weight factor is added to the loss function, and the Adam algorithm is used to iteratively update the weight factor, where the loss function expression is:

[0092]

[0093]

[0094] Where N is the number of input earthquake data points, and m+n=N; m is the positive sample, that is, the number of fault points; n is the negative sample, that is, the number of non-fault points; y i is the fault label data of the i-th feature image, where 1 represents a fault point and 0 represents a non-fault point; α∈[0,1] is a weight factor, which assigns a high weight value to the fault point sample and a low weight value to the non-fault point sample; x i ∈(0,1) is the fault prediction probability of the i-th feature image.

[0095] The embodiment of the present invention adopts a balanced cross entropy loss function to avoid the problem of gradient diffusion leading to a decrease in learning rate, so as to increase the model's learning ability of fault information. At the same time, the number of non-fault point samples is much larger than the number of fault point samples, which will lead to an imbalance of samples and then a tilt of the loss function. One of the most effective methods to address these problems is to add a weight factor α to the loss function, assigning a high weight value to the fault point samples and a low weight value to the non-fault point samples, thereby increasing the weight of a minority of samples in the loss function and balancing the distribution of the loss function.

[0096] The Adam algorithm can iteratively update the weights of the neural network based on the training data to achieve the purpose of optimization. Unlike the traditional stochastic gradient descent algorithm, which can only update all weights at a fixed learning rate and the learning rate does not change during the training process, the Adam algorithm assigns different learning rates to different parameters by calculating the first-order moment estimate and the second-order moment estimate of the gradient.

[0097] The embodiment of the present invention also verifies the performance of the trained fault recognition model, as follows:

[0098] The fault recognition effect of the trained fault recognition model is verified through the validation set. The results are as follows: Figure 6 As shown. Figure 6 As can be seen in (c), the model has high fault recognition accuracy on the validation set, which is basically consistent with the label data.

[0099] Fault identification is a binary classification problem, in which the positive class (Positive) is the fault point and the negative class (Negative) is the non-fault point. In the positive class, if the sample belongs to the positive class and is predicted to be the positive class, it is called a true positive class (True Positive, TP), and the sample belongs to the negative class but is predicted to be the positive class, it is called a false positive class (False Positive, FP); similarly, if the sample belongs to the negative class and is predicted to be the negative class, it is called a true negative class (True Negative, TN), and the sample belongs to the positive class but is predicted to be the negative class, it is called a false negative class (False Negative, FN).

[0100] According to the above classification characteristics, this embodiment uses four indicators, namely, accuracy (A), F1 value, true positive rate (TPR) and false positive rate (FPR), to judge the quality of the network for fault identification. The specific definitions of the four indicators are as follows.

[0101] Accuracy, also known as correctness, is the most common classification performance indicator. Its value is the percentage of correct predictions in the total samples. The calculation formula is as follows:

[0102] Acc=(TP+TN) / (TP+FP+TN+FN).

[0103] However, when the number of positive and negative samples is unbalanced, the samples with a large proportion usually become the main factor affecting the accuracy. In fault identification, the number of non-fault points is generally much larger than the number of fault points. The network can also obtain a high accuracy by predicting most non-fault points as non-fault points. Therefore, for fault points, the accuracy is often more biased towards non-fault points. Therefore, accuracy alone cannot accurately evaluate the quality of network performance under the condition of imbalanced positive and negative samples.

[0104] TPR, also known as recall rate (Recall, R), indicates the ratio of true positive samples to all positive samples. The larger the TPR, the more accurate it is. FPR indicates the ratio of false positive samples to all negative samples. The smaller the FPR, the more accurate it is. TPR and FPR can be expressed by the formula:

[0105]

[0106]

[0107] When the number of positive and negative samples is unbalanced, the F1 value can be used as a comprehensive evaluation standard for the network. The higher the F1 value, the more accurate the network. The F1 value is an indicator to measure the accuracy of the binary classification model, taking into account both precision (Precision, P) and recall. Precision represents the percentage of true positive samples to all samples predicted as positive.

[0108] The F1 value is the weighted harmonic average of precision and recall. The specific formula is as follows:

[0109]

[0110]

[0111] In order to verify the advantages of the Res-CBAM-UNet network model, this paper selects the Res34-UNet and Res50-UNet network models for comparative analysis.

[0112] In terms of accuracy, Figure 7 As shown in the figure, the accuracy curves of Res50-UNet and Res34-UNet networks fluctuate greatly, and the final values ​​are lower than the accuracy of Res-CBAM-UNet network. The accuracy of Res-CBAM-UNet network is around 0.98 at the 15th iteration, and stabilizes at around 0.988 after 50 iterations, with little curve fluctuation. Figure 8 As shown in the figure, the initial losses of Res50-UNet and Res34-UNet are both around 0.01, and the loss curves still fluctuate greatly after 25 iterations. The initial loss of Res-CBAM-UNet is around 0.008, and after 25 iterations, the loss is lower than that of other networks and stabilizes at around 0.0008, with little curve fluctuation.

[0113] In the fault identification task of the present invention, due to the limitation of accuracy rate on network performance evaluation, the network should also be evaluated by using the average TPR, average FPR and average F1 value. The average TPR, average FPR and average F1 value of different networks are shown in Table 3.

[0114] Table 3 Average TPR, average FPR and average F1 value of different network models

[0115] Network Type TPR(%) FPR(%) <![CDATA[F1 value (%)]]> Res-CBAM-UNet 98.95 2.34 63.35 Res50-UNet 99.40 3.00 52.03 Res34-UNet 99.34 4.02 44.56

[0116] In Table 3, although the TPR of the Res-CBAM-UNet network is less than 0.5 percentage points lower than that of the other two networks, the FPR of the other two networks is relatively high, indicating that the network predicts an increase in the number of non-fault points as fault points, which will cause the identified fault lines to be crowded and swollen, making the boundaries between the identified faults unclear. Compared with the other two networks, the F1 value of the network of the present invention reached 63.35%, which is the best performance. Overall, although the TPR of the network in this embodiment is slightly lower, the FPR and F1 values ​​are outstanding, which proves the superiority of the network of the present invention in identifying faults.

[0117] Therefore, the Res-CBAM-UNet network proposed in the present invention, which integrates the Res-50 residual block and the CBAM attention mechanism, can, to a certain extent, make the model pay more attention to the feature information related to the fault, thereby improving the network's ability to identify faults.

[0118] In addition, in actual testing, the present invention uses actual seismic data from a certain work area, trims and expands the data graph to make it the input of the network, and finally obtains the fault identification result graph after network testing.

[0119] In terms of average test time, Res-CBAM-UNet is 690ms / sheet, Res50-UNet is 620ms / sheet, and U-Net is the slowest, at 2670ms / sheet. In terms of memory usage, Res50-UNet is 130.14MB, Res-CBAM-UNet is 130.67MB, and U-Net is 137.90MB. It can be seen that the network of the present invention is more efficient and requires less memory.

[0120] A random graph of the results after Res-CBAM-UNet, Res50-UNet and U-Net tests was selected for comparative analysis. The results are as follows Fig. 9 shown. Fig. 9 (a) is the original seismic profile, which contains a relatively obvious large fault and some small faults. Fig. 9 In (b), the faults identified by U-Net have poor continuity and are randomly distributed, making it impossible to find effective information. Fig. 9 In (c), the recognition effect of Res50-UNet can roughly show the location of the large fault and the extension shape of the fault, but the faults are adhered to each other, and the fault direction cannot be carefully identified. In addition, some background areas are also identified as faults, and the continuity is poor. Fig. 9 In (d), the fault edge identified by Res-CBAM-UNet is clearer, the fault line thickness is moderate, the fault background is clean and tidy, and the fault continuity is good.

[0121] Fig.10The recognition results of a certain area in the actual work area data using Res50-UNet, U-Net and Res-CBAM-UNet are given. Fig.10 (a) shows that there is an obvious normal fault in the area, and the result identified by U-Net contains four discontinuities (see Fig.10 (b) arrow), the fault continuity is very poor. The fault continuity identified by Res50-UNet is much better than that of U-Net, but the extension shape of the fault is inconsistent with expectations, and the fault bends and extends in other directions (see Fig.10 (c) elliptical dotted box), and some interference appears in the background area (see Fig.10 (c) Square dotted box). Compared with the previous two networks, Res-CBAM-UNet improves the continuity of fault recognition. The identified faults are clearer and have less irrelevant interference, but there are still discontinuities in the recognition results (see Fig.10 (d) Arrow).

[0122] By comparing and analyzing the test results of the actual work area seismic data on different networks, it can be seen that the Res-CBAM-UNet network model provided by the present invention can obtain more accurate fault identification effects, and the predicted faults have high accuracy and good continuity, which verifies the effectiveness and practicality of the model.

[0123] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0124] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An automatic earthquake fault identification method integrating Res-50 and CBAM, characterized in that: include: S1. Based on the U-Net network, multiple groups of Res-50 residual blocks are introduced in the encoding part of the U-Net network, and multiple CBAM convolutional attention modules are introduced in the decoding part to build a fault recognition model; S2, in the encoding part, the original seismic image is converted into the original feature map; S3, after downsampling, the original feature map enters the first group of Res-50 residual blocks for convolution processing, and the convolution result is linearly superimposed with the input and activated by the ReLU function to obtain a new feature map; S4, repeatedly down-sample the feature map obtained in S3 as shown in S3, input the next set of Res-50 residual blocks, and repeat the convolution process as shown in S3 until all Res-50 residual blocks are passed to obtain the feature map F of the encoding part; S5, after deconvolution of the feature map F through the upsampling layer, enter the first CBAM convolution attention module to perform CBAM operation, assign different weights to the fault area and background in the feature map F, and obtain a feature map F″ that significantly contains channel position information and spatial position information; S6, performing feature mapping cascade on the feature map F″ and the feature map F, and performing a convolution operation on the cascaded feature map; S7, repeatedly performing upsampling, CBAM operation, feature map cascade and convolution operation as shown in S5-S6 on the feature map obtained in S6, until all CBAM convolution attention modules are passed to obtain a feature image containing key information of the fault; S8, restoring the feature image obtained in S7 to the original image size to obtain a fault recognition result image.

2. The method for automatic identification of earthquake faults integrating Res-50 and CBAM according to claim 1 is characterized in that: Each downsampling layer in the encoding part corresponds to a group of Res-50 residual blocks, and the number of groups of downsampling layers and Res-50 residual blocks is 4; the encoding part sets 3, 4, 6, and 3 groups of Res-50 residual blocks from shallow to deep layers.

3. The automatic earthquake fault identification method integrating Res-50 and CBAM according to claim 2 is characterized in that: Each group of Res-50 residual blocks contains three consecutive Res-50 residual blocks, and each Res-50 residual block contains three consecutive convolution kernels of 1×1, 3×3 and 1×1.

4. The automatic earthquake fault identification method integrating Res-50 and CBAM according to claim 1 is characterized in that: S2 includes: performing 3×3 convolution operation, normalization processing and maximum pooling processing on the original seismic image in sequence to obtain the original feature map; wherein, the 3×3 convolution operation is used to extract the feature information in the input image; the normalization processing is used to adjust the size of each eigenvalue to a similar range to make their distribution consistent; the maximum pooling processing is used to reduce the dimension of the eigenvalue.

5. The automatic identification method of earthquake faults integrating Res-50 and CBAM according to claim 1 is characterized in that: In S5, the CBAM convolutional attention module includes a channel attention module and a spatial attention module, and the CBAM operations performed include: S51, input the feature map F of the encoded part into the channel attention module, and obtain the feature map F′ reflecting the fault feature channel information by recalibrating the feature of F; S52. The feature map F′ enters the spatial attention module and multiplies it with the spatial weight coefficient matrix to obtain a feature map F″ that significantly contains channel position information and spatial position information.

6. The method for automatic identification of earthquake faults integrating Res-50 and CBAM according to claim 1 is characterized in that: The feature map cascade operation in S6 is: concatenating the feature map F″ of the decoding part with the feature map F of the corresponding position of the encoding part.

7. The automatic earthquake fault identification method integrating Res-50 and CBAM according to claim 1 is characterized in that: In S8, the feature image obtained in S7 is subjected to a 3×3 convolution, and then the feature image size is restored to the original image size through a 1×1 convolution kernel.

8. The automatic identification method of earthquake faults integrating Res-50 and CBAM according to claim 1 is characterized in that: In the training phase of the fault identification model, a weight factor is added to the loss function, and the Adam algorithm is used to iteratively update the weight factor, where the loss function expression is: Where N is the number of input earthquake data points, and m+n=N; m is the positive sample, that is, the number of fault points; n is the negative sample, that is, the number of non-fault points; y i is the fault label data of the i-th feature image, where 1 represents a fault point and 0 represents a non-fault point; α∈[0,1] is a weight factor, which assigns a high weight value to the fault point sample and a low weight value to the non-fault point sample; x i ∈(0,1) is the fault prediction probability of the i-th feature image.

9. The automatic identification method of earthquake faults integrating Res-50 and CBAM according to claim 1 is characterized in that: Before S2, the method further includes: after performing enhancement processing, trimming and normalization processing on the data set in sequence, combining the label data and taking them together as training samples of the fault recognition model, and taking each image of the training sample as an original seismic image.

Images