Seismic fault identification method based on Dual-Attention mechanism and multi-scale fusion network

Abstract

The invention discloses a seismic fault identification method based on a Dual-Attention mechanism and a multi-scale fusion network, and belongs to the technical field of seismic data interpretation, and the method comprises the following steps: S1, obtaining three-dimensional post-stack seismic data, and carrying out the slicing of a seismic data body, and obtaining multi-channel seismic data; s2, a DAMFault Net fault recognition model is constructed; s3, completing fault feature extraction in a coding stage; s4, obtaining an intermediate aggregation feature and an original same-scale feature from an encoder in a decoding stage; s5, adding a depth supervision mechanism to a tail end decoder; and S6, splicing and recovering prediction results to the size of the original seismic section. According to the seismic fault identification method based on the Dual-Attention mechanism and the multi-scale fusion network, excellent accuracy and fault continuity identification capability are shown in a fault identification task, and automatic, efficient and accurate identification of the fault in the seismic image is realized; and powerful technical support is provided for seismic analysis and research.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic data interpretation, and in particular to a seismic fault identification method based on a Dual-Attention mechanism and a multi-scale fusion network. Background Technique

[0002] A fault is an important geological structure formed by the fracture of a rock formation under tectonic stress. It is an important factor controlling the formation and occurrence location of ore bodies, and is also a storage space and migration channel for underground resources. Accurately and efficiently identifying faults has important significance and application value in multiple fields such as oil and gas reservoir development, geophysical exploration, and engineering geology.

[0003] Current conventional fault interpretation methods mainly use computer-aided technology combined with seismic attributes to achieve automated or semi-automated identification methods. Commonly used seismic attributes include coherence, variance, curvature, and texture, etc., which effectively improve the accuracy and efficiency of identifying and interpreting faults from seismic profiles. However, these methods usually contain more parameters that need to be set according to experience, and these parameters are sensitive to various noises existing in actual data. On the other hand, with the increase in the amount of available seismic data, the computational cost of extracting seismic attributes is also continuously increasing.

[0004] In the field of computer vision, seismic data interpretation tasks such as fault identification, horizon tracking, and salt dome identification can all be regarded as image classification or semantic segmentation problems. Given the advanced performance achieved by convolutional neural networks (CNNs) in a wide range of computer vision tasks in recent years, more and more researchers have begun to apply CNN-based methods to seismic data interpretation. These new methods not only reduce the dependence on professional seismic interpreters, but also can better adapt to the large-scale growth of data and effectively reveal the laws hidden inside the data. Classic image segmentation models include FPN, U-Net, DeepLab series, and PSPNet, etc., and they have achieved good results in the application of seismic exploration. For example, Wu et al. (2019) proposed a three-dimensional seismic fault model FaultSeg3D based on the U-Net network and trained and tested it on simulated data and actual data. Gao et al. (2022) adopted a nested U-Net structure and multi-scale fusion operation on the basis of FaultSeg3D to improve the accuracy and reliability of detecting faults in complex and high-noise seismic images. Li et al. (2023) proposed a multi-scale residual block for mining fine-grained fault features from a low-dimensional feature space.

[0005] Although U-Net and its variant networks have demonstrated excellent performance in many applications, especially achieving encouraging results in tomographic image interpretation tasks, they still face some inherent limitations. Although previous improvement efforts have alleviated these problems to some extent, they have also brought additional complexity and computational burden. Specifically, increasing network parameters and layers to pursue higher performance often makes the model more prone to overfitting, which not only damages the model's generalization ability but also significantly increases the computational complexity and time overhead. In addition, in the selection of training data, directly using 3D data for training can provide richer spatial information, but this approach greatly increases the training difficulty of the model and poses higher requirements for computational resources and algorithm optimization. Therefore, it is necessary to find a better balance between improving performance and maintaining model simplicity and reducing computational costs. Summary of the Invention

[0006] The object of the present invention is to provide a seismic fault identification method based on the Dual-Attention mechanism and multi-scale fusion network, which exhibits excellent accuracy and fault continuity identification ability in the fault identification task, realizes automatic, efficient, and accurate identification of faults in seismic images, and provides strong technical support for seismic analysis and research.

[0007] To achieve the above object, the present invention provides a seismic fault identification method based on the Dual-Attention mechanism and multi-scale fusion network, including the following steps:

[0008] S1. Obtain three-dimensional post-stack seismic data, slice the seismic data volume, and obtain multi-channel seismic data suitable for model input;

[0009] S2. Based on the encoder-decoder mode, introduce ResNeSt as the backbone network in the encoder part of the U-Net++ network, and introduce multiple dual attention modules in the decoder part to construct the DAMFaultNet fault identification model;

[0010] S3. In the encoding stage, the seismic slice data obtained in step S1 undergoes a series of downsampling operations consisting of convolution and max pooling to complete fault feature extraction;

[0011] S4. In the decoding stage, after the output of S3 is processed by multiple nested decoding units, they are respectively converged to the end decoder of the same layer, using dense connection. Each encoder node fuses the aggregated features of the same scale to obtain intermediate aggregated features and the original same-scale features from the encoder;

[0012] S5. Add a deep supervision mechanism to the end decoder, and each path uses a different upsampling rate;

[0013] S6. The predicted results are stitched and restored to the size of the original seismic profile.

[0014] Preferably, the seismic data in S1 includes synthetic seismic data and actual seismic data. Multi-channel two-dimensional seismic image slices are extracted according to the main survey line or connecting survey line, and the seismic data set is scaled to the interval [0, 1] using the normalization method.

[0015] Preferably, in the training stage of the model, a mixed loss function based on cross-entropy, Tversky loss, and Focal loss is used as the optimization objective, and the AdamW algorithm is used as the optimizer to iteratively update the model parameters. The loss function L is specifically defined as:

[0016] L = ω1L bce + ω2L tversky + ω3L focal ;

[0017] where ω1, ω2, and ω3 are the weight factors of each part of the loss, and L bce is the binary cross-entropy loss, L tversky is the Tversky loss, and L focal is the Focal loss;

[0018] Given the true class label y and the predicted class label y', L bce adopts the binary cross-loss function, which is defined as:

[0019]

[0020] where μ is the sample weight coefficient, N is the number of sample points, y i is the true category of the i-th sample point, and y' i is the predicted result of the i-th sample point;

[0021] The definition of the Tversky loss function is as follows:

[0022]

[0023] where α and β are balance parameters, and by adjusting α and β, the attention degree of the model to positive and negative samples is adjusted. N is the number of sample points;

[0024] The definition of the Focal loss is as follows:

[0025]

[0026] where ω and γ are both adjustable parameters, ω ≥ 0 is the focusing parameter, and γ is the adjustment factor.

[0027] Preferably, in S4, DAMFaultNet fuses the high-resolution feature map from the encoder network with the corresponding semantically rich feature map from the decoder network through a hierarchical nested network.

[0028] Preferably, the dual attention module in S4 has the following specific operation steps:

[0029] S41: Use 1×1 convolution and the softmax function to obtain attention weights, and then obtain global context features through attention pooling operations;

[0030] S42: Obtain Conv1 and Conv2 of two bottleneck layer 1×1 convolutions to capture the dependencies between channels, and adjust the feature map according to the dependencies between channels;

[0031] Its expression is as follows:

[0032]

[0033] In the formula, i ∈ {1, 2, …, H×W}, where H is the height of the feature map and W is the width of the feature map, α j represents the global attention weight, Conv1 and Conv2 represent two different convolutional layers, LN is the layer normalization operation, ReLU is the activation function, is the module output, x i is the input feature value, and j represents the position index;

[0034] S43: Compress the spatial features of the feature map, and calculate the importance of the spatial information at each position in the feature map through 1×1 convolution and the sigmoid function;

[0035] S44: Concatenate the feature submaps of the global context features and the important spatial features.

[0036] Preferably, in DAMFaultNet, the Swish activation function is used to replace the ReLU activation function, and the expression of the Swish activation function is as follows:

[0037] swish(x) = x * sigmoid(ln(1 + e x ));

[0038] In the formula, x is the independent variable; sigmoid is the activation function.

[0039] Preferably, in S5, DAMFaultNet introduces a deep supervision mechanism and adds supervision signals in each layer of the nested network.

[0040] Therefore, the present invention adopts the above-mentioned seismic fault identification method based on the Dual-Attention mechanism and multi-scale fusion network, and has the following beneficial effects:

[0041] (1) Based on the advanced U-Net++ network architecture, ResNeSt is incorporated as the core backbone network in the encoding stage, significantly enhancing the network's ability to extract seismic fault features;

[0042] (2) In the feature decoding stage, a dual attention mechanism is introduced, which can intelligently identify and assign a greater proportion to higher-weight information, effectively suppressing semantic noise while highlighting the key features of the fault;

[0043] (3) A deep supervision mechanism is added to the network. By nesting supervision signals at all levels of the network, this mechanism ensures that the network can accurately capture fault features at different scales, thus greatly enhancing the generalization performance of the model;

[0044] (4) In terms of data preparation, the present invention implements comprehensive data augmentation processing on synthetic seismic data to generate a high-quality training dataset.

[0045] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Description of the Drawings

[0046] Figure 1 is a flowchart of a seismic fault identification method based on the Dual-Attention mechanism and multi-scale fusion network of the present invention;

[0047] Figure 2 is a schematic diagram of synthetic seismic data and fault labels of a seismic fault identification method based on the Dual-Attention mechanism and multi-scale fusion network of the present invention;

[0048] Figure 3 is a network architecture diagram of the DAMFaultNet fault identification model of a seismic fault identification method based on the Dual-Attention mechanism and multi-scale fusion network of the present invention;

[0049] Figure 4 is a schematic diagram of the ResNeSt unit module of a seismic fault identification method based on the Dual-Attention mechanism and multi-scale fusion network of the present invention;

[0050] Figure 5 is a schematic diagram of the decoder unit module of a seismic fault identification method based on the Dual-Attention mechanism and multi-scale fusion network of the present invention;

[0051] Figure 6 Schematic diagram of the dual attention mechanism module of a seismic fault identification method based on the Dual-Attention mechanism and multi-scale fusion network according to the present invention;

[0052] Figure 7 Loss and IoU curve graph during the model training process of a seismic fault identification method based on the Dual-Attention mechanism and multi-scale fusion network according to the present invention; where Figure 7 (a) in it is the loss graph during the model training process, where Figure 7 (b) in it is the IoU curve graph during the model training process;

[0053] Figure 8 Fault identification result graph of the DAMFaultNet fault identification model of a seismic fault identification method based on the Dual-Attention mechanism and multi-scale fusion network according to the present invention for the validation set; Figure 8 (a) in it is the original seismic profile with fault labels, Figure 8 (b) in it is the PSP prediction result, Figure 8 (c) in it is the FPN prediction result, Figure 8 (d) in it is the DeepLabV3++ prediction result, Figure 8 (e) in it is the U-Net prediction result, Figure 8 (f) in it is the DAMFaultNet prediction result;

[0054] Figure 9 Prediction result graph of a seismic fault identification method based on the Dual-Attention mechanism and multi-scale fusion network according to the present invention on the GSB dataset; Figure 9 (a) in it is the seismic profile with fault markings, Figure 9 (b) in it is the DAMFaultNet prediction result, Figure 9 (c) in it is the U-Net prediction result, Figure 9 (d) in it is the CNN prediction result;

[0055] Figure 10 Prediction result graph of a seismic fault identification method based on the Dual-Attention mechanism and multi-scale fusion network according to the present invention on the main survey line of the F3 dataset; Figure 10 (a) in it is the seismic profile with fault markings, Figure 10 (b) in it is the DAMFaultNet prediction result, Figure 10 (c) in it is the U-Net prediction result;

[0056] Figure 11It is the prediction result diagram of a seismic fault identification method based on the Dual-Attention mechanism and multi-scale fusion network of the present invention in the connection survey line and the Z-axis direction of the F3 dataset;

[0057] Figure 11 In (a) is the prediction result of DAMFaultNet; Figure 11 In (b) is the prediction result of U-Net. Specific implementation manner

[0058] The technical solution of the present invention will be further described below through the accompanying drawings and embodiments.

[0059] Unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meaning understood by those of ordinary skill in the art to which the present invention belongs.

[0060] Embodiment 1

[0061] As Figure 1 shown, the present invention provides a seismic fault identification method based on the Dual-Attention mechanism and multi-scale fusion network, including the following steps:

[0062] S1. Obtain three-dimensional post-stack seismic data, slice the seismic data volume, and obtain multi-channel seismic data suitable for model input;

[0063] Use the Ricker wavelet to calculate the three-dimensional synthetic seismic response through the convolution model, and incorporate an appropriate amount of noise during this process to ensure that the obtained results are closer to the real seismic situation. In order to introduce necessary randomness, the model parameters including fault position, dip angle, dip direction and strike are randomly set. The fundamental purpose of this method is to accurately simulate the rich diversity and high complexity exhibited in seismic data.

[0064] As Figure 2 shown, Figure 2 shows the synthetic seismic data and fault label schematic diagram of the embodiment of the present invention. The first column and the third column in the figure are seismic slice data, and the corresponding fault interpretation is defined by the binary image corresponding to the slice data, and its resolution is the same as that of the seismic data of the pixels identified as fault (1) or non-fault (0).

[0065] The seismic data includes synthetic seismic data and actual seismic data. Multi-channel two-dimensional seismic image slices can be extracted according to the main survey line or the connection survey line, and the seismic data set is scaled to the interval [0, 1] using the normalization method.

[0066] The clipping of the original seismic data used an overlapping clipping strategy (16 - pixel overlap). Finally, a total of 11,108 pairs of samples with a size of 5×64×64 were obtained, and then the training samples were randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0067] The prepared training seismic data set used data augmentation techniques, including two variants of random rotation, grid distortion, horizontal and vertical flipping, transposition, a combination of vertical flipping and random rotation, random brightness, random contrast, random brightness contrast, and random gamma to increase the image diversity during the training process.

[0068] In the training stage of the model, a hybrid loss function based on cross - entropy, Tversky loss, and Focal loss was used as the optimization objective, and the AdamW algorithm was used as the optimizer to iteratively update the model parameters. The loss function L was specifically defined as:

[0069] L = ω1L bce +ω2L tversky +ω3L focal ;

[0070] where ω1, ω2, and ω3 are the weight factors of each part of the loss, L bce is the binary cross - entropy loss, L tversky is the Tversky loss, and L focal is the Focal loss;

[0071] Given the true class label y and the predicted class label y', L bce adopts the binary cross - loss function, which is defined as:

[0072]

[0073] where μ is the sample weight coefficient, N is the number of sample points, y i is the true class of the i - th sample point, and y' i is the prediction result of the i - th sample point;

[0074] The definition of the Tversky loss function is as follows:

[0075]

[0076] where α and β are balance parameters. By adjusting α and β, the attention degree of the model to positive and negative samples can be adjusted, and N is the number of sample points;

[0077] The definition of the Focal loss is as follows:

[0078]

[0079] Where ω and γ are both adjustable parameters, ω≥0 is the focusing parameter, and γ is the adjustment factor.

[0080] S2. Based on the encoder-decoder mode, ResNeSt is introduced as the backbone network in the encoder part of the U-Net++ network, and multiple Dual-Attention modules are introduced in the decoder part to construct the DAMFaultNet fault identification model.

[0081] As Figure 3 shown, Figure 3 Fig. shows the network architecture diagram of the DAMFaultNet fault identification model according to the embodiment of the present invention. The network realizes the utilization of multi-scale information through a nested structure and dense skip connections, overcoming the influence of the semantic differences between shallow features and deep features on the segmentation result. Different from the conventional symmetric encoder-decoder structure, DAMFaultNet fuses the high-resolution feature maps from the encoder network with the corresponding semantically rich feature maps from the decoder network through a hierarchical nested network, gradually enriching the model. And dense skip connections are introduced between the encoder and the decoder, and these connections can reduce the feature map gap between the encoding and decoding sub-networks.

[0082] S3. In the encoding stage, the seismic slice data obtained in step S1 undergoes a series of downsampling operations including convolution and max pooling to complete the extraction of fault features.

[0083] The specific encoder structure is as Figure 4 shown. The input is divided into K groups in the channel dimension. For each base group, it is further grouped, and different feature transformations are applied to each group, and then a voting mechanism is applied to form attention. Finally, all the outputs are accumulated to form the final output result.

[0084] S4. In the decoding stage, after the output of S3 is processed by a multi-level embedded decoding unit, it converges to the end decoder of the same layer respectively, and dense connections are adopted. Each encoder node fuses the aggregated features of the same scale, obtaining the intermediate aggregated features and the original features of the same scale from the encoder.

[0085] DAMFaultNet realizes the utilization of multi-scale information through a nested structure and dense skip connections, overcoming the influence of the semantic differences between shallow features and deep features on the segmentation result. Different from the conventional symmetric encoder-decoder structure, DAMFaultNet fuses the high-resolution feature maps from the encoder network with the corresponding semantically rich feature maps from the decoder network through a hierarchical nested network, gradually enriching the model.

[0086] Through the lateral output layer, the feature maps of different scales constructed by the encoder are fused layer by layer. Using x i,jDenote the output of the j-th node \(X\) in the \(i\)-th layer, then its corresponding mathematical expression is: i,j

[0087]

[0088] In the formula, DConv represents the decoder convolutional block, Concat represents the feature concatenation function, \([*]\) represents the object list, and Up represents the upsampling method.

[0089] As Figure 5 shown, Figure 5 shows the internal structure diagram of the decoder unit of the embodiment of the present invention. In the decoder unit, the information with a larger weight ratio is automatically selected through the dual attention mechanism to suppress semantic differences and highlight significant features.

[0090] The specific operation steps of the dual attention module are as follows:

[0091] S41. Obtain the attention weights by using 1×1 convolution and the softmax function, and then obtain the global context features through the attention pooling operation;

[0092] In DAMFaultNet, the Swish activation function is used to replace the ReLU activation function. The expression of the Swish activation function is as follows:

[0093] swish(x) = x * sigmoid(ln(1 + e x ));

[0094] In the formula, \(x\) is the independent variable; sigmoid is the activation function.

[0095] S42. Obtain the capture of the dependencies between channels through Conv1 and Conv2 of two bottleneck layers of 1×1 convolution, and adjust the feature map according to the dependencies between each channel;

[0096] Its expression is as follows:

[0097]

[0098] In the formula, \(i\in\{1,2,\ldots,H\times W\}\), where \(H\) is the height of the feature map and \(W\) is the width of the feature map, \(\alpha\) j represents the global attention weight, Conv1 and Conv2 represent two different convolutional layers, LN is the layer normalization operation, ReLU is the activation function, is the module output, \(x\) i is the input feature value, and \(j\) represents the position index;

[0099] ​S43. Compress the spatial features of the feature map, and calculate the importance of the spatial information at each position in the feature map through 1×1 convolution and the sigmoid function;

[0100] S44. Concatenate the feature submaps of the global context feature and the important spatial feature.

[0101] As Figure 6 shown, Figure 6 shows the internal structure diagram of the dual attention module of the embodiment of the present invention. The module enhances the network's learning ability of long-range dependence relationships and spatial features in the feature map by calculating the importance degrees of the global and spatial features of the feature map, and optimizes the object detection effect.

[0102] S5. Add a deep supervision mechanism to the end decoder, and each path uses a different upsampling rate;

[0103] DAMFaultNet introduces a deep supervision mechanism. By adding supervision signals in the nested networks of each layer, it improves the multi-scale detection performance of the network, and also enhances the adaptability to images of different sizes, and improves the generalization of the network.

[0104] S6. The prediction result is concatenated and restored to the size of the original seismic profile, and the size of the prediction result is 128×128.

[0105] In a specific example, the training process of the fault recognition model is as follows:

[0106] In this example, considering the unbalanced distribution of fault pixels, a stratified sampling method is adopted, and the proportion of fault pixels is used as the sampling basis. Subsequently, the dataset is randomly divided into three subsets, which are respectively designated as the training set, the validation set and the test set, with a ratio of 8:1:1. The AdamW optimizer is used for model training. Considering the limitation of the GPU memory capacity, the batch size is set to 64. The initial learning rate of the optimizer is set to 0.0008, and the weight decay coefficient is set to 0.01.

[0107] During the training process, the learning rate is dynamically adjusted by monitoring the performance index IoU. Once this value remains unchanged within six training epochs, the learning rate will be halved. In addition, an early stopping mechanism is also adopted to prevent the model from overfitting. Specifically, the training process is monitored to observe whether the validation set loss is decreasing. When the loss does not decrease after 12 epochs, the training process terminates. The maximum number of training epochs in the experiment is set to 200, and the evaluation metrics of the model are recorded in each epoch of the training set and the validation set.

[0108] As Figure 7 shown, Figure 7 shows the loss and IoU curves of the model training process of the embodiment of the present invention. Among them, Figure 7In (a) is the loss graph of the model training process, Figure 7 In (b) is the IoU curve graph of the model training process, as Figure 7 shown. Since the validation loss did not increase, the early stopping mechanism was triggered and the training stopped after 90 iterations.

[0109] To quantitatively compare the prediction results, we use Accuracy, Precision, F1-score, and Intersection over Union (IoU) as the four main metrics. In the task of fault prediction, we regard fault pixels as positive (1) and non-fault pixels as negative (0). As shown in Table 1, True Positive (TP) refers to the number of pixels accurately predicted as faults, while True Negative (TN) refers to the number of pixels accurately predicted as non-faults. False Positive (FP) and False Negative (FN) represent the mispredicted fault pixels and non-fault pixels respectively.

[0110] Table 1 Confusion Matrix

[0111]

[0112] The definitions of the four metrics are as follows:

[0113]

[0114] where Recall is the recall rate.

[0115] The embodiments of the present invention verified the performance of the fault recognition model as follows:

[0116] First, a comparative analysis was carried out on the synthetic seismic dataset and the existing SOTA method. For a fair comparison, the same parameters were used throughout the process, and the pre-trained ImageNet parameters were used to initialize the experiments and model parameters. The test set consists of 1000 slices, each with a size of 5×64×64. This dataset was used to quantitatively evaluate the model, and the results are shown in Table 2.

[0117] Table 2 Comparison of Experimental Results on the Test Set

[0118]

[0119] As shown in Table 2, DAMFaultNet performs best in four evaluation metrics, showing an accuracy value of 0.9843, an F1 score of 0.7742, an IoU of 0.6024, and an AUROC of 0.8759. It should be noted that DAMFaultNet is superior to the other four models in many aspects and has more significant improvements compared with the SOTA model. Although U-Net performs similarly to the proposed model in terms of accuracy, its lower F1 score and IoU indicate that it misses many fault structures and generates a large number of FNs in the prediction. In addition, after introducing the dual attention mechanism, the traditional U-Net model has been significantly improved. This demonstrates the effectiveness of the dual attention mechanism in enhancing feature fusion and fault recognition of the U-Net network.

[0120] To more intuitively understand the differences in fault recognition performance among various models, the proposed model was qualitatively evaluated using a dataset of test visualization of fault recognition results. As Figure 8 shown, Figure 8 shows the fault recognition results of the DAMFaultNet fault recognition model of the embodiment of the present invention for the validation set; among them, Figure 8 (a) in is the original seismic profile with fault labels, Figure 8 (b) in is the PSP prediction result, Figure 8 (c) in is the FPN prediction result, Figure 8 (d) in is the DeepLabV3++ prediction result, Figure 8 (e) in is the U-Net prediction result, Figure 8 (f) in is the DAMFaultNet prediction result.

[0121] To train and evaluate the proposed model, 2 publicly available real seismic datasets were used in the experiment. The first dataset is the GSB dataset, which comes from the Great South Basin in New Zealand and has a size of 5×76×484. The second dataset is the F3 dataset, which is provided by TNO and dGB Earth Sciences and has a size of 415×96×384.

[0122] As Figure 9 shown, Figure 9 is a comparison of the prediction results of different models in the GSB dataset in the embodiment of the present invention. As Figure 9 (a) in shows, the seismic data of GSB includes various seismic faults, most of which are vertical. These faults are particularly obvious on the reflection surface. Figure 9The subgraphs of (b)-(d) present the predicted fault results of the DAMFaultNet model, the U-Net model proposed by An et al. (2021), and the method used by Cunha et al. (2020). Compared with the actual labels, several unlabeled faults were also found and showed good continuity, as shown by the red boxes. In addition, in some areas without fault features, some possible faults were also found, as shown by the green arrows.

[0123] As Figure 10 shown, Figure 10 This is the prediction result on the main survey line of the F3 dataset in the embodiment of the present invention. Figure 10 (a) in depicts the true values of the 160 main survey line seismic data and the fault distribution. Figure 10 (b) in shows the results obtained when using the DAMFaultNet network proposed by us. Figure 10 (c) in is the prediction result of the U-Net method using ResNeSt101 as the encoder. As Figure 10 (c) in shows, the faults extracted by U-Net show discontinuous features, as shown by the green arrows. Compared with U-Net, the prediction results produced by DAMFaultNet are more consistent with the labels. Although DAMFaultNet can predict a wider range of faults than U-Net, as Figure 10 (b) in the red box shows, its prediction results do not result in more fault predictions than the labels. This highlights the significant impact of the training set on the network's recognition ability. Both U-Net and DAMFaultNet demonstrate the potential for practical applications in the data field. However, due to DAMFaultNet enhancing the multi-scale fusion ability, its predicted fault results are more complex than U-Net, and the fault lines are more complete.

[0124] As Figure 11 shown, Figure 11 This is the prediction result on the connecting survey line and the Z-axis direction of the F3 dataset in the embodiment of the present invention. Figure 11 (a) in is the prediction result of DAMFaultNet, Figure 11 (b) in is the prediction result of U-Net. Figure 11 (c) in is the prediction result of DAMFaultNet in the Z-axis direction; Figure 11 (d) in is the prediction result of U-Net in the Z-axis direction. The results verify the effectiveness of our model and highlight the significant advantages of DAMFaultNet in improving the model performance and prediction accuracy.

[0125] Therefore, the present invention adopts the above-mentioned seismic fault recognition method based on the Dual-Attention mechanism and multi-scale fusion network, which demonstrates excellent accuracy and fault continuity recognition ability in the fault recognition task, realizes the automatic, efficient and accurate recognition of faults in seismic images, and provides strong technical support for seismic analysis and research.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for earthquake fault identification based on dual-attention mechanism and multi-scale fusion network, characterized by: The following steps are involved: S1, obtaining three-dimensional post-stack seismic data, slicing the seismic data volume, and obtaining multi-channel seismic data suitable for model input; S2. Based on the encoder-decoder model, ResNeSt is introduced as the backbone network in the encoder part of the U-Net++ network, and multiple dual attention modules are introduced in the decoder part to build the DAMFaultNet fault recognition model; S3, in the encoding stage, the seismic slice data obtained in step S1 is subjected to a series of downsampling operations through convolution and maximum pooling to complete the fault feature extraction; S4, in the decoding stage, after the output of S3 is processed by the multi-level embedded decoding unit, it is respectively gathered at the end decoder of the same layer, using dense connection, each encoder node fuses the aggregated features of the same scale, and obtains the intermediate aggregated features and the original same-scale features from the encoder; S5. Add a deep supervision mechanism to the end decoder, and use a different upsampling rate for each path; S6. The prediction results are spliced ​​and restored to the original earthquake profile size.

2. The earthquake fault identification method based on the dual-attention mechanism and multi-scale fusion network according to claim 1, characterized in that: The seismic data in S1 include synthetic seismic data and actual seismic data. Multi-channel two-dimensional seismic image slices are extracted according to the main survey line or the contact survey line, and the seismic data set is scaled to the interval [0,1] using the normalization method.

3. The earthquake fault identification method based on the dual-attention mechanism and multi-scale fusion network according to claim 2 is characterized in that: In the training phase of the model, a hybrid loss function based on cross entropy, Tversky loss and Focal loss is used as the optimization target, and the AdamW algorithm is used as the optimizer to iteratively update the model parameters. The loss function L is specifically defined as: L=ω1L bce +ω2L tversky +ω3L focal ; In the formula, ω1, ω2, ω3 are the weight factors of each part of the loss, L bce is the binary cross entropy loss, L tversky For Tversky's loss, L focal For Focal loss; Given the true class label y and the predicted class label y', L bce The binary cross loss function is used, which is defined as: In the formula, μ is the sample weight coefficient, N is the number of sample points, and y i is the true category of the i-th sample point, y' i is the prediction result of the i-th sample point; The Tversky loss function is defined as follows: In the formula, α and β are balance parameters. By adjusting α and β, the model's attention to positive and negative samples can be adjusted. N is the number of samples. The definition of Focal loss is as follows: Wherein, ω and γ are both adjustable parameters, ω ≥ 0 is the focusing parameter, and γ is the adjustment factor.

4. The earthquake fault identification method based on dual-attention mechanism and multi-scale fusion network according to claim 1, characterized in that: DAMFaultNet in S4 fuses the high-resolution feature maps from the encoder network with the corresponding semantically rich feature maps from the decoder network through a hierarchical nested network.

5. The earthquake fault identification method based on dual-attention mechanism and multi-scale fusion network according to claim 4 is characterized in that: The specific operation steps of the dual attention module in S4 are as follows: S41, use 1×1 convolution and softmax function to obtain attention weights, and then obtain global context features through attention pooling operation; S42, Conv1 and Conv2 through two bottleneck layers of 1×1 convolution obtain the dependency between the captured channels, and adjust the feature map according to the dependency between the channels; Its expression is as follows: In the formula, i∈{1,2,…,H×W}, where H is the feature map height, W is the feature map width, and α j Represents the global attention weight, Conv1 and Conv2 represent two different convolutional layers, LN is the layer normalization operation, and ReLU is the activation function. is the module output, x i is the input feature value, j represents the position index; S43, compress the spatial features of the feature map, and calculate the importance of spatial information at each position in the feature map through 1×1 convolution and sigmoid function; S44, feature subgraph that concatenates global context features and important spatial features.

6. The earthquake fault identification method based on dual-attention mechanism and multi-scale fusion network according to claim 5, characterized in that: In DAMFaultNet, the Swish activation function is used instead of the ReLU activation function. The expression of the Swish activation function is as follows: swish(x)=x*sigmoid(ln(1+e x )); In the formula, x is the independent variable and sigmoid is the activation function.

7. The earthquake fault identification method based on dual-attention mechanism and multi-scale fusion network according to claim 1, characterized in that: In S5, DAMFaultNet introduces a deep supervision mechanism and adds supervision signals in each layer of the nested network.

Images