Three-dimensional seismic fault identification method based on ResSC-UNet

Abstract

The invention relates to a seismic fault identification method, in particular to a three-dimensional seismic fault identification method based on ResSC-UNet. The method comprises the following steps: 1, making a three-dimensional seismic fault data set, and generating a one-dimensional horizontal reflectivity model; using a vertical shear model to act on the reflectivity model to create a wrinkle structure; generating an inclined structure by using the linear displacement diagram; a plane fault is added into the wrinkle model, and the generated fault has diversity by controlling the dip angle, the trend and the fault displacement range of the fault and selecting random reference points; carrying out convolution on the reflection model by using wavelets to obtain sharp discontinuity near the fault of the model; gaussian noise with different signal-to-noise ratios is added; 2, constructing a deep learning network model; and 3, establishing a ResSC-UNet network model. According to the three-dimensional seismic fault identification method, improvement is carried out based on a U-Net network model, an efficient model is finally trained, and fault information can be efficiently and accurately identified from seismic data.

Description

Technical Field

[0001] The present invention relates to a method for identifying earthquake faults, and in particular to a three-dimensional earthquake fault identification method based on ResSC-UNet. Background Art

[0002] In underground exploration, faults are fracture structures caused by relative movement between strata in the crust. They range in size from meters to kilometers and are one of the important structures in the crust. Faults can intuitively reflect the geological and oil and gas reservoir conditions such as underground reservoir description, trap evaluation, and well layout, which greatly affects the efficiency and benefits of oil and gas reservoir development. With the development of geophysical exploration, the collected seismic data are becoming more and more large and complex, so the requirements for fault identification technology are also constantly increasing. The existing conventional identification methods and identification methods based on seismic attributes are no longer applicable to existing seismic data. For this reason, the scientific and industrial communities are actively exploring and promoting emerging technology paths, such as deep learning methods driven by big data, which are being gradually applied.

[0003] Existing conventional fault identification technologies can be divided into three categories: seismic profile analysis, well breakpoint identification, and well-seismic joint identification technology. Seismic profile identification is one of the most commonly used fault identification methods. When seismic waves encounter underground lithology interfaces or geological anomalies such as faults, reflection, refraction, and diffraction will occur. These phenomena will be received and recorded by detectors arranged on the surface to form seismic recording profiles. Seismic interpreters analyze seismic reflection profiles and manually pick up faults along the survey line in the direction of the main survey line or the survey line perpendicular to the fault strike. However, this method relies on seismic characteristics such as reflection and diffraction generated by seismic waves when encountering faults to identify the existence and location of faults, which is greatly affected by the quality of seismic data. If the seismic data has noise interference or insufficient resolution, the accuracy of the identification results may be affected. Well breakpoint identification is a technology that uses underground breakpoint information directly obtained during drilling to identify faults. During the drilling process, when the drill bit passes through the fault, special lithology changes, missing or repeated formations and other geological phenomena will be recorded. This information is of great value for the identification and interpretation of faults. However, in practical applications, it is necessary to pay attention to its shortcomings such as being limited by drilling density, strong subjectivity in interpretation, difficulty in identifying low-order faults, and high cost. Well-seismic joint identification is a technical method that combines well data and seismic data to identify underground geological structures such as faults. This method makes full use of the high resolution of well data in the vertical direction (depth direction) and the high accuracy of seismic data in the horizontal direction (plane direction). By constructing synthetic seismic records and performing deep-time conversion, accurate matching of well-seismic data is achieved, and the three-dimensional morphology and spatial distribution of faults are then portrayed. However, this method has high requirements on the professional quality and technical level of the interpreters, and requires a lot of data processing and calculation work.

[0004] The identification method using seismic attributes is mainly because faults often appear as sudden changes or discontinuities in the reflection layer in seismic data. The seismic attributes are extracted and calculated through seismic data, and the unreasonable fault picking is supplemented accordingly based on the experience of the interpreters, so as to finally obtain a clear fault interpretation image. These attributes include coherence volume attributes, curvature attributes, variance attributes and ant volume recognition methods. For coherence volume technology, the similarity of seismic data is mainly calculated, and the results are displayed through coherence attributes. In the displayed calculation results, through coherence changes, the faults can be highlighted, and the overall spatial development characteristics of the geological structure can be characterized. For this method, not only the quantitative identification of faults is realized, but also the interpretation errors can be greatly avoided, and the accuracy of fault interpretation can be further improved. The curvature attribute is also one of the better attributes used in many fault identifications in recent years, with high noise resistance and three-dimensional visualization interpretation functions. The variance volume technology calculates the variance between adjacent seismic traces by calculating the probability variance analysis. By characterizing the differences in the reflection characteristics of each seismic trace, it further realizes the identification of stratigraphic discontinuities, and finally the identification of faults. The ant volume recognition method uses the amplitude, phase, frequency and other attributes in the seismic data to identify and track fault lines in seismic images by simulating the path selection and pheromone marking mechanism of ant colonies during foraging. In the algorithm, "ants" act as search units, moving in the seismic data volume and releasing "pheromones" to indicate possible fault paths. Although the above attribute volume-based recognition methods have made some progress compared to manual recognition, they still have problems of low efficiency and discontinuous recognition results, and are still interfered by subjective factors.

[0005] With the vigorous development of artificial intelligence technology, machine learning algorithms have also been used in the field of fault identification, including BP neural network, support vector machine, generative adversarial network and clustering algorithm. In the fault identification task, by extracting multi-dimensional seismic wave dynamic characteristic parameters, the BP neural network model is trained to achieve accurate identification of faults. As a pioneer in the field of neural networks, BP neural network can effectively and excellently perform fault automatic identification tasks with its excellent generalization ability. However, with the rapid development of science and technology, more stringent requirements are put forward for the accuracy and efficiency of fault identification. In this context, the limitations of BP neural network gradually emerge, such as slow network convergence speed, high sensitivity to initial weights and lack of stability. Support vector machine also plays an important role in the field of fault automatic identification. Its core advantage lies in the accurate selection of kernel function and adjustment of penalty parameters to optimize the model. However, support vector machine essentially focuses on constructing the optimal classification plane, which makes it more suitable for processing small-scale data sets. When facing large-scale and complex three-dimensional seismic data, it is difficult to maintain efficient and accurate automatic identification. The rise of generative adversarial network (GAN) has brought a new perspective to seismic image processing. GAN consists of a generator and a discriminator. Through the dynamic game mechanism between the two, the generator is prompted to accurately capture the intrinsic characteristics of the data set. In recent years, GAN has been increasingly used in the field of seismic image preprocessing, showing great potential. The clustering algorithm uses unsupervised learning to identify discontinuities in seismic event axes and effectively classify them, trying to merge discontinuities belonging to the same fault into the same cluster, so as to achieve a clear display of the fault line. Although the clustering algorithm is easy to operate and does not require a tedious fault labeling process, its characteristics as an unsupervised learning method also limit its further improvement in fault identification accuracy.

[0006] The rapid development of computer software and hardware has become the core force driving innovation and development in all walks of life. In the field of fault recognition, deep learning has demonstrated unprecedented potential and advantages by virtue of its ability to automatically extract deep features from massive data. Wang Jing et al. (2021) combined the residual network with the 3D U-Net network and used the knowledge distillation method to greatly improve the accuracy of fault recognition (WANG J, ZHANG JH, ZHANG JL, et al. Research on fault recognition method combining 3D Res-UNet and knowledge distillation [J] Applied Geophysics, 2021, 18 (2): 199-212). Yu (2022) added a lightweight channel attention module SE to the end of each layer of the left encoder. Without significantly increasing the parameters, the model also has good prediction accuracy and recognition continuity (YU T, WANG X, CHEN TJ, et al. Fault Recognition Method Based on Attention Mechanism and the 3D-UNet[J]. Computational Intelligence and Neuroscience, 2022, 2022). Li Qingwu et al. (2023) used the 3D U-Net++ method to identify small faults, which improved the efficiency for both conventional and complex seismic data (Li Qingwu, Wang Xingjian, Zhang Yongheng, et al. Fault Recognition Method and Application Based on 3D U-Net++ Convolutional Neural Network[J]. Computational Technology of Geophysical and Geochemical Exploration, 2024, 46(03):284-291). Summary of the invention

[0007] In order to solve the problems existing in the prior art, the present invention provides a three-dimensional seismic fault recognition method based on ResSC-UNet. The three-dimensional seismic fault recognition method is improved based on the U-Net network model. A diversified three-dimensional seismic data set is constructed through earthquake folding, fault structure, wavelet peak frequency and noise. The model is trained with an improved hybrid loss function. Through the powerful nonlinear ability of deep learning, the intrinsic correlation between these data sets is deeply mined and mapped, and finally an efficient model is trained.

[0008] The technical solution adopted by the present invention is: a three-dimensional seismic fault identification method based on ResSC-UNet, and the three-dimensional seismic fault identification method comprises the following steps: The first step is to create a three-dimensional seismic fault data set. In the training process of the neural network model, the data set is particularly important. The core of deep learning is to deeply analyze the internal logic and relationship between data, and the final performance of the model depends directly on the quality of the features learned from the training data. Therefore, the quality of the data set is the key to the success of neural network training. Creating a diverse and accurate training data set can guide the network to capture more accurate and comprehensive data features and show better recognition effects.

[0009] The production of 3D seismic fault dataset includes the following steps: 1) Generate a one-dimensional horizontal reflectivity model r ( x, y, z ), which consists of a series of random values ​​in the range [–1, 1]; 2) Use the vertical shear model to act on the reflectivity model to create some wrinkle structures and vertical displacement maps S The formula for 1 is: (1) S 1 by N It consists of a two-dimensional Gaussian function and a linear scalar function, where the parameters a0, b k , c k , d k , σ k Through random combination within a certain range of values, it is used to produce folding structures with some specific spatial variations; x , y , z They are respectively x axis, y axis, z axis, z max for z Maximum value of the direction, 1.5z / z max As a linear scalar, z Scaling can suppress folding from bottom to top; k Denote the number of Gaussian functions by, b k It is k The weight coefficient of the item, ( x−c k ) 2 + ( y−d k ) 2 is a two-dimensional Euclidean distance square, indicating the coordinate point (x , y ) and the center point ( c k , d k ), σ k For the k The scale parameter of the item controls the width of the Gaussian distribution; using the displacement map S 1. Use sinc Interpolation is used to vertically displace the original reflectivity model r to obtain the wrinkle model r (x, y ,z+ S 1(x, y ,z)); 3) Based on step 2, use the linear displacement map S 2 Generate inclined structure, the formula is defined as: (2) e 0 is a constant term, representing the base bias value, f 0 and g 0 is a random number within the limit range, controlling x Coordinates and y The slope of the coordinates will eventually generate an inclined surface of a two-dimensional plane. S 2 loops are applied to the wrinkle model r to obtain a new reflectivity model r (x, y ,z+ S 1(x, y ,z)+ S 2(x, y ,z)); 4) Add plane faults to the fold model, control the fault dip, strike, fault range and select random reference points ( x 0. y 0. z 0), so that the generated faults are diverse; 5) Use wavelets to convolve the reflection model and model sharp discontinuities near faults; 6) Add Gaussian noise with different signal-to-noise ratios; The second step is to build a deep learning network model, integrate the multi-scale residual module into the U-Net framework, and further enhance the retention and transmission of spatial feature information through the residual connection strategy of the 1×1×1 convolutional layer; When building a deep network, a deeper network architecture is used to capture richer information, but simply stacking network layers may encounter the challenge of network degradation, that is, as the number of layers increases, the performance decreases instead of increases. The deep residual network introduces the residual module, which cleverly transforms some convolutional layers in the network. The core of the residual module is that it allows the network to automatically identify and remove redundant layers during training, and avoids the occurrence of network degradation by achieving identity mapping. This not only successfully overcomes the bottleneck of deep network degradation, but also greatly improves the learning efficiency and stability of the network. The residual function formula is: (3) Among them, x represents the input feature, F (x) contains multiple convolutional layers and activation functions, H (x) is the result obtained after residual error; Seismic fault images often present complex shapes with intricate folds and dense folds. The boundaries of the rock layers are blurred and difficult to define, which can easily lead to confusion with non-seismic related geological structures. To solve this problem, a multi-scale residual module is incorporated into the U-Net framework to replace the traditional encoder-decoder hierarchical unit. The purpose is to achieve in-depth analysis of multiple layers of resolution, streamline model parameters, and further enhance the retention and transmission of spatial feature information through the residual connection strategy of the 1×1×1 convolutional layer.

[0010] The core of the multi-scale residual module is that it can use convolution kernels of different sizes (such as 3×3×3, 5×5×5, etc.) to perform multi-scale convolution operations on the input feature map, thereby capturing and extracting the feature information of the image at different scales. These features not only enrich the dimensions of image representation, but also promote a deeper understanding of the content of earthquake fault images. The residual connection mechanism maps the input features directly to the output. This connection design effectively alleviates the common gradient vanishing and gradient exploding problems in the deep network training process, and significantly improves the training stability and performance of the network.

[0011] In the neural network architecture, the skip connection directly fuses the features of the lower level with the features of the higher level, thereby enhancing the network's information transfer efficiency and feature representation capabilities. However, given the significant differences between low-level and high-level features, directly fusing these two types of feature maps at the same level will lead to the loss of key spatial information, especially in fault identification tasks that require extremely high spatial details. Every subtle spatial feature carries information that cannot be ignored. Introducing the residual mechanism in the skip connection, this strategy can not only effectively improve the resolution of the low-level feature map so that it still maintains sufficient detail information in the deep network, but also alleviate the gradient vanishing problem in the deep network training through the introduction of the residual path, prevent network overfitting, and accelerate network convergence. The residual mechanism is introduced into the skip connection process and improved. The formula is: (4) Among them, x represents the input feature, F (x) represents a 3×3×3 convolution and ReLu activation function, f (x) represents a 1×1×1 convolution, H (x) is the result obtained after residual error; The attention mechanism simulates the degree of human attention to information, allowing the model to prioritize and deeply process those pieces of information that are more important to the task. This mechanism dynamically adjusts the weight distribution of the input data, giving higher weights to the key parts of the input data, while giving lower weights to the relatively unimportant parts, so that the model can prioritize and deeply process those information that are critical to the current task under limited resources. The improved attention module is used in the last layer of the encoder; The attention module is composed of a spatial attention module and a channel attention module in parallel. The input feature map is learned through the spatial attention mechanism and the channel attention mechanism respectively, and the results are multiplied to obtain the output feature map. This parallel processing mechanism not only improves the efficiency of information processing, but also enhances the model's ability to capture complex data features. The calculation formula is: (5) M s ( F )and M c ( F ) are spatial attention module and channel attention module respectively, is the input feature map, where C is the number of channels, H, W, D They are the three dimensions of the feature map; In the spatial attention module, the pooling results are added and multiplied with the input features after passing through the Sigmoid activation function. The calculation formula is: (6) In the formula, AvgPool(F) is the global average pooling, MaxPool(F) is the global maximum pooling, and "+" represents the addition operation. The two pooling results are added through the Concatenate function, and the output is limited to 0-1 through the σ (·) Sigmoid activation function; The global average pooling of the channel attention module passes through two full connections and Sigmoid activation functions and is multiplied by the input features. The calculation formula is: (7) Both global average pooling and global maximum pooling compress the information in the feature map into a real number, so that important information can be captured from a global perspective, but the emphasis of their compression methods is different. Average pooling takes the average value in a region, and maximum pooling takes the maximum value. Combining the two can obtain a richer global feature representation. The Sigmoid activation function can control the output between 0 and 1, making it an ideal activation function in classification problems, especially for binary classification problems, which can be directly interpreted as probability output.

[0012] Step 3: Establish the ResSC-UNet network model The unique "U"-shaped architecture of the U-Net convolutional neural network can play a good role in classification tasks. The U-Net network cleverly combines the functions of the encoder on the left and the decoder on the right. The encoder on the left, through a series of convolution, pooling and downsampling operations, deeply analyzes the data, gradually mines and refines deep features; the decoder on the right performs reverse operations, using convolution and upsampling operations to gradually restore the resolution of the feature map, and through a clever jump connection mechanism, integrates the low-level position information with the deep semantic information, and finally outputs the fault probability end-to-end.

[0013] Convolution is used to extract features from input images. The basic idea of ​​convolution is to perform calculations on local areas of input data by sliding a small window. The specific operation is to multiply the weight of the convolution kernel by the value of the corresponding position of the input data, and then add all these products to get the output of the convolution. The common operation of downsampling is pooling, which is to reduce the dimension of the data, thereby reducing the computational complexity and number of parameters of the model and improving the computational efficiency of the model. The pooling operations mainly include maximum pooling and average pooling. The basic idea of ​​maximum pooling is to take the maximum value of the local area of ​​the input data, operate on the input data by sliding the window, and take the maximum value in each window as the output. Through maximum pooling, the spatial dimension of the feature map can be gradually reduced at different scales to form a hierarchical feature representation, so that the network can more effectively capture the structure of the input data. Common upsampling methods include linear interpolation, bilinear interpolation and cubic interpolation. Upsampling can enable the model to better learn and reconstruct the details of the data. The basic principle of the UpSampling3D method is to increase the sampling points in the three-dimensional space through the interpolation method to increase the details and resolution of the data, thereby improving the performance and expressiveness of the model.

[0014] The ResSC-UNet model is a deep optimization and innovation of the classic three-dimensional U-Net network structure. While retaining the original advantages of U-Net, it incorporates residual learning mechanism and advanced attention module. The encoder consists of 3×3×3 convolution blocks and downsampling. The decoder includes the same 3×3×3 convolution blocks and upsampling and the Sigmoid activation function of the last layer. The residual jump connection encoder and decoder have the same layer features. Each layer of the ResSC-UNet network structure is a multi-scale residual module with 16, 32, 64, and 128 channels. Downsampling uses maximum pooling, and a residual mechanism is added in the jump connection process. Upsampling uses the UpSampling3D method in the Tensorflow framework to restore the resolution, and an attention module is added to the last layer of the decoder. The original U-Net model is widely used for its unique encoder-decoder architecture and the jump connection between the two. This recognition method not only ensures the effective transmission of information, but also promotes the network's ability to capture detailed features.

[0015] Furthermore, to ensure the richness of the training data, the training set and the corresponding label data were rotated 90°, 180°, and 270° through a simple data augmentation method, and expanded to four times the original size as the final training set. The final data set met the model training requirements in terms of diversity, authenticity, and sample quantity.

[0016] Beneficial effects of the present invention: A three-dimensional seismic fault identification method based on ResSC-UNet is provided. The three-dimensional seismic fault identification method is improved based on the U-Net network model, and finally an efficient model is trained, which can efficiently and accurately identify fault information from seismic data. The model can receive the seismic data to be processed as input, and directly output the predicted fault data. The final model is applied to the test data and actual data, and higher evaluation indicators and identification results are achieved. Compared with traditional fault identification methods and other deep learning methods, the application of this method greatly simplifies the technical operation process. Its advantages are as follows: (1) This network uses an end-to-end model that is completely data-driven, avoiding data preprocessing and parameter setting.

[0017] (2) Technical personnel do not need a deep theoretical background or long-term practical experience. They only need to use pre-trained models and directly input raw seismic data to quickly obtain fault identification results, which reduces the reliance on complex theoretical knowledge, improves the convenience of overall operation, and greatly reduces processing time. In addition, the pre-trained model does not limit the size of input data during use, and is highly operable.

[0018] (3) The network adopts the residual mechanism and attention module, and achieves the ability to accurately focus on key areas on the basis of introducing only a small number of additional parameters, and ensures that the model can still maintain high performance when the number of layers increases significantly, effectively avoiding the degradation problem faced by traditional deep networks. Therefore, this 3D seismic fault identification method can meet the requirements of fault identification in oil and gas exploration and development. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the 3D seismic fault identification method based on ResSC-UNet.

[0020] Figure 2 It is a flow chart for the production of 3D seismic data sets.

[0021] Figure 3 It is the flowchart of the multi-scale residual module.

[0022] Figure 4 It is the flowchart of the residual skip connection module.

[0023] Figure 5 It is the flowchart of the spatial channel attention module.

[0024] Figure 6 This is the ResSC-UNet network structure diagram.

[0025] Figure 7 It is a comparison chart of the 3D recognition results of the original test data image, test data slice labels and ResSC-UNet recognition results.

[0026] Figure 8 It is a slice comparison diagram of the original seismic image, fault label image and fault identification image in the Time direction of the fault identification result.

[0027] Fig. 9 It is a slice comparison diagram of the original seismic image, fault label image and fault identification image in the XLine direction of the fault identification result.

[0028] Fig.10 It is a slice comparison diagram of the original seismic image, fault label image and fault identification image in the Inline direction of the fault identification result.

[0029] Fig.11 It is the accuracy curve of the recognition results.

[0030] Fig.12 It is the error curve of the recognition result.

[0031] Fig.13It is a slice comparison diagram of the original fault image of the actual earthquake fault identification result in the Time direction and the ResSC-UNet identification result image.

[0032] Fig.14 It is a slice comparison diagram of the original fault image of the actual earthquake fault identification result in the XLine direction and the ResSC-UNet identification result image.

[0033] Fig.15 It is a slice comparison diagram of the original fault image of the actual earthquake fault identification result in the Inline direction and the ResSC-UNet identification result image.

[0034] Fig.16 It is a three-dimensional stereogram of the original fault image of the actual earthquake fault identification result and the ResSC-UNet identification result image. DETAILED DESCRIPTION Example

[0035] A three-dimensional seismic fault identification method based on ResSC-UNet, the three-dimensional seismic fault identification method comprises the following steps: The first step is to create a three-dimensional seismic fault data set. In the training process of the neural network model, the data set is particularly important. The core of deep learning is to deeply analyze the internal logic and relationship between data, and the final performance of the model depends directly on the quality of the features learned from the training data. Therefore, the quality of the data set is the key to the success of neural network training. Creating a diverse and accurate training data set can guide the network to capture more accurate and comprehensive data features and show better recognition effects.

[0036] The production of 3D seismic fault dataset includes the following steps: 1) Generate a one-dimensional horizontal reflectivity model r ( x, y, z ), which consists of a series of random values ​​in the range [–1, 1]; 2) Use the vertical shear model to act on the reflectivity model to create some wrinkle structures and vertical displacement maps S The formula for 1 is: (1) S 1 by N It consists of a two-dimensional Gaussian function and a linear scalar function, where the parameters a0, b k , c k , d k , σ kThrough random combination within a certain range of values, it is used to produce folding structures with some specific spatial variations; x , y , z They are respectively x axis, y axis, z Axis, z max for z Maximum value of the direction, 1.5z / z max As a linear scalar, z Scaling can suppress folding from bottom to top; k Denote the number of Gaussian functions by, b k It is k The weight coefficient of the item, ( x−c k ) 2 + ( y−d k ) 2 is a two-dimensional Euclidean distance square, indicating the coordinate point ( x , y ) and the center point ( c k , d k ), σ k For the k The scale parameter of the item controls the width of the Gaussian distribution; using the displacement map S 1. Use sinc Interpolation is used to vertically displace the original reflectivity model r to obtain the wrinkle model r (x, y ,z+ S 1(x, y ,z)); 3) Based on step 2, use the linear displacement map S 2 Generate inclined structure, the formula is defined as: (2) e 0 is a constant term, representing the base bias value, f 0 and g 0 is a random number within the limit range, controlling x Coordinates and y The slope of the coordinates will eventually generate an inclined surface of a two-dimensional plane. S 2 loops are applied to the wrinkle model r to obtain a new reflectivity model r (x, y ,z+ S 1(x,y ,z)+ S 2(x, y ,z)); 4) Add plane faults to the fold model, control the fault dip, strike, fault range and select random reference points ( x 0. y 0. z 0), so that the generated faults are diverse; 5) Use wavelets to convolve the reflection model and model sharp discontinuities near faults; 6) Add Gaussian noise with different signal-to-noise ratios; To ensure the richness of the training data, the training set and the corresponding label data were rotated 90°, 180°, and 270° through a simple data augmentation method, and expanded to four times the original size as the final training set. The final data set met the model training requirements in terms of diversity, authenticity, and sample quantity.

[0037] The second step is to build a deep learning network model, integrate the multi-scale residual module into the U-Net framework, and further enhance the retention and transmission of spatial feature information through the residual connection strategy of the 1×1×1 convolutional layer; When building a deep network, a deeper network architecture is used to capture richer information, but simply stacking network layers may encounter the challenge of network degradation, that is, as the number of layers increases, the performance decreases instead of increases. The deep residual network introduces the residual module, which cleverly transforms some convolutional layers in the network. The core of the residual module is that it allows the network to automatically identify and remove redundant layers during training, and avoids the occurrence of network degradation by achieving identity mapping. This not only successfully overcomes the bottleneck of deep network degradation, but also greatly improves the learning efficiency and stability of the network. The residual function formula is: (3) Among them, x represents the input feature, F (x) contains multiple convolutional layers and activation functions, H (x) is the result obtained after residual error; Seismic fault images often present complex shapes with intricate folds and dense folds. The boundaries of the rock layers are blurred and difficult to define, which can easily lead to confusion with non-seismic related geological structures. To solve this problem, a multi-scale residual module is incorporated into the U-Net framework to replace the traditional encoder-decoder hierarchical unit. The purpose is to achieve in-depth analysis of multiple layers of resolution, streamline model parameters, and further enhance the retention and transmission of spatial feature information through the residual connection strategy of the 1×1×1 convolutional layer.

[0038] The core of the multi-scale residual module is that it can use convolution kernels of different sizes (such as 3×3×3, 5×5×5, etc.) to perform multi-scale convolution operations on the input feature map, thereby capturing and extracting the feature information of the image at different scales. These features not only enrich the dimensions of image representation, but also promote a deeper understanding of the content of earthquake fault images. The residual connection mechanism maps the input features directly to the output. This connection design effectively alleviates the common gradient vanishing and gradient exploding problems in the deep network training process, and significantly improves the training stability and performance of the network.

[0039] In the neural network architecture, the skip connection directly fuses the features of the lower level with the features of the higher level, thereby enhancing the network's information transfer efficiency and feature representation capabilities. However, given the significant differences between low-level and high-level features, directly fusing these two types of feature maps at the same level will lead to the loss of key spatial information, especially in fault identification tasks that require extremely high spatial details. Every subtle spatial feature carries information that cannot be ignored. Introducing the residual mechanism in the skip connection, this strategy can not only effectively improve the resolution of the low-level feature map so that it still maintains sufficient detail information in the deep network, but also alleviate the gradient vanishing problem in the deep network training through the introduction of the residual path, prevent network overfitting, and accelerate network convergence. The residual mechanism is introduced into the skip connection process and improved. The formula is: (4) Among them, x represents the input feature, F (x) represents a 3×3×3 convolution and ReLu activation function, f (x) represents a 1×1×1 convolution, H (x) is the result obtained after residual error; The attention mechanism simulates the degree of human attention to information, allowing the model to prioritize and deeply process those pieces of information that are more important to the task. This mechanism dynamically adjusts the weight distribution of the input data, giving higher weights to the key parts of the input data, while giving lower weights to the relatively unimportant parts, so that the model can prioritize and deeply process those information that are critical to the current task under limited resources. The improved attention module is used in the last layer of the encoder; The attention module is composed of a spatial attention module and a channel attention module in parallel. The input feature map is learned through the spatial attention mechanism and the channel attention mechanism respectively, and the results are multiplied to obtain the output feature map. This parallel processing mechanism not only improves the efficiency of information processing, but also enhances the model's ability to capture complex data features. The calculation formula is: (5) Ms ( F )and M c ( F ) are spatial attention module and channel attention module respectively, is the input feature map, where C is the number of channels, H, W, D They are the three dimensions of the feature map; In the spatial attention module, the pooling results are added and multiplied with the input features after passing through the Sigmoid activation function. The calculation formula is: (6) In the formula, AvgPool(F) is the global average pooling, MaxPool(F) is the global maximum pooling, and "+" represents the addition operation. The two pooling results are added through the Concatenate function, and the output is limited to 0-1 through the σ (·) Sigmoid activation function; The global average pooling of the channel attention module passes through two full connections and Sigmoid activation functions and is multiplied by the input features. The calculation formula is: (7) Both global average pooling and global maximum pooling compress the information in the feature map into a real number, so that important information can be captured from a global perspective, but the emphasis of their compression methods is different. Average pooling takes the average value in a region, and maximum pooling takes the maximum value. Combining the two can obtain a richer global feature representation. The Sigmoid activation function can control the output between 0 and 1, making it an ideal activation function in classification problems, especially for binary classification problems, which can be directly interpreted as probability output.

[0040] Step 3: Establish the ResSC-UNet network model The unique "U"-shaped architecture of the U-Net convolutional neural network can play a good role in classification tasks. The U-Net network cleverly combines the functions of the encoder on the left and the decoder on the right. The encoder on the left, through a series of convolution, pooling and downsampling operations, deeply analyzes the data, gradually mines and refines deep features; the decoder on the right performs reverse operations, using convolution and upsampling operations to gradually restore the resolution of the feature map, and through a clever jump connection mechanism, integrates the low-level position information with the deep semantic information, and finally outputs the fault probability end-to-end.

[0041] Convolution is used to extract features from input images. The basic idea of ​​convolution is to perform calculations on local areas of input data by sliding a small window. The specific operation is to multiply the weight of the convolution kernel by the value of the corresponding position of the input data, and then add all these products to get the output of the convolution. The common operation of downsampling is pooling, which is to reduce the dimension of the data, thereby reducing the computational complexity and number of parameters of the model and improving the computational efficiency of the model. The pooling operations mainly include maximum pooling and average pooling. The basic idea of ​​maximum pooling is to take the maximum value of the local area of ​​the input data, operate on the input data by sliding the window, and take the maximum value in each window as the output. Through maximum pooling, the spatial dimension of the feature map can be gradually reduced at different scales to form a hierarchical feature representation, so that the network can more effectively capture the structure of the input data. Common upsampling methods include linear interpolation, bilinear interpolation and cubic interpolation. Upsampling can enable the model to better learn and reconstruct the details of the data. The basic principle of the UpSampling3D method is to increase the sampling points in the three-dimensional space through the interpolation method to increase the details and resolution of the data, thereby improving the performance and expressiveness of the model.

[0042] The ResSC-UNet model is a deep optimization and innovation of the classic three-dimensional U-Net network structure. While retaining the original advantages of U-Net, it incorporates residual learning mechanism and advanced attention module. The encoder consists of 3×3×3 convolution blocks and downsampling. The decoder includes the same 3×3×3 convolution blocks and upsampling and the Sigmoid activation function of the last layer. The residual jump connection encoder and decoder have the same layer features. Each layer of the ResSC-UNet network structure is a multi-scale residual module with 16, 32, 64, and 128 channels. Downsampling uses maximum pooling, and a residual mechanism is added in the jump connection process. Upsampling uses the UpSampling3D method in the Tensorflow framework to restore the resolution, and an attention module is added to the last layer of the decoder. The original U-Net model is widely used for its unique encoder-decoder architecture and the jump connection between the two. This recognition method not only ensures the effective transmission of information, but also promotes the network's ability to capture detailed features.

[0043] Example 2 (Test 1) In order to verify the effect of the core modules on improving the model performance, a comparative experiment was conducted. On the basis of retaining UNet as the benchmark model, different modules were gradually and systematically introduced. All experiments were conducted on the same dataset, the same partitioning method, the same number of iterations, and the same experimental environment, and Accuracy, Recall, F1, and IOU were used as evaluation indicators.

[0044] The order of model performance testing is: ordinary four-layer Unet network, Unet network with only attention mechanism, Unet network with only skip residual structure, Unet network with only multi-scale residual module, and finally the Unet network with attention mechanism, skip residual and multi-scale residual module. From the test results, it can be seen that after adding each module, the evaluation indicators have been improved to a certain extent. After adding all modules, all indicators have reached the optimal results.

[0045] Example 3 (Test 2) Synthetic earthquake fault data is used to verify the fault recognition effectiveness of the network model that integrates the residual structure and dual attention mechanism. The samples that are not involved in the training process in the synthetic 3D dataset are input into the ResSC-UNet network model to observe the recognition effect of the network. Figure 7 The following are the three-dimensional image recognition results. From left to right, they are the original test data image, the test data slice label image, and the ResSC-UNet recognition result image. Figure 8-10 The slice display comparison of the fault recognition results in the Time, XLine and Inline directions are shown in Figure 1. From left to right, they are the original seismic image, fault label and fault recognition image. From the recognition results, it can be seen that both the location of the fault and the shape and distribution of the fault are almost completely consistent with the corresponding label, which also verifies that the fault recognition accuracy of the neural network model based on the ResSC-UNet 3D seismic fault recognition method is extremely high.

[0046] Fig.11 and Fig.12 They are the accuracy curve and error curve of the recognition results. The curves of the two key performance indicators of the model, loss rate and accuracy, are optimal. The accuracy curve reaches a significant rising plateau at about the 30th iteration, showing a trend of gradual and steady convergence after a sharp rise in the early stage. The trend of the loss rate curve is in sharp contrast to the accuracy curve, showing a change trajectory in the opposite direction of the accuracy improvement.

[0047] Example 4 (Test 3) Actual seismic data is used to verify the processing effect of the network on actual data. The subset size is 511 main survey lines (Line100-611), 374 cross survey lines (Crossline300-683), the sampling interval is 4ms, and each channel has 128 sampling points (Time 1336-1844ms). The actual seismic data is input into the ResSC-UNet network model to observe the recognition effect of the network. Figure 13-15The slice displays of fault identification results of actual seismic data in the Time, XLine and Inline directions are shown respectively. The left figure is the original fault image, and the right figure is the ResSC-UNet identification result. Fig.16 The three-dimensional stereogram recognition result. From the recognition results, it can be seen that the network shows good continuity in this data set, can accurately capture the characteristics of the fault, accurately locate and describe the position and shape of the fault in the region, and the identified fault direction is basically consistent with the label. Overall, the network model not only achieves high-precision depiction of the fault boundary, but also greatly reduces the phenomenon of missed and misidentified faults, proving that the ResSC-UNet model has high fault recognition accuracy and robustness in actual seismic data, and shows excellent performance in actual work area seismic data.

[0048] The three-dimensional seismic fault recognition method is an improvement based on the U-Net network model. Compared with traditional fault recognition methods and other deep learning methods, the application of this method greatly simplifies the technical operation process. Traditional fault recognition methods have a long cycle and are easily affected by subjective factors. The use of deep learning methods can automatically identify faults, greatly shorten the recognition time and ensure high accuracy. In the encoder and decoder modules, the introduction of multi-scale residual modules can fuse feature information of different scales at different levels. Compared with the U-Net model framework, the network can capture the details and global information in the image more comprehensively, thereby improving the accuracy and robustness of recognition. In the decoder and connection, the spatial channel attention and residual modules are added, which can solve the problem of semantic information loss during the jump connection process and can focus on the more important parts of the image. Compared with the U-Net model framework, the network has a higher fault recognition accuracy.

Claims

1. A three-dimensional seismic fault identification method based on ResSC-UNet, characterized in that: The three-dimensional seismic fault identification method comprises the following steps: Step 1: Preparation of 3D seismic fault dataset The production of 3D seismic fault dataset includes the following steps: 1) Generate a one-dimensional horizontal reflectivity model r ( x, y, z ), which consists of a series of random values ​​in the range [–1, 1]; 2) Use the vertical shear model to act on the reflectivity model to create a wrinkle structure and vertical displacement map S The formula for 1 is: （1） S 1 by N It consists of a two-dimensional Gaussian function and a linear scalar function, where the parameters a0, b k , c k , d k , σ k Through random combination within a certain value range, it is used to generate folding structures with specific spatial variations; x , y , z They are respectively x axis, y axis, z Axis, z max for z Maximum value of the direction, 1.5z / z max As a linear scalar, z Scaling can suppress folding from bottom to top; k Denote the number of Gaussian functions by, b k It is k The weight coefficient of the item, ( x−c k ) 2 + ( y−d k ) 2 is a two-dimensional Euclidean distance square, indicating the coordinate point ( x , y ) and the center point ( c k , d k ), σ k For the k The scale parameter of the item controls the width of the Gaussian distribution; using the displacement map S 1. Use sinc Interpolation is used to vertically displace the original reflectivity model r to obtain the wrinkle model r (x, y ,z+ S 1(x, y ,z)); 3) Based on step 2, use the linear displacement map S 2 Generate inclined structure, the formula is defined as: （2） e 0 is a constant term, representing the base bias value, f 0 and g 0 is a random number within the limit range, controlling x Coordinates and y The slope of the coordinates eventually generates an inclined surface of a two-dimensional plane; the linear displacement S 2 loops are applied to the wrinkle model r to obtain a new reflectivity model r (x, y ,z+ S 1(x, y ,z)+ S 2(x, y ,z)); 4) Add plane faults to the fold model, control the fault dip, strike, fault range and select random reference points ( x 0. y 0. z 0), so that the generated faults are diverse; 5) Use wavelets to convolve the reflection model and model sharp discontinuities near faults; 6) Add Gaussian noise with different signal-to-noise ratios; Step 2: Build a deep learning network model The multi-scale residual module is integrated into the U-Net framework, and the residual connection strategy of the 1×1×1 convolutional layer is used to further enhance the retention and transmission of spatial feature information; the residual function formula is: （3） Among them, x represents the input feature, F (x) contains multiple convolutional layers and activation functions, H (x) is the result obtained after residual error; The residual mechanism is introduced in the skip connection. By introducing the residual path, the gradient vanishing problem in deep network training is alleviated, network overfitting is prevented, and network convergence is accelerated. The residual mechanism is introduced into the skip connection process and improved. The formula is: （4） Among them, x represents the input feature, F (x) represents a 3×3×3 convolution and ReLu activation function, f (x) represents a 1×1×1 convolution, H (x) is the result obtained after residual error; The attention mechanism simulates the degree of human attention to information, enabling the model to prioritize and deeply process information fragments that are more important to the task. The improved attention module is applied to the last layer of the encoder. The attention module is composed of a spatial attention module and a channel attention module in parallel. The input feature map is learned through the spatial attention mechanism and the channel attention mechanism respectively, and the output feature map is obtained by multiplying the results. The calculation formula is: （5） M s ( F )and M c ( F ) are spatial attention module and channel attention module respectively, is the input feature map, where C is the number of channels, H, W, D They are the three dimensions of the feature map; In the spatial attention module, the pooling results are added and multiplied by the input features after passing through the Sigmoid activation function. The calculation formula is: （6） In the formula, AvgPool(F) is the global average pooling, MaxPool(F) is the global maximum pooling, "+" represents the addition operation, the two pooling results are added through the Concatenate function, and the output is limited to 0-1 through the σ (·) Sigmoid activation function; The global average pooling of the channel attention module passes through two full connections and Sigmoid activation functions and is multiplied by the input features. The calculation formula is: （7） Step 3: Establish the ResSC-UNet network model The encoder part on the left side of U-Net deeply analyzes the data through a series of convolution, pooling and downsampling operations, gradually mining and refining deep features; the decoder part on the right side performs the reverse operation, using convolution and upsampling operations to gradually restore the resolution of the feature map, and through the skip connection mechanism, it integrates the low-level position information with the deep semantic information, and finally outputs the fault probability end-to-end; Through maximum pooling, the spatial dimensions of feature maps can be gradually reduced at different scales to form a hierarchical feature representation, allowing the network to more effectively capture the structure of input data; The ResSC-UNet model is a deep optimization of the three-dimensional U-Net network structure. While retaining the original advantages of U-Net, it incorporates the residual learning mechanism and attention module. The encoder consists of 3×3×3 convolution blocks and downsampling. The decoder includes the same 3×3×3 convolution blocks and upsampling as well as the Sigmoid activation function of the last layer. The residual jump connection encoder and decoder have the same layer features. Each layer of the ResSC-UNet network structure is a multi-scale residual module with 16, 32, 64, and 128 channels, respectively. Downsampling uses maximum pooling, and a residual mechanism is added in the jump connection process. Upsampling uses the UpSampling3D method in the Tensorflow framework to restore the resolution, and an attention module is added to the last layer of the decoder.

2. The three-dimensional seismic fault identification method based on ResSC-UNet according to claim 1, characterized in that: To ensure the richness of the training data, the training set and the corresponding label data are rotated 90°, 180°, and 270° through data augmentation methods, and expanded to four times the original size as the final training set.

Images