A 3D fault identification method integrating seismic data gradient attributes and semantic enhancement

Abstract

The present invention provides a three-dimensional fault recognition method integrating seismic data gradient attributes and semantic enhancement, and relates to the field of seismic exploration. The method includes: obtaining seismic data, preprocessing the seismic data, establishing a three-dimensional fault recognition model integrating seismic gradient attributes and semantic enhancement technology, training the model using seismic data, inputting seismic data into the trained model, and obtaining fault recognition results. The beneficial effects of the present invention are: proposing a three-dimensional fault recognition model integrating seismic gradient attributes and semantic enhancement technology, obtaining fault recognition results, and improving continuity, resolution and accuracy.

Description

Technical Field

[0001] The present invention relates to the field of seismic exploration, and in particular to a three-dimensional fault recognition method integrating seismic data gradient attributes with semantic enhancement. Background Art

[0002] In the field of geophysical exploration, accurate and efficient fault identification is essential for precise mapping of underground structures. This technology not only provides key support for exploration and development decisions, but also plays an important role in earthquake risk assessment, urban planning, and disaster prevention and mitigation. Before the emergence of automatic fault identification methods, faults were usually manually marked based on expert reasoning and prior geological knowledge related to a certain area. This method is not only limited by the subjective bias of interpreters and prone to misjudgment, but also with the continuous improvement of seismic data acquisition and processing technology, the amount of seismic data has increased exponentially, making it difficult for this method to meet the processing efficiency requirements, resulting in an extension of the exploration cycle and an increase in the risk of missing or misjudgment of key geological information, affecting the accuracy of decision-making.

[0003] In recent years, deep learning has made great achievements in a wide range of scientific and engineering fields, such as image processing, language translation, and intelligent driving, due to its outstanding capabilities in information processing, data analysis, and pattern recognition. The rapid development of deep learning has catalyzed a paradigm shift in the research of geophysicists. At present, algorithms based on deep learning are increasingly being used to solve challenging problems in seismic data processing, interpretation, and inversion. In the task of fault identification, given the structural similarities between seismic data and image data, supervised deep learning-based image semantic segmentation models have been widely used and developed rapidly.

[0004] Existing fault identification methods have limited recognition effect and resolution on faults as a whole, especially when dealing with small fault throw faults, which limits their application effect under complex geological conditions. When processing seismic data, they fail to fully utilize the unique properties of the data, especially the failure to effectively identify and utilize the lateral discontinuities of faults in seismic images. The application of enhanced seismic data gradient information can significantly improve the recognition ability of these discontinuities, but existing methods fail to effectively achieve this, limiting the accuracy and sensitivity of recognition. There are deficiencies in dealing with the long-distance continuity characteristics of faults. The network design that relies on traditional convolution and pooling operations has a limited local receptive field that reduces the effectiveness of the model in capturing long-distance correlations of faults. Summary of the invention

[0005] The purpose of the present invention is to solve the problem of limited recognition effect and resolution of existing fault recognition methods. The present invention provides a three-dimensional fault recognition method integrating seismic data gradient attributes and semantic enhancement. The method mainly includes the following steps:

[0006] S1. Obtain seismic data;

[0007] S2, preprocessing the seismic data;

[0008] S3, establish a three-dimensional fault recognition model integrating seismic gradient attributes and semantic enhancement technology, and use seismic data to train the model;

[0009] S4. Input the seismic data into the trained model to obtain the fault identification result.

[0010] Furthermore, the seismic data includes synthetic seismic data and real seismic data; the synthetic seismic data is used to train and verify the model, and the real seismic data is the data to be identified.

[0011] Furthermore, the preprocessing methods are different for different types of synthetic seismic data and real seismic data:

[0012] The preprocessing of synthetic seismic data is standardized processing;

[0013] The preprocessing of real seismic data includes data cleaning, cropping and standardization.

[0014] Furthermore, the 3D fault recognition model integrating seismic gradient attributes and semantic enhancement technology uses a simplified U-Net architecture as the backbone network, including a data processing module and a semantic enhancement module embedded with seismic gradient attributes;

[0015] The simplified U-Net architecture consists of a contraction path, an expansion path, and jump connections between the paths. Both the contraction path and the expansion path have three levels.

[0016] The contraction path is a three-stage encoder structure, each stage of which contains downsampling, 3D convolution, batch normalization, and ReLU activation function in sequence;

[0017] The expansion path is a three-stage decoder structure, each stage includes upsampling, 3D convolution, batch normalization and ReLU activation function in sequence;

[0018] The encoder structure and the corresponding decoder structure are linked by jump connection, and the encoding result of each level of encoder is spliced ​​with the up-sampled output result of the decoder of the corresponding level for subsequent decoding processing.

[0019] Furthermore, the specific working steps of the data processing module embedded with seismic gradient attributes are as follows:

[0020] S31, calculating the gradient data of the seismic data in two directions;

[0021] The calculation of gradient data is carried out from the horizontal line and the main survey line of the seismic data, based on the central difference, and the calculation formula is as follows:

[0022] Δ cl f[i,j,k]=(f[i,j+1,k]-f[i,j-1,k]) / 2

[0023] Δ il f[i,j,k]=(f[i,j,k+1]-f[i,j,k-1] / 2

[0024] Among them, Δ cl , Δ il They represent the differential operations in the main survey line and the horizontal line directions of the seismic data respectively; f[i, j, k] represents an element in the seismic data, and [i, j, , k] is used to specify the position of the element in the seismic data;

[0025] S32, concatenate the seismic data and the gradient data sets in two directions along the channel dimension to form the final input data of the semantic enhancement module.

[0026] Furthermore, the semantic enhancement module includes strip pooling and overall nesting modules;

[0027] Strip pooling is set at the third stage of the three-stage encoder structure, which is used for maximum pooling, which is different from the square pooling used in the first and second stages;

[0028] The overall nested module includes side output, channel splicing, three-dimensional convolution and Softmax activation function; there are four side outputs in total, which are located at the end of the third-level encoder in the three-level encoder structure and at the end of each level of the decoder structure. All side outputs are spliced ​​in the channel dimension and processed by three-dimensional convolution and Softmax activation function. The final output is mapped to the probability value of the fault, so as to achieve accurate identification of the fault.

[0029] Furthermore, the convolution kernel selected for square pooling is 2×2×2, and the convolution kernel selected for strip pooling is 4×2×2. The process of strip pooling is expressed as follows:

[0030] Y(i,j,k)=max(X(i×4: i×4+3, j×2: j×2+1, k×2: k×2+1))

[0031] Among them, Y represents the strip pooling output; X is the convolution input; max is the maximum value function.

[0032] Furthermore, the expression of the overall nested module is:

[0033]

[0034] Among them, S represents the final splicing result of the side output; represents the splicing along the channel dimension; Di represents the i-th side output, there are m side outputs in total; C(D i ) represents the convolution operation on the input of the side output; U(C(D i )) indicates that C(D i ) performs upsampling operation; α i Denote as the weight of the upsampling result.

[0035] A storage medium stores instructions and data for realizing a three-dimensional fault recognition method integrating seismic data gradient attributes and semantic enhancement.

[0036] A computer device comprises: a processor and the storage medium; the processor loads and executes instructions and data in the storage medium to implement a three-dimensional fault identification method integrating seismic data gradient attributes and semantic enhancement.

[0037] The beneficial effects of the technical solution provided by the present invention are as follows: the present invention proposes a three-dimensional fault recognition model that integrates seismic gradient attributes and semantic enhancement technology, thereby improving the continuity, resolution and accuracy of fault recognition results; proposes a data preprocessing module that embeds seismic gradient attributes, which is used to extract and fuse the gradient information of seismic data, and embed the seismic gradient attributes into the input data to cope with the characteristics of discontinuous lateral reflectivity of faults in seismic images; through a semantic enhancement module specially designed for fault recognition, the pooling kernel in the final downsampling stage is adjusted from a square to a strip shape to effectively capture the fault features spanning a larger area, and at the same time, a global nested module is added to the decoder part to integrate global semantic information. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0039] Figure 1 It is a flow chart of a three-dimensional fault identification method integrating seismic data gradient attributes and semantic enhancement in an embodiment of the present invention;

[0040] Figure 2 is a schematic diagram of a SeisUNet model in an embodiment of the present invention;

[0041] Figure 3 is a schematic diagram of the evolution of the pooling kernel during the downsampling process in an embodiment of the present invention;

[0042] Figure 4 It is a schematic diagram comparing the synthetic seismic data recognition performance of the NRU model, the FaultSeg3D model and the SeisUNet model in an embodiment of the present invention;

[0043] Figure 5It is a schematic diagram comparing fault recognition results of the F3 dataset subset of the NRU model, the FaultSeg3D model and the SeisUNet model in an embodiment of the present invention;

[0044] Figure 6 It is a schematic diagram comparing fault recognition results of a subset of a Kerry data set using the NRU model, the FaultSeg3D model, and the SeisUNet model in an embodiment of the present invention;

[0045] Figure 7 It is a schematic diagram of the operation of the hardware device in the embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0047] An embodiment of the present invention provides a three-dimensional fault identification method that integrates seismic data gradient attributes and semantic enhancement.

[0048] Please refer to Figure 1 , Figure 1 : is a flow chart of a three-dimensional fault identification method integrating seismic data gradient attributes and semantic enhancement in an embodiment of the present invention, which specifically includes the following steps:

[0049] The first step is to obtain seismic data.

[0050] The seismic data includes synthetic seismic data and real seismic data; the synthetic seismic data is used to train and verify the model, and the real seismic data is the data to be identified.

[0051] A total of 220 synthetic 3D seismic data were collected, of which 200 were used for model training and the remaining 20 were used as validation data to ensure the accuracy and robustness of the model. The use of synthetic data has been confirmed by previous studies to effectively assist deep neural networks in capturing key features of faults and to achieve model generalization to a certain extent.

[0052] To further enhance the usefulness and predictive power of the model, the present invention has predicted several field seismic data sets with significant differences in subsurface geology, sedimentary environment and seismic characteristics. This includes the Netherlands F3 data set, which covers the Netherlands offshore central basin, an area of ​​about 384 square kilometers, about 180 kilometers from the Dutch coastline. Another data set is the New Zealand Kerry data set, which is available on the SEG Wiki website.

[0053] The second step is to preprocess the seismic data.

[0054] The preprocessing methods for different types of synthetic seismic data and real seismic data are different:

[0055] The preprocessing of synthetic seismic data is standardization, which aims to adjust the data to a standard normal distribution, that is, the data has a mean of 0 and a standard deviation of 1. This process is achieved by calculating the difference between each data point and the mean of the overall data and dividing it by the standard deviation of the data, as shown in the following formula:

[0056]

[0057] Among them, represents the data after standardization, represents the data before standardization, and represents the mean and standard deviation of the data before standardization. This step is helpful to reduce the impact of dimension, improve the comparability between different features, and promote the efficiency of model training.

[0058] The preprocessing of real seismic data includes data cleaning, cropping and standardization.

[0059] For the Dutch F3 data, the data was cleaned and trimmed, and the parts with rich faults were selected for retention to focus on analyzing the areas with the largest amount of information, and the data size was optimized to 128 (Depth) × 512 (Crossline) × 384 (Inline). Similarly, for the New Zealand Kerry data, the data size was adjusted to 480 (Depth) × 730 (Crossline) × 286 (Inline) after trimming to exclude areas with low seismic reflection in the lower space and reduce unnecessary noise interference. The trimmed F3 and Kerry data were then standardized to ensure the consistency of the scale of each data set and reduce uncertainty during model training.

[0060] The third step is to establish a three-dimensional fault recognition model that integrates seismic gradient attributes and semantic enhancement technology, and use seismic data to train the model.

[0061] like Figure 2 As shown, Figure 2 Schematic diagram of a three-dimensional fault recognition model (SeisUNet model) integrating seismic gradient attributes and semantic enhancement technology in an embodiment of the present invention; the SeisUNet model uses a simplified U-Net architecture as a backbone network, and includes a data processing module and a semantic enhancement module embedded with seismic gradient attributes;

[0062] Based on the powerful feature learning and precise segmentation capabilities of U-Net, the SeisUNet model adopts a simplified variant of this structure as its backbone network to adapt to the specific needs of fault recognition. Specifically, the contraction path reduces one level, that is, the four downsamplings are modified to three, and the number of subsequent convolution operations corresponding to it is also reduced. In addition, the SeisUNet model retains the jump connection structure in U-Net. These connections ensure that the feature maps between the encoder and the decoder can effectively transmit information and help maintain contextual information. The expansion path is designed to gradually restore the feature map to the original size of the input image in order to generate high-resolution recognition results.

[0063] The contraction path is a three-level encoder structure, each of which includes downsampling, 3D convolution, batch normalization, and ReLU activation function. The first and second level maximum pooling uses square pooling (to distinguish strip pooling), and both follow the standard downsampling process. The convolution kernel size used is 2×2×2, and the downsampling step size is set to 2; the third level maximum pooling uses strip pooling SPM. Figure 3 As shown, Figure 3 A schematic diagram showing the evolution of the pooling kernel during the downsampling process in an embodiment of the present invention is shown.

[0064] The expansion path is a three-stage decoder structure, each stage includes upsampling, 3D convolution, batch normalization and ReLU activation function. In particular, the first stage upsampling module is specially designed to match the last stage of the encoder, using a long-distance upsampling mechanism. The other decoders perform standard square upsampling steps with a step size of 2 to refine the reconstruction of the fault details.

[0065] The encoder structure and the corresponding decoder structure are linked by jump connection, and the encoding result of each level of encoder is spliced ​​with the up-sampled output result of the decoder of the corresponding level for subsequent decoding processing.

[0066] The simplified backbone network can reduce the computational complexity of the model by reducing the number of downsampling and convolution operations, and speed up model training and inference. This is particularly beneficial in resource-limited environments. In addition, reducing the number of layers and parameters may help reduce the risk of overfitting.

[0067] The data processing module DPM embedded with seismic gradient attributes places special emphasis on the importance of accurately locating fault discontinuities in space. This fine processing of seismic attributes is essential for building accurate seismic interpretation models. In three-dimensional seismic image analysis, faults often appear as discontinuities in lateral reflections. This discontinuity is closely related to the image gradient, because the gradient can characterize the direction and amplitude of the color or brightness changes of pixels in the image. Therefore, gradient analysis has become a key technology for identifying discontinuous areas or edges in images, and is also regarded as an important indicator. For three-dimensional seismic data, the gradient is usually obtained by calculating the directional derivative of the image on the horizontal line and the main line. When the image shows the existence of a fault, the derivatives of these discontinuous points will change significantly, that is, the gradient value will change significantly. By analyzing the gradient of seismic data, the discontinuous areas in the data can be effectively identified, thereby providing important auxiliary information for fault identification based on deep learning.

[0068] The specific working steps of DPM are as follows:

[0069] Step 1, calculating the gradient data of the seismic data in two directions;

[0070] A limited number of synthetic seismic data were used in the training and validation process. The dimensions of the synthetic seismic data are 1 (Channel) × 128 (Depth) × 128 (Crossline) × 128 (Inline). The gradient data is calculated from the horizontal line and the main line of the seismic data, based on the central difference. The calculation formula is as follows:

[0071] Δ cl f[i,j,k]=(f[i,j+1,k]-f[i,j-1,k]) / 2

[0072] Δ il f[i,j,k]=(f[i,j,k+1]-f[i,j,k-1]) / 2

[0073] Among them, Δ cl , Δ il They represent the differential operations in the main survey line and the horizontal line directions of the seismic data respectively; f[i, j, k] represents an element in the seismic data, and [i, j, k] is used to specify the position of the element in the seismic data;

[0074] In step 2, the seismic data are concatenated with the gradient datasets in two directions along the channel dimension to form the final input data of the semantic enhancement module.

[0075] The final input data dimensions are 3 (Channel) × 128 (Depth) × 128 (Crossline) × 128 (Inline). This process embeds discontinuous areas in the seismic data into the input data, which helps the neural network better understand the characteristics of the seismic data.

[0076] The semantic enhancement module includes strip pooling SPM and overall nested module HNM, which aims to expand the local receptive field and integrate global semantic information, and jointly improve the feature extraction and analysis capabilities of seismic data.

[0077] Downsampling a larger pooling kernel in the vertical direction helps to cover a wider area, which is crucial for capturing the narrow and long fault linear features in the vertical direction, thereby enhancing the feature extraction ability of the network model. At the same time, using a smaller pooling kernel in the horizontal direction helps to maintain certain horizontal details. Based on this, the present invention proposes an SPM, which is set at the third level of the three-level encoder structure for maximum pooling. Different from the square pooling used in the first and second levels, the convolution kernel selected by SPM is 4 (Depth) × 2 (Crossline) × 2 (Inline), and the process is expressed as follows:

[0078] Y(i,j,k)=max(X(i×4: i×4+3, j×2: j×2+1, k×2: k×2+1))

[0079] Among them, Y represents the strip pooling output; X is the convolution input; max is the maximum value function.

[0080] In deep learning, it is common to start with a smaller pooling kernel and then gradually increase it to avoid premature information loss, which can blur details and cause small changes in the network's receptive field. This conservative approach is essential for effectively extracting multi-scale features. Therefore, SPM is only applied in the third downsampling stage of the network to improve the extraction of fault features and improve recognition performance.

[0081] The overall nested module includes side output, channel splicing, three-dimensional convolution and Softmax activation function; there are four side outputs in total, which are located at the end of the third-level encoder in the three-level encoder structure and at the end of each level of the decoder structure. All side outputs are spliced ​​in the channel dimension and processed by three-dimensional convolution and Softmax activation function. The final output is mapped to the probability value of the fault, so as to achieve accurate identification of the fault.

[0082] During model training, the SeisUNet model uses a loss function that combines Dice Loss and L2 regularization. Dice Loss is a loss function used in image segmentation tasks to quantify the similarity between the segmentation output and the actual segmentation. It focuses on the intersection between positive and negative classes rather than relying solely on the overall class distribution, so it performs better in the case of class imbalance. The general form of Dice Loss is as follows:

[0083] DiceLoss=1-(2*|A∩B|) / (|A|+|B|)

[0084] Among them, A represents the segmentation area predicted by the model, B represents the actual segmentation area, |A∩B| represents the number of pixels intersecting between A and B, and |A| and |B| represent the total number of pixels in A and B, respectively.

[0085] L2 regularization is a technique for deep neural networks to avoid overfitting. It aims to encourage the weights of the model to remain relatively small. By incorporating the regularization term into the model's loss function, it controls the complexity of the model, prevents overfitting, and enhances the model's generalization ability. The general form of L2 regularization is as follows:

[0086]

[0087] Among them, λ represents the regularization strength used to control the weight of the regularization term. The value of λ is set to 1000, w i Represents the weight parameter of the model. In summary, the final form of the loss function L is as follows:

[0088] L = Dice Loss + L2

[0089] Due to limited training data, in order to prevent overfitting, an early stopping strategy was used during training, and only 30 cycles were trained. The AdamW optimizer was used during training, and the initial learning rate was set to 0.0001. The cosine annealing learning rate adjuster was used to adjust the learning rate, the cycle length was set to 10, and the lower limit of the learning rate was set to 10% of the initial learning rate. Throughout the training process, the intersection over union (IoU) and Dice coefficient (Dice) were used as quantitative evaluation indicators to comprehensively evaluate the fault identification performance on synthetic seismic data.

[0090] In order to verify the effectiveness of the SeisUNet model, a comprehensive comparative experiment was conducted to compare it with two widely recognized methods in the field of automatic fault identification: FaultSeg3D model and NRU model. Table 1 shows the experimental results on the validation set. To ensure the comparability of the results, the same loss function as the original paper was used.

[0091] Table 1 Performance evaluation statistics of different models on the synthetic seismic data validation set

[0092]

[0093] Specifically, in terms of the choice of loss function, the NRU model uses Dice Loss as the loss function, while the FaultSeg3D model uses balanced cross entropy as the loss function. When comparing the loss values, the SeisUNet model shows a significant performance improvement, with its loss value reduced by 48.61% relative to the FaultSeg3D model. Although the loss value of the NRU model is similar to that of the SeisUNet model, it is worth noting that the SeisUNet model also introduces an L2 regularization term. This shows the potential advantages of the SeisUNet model in suppressing overfitting and enhancing the generalization ability of the model. In terms of the IoU value of the validation set, the SeisUNet model improves it to 82.48, which is 7.99 and 14.62 higher than the FaultSeg3D model and the NRU model, respectively. This result not only shows the advantage of the SeisUNet model in numerical performance, but also highlights its significant improvements in model reliability and accuracy.

[0094] By observing the convergence curve of the verification set IoU during the training process, the SeisUNet model exhibits rapid convergence characteristics, showing the efficiency of its training process. Figure 4 The recognition results of the above models on synthetic seismic data are shown: (a) original labels, (b) NRU model, (c) FaultSeg3D model, and (d) SeisUNet model. Compared with the FaultSeg3D model and the NRU model, the SeisUNet model proposed in the present invention not only achieves a higher recognition accuracy, but also excels in detail preservation, especially in terms of coherence and resolution of recognition results. In summary, these experimental results fully demonstrate the efficiency and accuracy of the SeisUNet model in processing the fault identification task of synthetic seismic data.

[0095] The fourth step is to input the seismic data into the trained model to obtain the fault identification results.

[0096] The model trained on the synthetic data was then applied to fault prediction of real earthquake data. The first real earthquake dataset is a subset of the F3 dataset with a dimension of 128 (Depth) × 512 (Crossline) × 384 (Inline), as shown in Figure 5As shown in the figure, (a) the original image of the F3 dataset subset, (b) the fault identification result of the NRU model, (c) the fault identification result of the FaultSeg3D model, and (d) the fault identification result of the SeisUNet model. In this three-dimensional seismic volume, multi-directional faults are clearly visible. From the fault identification results of the NRU model on the F3 dataset subset, it can be clearly observed that the recognition performance of the NRU model is relatively poor. The NRU model failed to accurately capture the complex fault structure in the seismic data, resulting in unclear fault boundaries and weak performance in terms of resolution and continuity. It can be observed from the fault identification results of the FaultSeg3D model and the SeisUNet model on the F3 subset that the SeisUNet model shows higher resolution and continuity. This improvement in resolution ensures that the fault details are clearly visible, making the identification of tiny faults possible, while the enhancement of continuity helps to maintain the integrity of the fault structure and avoid fragmentation during the identification process, which is particularly important for the overall interpretation and subsequent analysis of the fault. Especially in the area pointed by the white arrow, the SeisUNet model successfully achieved more continuous fault identification, and in the area pointed by the black arrow, the misclassification was significantly reduced.

[0097] The second real earthquake dataset is a subset of the Kerry dataset in New Zealand, such as Figure 6 As shown: (a) original image of Kerry dataset subset, (b) fault recognition result of NRU model, (c) fault recognition result of FaultSeg3D model, (d) fault recognition result of SeisUNet model. Its dimension is 480 (Depth) × 730 (Crossline) × 286 (Inline). From the fault recognition results of NRU model on Kerry dataset subset, it can be observed that the recognition performance of NRU model is obviously insufficient, which is characterized by fragmentation and discontinuity, including the wrong merging of small displacement faults, unclear fault dip and path division, etc. The recognition results of FaultSeg3D model and SeisUNet model on Kerry subset are compared. Compared with FaultSeg3D model, SeisUNet model shows higher recognition accuracy and better fault boundary retention ability. The boundary of its recognition result is highly consistent with the real fault line, especially in the area marked by gray arrow, which shows the obvious advantage of SeisUNet model in detail retention, and significantly improves the accuracy of fault recognition of seismic data.

[0098] In summary, the SeisUNet model performs well in fault identification tasks on different field earthquake datasets, effectively verifying its accuracy and reliability in such applications.

[0099] See also Figure 7 , Figure 740 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, wherein the hardware device specifically comprises: a computer device 401, a processor 402 and a storage medium 403.

[0100] A computer device 401: The computer device 401 implements the three-dimensional fault identification method that integrates seismic data gradient attributes and semantic enhancement.

[0101] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the three-dimensional fault identification method integrating seismic data gradient attributes and semantic enhancement.

[0102] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the three-dimensional fault identification method that integrates seismic data gradient attributes and semantic enhancement.

[0103] The beneficial effects of the present invention are as follows: the present invention proposes a three-dimensional fault recognition model that integrates seismic gradient attributes and semantic enhancement technology, thereby improving the continuity, resolution and accuracy of fault recognition results; proposes a data preprocessing module that embeds seismic gradient attributes, which is used to extract and fuse the gradient information of seismic data, and embed the seismic gradient attributes into the input data to cope with the characteristics of discontinuous lateral reflectivity of faults in seismic images; through a semantic enhancement module specially designed for fault recognition, the pooling kernel of the final downsampling stage is adjusted from a square to a strip shape to effectively capture the fault features spanning a larger area, and at the same time, a global nested module is added to the decoder part to integrate global semantic information.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A three-dimensional fault identification method integrating seismic data gradient attributes and semantic enhancement, characterized in that: The specific steps include: S1. Obtain seismic data; The seismic data includes synthetic seismic data and real seismic data; the synthetic seismic data is used to train and verify the model, and the real seismic data is the data to be identified; S2, preprocessing the seismic data; The preprocessing methods are different for different types of synthetic seismic data and real seismic data: The preprocessing of synthetic seismic data is standardized processing; The preprocessing of real seismic data includes: data cleaning, clipping and standardization; S3, establish a three-dimensional fault recognition model integrating seismic gradient attributes and semantic enhancement technology, and use seismic data to train the model; The three-dimensional fault recognition model integrating seismic gradient attributes and semantic enhancement technology uses a simplified U-Net architecture as a backbone network, including a data processing module and a semantic enhancement module embedded with seismic gradient attributes; The simplified U-Net architecture includes a contraction path, an expansion path, and a jump connection between the paths, and both the contraction path and the expansion path have three levels; The contraction path is a three-stage encoder structure, each stage sequentially including downsampling, three-dimensional convolution, batch normalization and ReLU activation function; The expansion path is a three-stage decoder structure, each stage sequentially including upsampling, three-dimensional convolution, batch normalization and ReLU activation function; The jump connection links the encoder structure and the corresponding decoder structure, and splices the encoding result of each level of encoder with the up-sampled output result of the decoder of the corresponding level for subsequent decoding processing; The specific working steps of the data processing module embedded with seismic gradient attributes are as follows: S31, calculating the gradient data of the seismic data in two directions; The calculation of gradient data is carried out from the horizontal line and the main survey line of the seismic data, based on the central difference, and the calculation formula is as follows: Δ cl f[i,j,k]=(f[i,j+1,k]-f[i,j-1,k]) / 2 Δ il f[i,j,k]=(f[i,j,k+1]-f[i,j,k-1]) / 2 Among them, Δ cl , Δ il They represent the differential operations in the main survey line and the horizontal line directions of the seismic data respectively; f[i, j, k] represents an element in the seismic data, and [i, j, k] is used to specify the position of the element in the seismic data; S32, splicing the seismic data and the gradient data sets in two directions along the channel dimension to form the final input data of the semantic enhancement module; The semantic enhancement module includes strip pooling and overall nesting modules; The strip pooling is arranged at the third stage of the three-stage encoder structure, and is used for maximum pooling, which is different from the square pooling used in the first and second stages; The overall nested module includes side output, channel splicing, three-dimensional convolution and Softmax activation function; there are four side outputs in total, which are respectively located at the end of the third level encoder in the three-level encoder structure and the end of each level decoder structure. All side outputs are spliced ​​in the channel dimension, and processed by three-dimensional convolution and Softmax activation function. The final output is mapped to the probability value of the fault, so as to realize accurate identification of the fault; The convolution kernel selected for the square pooling is 2×2×2, and the convolution kernel selected for the strip pooling is 4×2×2. The process of strip pooling is expressed as follows: Y(i,j,k)=max(X(i×4:i×4+3,j×2:j×2+1,k×2:k×2+1)) Among them, Y represents the strip pooling output; X is the convolution input; max is the maximum value function; The expression of the overall nested module is: S represents the final splicing result of the side output; represents the splicing along the channel dimension; D i represents the i-th side output, there are m side outputs in total; C(D i ) represents the convolution operation on the input of the side output; U(C(D i )) indicates that C(D i ) performs upsampling operation; α i Denote as the weight of the upsampling result; S4. Input the seismic data into the trained model to obtain the fault identification result.

2. A storage medium, characterized in that: The storage medium stores instructions and data for implementing a three-dimensional fault identification method integrating seismic data gradient attributes and semantic enhancement as described in claim 1.

3. A computer device, characterized in that: include: Processor and storage medium; the processor loads and executes instructions and data in the storage medium to implement the three-dimensional fault identification method that integrates seismic data gradient attributes and semantic enhancement as described in claim 1.

Images