A digital detection method for strong earthquake faults using convolutional neural networks

Abstract

The present invention belongs to the field of seismic data processing technology and provides a convolutional neural network method for digitally detecting strong earthquake faults. The method comprises: preprocessing acquired three-dimensional seismic volume data of strong earthquake faults to obtain preprocessed data; processing the preprocessed data using an improved three-dimensional convolutional neural network and a multi-interpolation algorithm fusion technique to obtain data identification results for strong earthquake faults; and, based on the data identification results, using digital twin technology to create digital twin modeling of the seismic faults, which is then used for digital detection of the seismic faults. The present invention can effectively capture deep-level features in seismic data, improve the accuracy of fault detection, and provide strong support for early warning and prevention of earthquake disasters.

Description

Technical Field

[0001] The present invention relates to the technical field of earthquake data processing, and in particular to a convolutional neural network strong earthquake fault digital detection method. Background Art

[0002] Fault interpretation is fundamental to seismic data interpretation. Accurate fault interpretation directly impacts the accuracy of structural mapping. Currently, a wide variety of fault identification methods exist, broadly categorized as conventional seismic fault interpretation, seismic attribute-based methods, automatic fault identification, and image processing-based methods. Conventional fault interpretation methods involve manual analysis of reflection event anomalies (such as misalignment, localized changes, sudden increases or disappearances) in seismic profiles and analysis of seismic attributes (such as coherence volumes, variance, and curvature) to further determine fault strike, dip, and extension. However, with the continuous advancement of seismic exploration technology, the volume of seismic data is increasing. Using conventional methods for fault interpretation is not only tedious and time-consuming, but also difficult and non-reproducible. The complex interpretation process also requires a high level of expertise from the interpreter.

[0003] Existing tomographic image processing technologies often encounter the problem of jagged edges, which is caused by the inherent defects of the nearest neighbor interpolation method. Although this interpolation method is simple to calculate, it cannot smoothly handle the transition between pixels when the image is magnified, resulting in visually jagged edges. In addition, the problem of blurred details has also been plaguing researchers. Although bilinear interpolation improves the smoothness of the image to a certain extent, it still seems incapable of processing details, causing the fine structures in the image to become blurred.

[0004] In the field of deep learning, traditional two-dimensional convolutional networks often struggle to capture the continuity features of image space when processing fault images. However, due to the limitations of two-dimensional convolutional networks, these continuity features are often ignored or lost, affecting the quality of image processing. By combining the smooth transition capabilities of interpolation algorithms with the feature extraction capabilities of deep learning, it may be possible to achieve more accurate processing and recognition of strong earthquake fault images, providing more reliable technical support for seismological research and disaster prevention and mitigation. However, currently no method can achieve a perfect combination of the two. Existing methods often use interpolation algorithms or deep learning techniques alone, without fully exploring the potential of combining the two.

[0005] Therefore, it is necessary to provide a convolutional neural network strong earthquake fault digital detection method. Summary of the Invention

[0006] The present invention provides a convolutional neural network strong earthquake fault digital detection method, which can effectively capture the deep-level features in earthquake data, improve the accuracy of fault detection, and provide strong support for early warning and prevention of earthquake disasters.

[0007] The present invention provides a convolutional neural network strong earthquake fault digital detection method, comprising:

[0008] Preprocessing the acquired three-dimensional seismic volume data of the strong earthquake fault to obtain preprocessed data;

[0009] Using an improved three-dimensional convolutional neural network and multi-interpolation algorithm fusion technology, the pre-processed data is processed to obtain the data identification results of strong earthquake faults;

[0010] Based on the data identification results, digital twin technology is used to generate a digital twin model of the earthquake fault.

[0011] Furthermore, the acquired three-dimensional seismic volume data of the strong earthquake fault is preprocessed to obtain preprocessed data, including:

[0012] Obtain 3D seismic volume data of strong earthquake faults;

[0013] The three-dimensional seismic volume data is preprocessed to obtain preprocessed data.

[0014] Furthermore, the 3D seismic volume data is preprocessed to obtain preprocessed data, including:

[0015] The mean and standard deviation of each slice of 3D seismic volume data are calculated independently and normalized;

[0016] Multi-scale denoising based on wavelet transform is used to denoise the normalized data;

[0017] The denoised data are preliminarily enhanced by random rotation, horizontal mirroring and elastic deformation.

[0018] For the data after the initial data enhancement processing, the second data enhancement processing data is obtained by generating samples of fault dislocation, formation tilt and porosity change through the adversarial network;

[0019] The second data enhancement processing data is processed to eliminate data segments with a signal-to-noise ratio lower than a set signal-to-noise ratio threshold or with broken continuity, and to retain effective geological boundaries through three-dimensional gradient detection to obtain preprocessed data.

[0020] Furthermore, the improved three-dimensional convolutional neural network and multi-interpolation algorithm fusion technology are used to process the preprocessed data to obtain the data identification results of strong earthquake faults, including:

[0021] Based on the preprocessed data, the multi-interpolation algorithm fusion technology is used to extract the feature data set of the preprocessed data;

[0022] Build an improved 3D convolutional neural network;

[0023] According to the characteristic data set, the improved three-dimensional convolutional neural network is used for data processing to obtain the data identification results of strong earthquake faults.

[0024] Furthermore, based on the preprocessed data, a multi-interpolation algorithm fusion technology is used to extract a feature data set of the preprocessed data, including:

[0025] In an upsampling layer of the three-dimensional convolutional neural network, a bilinear interpolation method is used to perform a first sampling process on the preprocessed data to obtain a first sampling process image;

[0026] Based on the set gradient threshold of the pre-processed data, if the calculated local gradient of the pre-processed data is greater than the gradient threshold, the nearest neighbor interpolation method is used to perform a second sampling process on the pre-processed data to obtain a second sampling process image;

[0027] The first sampled processed image and the second sampled processed image are aggregated to obtain a feature data set.

[0028] Furthermore, based on the set gradient threshold of the preprocessed data, if the calculated local gradient of the preprocessed data is greater than the gradient threshold, a nearest neighbor interpolation method is used to perform a second sampling process on the preprocessed data to obtain a second sampled processed image, including:

[0029] Performing anisotropic normalization processing on the input pre-processed data to obtain anisotropic normalized processed data;

[0030] The improved three-dimensional Sobel operator is used to calculate the local gradient value of the anisotropic normalized data;

[0031] A gradient threshold of the preprocessed data is set. If the local gradient value is greater than the gradient threshold, the nearest neighbor interpolation method is used to perform a second preprocessing on the preprocessed data to obtain a second preprocessed image.

[0032] Furthermore, the improved three-dimensional Sobel operator includes:

[0033] Reduce the Z-direction convolution kernel length to the set value, configure the weights of the X-direction, Y-direction, and Z-direction convolution kernel lengths, and set the weight ratio to 1:2:1;

[0034] Set the convolution kernel lengths in the X and Y directions to be asymmetric;

[0035] Three scales are distinguished: original resolution, 1 / 2 downsampling, and 1 / 4 downsampling. Corresponding scale weights are configured. The gradient fields are calculated separately and summarized to obtain the final calculated gradient. A three-dimensional non-local mean filter is applied before the gradient calculation.

[0036] Furthermore, the improved three-dimensional convolutional neural network is a multi-scale feature pyramid fusion structure. The multi-scale feature pyramid fusion structure uses a dilation convolution layer with expansion rates of 2, 4, and 6 to extract multi-scale features. Its output is a 1×1×1 convolution layer to unify the number of channels.

[0037] The encoder and decoder parts of the improved 3D convolutional neural network include an improved multi-scale residual module. The improved multi-scale residual module includes three encoding blocks, each of which consists of a 3×3×3 convolutional layer, a batch normalization layer, and a LeakyReLU activation function.

[0038] The decoder part of the improved 3D convolutional neural network uses deconvolution layers for upsampling and channel splicing for skip connections.

[0039] Furthermore, constructing an improved three-dimensional convolutional neural network also includes: using a mixed loss function during network training; the mixed loss function includes a cross-entropy loss function, a Hausdorff distance loss function, and a dynamic regularization method based on interpolation; among them, the cross-entropy loss function is used for fault classification, the Hausdorff distance loss function is used to optimize the fault boundary accuracy, and the dynamic regularization method based on interpolation is used to balance computational efficiency and accuracy.

[0040] Furthermore, based on the data recognition results, digital twin technology is used to generate a digital twin model of the earthquake fault, including:

[0041] Based on the identification results, a digital twin model of the strong earthquake fault is constructed using digital twin technology;

[0042] The probability and spatial distribution of earthquakes of different magnitudes are used as supplementary data to improve the digital twin model of strong earthquake faults;

[0043] The improved digital twin model of strong earthquake faults is used for digital detection of strong earthquake faults.

[0044] Compared with the existing technology, the present invention has the following advantages and beneficial effects: it improves the accuracy and efficiency of digital detection of strong earthquake faults; through the improved three-dimensional convolutional neural network and multi-interpolation algorithm fusion technology, it can more effectively extract and process the characteristics of seismic volume data, thereby more accurately identifying strong earthquake faults; at the same time, the earthquake fault digital twin model constructed using digital twin technology can intuitively display the characteristics and spatial distribution of strong earthquake faults, providing strong support for earthquake early warning and disaster assessment; in addition, the method also has high flexibility and adaptability, and can be applied to earthquake fault detection tasks of different scales and complexities, providing new ideas and methods for earthquake science research.

[0045] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0046] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0048] Figure 1 This is a schematic diagram of the steps of a convolutional neural network strong earthquake fault digital detection method;

[0049] Figure 2 A schematic diagram of the steps for preprocessing the acquired 3D seismic volume data of a strong earthquake fault to obtain the preprocessed data;

[0050] Figure 3 Schematic diagram of the method steps for obtaining data identification results of strong earthquake faults by using the improved three-dimensional convolutional neural network and adopting multi-interpolation algorithm fusion technology to process the preprocessed data. DETAILED DESCRIPTION

[0051] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0052] The present invention provides a convolutional neural network strong earthquake fault digital detection method, such as Figure 1 As shown, including:

[0053] Preprocessing the acquired three-dimensional seismic volume data of the strong earthquake fault to obtain preprocessed data;

[0054] Using an improved three-dimensional convolutional neural network and multi-interpolation algorithm fusion technology, the pre-processed data is processed to obtain the data identification results of strong earthquake faults;

[0055] Based on the data identification results, digital twin technology is used to generate a digital twin model of the earthquake fault.

[0056] The working principle of the above technical solution is as follows: First, by preprocessing the 3D seismic volume data of strong earthquake faults, noise and outliers can be removed from the data, improving data quality and reliability. Next, the preprocessed data is processed using an improved 3D convolutional neural network. This network uses multi-interpolation algorithm fusion technology to more accurately capture earthquake fault characteristics and improve the accuracy and efficiency of data identification. Through this step, data identification results of strong earthquake faults can be obtained, providing a basis for the subsequent generation of digital twin models. Finally, based on the data identification results, digital twin technology is used to generate a digital twin model of the earthquake fault. This model can truly reflect the morphology and characteristics of the earthquake fault, providing strong support for earthquake prediction and disaster assessment. The digital twin model can enable more in-depth analysis and research of earthquake faults.

[0057] The beneficial effects of the above technical solution are: the solution provided in this embodiment can significantly improve the accuracy and efficiency of digital detection of strong earthquake faults; first, the preprocessing step ensures the quality and reliability of the data, providing a solid foundation for subsequent processing; second, the improved three-dimensional convolutional neural network combined with multi-interpolation algorithm fusion technology can more accurately capture the characteristics of earthquake faults, which greatly improves the accuracy and efficiency of data recognition; finally, through the application of digital twin technology, the generated earthquake fault digital twin model not only truly reflects the morphology and characteristics of the earthquake fault, but also provides an intuitive and accurate reference for earthquake prediction and disaster assessment.

[0058] In one embodiment, Figure 2 As shown, the acquired three-dimensional seismic volume data of the strong earthquake fault is preprocessed to obtain preprocessed data, including:

[0059] Obtain 3D seismic volume data of strong earthquake faults;

[0060] The three-dimensional seismic volume data is preprocessed to obtain preprocessed data.

[0061] The working principle of the above technical solution is as follows: the preprocessing step is the basis of the entire digital detection process, and its purpose is to improve the purity and consistency of the data so that subsequent processing steps can more accurately identify earthquake fault characteristics; the preprocessing process may include a series of operations such as denoising, enhancement, and standardization. These operations are aimed at eliminating interference factors in the original data, such as background noise, instrument errors, etc., while enhancing the earthquake fault signal to make it more prominent; through preprocessing, it can ensure that the data input into the three-dimensional convolutional neural network is of high quality, thereby improving the accuracy and stability of the entire digital detection process.

[0062] The beneficial effects of the above technical solution are: the solution provided in this embodiment can significantly improve the stability and consistency of data processing, reduce noise interference in the data, enhance the visibility of useful signals, and ensure the consistency of data in different dimensions. Through preprocessing, the efficiency of subsequent neural network processing can be improved.

[0063] In one embodiment, preprocessing the 3D seismic volume data to obtain preprocessed data includes:

[0064] The mean and standard deviation of each slice of 3D seismic volume data are calculated independently and normalized;

[0065] Multi-scale denoising based on wavelet transform is used to denoise the normalized data;

[0066] The denoised data are preliminarily enhanced by random rotation, horizontal mirroring and elastic deformation.

[0067] For the data after the initial data enhancement processing, the second data enhancement processing data is obtained by generating samples of fault dislocation, formation tilt and porosity change through the adversarial network;

[0068] The second data enhancement processing data is processed to eliminate data segments with a signal-to-noise ratio lower than a set signal-to-noise ratio threshold or with broken continuity, and to retain effective geological boundaries through three-dimensional gradient detection to obtain preprocessed data.

[0069] The working principle of the above technical solution is as follows: normalization processing calculates the mean and standard deviation of each slice layer, so that the data can be compared at a unified scale, reducing the error caused by differences in data range; the multi-scale denoising method based on wavelet transform effectively removes background noise from the data and improves the clarity of the signal; the initial data enhancement processing increases the diversity of the data through random rotation, horizontal mirroring and elastic deformation, which helps to improve the generalization ability of the model; further, by generating adversarial network to synthesize samples of fault dislocation, stratum tilt and porosity changes, it not only enriches the data set but also enables the model to learn more complex geological features; the final screening step eliminates data segments with low signal-to-noise ratio or continuity damage, and retains effective geological boundaries through three-dimensional gradient detection, ensuring that the data input to subsequent processing steps are both pure and contain important geological information.

[0070] The beneficial effects of the above technical solution are as follows: the solution provided in this embodiment can significantly improve the accuracy and efficiency of digital detection of strong earthquake faults; normalization processing ensures the comparability of data and reduces the source of error; the multi-scale denoising method significantly improves the clarity of the signal, laying a solid foundation for subsequent analysis; the initial data enhancement processing and the synthetic sample strategy of the generative adversarial network not only increase the diversity of the data, but also enable the model to cope with more complex geological conditions, thereby improving the generalization ability and detection accuracy of the model; the final screening step further ensures the purity of the data and the importance of geological information, providing a reliable basis for subsequent geological interpretation and disaster prediction.

[0071] In one embodiment, Figure 3 As shown in the figure, the improved three-dimensional convolutional neural network is used to process the preprocessed data using the multi-interpolation algorithm fusion technology to obtain the data identification results of strong earthquake faults, including:

[0072] Based on the preprocessed data, the multi-interpolation algorithm fusion technology is used to extract the feature data set of the preprocessed data;

[0073] Build an improved 3D convolutional neural network;

[0074] According to the characteristic data set, the improved three-dimensional convolutional neural network is used for data processing to obtain the data identification results of strong earthquake faults.

[0075] The working principle of the above technical solution is as follows: the multi-interpolation algorithm fusion technology can comprehensively consider the advantages of different interpolation algorithms, so as to more accurately extract the features of the preprocessed data and reduce information loss; the improved three-dimensional convolutional neural network can better capture complex spatial features compared to the traditional two-dimensional convolutional neural network, because the three-dimensional convolution kernel can extract features in three dimensions, which is particularly important for the analysis of geological data; the network gradually abstracts high-level features through multi-layer convolution and pooling operations, thereby realizing accurate identification of strong earthquake faults; in the data processing stage, the feature data set is input into the improved three-dimensional convolutional neural network, the network calculates the response of the feature map through forward propagation, and optimizes the network parameters through back propagation, and finally outputs the data identification results of the strong earthquake fault.

[0076] The beneficial effects of the above technical solution are as follows: the solution provided in this embodiment can significantly improve the accuracy and efficiency of strong earthquake fault identification; the application of multi-interpolation algorithm fusion technology ensures the comprehensiveness and accuracy of data feature extraction, and reduces the information loss or error that may be caused by a single interpolation algorithm; the improved three-dimensional convolutional neural network structure enhances the model's ability to process complex geological data, especially showing significant advantages in capturing spatial features, which is crucial for identifying strong earthquake faults, a geological phenomenon with spatial distribution characteristics; in addition, the solution further improves the accuracy and stability of identification by optimizing network parameters and the response of characteristic maps, providing strong technical support for the prediction and prevention of earthquake disasters.

[0077] In one embodiment, based on the preprocessed data, a multi-interpolation algorithm fusion technique is used to extract a feature data set of the preprocessed data, including:

[0078] In an upsampling layer of the three-dimensional convolutional neural network, a bilinear interpolation method is used to perform a first sampling process on the preprocessed data to obtain a first sampling process image;

[0079] Based on the set gradient threshold of the pre-processed data, if the calculated local gradient of the pre-processed data is greater than the gradient threshold, the nearest neighbor interpolation method is used to perform a second sampling process on the pre-processed data to obtain a second sampling process image;

[0080] The first sampled processed image and the second sampled processed image are aggregated to obtain a feature data set.

[0081] The working principle of the above technical solution is as follows: bilinear interpolation can smoothly transition pixel values, avoiding obvious jagged edges when the image is zoomed in or out, which is crucial for preserving the detailed information of geological data; the nearest neighbor interpolation method is highly efficient when processing image edges or sudden changes because it directly copies the nearest pixel value, reducing the amount of calculation; by combining these two interpolation algorithms, the most appropriate processing method can be selected based on the local characteristics of the data, thereby ensuring efficiency while also ensuring the accuracy of data features; in addition, the set gradient threshold serves as a judgment basis, which can intelligently switch the interpolation algorithm, so that a more refined processing method is used in areas with drastic changes in data features, while a faster processing method is used in relatively flat areas. This adaptive processing strategy further improves the accuracy and efficiency of data feature extraction; the first and second sampled processed images obtained by aggregation not only contain rich geological data features, but also process the local features of the data through different interpolation algorithms, making the final feature dataset more comprehensive and accurate, providing a solid foundation for subsequent 3D convolutional neural network processing.

[0082] The beneficial effects of the above technical solution are as follows: the solution provided by this embodiment can significantly improve the accuracy and efficiency of digital detection of strong earthquake faults; by intelligently combining bilinear interpolation and nearest neighbor interpolation, this method can optimize the processing of different local features of geological data, retaining detailed information while reducing unnecessary calculations; the setting of the gradient threshold realizes the adaptive switching of the interpolation algorithm, so that when processing complex geological data, it can respond more flexibly to changes in data features and further improve the quality of data feature extraction; in addition, the feature data set processed by this method is more comprehensive and accurate, providing a more reliable data basis for subsequent three-dimensional convolutional neural network processing, thereby improving the accuracy and efficiency of the entire digital detection of strong earthquake faults.

[0083] In one embodiment, based on a set gradient threshold of the preprocessed data, if the calculated local gradient of the preprocessed data is greater than the gradient threshold, a nearest neighbor interpolation method is used to perform a second sampling process on the preprocessed data to obtain a second sampled image, including:

[0084] Performing anisotropic normalization processing on the input pre-processed data to obtain anisotropic normalized processed data;

[0085] The improved three-dimensional Sobel operator is used to calculate the local gradient value of the anisotropic normalized data;

[0086] A gradient threshold of the preprocessed data is set. If the local gradient value is greater than the gradient threshold, the nearest neighbor interpolation method is used to perform a second preprocessing on the preprocessed data to obtain a second preprocessed image.

[0087] The working principle of the above technical solution is as follows: the present invention effectively balances the geological features in different directions in the preprocessed data through anisotropic normalization processing, making the subsequent calculation of local gradients more accurate; the improved three-dimensional Sobel operator can more accurately capture subtle changes in geological data, especially the edge information of key features such as faults; the setting of the gradient threshold, as a key link in the adaptive processing strategy, ensures that when the local features of the data change drastically, it can quickly switch to a more appropriate interpolation method, namely the nearest neighbor interpolation method. This method performs better when processing features with sharp edges, can effectively reduce the blurring effect, and retain more detailed information; through the above steps, the second sampling processed image is not only clearer, but also has more prominent geological features.

[0088] The beneficial effects of the above technical solution are as follows: the solution provided in this embodiment can significantly improve the accuracy and speed of digital detection of strong earthquake faults; anisotropic normalization processing balances the geological features in different directions, making the geological data more balanced, which is conducive to subsequent processing; the improved three-dimensional Sobel operator can accurately capture subtle changes in geological data, especially the edge information of key features such as faults, which is crucial for accurate identification of faults; the adaptive processing strategy of the gradient threshold ensures that the most appropriate interpolation method can be used under different geological features, effectively reducing the blurring effect and retaining more detailed information; finally, after these preprocessing steps, the second sampled processed image provides high-quality data input for subsequent three-dimensional convolutional neural network processing, thereby further improving the accuracy and efficiency of digital detection of strong earthquake faults.

[0089] In one embodiment, the improved three-dimensional Sobel operator includes:

[0090] Reduce the Z-direction convolution kernel length to the set value, configure the weights of the X-direction, Y-direction, and Z-direction convolution kernel lengths, and set the weight ratio to 1:2:1;

[0091] Set the convolution kernel lengths in the X and Y directions to be asymmetric;

[0092] Three scales are distinguished: original resolution, 1 / 2 downsampling, and 1 / 4 downsampling. Corresponding scale weights are configured. The gradient fields are calculated separately and summarized to obtain the final calculated gradient. A three-dimensional non-local mean filter is applied before the gradient calculation.

[0093] The working principle of the above technical solution is as follows: The improved 3D Sobel operator optimizes the feature capture ability of geological data in three dimensions by adjusting the configuration of the convolution kernel. The length of the convolution kernel in the Z direction is reduced to a set value, and the convolution kernel length weights in the X, Y, and Z directions are reasonably configured. While maintaining sensitivity to the overall characteristics of geological data, the operator pays more attention to changes in geological characteristics in the X and Y directions (i.e., in the horizontal plane). This helps to more accurately identify the distribution and direction of geological structures such as faults on the horizontal plane.

[0094] The asymmetric convolution kernel length settings in the X and Y directions further enhance the operator's ability to capture the anisotropy of geological data on the horizontal plane, making the information of geological features such as faults more prominent at the edges, which is conducive to the accurate identification of these key features in subsequent processing;

[0095] The gradient fields are calculated at three scales: original resolution, 1 / 2 downsampling, and 1 / 4 downsampling. Corresponding scale weights are assigned to each scale. This strategy takes into account the characteristic performance of geological data at different scales. By summarizing the gradient fields at different scales, more comprehensive and detailed geological characteristic information can be obtained, providing richer and more accurate data input for subsequent three-dimensional convolutional neural network processing. Applying three-dimensional non-local mean filtering before gradient calculation helps reduce noise interference in geological data and improve the accuracy and stability of gradient calculation. This step is of great significance for preserving the true characteristic information in geological data and reducing misjudgments and missed judgments.

[0096] The beneficial effects of the above technical solution are as follows: by adopting the solution provided in this embodiment, the improved three-dimensional Sobel operator significantly improves the ability to capture subtle changes and key features in geological data through a series of optimized configurations and innovative strategies, providing a high-quality data foundation for subsequent three-dimensional convolutional neural network processing, thereby further improving the accuracy and efficiency of digital detection of strong earthquake faults.

[0097] In one embodiment, the improved three-dimensional convolutional neural network is a multi-scale feature pyramid fusion structure, which uses a dilated convolution layer with dilation rates of 2, 4, and 6 to extract multi-scale features, and its output is a 1×1×1 convolution layer to unify the number of channels;

[0098] The encoder and decoder parts of the improved 3D convolutional neural network include an improved multi-scale residual module. The improved multi-scale residual module includes three encoding blocks, each of which consists of a 3×3×3 convolutional layer, a batch normalization layer, and a LeakyReLU activation function.

[0099] The decoder part of the improved 3D convolutional neural network uses deconvolution layers for upsampling and channel splicing for skip connections.

[0100] The working principle of the above technical solution is as follows: through the multi-scale feature pyramid fusion structure, the network can capture geological features at different levels, from local details to global structures, enhancing the model's understanding and analysis capabilities of complex geological phenomena; the use of the dilated convolutional layer not only expands the receptive field but also maintains the resolution of the feature map, which is crucial for fine geological structure identification; the 3×3×3 convolutional layer in the encoding block is responsible for extracting local features, the batch normalization layer helps accelerate the training process and improve the stability of the model, and the LeakyReLU activation function introduces nonlinearity, enhancing the expressiveness of the model. This combination enables each encoding block to effectively learn meaningful feature representations from the input data; the decoder part achieves upsampling through the deconvolution layer, gradually restoring the resolution of the feature map to a size close to the original input data, and the jump connection uses channel splicing to directly merge the feature map of the encoder part with the feature map of the corresponding level of the decoder. This not only retains information from different scales, but also promotes the flow of information between high-level and low-level features, which helps to improve the reconstruction accuracy and generalization ability of the model.

[0101] The beneficial effects of the above technical solution are: by adopting the solution provided in this embodiment, the improved three-dimensional convolutional neural network structure effectively improves the accuracy and efficiency of digital detection of strong earthquake faults through careful design and optimization, providing strong technical support for earthquake disaster warning and prevention.

[0102] In one embodiment, constructing an improved three-dimensional convolutional neural network further includes: using a hybrid loss function during network training; the hybrid loss function includes a cross-entropy loss function, a Hausdorff distance loss function, and an interpolation-based dynamic regularization method; wherein the cross-entropy loss function is used for fault classification, the Hausdorff distance loss function is used for optimizing fault boundary accuracy, and the interpolation-based dynamic regularization method is used for balancing computational efficiency and accuracy;

[0103] To address the extreme imbalance of positive and negative samples (fault / non-fault areas) in fault classification, this paper proposes a dynamic weighted cross entropy loss for the cross entropy loss function. The specific calculation formula is:

[0104]

[0105] In the above formula, represents the dynamic weight cross entropy loss, represents the total number of samples, Represents the true label of the i-th sample; ∈{0,1} (1 represents the fault area, 0 represents the non-fault area); represents the probability that the i-th sample belongs to the fault area predicted by the model; ∈{0,1}; Represents the current training round; represents the total number of training rounds; represents the proportion of the fault area; Represents the hyperbolic tangent function, which maps the input to the interval (-1,1) for smooth transition; if S is large, that is, when the data has a large proportion of discontinuities, then The adjustment range is reduced to avoid oscillation caused by excessive weighting; if S is small, that is, when the proportion of data fault layers is small, the offset of the tanh function increases, extending the training phase dominated by negative samples;

[0106] The working principle of the above scheme is as follows: the cross-entropy loss function guides the network to learn more accurate fault classification capabilities by measuring the error between the predicted fault category and the actual label. In the early stages of training, due to the small proportion of faults, negative samples (non-fault areas) dominate. At this time, the dynamic weighted cross-entropy loss function increases the weight of negative samples through the characteristics of the tanh function, making the model focus more on learning the characteristics of non-fault areas. With the increase of training rounds, the proportion of fault areas S gradually increases, the offset of the tanh function decreases, and the dynamic weighted cross-entropy loss function tends to balance the weighting of positive and negative samples, avoiding the problem of overweighting caused by an excessively large proportion of faults, thereby ensuring the stability and accuracy of the model in the fault classification task. In addition, the introduction of the Hausdorff distance loss function further optimizes the accuracy of fault boundaries, making the detected fault boundaries closer to the actual situation. The dynamic regularization method based on interpolation effectively improves the generalization ability of the model while ensuring computational efficiency and avoiding the occurrence of overfitting.

[0107] The beneficial effects of the above technical solution are as follows: by adopting the solution provided in this embodiment, the introduction of this hybrid loss function enables the entire network to comprehensively consider multiple aspects such as classification accuracy, boundary accuracy and computational efficiency during the training process, thereby further improving the accuracy and efficiency of digital detection of strong earthquake faults; the dynamic weight cross entropy loss breaks through the limitations of traditional fixed weights and can achieve phased optimization of the training process; by integrating geological prior knowledge into the loss function through the area ratio of the fault region, the adaptability of the model to specific data sets can be improved; based on the smoothness of the tanh function, the gradient of the loss function can be guaranteed to be continuous, avoiding optimization oscillations.

[0108] In one embodiment, based on the data recognition results, a digital twin model of an earthquake fault is generated using digital twin technology, including:

[0109] Based on the identification results, a digital twin model of the strong earthquake fault is constructed using digital twin technology;

[0110] The probability and spatial distribution of earthquakes of different magnitudes are used as supplementary data to improve the digital twin model of strong earthquake faults;

[0111] The improved digital twin model of strong earthquake faults is used for digital detection of strong earthquake faults.

[0112] The working principle of this technical solution is as follows: Through digital twin technology, a virtual model highly similar to a real strong earthquake fault can be created. This model not only reflects the fault geometry but also incorporates its physical properties and dynamic behavior, providing a comprehensive and in-depth fault analysis platform. In the process of building the digital twin model, we fully consider the diversity and complexity of seismic activity. To further improve the accuracy and practicality of the model, the probability and spatial distribution of earthquakes of different magnitudes are incorporated as important supplementary data. This data helps to more accurately simulate the potential impact of seismic activity and the response mechanism of the fault under different earthquake intensities.

[0113] The beneficial effects of the above technical solution are: by adopting the solution provided in this embodiment, a series of digital detection work can be carried out based on this improved digital twin model; this not only includes the precise measurement and analysis of the fault morphology, but also involves the prediction and evaluation of the dynamic response of the fault under the action of an earthquake; through this method, we can have a deeper understanding of the characteristics and behavior of strong earthquake faults, and provide strong support for earthquake early warning and disaster prevention.

[0114] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A convolutional neural network strong earthquake fault digital detection method, characterized in that: include: Preprocessing the acquired three-dimensional seismic volume data of the strong earthquake fault to obtain preprocessed data; Using an improved three-dimensional convolutional neural network and multi-interpolation algorithm fusion technology, the pre-processed data is processed to obtain the data identification results of strong earthquake faults; Based on the data identification results, digital twin technology is used to generate a digital twin model of the earthquake fault; Using an improved three-dimensional convolutional neural network and multi-interpolation algorithm fusion technology, the pre-processed data is processed to obtain the data identification results of strong earthquake faults, including: In an upsampling layer of the three-dimensional convolutional neural network, a bilinear interpolation method is used to perform a first sampling process on the preprocessed data to obtain a first sampling process image; Based on the set gradient threshold of the pre-processed data, if the calculated local gradient of the pre-processed data is greater than the gradient threshold, the nearest neighbor interpolation method is used to perform a second sampling process on the pre-processed data to obtain a second sampling process image; Summarizing the first sampled processed image and the second sampled processed image to obtain a feature data set; Build an improved 3D convolutional neural network; The improved 3D convolutional neural network is a multi-scale feature pyramid fusion structure. The multi-scale feature pyramid fusion structure uses a dilation convolution layer with dilation rates of 2, 4, and 6 to extract multi-scale features. Its output is a 1×1×1 convolution layer to unify the number of channels. The encoder and decoder parts of the improved 3D convolutional neural network include an improved multi-scale residual module. The improved multi-scale residual module includes three encoding blocks, each of which consists of a 3×3×3 convolutional layer, a batch normalization layer, and a LeakyReLU activation function. The decoder part of the improved 3D convolutional neural network uses deconvolution layers for upsampling and channel splicing for skip connections; According to the characteristic data set, the improved three-dimensional convolutional neural network is used for data processing to obtain the data identification results of strong earthquake faults.

2. The method for digitally detecting strong earthquake faults using a convolutional neural network according to claim 1, characterized in that: The acquired three-dimensional seismic volume data of the strong earthquake fault are preprocessed to obtain preprocessed data, including: Obtain 3D seismic volume data of strong earthquake faults; The three-dimensional seismic volume data is preprocessed to obtain preprocessed data.

3. The method for digitally detecting strong earthquake faults using a convolutional neural network according to claim 2, characterized in that: Preprocess the 3D seismic volume data to obtain preprocessed data, including: The mean and standard deviation of each slice of 3D seismic volume data are calculated independently and normalized; Multi-scale denoising based on wavelet transform is used to denoise the normalized data; The denoised data are preliminarily enhanced by random rotation, horizontal mirroring and elastic deformation. For the data after the initial data enhancement processing, the second data enhancement processing data is obtained by generating samples of fault dislocation, formation tilt and porosity change through the adversarial network; The second data enhancement processing data is processed to eliminate data segments with a signal-to-noise ratio lower than a set signal-to-noise ratio threshold or with broken continuity, and to retain effective geological boundaries through three-dimensional gradient detection to obtain preprocessed data.

4. The method for digitally detecting strong earthquake faults using a convolutional neural network according to claim 1, wherein: Based on a set gradient threshold of the preprocessed data, if the calculated local gradient of the preprocessed data is greater than the gradient threshold, a nearest neighbor interpolation method is used to perform a second sampling process on the preprocessed data to obtain a second sampled processed image, including: Performing anisotropic normalization processing on the input pre-processed data to obtain anisotropic normalized processed data; The improved three-dimensional Sobel operator is used to calculate the local gradient value of the anisotropic normalized data; A gradient threshold of the preprocessed data is set. If the local gradient value is greater than the gradient threshold, the nearest neighbor interpolation method is used to perform a second preprocessing on the preprocessed data to obtain a second preprocessed image.

5. The method for digitally detecting strong earthquake faults using a convolutional neural network according to claim 4, characterized in that: Improved 3D Sobel operator, including: Reduce the Z-direction convolution kernel length to the set value, configure the weights of the X-direction, Y-direction, and Z-direction convolution kernel lengths, and set the weight ratio to 1:2:1; Set the convolution kernel lengths in the X and Y directions to be asymmetric; Three scales are distinguished: original resolution, 1 / 2 downsampling, and 1 / 4 downsampling. Corresponding scale weights are configured. The gradient fields are calculated separately and summarized to obtain the final calculated gradient. A three-dimensional non-local mean filter is applied before the gradient calculation.

6. The method for digitally detecting strong earthquake faults using a convolutional neural network according to claim 1, characterized in that: Constructing an improved three-dimensional convolutional neural network also includes: using a mixed loss function during network training; the mixed loss function includes a cross-entropy loss function, a Hausdorff distance loss function, and a dynamic regularization method based on interpolation; among them, the cross-entropy loss function is used for fault classification, the Hausdorff distance loss function is used to optimize the fault boundary accuracy, and the dynamic regularization method based on interpolation is used to balance computational efficiency and accuracy.

7. The method for digitally detecting strong earthquake faults using a convolutional neural network according to claim 1, characterized in that: Based on the data identification results, digital twin technology is used to generate a digital twin model of the earthquake fault, including: Based on the identification results, a digital twin model of the strong earthquake fault is constructed using digital twin technology; The probability and spatial distribution of earthquakes of different magnitudes are used as supplementary data to improve the digital twin model of strong earthquake faults; The improved digital twin model of strong earthquake faults is used for digital detection of strong earthquake faults.

Images