Deep learning-based rock fracture evolution prediction method and equipment

Abstract

The invention discloses a rock fracture evolution prediction method and equipment based on deep learning, and the method comprises the steps: applying stress to a rock under a stress condition, obtaining a three-dimensional CT scanning image of the rock changing with the stress, and obtaining a plurality of groups of rock time sequence image data, the rock time sequence image data comprising a plurality of rock time sequence images; based on each rock time sequence image in the rock time sequence image data, determining rock time sequence characteristic data corresponding to the rock time sequence image data, the rock time sequence characteristic data including rock time sequence characteristics of each rock time sequence image; determining fracture label data of the rock time sequence image data based on each rock time sequence image in the rock time sequence image data; based on the rock time sequence characteristics of the rock time sequence images, forming input of a rock fracture evolution prediction model, training the rock fracture evolution prediction model, and predicting a current rock fracture stage corresponding to the rock time sequence image data in current iteration; and based on the current rock fracture stage and the fracture label data, training the rock fracture evolution prediction model.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and computing device for predicting rock fracture evolution based on deep learning. Background Technology

[0002] The damage and fracture evolution of rocks is a key research focus in rock mechanics. Existing fracture analysis methods are mostly based on feature extraction at a single time point, failing to reflect the dynamic evolution of fractures. Furthermore, most methods can only describe the current state and cannot predict the next stage. There is currently no method to establish a quantitative mapping relationship between the temporal evolution trajectory of fracture matrix eigenvalues ​​and fracture development stages, and there is also a lack of technology to predict future evolution trends based on trajectory features. This paper proposes a method that can identify fracture evolution stages in real-time, quantitatively, and automatically, and predict the characteristics of the next stage, addressing the problems of strong subjectivity and insufficient predictive ability in existing technologies. Summary of the Invention

[0003] To address existing technical problems, this invention provides a deep learning-based method and computing device for predicting rock fracture evolution, which can adapt to the complex changes at different fracture stages, thereby providing reliable technical support for rock stability monitoring and disaster early warning.

[0004] In a first aspect, a deep learning-based method for predicting rock fracture evolution is provided, comprising: applying stress to a rock under stress conditions, acquiring three-dimensional CT scan images of the rock as the stress changes, and obtaining multiple sets of time-series rock image data, wherein the time-series rock image data includes multiple time-series rock images; determining rock time-series feature data corresponding to each of the time-series rock images, wherein the rock time-series feature data includes the rock time-series features of each of the time-series rock images; determining fracture label data of each of the time-series rock images based on each of the time-series rock images; forming the input of a rock fracture evolution prediction model based on the rock time-series features of each of the time-series rock images, training the rock fracture evolution prediction model, and predicting the current rock fracture stage corresponding to the time-series rock image data in the current iteration; and training the rock fracture evolution prediction model based on the current rock fracture stage and the fracture label data.

[0005] In a second aspect, a computing device is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the deep learning-based rock fracture evolution prediction method as provided in the first aspect of this application.

[0006] This application enables the training of a dynamic rock fracture evolution prediction model by extracting temporal features from time-series rock image data and combining them with fracture label data. This model not only captures the evolutionary patterns of fractures over time, improving prediction accuracy and robustness, but also continuously optimizes itself through iterative training to adapt to complex changes at different fracture stages, thus providing reliable technical support for rock stability monitoring and disaster early warning. Attached Figure Description

[0007] Figure 1 This is an application environment diagram of a deep learning-based rock fracture evolution prediction method in one embodiment; Figure 2 This is a flowchart of a deep learning-based method for predicting rock fracture evolution in one embodiment; Figure 3 This is a flowchart of rock time series feature data corresponding to rock time series image data in one embodiment; Figure 4 This is a network structure diagram of a rock fracture evolution prediction model in one embodiment; Figure 5 This is a schematic diagram of a rock fracture evolution prediction device based on deep learning in one embodiment; Figure 6 This is a schematic diagram of a computing device in one embodiment. Detailed Implementation

[0008] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0009] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0010] In the following description, the expression “some embodiments” refers to a subset of all possible embodiments. However, it should be understood that “some embodiments” can be the same subset or different subsets of all possible embodiments and can be combined with each other without conflict.

[0011] See Figure 1This diagram illustrates the application environment of a deep learning-based rock fracture evolution prediction method in one embodiment. The deep learning-based rock fracture evolution prediction method is applied to a computing device 10, which can acquire multiple sets of time-series rock image data. Each time-series rock image is a three-dimensional computed tomography (CT) image. During the scanning process, stress is applied to the rock according to stress conditions, and the rock is scanned simultaneously to obtain a three-dimensional CT scan image. The three-dimensional CT scan image is a three-dimensional image within a three-dimensional image, reconstructed from a three-dimensional measurement projection map. The three-dimensional measurement projection map can be acquired by a detection device performing a three-dimensional CT scan on the rock. The X-ray source in the detection device rotates around the rock, acquiring three-dimensional measurement projection maps of the rock from different angles. These projection data are then used to reconstruct a three-dimensional image of the scanned area. The acquisition process of the measurement projection map involves the attenuation of the X-ray source; tissues of different densities absorb X-rays to varying degrees, and this difference is used to distinguish different tissue structures. During the scan, the X-ray tube emits a narrow beam of X-rays, which is attenuated by a detector on the opposite side of the tube. The data collected by the detection equipment are X-ray projection and contour data.

[0012] Please see Figure 2 This is a flowchart illustrating a deep learning-based method for predicting rock fracture evolution according to an embodiment of this application. The deep learning-based method for predicting rock fracture evolution is applied in a computing device and includes the following steps: S11. Apply stress to the rock under stress conditions, obtain three-dimensional CT scan images of the rock as the stress changes, and obtain multiple sets of time-series image data of the rock.

[0013] In this embodiment, the rock time-series image data includes multiple rock time-series images, each corresponding to a specific time point, which is the time when stress was applied. The rock time-series image data is a sequence of images with a temporal order. By applying stress, the changes in rock cracks are observed, thereby obtaining training samples for subsequent training of the rock crack evolution prediction model.

[0014] S12. Based on each rock time series image in the rock time series image data, determine the rock time series feature data corresponding to the rock time series image data.

[0015] In this embodiment, the rock time-series feature data includes the rock time-series features of each rock time-series image, which indicate the data of fracture evolution state. Each rock time-series image corresponds to its own rock time-series features, so the data formed by the rock time-series features of each rock time-series image is also a set of time-series sequence data.

[0016] S13. Based on each of the rock time series images in the rock time series image data, determine the fracture label data of the rock time series image data.

[0017] In this embodiment, the fracture label data is used as the label data for training the rock fracture evolution prediction model. Features can be extracted from the time-series rock image data, and fracture label data is obtained based on these features.

[0018] S14. Based on the rock time series features of each rock time series image, form the input of the rock fracture evolution prediction model, train the rock fracture evolution prediction model, and predict the current rock fracture stage corresponding to the rock time series image data in the current iteration.

[0019] In this embodiment, the rock temporal features of each rock time-series image and the fracture label data of the rock time-series image data form a training sample. Since many sets of rock time-series image data are collected, the rock temporal feature data and fracture label data of these multiple sets of rock time-series image data form multiple sets of training samples. Based on these multiple sets of training samples, a rock fracture evolution prediction model is obtained.

[0020] In each current iteration, based on the rock time series features of each input rock time series image, the corresponding predicted current rock fracturing stage is output.

[0021] S15. Based on the current rock fracture stage and fracture label data, train a rock fracture evolution prediction model.

[0022] In this embodiment, based on the current rock fracture stage and fracture label data, the current loss value is calculated. If the current loss value is greater than a preset loss value, backpropagation is performed to continue the rock fracture evolution prediction model. If the current loss value is less than or equal to the preset loss value, the rock fracture evolution prediction model after stopping iterations is used as the pre-trained rock fracture evolution prediction model.

[0023] In this embodiment, time-series image data is used, meaning the model can capture the evolution of cracks over time. Extracting temporal features of the rock helps the model understand dynamic changes. Fracture label data is used as a supervisory signal to ensure the model learns the correct fracture stages. During training, the model iteratively predicts the current fracture stage and compares it with the labels to optimize the model. Through rock temporal features, the model can learn patterns and trends in crack development, improving prediction accuracy. Utilizing fracture label data provides a clear learning objective, making training more effective. The iterative training process allows the model to gradually optimize and adapt to changes in different fracture stages. Overall, this helps build a robust rock fracture evolution prediction model that can be used for practical monitoring and early warning.

[0024] In the above embodiments, by extracting temporal features from rock time-series image data and combining them with fracture label data, a dynamic rock fracture evolution prediction model can be trained. This model can not only capture the evolutionary patterns of fractures over time, improving the accuracy and robustness of predictions, but also be continuously optimized through iterative training to adapt to the complex changes in different fracture stages, thereby providing reliable technical support for rock stability monitoring and disaster early warning.

[0025] In some embodiments, such as Figure 3 As shown, Figure 3 This is a flowchart illustrating rock time-series feature data corresponding to rock time-series image data in one embodiment. The step of determining the rock time-series feature data corresponding to each rock time-series image in the rock time-series image data includes: S31. Based on the time series images of rocks, generate the crack intensity matrix of the time series images of rocks.

[0026] In this embodiment, the crack intensity matrix represents the crack intensity of the time-series rock image. The crack intensity matrix is ​​a binary matrix. Voxel representation The crack strength at time t, where This represents the spatial coordinates of a pixel. In the binary matrix, 0 represents intact material and 1 represents a crack; t represents the time when the image was acquired. Represents the coordinates on the X-axis. This represents the coordinates on the Y-axis. This represents the coordinates on the Z-axis. The X-axis is the horizontal direction, the Y-axis is the direction perpendicular to the X-axis on the horizontal plane, and the Z-axis is the direction perpendicular to both the X and Y axes.

[0027] S32. Based on the crack intensity matrix of the rock time series image, determine the crack gradient matrix of the rock time series image.

[0028] In this embodiment, the crack gradient matrix represents the direction of crack propagation front and the rate of intensity change. The crack gradient matrix is... This represents the expression, where G is a three-dimensional gradient vector field. The intensity change rate in the x-direction represents the gradient at the crack front. The rate of change of intensity in the Y direction. Let Z be the rate of change of intensity in the Z direction.

[0029] S33. Based on the crack gradient matrix of the rock time series image, determine the local structure tensor of the rock time series image.

[0030] In this embodiment, the local structure tensor represents the gradient information within the local neighborhood of the crack, and is used to capture the local structural features of the crack.

[0031] Optionally, determining the local structure tensor of the rock time-series image based on the crack gradient matrix includes: Calculate the gradient magnitude of each voxel in the crack gradient matrix; Voxels with gradient magnitudes greater than a preset magnitude are selected as target voxels. With each target voxel as a center point, neighboring voxels within a preset neighborhood of the target voxel are determined. Based on the voxel value of the target voxel and the voxel values ​​of each neighboring voxel, calculate the Gaussian weight of each neighboring voxel. Based on the Gaussian weights of each neighboring voxel and the crack gradient values ​​of each neighboring voxel, the local structure value of the target voxel is calculated, wherein the local structure tensor of the rock time series image includes the local structure values ​​of each target voxel in the rock time series image.

[0032] In this embodiment, the gradient magnitude of each voxel is used This indicates that a significance threshold is set based on the gradient magnitude. That is, the preset amplitude. Only for values ​​that meet the following criteria... The voxel is used as the target voxel. A target voxel... For example, the neighborhood of the target voxel is used Indicates, such as , or The neighborhood size can be adaptively selected based on the gradient magnitude of the central voxel: a smaller neighborhood is used when the gradient magnitude is large to preserve detailed features, and a larger neighborhood is used when the gradient magnitude is small to enhance noise resistance. Target voxel The expression for calculating the local structure value is as follows: .

[0033] Weighting function To target voxels Gaussian weights with a core: ,in is the Gaussian kernel scale parameter.

[0034] In the above embodiments, by focusing on significant gradient regions (target voxels) and utilizing Gaussian-weighted neighborhood analysis, the local structure and orientation features of the crack edge can be accurately quantified. This method not only effectively suppresses the interference of image noise and enhances the robustness of feature extraction, but also provides a solid basis for subsequent analysis with the obtained local structure tensor, which can more reliably describe the crack's direction, intensity, and evolution trend, thereby significantly improving the accuracy of crack detection and evolution prediction.

[0035] S34. Based on the local structure tensor of the rock time series image, determine the rock time series features of the rock time series image.

[0036] In this embodiment, the rock time series features include: crack planarity, crack linearity, angle between adjacent time series vectors, crack change rate, and crack change acceleration, wherein the angle between adjacent time series vectors is used to quantify the stability of the crack propagation direction of the rock time series image; The determination of the rock temporal features of the rock temporal image based on the local structure tensor of the rock temporal image includes: Based on the local structure tensor of the rock time series image, multiple feature values ​​of the local structure tensor and the feature vector corresponding to each feature value are calculated. The feature values ​​include a first feature value, a second feature value, and a third feature value, wherein the first feature value is greater than the second feature value, the second feature value is greater than the third feature value, and the feature value represents the intensity of the gradient change, and the feature vector represents the gradient distribution direction of the crack. The ratio of the second feature value to the first feature value is determined as the degree of crack plane. The ratio of the third characteristic value to the second characteristic value is determined as the degree of crack linearity; Obtain neighboring time series images that are temporally adjacent to the rock time series image, and determine the angle between the first feature vector of the neighboring time series image and the first feature vector of the rock time series image; The crack change rate of the rock time series image is determined based on the difference between the feature values ​​of the neighboring time series image and the feature values ​​of the rock time series image, and the difference between the first feature vector of the neighboring time series image and the first feature vector of the rock time series image. The crack change acceleration in the rock time series image is determined based on the crack change rate in the rock time series image and the crack change rate in the adjacent time series images.

[0037] In this embodiment, calculation eigenvalues ​​of a matrix and corresponding feature vectors .in The first eigenvalue, The second eigenvalue, It is the third eigenvalue. The first eigenvector corresponding to the first eigenvalue. This is the second eigenvector corresponding to the second eigenvalue. This is the third eigenvector corresponding to the third eigenvalue. The first eigenvalue represents the intensity of the gradient change in the direction of strongest change. (And the largest eigenvalue...) The corresponding eigenvector points in the direction of the most dramatic gradient change (i.e., the normal direction). This is related to the minimum eigenvalue. The corresponding eigenvector points to the direction in which the gradient change is most gradual, i.e. the direction of the structure (for linear structures such as cracks, this is the tangent direction of the crack).

[0038] For each time-series rock image, the respective time-series characteristics can be obtained by following the steps described above. The degree of crack planarity is used as... express, , This refers to the degree to which a crack structure approaches a plane. Time refers to the two-dimensional extension being dominant. This refers to cracks that extend in a linear or dotted pattern.

[0039] Crack linearity is used express, Used to describe the degree to which the structure exhibits linearity. This indicates that the crack is highly linear. This indicates that the crack is in a planar state.

[0040] Used to describe the degree of isotropy of cracks This indicates that the development of cracks is isotropic. This indicates the anisotropy of crack development. By constructing these three parameters, absolute eigenvalues ​​are transformed into relative contrast values, eliminating the influence of the strength scale and highlighting the shape parameters of the structure.

[0041] This represents the trajectory velocity of the crack change. The calculation formula is as follows: .

[0042] in The degree of crack plane in adjacent time series images, The linearity of the cracks in adjacent time-series images. This represents the time difference between neighboring time series images and rock time series images. , The first feature vector representing the time series image of rocks The first feature vector of neighboring time series images The angle between two adjacent time steps is the angle between them.

[0043] Indicates the acceleration of the crack change. .in This represents the trajectory velocity of crack changes in adjacent time-series images.

[0044] In the above embodiments, this method quantifies the geometric properties (planarity, linearity) and dynamic evolution characteristics (rate of change, acceleration) of cracks, enabling a multi-dimensional characterization of rock cracks from static structure to temporal evolution. It can not only accurately describe the spatial morphological characteristics of cracks (such as distinguishing between planar and linear structures), but also capture their extension, turning and accelerated fracturing behavior over time, thereby significantly improving the understanding and prediction ability of crack development process and providing a more reliable quantitative basis for geological stability assessment and disaster early warning.

[0045] In some embodiments, the fracture label data includes fracture stage labels, and determining the fracture label data of the rock time-series image data based on each of the rock time-series images includes: When the angle between adjacent time-series vectors in the rock time-series image data is random, the first feature value of the first rock time-series image in the rock time-series image data is less than the first feature threshold, the second feature value of the first rock time-series image is less than the second feature threshold, and the second feature value of the first rock time-series image is less than the first feature value by a preset percentage, the fracture stage label of the rock time-series image data is determined to be the fracture initiation period, wherein the first rock time-series image is the latest image in the rock time-series image data sorted according to the evolution time sequence; When the angle between each adjacent time-series vector in the rock time-series image data is within the error range of 0°, and the first feature value of the first rock time-series image in the rock time-series image data is less than the third feature threshold, the second feature value of the first rock time-series image is less than the fourth feature threshold, and the third feature value of the first rock time-series image is less than the fifth feature threshold, the fracture stage label of the rock time-series image data is determined to be the linear fracture propagation period. When the angle between each adjacent time-series vector in the rock time-series image data is less than a first angle threshold, the first feature value of the first rock time-series image in the rock time-series image data is greater than a sixth feature threshold, the second feature value of the first rock time-series image is less than a seventh feature threshold, and the change of each first feature vector in the rock time-series image data is within a preset change range, the fracture stage label of the rock time-series image data is determined to be the fracture plane propagation period. When the angle between adjacent time-series vectors in the rock time-series image data gradually increases as the evolution time progresses, the first feature value in the rock time-series image data gradually increases as the evolution time progresses, the second feature value of the first rock time-series image gradually increases as the evolution time progresses, and the change in the angle between adjacent time-series vectors in the rock time-series image data is greater than a preset angle value, the fracture stage label of the rock time-series image data is determined to be the accurate period of fracture bifurcation; When the angle between adjacent time-series vectors in the rock time-series image data is greater than a second angle threshold, the first feature value of the first rock time-series image in the rock time-series image data is greater than an eighth feature threshold, the second feature value of the first rock time-series image is greater than a ninth feature threshold, and the difference between each first feature vector in the rock time-series image data is greater than a second preset change value, the fracture stage label of the rock time-series image data is determined to be the fracture bifurcation occurrence period. When the fracture trajectory of the first rock time series image tends to converge, the fracture stage label of the rock time series image data is determined to be the fracture connection period; When the crack change rate of the first rock time series image is within the error range of 0, and the crack change acceleration of the first rock time series image is less than 0, the cracking stage label of the rock time series image data is determined to be the crack propagation cessation period.

[0046] In this embodiment, the rock time-series image data includes N rock time-series images, each rock time-series image corresponding to a time point, for example, from the first time point to the fourth time point, where the first time point is earlier than the fourth time point, and the rock time-series image corresponding to the fourth time point is the first rock time-series image.

[0047] For example when Random, and When the second eigenvalue is less than 20% of the first eigenvalue, it means that one-dimensional crack propagation is dominant, and this is the crack initiation period, with cracks spreading in a point-like manner.

[0048] The third feature threshold is greater than the first feature threshold, and the third feature threshold is greater than the second feature threshold. When the third feature value of the first rock time series image is less than the fifth feature threshold, it indicates that the fracture third-order orientation is very weak. For example, when ,and At this stage, the third-dimensional direction of the crack is very weak, exhibiting a planar or linear structure, and the crack expands stably in one dimension. This is the period of linear crack expansion.

[0049] For example, when ,and When the crack extends stably in two dimensions, and the changes in each first feature vector in the rock time-series image data are within a preset range, it indicates that the direction is basically stable, and this belongs to the planar extension period of the crack. The changes in each first feature vector can be determined by comparing the first feature vectors of adjacent time steps.

[0050] when It starts to grow. When all begin to rise, the crack begins to have three-dimensional features. When the change in the angle between each adjacent time series vector in the rock time series image data is greater than the preset angle value, it indicates that the angle has changed to a certain extent, the crack is about to bifurcate, and it belongs to the crack bifurcation preparation period. The change in the angle between each adjacent time series vector can be determined by comparing the angle between adjacent time series vectors in adjacent time steps.

[0051] when ,and At that time, the crack has significant three-dimensional features. The angle between each adjacent time vector in the rock time series image data is greater than the second angle threshold, indicating that the angle has changed significantly. The crack extends in multiple directions and belongs to the crack bifurcation period. The second angle threshold is greater than the preset angle value.

[0052] When the crack trajectory converges towards a nearby crack, a crack network is formed, which is the crack connection period. The crack trajectory in the first rock time-series image can be compared with the actual cracks to determine whether the crack trajectory tends to converge. Convergence is determined when the difference between the position of the crack trajectory and the actual crack is less than a preset distance. The crack trajectory in the first rock time-series image can be obtained from the crack intensity matrix.

[0053] when At that point, the crack propagation stopped.

[0054] In some embodiments, the step of using the rock temporal features of each of the rock temporal images as input to form a rock fracture evolution prediction model, training the rock fracture evolution prediction model, and predicting the current rock fracture stage corresponding to the rock temporal image data in the current iteration includes: Based on the rock temporal features of each of the rock temporal images, the rock temporal enhancement features of each of the rock fracture evolution prediction models are obtained using the feature enhancement layer of the rock fracture evolution prediction model. Based on the temporal enhancement features of each of the rock temporal images, the inputs of multiple temporal convolutional networks in the rock fracture evolution prediction model are formed. Temporal feature tensors are extracted through the temporal convolutional networks to determine multiple temporal feature tensors, wherein different temporal convolutional networks correspond to different scales. Based on multiple temporal feature tensors, the input of the fusion network in the rock fracture evolution prediction model is formed, and the output is a fusion feature tensor; Based on the fused feature tensor, the input of the dual-pooling network in the rock fracture evolution prediction model is formed, and the output is an enhanced fused feature tensor. Based on the enhanced fusion feature tensor, the input to the prediction network in the rock fracture evolution prediction model is formed, and the current rock fracture stage corresponding to the rock time series image data is output. Based on the enhanced fusion feature tensor, the input to the regression network in the rock fracture evolution prediction model is formed to predict the fracture prediction time corresponding to the rock time series image data.

[0055] In this embodiment, the rock time series characteristics are used Indicates. For example... Figure 6 As shown, the input layer of the rock fracture evolution prediction model receives 5-dimensional rock time-series features. For example, if t=50, the input data is a 50×5 matrix. Then, the feature enhancement layer preprocesses and enhances the rock time-series features output from the input layer.

[0056] Optionally, the step of obtaining enhanced rock time-series features for each of the rock time-series images based on their respective rock time-series features, using the feature enhancement layer of the rock fracture evolution prediction model, includes: For the rock temporal features of the rock temporal image, the crack planarity, the crack linearity, and the angle between adjacent temporal vectors are coupled to obtain coupled feature values; The rock temporal features and the coupled feature values ​​are formed into a vector, which is used as the rock temporal enhancement feature of the rock temporal image.

[0057] In this embodiment, it is proposed , , Perform coupling to obtain coupling eigenvalues This approach captures the physical consensus of fracture planarity, linear synchronous uplift, and abrupt changes in direction. While ensuring that the input data conforms to fundamental physical laws, it also enhances the geological characterization capabilities of the features, enabling the model to learn richer features. This involves combining time-series rock images... , , , , , Formation vectors are used as rock temporal enhancement features for rock temporal images.

[0058] like Figure 6 As shown, this application constructs a multi-scale temporal convolutional network for rock fracture prediction. For example, it uses three parallel temporal convolutional networks (TCNs): a short-term branch, a medium-term branch, and a long-term branch. Their convolutional kernels, ranging from small to large (3, 7, and 15), respectively focus on capturing short-term fluctuations, medium-term trends, and long-term patterns. By using convolutional kernels of different sizes and dilation coefficients, the model can simultaneously understand the instantaneous changes, staged development, and overall patterns of fracture evolution. The three output feature vectors are fused and then fed into a fusion network. For example, the fusion network includes a multi-head attention mechanism, which allows the model to focus on important events, significantly improving prediction accuracy and interpretability. Further enhancements are then made to improve the model's feature extraction capabilities.

[0059] This application designs a physical-feature fusion layer, which implements physical penalty and feature compression through the same dual-pooling network. The dual-pooling network includes max-pooling and average-pooling layers. For example, if the fused feature tensor is 64-dimensional, using classic dual-pooling, the average is applied over t time steps to obtain the overall trend of the 64-dimensional feature vector. Simultaneously, the maximum value is taken for each channel over t time steps to obtain the most salient 64-dimensional feature, also known as the most salient event. These two 64-dimensional vectors are concatenated along the channel dimension, and the dual-pooling network ultimately outputs a 128-dimensional enhanced fused feature tensor for prediction and a scalar physical loss.

[0060] A multi-task architecture with shared representation is adopted, namely, predicting the current rock fracture stage corresponding to the time series image data of rocks through a prediction network, and predicting the fracture prediction time corresponding to the time series image data of rocks through a regression network. The two tasks share the physical features extracted by the network, but perform specialized predictions through independent heads. For example, in the prediction network, the 128-dimensional shared features are processed by independent branches, the dense network (64) extracts semantic abstraction, ReLU activation is then mapped to 7 stages logits by Dense (7), and finally Softmax normalization is used to obtain the probability of each rock fracture stage. The rock fracture stage with the highest probability is taken as the current rock fracture stage. At the same time, the same shared feature enters the regression path. In the regression network, the semantics are first compressed by the dense network (32), and after ReLU activation, the scalar is output by Dense (1). Then, the ReLU is truncated to a positive value to obtain the fracture time estimate and output the fracture prediction time. The classification head focuses on the semantic division of fracture stages, and the regression head focuses on the accurate prediction of fracture time. This architecture not only improves the computational efficiency, but also enhances the overall performance through mutual promotion between tasks. Through this series of innovative designs, PGAT-Net has achieved a perfect combination of data-driven approaches and geology in the field of rock fracture prediction, providing new ideas and technologies for fracture prediction in geological engineering.

[0061] Optionally, the rupture tag data includes rupture stage tags and rupture time tags; The process of training the rock fracture evolution prediction model based on the current rock fracture stage and the fracture label data includes: Calculate the current classification loss based on the current rock fracture stage and the fracture stage label; Based on the predicted rupture time and the rupture time label, calculate the regression prediction loss; Calculate the crack boundary change intensity of each of the rock time-series image data, and calculate the physical loss corresponding to the rock time-series image data based on the crack boundary change intensity of each of the rock time-series image data; Calculate the current loss value based on the current classification loss, the regression prediction loss, and the physical loss; Based on the current loss value, backpropagation is performed to train the rock fracture evolution prediction model until the training termination condition is met.

[0062] In this embodiment, when the current rock fracture stage is different from the fracture stage label, the current classification loss is assigned a first preset classification loss value, such as 1; when the current rock fracture stage is the same as the fracture stage label, the current classification loss is assigned a second preset classification loss value, such as 0. The regression prediction loss is determined based on the distance between the fracture prediction time and the fracture time label; the greater the distance, the greater the regression prediction loss.

[0063] Optionally, calculating the physical loss corresponding to the rock time-series image data based on the crack boundary change intensity of each of the rock time-series images includes: Obtain the load-displacement curves corresponding to the time-series images of the rock; Based on the load-displacement curve, the external force work with the same time sequence as each of the rock time series images is obtained, and the external force work corresponding to each of the rock time series images is determined; The number of images in which the crack boundary change intensity of the rock time series image is greater than the external force work corresponding to the rock time series image is determined; The number of images is multiplied by a preset loss value to obtain the physical loss corresponding to the time-series rock image data.

[0064] In this embodiment, the load-displacement curve refers to the stress multiplied by the displacement used for synchronous load, and the load-displacement curve can be obtained under experimental conditions. The load-displacement curve represents the external work done at different times. Therefore, based on the time corresponding to the rock time series image, the external work done corresponding to the rock time series image can be determined from the load-displacement curve. For a rock time series image, the Sobel operator is used to calculate the gray-level gradient amplitude of the crack boundary of the rock time series image pixel by pixel to determine the maximum jump amplitude of the rock time series image. The boundary change intensity is equal to the maximum jump amplitude multiplied by the voxel volume at the position corresponding to the maximum jump amplitude. When the crack boundary change intensity of the rock time series image is greater than the external work done corresponding to the rock time series image, it indicates that the system energy is negative, violating conservation, and a positive penalty, i.e., physical loss, is given. Otherwise, there is no penalty. The physical loss is used as part of the current loss value. There are multiple rock time series images in the rock time series image data. The rock time series image that satisfies the condition that the crack boundary change intensity is greater than the external work is selected.

[0065] First, classification loss ensures the model correctly identifies the current crack stage. Regression loss allows the model to predict the rupture time, providing specific time warnings. However, the key is the physical loss, which uses the intensity of crack boundary changes as a constraint, forcing the model to consider actual physical changes during the learning process, such as energy dissipation or structural evolution, ensuring predictions don't deviate from physical realities. To enhance generalization ability in real-world scenarios or avoid physically unreasonable predictions, the physical loss is integrated. This ensures the model not only fits the data but also follows known physical laws, such as the energy principles of crack propagation. This enhances the model's interpretability and reliability, especially in situations with limited or noisy data.

[0066] Furthermore, multi-task training with shared features can improve efficiency and avoid overfitting. Classification and regression tasks can learn more robust feature representations from physical constraints. Physical loss ensures that model predictions are consistent with energy changes, which helps the model understand the dynamics of crack evolution.

[0067] This training method cleverly integrates fracture stage identification, fracture time prediction, and physical consistency constraints by constructing a multi-task loss function. Specifically, the classification loss ensures that the model accurately determines the physical stage of the current fracture, such as initiation, propagation, or bifurcation; the regression loss drives the model to learn the time mapping relationship from the current state to the final fracture, providing a quantitative early warning indicator; and the most innovative physical loss, by forcing the evolution trend predicted by the model to be consistent with the physical observation of the fracture boundary change intensity calculated by gradient energy in time, injects prior knowledge of rock fracture mechanics into the model. This approach offers three core benefits: First, it enhances the physical interpretability and generalization ability of the model, making predictions not merely data fitting but inferences consistent with physical laws, thus increasing reliability when facing complex geological conditions not covered by training data. Second, the physical loss, as a powerful regularization term, effectively prevents the model from overfitting to data noise or spurious correlations, guiding the model to focus on essential characteristics related to the actual fracturing process. Finally, it achieves closed-loop verification of microscopic image features and macroscopic mechanical behavior, enabling the model's predictions (such as impending bifurcation) to be intrinsically explained and verified through corresponding physical signal changes (such as a surge in boundary energy), greatly enhancing the credibility and engineering practical value of the entire prediction system.

[0068] In another aspect, this application provides a computer program product, including a computer program that, when executed by a processor, implements the deep learning-based rock fracture evolution prediction method described in any embodiment of this application.

[0069] In the computer program product, the optional implementation form of the program module architecture of the computer program that implements each step of the target identification method can be a rock fracture evolution prediction device based on deep learning.

[0070] Please see Figure 5 This application provides a deep learning-based rock fracture evolution prediction device, comprising: an acquisition module 51, configured to apply stress to a rock under stress conditions, acquire three-dimensional CT scan images of the rock as the stress changes, and obtain multiple sets of rock time-series image data, the rock time-series image data including multiple rock time-series images; a determination module 52, configured to determine rock time-series feature data corresponding to each rock time-series image data based on each rock time-series image in the rock time-series image data, the rock time-series feature data including rock time-series features of each rock time-series image; the determination module 52 is further configured to determine fracture label data of the rock time-series image data based on each rock time-series image in the rock time-series image data; a training module 53, configured to form the input of a rock fracture evolution prediction model based on the rock time-series features of each rock time-series image, train the rock fracture evolution prediction model, and predict the current rock fracture stage corresponding to the rock time-series image data in the current iteration; the training module 53 is further configured to train the rock fracture evolution prediction model based on the current rock fracture stage and the fracture label data.

[0071] Optionally, the determining module 52 is also used for: Based on the time-series rock images, a crack intensity matrix of the time-series rock images is generated; Based on the crack intensity matrix of the rock time series image, the crack gradient matrix of the rock time series image is determined; Based on the crack gradient matrix of the rock time series image, the local structure tensor of the rock time series image is determined; Based on the local structure tensor of the rock time series image, the rock time series features of the rock time series image are determined.

[0072] Optionally, the determining module 52 is also used for: Calculate the gradient magnitude of each voxel in the crack gradient matrix; Voxels with gradient magnitudes greater than a preset magnitude are selected as target voxels. With each target voxel as a center point, neighboring voxels within a preset neighborhood of the target voxel are determined. Based on the voxel value of the target voxel and the voxel values ​​of each neighboring voxel, calculate the Gaussian weight of each neighboring voxel. Based on the Gaussian weights of each neighboring voxel and the crack gradient values ​​of each neighboring voxel, the local structure value of the target voxel is calculated, wherein the local structure tensor of the rock time series image includes the local structure values ​​of each target voxel in the rock time series image.

[0073] Optionally, the determining module 52 is also used for: Based on the local structure tensor of the rock time series image, multiple feature values ​​of the local structure tensor and the feature vector corresponding to each feature value are calculated. The feature values ​​include a first feature value, a second feature value, and a third feature value, wherein the first feature value is greater than the second feature value, the second feature value is greater than the third feature value, and the feature value represents the intensity of the gradient change, and the feature vector represents the gradient distribution direction of the crack. The ratio of the second feature value to the first feature value is determined as the degree of crack plane. The ratio of the third characteristic value to the second characteristic value is determined as the degree of crack linearity; Obtain neighboring time series images that are temporally adjacent to the rock time series image, and determine the angle between the first feature vector of the neighboring time series image and the first feature vector of the rock time series image; The crack change rate of the rock time series image is determined based on the difference between the feature values ​​of the neighboring time series image and the feature values ​​of the rock time series image, and the difference between the first feature vector of the neighboring time series image and the first feature vector of the rock time series image. The crack change acceleration in the rock time series image is determined based on the crack change rate in the rock time series image and the crack change rate in the adjacent time series images.

[0074] Optionally, the determining module 52 is also used for: When the angle between adjacent time-series vectors in the rock time-series image data is random, the first feature value of the first rock time-series image in the rock time-series image data is less than the first feature threshold, the second feature value of the first rock time-series image is less than the second feature threshold, and the second feature value of the first rock time-series image is less than the first feature value by a preset percentage, the fracture stage label of the rock time-series image data is determined to be the fracture initiation period, wherein the first rock time-series image is the latest image in the rock time-series image data sorted according to the evolution time sequence; When the angle between each adjacent time-series vector in the rock time-series image data is within the error range of 0°, and the first feature value of the first rock time-series image in the rock time-series image data is less than the third feature threshold, the second feature value of the first rock time-series image is less than the fourth feature threshold, and the third feature value of the first rock time-series image is less than the fifth feature threshold, the fracture stage label of the rock time-series image data is determined to be the linear fracture propagation period. When the angle between each adjacent time-series vector in the rock time-series image data is less than a first angle threshold, the first feature value of the first rock time-series image in the rock time-series image data is greater than a sixth feature threshold, the second feature value of the first rock time-series image is less than a seventh feature threshold, and the change of each first feature vector in the rock time-series image data is within a preset change range, the fracture stage label of the rock time-series image data is determined to be the fracture plane propagation period. When the angle between adjacent time-series vectors in the rock time-series image data gradually increases as the evolution time progresses, the first feature value in the rock time-series image data gradually increases as the evolution time progresses, the second feature value of the first rock time-series image gradually increases as the evolution time progresses, and the change in the angle between adjacent time-series vectors in the rock time-series image data is greater than a preset angle value, the fracture stage label of the rock time-series image data is determined to be the accurate period of fracture bifurcation; When the angle between adjacent time-series vectors in the rock time-series image data is greater than a second angle threshold, the first feature value of the first rock time-series image in the rock time-series image data is greater than an eighth feature threshold, the second feature value of the first rock time-series image is greater than a ninth feature threshold, and the difference between each first feature vector in the rock time-series image data is greater than a second preset change value, the fracture stage label of the rock time-series image data is determined to be the fracture bifurcation occurrence period. When the fracture trajectory of the first rock time series image tends to converge, the fracture stage label of the rock time series image data is determined to be the fracture connection period; When the crack change rate of the first rock time series image is within the error range of 0, and the crack change acceleration of the first rock time series image is less than 0, the cracking stage label of the rock time series image data is determined to be the crack propagation cessation period.

[0075] Optionally, training module 53 is also used for: Based on the rock temporal features of each of the rock temporal images, the rock temporal enhancement features of each of the rock fracture evolution prediction models are obtained using the feature enhancement layer of the rock fracture evolution prediction model. Based on the temporal enhancement features of each of the rock temporal images, the inputs of multiple temporal convolutional networks in the rock fracture evolution prediction model are formed. Temporal feature tensors are extracted through the temporal convolutional networks to determine multiple temporal feature tensors, wherein different temporal convolutional networks correspond to different scales. Based on multiple temporal feature tensors, the input of the fusion network in the rock fracture evolution prediction model is formed, and the output is a fusion feature tensor; Based on the fused feature tensor, the input of the dual-pooling network in the rock fracture evolution prediction model is formed, and the output is an enhanced fused feature tensor. Based on the enhanced fusion feature tensor, the input to the prediction network in the rock fracture evolution prediction model is formed, and the current rock fracture stage corresponding to the rock time series image data is output. Based on the enhanced fusion feature tensor, the input to the regression network in the rock fracture evolution prediction model is formed to predict the fracture prediction time corresponding to the rock time series image data.

[0076] Optionally, training module 53 is also used for: For the rock temporal features of the rock temporal image, the crack planarity, the crack linearity, and the angle between adjacent temporal vectors are coupled to obtain coupled feature values; The rock temporal features and the coupled feature values ​​are formed into a vector, which is used as the rock temporal enhancement feature of the rock temporal image.

[0077] Optionally, training module 53 is also used for: Calculate the current classification loss based on the current rock fracture stage and the fracture stage label; Based on the predicted rupture time and the rupture time label, calculate the regression prediction loss; Calculate the crack boundary change intensity of each of the rock time-series image data, and calculate the physical loss corresponding to the rock time-series image data based on the crack boundary change intensity of each of the rock time-series image data; Calculate the current loss value based on the current classification loss, the regression prediction loss, and the physical loss; Based on the current loss value, backpropagation is performed to train the rock fracture evolution prediction model until the training termination condition is met.

[0078] Optionally, training module 53 is also used for: Obtain the load-displacement curves corresponding to the time-series images of the rock; Based on the load-displacement curve, the external force work with the same time sequence as each of the rock time series images is obtained, and the external force work corresponding to each of the rock time series images is determined; The number of images in which the crack boundary change intensity of the rock time series image is greater than the external force work corresponding to the rock time series image is determined; The number of images is multiplied by a preset loss value to obtain the physical loss corresponding to the time-series rock image data.

[0079] Please see Figure 6In another aspect of this application, a computing device 10 is also provided, including a memory 3011 and a processor 3012. The memory 3011 stores a computer program, and when the computer program is executed by the processor, the processor 3012 performs the steps of the deep learning-based rock fracture evolution prediction method provided in any of the above embodiments of this application. The computing device 10 can be (e.g., a desktop computer, laptop computer, tablet computer, handheld computer, smart speaker, server, etc.), a mobile phone (e.g., a smartphone, cordless phone, etc.), a wearable device (e.g., a pair of smart glasses or a smartwatch), or a similar device.

[0080] The processor 3012 is the control center, connecting various parts of the computer device via various interfaces and lines. It executes software programs and / or modules stored in the memory 3011, and calls data stored in the memory 3011 to perform various functions and process data. Optionally, the processor 3012 may include one or more processing cores; preferably, the processor 3012 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user page, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 3012.

[0081] The memory 3011 can be used to store software programs and modules. The processor 3012 executes various functional applications and data processing by running the software programs and modules stored in the memory 3011. The memory 3011 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 3011 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 3011 may also include a memory controller to provide the processor 3012 with access to the memory 3011.

[0082] In another aspect, this application also provides a storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the deep learning-based rock fracture evolution prediction method provided in any of the above embodiments of this application.

[0083] Those skilled in the art will understand that all or part of the processes in the methods provided in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0084] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting rock fracture evolution based on deep learning, characterized in that, include: Stress is applied to the rock under stress conditions, and three-dimensional CT scan images of the rock as the stress changes are obtained to obtain multiple sets of time-series rock image data, which include multiple time-series rock images; Based on each of the rock time-series images in the rock time-series image data, rock time-series feature data corresponding to the rock time-series image data is determined, and the rock time-series feature data includes the rock time-series features of each of the rock time-series images; Based on each of the rock time-series images in the rock time-series image data, determine the fracture label data of the rock time-series image data; Based on the rock time series features of each of the rock time series images, the input of the rock fracture evolution prediction model is formed, the rock fracture evolution prediction model is trained, and the current rock fracture stage corresponding to the rock time series image data in the current iteration is predicted. Based on the current rock fracture stage and the fracture label data, the rock fracture evolution prediction model is trained.

2. The rock fracture evolution prediction method based on deep learning as described in claim 1, characterized in that, The step of determining the rock time-series feature data corresponding to each of the rock time-series image data includes: Based on the time-series rock images, a crack intensity matrix of the time-series rock images is generated; Based on the crack intensity matrix of the rock time series image, the crack gradient matrix of the rock time series image is determined; Based on the crack gradient matrix of the rock time series image, the local structure tensor of the rock time series image is determined; Based on the local structure tensor of the rock time series image, the rock time series features of the rock time series image are determined.

3. The rock fracture evolution prediction method based on deep learning as described in claim 2, characterized in that, The determination of the local structure tensor of the rock time-series image based on the crack gradient matrix includes: Calculate the gradient magnitude of each voxel in the crack gradient matrix; Voxels with gradient magnitudes greater than a preset magnitude are selected as target voxels. With each target voxel as a center point, neighboring voxels within a preset neighborhood of the target voxel are determined. Based on the voxel value of the target voxel and the voxel values ​​of each neighboring voxel, calculate the Gaussian weight of each neighboring voxel. Based on the Gaussian weights of each neighboring voxel and the crack gradient values ​​of each neighboring voxel, the local structure value of the target voxel is calculated, wherein the local structure tensor of the rock time series image includes the local structure values ​​of each target voxel in the rock time series image.

4. The rock fracture evolution prediction method based on deep learning as described in claim 2, characterized in that, The rock time series features include: crack planarity, crack linearity, angle between adjacent time series vectors, crack change rate, and crack change acceleration, wherein the angle between adjacent time series vectors is used to quantify the stability of the crack propagation direction of the rock time series image; The determination of the rock temporal features of the rock temporal image based on the local structure tensor of the rock temporal image includes: Based on the local structure tensor of the rock time series image, multiple feature values ​​of the local structure tensor and the feature vector corresponding to each feature value are calculated. The feature values ​​include a first feature value, a second feature value, and a third feature value, wherein the first feature value is greater than the second feature value, the second feature value is greater than the third feature value, and the feature value represents the intensity of the gradient change, and the feature vector represents the gradient distribution direction of the crack. The ratio of the second feature value to the first feature value is determined as the degree of crack plane. The ratio of the third characteristic value to the second characteristic value is determined as the degree of crack linearity; Obtain neighboring time series images that are temporally adjacent to the rock time series image, and determine the angle between the first feature vector of the neighboring time series image and the first feature vector of the rock time series image; The crack change rate of the rock time series image is determined based on the difference between the feature values ​​of the neighboring time series image and the feature values ​​of the rock time series image, and the difference between the first feature vector of the neighboring time series image and the first feature vector of the rock time series image. The crack change acceleration in the rock time series image is determined based on the crack change rate in the rock time series image and the crack change rate in the adjacent time series images.

5. The rock fracture evolution prediction method based on deep learning as described in claim 4, characterized in that, The fracture label data includes fracture stage labels. The method for determining the fracture label data of the rock time-series image data based on each of the rock time-series images includes: When the angle between adjacent time-series vectors in the rock time-series image data is random, the first feature value of the first rock time-series image in the rock time-series image data is less than the first feature threshold, the second feature value of the first rock time-series image is less than the second feature threshold, and the second feature value of the first rock time-series image is less than the first feature value by a preset percentage, the fracture stage label of the rock time-series image data is determined to be the fracture initiation period, wherein the first rock time-series image is the latest image in the rock time-series image data sorted according to the evolution time sequence; When the angle between each adjacent time-series vector in the rock time-series image data is within the error range of 0°, and the first feature value of the first rock time-series image in the rock time-series image data is less than the third feature threshold, the second feature value of the first rock time-series image is less than the fourth feature threshold, and the third feature value of the first rock time-series image is less than the fifth feature threshold, the fracture stage label of the rock time-series image data is determined to be the linear fracture propagation period. When the angle between each adjacent time-series vector in the rock time-series image data is less than a first angle threshold, the first feature value of the first rock time-series image in the rock time-series image data is greater than a sixth feature threshold, the second feature value of the first rock time-series image is less than a seventh feature threshold, and the change of each first feature vector in the rock time-series image data is within a preset change range, the fracture stage label of the rock time-series image data is determined to be the fracture plane propagation period. When the angle between adjacent time-series vectors in the rock time-series image data gradually increases as the evolution time progresses, the first feature value in the rock time-series image data gradually increases as the evolution time progresses, the second feature value of the first rock time-series image gradually increases as the evolution time progresses, and the change in the angle between adjacent time-series vectors in the rock time-series image data is greater than a preset angle value, the fracture stage label of the rock time-series image data is determined to be the accurate period of fracture bifurcation; When the angle between adjacent time-series vectors in the rock time-series image data is greater than a second angle threshold, the first feature value of the first rock time-series image in the rock time-series image data is greater than an eighth feature threshold, the second feature value of the first rock time-series image is greater than a ninth feature threshold, and the difference between each first feature vector in the rock time-series image data is greater than a second preset change value, the fracture stage label of the rock time-series image data is determined to be the fracture bifurcation occurrence period. When the fracture trajectory of the first rock time series image tends to converge, the fracture stage label of the rock time series image data is determined to be the fracture connection period; When the crack change rate of the first rock time series image is within the error range of 0, and the crack change acceleration of the first rock time series image is less than 0, the cracking stage label of the rock time series image data is determined to be the crack propagation cessation period.

6. The rock fracture evolution prediction method based on deep learning as described in claim 1, characterized in that, The rock temporal features based on each of the rock temporal images are used as input to form a rock fracture evolution prediction model. This model is then trained to predict the current rock fracture stage corresponding to the rock temporal image data in the current iteration, including: Based on the rock temporal features of each of the rock temporal images, the rock temporal enhancement features of each of the rock fracture evolution prediction models are obtained using the feature enhancement layer of the rock fracture evolution prediction model. Based on the temporal enhancement features of each of the rock temporal images, the inputs of multiple temporal convolutional networks in the rock fracture evolution prediction model are formed. Temporal feature tensors are extracted through the temporal convolutional networks to determine multiple temporal feature tensors, wherein different temporal convolutional networks correspond to different scales. Based on multiple temporal feature tensors, the input of the fusion network in the rock fracture evolution prediction model is formed, and the output is a fusion feature tensor; Based on the fused feature tensor, the input of the dual-pooling network in the rock fracture evolution prediction model is formed, and the output is an enhanced fused feature tensor. Based on the enhanced fusion feature tensor, the input to the prediction network in the rock fracture evolution prediction model is formed, and the current rock fracture stage corresponding to the rock time series image data is output. Based on the enhanced fusion feature tensor, the input to the regression network in the rock fracture evolution prediction model is formed to predict the fracture prediction time corresponding to the rock time series image data.

7. The rock fracture evolution prediction method based on deep learning as described in claim 6, characterized in that, The rock temporal features include crack planarity, crack linearity, angle between adjacent temporal vectors, crack change rate, and crack change acceleration, wherein the angle between adjacent temporal vectors is used to quantify the stability of the crack propagation direction in the rock temporal image; The process of obtaining enhanced rock time-series features for each of the rock time-series images using the feature enhancement layer of the rock fracture evolution prediction model, based on the rock time-series features of each of the rock time-series images, includes: For the rock temporal features of the rock temporal image, the crack planarity, the crack linearity, and the angle between adjacent temporal vectors are coupled to obtain coupled feature values; The rock temporal features and the coupled feature values ​​are formed into a vector, which is used as the rock temporal enhancement feature of the rock temporal image.

8. The rock fracture evolution prediction method based on deep learning as described in claim 6, characterized in that, The rupture label data includes rupture stage labels and rupture time labels; The process of training the rock fracture evolution prediction model based on the current rock fracture stage and the fracture label data includes: Calculate the current classification loss based on the current rock fracture stage and the fracture stage label; Based on the predicted rupture time and the rupture time label, calculate the regression prediction loss; Calculate the crack boundary change intensity of each of the rock time-series image data, and calculate the physical loss corresponding to the rock time-series image data based on the crack boundary change intensity of each of the rock time-series image data; Calculate the current loss value based on the current classification loss, the regression prediction loss, and the physical loss; Based on the current loss value, backpropagation is performed to train the rock fracture evolution prediction model until the training termination condition is met.

9. The rock fracture evolution prediction method based on deep learning as described in claim 8, characterized in that, The calculation of the physical loss corresponding to the rock time-series image data based on the crack boundary change intensity of each of the rock time-series images includes: Obtain the load-displacement curves corresponding to the time-series images of the rock; Based on the load-displacement curve, the external force work with the same time sequence as each of the rock time series images is obtained, and the external force work corresponding to each of the rock time series images is determined; The number of images in which the crack boundary change intensity of the rock time series image is greater than the external force work corresponding to the rock time series image is determined; The number of images is multiplied by a preset loss value to obtain the physical loss corresponding to the time-series rock image data.

10. A computing device, characterized in that, It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the deep learning-based rock fracture evolution prediction method as described in any one of claims 1 to 9.

Images