A method and system for predicting dynamic propagation of a fracture surface in an engineered rock mass

Abstract

The application discloses an engineering rock mass fracture surface dynamic expansion prediction method and system, and the prediction method comprises the following steps: obtaining original acoustic wave data, performing positioning calculation and source mechanism inversion on the original acoustic wave data, obtaining a coal rock breaking source core parameter at each time and calculating a breaking through possibility index BCI. The fracture similarity is calculated through multi-dimensional characteristics, density clustering is performed by using a DBSCAN clustering algorithm, and a reconstructed fracture surface time sequence image is obtained by combining BCI optimization and B-spline surface fitting technology. The fracture surface image space-time data set is obtained through preprocessing and is divided, an improved SimVP space-time neural network is used to construct a fracture surface image sequence prediction model for training, a multi-head attention mechanism and a local residual connection are introduced into an encoder and a decoder, a Poolformer network is used as a translator, and finally the trained prediction model is used to accurately predict the future time fracture surface dynamic expansion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fracture prediction, in particular to an engineering rock mass fracture surface dynamic expansion prediction method and system. BACKGROUND

[0002] As an important engineering measure, hydraulic measures have the advantages of pressure relief, permeability improvement, and modification, and have been widely used in deep oil and gas resource exploitation, hot dry rock development and utilization, carbon storage layer modification, etc. In the implementation process of hydraulic measures, the expansion characteristics and spatial distribution of fractures are related to the implementation effect and engineering safety of the measures. By accurately predicting the dynamic expansion process of the fractures, the recoverability of the reservoir can be effectively predicted, the hydraulic fracturing measures can be optimized, and the risk of fracturing-induced earthquakes can be reduced, thereby improving the economic efficiency and safety of green resource development.

[0003] Traditional fracture expansion prediction methods rely on physical models and empirical formulas, which often assume that fracture expansion follows certain fixed rules, making it difficult to fully capture the complexity of the underground environment and the dynamic changes of fracture expansion. Therefore, as an advanced data-driven technology, deep learning has gradually become a research hotspot in the field of fracture expansion prediction. Patent CN119228748A provides a rock mass fracture prediction method, device and storage medium, which optimizes data continuity by constructing a data set and using an image interpolation algorithm, and uses a spatio-temporal prediction network to capture the spatial and temporal dependence between feature maps through a self-attention mechanism, achieving accurate capture of dynamic changes in fractures. Patent CN118506104A discloses a rock fracture identification and expansion prediction method and system, which inputs the processed CT scan slice images into a SAM segmentation model for rock fracture identification, and establishes a graph convolutional neural network to analyze the connection between fractures and predict the fracture expansion path. Patent CN117557874A proposes an intelligent prediction method for coal rock fatigue deformation fracture expansion, which uses CT images of coal rock patterns as time series data sets to construct a coal rock fatigue fracture expansion regression model to predict the development and expansion of pore fractures and the deformation of coal rock in real time. Although the above methods have made significant progress in fracture expansion prediction, they still face the challenge of processing high-dimensional and complex spatio-temporal data. In this context, SimVP spatio-temporal neural network, as a video prediction network based entirely on CNN architecture, can better handle the complexity of spatio-temporal data with its simple and efficient design, providing a new technical path for dynamic prediction of fracture expansion. SUMMARY

[0004] The present application proposes an engineering rock mass fracture surface dynamic expansion prediction method and system to address the problems and needs mentioned above. The technical features adopted enable the realization of the above technical purposes and bring about other technical effects.

[0005] The application provides an engineering rock mass crack face dynamic expansion prediction method.

[0006] S10: original acoustic wave data is acquired and positioning calculation and a focal mechanism inversion are performed, to obtain a coal rock rupture source core parameter directly related to rupture at each moment, and a rupture breakthrough possibility index BCI of a crack at each moment is calculated in combination with a standardized crack volume, a focal radius and a spatial distance;

[0007] S20: similarity between cracks at each moment is calculated according to multi-dimensional characteristics of the cracks, a DBSCAN clustering algorithm is used to perform density clustering on the cracks, a potential crack face with high breakthrough and expansion potential is identified in combination with the rupture breakthrough possibility index BCI, a B-spline surface fitting algorithm is applied to perform smooth fitting on the crack face, and a reconstructed crack face image is obtained;

[0008] S30: the reconstructed crack face image is arranged in time sequence to form a crack face time sequence image; after pre-processing, a crack face image space-time data set is constructed, and the crack face image space-time data set is divided into a training set, a verification set and a test set;

[0009] S40: an improved SimVP space-time neural network is used to construct a crack face image sequence prediction model;

[0010] S50: the training set is input into the crack face image sequence prediction model for training, the verification set is used to test and evaluate the trained prediction model, parameters are continuously adjusted and trained, after the training is completed, crack face sequence images of the test set are input into the trained prediction network, to obtain predicted crack face image sequences, and corresponding evaluation indexes are used to systematically evaluate the prediction results.

[0011] In addition, the engineering rock mass crack face dynamic expansion prediction method according to the application can also have the following technical features.

[0012] In one example of the application, in the step S10, original acoustic wave data is subjected to positioning calculation and focal mechanism inversion, to obtain a focal mechanism core parameter directly related to rupture at each moment, and a rupture breakthrough possibility index BCI of a crack at each moment is calculated, including the following steps.

[0013] S11: original acoustic wave data collected by a microseismic monitoring system is post-processed, and a first arrival time and amplitude information extraction is performed by using the Akaike information criterion method;

[0014] S12: acoustic wave first arrival time data at different position sensors of the microseismic monitoring system are used, and a simplex positioning robust algorithm is used to perform spatial positioning calculation on an acoustic emission source;

[0015] S13: Based on the shear fracture source model constraint condition, the amplitude data of the first wave of the sound wave collected by the sensors at different positions are used to solve the 6 components M of the source rupture moment tensor pq ;

[0016] S14: According to the calculated moment tensor components, the core parameters of the coal rock fracture source are quantitatively inverted and calculated, wherein the core parameters of the coal rock fracture source include the spatial orientation of the crack, the crack volume and the crack energy;

[0017] S15: The critical distance D of the fracture penetration is determined according to the source radius of the crack point i , the data are converted into a standard normal distribution with zero mean and unit variance by the Z-score standardization method;

[0018] S16: When the adjacent two crack points i and j meet the fracture penetration condition, the fracture penetration possibility index BCI between the adjacent two crack points is calculated ij , and the total fracture penetration index BCI of the crack point i is calculated i .

[0019] In one example of the present application, in the step S20, the DBSCAN algorithm is used for density clustering of the cracks, including the following steps:

[0020] S211: Multidimensional feature vector reading and Euclidean distance calculation: the multidimensional feature vector of each crack point obtained by the source mechanism inversion at each time is read, and the Euclidean distance between the crack points is calculated;

[0021] S212: Determining the neighborhood distance and the minimum point number: selecting appropriate neighborhood distance ∈ and minimum point number MinPts, based on the Euclidean distance and neighborhood distance ∈, the neighborhood of each crack point is calculated; if the number of points contained in the neighborhood of a crack point is greater than or equal to MinPts, the crack point is a core point; otherwise, it is a boundary point or a noise point;

[0022] S213: Cluster expansion and core point processing: selecting one from the unmarked core points as an initial point, marking it as the starting point of a new cluster, and assigning a cluster number to the point; then, expanding the neighborhood of the core point, and putting all the points in the neighborhood into the same cluster; for the newly added points in the expanded cluster, check whether they are core points; if they are core points, continue to expand the cluster; if they are boundary points, keep their cluster labels unchanged; continue to select new points from the unmarked core points for cluster expansion until all the core points are processed;

[0023] S214: Cluster number and cluster result check: according to the result of DBSCAN clustering, a unique crack surface number is assigned to each cluster, and each cluster represents an independent crack surface; if a certain fracture point belongs to the neighborhood of multiple core points at the same time, the point will be assigned to the crack surface number of the first discovered cluster; then check the clustering result to ensure that each cluster is assigned a unique crack surface number, and the number of noise points is-1; if it is found that the cluster number assignment is wrong or missing, necessary adjustments are made to ensure that the final crack surface number is accurate.

[0024] In an example of the present application, in the step S20, the potential fracture surface with higher penetration and expansion potential is identified by combining the fracture penetration possibility index BCI optimization, and a B-spline surface fitting algorithm is applied to smooth fit the fracture surface to obtain a reconstructed fracture surface image, including the following steps:

[0025] S221: Optimize and screen fractures based on BCI value: for each fracture cluster, calculate the fracture penetration possibility index BCI of each fracture, and further visualize the BCI value distribution of the fracture, set a suitable threshold, and select the fracture points with higher BCI value as potential fracture surfaces;

[0026] S222: Reconstruct fracture surface time series image: read the fracture data obtained by clustering analysis at all times, combine the spatial coordinates, BCI value, volume and normal direction feature information of each fracture at each time, and perform fitting processing on each fracture at each time; through B-spline curve and triangular mesh division method, the fracture surface at each time is accurately reconstructed, and finally the fracture surface reconstruction results at all times are integrated to generate a complete fracture surface time series image.

[0027] In an example of the present application, in the step S30, the data is preprocessed and the data set is divided, including:

[0028] The fracture penetration possibility index of each fracture point on the fracture surface obtained by clustering fitting is converted into a corresponding gray value, and is interpolated into a uniform 128x128x128 voxel grid, and the interpolated image is normalized to construct a fracture surface image spatiotemporal data set. For example, 80% of the data set is used as the training set, 10% is used as the validation set, and 10% is used as the test set.

[0029] In an example of the present application, in the step S40, the fracture surface image sequence prediction model includes a spatial encoder, a translator and a spatial decoder,

[0030] The spatial encoder is configured to map the input high-dimensional fracture graph data to a low-dimensional latent space, capture spatial features in three-dimensional fracture graph data using three-dimensional convolution operations, and introduce a multi-head attention mechanism before convolution to capture spatiotemporal dependencies;

[0031] The translator is configured to use a Poolformer network as a prediction network to learn the spatial dependence and temporal variation of the fracture face image sequence from the latent space, stack the output tensor of the encoder along the time axis direction, and then use the Poolformer to learn from the stacked multi-frame features to capture the inherent temporal evolution law of the fracture face image sequence data.

[0032] The spatial decoder is configured to decode the information in the latent space into predicted future fracture face images, uses a three-dimensional deconvolution operation to gradually restore the details of the image, and introduces a multi-head attention mechanism before deconvolution to capture the spatio-temporal dependence of the decoder features.

[0033] In an example of the present application, the spatial encoder comprises:

[0034] A first multi-head attention mechanism module is configured to map the input fracture map feature to different feature space queries Q, keys K and values V, respectively package them into matrices Q1, K1 and V1, and perform linear transformation on Q1, K1 and V1 through different weight matrices to obtain Q1', K1' and V1'; then calculate the attention value of each attention head independently; finally, the outputs of all heads are spliced and subjected to a final linear transformation to obtain the final multi-head attention output.

[0035] A first local residual connection module is arranged at the input and output ends of the first multi-head attention mechanism module, and is configured to maintain the transmission of spatial information.

[0036] A three-dimensional convolution module comprises a plurality of convolution layers connected in series, and each convolution layer comprises convolution, normalization operation and nonlinear transformation. The three-dimensional convolution module is configured to perform feature extraction on the output z'0 of the first multi-head attention mechanism module after the first local residual connection module, and performs downsampling operation on every two convolution layers.

[0037] In an example of the present application, the spatial decoder comprises:

[0038] A second multi-head attention mechanism module is configured to map the output tensor w0 of the translator to different feature space queries Q, keys K and values V, respectively package them into matrices Q2, K2 and V2, and perform linear transformation on Q2, K2 and V2 through different weight matrices to obtain Q2', K2' and V2'; then calculate the attention value of each attention head independently; finally, the outputs of all heads are spliced and subjected to a final linear transformation to obtain the final multi-head attention output.

[0039] a second local residual connection module, the second local residual connection module being arranged at the input and output ends of the second multi-head attention mechanism module, and configured to maintain the transmission of spatial information;

[0040] a three-dimensional deconvolution module, including a plurality of serially connected deconvolution layers, each of the deconvolution layers including deconvolution, normalization operation and nonlinear transformation, and the three-dimensional deconvolution module being configured to extract features and restore spatial resolution of w'0 rows output by the second multi-head attention mechanism module after the second local residual connection module, wherein an upsampling operation is performed once by PixelShuffle for every two deconvolution layers.

[0041] In an example of the present application, the step S50 specifically comprises the following steps:

[0042] S51: training the fracture surface image sequence prediction model by using a training set, adjusting batchsize, learning rate, optimizer and epoch number, and adopting L2 loss function and motion perception loss function L m A total loss function L is constructed to supervise the training of the model;

[0043] S52: inputting a test set into the trained fracture surface image sequence prediction model to obtain a predicted fracture surface image sequence, and evaluating the prediction result by using MSE, SSIM and PSNR evaluation indexes, wherein the evaluation indexes are as follows:

[0044]

[0045] In the formula, Y i and respectively represent the true value and the predicted value of the i-th time step, μ Y and respectively represent the average value of Y i and , σ Y and respectively represent the standard deviation of Y i and , and represents the covariance of Y i and , c1 and c2 are constants, and I MAX represents the maximum pixel value of the image.

[0046] In an example of the present application, in the step S52, the test set is inputted into the trained fracture surface image sequence prediction model to obtain a predicted fracture surface image sequence, and the step specifically comprises the following steps:

[0047] inputting a fracture surface image sequence X∈RT×C×H×W×D where T represents a time step, C represents a number of channels of each image, H, W, D represent height, width and depth of each image respectively,

[0048] A group of dynamic sequences is formed according to the changes X of T time steps, and a crack surface image sequence prediction predicts T' frames in the future according to the given T frames of images, and the modeling formula is as follows:

[0049]

[0050] In the formula: X T×C×H×W×D is an input crack surface image sequence, is an output predicted crack surface image sequence, Y T′×C×H×W×D is a real crack surface image sequence, and Θ represents a series of parameters required to be optimized;

[0051] Another object of the present application is to provide an engineering rock mass crack surface dynamic expansion prediction system, comprising:

[0052] A parameter acquisition and calculation module is configured to obtain original acoustic wave data and perform positioning calculation and source mechanism inversion on the original acoustic wave data, to obtain a coal rock rupture source core parameter directly related to rupture at each time, and to calculate a rupture breakthrough possibility index BCI of each crack at each time in combination with a standardized crack volume, a source radius and a spatial distance;

[0053] A crack surface image reconstruction module is configured to calculate the similarity between cracks at each time according to the multi-dimensional characteristics of the cracks, and to perform density clustering on the cracks by using a DBSCAN clustering algorithm, to identify potential crack surfaces with high breakthrough and expansion potential in combination with the rupture breakthrough possibility index BCI, to perform smooth fitting on the crack surfaces by using a B-spline surface fitting algorithm, and to obtain a reconstructed crack surface image;

[0054] A data set processing and division module is configured to arrange the reconstructed crack surface images in chronological order to form a crack surface time sequence image; after preprocessing, a crack surface image space-time data set is constructed, and the crack surface image space-time data set is divided into a training set, a validation set and a test set;

[0055] A prediction model construction module is configured to construct a crack surface image sequence prediction model by using an improved SimVP space-time neural network;

[0056] A crack surface image prediction module is configured to input the training set into the crack surface image sequence prediction model for training, to use the validation set to test and evaluate the trained prediction model, and to continuously adjust parameters for training; after the training is completed, the crack surface sequence images of the test set are input into the trained prediction network to obtain a predicted crack surface image sequence, and the prediction results are systematically evaluated by using corresponding evaluation indexes.

[0057] The technical scheme has the following beneficial effects relative to the prior art:

[0058] The present application can accurately capture the source characteristics of the fracture by microseismic monitoring inversion to obtain the core parameters of the source mechanism, and use the same as the input data of the DBSCAN clustering algorithm, thereby significantly improving the accuracy and reliability of the reconstruction of the fracture surface time sequence image, and providing more scientific technical support for further research on the dynamic expansion prediction of the fracture.

[0059] Unlike the traditional time sequence prediction model, the SimVP network used in the present application is a video prediction network based entirely on the CNN architecture, without complex cycle structure or additional modules, thereby avoiding the introduction of too many skills and strategies, and achieving efficient prediction of the fracture expansion process while keeping the network structure simple.

[0060] In the present application, the multi-head attention mechanism is introduced in the encoder and decoder when constructing the fracture surface image sequence prediction model, thereby effectively extracting the high-dimensional features of the fracture surface image sequence data, making the model learn the time sequence correlation while keeping the spatial information of the sequence, and thereby improving the accuracy and reliability of the fracture surface prediction.

[0061] The present application fully integrates the advantages of multiple technologies by the comprehensive application of the source mechanism inversion, the DBSCAN clustering algorithm and the spatio-temporal neural network prediction model, and effectively captures the spatio-temporal correlation in the fracture expansion process. The prediction model can deeply mine the time sequence features and spatial distribution patterns in the fracture expansion process, thereby significantly improving the prediction accuracy of the fracture surface dynamic evolution, and providing a basis for optimizing the hydraulic fracturing design and implementation.

[0062] The optimal embodiments of the present application will be described in more detail below with reference to the accompanying drawings, so that the features and advantages of the present application can be easily understood. BRIEF DESCRIPTION OF DRAWINGS

[0063] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings of the embodiments of the present application will be briefly introduced below. The drawings are only used to show some embodiments of the present application, and not to limit all embodiments of the present application to this.

[0064] Figure 1 The flowchart of the engineering rock mass fracture surface dynamic expansion prediction method according to the embodiment of the present application;

[0065] Figure 2 The acoustic data positioning inversion calculation flowchart according to the embodiment of the present application;

[0066] Figure 3 The DBSCAN clustering flowchart according to the embodiment of the present application;

[0067] Figure 4 A flowchart of a crack surface image sequence prediction model according to an embodiment of the present application;

[0068] Figure 5 A structural diagram of a crack surface image sequence prediction model according to an embodiment of the present application;

[0069] Figure 6 A structural diagram of a multi-head attention mechanism module according to an embodiment of the present application. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical scheme and advantages of the technical scheme of the present application more clear, the technical scheme of the embodiments of the present application will be described clearly and completely below in combination with the drawings of the embodiments of the present application. The same reference signs in the drawings represent the same components. It should be noted that the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the described embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0071] Unless otherwise defined, technical terms or scientific terms used herein should be understood as having the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terms "first", "second", and similar terms used in the description and the claims of the present patent application do not necessarily mean any order, number, or importance, but are only used to distinguish different components. Similarly, the terms "one" or "a" and the like do not necessarily mean a quantity limitation. The terms "include" or "contain" and the like mean that the elements or objects before the term encompass the elements or objects listed after the term and their equivalents, without excluding other elements or objects. The terms "connect" or "connected" and the like do not necessarily mean a physical or mechanical connection, but can include an electrical connection, whether direct or indirect. The terms "up", "down", "left", "right", and the like are only used to represent relative positional relationships, and when the absolute positions of the described objects change, the relative positional relationships can also change accordingly.

[0072] According to the first aspect of the present application, a method for predicting dynamic expansion of a crack surface of an engineering rock mass, as shown in Figure 1 , comprises the following steps:

[0073] S10, as shown in Figure 2 , by extracting the original acoustic data acquired by the microseismic monitoring system and performing positioning calculation and source mechanism inversion, the coal and rock rupture source core parameters directly related to each time are obtained, and the crack rupture through possibility index BCI at each time is calculated in combination with the standardized crack volume, source radius and spatial distance.

[0074] S20: Calculate the similarity between each time crack according to the multi-dimensional characteristics of crack such as spatial distance, orientation, volume and energy, and use DBSCAN clustering algorithm to cluster the density of cracks, combine with the crack breakthrough possibility index BCI to optimize and identify the potential crack surface with high breakthrough and expansion potential, apply B-spline surface fitting algorithm to smooth the crack surface, and get the reconstructed crack surface image;

[0075] S30: Arrange the reconstructed crack surface image in time sequence to form a crack surface time sequence image; after pretreatment, construct the crack surface image spatio-temporal dataset, and divide it into training set, validation set and test set;

[0076] S40: Use the improved SimVP spatio-temporal neural network to construct the crack surface image sequence prediction model, including introducing multi-head self-attention mechanism and local residual connection in the encoder and decoder, and replacing the translator with Poolformer network;

[0077] S50: Input the training set into the crack surface image sequence prediction model for training, use the validation set to test and evaluate the trained prediction model, and constantly adjust the parameters for training; after training, input the crack surface sequence image of the test set into the trained prediction network to obtain the predicted crack surface image sequence, and use the corresponding evaluation index to systematically evaluate the prediction results.

[0078] The prediction method can accurately capture the source characteristics of the crack by obtaining the core parameters of the seismic source mechanism through microseismic monitoring and inversion, and use them as input data for the DBSCAN clustering algorithm, which significantly improves the accuracy and reliability of the crack surface time sequence image reconstruction, and provides more scientific technical support for further research on the dynamic expansion prediction of the crack.

[0079] Unlike traditional time series prediction models, the SimVP network used in this prediction method is a video prediction network based entirely on CNN architecture, without complex loop structures or additional modules, avoiding the introduction of too many techniques and strategies, thus achieving efficient prediction of the crack expansion process while keeping the network structure simple.

[0080] When constructing the crack surface image sequence prediction model, the prediction method introduces multi-head attention mechanism in the encoder and decoder, effectively extracting the high-dimensional features of the crack surface image sequence data, so that the model can learn the temporal correlation while maintaining the spatial information of the sequence, thereby improving the accuracy and reliability of the crack surface prediction.

[0081] The prediction method fully integrates the advantages of multiple technologies by the comprehensive application of the focal mechanism inversion, the DBSCAN clustering algorithm and the spatio-temporal neural network prediction model, effectively captures the spatio-temporal correlation in the crack propagation process. Moreover, the prediction model can deeply mine the time series features and spatial distribution patterns in the crack propagation process, significantly improves the prediction accuracy of the crack surface dynamic evolution, and provides a basis for optimizing the hydraulic fracturing design and implementation.

[0082] In one example of the present application, in the step S10, the original acoustic wave data is subjected to positioning calculation and focal mechanism inversion to obtain the focal mechanism core parameters directly related to the rupture at each time, and the crack rupture breakthrough possibility index BCI at each time is calculated, including the following steps:

[0083] S11: The original acoustic wave data collected by the microseismic monitoring system is subjected to post-processing, and the Akaike information criterion method is used for first arrival time and amplitude information extraction;

[0084] S12: The acoustic wave first arrival time data at different positions of the sensor of the microseismic monitoring system is used to calculate the spatial positioning of the acoustic emission source by using the simplex positioning robust algorithm;

[0085] S13: Based on the constraint conditions of the tensile-shear rupture source model, the 6 components M pq of the rupture moment tensor are solved from the acoustic wave first arrival amplitude data collected by different position sensors;

[0086] S14: The core parameters of the coal and rock rupture source are quantitatively inverted and calculated according to the calculated moment tensor components, wherein the core parameters of the coal and rock rupture source include crack spatial orientation, crack volume and crack energy;

[0087] S15: The rupture breakthrough critical distance D i is determined according to the focal radius of the crack point to determine whether the crack points have a rupture breakthrough relationship. The data is converted into a standard normal distribution with zero mean and unit variance by the Z-score standardization method to standardize the volume of the crack point;

[0088] S16: When the adjacent two crack points i and j satisfy the rupture breakthrough condition, the rupture breakthrough possibility index BCI ij between the adjacent two crack points is calculated, and the total rupture breakthrough index BCI i of the crack point i is calculated, and the formula is as follows:

[0089]

[0090] In the formula, Z(V i ) and Z(V j ) are the standardized volume values of the i-th crack point and the j-th crack point, respectively; dij is the Euclidean distance between the ith crack point and the jth crack point; I ij is an indicator function, which is 1 when d ij ≤D i , and 0 otherwise, indicating whether the two points satisfy the crack-through condition.

[0091] In one example of the present application, in the step S20, the DBSCAN algorithm is used to perform density clustering on the cracks, including the following steps: Figure 3

[0092] S211: Multidimensional feature vector reading and Euclidean distance calculation: read the multidimensional feature vector (spatial coordinates, spatial orientation, volume and energy) of each crack point obtained by the focal mechanism inversion at each time, and calculate the Euclidean distance between the crack points; wherein, for each pair of crack points P i = (x i , y i , z i , nx i , ny i , nz i , V i , E i ) and P j = (x j , y j , z j , nx j , ny j , nz j , V j , E j ), the Euclidean distance calculation formula is as follows:

[0093]

[0094] In the formula: n is the dimension of the feature vector, P i,k and P i,k are the values of the kth dimension in the multidimensional feature vectors of the two crack points, w k is the weight of each feature (which can be adjusted according to the importance of different features);

[0095] S212: Determine the neighborhood distance and the minimum number of points: select appropriate neighborhood distance ∈ and minimum number of points MinPts, ∈ determines the neighborhood radius of a crack point in the Euclidean space, and MinPts is the minimum number of points that need to be contained in the neighborhood, which is commonly used to determine whether a point is a core point. Based on the Euclidean distance and the neighborhood distance ∈, the neighborhood of each crack point is calculated; if the number of points contained in the neighborhood of a crack point is greater than or equal to MinPts, the crack point is a core point; otherwise, it is a boundary point or a noise point;

[0096] ​S213: Cluster expansion and core point processing: select one from the unmarked core points as an initial point, mark it as the starting point of a new cluster, and assign a cluster number to the point; then, expand the neighborhood of the core point, and assign all points in the neighborhood (including core points and density-reachable boundary points) to the same cluster; for the newly added points in the expanded cluster, check whether they are core points; if they are core points, continue to expand the cluster; if they are boundary points, keep their cluster labels unchanged; continue to select new points from unmarked core points for cluster expansion until all core points are processed;

[0097] That is, as shown in Figure 3 , the core point needs to determine whether each core point has a label and whether all core points have been read at the same time, if the core point has no label, the neighborhood needs to be expanded and the label needs to be set, otherwise the next core point is read, if there is a point that has not been read among all core points, it is determined whether the core point has a label, otherwise, all crack clusters are generated; wherein, after the core point is expanded and the label is set, it is further determined whether the newly added points in the neighborhood are core points, if they are core points, the expansion and setting of the same label as the initial core point are continued, and it is further determined whether the newly added points in the neighborhood are core points; otherwise, the label is kept unchanged, and a new crack cluster is generated to form all crack clusters.

[0098] S214: Cluster number and cluster result checking: according to the DBSCAN clustering result, a unique crack surface number is assigned to each cluster, and each cluster represents an independent crack surface; if a certain crack point belongs to the neighborhood of multiple core points (i.e. the point is a density-reachable point), the point will be assigned to the crack surface number of the first discovered cluster; then, the clustering result is checked to ensure that each cluster is assigned a unique crack surface number, and the number of noise points is -1; if it is found that the cluster number assignment is incorrect or missing, necessary adjustments are made to ensure that the final crack surface number is accurate.

[0099] In an example of the present application, in the step S20, the potential crack surface with higher penetration and expansion potential is identified by combining the fracture penetration possibility index BCI optimization, and a B-spline surface fitting algorithm is applied to smooth fit the crack surface to obtain a reconstructed crack surface image, including the following steps:

[0100] S221: Optimize and select cracks based on BCI value: for each crack cluster, calculate the fracture penetration possibility index BCI of each crack, and further visualize the BCI value distribution of the crack, set a suitable threshold, and select the crack points with higher BCI value as potential crack surfaces;

[0101] S222: Reconstructing the crack surface time series image: read the crack data obtained by cluster analysis at all times, combine the spatial coordinates, BCI values, volumes and normal direction feature information of the cracks at each time, and perform fitting processing on each crack at each time; through B-spline curve and triangular mesh division method, the crack surface at each time is accurately reconstructed, and finally the crack surface reconstruction results at all times are integrated to generate a complete crack surface time series image.

[0102] In one example of the present application, in the step S30, the data is preprocessed and the data set is divided, including:

[0103] The crack penetration possibility index of each crack point on the crack surface obtained by cluster fitting is converted into a corresponding gray value, and is interpolated into a uniform 128x128x128 voxel grid. The interpolated image is normalized to construct a crack surface image spatiotemporal data set. 80% of the data set is used as the training set, 10% is used as the validation set, and 10% is used as the test set.

[0104] In one example of the present application, in the step S40, as shown in Figure 4 and Figure 5 The crack surface image sequence prediction model includes a spatial encoder, a translator and a spatial decoder,

[0105] The spatial encoder is configured to map the input high-dimensional crack graph data to a low-dimensional latent space, capture the spatial features in the three-dimensional crack graph data using three-dimensional convolution operation, and introduce a multi-head attention mechanism before convolution to capture the spatiotemporal dependence relationship;

[0106] The translator is configured to use a Poolformer network as a translator of the prediction network, to learn the spatial dependence relationship and time variation of the crack surface image sequence from the latent space, take the output tensor of the encoder as the input of the Poolformer, then stack multiple sequence frames along the time axis direction, and then use the Poolformer to learn from the stacked multiple frame features, capture the inherent time evolution law of the crack surface image sequence data, and finally take the output result as the input of the decoder;

[0107] The spatial decoder is configured to decode the information in the latent space into the predicted future crack surface image, uses a three-dimensional deconvolution operation to gradually restore the details of the image, and introduces a multi-head attention mechanism before deconvolution to capture the spatiotemporal dependence relationship of the decoder features.

[0108] In one example of the present application, the spatial encoder includes:

[0109] The first multi-head attention mechanism module, as Figure 6As shown, the input crack map original features are mapped to different feature space queries Q, keys K and values V, and are packaged into matrices Q1, K1 and V1, respectively. Q1, K1 and V1 are linearly transformed by different weight matrices to obtain Q1', K1' and V1'. Assuming there are h attention heads, each attention head has its own transformation matrix of Q, K and V. For the i-th attention head, the transformation formulas of Q, K and V are as follows:

[0110] Q′ i =QW Qi

[0111] K′ i =KW Ki

[0112] V i ′=VW Vi

[0113] (Q′ i ,K′ i ,V i ′)=(QW Qi ,KW Ki ,VW Vi )

[0114] In the formula, W Qi , W Ki and W Vi are weight matrices of the i-th attention head.

[0115] Then, the attention value of each attention head is calculated independently, and the formula is as follows:

[0116]

[0117] In the formula, head i is the attention value of the i-th attention head, Attention represents the attention function, softmax represents the activation function, d k is a scaling factor of the vector dimension to avoid the problem of gradient disappearance caused by too large values.

[0118] Finally, the outputs of all heads are spliced and linearly transformed to obtain the final multi-head attention output, and the formula is as follows:

[0119]

[0120] In the formula: is the output tensor calculated by the multi-head attention mechanism, MultiHead represents the multi-head attention output, Concat represents splicing, and W O is the linear transformation matrix of the output.

[0121] The first local residual connection module is arranged at the input and output ends of the first multi-head attention mechanism module, configured to maintain the transmission of spatial information, and the formula is as follows:

[0122]

[0123] In the formula, z'0 is a feature tensor obtained through residual connection, z0 is an initial feature tensor input into the multi-head attention mechanism, is an output tensor calculated through the multi-head attention mechanism;

[0124] The three-dimensional convolution module includes a plurality of convolution layers connected in series, and each convolution layer includes convolution, normalization operation and nonlinear transformation. The three-dimensional convolution module is configured to perform feature extraction on z'0 output by the first multi-head attention mechanism module after the first local residual connection module, and a downsampling operation is performed once for each two convolution layers.

[0125] That is, in order to input z'0 output by the first multi-head attention mechanism module through residual connection into a plurality of convolution layers for feature extraction, 6 convolution layers connected in series, normalization operation and nonlinear transformation are included, and the formula is as follows:

[0126] z i =SiLU(Norm3d(Conv3d(z i-1 )))1≤i≤6

[0127] In the formula, SiLU is a linear logic function for nonlinear transformation, Norm3d is a three-dimensional normalization layer, Conv3d is a three-dimensional convolution operation, z i represents a tensor input into the convolution layer, and the stride size of the convolution layer is set to 1.

[0128] A downsampling operation is performed once for each two convolution layers, and the stride size is set to 2, and the formula is as follows:

[0129] z i =MaxPool3d(z i ,stride=2)

[0130] In the formula, MaxPool3d represents a three-dimensional maximum pooling operation.

[0131] In an example of the present application, the spatial decoder includes:

[0132] The second multi-head attention mechanism module, as Figure 6As shown, the output tensor w0 of the translator is configured to be mapped into different feature space queries Q, keys K and values V, which are packed into matrices Q2, K2 and V2 respectively, and Q2, K2 and V2 are linearly transformed by different weight matrices to obtain Q2', K2' and V2'. Assuming there are h attention heads, each attention head has its own transformation matrix of Q, K and V. For the i-th attention head, the transformation formulas of Q, K and V are as follows:

[0133] Q i ′=QW Qi

[0134] K i ′=KW Ki

[0135] V i ′=VW Vi

[0136] (Q′ i ,K′ i ,V i ′)=(QW Qi ,KW Ki ,VW Vi )

[0137] In the formula, W Qi , W Ki and W Vi are weight matrices of the i-th attention head;

[0138] Then, for each attention head, its attention value is independently calculated, and the formula is as follows:

[0139]

[0140] In the formula, head i is the attention value of the i-th attention head, Attention represents the attention function, softmax represents the activation function, and d k is a scaling factor of the vector dimension to avoid the problem of gradient disappearance caused by too large values;

[0141] Finally, the outputs of all heads are spliced and linearly transformed to obtain the final multi-head attention output, and the formula is as follows:

[0142]

[0143] In the formula: is the final output tensor calculated by the multi-head attention mechanism, MultiHead represents the multi-head attention output, Concat represents splicing, and W O is the linear transformation matrix of the output.

[0144] A second local residual connection module is arranged at the input and output ends of the second multi-head attention mechanism module, configured to maintain the transmission of spatial information, and the formula is as follows:

[0145]

[0146] In the formula, w'0 is a feature tensor obtained through residual connection, w0 is a translator output tensor input into the multi-head attention mechanism, is a final output tensor calculated through the multi-head attention mechanism;

[0147] A three-dimensional deconvolution module includes a plurality of serially connected deconvolution layers, and each deconvolution layer includes deconvolution, normalization operation and nonlinear transformation. The three-dimensional deconvolution module is configured to extract features and restore spatial resolution of w'0 row features output by the second multi-head attention mechanism module through the second local residual connection module, and the up-sampling operation is performed once on each two deconvolution layers by using PixelShuffle.

[0148] That is, in order to transmit the w'0 output by the second multi-head attention mechanism module with residual connection into a plurality of deconvolution layers for feature extraction and spatial resolution restoration, six serially connected deconvolution, normalization operation and nonlinear transformation are included, and the formula is as follows:

[0149] w k =SiLU(Norm3d(UnConv3d(w k-1 )))1≤k≤6

[0150] In the formula, SiLU is a linear logistic function for nonlinear transformation, Norm3d is a three-dimensional normalization layer, UnConv3d is a three-dimensional deconvolution operation, w i represents a tensor input into the deconvolution layer, and the stride size of the deconvolution layer is set to 1.

[0151] The up-sampling operation is performed once on each two deconvolution layers by using PixelShuffle, and the stride size is set to 2, and the formula is as follows:

[0152] w k =PixelShuffle3d(w k ,stride=2)

[0153] In one example of the present application, the step S50 specifically includes the following steps:

[0154] S51: Train the crack surface image sequence prediction model through the training set, adjust the batch size, learning rate, optimizer and epoch number, and adopt L2 loss function and motion perception loss function L m The total loss function L is constructed to supervise the training of the model, and the expression of the loss function is as follows:

[0155]

[0156] L = λ1L2 + λ2L m

[0157] In the formula, N is the length of the prediction sequence, Y i and respectively represent the true value and the predicted value of the i-th time step, COF represents the function of calculating the dense optical flow to calculate the crack motion change between adjacent two frames, λ1 and λ2 are respectively the L2 loss coefficient and the motion perception loss coefficient;

[0158] S52: input the test set into the trained crack surface image sequence prediction model to obtain the predicted crack surface image sequence, and evaluate the prediction result using the MSE, SSIM and PSNR evaluation indexes, and the evaluation indexes are as follows:

[0159]

[0160] In the formula, Y i and respectively represent the true value and the predicted value of the i-th time step, μ Y and respectively represent the average value of Y i and , σ Y and respectively represent the standard deviation of Y i and , and represents the covariance of Y i and , c1 and c2 are constants, and I MAX represents the maximum pixel value of the image.

[0161] In one example of the present application, in the step S52, the test set is input into the trained crack surface image sequence prediction model to obtain the predicted crack surface image sequence, which specifically includes the following steps:

[0162] The crack surface image sequence X ∈ R T×C×H×W×D is input into the prediction model, where T represents the time step, C represents the channel number of each image, H, W and D respectively represent the height, width and depth of each image,

[0163] A set of dynamic sequences is formed according to the changes X of T time steps, and the crack surface image sequence prediction predicts T' frames in the future according to the given T frames of images, and the modeling formula is as follows:

[0164]

[0165] In the formula, X T×C×H×W×D is the input crack surface image sequence, is the output predicted crack surface image sequence, Y T′×C×H×W×D is the real crack surface image sequence, and Θ represents a series of parameters to be optimized;

[0166] The engineering rock mass crack surface dynamic expansion prediction system according to the second aspect of the present application comprises:

[0167] A parameter acquisition and calculation module is configured to extract original acoustic wave data acquired by a microseismic monitoring system, and to perform positioning calculation and source mechanism inversion on the original acoustic wave data, to obtain a coal rock rupture source core parameter directly related to rupture at each time, and to calculate a rupture breakthrough possibility index BCI of the crack at each time in combination with a standardized crack volume, a source radius and a spatial distance;

[0168] A crack surface image reconstruction module is configured to calculate the similarity between cracks at each time according to multi-dimensional features such as spatial distance, orientation, volume and energy of the cracks, and to perform density clustering on the cracks by using a DBSCAN clustering algorithm, to identify potential crack surfaces with high breakthrough and expansion potential in combination with the rupture breakthrough possibility index BCI, to perform smooth fitting on the crack surfaces by using a B-spline surface fitting algorithm, and to obtain reconstructed crack surface images;

[0169] A data set processing and division module is configured to arrange the reconstructed crack surface images in chronological order to form crack surface time sequence images; after preprocessing, a crack surface image spatiotemporal data set is constructed, and the crack surface image spatiotemporal data set is divided into a training set, a validation set and a test set;

[0170] A prediction model construction module is configured to construct a crack surface image sequence prediction model by using an improved SimVP spatiotemporal neural network, including introducing a multi-head self-attention mechanism and a local residual connection in an encoder and a decoder, and replacing a translator with a Poolformer network;

[0171] A crack surface image prediction module is configured to input the training set into the crack surface image sequence prediction model for training, to use the validation set to test and evaluate the trained prediction model, and to continuously adjust parameters for training; after the training is completed, crack surface sequence images of the test set are input into the trained prediction network to obtain predicted crack surface image sequences, and corresponding evaluation indexes are used to systematically evaluate the prediction results.

[0172] The prediction system obtains the core parameters of the focal mechanism by microseismic monitoring inversion, can accurately capture the source characteristics of the fracture, and takes the source characteristics as the input data of the DBSCAN clustering algorithm, thereby significantly improving the accuracy and reliability of the reconstruction of the fracture surface time sequence image, and providing more scientific technical support for further research on the dynamic expansion prediction of the fracture.

[0173] Unlike traditional time series prediction models, the SimVP network used by the prediction system is a video prediction network based entirely on the CNN architecture, without complex loop structures or additional modules, avoiding the introduction of too many skills and strategies, thereby achieving efficient prediction of the fracture expansion process while keeping the network structure simple.

[0174] When constructing the fracture surface image sequence prediction model, the prediction system introduces a multi-head attention mechanism in the encoder and decoder, effectively extracts the high-dimensional features of the fracture surface image sequence data, and enables the model to learn the temporal correlation while maintaining the spatial information of the sequence, thereby improving the accuracy and reliability of the fracture surface prediction.

[0175] The prediction system fully integrates the advantages of multiple technologies by comprehensively applying the focal mechanism inversion, DBSCAN clustering algorithm, and spatio-temporal neural network prediction model, effectively capturing the spatio-temporal correlation in the fracture expansion process. The prediction model can deeply mine the time series features and spatial distribution patterns in the fracture expansion process, significantly improving the prediction accuracy of the dynamic evolution of the fracture surface, and providing a basis for optimizing the design and implementation of hydraulic fracturing.

[0176] The above describes the exemplary embodiments of the engineering rock mass fracture surface dynamic expansion prediction method and system proposed by the present application in detail with reference to the preferred embodiments, however, those skilled in the art can understand that various modifications and improvements can be made to the above specific embodiments without departing from the concept of the present application, and various technical features and structures proposed by the present application can be combined without exceeding the protection scope of the present application, and the protection scope of the present application is determined by the appended claims.

Claims

1. A method for predicting dynamic propagation of a fracture surface in an engineered rock mass, characterized by, Comprise the following steps: S10: Obtain original acoustic wave data and perform positioning calculation and source mechanism inversion to obtain coal rock fracture source core parameters directly related to each time point of the rupture, including crack spatial orientation, crack volume and crack energy, and calculate the crack fracture breakthrough possibility index BCI of each time point of the crack in combination with the standardized crack volume, source radius and spatial distance; wherein, the calculation of the crack fracture breakthrough possibility index BCI of each time point of the crack includes: determining the fracture breakthrough critical distance D i according to the source radius of the crack point, converting the data into a standard normal distribution with zero mean and unit variance by a Z-score standardization method; when the adjacent two crack points i and j meet the fracture breakthrough condition, the fracture breakthrough possibility index BCI ij between the adjacent two crack points is calculated, and the total fracture breakthrough index BCI i of the crack point i is calculated. S20: According to the multi-dimensional characteristics of the crack, the similarity between the cracks at each time is calculated, and the DBSCAN clustering algorithm is used for density clustering of the cracks, combined with the crack breakthrough possibility index BCI to optimize the identification of potential crack surfaces with high connectivity and expansion potential, and a B-spline surface fitting algorithm is used to smooth the crack surface and obtain the reconstructed crack surface image; S30: The reconstructed crack surface image is arranged in time sequence to form a crack surface time sequence image; after pretreatment, a crack surface image spatio-temporal dataset is constructed, and it is divided into a training set, a validation set and a test set; S40: An improved SimVP spatio-temporal neural network is used to construct a crack surface image sequence prediction model; S50: The training set is input into the crack surface image sequence prediction model for training, the validation set is used to test and evaluate the trained prediction model, and the parameters are adjusted for training; after training, the crack surface sequence image of the test set is input into the trained prediction network to obtain the predicted crack surface image sequence, and the prediction results are systematically evaluated using the corresponding evaluation index.

2. The engineering rock mass crack surface dynamic expansion prediction method according to claim 1, characterized in that, In the step S10, the original acoustic wave data is calculated and the source mechanism inversion is performed to obtain the source mechanism core parameters directly related to the rupture at each time, and the crack rupture breakthrough possibility index BCI at each time is calculated, comprising the following steps: S11: The original acoustic wave data collected by the microseismic monitoring system is post-processed, and the Akaike information criterion method is used to extract the first arrival time and amplitude information; S12: The acoustic wave first arrival time data at different positions of the sensor are used to calculate the spatial location of the acoustic emission source by using the simplex positioning robust algorithm; S13: Based on the shear crack source model constraint condition, the amplitude data of the first wave of the sound wave collected by the sensors at different positions are used to solve the 6 components M of the source rupture moment tensor pq ; S14: According to the calculated moment tensor components, the coal and rock rupture source core parameters are quantitatively inverted and calculated, wherein the coal and rock rupture source core parameters include crack spatial orientation, crack volume and crack energy; S15: determining the rupture through critical distance D according to the focal radius of the crack point i The data were transformed into a standard normal distribution with zero mean and unit variance by the Z-score standardization method. S16: When the two adjacent fracture points i and j satisfy the fracture breakthrough condition, calculate the fracture breakthrough possibility index BCI between the two adjacent fracture points ij , and thereby calculate the total fracture breakthrough index BCI of the fracture point i i .

3. The engineering rock mass crack surface dynamic expansion prediction method according to claim 1, characterized in that, In the step S20, the DBSCAN algorithm is used for density clustering of the cracks, comprising the following steps: S211: Multi-dimensional feature vector reading and Euclidean distance calculation: read the multi-dimensional feature vector of each crack point obtained by source mechanism inversion at each time, and calculate the Euclidean distance between the crack points; S212: Determine the neighborhood distance and the minimum number of points: select appropriate neighborhood distance ∈ and minimum number of points MinPts, based on the Euclidean distance and neighborhood distance ∈, calculate the neighborhood of each crack point; If the number of points contained in the neighborhood of a crack point is greater than or equal to MinPts, the crack point is a core point; otherwise, it is a boundary point or a noise point; S213: Cluster expansion and core point processing: select one from the unmarked core points as the initial point, mark it as the starting point of a new cluster, and assign a cluster number to it; Subsequently, the neighborhood of the core point is expanded, and all points in the neighborhood are classified into the same cluster; for the newly added points in the expanded cluster, it is checked whether they are core points; if they are core points, the cluster is continuously expanded; if they are boundary points, their cluster labels remain unchanged; continue to select new points from the unmarked core points for cluster expansion until all core points are processed; S214: Cluster number and clustering result checking: according to the clustering result of DBSCAN, a unique crack surface number is assigned to each cluster, and each cluster represents an independent crack surface; if a certain crack point belongs to the neighborhood of multiple core points at the same time, the point will be assigned to the crack surface number of the first discovered cluster; then check the clustering result to ensure that each cluster is assigned a unique crack surface number, and the number of noise points is-1; if it is found that the cluster number assignment is wrong or missing, necessary adjustments are made to ensure that the final crack surface number is accurate.

4. The engineering rock mass crack surface dynamic expansion prediction method according to claim 1, characterized in that, In the step S20, the potential crack surface with high connectivity and expansion potential is identified by combining the fracture breakthrough possibility index BCI optimization, and the B-spline surface fitting algorithm is applied for smooth fitting of the crack surface to obtain a reconstructed crack surface image, including the following steps: S221: Optimize and select cracks based on BCI value: For each crack cluster, calculate the fracture breakthrough possibility index BCI of each crack, and further visualize the BCI value distribution of the crack, set a suitable threshold, and select the crack points with higher BCI value as potential crack surfaces; S222: Reconstruct crack surface time sequence image: read the crack data obtained by clustering analysis at all times, combine the spatial coordinates, BCI value, volume and normal direction feature information of each crack at each time, and perform fitting processing on each crack at each time; through B-spline curve and triangular mesh division method, the crack surface at each time is accurately reconstructed, and finally the crack surface reconstruction results at all times are integrated to generate a complete crack surface time sequence image.

5. The engineering rock mass crack surface dynamic expansion prediction method according to claim 1, characterized in that, In the step S40, the crack surface image sequence prediction model includes a spatial encoder, a translator and a spatial decoder, The spatial encoder is configured to map the input high-dimensional crack graph data to a low-dimensional latent space, capture the spatial features in the three-dimensional crack graph data using three-dimensional convolution operation, and introduce a multi-head attention mechanism before convolution to capture the spatio-temporal dependence relationship; The translator is configured to use a Poolformer network as the translator of the prediction network to learn the spatial dependence relationship and time variation of the crack surface image sequence from the latent space, take the output tensor of the encoder as the input of the Poolformer, then stack multiple sequence frames along the time axis direction, and then use the Poolformer to learn from the stacked multiple frame features to capture the inherent time evolution law of the crack surface image sequence data, and finally take the output result as the input of the decoder; The spatial decoder is configured to decode information in the latent space into a predicted future fracture surface image, gradually restores details of the image using a three-dimensional deconvolution operation, and introduces a multi-head attention mechanism before deconvolution to capture the spatio-temporal dependence of decoder features.

6. The method according to claim 5, wherein, The spatial encoder comprises: A first multi-head attention mechanism module is configured to map input fracture map original features into different feature space queries Q, keys K and values V, and package them into matrices Q1, K1 and V1 respectively, and perform linear transformation on Q1, K1 and V1 through different weight matrices to obtain Q1', K1' and V1'; then, for each attention head, the attention value is independently calculated; finally, the outputs of all heads are spliced and subjected to a final linear transformation to obtain the final multi-head attention output. A first local residual connection module is arranged at the input and output ends of the first multi-head attention mechanism module, and is configured to maintain the transmission of spatial information. The three-dimensional convolution module comprises a plurality of convolution layers connected in series, and each convolution layer comprises convolution, normalization operation and nonlinear transformation. The three-dimensional convolution module is configured to perform feature extraction on the output z'0 of the first multi-head attention mechanism module after the first local residual connection module, and performs downsampling operation once for each two convolution layers.

7. The method according to claim 5, wherein, The spatial decoder comprises: A second multi-head attention mechanism module is configured to map the output tensor w0 of the translator into different feature space queries Q, keys K and values V, and package them into matrices Q2, K2 and V2 respectively, and perform linear transformation on Q2, K2 and V2 through different weight matrices to obtain Q2', K2' and V2'; then, for each attention head, the attention value is independently calculated; finally, the outputs of all heads are spliced and subjected to a final linear transformation to obtain the final multi-head attention output. A second local residual connection module is arranged at the input and output ends of the second multi-head attention mechanism module, and is configured to maintain the transmission of spatial information. The three-dimensional deconvolution module comprises a plurality of deconvolution layers connected in series, and each deconvolution layer comprises deconvolution, normalization operation and nonlinear transformation. The three-dimensional deconvolution module is configured to perform feature extraction and spatial resolution recovery on the output w'0 of the second multi-head attention mechanism module after the second local residual connection module, and performs upsampling operation once for each two deconvolution layers using PixelShuffle.

8. The method according to claim 1, wherein, The step S50 specifically comprises the following steps: S51: Train the crack surface image sequence prediction model through the training set, adjust the batch size, learning rate, optimizer and epoch number, and use L2 loss function and motion perception loss function L m The total loss function L is constructed to supervise the training of the model. S52: inputting the test set into the trained fracture surface image sequence prediction model to obtain a predicted fracture surface image sequence, and evaluating the prediction result using MSE, SSIM and PSNR evaluation indexes, wherein the evaluation indexes are as follows: Where Y i and Represent the true value and predicted value of the i-th time step, μ Y and Represents Y i and The average value, σ Y and Represents Y i and The standard deviation of Indicates Y i and covariance, c1 and c2 are constants, I MAX Indicates the maximum pixel value of the image.

9. The method of claim 8, wherein, In the step S52, the test set is input into the trained crack face image sequence prediction model to obtain the predicted crack face image sequence, specifically including the following steps: Input a sequence of fracture surface images X e R T×C×H×W×D where T represents the time step, C represents the number of channels of each image, H, W, D represent the height, width and depth of each image respectively; A set of dynamic sequences is formed according to the changes X of T time steps, and the crack face image sequence prediction predicts T' frames in the future according to the given T frames, and the modeling formula is as follows: where X T×C×H×W×D is the input sequence of slit face images, is the output predicted sequence of slit face images, Y T′×C×H×W×D is the real sequence of slit face images, and Θ represents a set of parameters to be optimized.

10. An engineering rock mass fracture plane dynamic propagation prediction system, characterized in that, comprising: The parameter acquisition and calculation module is configured to obtain original acoustic wave data and perform positioning calculation and source mechanism inversion on the original acoustic wave data to obtain coal rock rupture source core parameters directly related to rupture at each time, the coal rock rupture source core parameters including crack spatial orientation, crack volume and crack energy, and to calculate a crack rupture breakthrough possibility index BCI of each crack at each time in combination with standardized crack volume, source radius and spatial distance; wherein, the crack rupture breakthrough possibility index BCI of each crack at each time includes: determining a rupture breakthrough critical distance D i by the source radius of the crack point, converting the data into a standard normal distribution with zero mean and unit variance by a Z-score standardization method, calculating a crack rupture breakthrough possibility index BCI ij between the two adjacent crack points i and j when the two adjacent crack points i and j satisfy the rupture breakthrough condition, and calculating a total crack rupture breakthrough index BCI i of the crack point i. a crack face image reconstruction module configured to calculate the similarity between cracks at each time according to the multi-dimensional features of the cracks, and to perform density clustering on the cracks using a DBSCAN clustering algorithm, and to identify potential crack faces with high connectivity and expansion potential by combining a breakthrough possibility index BCI, and to apply a B-spline surface fitting algorithm to smooth fit the crack faces to obtain a reconstructed crack face image; a data set processing and division module configured to arrange the reconstructed crack face images in chronological order to form a crack face time sequence image, and to construct a crack face image spatio-temporal data set after preprocessing, and to divide the data set into a training set, a validation set and a test set; a prediction model construction module configured to construct a crack face image sequence prediction model using an improved SimVP spatio-temporal neural network; a crack face image prediction module configured to input the training set into the crack face image sequence prediction model for training, to use the validation set to test and evaluate the trained prediction model, and to continuously adjust the parameters for training; after the training is completed, the crack face sequence image of the test set is input into the trained prediction network to obtain the predicted crack face image sequence, and the prediction results are systematically evaluated using the corresponding evaluation indicators.

Images