A deep learning-based complex fissure rock mass physical reconstruction method

Abstract

The application discloses a complex fissure rock mass physical reconstruction method based on deep learning, and belongs to the technical field of rock mass sample preparation. The method comprises the following steps: S1, quantifying rock mass characteristics by using mechanical performance target values and fissure structure characteristic vectors, and establishing a database; S2, constructing a multi-task deep neural network model based on an attention mechanism, and training the model; S3, determining the mechanical characteristics and fissure structure of a required rock sample, and using the trained multi-task deep neural network model based on the attention mechanism to obtain corresponding 3D printing process parameters; and according to the obtained 3D printing process parameters, a rock sample with the required mechanical characteristics and fissure structure is prepared by using a 3D printing method. The method balances the mechanical characteristic similarity and fissure structure fidelity, ensures high consistency between the prepared sample and the original rock mass in structure, and is repeatable in the preparation process, thereby providing a reliable sample for rock mechanics tests.

Description

Technical Field

[0001] This invention relates to the field of rock mass sample preparation technology, and more specifically to a method for physical reconstruction of complex fractured rock masses based on deep learning. Background Technology

[0002] In fields such as deep resource extraction and tunnel engineering, rock masses commonly exhibit complex distributions of joints, fissures, and other defect structures, which significantly control the mechanical properties and stability of the rock mass. Indoor physical testing is a key means of studying the mechanical behavior of fractured rock masses, but existing methods struggle to accurately, efficiently, and reproducibly prepare specimens or models that are highly similar to real rock masses in both mechanical properties (similar physical properties) and internal structure (structural consistency).

[0003] For example, processing existing samples using the original rock (such as waterjet cutting) can introduce secondary damage and make it difficult to create complex fracture networks; traditional similar material casting methods make it difficult to accurately control the geometry and spatial distribution of internal fractures. Furthermore, when preparing rock-like samples using 3D printing technology, their mechanical properties (such as strength and elastic modulus) are complexly affected by various factors such as printing materials, parameters, and post-processing techniques. The lack of a systematic "material-parameter-mechanical property" mapping model makes it difficult to achieve controllable strength preparation for the target lithology. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention proposes a method for physical reconstruction of complex fractured rock masses based on deep learning.

[0005] The technical solution adopted in this invention is:

[0006] A deep learning-based method for physical reconstruction of complex fractured rock masses includes the following steps:

[0007] S1. Quantify the characteristics of the rock mass and establish a database;

[0008] Rock mass properties are quantified using target values ​​of mechanical properties and characteristic vectors of fracture structure.

[0009] Establish a database including historical test data, with each record being {X, Y, Q}; where X is the input, Y and Q are the outputs, X is the 3D printing process parameters, Y is the measured mechanical properties of the specimen prepared under the 3D printing process parameters, and Q is the fidelity evaluation index of the crack structure of the specimen prepared under the 3D printing process parameters.

[0010] S2. Construct a multi-task deep neural network model based on the attention mechanism and train the model;

[0011] The attention-based multi-task deep neural network model includes:

[0012] Input layer, receiving 3D printing process parameters;

[0013] A shared feature extraction layer extracts shared low-level feature representations from 3D printing process parameters;

[0014] The attention mechanism layer introduces a multi-head attention mechanism after the shared feature extraction layer to automatically learn the influence weights of different process parameters on mechanical properties and the fidelity of crack structure.

[0015] After the attention mechanism layer, it splits into two independent branches: the mechanical performance prediction branch and the structural fidelity prediction branch. The mechanical performance prediction branch is used to predict mechanical parameters, while the structural fidelity prediction branch is used to predict the fidelity index of the fracture structure.

[0016] Finally, the prediction results are output through a linear layer;

[0017] The database established using S1 was used to train the constructed attention-based multi-task deep neural network model.

[0018] S3. Determine the mechanical properties and fracture structure of the rock sample to be prepared, and use the multi-task deep neural network model based on the attention mechanism trained in S2 to obtain the corresponding 3D printing process parameters; based on the obtained 3D printing process parameters, use 3D printing methods to prepare the rock sample with the required mechanical properties and fracture structure.

[0019] The beneficial technical effects of the present invention are as follows:

[0020] (1) Intelligentization and automation: The trial-and-error process that relies on human experience is transformed into data-driven intelligent decision-making, which greatly improves the efficiency and accuracy of physical reconstruction of complex fractured rock masses.

[0021] (2) "Property-Structure" Co-optimization: By using a multi-task deep learning model, the similarity of mechanical properties and the fidelity of the crack structure are taken into account at the same time, which solves the problem that traditional methods cannot balance the relationship between the two and achieves the true meaning of "similar property, controllable strength and consistent structure".

[0022] (3) High fidelity and repeatability: Based on real fracture data from CT scans and high-precision 3D printing, the prepared samples are structurally highly consistent with the original rock mass, and the preparation process is repeatable, providing reliable samples for rock mechanics tests.

[0023] (4) Adaptation and continuous evolution: Through the feedback mechanism of preparation results, the database is continuously enriched and the model is optimized, so that the system has the ability to learn and continuously improve, and can adapt to the preparation needs of different mining areas and different lithologies. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the deep learning-based physical reconstruction method for complex fractured rock masses according to the present invention.

[0025] Figure 2 This is a schematic diagram of the architecture of the multi-task deep neural network model based on the attention mechanism used in the method of the present invention;

[0026] Figure 3 This is a flowchart illustrating the training process of the multi-task deep neural network model based on the attention mechanism in the method of this invention. Detailed Implementation

[0027] Currently, determining the optimal 3D printing fabrication scheme (including material selection and parameter settings) based on specific rock masses (given mechanical properties and fracture structures) at an engineering site mainly relies on the experience of experimental personnel, which is cumbersome, inefficient, and difficult to guarantee accuracy. Therefore, this invention proposes a deep learning-based physical reconstruction method for complex fractured rock masses to solve the challenges of "precise control of mechanical properties" and "high-fidelity reproduction of fracture structures" in the physical reconstruction process, achieving intelligent, efficient, and accurate decision-making from engineering rock mass information to 3D printing fabrication process parameters. This invention can achieve intelligent and standardized physical reconstruction of fractured rock masses with "similar physical properties, controllable strength, and consistent structure."

[0028] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0029] like Figure 1 As shown, a deep learning-based physical reconstruction method for complex fractured rock masses includes the following steps:

[0030] S1. Goal definition and data preparation (establishing optimization goals);

[0031] This stage transforms the engineering problem into a quantifiable objective that can be handled by a mathematical model.

[0032] S11, Quantification of rock mass characteristics;

[0033] Using the target value of mechanical properties (Y) target ) and fracture structure eigenvectors (S target The properties of the rock mass are quantified. Target values ​​for mechanical properties are obtained directly from rock mass tests, such as uniaxial compressive strength (UCS) and elastic modulus (E). The fracture structure feature vectors are extracted from the three-dimensional model reconstructed from CT scans of the rock mass, such as fracture density (P0). 21 ) and average trace length (L aveThis vector is used for initial screening of the range of process parameters that can achieve high structural fidelity. For example, by calculating the Euclidean distance between the target fracture structure feature vector (Starget) and the fracture structure feature vectors (Ssample) of each sample in the historical database, the process parameters corresponding to the K samples with the smallest distance (K takes the value of 20~50, adjusted according to the database size) are selected as the initial search space for subsequent optimization. The formula is as follows:

[0034] ;

[0035] In the formula: m is the dimension of the fracture structure feature vector (such as fracture density, average trace length, etc.). Let i be the i-th eigenvalue of the target vector. Let be the i-th feature value of the sample vector.

[0036] S12, Historical Database Construction;

[0037] Establish a database that includes historical experimental data, with each record consisting of {X, Y, Q}; the data format is uniformly CSV to facilitate subsequent model reading and training.

[0038] X is the model input, which consists of 3D printing process parameters, such as sand powder type (silica sand, GS19 sand, quartz sand, etc.) and gradation (reduction coefficient n=0.3~0.5), binder type (furan resin and phenolic resin) and saturation (C). b =0%~100%), Printing layer thickness (T) l =0.2~0.4mm), type of reinforcing fiber (glass fiber and carbon fiber, etc.) and dosage (V f =0%~12%), post-treatment temperature (T) p =50~250℃) and blower drying time (t) p =1~4h), etc.

[0039] Y and Q are the model outputs, where Y is the measured mechanical property value of the sample prepared under this set of 3D printing process parameters (such as uniaxial compressive strength UCS). print Elastic modulus E print Q represents the fidelity evaluation index of the fracture structure of the sample prepared under the 3D printing process parameters (such as porosity error and fracture position deviation compared to the original CT model). The fidelity evaluation index of fracture structure includes a comprehensive score of fracture density error, average trace length deviation, porosity error, and fracture position deviation.

[0040] S2. Construct a multi-task deep neural network model based on the attention mechanism and train the model (learn the mapping relationship).

[0041] S21. The attention-based multi-task deep neural network model adopts an architecture design of "shared low-level features + task-specific branches", and the overall framework is as follows: Figure 2 As shown, it includes:

[0042] Input layer: Receives a 3D printing process parameter vector X, with dimension d. input Key parameters include sand powder type, binder type and saturation, printing layer thickness, reinforcing fiber type and dosage, post-processing temperature and drying time.

[0043] The 3D printing process parameters received by the input layer and the 3D printing process parameter vector X are the same concept, with vector dimension d. input If d equals the number of process parameters (e.g., sand powder type, binder saturation, etc., a total of 8 parameters), then d input =8). Clearly defining the vector dimension is used to define the number of neurons in the neural network input layer, ensuring a consistent model input format, and providing a basis for parameter initialization in the feature extraction layer.

[0044] The shared feature extraction layer consists of three fully connected layers, each containing 512 neurons. It uses the ReLU activation function and is responsible for extracting the shared low-level feature representation, i.e., the low-level shared features, from the original process parameters, i.e. the 3D printing process parameters received from the input layer. This provides the basic features for subsequent multi-task learning.

[0045] The attention mechanism layer introduces a multi-head attention mechanism after the shared feature layer. Through a query-key-value (QKV) calculation method, it automatically learns the weights of different process parameters on mechanical properties and structural fidelity. The specific calculation formula is as follows:

[0046] ;

[0047] In the formula: Q, K, and V are obtained from the shared features at the bottom layer through linear transformations. The specific steps are: through three independent linear transformation matrices ( , , The shared features (dimension d) output by the shared feature extraction layer are the underlying shared features. model These are mapped to query (Q), key (K), and value (V) vectors, respectively, as shown in the following formulas:

[0048] ;

[0049] ;

[0050] ;

[0051] In the formula: The dimension of the key vector. For value vector dimensions, typically h represents the number of multi-head attention heads.

[0052] The multi-head attention mechanism repeats the attention calculation h times (usually h=8), then concatenates the results and performs a linear transformation to obtain the final output. The attention mechanism layer enables the model to automatically focus on key process parameters, improving prediction accuracy and interpretability.

[0053] After the attention mechanism layer, it splits into two independent branches:

[0054] Mechanical property prediction branch: Predicting mechanical parameters such as uniaxial compressive strength (UCS) and elastic modulus (E);

[0055] Structural fidelity prediction branch: Predicts the structural fidelity index (Q) of fracture structures.

[0056] Each branch contains two fully connected layers with 256 and 128 neurons, respectively, and finally outputs the prediction results through a linear output layer. Among them, the mechanical property prediction branch outputs uniaxial compressive strength (UCS), elastic modulus (E), etc.; the structural fidelity prediction branch outputs the fracture structure fidelity index (Q).

[0057] S22. The database established in S12 is used to train the constructed multi-task deep neural network model based on the attention mechanism.

[0058] The model learns by optimizing the loss function during training, and the process is as follows: Figure 3 As shown, it includes the following steps:

[0059] S221, Data preprocessing;

[0060] Standardize the process parameters:

[0061] ;

[0062] In the formula: X is the original process parameter vector (i.e., 3D printing process parameters); μ is the mean of the process parameter vector; σ is the standard deviation of the process parameter vector; X norm This is the standardized process parameter vector.

[0063] Normalize the mechanical property labels.

[0064] The dataset in the database is divided into training set, validation set and test set in a ratio of 7:2:1.

[0065] S222, Model initialization;

[0066] Initialize network weights using the Xavier initialization method.

[0067] Set a learning rate decay strategy: the initial learning rate is 0.001, and it decays to 0.9 times the original rate every 50 epochs.

[0068] Using the Adam optimizer, β1=0.9, β2=0.999.

[0069] S223, Training Loop;

[0070] Forward propagation: Calculates predictions for two tasks using the training set data. Specifically, it involves normalizing the 3D printing process parameter vector (X) from the training set. norm The input model is processed by a shared feature extraction layer to extract common features from the underlying layers and an attention mechanism layer to enhance the influence of key parameters. These are then fed into the mechanical performance prediction branch and the structural fidelity prediction branch in parallel. The mechanical performance prediction branch uses a two-layer fully connected network to calculate and output continuous predicted values ​​such as uniaxial compressive strength and elastic modulus. The structure fidelity prediction branch calculates and outputs the structure fidelity prediction probability in the 0~1 interval through a 2-layer fully connected network. These two together constitute the two task prediction values ​​for forward propagation, which are used for subsequent loss calculation.

[0071] The formula for calculating the loss in the mechanical performance prediction task is as follows (mean square error):

[0072] ;

[0073] In the formula: N is the sample size; y i This represents the true mechanical property value of the i-th sample. Let be the predicted mechanical property value of the i-th sample.

[0074] The loss calculation formula for the structure fidelity prediction task is as follows (cross-entropy loss):

[0075] ;

[0076] In the formula: The true structural fidelity label (0 or 1) for the i-th sample. The predicted structure fidelity probability (between 0 and 1) is the probability of the i-th sample.

[0077] Fidelity label, setting fidelity threshold (Usually taken as 0.8), when the sample measured fidelity index When the value is 1, it is marked as 1 (high fidelity); otherwise, it is marked as 0 (low fidelity). Fidelity probability is the probability (value 0~1) that the sample output by the model belongs to the high fidelity category. It is used to calculate the cross-entropy loss and measures the difference between the predicted result and the true label.

[0078] The formula for calculating the total loss is as follows:

[0079] ;

[0080] In the formula: α and β are hyperparameters, with values ​​ranging from [0,1], used to balance the importance of the two tasks (i.e., the two prediction branches); The total loss of the model, For the task of predicting mechanical properties, To predict the loss of the structure fidelity prediction task.

[0081] Backpropagation: Calculates the gradient and updates the network parameters; the gradient here represents the total loss. The partial derivatives of the weight parameters of each layer of the network (i.e., the loss gradient) are calculated using the chain rule, starting from the output layer, sequentially calculating the gradient of the total loss with respect to the weights of the linear output layer, fully connected layer, attention mechanism layer, and shared feature extraction layer. Then, the Adam optimizer updates all parameters based on the gradients, as shown in the following formula:

[0082] ;

[0083] In the formula: These are the network weight parameters. The current learning rate, For the total loss The gradient.

[0084] Validation set evaluation: The model performance is evaluated on the validation set every 10 epochs (epoch is the training round, and 1 epoch means that all samples in the training set participate in one forward propagation and one back propagation).

[0085] Early stopping strategy: Stop training when the total loss of the validation set no longer decreases for 10 consecutive epochs.

[0086] Model saving:

[0087] Save the model parameters that achieve the best performance on the validation set, i.e., the total loss on the validation set. The minimum model weight parameter (independent of 3D printing process parameters) is saved to ensure optimal accuracy of subsequent inference; the loss curve and evaluation metrics during training are recorded to analyze overfitting / underfitting during model training and to provide a basis for hyperparameter adjustment (such as learning rate, α, β).

[0088] S23, Visualization of attention weights;

[0089] After training, the key process parameters with the greatest impact on mechanical properties and structural fidelity are identified by analyzing the attention weight matrix. The attention weight matrix is ​​generated during the calculation process of the attention mechanism layer. Each row corresponds to a process parameter, and each column corresponds to a sample. The matrix elements are the influence weights of the process parameter on the prediction results of that sample. The average weight of each process parameter in the attention mechanism is calculated, that is, each element in the 3D printing process parameter vector X (such as sand powder gradation, binder saturation, etc.). By calculating the average attention weight, key parameters can be identified (the higher the weight, the greater the impact on the prediction results).

[0090] ;

[0091] In the formula: The average attention weight for the i-th process parameter; Attention i,j is the attention weight of the i-th process parameter on the j-th sample; N is the number of samples.

[0092] S3. Determine the mechanical properties and fracture structure of the rock sample to be prepared, and use the multi-task deep neural network model based on the attention mechanism trained in S2 to obtain the corresponding 3D printing process parameters; based on the obtained 3D printing process parameters, use 3D printing methods to prepare the rock sample with the required mechanical properties and fracture structure.

[0093] In other words, once the model F(X) is trained, F(X) is the trained multi-task deep neural network model based on the attention mechanism, which contains two sub-models: the mechanical performance prediction sub-model F Y (X) and the structure fidelity prediction sub-model F Q (X), the core problem becomes: given the new target mechanical properties of the rock mass and the characteristic vector of the fracture structure, solve for the optimal 3D printing process parameters. In the constrained optimization stage, the target mechanical properties Y... target and the eigenvector S of the fracture structure target The optimization objective (rather than the model input) is the 3D printing process parameter X to be optimized, and the output is Y. pred =F Y (X) and Q pred =F Q (X), the optimization objective is to find the value that makes Y... pred Approaching Y target And Q pred ≥Q th X.

[0094] Constraint optimization is performed using the following formula:

[0095] ;

[0096] In the formula: It is the L2 norm squared (a measure of the difference between the predicted mechanical properties and the target value). For the fidelity threshold, This refers to the physically feasible range of process parameters (e.g., a printing layer thickness of 0.2~0.4mm).

[0097] Model evaluation metrics are used to quantify the accuracy of model predictions and ensure the reliability of optimized 3D printing process parameters. RMSE and R0 are used to measure the predicted mechanical properties. 2 Used to verify F Y The predictive accuracy (R) of (X) 2 (The closer to 1, the better; the smaller the RMSE, the better); the accuracy of structure fidelity prediction and the F1 score are used to validate the F... Q The classification accuracy of (X) (the higher the better) is only considered when the evaluation metrics of both tasks meet the preset threshold (e.g., R). 2 Only models with an F1 score ≥0.9 and an F1 score ≥0.85 are used for subsequent process parameter optimization.

[0098] Mechanical property prediction and evaluation:

[0099] Root mean square error:

[0100] ;

[0101] In the formula: R is the root mean square error (RMSE); N is the sample size; y i This represents the true mechanical property value of the i-th sample. Let be the predicted mechanical property value of the i-th sample.

[0102] Coefficient of determination:

[0103] ;

[0104] In the formula: This is the sum of squared residuals (prediction error); This is the total sum of squares (the variance of the true values).

[0105] Structural fidelity prediction and assessment:

[0106] Accuracy:

[0107] ;

[0108] In the formula:

[0109] TP is a true positive case, predicted as positive and actually positive; TN is a true negative case, predicted as negative and actually negative; FP is a false positive case, predicted as positive but actually negative; FN is a false negative case, predicted as negative but actually positive.

[0110] F1 score:

[0111] ;

[0112] In the formula:

[0113] Precision: TP / (TP+FP), the proportion of samples that were predicted to be positive but were actually positive;

[0114] Recall: TP / (TP+FN), the proportion of samples that are actually positive that are predicted to be positive.

[0115] The aforementioned deep learning-based physical reconstruction method for complex fractured rock masses also includes steps of physical verification and model evolution (closed-loop feedback):

[0116] The optimal 3D printing process parameters (decision result X) obtained are used to solve for the optimal 3D printing process parameters. opt The sample was used for 3D printing and the resulting finished product was measured to obtain the actual result Y. real and Q real ; Treat the actual results as new data for {X opt ,Y real Q real The model is incrementally learned by adding data to the historical database. This forms a self-optimizing closed-loop system, the process of which can be represented as:

[0117] ;

[0118] In the formula: To optimize the obtained optimal 3D printing process parameters, These are the measured mechanical property values ​​of the sample. This serves as the fidelity index for the measured structure of the sample; after new data pairs are added to the database, the model parameters are updated through incremental learning to obtain a new model F with higher accuracy. new (X).

[0119] Furthermore, the model optimization strategy is as follows:

[0120] Dynamic weight adjustment: α and β are dynamically adjusted based on the rate of change of loss for the two tasks; the rate of change of loss for the two tasks, i.e., the rate of decrease of mechanical performance loss and structural fidelity loss (not the loss itself), reflects the convergence speed of task training. α (t) and β (t) It is the weight in the total loss formula that changes dynamically with the training step size (replacing the fixed hyperparameter), so that the model focuses on the task with slower convergence in the early stage of training, and balances the loss of the two tasks in the later stage.

[0121] ;

[0122] In the formula: α (t)β represents the weight of the mechanical performance task at the t-th training step. (t) Let be the weights for the structure fidelity task at the t-th training step. η is a temperature parameter that controls the smoothness of the weight distribution; Let be the rate at which the mechanical property loss decreases at step t; Let be the rate at which the structural fidelity loss decreases at step t;

[0123] ;

[0124] ;

[0125] In the formula: The mechanical property loss at step t; This represents the loss of mechanical properties at step t-1. The structural fidelity loss at step t; This represents the structural fidelity loss at step t-1.

[0126] Gradient normalization: To prevent any one task from dominating the training process, the gradients for each task are normalized. Specifically, the loss gradients for the mechanical performance prediction branch and the structural fidelity prediction branch are normalized separately. For example, L2 gradient normalization normalizes the gradient vector for each task to the unit norm, as shown in the following formula:

[0127] ;

[0128] In the formula: This represents the original loss gradient for the task. To minimize the value (avoiding a denominator of 0), ensure that the gradient magnitudes of the two tasks are appropriate, prevent a single task from dominating parameter updates, and ensure that the contributions of each task to parameter updates are relatively balanced.

[0129] In summary, the method of this invention solves the problem of balancing the similarity of mechanical properties and the fidelity of fracture structure in traditional methods by integrating CT scanning, 3D printing and constructing a multi-task deep neural network model based on attention mechanism. It ensures a high degree of structural consistency between the prepared sample and the original rock mass, and the preparation process is repeatable, providing reliable samples for rock mechanics tests.

[0130] For any parts not mentioned above, existing technologies can be adopted or referenced.

[0131] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A method for physical reconstruction of complex fractured rock masses based on deep learning, characterized in that... Includes the following steps: S1. Quantify the characteristics of the rock mass and establish a database; Rock mass properties are quantified using target values ​​of mechanical properties and characteristic vectors of fracture structure. Establish a database including historical test data, with each record being {X, Y, Q}; where X is the 3D printing process parameters, which serve as input; Y is the measured mechanical properties of the specimen prepared under the 3D printing process parameters; Q is the fidelity evaluation index of the crack structure of the specimen prepared under the 3D printing process parameters; and Y and Q serve as outputs. S2. Construct a multi-task deep neural network model based on the attention mechanism and train the model; The attention-based multi-task deep neural network model includes: Input layer, receiving 3D printing process parameters; A shared feature extraction layer extracts shared low-level feature representations from 3D printing process parameters; The attention mechanism layer introduces a multi-head attention mechanism after the shared feature extraction layer to automatically learn the influence weights of different process parameters on mechanical properties and the fidelity of crack structure. After the attention mechanism layer, it splits into two independent branches: the mechanical performance prediction branch and the structural fidelity prediction branch. The mechanical performance prediction branch is used to predict mechanical parameters, while the structural fidelity prediction branch is used to predict the fidelity index of the fracture structure. Finally, the prediction results are output through a linear layer; The database established using S1 was used to train the constructed attention-based multi-task deep neural network model. S3. Determine the mechanical properties and fracture structure of the rock sample to be prepared, and use the multi-task deep neural network model based on the attention mechanism trained in S2 to obtain the corresponding 3D printing process parameters; based on the obtained 3D printing process parameters, use 3D printing methods to prepare the rock sample with the required mechanical properties and fracture structure.

2. The method for physical reconstruction of complex fractured rock masses based on deep learning according to claim 1, characterized in that, In S1: the target mechanical properties are obtained directly from rock mass tests, including uniaxial compressive strength and elastic modulus; the fracture structure feature vector is extracted from the three-dimensional model reconstructed from the rock mass by CT scan, including fracture density and average trace length.

3. The method for physical reconstruction of complex fractured rock masses based on deep learning according to claim 2, characterized in that: The 3D printing process parameters include sand powder type and gradation, binder type and saturation, printing layer thickness, reinforcing fiber type and dosage, post-processing temperature and drying time. The fidelity evaluation index of the fracture structure includes the porosity error and fracture location deviation between the fracture structure of the sample prepared under the 3D printing process parameters and the three-dimensional model reconstructed from the CT scan rock mass.

4. The method for physical reconstruction of complex fractured rock masses based on deep learning according to claim 1, characterized in that, In S2: The shared feature extraction layer consists of 3 fully connected layers, each containing 512 neurons. The ReLU activation function is used to extract the shared low-level feature representation from the 3D printing process parameters to obtain the low-level shared features. The attention mechanism layer automatically learns the influence weights of different process parameters on mechanical properties and structural fidelity through a query-key-value calculation method. The specific calculation formula is as follows: ； In the formula: Q, K, and V are obtained from the shared features at the bottom layer through linear transformation. The dimension of the key vector; the multi-head attention mechanism repeats the attention calculation multiple times, then concatenates the results and performs a linear transformation to obtain the final output; Both the mechanical property prediction branch and the structural fidelity prediction branch contain two fully connected layers with 256 and 128 neurons respectively, and finally output the prediction results through a linear output layer.

5. The method for physical reconstruction of complex fractured rock masses based on deep learning according to claim 1, characterized in that, In S2, the steps for training the model are as follows: S21. Data preprocessing; Standardize 3D printing process parameters: ； In the formula: X is the original process parameter vector; μ is the mean of the process parameter vector; σ is the standard deviation of the process parameter vector; X norm This is the standardized process parameter vector; S22, Model initialization; S23, Training Loop; The formula for calculating total loss is as follows: ； In the formula: α and β are hyperparameters, with values ​​ranging from [0,1], used to balance the importance of the two tasks; The total loss of the model, For the task of predicting mechanical properties, To predict the loss of the structural fidelity prediction task; The formula for calculating the loss in the mechanical performance prediction task is as follows: ； In the formula: N is the sample size; y i This represents the true mechanical property value of the i-th sample. Let be the predicted mechanical property value of the i-th sample; The formula for calculating the loss in the structure fidelity prediction task is as follows: ； In the formula: The true structural fidelity label for the i-th sample; Let be the predicted structure fidelity probability of the i-th sample.

6. The method for physical reconstruction of complex fractured rock mass based on deep learning according to claim 5, characterized in that: After training, the key process parameters that have the greatest impact on mechanical properties and structural fidelity are identified by analyzing the attention weight matrix; the average weight of each process parameter in the attention mechanism is calculated. ； In the formula: The average attention weight for the i-th process parameter; Attention i,j The attention weight of the i-th process parameter on the j-th sample; N is the number of samples.

7. The method for physical reconstruction of complex fractured rock masses based on deep learning according to claim 1, characterized in that, In S3: Once the model training is complete, given new mechanical performance target values ​​and crack structure feature vectors, the optimal 3D printing process parameters are solved. Constraint optimization is performed using the following formula: ； In the formula: The square of the L2 norm, For the fidelity threshold, This refers to the physically feasible range of process parameters; The target mechanical properties Y target and the eigenvector S of the fracture structure target The model input is the 3D printing process parameter X to be optimized, and the output is Y. pred =F Y (X) and Q pred =F Q (X), the optimization objective is to find the value that makes Y... pred Approaching Y target And Q pred ≥Q th The 3D printing process parameters X.

8. The method for physical reconstruction of complex fractured rock masses based on deep learning according to claim 7, characterized in that, S3 also includes the following steps: The optimal 3D printing process parameters obtained are applied to 3D printing fabrication, and the resulting sample is measured to obtain actual results. These actual results are then added to the historical database as new data to incrementally learn the model, forming a self-optimizing closed-loop system. The process is represented as follows: ； In the formula: X opt For optimal 3D printing process parameters, Y real Q represents the measured mechanical property value. real This is a measure of the fidelity of the measured structure. The model is then updated with an expanded database, allowing it to continuously evolve.

9. The method for physical reconstruction of complex fractured rock masses based on deep learning according to claim 8, characterized in that, S3 also includes: α is dynamically adjusted based on the rate of decrease of mechanical property loss and structural fidelity loss. (t) and β (t) ; ； In the formula: α (t) β represents the weight of the mechanical performance task at the t-th training step. (t) Let be the weights for the structure fidelity task at the t-th training step. η is a temperature parameter that controls the smoothness of the weight distribution; Let be the rate of decrease of mechanical property loss at step t; Let be the rate at which the structural fidelity loss decreases at step t; ； ； In the formula: The mechanical property loss at step t; This represents the loss of mechanical properties at step t-1. The structural fidelity loss at step t; This represents the structural fidelity loss at step t-1.

Images