Coal rock material mesoscopic modeling and multi-scale simulation method based on active deep learning

Abstract

The invention relates to a coal rock material mesoscopic modeling and multi-scale simulation method based on active deep learning. The method comprises the following steps: firstly, drawing up a complete label-free data sample space, and establishing a labeled data sample set based on the complete label-free data sample space; constructing a multi-scale analysis offline proxy model based on a deep learning model, and training the multi-scale analysis offline proxy model based on the labeled data sample set to obtain a basic predictor; inputting an unlabeled data sample into the basic predictor to obtain a sample false label, and determining an unlabeled data information measure by adopting an improved active learning strategy based on the sample false label; expanding the annotated data sample set based on the unannotated data information measurement, and performing iterative training to obtain a final offline agent meta-model; and finally, embedding the final offline agent element model into a finite element, and obtaining a strain-stress response under coal rock loading multi-scale simulation. And the multi-scale simulation speed is obviously improved.

Description

Technical Field

[0001] This application relates to the technical field of mesoscopic modeling and numerical calculation of coal and rock materials, and particularly to a mesoscopic modeling and multi-scale simulation method of coal and rock materials based on active deep learning. Background Art

[0002] As a typical underground engineering medium, the mechanical properties and damage failure laws of coal and rock are directly affected by the mesoscopic particle state and contact conditions in coal and rock masses. Accurately evaluating the mechanical response of coal and rock masses is crucial for ensuring the safe mining of coal mines.

[0003] Under such complex conditions, numerical simulation technology (finite element) has become the core means to evaluate the mechanical response of materials, and the constitutive model directly determines the reliability and accuracy of the simulation. However, it is very difficult to establish an accurate phenomenological coal and rock constitutive model. Strict mathematical derivations and rich basic theoretical knowledge limit the development of traditional models. The corresponding discrete element method can reproduce the mechanical properties of rock masses from the mesoscopic structure without complex model derivations and necessary assumptions, and it is a reliable numerical simulation technology. However, the huge amount of computing resources required is often unacceptable. On this basis, multi-scale simulation provides the possibility to balance this phenomenon. It analyzes the mechanical behavior of complex materials or structures by coupling the microscopic and macroscopic scales. This method divides the network at the macroscopic scale and calculates variables such as the displacement and stress of nodes; a microscopic discrete element model (generally called a representative volume element) is established at the integration point of each macroscopic element to capture the mesoscopic structural behavior of the materials. The stress and other response quantities obtained by solving the decoupled equations of the microscopic model are transmitted back to the macroscopic model. However, in practical applications, the computing power required for such a large-scale microscopic discrete element model calculation is still extremely large, which limits the development and application of multi-scale simulation.

[0004] Machine learning, especially deep learning, provides a simpler and faster way for multi-scale numerical simulation as an offline surrogate model. As the core content, the construction of an offline deep learning surrogate model inevitably requires a large amount of data, and it is usually difficult to obtain this data through experiments. Discrete simulation can directly reproduce the macroscopic mechanical properties of materials from mesoscopic contacts, which provides the possibility to collect large-scale mechanical data of materials. However, the stress state of underground coal and rock masses is complex, and even using a single unit model, it is still very difficult to fully realize these complex stress states.

[0005] Therefore, in the related technologies, there is an urgent need for a method that can achieve low-cost and high-precision multi-scale simulation based on as little high-cost data as possible and improve the simulation accuracy and efficiency. Summary of the Invention

[0006] Based on this, it is necessary to provide a mesoscopic modeling and multi-scale simulation method based on active deep learning for the above technical problems, which can achieve low-cost and high-precision multi-scale simulation based on as little high-cost data as possible and improve the simulation accuracy and efficiency.

[0007] In a first aspect, the present application provides a mesoscopic modeling and multi-scale simulation method for coal and rock materials based on active deep learning. The method includes:

[0008] Formulate a complete unlabeled data sample space, and establish a labeled data sample set based on the complete unlabeled data sample space;

[0009] Construct a multi-scale analysis offline proxy model based on a deep learning model, and train the multi-scale analysis offline proxy model based on the labeled data sample set to obtain a basic predictor;

[0010] Input unlabeled data samples into the basic predictor to obtain sample pseudo-labels, and determine the unlabeled data information metric based on the sample pseudo-labels using an improved active learning strategy, where the improved active learning strategy includes a distance-weighted diversity metric query strategy for the input space and a data difference metric query strategy based on gradients;

[0011] Expand the labeled data sample set based on the unlabeled data information metric and perform iterative training to obtain a final offline proxy meta-model;

[0012] Embed the final offline proxy meta-model into the finite element to obtain the strain-stress response under multi-scale simulation of coal and rock loading.

[0013] Optionally, in an embodiment of the present application, the establishing a labeled data sample set based on the complete unlabeled data sample space includes:

[0014] Perform numerical calibration on the discrete element to clarify the variables included in the representative volume element;

[0015] Formulate a complete unlabeled data sample space, use the representative volume element to simulate typical loading paths, obtain material offline data, and establish a labeled data sample set.

[0016] Optionally, in an embodiment of the present application, the training the multi-scale analysis offline proxy model based on the labeled data sample set includes:

[0017] Use the gradient change method to determine the training balance judgment condition of the multi-scale analysis offline proxy model.

[0018] Optionally, in an embodiment of the present application, the iterative gradient norm calculation formula is:

[0019]

[0020] Among them, is the gradient norm, L is the loss function, and ω i are the weights of the model.

[0021] Optionally, in an embodiment of the present application, the determining of the unlabeled data information metric based on the sample pseudo-label by using an improved active learning strategy includes:

[0022] Calculating the input data difference metric to evaluate data diversity;

[0023] Determining the samples to be labeled by evaluating the magnitude or change of the gradient.

[0024] Optionally, in an embodiment of the present application, the calculation formula for the diversity metric of the input space is:

[0025]

[0026] where x i , ε j are the strain data in the input space, x i are the data in the unlabeled sample space, is the unlabeled sample space, ε j are the labeled sample data, D i is the labeled sample space, and Cosine Similarity(x i , ε j ) is the cosine similarity calculation.

[0027] Optionally, in an embodiment of the present application, the calculation formula for the data difference metric based on the gradient is:

[0028]

[0029] where is the gradient norm of the root mean square error of the basic predictor, is the maximum gradient norm.

[0030] In a second aspect, the present application further provides a mesoscopic modeling and multi-scale simulation device for coal and rock materials based on active deep learning. The device includes:

[0031] A sample set establishment module, configured to formulate a complete unlabeled data sample space and establish a labeled data sample set based on the complete unlabeled data sample space;

[0032] Multi-scale analysis offline proxy model construction and initial training module, which is used to construct a multi-scale analysis offline proxy model based on a deep learning model, and train the multi-scale analysis offline proxy model based on the labeled data sample set to obtain a basic predictor;

[0033] Data difference measurement module, which is used to input unlabeled data samples into the basic predictor to obtain sample pseudo-labels, and determine the information measurement of unlabeled data by adopting an improved active learning strategy based on the sample pseudo-labels. The improved active learning strategy includes a distance-weighted diversity measurement query strategy in the input space and a data difference measurement query strategy based on gradients;

[0034] Multi-scale analysis offline proxy model iterative training module, which is used to expand the labeled data sample set based on the information measurement of unlabeled data and perform iterative training to obtain a final offline proxy meta-model;

[0035] Multi-scale analysis offline proxy model application module, which is used to embed the final offline proxy meta-model into the finite element to obtain the strain-stress response under multi-scale simulation of coal and rock loading.

[0036] In a third aspect, the present application also provides a computer device. The computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the steps of the methods in the above respective embodiments.

[0037] In a fourth aspect, the present application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the methods in the above respective embodiments are implemented.

[0038] The above-mentioned mesoscopic modeling and multi-scale simulation method of coal and rock materials based on active deep learning, first, formulate a complete unlabeled data sample space, and establish a labeled data sample set based on the complete unlabeled data sample space; then, construct a multi-scale analysis offline proxy model based on a deep learning model, and train the multi-scale analysis offline proxy model based on the labeled data sample set to obtain a basic predictor; then, input the unlabeled data sample into the basic predictor to obtain a sample pseudo-label, and determine the unlabeled data information metric based on the sample pseudo-label using an improved active learning strategy, where the improved active learning strategy includes a distance-weighted diversity metric query strategy for the input space and a data difference metric query strategy based on gradients; then, expand the labeled data sample set based on the unlabeled data information metric and perform iterative training to obtain a final offline proxy meta-model; finally, embed the final offline proxy meta-model into the finite element to obtain the strain-stress response under the multi-scale simulation of coal and rock loading. That is to say, by using the data optimization method, the data with the most learning value in the data space is intelligently selected, and an offline proxy model of coal and rock materials is constructed at the cost of the least computational resources, significantly reducing the human and material resources consumed by data collection or generation; secondly, compared with traditional multi-scale numerical simulation, the mesoscopic scale analysis process in multi-scale simulation is replaced by an offline proxy model, significantly improving the multi-scale simulation speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 FIG. is an application environment diagram of the mesoscopic modeling and multi-scale simulation method of coal and rock materials based on active deep learning in an embodiment;

[0040] Figure 2 FIG. is a flowchart of the mesoscopic modeling and multi-scale simulation method of coal and rock materials based on active deep learning in an embodiment;

[0041] Figure 3 FIG. is a structural diagram of a multi-scale analysis offline proxy model in an embodiment;

[0042] Figure 4 FIG. is a schematic diagram of iterative training of a multi-scale analysis offline proxy model in an embodiment;

[0043] Figure 5 FIG. is a schematic diagram of comparing the amount of labeled data of improved active learning in an embodiment;

[0044] Figure 6 FIG. is a schematic diagram of comparing the training loss of an improved active deep learning model and a general training loss in an embodiment;

[0045] Figure 7 FIG. is a schematic diagram of the application of a multi-scale analysis offline proxy model in an embodiment;

[0046] Figure 8 Schematic diagram for comparison with traditional numerical simulation results in an embodiment;

[0047] Figure 9 Schematic diagram of the stress-strain relationship of a certain unit in an embodiment;

[0048] Figure 10 Structural block diagram of a mesoscopic modeling and multi-scale simulation device for coal and rock materials based on active deep learning in an embodiment;

[0049] Figure 11 Internal structure diagram of a computer device in an embodiment. Detailed implementation manners

[0050] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0051] The mesoscopic modeling and multi-scale simulation method for coal and rock materials based on active deep learning provided by the embodiments of the present application can be applied to, for example, Figure 1 the application environment shown in the figure. Among them, the terminal communicates with the server through the network. The data storage system can store the data that the server needs to process. The data storage system can be integrated on the server, or placed in the cloud or other network servers. Among them, the terminal can be, but is not limited to, various personal computers, laptop computers, smart phones, tablet computers, Internet of Things devices and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. The portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server can be implemented by an independent server or a server cluster composed of multiple servers.

[0052] In one embodiment, as shown in Figure 2 the figure, a mesoscopic modeling and multi-scale simulation method for coal and rock materials based on active deep learning is provided. Taking the method applied to the Figure 1 server in the figure as an example, the method includes the following steps:

[0053] S201: Draw up a complete unlabeled data sample space, and establish a labeled data sample set based on the complete unlabeled data sample space.

[0054] In the embodiments of the present application, first, a complete unlabeled data sample space is drawn up and based on the complete unlabeled data sample space, a small number of typical loading paths are selected to establish a labeled data sample set

[0055] Specifically, in an embodiment of the present application, the establishment of the labeled data sample set based on the complete unlabeled data sample space includes:

[0056] S301: Numerically calibrate the discrete elements to clarify the variables included in the representative volume element.

[0057] S303: Draw up a complete unlabeled data sample space, use the representative volume element to simulate the typical loading path, obtain the offline data of the material, and establish a labeled data sample set.

[0058] In an embodiment of the present application, first, numerically calibrate the discrete elements to clarify the number of particles, particle gradation, particle density, loading amplitude, and loading speed included in the representative volume element. In specific applications, without loss of generality, the loading is carried out in plane stress to simplify the mathematical expression and computing resources. It should be specially noted that the loading in three-dimensional space is still applicable. The three strain and stress components under plane stress are convenient to be displayed in the form of a graph in space, so the plane stress model is adopted. According to relevant theoretical research, when the number of particle units is greater than 4000, the mesoscopic model can represent the macroscopic characteristics of the material to a great extent. Therefore, the number of unit particles is 6000, the average particle diameter is 1.2 mm, the particle density is 1.8 g / cm3, and the particle gradation adopts a uniform distribution; the loading speed is set to 10 -2 m / s, the time step is 2e-6, and the maximum strain is 0.1. Optionally, triaxial loading can also be used, and more computing resources can be omitted in triaxial loading.

[0059] After that, in multi-scale simulation, the force conditions of the elements at the mesoscopic level are not clear, and as many paths as possible should be considered when using the deep learning proxy model. Taking the condition of pure loading as an example, a small number of typical loading paths are expressed as typical model loading conditions such as uniaxial compression and biaxial compression in the plane mechanics space. Use the representative volume element to simulate the typical loading path, and the relevant simulations can be batch-executed through the Python language to calculate and obtain the offline data of the material. Extract the two principal direction strain components ε XX , ε YY and the stress components σ XX , σ YY during the loading process to form a labeled data sample set Without loss of generality, the path is relatively simple. When multi-scale simulation under complex paths is required, the loading conditions should be expanded. It should be noted that the same applicability also applies to cyclic loading.

[0060] In the plane strain stress space, the general expressions of strain and stress are:

[0061]

[0062] To further simplify the expression, the strain-stress components are projected into the principal strain-stress space through a transformation method. This transformation follows the transformation from the general strain-stress state to the principal strain-stress space, reducing the analysis dimension and saving computational resources. The calculation method for converting traditional stress components into the principal stress space is as follows:

[0063]

[0064] The inverse transformation formula is as follows:

[0065]

[0066] Similarly, the transformation of strain components has the same method.

[0067] S203: Construct a multi-scale analysis offline proxy model based on a deep learning model, and train the multi-scale analysis offline proxy model based on the labeled data sample set to obtain a basic predictor.

[0068] In the embodiments of the present application, a multi-scale analysis offline proxy model NN is constructed based on a deep learning model ini , specifically, the deep learning model can select a multi-layer perceptron model MLP, a recurrent neural network model RNN, and a Kolmogorov–Arnold network model KAN. Considering that the data is independent of the path, taking the multi-layer perceptron model MLP as an example, it is constructed using the Python deep learning package Pytorch, and after detailed architecture design and parameter optimization, as Figure 3 shown, it is set to [2, 64, 64, 64, 64, 2], that is, the input layer contains 2 neurons, the output layer also contains 2 neurons, and each of the middle 4 hidden layers contains 64 neurons. This design fully considers the complexity and balance of the problem, ensuring that the model can effectively capture the non-linear mapping relationship between the input and the output. The hidden layer uses the ReLU activation function (Rectified Linear Unit), which has good non-linear expression ability and can avoid the problem of gradient disappearance, thereby improving the convergence efficiency and performance of the model. Then, the labeled data sample set is input into the multi-scale analysis offline proxy model for model training to construct a basic predictor. During the training process, the mean squared error (MSE) is selected as the loss function, which is defined as the average of the squared differences between the predicted value and the true value, and is used to measure the fitting error of the model. The optimizer uses Adam (Adaptive Moment Estimation), which has the advantage of fast convergence and automatically adjusts the learning rate to meet the needs of different gradient updates. The initial value of the learning rate is set to 0.001. The calculation of the loss function is as follows:

[0069]

[0070] Among them, N is the number of samples, and y i is the predicted value, and is the labeled value.

[0071] In addition, in an embodiment of the present application, training the multi-scale analysis offline proxy model based on the labeled data sample set includes:

[0072] Using the gradient change method to determine the training balance judgment condition of the multi-scale analysis offline proxy model.

[0073] In an embodiment of the present application, the gradient change method is used as the criterion for determining model convergence. When the gradient norm is less than the preset threshold in 5 consecutive training rounds, it can be considered that the model has converged, the training ends, and the current multi-scale analysis offline proxy model is saved. This method can effectively avoid overtraining, reduce unnecessary computational overhead, and ensure that the prediction ability of the model reaches an ideal level. The preset gradient norm threshold is 10 -5 .

[0074] In an embodiment of the present application, the iterative gradient norm calculation formula is:

[0075]

[0076] Among them, is the gradient norm, L is the loss function, and ω i is the weight of the model.

[0077] In an embodiment of the present application, the iterative gradient norm calculation formula quantifies the gradient change amplitude of the current model by calculating the square root of the sum of the squares of the gradients of the loss function with respect to all weight parameters. The gradient norm can directly reflect the parameter update amplitude of the model in the current training round. A smaller gradient norm indicates that the parameters of the model have tended to be stable and the training is approaching convergence.

[0078] S205: Input the unlabeled data sample into the basic predictor to obtain a sample pseudo-label, and determine the unlabeled data information metric based on the sample pseudo-label. The improved active learning strategy includes a distance-weighted diversity metric query strategy in the input space and a data difference metric query strategy based on gradients.

[0079] In the embodiment of the present application, the basic predictor is used to infer the unlabeled data sample to obtain the sample output that is, the sample pseudo-label i (ε 1 , ε 2 , …, ε s ) and the one obtained by the basic predictor Analyze the information content contained in unlabeled data samples using an improved active learning strategy. The improved active learning strategy is an improved active deep learning query strategy that combines distance-weighted diversity metric queries in the input space with gradient-based data dissimilarity metric queries, making up for the characteristic that traditional gradient-based methods are prone to falling into local optima. That is, datasets with similar input sample features have similar results when calculated based on gradients, which makes the gradient-based query strategy easily batch-select data in some concentrated regions, resulting in the loss of the best information-rich data during intelligent selection. At the same time, this also provides a basis for batch expanding the dataset.

[0080] Specifically, in an embodiment of the present application, determining the unlabeled data information metric using the improved active learning strategy based on the sample pseudo-label includes:

[0081] S401: Calculate the input data dissimilarity metric to evaluate data dissimilarity.

[0082] S403: Determine the samples to be labeled by evaluating the magnitude or change of the gradient.

[0083] In an embodiment of the present application, the distance-weighted diversity metric query strategy in the input space is distance-weighted diversity: directly calculate the input data dissimilarity metric to evaluate data dissimilarity, and its core idea is to judge which samples are most valuable for model training by measuring the dissimilarity between different input samples. The gradient-based data dissimilarity metric query strategy: intelligently select more appropriate samples for labeling by evaluating the magnitude or change of the gradient. Its core idea is that when some data can significantly affect the model gradient update, these data samples often contain more information.

[0084] In an embodiment of the present application, the diversity metric calculation formula in the input space is:

[0085]

[0086] where x i , ε j are strain data in the input space, x i is the data in the unlabeled sample space, is the unlabeled sample space, ε j is the labeled sample data, D i is the labeled sample space, and Cosine Similarity(x i , ε j ) is the cosine similarity calculation.

[0087] In an embodiment of the present application, the distance-weighted diversity calculation method in the input space is cosine similarity, and without loss of generality, this method can be used to evaluate the similarity between high-dimensional vectors. Taking the two-dimensional main space components (strain components ε XX , ε YY ) as an example, the cosine similarity calculation method is as follows:

[0088]

[0089] where x i , x j are vectors in the data space, · represents the vector dot product, and ||·|| represents the vector norm.

[0090] The diversity metric calculation formula in the input space is:

[0091]

[0092] where x i , ε j are strain data in the input space, x i is the data in the unlabeled sample space, is the unlabeled sample space, ε j is the labeled sample data, D i is the labeled sample space, and Cosine Similarity(x i , ε j ) is the cosine similarity calculation.

[0093] In this embodiment, this strategy directly calculates the difference between input samples, and by measuring the distribution difference of samples in the feature space, it identifies samples with relatively large changes in feature values.

[0094] In an embodiment of the present application, the formula for calculating the gradient-based data difference metric is:

[0095]

[0096] where is the gradient norm of the root mean square error of the basic predictor, is the maximum gradient norm.

[0097] In an embodiment of the present application, the calculation process of the gradient-based data difference metric query strategy is as follows: for the basic predictor NN ini , according to the loss function - root mean square error MSE defined during the model training process as L(x i ); for the input sample , calculate its gradient norm

[0098]

[0099] In the formula, is the partial derivative of the loss function with respect to the j-th feature of the sample x i of.

[0100] Furthermore, calculate its maximum norm:

[0101]

[0102] where N is the total number of unlabeled samples.

[0103] Furthermore, in order to normalize and calculate the diversity of the input space to a unified dimension, normalization calculation is performed, and the calculation formula of the gradient-based data difference metric is expressed as:

[0104]

[0105] where is the gradient norm of the root mean square error of the basic predictor, is the maximum gradient norm.

[0106] The overall data information metric is calculated as follows:

[0107] Score = α × NGrad(x i ) + (1 - α) × InputDiff(x i , ε j )

[0108] where α is a parameter for controlling the weights of the two, and 0.7 can be selected in specific applications.

[0109] S207: Expand the labeled data sample set based on the unlabeled data information metric, and perform iterative training to obtain a final offline proxy meta-model.

[0110] In the embodiments of the present application, according to the unlabeled data information metric, unlabeled data samples with more information are extracted, discrete element RVE offline numerical calculations are performed, corresponding sample annotation information is obtained, the labeled data sample set is expanded, and a new data set is obtained In specific applications, the top 5% of the data with the most extracted information is used as augmented data, and the RVE unit is used for simulation to extract strain and stress data to augment the existing dataset. Without loss of generality, the selection of the amount of information samples can be determined according to the specific data space, rather than being fixed. At the same time, iterative training is carried out, that is, based on the augmented labeled data sample set, the multi-scale analysis offline proxy model is trained continuously, and the data information is measured again to augment the labeled data sample set. These steps are repeated iteratively until the information volume of the unlabeled dataset is less than the threshold or the number of iterations is greater than the preset value, and the final offline proxy meta-model is saved. The iterative loop execution process is as follows Figure 4 shown. The flowchart loops through the existing data space until the prediction gap of the proxy model is less than the threshold or the maximum number of iterations is reached. The preset proxy model prediction gap threshold is 0.05; the preset maximum number of iterations is 8 times, that is, the maximum learning path is 40% of all paths. In this embodiment, the final number of iterations is 5, and the augmented data is 25% of the data volume of the general data-driven method. As Figure 5 shown, it is a comparison of the different data volume requirements between the general deep learning proxy model and the improved active deep learning proposed in the present invention. Figure 5 On the left is the data requirement of the general deep learning proxy model as a control group. The preset maximum loading value is 0.15 strain, and the path loading end point is considered to select evenly distributed points in the ε x , ε y , ε xy space.

[0111] As Figure 6 shown, it is the change curve of the general deep learning training loss and the change trend of the optimal loss function under different numbers of iterations of the improved active deep learning proposed in the present invention. It can be seen that only one-fourth of the data is used to achieve an accuracy extremely close to that of the general deep learning method. The data after the improved active deep learning iteration is input into the multi-scale analysis offline proxy model NN ini , and it is trained using the same training method as the offline proxy model to obtain the final offline proxy meta-model

[0112] S209: Embed the final offline proxy meta-model into the finite element to obtain the strain-stress response under the multi-scale simulation of coal and rock loading.

[0113] In the embodiment of the present application, as Figure 7 shown, it is a schematic diagram of the application of the multi-scale analysis offline proxy model. Using the finite element analysis secondary development interface, Abaqus is used to build and implement a user subroutine VUMAT for explicit analysis. This program uses C++ as the basic framework and realizes the embedding of the designed final offline proxy meta-model by calling a Python program. Based on previous training and data fitting, this model can effectively predict and analyze the strain-stress response of coal and rock materials. By embedding this subroutine, seamless connection with the finite element calculation environment can be achieved, enabling users to directly call this model during subsequent simulation processes for real-time stress-strain analysis and result visualization. Meanwhile, a loading model of coal and rock materials is established in the finite element software to simulate the stress-strain behavior of coal and rock under different loading conditions in actual engineering. Specifically, a two-dimensional model is constructed, uniaxial loading is implemented, the maximum strain is less than 0.15, an inp input file is generated, and finally the input file and the developed user subroutine VUMAT are called through Abaqus commands to complete the numerical calculation. As Figure 8 shown, the comparison of the misess stress of uniaxial loading of coal and rock between the present invention and traditional numerical simulation calculation is presented. As Figure 9 shown, the force conditions of the same element in the calculation model are compared. It is not difficult to find that the errors in the macroscopic mechanical response of coal and rock under compression and the stress changes of the element are very small, fully meeting the engineering requirements. Through this mechanism, the microscopic and macroscopic characteristics of coal and rock materials can be considered simultaneously at multiple scales, capturing their complex stress-strain responses. Finally, the strain-stress relationship of coal and rock under multi-scale loading can be obtained, further providing an efficient numerical tool for the research and application of the mechanical behavior of coal and rock.

[0114] In the above mesoscopic modeling and multi-scale simulation method of coal and rock materials based on active deep learning, first, a complete unlabeled data sample space is formulated, and a labeled data sample set is established based on the complete unlabeled data sample space; then, a multi-scale analysis offline proxy model is constructed based on a deep learning model, and the multi-scale analysis offline proxy model is trained based on the labeled data sample set to obtain a basic predictor; then, unlabeled data samples are input into the basic predictor to obtain sample pseudo-labels, and an improved active learning strategy is used to determine the unlabeled data information metric based on the sample pseudo-labels. The improved active learning strategy includes a distance-weighted diversity metric query strategy for the input space and a data difference metric query strategy based on gradients; then, the labeled data sample set is expanded based on the unlabeled data information metric, and iterative training is performed to obtain a final offline proxy meta-model; finally, the final offline proxy meta-model is embedded into the finite element to obtain the strain-stress response under multi-scale simulation of coal and rock loading. That is to say, by using the data optimization method, the data with the most learning value in the data space is intelligently selected, and an offline proxy model of coal and rock materials is constructed at the cost of the least computational resources, significantly reducing the human and material resources consumed by data collection or generation; secondly, compared with traditional multi-scale numerical simulation, the mesoscopic scale analysis process in multi-scale simulation is replaced by an offline proxy model, significantly improving the multi-scale simulation speed.

[0115] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are sequentially shown according to the indications of the arrows, these steps are not necessarily executed sequentially in the order indicated by the arrows. Unless there is a clear indication in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least a part of the steps or stages in other steps or other steps.

[0116] Based on the same inventive concept, an embodiment of the present application further provides an apparatus for implementing the mesoscopic modeling and multi-scale simulation method of coal-rock materials based on active deep learning described above. The solution provided by this apparatus to solve the problem is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the apparatus for mesoscopic modeling and multi-scale simulation of coal-rock materials based on active deep learning provided below can refer to the limitations on the mesoscopic modeling and multi-scale simulation method of coal-rock materials based on active deep learning in the above text, and will not be repeated here.

[0117] In one embodiment, as Figure 10 shown, an apparatus 1000 for mesoscopic modeling and multi-scale simulation of coal-rock materials based on active deep learning is provided, including: a sample set establishment module 1001, a multi-scale analysis offline proxy model construction and initial training module 1003, a data difference measurement module 1005, a multi-scale analysis offline proxy model iterative training module 1007, and a multi-scale analysis offline proxy model application module 1009, wherein:

[0118] The sample set establishment module 1001 is used to draw up a complete unlabeled data sample space and establish a labeled data sample set based on the complete unlabeled data sample space.

[0119] The multi-scale analysis offline proxy model construction and initial training module 1003 is used to construct a multi-scale analysis offline proxy model based on a deep learning model and train the multi-scale analysis offline proxy model based on the labeled data sample set to obtain a basic predictor.

[0120] The data difference measurement module 1005 is configured to input unlabeled data samples into the basic predictor to obtain sample pseudo-labels, and determine the information measurement of the unlabeled data based on the sample pseudo-labels by adopting an improved active learning strategy, where the improved active learning strategy includes a distance-weighted diversity measurement query strategy in the input space and a data difference measurement query strategy based on gradients.

[0121] The multi-scale analysis offline proxy model iterative training module 1007 is configured to expand the labeled data sample set based on the information measurement of the unlabeled data and perform iterative training to obtain a final offline proxy meta-model.

[0122] The multi-scale analysis offline proxy model application module 1009 is configured to embed the final offline proxy meta-model into the finite element to obtain the strain-stress response under multi-scale simulation of coal and rock loading.

[0123] In an embodiment of the present application, the sample set establishment module is further configured to:

[0124] Perform numerical calibration on the discrete element to clarify the variables included in the representative volume element;

[0125] Formulate a complete unlabeled data sample space, use the representative volume element to simulate typical loading paths, obtain material offline data, and establish a labeled data sample set.

[0126] In an embodiment of the present application, the multi-scale analysis offline proxy model construction and initial training module is further configured to:

[0127] Adopt the gradient change method to determine the training balance judgment condition of the multi-scale analysis offline proxy model.

[0128] In an embodiment of the present application, the iterative gradient norm calculation formula is:

[0129]

[0130] Wherein, is the gradient norm, L is the loss function, and ω i is the weight of the model.

[0131] In an embodiment of the present application, the data difference measurement module is further configured to:

[0132] Calculate the input data difference measurement to evaluate the data difference;

[0133] Determine the samples to be labeled by evaluating the magnitude or change of the gradient.

[0134] In an embodiment of the present application, the diversity measurement calculation formula in the input space is:

[0135]

[0136] Among them, x i , ε j is the strain data in the input space, and x i is the data in the unlabeled sample space, is the unlabeled sample space, and ε j is the labeled sample data, and D i is the labeled sample space. CosineSimilarity(x i , ε j ) is the cosine similarity calculation.

[0137] In an embodiment of the present application, the calculation formula for the gradient-based data difference metric is as follows:

[0138]

[0139] Among them, is the gradient norm of the root mean square error of the basic predictor, is the maximum gradient norm.

[0140] Each module in the above mesoscopic modeling and multi-scale simulation device for coal and rock materials based on active deep learning can be implemented in whole or in part by software, hardware, and their combination. Each of the above modules can be embedded in the processor of the computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each of the above modules.

[0141] In one embodiment, a computer device is provided. The computer device can be a terminal, and its internal structure diagram can be as shown in Figure 11As shown in the figure. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner, and the wireless manner can be achieved through WIFI, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it realizes a mesoscopic modeling and multi-scale simulation method for coal and rock materials based on active deep learning. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, trackball, or touchpad set on the computer device housing, or an external keyboard, touchpad, or mouse, etc.

[0142] Those skilled in the art can understand that Figure 11 the structure shown in the figure is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0143] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above method embodiments are realized.

[0144] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by the processor, the steps in the above method embodiments are realized.

[0145] In one embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by the processor, the steps in the above method embodiments are realized.

[0146] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0147] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memory can include Read-Only Memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., without limitation.

[0148] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0149] The above-described embodiments merely represent several implementation manners of the present application. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for microscopic modeling and multi-scale simulation of coal and rock materials based on active deep learning, characterized in that: The method comprises: Formulate a complete unlabeled data sample space, and establish a labeled data sample set based on the complete unlabeled data sample space; Building a multi-scale analysis offline proxy model based on the deep learning model, and training the multi-scale analysis offline proxy model based on the labeled data sample set to obtain a basic predictor; Inputting unlabeled data samples into the basic predictor to obtain sample pseudo labels, and determining unlabeled data information metrics based on the sample pseudo labels using an improved active learning strategy, wherein the improved active learning strategy includes a distance-weighted diversity metric query strategy of an input space and a gradient-based data difference metric query strategy; Expanding the labeled data sample set based on the unlabeled data information metric, and performing iterative training to obtain a final offline proxy meta-model; The final offline proxy element model is embedded into the finite element to obtain the strain-stress response of coal and rock under multi-scale simulation of loading.

2. The method for mesoscopic modeling and multi-scale simulation of coal and rock materials based on active deep learning according to claim 1 is characterized in that: The establishing of a labeled data sample set based on the complete unlabeled data sample space comprises: Numerical calibration of discrete elements is performed to clarify the variables contained in the representative volume unit; A complete unlabeled data sample space is proposed, and the representative volume unit is used to simulate a typical loading path, obtain material offline data, and establish a labeled data sample set.

3. The method for mesoscopic modeling and multi-scale simulation of coal and rock materials based on active deep learning according to claim 1 is characterized in that: The training of the multi-scale analysis offline proxy model based on the labeled data sample set includes: The gradient change method is used to determine the training balance judgment condition of the multi-scale analysis offline proxy model.

4. The method for mesoscopic modeling and multi-scale simulation of coal and rock materials based on active deep learning according to claim 3 is characterized in that: The formula for calculating the iterative gradient norm is: in, is the gradient norm, L is the loss function, ω i are the weights of the model.

5. The method for mesoscopic modeling and multi-scale simulation of coal and rock materials based on active deep learning according to claim 1 is characterized in that: The method of determining the unlabeled data information metric based on the sample pseudo-label by adopting an improved active learning strategy includes: Calculate input data difference metrics to assess data differences; The samples to be labeled are determined by evaluating the size or change of the gradient.

6. The method for mesoscopic modeling and multi-scale simulation of coal and rock materials based on active deep learning according to claim 1 is characterized in that: The diversity metric calculation formula of the input space is: Among them, x i , ε j is the strain data in the input space, x i is the data in the unlabeled sample space, is the unlabeled sample space, ε j is the labeled sample data, D i is the labeled sample space, Cosine Similarity(x i ,ε j ) is the cosine similarity calculation.

7. The method for mesoscopic modeling and multi-scale simulation of coal and rock materials based on active deep learning according to claim 1 is characterized in that: The gradient-based data difference measurement calculation formula is: in, is the gradient norm of the root mean square error of the basic predictor, is the maximum gradient norm.

8. A coal and rock material microscopic modeling and multi-scale simulation device based on active deep learning, characterized in that: The device comprises: A sample set establishment module, used to formulate a complete unlabeled data sample space, and establish a labeled data sample set based on the complete unlabeled data sample space; A multi-scale analysis offline proxy model construction and initial training module is used to construct a multi-scale analysis offline proxy model based on a deep learning model, train the multi-scale analysis offline proxy model based on the labeled data sample set, and obtain a basic predictor; A data difference measurement module, used for inputting unlabeled data samples into the basic predictor to obtain sample pseudo labels, and determining the unlabeled data information measurement based on the sample pseudo labels by adopting an improved active learning strategy, wherein the improved active learning strategy includes a distance-weighted diversity measurement query strategy of the input space and a gradient-based data difference measurement query strategy; A multi-scale analysis offline proxy model iterative training module is used to expand the labeled data sample set based on the unlabeled data information metric and perform iterative training to obtain a final offline proxy meta-model; The multi-scale analysis offline proxy model application module is used to embed the final offline proxy element model into the finite element to obtain the strain-stress response under multi-scale simulation of coal and rock loading.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images