Large-scale coal rock multi-physics field imaging method and system based on iterative enhanced distillation

Abstract

The invention discloses a large-scale coal rock multi-physics field imaging method and system based on iteration intensified distillation, and the method comprises the following steps: obtaining detection, monitoring and finite physics field actual measurement data in the loading process of small and large-scale coal rock samples, building a numerical model according to a test, and simulating to obtain multi-physics field simulation data in the loading process; obtaining a full physical field data set; training and verifying the feature extraction network based on the data mechanism joint guidance architecture through a full-physical field data set to obtain a feature extraction model; testing a feature extraction model through detection and monitoring data to obtain physical field data features; the large-scale multi-physics-field coupling imaging network based on the enhanced distillation architecture extracts coupling factors through data features and optimizes a distillation strategy, so that small-scale sample data features are fused in the large-scale sample physics-field imaging process, and a large-scale multi-physics-field coupling imaging model is obtained; the model is tested through detection and monitoring data of a large-scale sample, and a multi-physical-field imaging result is obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of coal rock monitoring, and particularly relates to a large-scale coal rock multi-physical field imaging method and system based on iterative reinforcement distillation. BACKGROUND

[0002] Coal rock dynamic disaster is a complex disaster-creating evolution process in which coal rock mass develops from static to dynamic and transforms from balance to instability under the combined action of external force and excavation disturbance. At the same time, deep coal rock mass is subjected to the combined action of high ground stress, high ground temperature, gas-liquid seepage and other fields in actual engineering. Although domestic and foreign scholars have carried out a large number of researches and explorations on the mechanism of typical coal rock dynamic disaster, the multi-phase and multi-field evolution process and coupling law of coal rock mass are still unclear, and it is impossible to realize the coupling evolution process reproduction of stress, fracture, seepage, strain and temperature and other multi-physical fields of large-scale coal rock mass in the loading process. Therefore, it is necessary to carry out in-depth research on the multi-physical field imaging of large-scale coal rock in the loading process.

[0003] At present, certain progress has been made based on neural network and numerical simulation method, and these technologies can be used for the coupling evolution process reproduction of multi-physical field, especially showing potential in constructing three-dimensional digital model of discontinuous structure. For example, patent CN119129408A discloses an intelligent early warning method for coal and gas outburst based on test-theory-data hybrid driving, patent CN115879335A discloses a fluid multi-physical field parameter prediction method based on graph generation neural network, and patent CN115081354A discloses a numerical simulation analysis method for coal mine spent air phase change heat storage catalytic combustion based on multi-physical field coupling. However, these methods still have some limitations. They usually have complicated model reconstruction methods, and cannot realize the coupling imaging of large-scale coal rock mass multi-physical field based on detection and monitoring data, which limits their wide application in actual engineering.

[0004] In recent years, some experimental methods combining detection and monitoring technology with numerical simulation have been applied in the laboratory. However, the current utilization of these detection and monitoring data still mainly focuses on manually analyzing the time-varying law thereof. In addition, completely relying on numerical simulation for multi-physical field imaging of large-scale coal rock samples requires extremely high calculation cost, and even a slight modification in the model is likely to cause a large negative weight in the model calculation process. Therefore, how to utilize the detection and monitoring data to realize multi-physical field imaging of large-scale coal rock samples and make up for the limitations of traditional methods still needs further research. With the continuous progress and development of science and technology, the breakthrough progress of the reinforcement distillation method has attracted people's attention in the multi-physical field imaging of large-scale coal rock samples. Especially by using the reinforcement learning method, the neural network can be guided by the reward signal to gradually learn how to make the optimal decision without explicit supervision signal. At the same time, the distillation method transfers the complex data features to other models, so as to realize accelerated reasoning and higher optimization accuracy.

[0005] This provides the possibility for realizing large-scale coal rock physical field imaging by using the above method, but there is still a lack of specific feasible ideas and implementation methods at present. Therefore, it is urgent to carry out actual measurement of the detection and monitoring signals in the loading process of the coal rock sample, propose a large-scale coal rock multi-physical field imaging method and system based on iterative reinforcement distillation, realize multi-physical field coupling imaging based on the detection and monitoring data, and have a significant promoting effect on revealing the "black box" evolution process of the coal rock mass dynamic disaster and realizing the monitoring and early warning of the coal rock dynamic disaster. SUMMARY

[0006] The present scheme proposes a large-scale coal rock multi-physical field imaging method and system based on iterative reinforcement distillation to solve the problems and needs mentioned above. The technical purposes mentioned above can be achieved due to the adoption of the following technical features, and other technical effects can also be brought.

[0007] One object of the present application is to propose a large-scale coal rock multi-physical field imaging method based on iterative reinforcement distillation, characterized in that it comprises the following steps:

[0008] S10: arranging stress and optical fiber micro-nano monitoring sensors on the surface and inside of a small-scale coal rock sample, loading the coal rock sample by a testing machine, and recording mechanical parameters, constitutive equation, loading boundary condition data and measured data collected by the detection and monitoring sensors during the test;

[0009] S20: constructing a numerical simulation model according to the mechanical parameters, constitutive equation and loading boundary condition data recorded during the loading test of the small-scale coal rock sample, and obtaining full-time and space physical field data by calculating the numerical simulation model;

[0010] S30: Correspond the collected detection and monitoring data, finite physical field data and full-time-space physical field data through the boundary conditions and time data in the loading test process of the small-scale coal rock sample, obtain a small-scale full-physical field data set, and divide the small-scale full-physical field data set into a training set 1, a verification set 1 and a test set 1;

[0011] S40: Arranging stress and optical fiber micro-nano monitoring sensors on the surface and inside of the large-scale coal rock sample, loading the coal rock sample by a testing machine, and recording mechanical parameters, a constitutive equation, loading boundary condition data and measured data collected by the monitoring sensors in the test process;

[0012] S50: Constructing a numerical simulation model according to the mechanical parameters, the constitutive equation and the loading boundary condition data recorded in the loading test process of the large-scale coal rock sample, and obtaining full-time-space physical field data by calculating the numerical simulation model;

[0013] S60: Corresponding the collected detection and monitoring data and full-time-space physical field data through the boundary conditions and time data in the loading test process of the large-scale coal rock sample, obtaining a large-scale physical field data set, and dividing the large-scale physical field data set into a training set 2, a verification set 2, a test set 2 and a test set 3;

[0014] S70: Constructing a feature extraction network based on a data mechanism joint guidance architecture, training and verifying the feature extraction network through the small-scale and large-scale physical field data sets to obtain small-scale and large-scale physical field feature extraction models respectively, and inputting measured data of the test set 1 in the small-scale and large-scale physical field data sets into the small-scale and large-scale physical field imaging guidance models to obtain physical field imaging data features;

[0015] S80: A large-scale multi-physical field coupling imaging network based on a reinforced distillation architecture extracts coupling factors based on data features and optimizes a distillation strategy, realizes the fusion of small-scale sample data features in the large-scale sample physical field imaging process, and obtains a large-scale multi-physical field coupling imaging model;

[0016] S90: Inputting measured data of the test set 2 in the small-scale and large-scale physical field data sets into the multi-physical field coupling imaging model to obtain full-time-space physical field imaging results of the large-scale coal rock sample.

[0017] In addition, the large-scale coal rock multi-physical field imaging method and system based on iterative reinforced distillation according to the application can also have the following technical features:

[0018] In one example of the present application, in the step S30, the collected detection and monitoring data, the finite physical field data and the full-time-space physical field data are corresponded through the boundary conditions and the time data in the loading test process of the small-scale coal rock sample, to obtain a small-scale full-physical field data set, including the following steps:

[0019] Firstly, the seepage and fracture measured data are processed to obtain the finite physical field data at the stage time;

[0020] Then, the measured data are aligned according to the time to obtain the aligned detection and monitoring measured data; the finite physical field data and the full-time-space physical field data are aligned according to the time and the loading boundary conditions to obtain the aligned physical field data;

[0021] Finally, the aligned detection and monitoring measured data and the aligned physical field data are aligned according to the time and the loading boundary conditions to obtain the small-scale full-physical field data set.

[0022] In one example of the present application, in the step S50, a numerical simulation model is constructed according to the recorded mechanical parameters, the constitutive equation and the loading boundary condition data in the loading test process of the large-scale coal rock sample, and the full-time-space physical field data are obtained by calculating the numerical simulation model, including the following steps:

[0023] Firstly, a complete small-scale coal rock sample model is generated according to the size of the small-scale coal rock sample by using a numerical simulation method;

[0024] Then, the physical constants of the large-scale coal rock sample model are set according to the mechanical parameters and the constitutive equation of the large-scale coal rock sample loading test; the boundary conditions of the large-scale coal rock sample model are set according to the boundary conditions of the large-scale coal rock sample loading test;

[0025] Finally, the large-scale coal rock sample model is calculated to obtain the simulation data of the fracture field, the seepage field, the temperature field, the stress field and the strain field in the large-scale coal rock loading test.

[0026] In one example of the present application, in the step S70, an extraction network based on a data mechanism joint guiding architecture is constructed, a small-scale physical field feature extraction model is obtained by training and verifying the feature extraction network through the small-scale physical field data set, including the following steps:

[0027] S711: The small-scale physical field feature guiding learning module encodes the stress measured data σ n , the strain measured data ∈ n , the micro-current measured data e n , the fracture measured data f n , the resistivity measured data ρ n , and the potential measured data and wave velocity measured data v n Parameter enhanced stress feature σ fse , strain feature ∈ fse , micro-current feature e fse , fracture feature f fse , resistivity feature ρ fse , potential feature and wave velocity feature v fse Feature; the coded features are spliced into the features σ b , ∈ b , e b , f b , ρ b , and v b obtained by the first linear feature extraction layer B; the parameter enhanced features σf se , ∈f se , ef se , ff se , ρf se , and vf se are regularized by the first Normalization function to obtain features σ nn , ∈ nn , e nn , f nn , ρ nn , and v nn ;

[0028] S712: The features σ nn , ∈ nn , e nn , f nn , ρ nn , and v nn regularized are feature extracted by the second linear feature extraction layer A and the second activation function layer composed of the residual connection to obtain σ a+s , ∈ a+s , e a+s , f a+s , ρ a+s , and v a+s spliced into the features σ sssm , ∈ sssm , e sssm , f sssm , ρ sssm , and v sssm output by the first selective state space architecture to obtain σ SSSM , ∈ SSSM , eSSSM f SSSM ρ SSSM , and v SSSM The feature σ output by the first linear feature extraction layer A a ,∈ a e a f a ρ a , and v a Higher-dimensional features σ are obtained through feature extraction via the first convolutional layer and the first activation function layer. conv+s ,∈ conv+s e conv+s f conv+s ρ conv+s , and v conv+s ;

[0029] S713: Gated computation of high-dimensional features is performed through the first selective state-space architecture, and then concatenated with the features extracted by the first convolutional layer and the first activation function layer to obtain feature σ. SSSM ,∈ SSSM e SSSM f SSSM ρ SSSM , and v SSSM The first linear feature extraction layer B extracts features from the concatenated features and concatenates the extracted features with the features of the encoded data obtained by the first residual connection layer to obtain feature σf. b+se ,∈f b+se ef b+se ff b+se ,ρf b+se , and vf b+se ;

[0030] S714: Regularize the concatenated features using the second Normalization function to obtain feature σf. n ,∈f n ef n ff n ,ρf n , and vf n The stress characteristic σ is calculated by using the features extracted from the first fully connected layer. m,pred Fracturing characteristics F m,pred , seepage characteristics P m,pred Strain characteristics ∈ m,pred and temperature characteristics T m,pred ;

[0031] S715: constraint calculation of the features σ m,pred , F m,pred , P m,pred , ∈ m,pred and T m,pred of the neural network through the data-driven equation and the evolution mechanism equation, and finally obtaining a small-scale physical field feature extraction model.

[0032] In an example of the present application, in the step S713, the selective state space architecture implementation process is as follows:

[0033] First, a state space architecture is built through a continuous state space equation, wherein the expression of the continuous state space equation is as follows:

[0034] h t = Ah t-1 +Bx t

[0035] y t = Ch t

[0036] wherein h t-1 represents the system state; h t represents the updated system state; the matrix A represents the state transition matrix; B represents the input gate matrix; C represents the output mapping matrix; and x t represents the data of the current state input architecture.

[0037] Second, the state space architecture is discretized through a discrete state space equation; wherein the equation expression of the state transition matrix A is as follows:

[0038]

[0039] wherein n is the row index related to the orthogonal basis function in the polynomial space, and k is the column index related to the orthogonal basis function in the polynomial space.

[0040] In an example of the present application, in the step S715, the constraint calculation of the features σ m,pred , F m,pred , P m,pred , ∈ m,pred and T m,pred of the neural network through the data-driven equation and the evolution mechanism equation is as follows:

[0041] First, the data-driven equation is used to constrain the features σ m,pred , F m,pred , P m,pred , ∈ m,pred and T m,pred of the neural network.The calculation is constrained, and the expression is as follows:

[0042]

[0043] In the formula, and respectively represent the deviation calculation equations of stress, fracture, seepage, strain and temperature characteristics; m represents the sample quantity; i represents the serial index of the characteristics; σ m,pred , F m,pred , P m,pred , ∈ m,pred and T m,pred respectively represent the predicted stress, fracture, seepage, strain and temperature characteristics; σ m,true , F m,true , P m,true , ∈ m,true and T m,true respectively represent the real stress, fracture, seepage, strain and temperature characteristics.

[0044] Then, the mechanism equation is used to calculate the characteristics F m of the neural network. The calculation is constrained, and the expression is as follows:

[0045]

[0046] In the formula, represents the mechanism constraint equation of the predicted characteristics; M represents the total number of samples; and α represents a weight coefficient.

[0047] In one example of the present application, in the step S70, a physical field imaging guidance network based on a data mechanism joint guidance architecture is constructed, a large-scale physical field imaging guidance model is obtained by training and verifying the physical field imaging guidance network through a large-scale physical field data set, and the following steps are included:

[0048] S721: A large-scale physical field characteristic guidance learning module extracts parameter-enhanced stress characteristics σ Nfse , strain characteristics ∈ Nfse , micro-current characteristics e Nfse , fracture characteristics f Nfse , resistivity characteristics ρ Nfse , potential characteristics and wave velocity characteristics v Nfse from the stress measured data σ N , strain measured data ∈ N , micro-current measured data e N , fracture measured data f N , resistivity measured data ρ N , potential measured data and wave velocity measured data v N through a second data encoding layer.characteristics; the coded characteristics are spliced into the characteristics obtained by the second linear feature extraction layer B to obtain the characteristics Nb , Nb , Nb , Nb , Nb , , Nb ; the parameter-enhanced characteristics are regularized by the third Normalization function to obtain the characteristics Nse , Nse , Nse , Nse , Nse , , Nse ; Nn , Nn , Nn , Nn , Nn , , Nn ;

[0049] The regularized characteristics are spliced into the characteristics Nn , Nn , Nn , Nn , Nn , , Nn obtained by the fourth linear feature extraction layer A and the fourth activation function layer to obtain the characteristics Na+s , Na+s , Na+s , Na+s , Na+s , , Na+s ; Nsssm , Nsssm , Nsssm , Nsssm , Nsssm , , Nsssm spliced into the characteristics NSSSM , NSSSM , NSSSM , NSSSM , NSSSM , , NSSSM obtained by the second selective state space architecture to obtain the characteristics Na , Na , Na , Na , Na , ;and v a Feature extraction is performed through the second convolutional layer and the third activation function layer to obtain higher-dimensional features σ Nconv+s , Nconv+s , Nconv+s , Nconv+s , Nconv+s , and v Nconv+s ;

[0050] S723: Gating calculation is performed on the high-dimensional features by the second selective state space architecture, and the features extracted by the second convolutional layer and the third activation function layer are spliced to obtain features σ NSSSM , NSSSM , NSSSM , NSSSM , NSSSM , and v NSSSM ; The second linear feature extraction layer B performs feature extraction on the spliced features, and splices the extracted features with the features of the encoded data by the second residual connection layer to obtain features σf Nb+se , Nb+se , Nb+se , Nb+se , Nb+se , and vf Nb+se ;

[0051] S724: The spliced features are regularized by the fourth Normalization function to obtain features σf Nn , Nn , Nn , Nn , Nn , and vf Nn ; The extracted features by the second fully connected layer are calculated to obtain stress features σ M,pred , fracture features F M,pred , seepage features P M,pred , strain features ∈ M,pred and temperature features T M,pred ;

[0052] S725: The features σ M,pred , F M,pred , P M,pred , ∈ M,pred and T M,pred of the neural network are calculated by the data-driven equation and the evolution mechanism equation to obtain the large-scale physical field imaging guide model.

[0053] In one example of the present application, in the step S725, the features σ M,pred M,pred M,pred M,pred M,pred and T M,pred of the neural network are calculated by the data-driven equation and the evolution mechanism equation, and the calculation is constrained as follows:

[0054] Firstly, the features σ M,pred M,pred M,pred M,pred of the neural network are calculated by the data-driven equation and the calculation is constrained, and the expression is as follows:

[0055]

[0056] In the formula, and respectively represent the deviation calculation equations of stress, fracture, seepage, strain and temperature features; M represents the number of samples; i represents the serial index of the features; σ M,pred M,pred M,prea M,pred M,pred respectively represent the predicted stress, fracture, seepage, strain and temperature features; σ M,true M,true M,true M,true M,true respectively represent the real stress, fracture, seepage, strain and temperature features.

[0057] Then, the features F m of the neural network are calculated by the mechanism equation and the calculation is constrained, and the expression is as follows:

[0058]

[0059] In the formula, represents the mechanism constraint equation of the predicted features; M represents the total number of samples; and α represents the weight coefficient.

[0060] In one example of the present application, in the step S80, the multi-physical field coupling imaging network based on the reinforced distillation architecture extracts coupling factors based on data features and optimizes a distillation strategy, realizes the fusion of small-scale sample data features in the large-scale sample physical field imaging process, and obtains a large-scale multi-physical field coupling imaging model, including the following steps:

[0061] S81: The multi-physical field coupling imaging network extracts and monitors the measured data x​​​​​​​​​​​​​​​i N rounds of distillation to obtain small-scale physical field imaging results p(x i ; Θ p ), and the expression of the distillation equation is as follows:

[0062]

[0063] In the formula, m represents the number of samples; x represents the equation for N rounds of distillation for small-scale physical field imaging; x i represents the input detection and monitoring measured data; Θ g represents the weight parameter of the small-scale physical field feature extraction model; p(x i ; Θ g ) represents the probability distribution of the small-scale physical field imaging data features based on the current weight parameter; Θ p represents the weight parameter of the multi-physical field coupling imaging network; p(x i ; Θ p ) represents the probability distribution of the small-scale physical field imaging results based on the current weight parameter;

[0064] S82: The distillation strategy of the multi-physical field coupling imaging network is optimized through the reinforcement learning architecture to obtain a small-scale multi-physical field coupling imaging model;

[0065] S83: The small-scale multi-physical field coupling imaging model performs N rounds of distillation on the detection and monitoring measured data X I in the large-scale physical field imaging data features to obtain large-scale physical field imaging results P(X I ; Θ P ), and the expression of the distillation equation is as follows:

[0066]

[0067] In the formula, M represents the number of samples; X represents the equation for N rounds of distillation for large-scale physical field imaging; X I represents the input detection and monitoring measured data; Θ G represents the weight parameter of the large-scale physical field feature extraction model; P(x I ; Θ G ) represents the probability distribution of the large-scale physical field imaging data features based on the current weight parameter; Θ P represents the weight parameter of the small-scale multi-physical field coupling imaging model; P(x I ; Θ P ) represents the probability distribution of the large-scale physical field imaging results based on the current weight parameter;

[0068] S84: A distillation strategy of a multi-physical field coupling imaging network is optimized through a reinforcement learning architecture, and a large-scale multi-physical field coupling imaging model is obtained.

[0069] Another object of the present application is to provide a large-scale coal rock multi-physical field imaging system based on iterative reinforcement distillation, comprising:

[0070] The first data acquisition module is configured to arrange stress and optical fiber micro-nano monitoring sensors on the surface and inside of the small-scale coal rock sample, and to record the mechanical parameters, constitutive equation, loading boundary condition data and measured data collected by the monitoring sensors during the loading test of the coal rock sample by the testing machine.

[0071] The first numerical simulation model module is configured to construct a numerical simulation model according to the mechanical parameters, constitutive equation and loading boundary condition data recorded during the loading test of the small-scale coal rock sample, and to obtain full-time-space physical field data by calculating the numerical simulation model.

[0072] The first data division module is configured to correspond the collected detection and monitoring data, finite physical field data and full-time-space physical field data by the boundary conditions and time data during the loading test of the small-scale coal rock sample, to obtain a small-scale full-physical field data set, and to divide the small-scale full-physical field data set into a training set 1, a validation set 1 and a test set 1.

[0073] The second data acquisition module is configured to arrange stress and optical fiber micro-nano monitoring sensors on the surface and inside of the large-scale coal rock sample, and to record the mechanical parameters, constitutive equation, loading boundary condition data and measured data collected by the monitoring sensors during the loading test of the coal rock sample by the testing machine.

[0074] The second numerical simulation model module is configured to construct a numerical simulation model according to the mechanical parameters, constitutive equation and loading boundary condition data recorded during the loading test of the large-scale coal rock sample, and to obtain full-time-space physical field data by calculating the numerical simulation model.

[0075] The second data division module is configured to correspond the collected detection and monitoring data and full-time-space physical field data by the boundary conditions and time data during the loading test of the large-scale coal rock sample, to obtain a large-scale physical field data set, and to divide the large-scale physical field data set into a training set 2, a validation set 2, a test set 2 and a test set 3.

[0076] The imaging data feature module is configured to construct a feature extraction network based on a data mechanism joint guidance architecture, and train and verify the feature extraction network through small-scale and large-scale physical field data sets to obtain small-scale and large-scale physical field feature extraction models respectively; and the measured data of test set 1 in the small-scale and large-scale physical field data sets are respectively input into the small-scale and large-scale physical field imaging guidance models to obtain physical field imaging data features.

[0077] The coupled imaging model module is configured to extract coupling factors through data features and optimize distillation strategies based on a large-scale multi-physical field coupled imaging network based on a reinforcement distillation architecture, so that small-scale sample data features are integrated into the large-scale sample physical field imaging process, and a large-scale multi-physical field coupled imaging model is obtained.

[0078] The imaging result module is configured to input the measured data of test set 2 in the small-scale and large-scale physical field data sets into the multi-physical field coupled imaging model to obtain the full-time-space physical field imaging result of the large-scale coal rock sample.

[0079] Compared with the prior art, the present application has the following beneficial effects:

[0080] 1. The present application focuses on realizing large-scale coal rock multi-physical field imaging based on detection and monitoring data in the large-scale coal rock sample loading test process through the large-scale coal rock multi-physical field imaging method and system based on iterative reinforcement distillation. Compared with the traditional numerical simulation and neural network method, this large-scale coal rock multi-physical field imaging method based on iterative reinforcement distillation integrates the detection, monitoring and physical field data features of small-scale coal rock samples in the large-scale coal rock sample fracture imaging process by means of reinforcement learning and knowledge distillation, so as to realize the organic integration of multi-scale coal rock sample data collaborative learning and dynamic distillation optimization, significantly improve the physical field imaging spatial resolution of large-scale coal rock samples, and construct a new paradigm of coal rock physical field imaging based on the reinforcement distillation architecture.

[0081] 2. The small-scale and large-scale physical feature extraction model in the present application is a selective state space neural network based on a data mechanism joint guidance architecture. The network simultaneously uses evolutionary mechanism equations and data-driven equations as constraints, which not only ensures that the output of the guidance model follows the physical law, but also fully excavates the nonlinear features in the multi-source experimental data, so as to exhibit excellent physical feasibility and robustness under the test data distribution. The selective state space architecture neural network greatly reduces the dependence on computing resources through fine division and selection of state variables and observation variables, so that the model can efficiently use the data collected in the laboratory coal rock sample loading process and the data obtained by numerical model simulation, and extract high-credibility physical field imaging data features for the large-scale multi-physical field coupled imaging network.

[0082] 3、The application realizes the fusion of small-scale sample data characteristics in the large-scale sample physical field imaging process through knowledge distillation and reinforcement learning, realizes the deep fusion of measured data and numerical simulation results, and ensures that the large-scale fracture imaging results can truly reflect the physical field distribution characteristics of the coal rock sample. Meanwhile, the equation optimization network based on reinforcement learning calculates the imaging coupling factors between multiple physical fields through the distillation equation and strategy module, realizes the collaborative innovation of structured knowledge distillation and dynamic coupling strategy optimization, refines the physical field imaging mechanism knowledge of samples of different scales and deeply fuses it into the distillation strategy of reinforcement learning, and effectively improves the large-scale coal rock sample physical field imaging precision based on detection and monitoring data.

[0083] The optimal embodiments of implementing the present application will be described in more detail below with reference to the accompanying drawings, so that the features and advantages of the present application can be easily understood. BRIEF DESCRIPTION OF DRAWINGS

[0084] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings of the embodiments of the present application will be briefly introduced below. The drawings are only used to show some embodiments of the present application, and not to limit all embodiments of the present application to this.

[0085] Figure 1 The flow chart of the large-scale coal rock multi-physical field imaging method based on iterative reinforcement distillation according to the embodiments of the present application;

[0086] Figure 2 The structural schematic diagram of the small-scale physical field characteristic guided learning module according to the embodiments of the present application;

[0087] Figure 3 The structural schematic diagram of the large-scale physical field characteristic guided learning module according to the embodiments of the present application;

[0088] Figure 4 The structural schematic diagram of the small-scale physical field imaging reinforcement distillation stage according to the embodiments of the present application;

[0089] Figure 5 The structural schematic diagram of the test machine according to the embodiments of the present application;

[0090] Figure 6 The structural schematic diagram of the test machine (X-ray source and CT detection sensor are hidden) according to the embodiments of the present application;

[0091] Figure 7 The structural schematic diagram of the internal sensor arrangement of the test machine according to the embodiments of the present application.

[0092] LIST OF REFERENCE NUMERALS:

[0093] Test machine 100;

[0094] X-ray source 110;

[0095] CT detection sensor 120;

[0096] Multi-frequency acoustic wave monitoring sensor 130;

[0097] Ultrasonic phased array detection sensor 140;

[0098] Surface fiber micro-nano monitoring sensor 150;

[0099] Surface stress monitoring sensor 160;

[0100] Electrode monitoring sensor 170;

[0101] Micro-current monitoring sensor 180;

[0102] Internal fiber nano monitoring sensor 190;

[0103] Internal stress monitoring sensor 200. DETAILED DESCRIPTION

[0104] In order to make the purpose, technical scheme and advantages of the technical scheme of the present application more clear, the technical scheme of the present application will be described clearly and completely below in combination with the drawings of the specific embodiments of the present application. The same reference signs in the drawings represent the same parts. It should be noted that the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the described embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0105] Unless otherwise defined, technical or scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terms "first", "second", and similar terms, as used in the specification and claims of this application, do not imply any order, quantity, or importance, but are used to distinguish different components. Similarly, the terms "one" or "a" or "an" do not necessarily mean "one", "single" or "only one". The terms "including", "containing", and similar terms, mean that the elements or objects before the term encompass the elements or objects listed after the term and their equivalents, without excluding other elements or objects. The terms "connected" or "connected" and similar terms do not necessarily mean physical or mechanical connections, but can include electrical connections, whether direct or indirect. The terms "up", "down", "left", "right", and the like only represent relative positional relationships, which may change when the absolute positions of the described objects change.

[0106] According to a first aspect of the present invention, a large-scale coal and rock multiphysics imaging method based on iterative enhanced distillation is provided, such as... Figure 1 As shown, it includes the following steps:

[0107] S10: Stress and fiber optic micro / nano monitoring sensors are deployed on the surface and inside of small-scale coal and rock samples. Microcurrent monitoring, multi-frequency acoustic wave monitoring, electrode monitoring, ultrasonic phased array detection, and CT detection sensors are deployed on the surface of the small-scale coal and rock samples. The coal and rock samples are subjected to loading tests using a testing machine. During the test, mechanical parameters, constitutive equations, loading boundary condition data, and measured data collected by the detection and monitoring sensors are recorded, including full-time stress at local measuring points, full-time strain at local measuring points, full-time microcurrent at local measuring points, full-time fracture, full-time resistivity at local measuring points, full-time potential at local measuring points, full-time wave velocity, and staged full-space seepage and fracture (finite physics field data). Specifically, such as... Figure 5 , Figure 6 and Figure 7As shown, the testing machine 100 comprises: a loading device 101, a coal rock sample 102, an X-ray source 110, a CT detection sensor 120, a multi-frequency acoustic wave monitoring sensor 130, an ultrasonic phased array detection sensor 140, a surface optical fiber micro-nano monitoring sensor 150, a surface stress monitoring sensor 160, an electrode monitoring sensor 170, a micro-current monitoring sensor 180, an internal optical fiber monitoring sensor 190, and an internal stress monitoring sensor 200; the loading device 101 is located on both sides of the coal rock sample 102 in the three-dimensional height direction Z, the X-ray source 110 and the CT detection sensor 120 are located on both sides of the coal rock sample 102 in the lateral direction X, the multi-frequency acoustic wave monitoring sensor 130, the ultrasonic phased array detection sensor 140, the surface optical fiber micro-nano monitoring sensor 150, the surface stress monitoring sensor 160, the electrode monitoring sensor 170, the micro-current monitoring sensor 180, the internal optical fiber monitoring sensor 190, and the internal stress monitoring sensor 200 are located on the coal rock sample 102; the loading device 101 performs loading test on the coal sample 102, the X-ray source 110 emits X-rays and the CT-ray detector 120 receives the X-rays, thereby obtaining full-space seepage and fracture measurement data of the coal rock sample during the breaking process, the multi-frequency acoustic wave monitoring sensor 130 collects full-time and space breaking measurement data of the coal rock test, and the ultrasonic phased array detection sensor 140 collects full-time and space wave velocity measurement data of the coal rock sample; the surface optical fiber micro-nano monitoring sensor 150 and the internal optical fiber monitoring sensor 190 are used to monitor full-time strain measurement data of the local measurement point position of the coal rock sample; the surface stress monitoring sensor 160 and the internal stress monitoring sensor 200 are used to monitor full-time stress measurement data of the local measurement point position of the coal rock sample; the electrode monitoring sensor 170 is used to monitor full-time resistivity and potential measurement data of the local measurement point position of the coal rock sample; the micro-current monitoring sensor 180 is used to monitor full-time micro-current measurement data of the local measurement point position of the coal rock sample. The lateral direction X, the longitudinal direction Y, and the height direction Z are perpendicular to each other.

[0108] S20: Construct a numerical simulation model according to the mechanical parameters, the constitutive equation, and the loading boundary condition data recorded in the loading test process of the small-scale coal rock sample, and obtain full-time and space physical field data (stress field simulation, strain field simulation data, and temperature field simulation data) by calculating the numerical simulation model;

[0109] S30: Correspond the collected detection and monitoring data (measured data such as micro-current, fracture, resistivity, potential, wave velocity, stress and strain), finite physical field data (measured data of seepage and fracture) and full-time-space physical field data (simulated data of temperature field, stress field and strain field) through the boundary conditions and time data in the loading test process of the small-scale coal rock sample, to obtain a small-scale full-physical field data set, and divide the small-scale full-physical field data set into a training set 1, a verification set 1 and a test set 1;

[0110] S40: Arranging stress and optical fiber micro-nano monitoring sensors on the surface and inside of the large-scale coal rock sample. Arranging micro-current monitoring, multi-frequency acoustic wave monitoring, electrode monitoring, CT detection and ultrasonic phased array detection sensors on the surface of the large-scale coal rock sample. The test machine performs a loading test on the coal rock sample, and records mechanical parameters, constitutive equations, loading boundary condition data and local measurement point position full-time stress, local measurement point position full-time strain, local measurement point position full-time micro-current, full-time-space fracture, local measurement point position full-time resistivity, local measurement point position full-time potential and full-time-space wave velocity and other measured data collected by the monitoring sensors during the test;

[0111] S50: Constructing a numerical simulation model according to the mechanical parameters, constitutive equations and loading boundary condition data recorded during the loading test of the large-scale coal rock sample, and obtaining full-time-space physical field data (simulated data of stress field, strain field, fracture field, seepage field and temperature field) by calculating the numerical simulation model;

[0112] S60: Corresponding the collected detection and monitoring data (measured data such as micro-current, fracture, resistivity, potential, wave velocity, stress and strain) and full-time-space physical field data (simulated data of fracture field, seepage field, temperature field, stress field and strain field) through the boundary conditions and time data in the loading test process of the large-scale coal rock sample, to obtain a large-scale physical field data set, and divide the large-scale physical field data set into a training set 2, a verification set 2, a test set 2 and a test set 3;

[0113] S70: Constructing a feature extraction network based on a data mechanism joint guidance architecture, and training and verifying the feature extraction network through the small-scale and large-scale physical field data sets to obtain small-scale and large-scale physical field feature extraction models respectively; inputting the local measurement point position full-time micro-current, full-time-space fracture, local measurement point position full-time resistivity, local measurement point position full-time temperature, full-time-space wave velocity, local measurement point position full-time stress and local measurement point position full-time strain and other measured data of the test set 1 in the small-scale and large-scale physical field data sets into the small-scale and large-scale physical field imaging guidance models respectively to obtain physical field imaging data features;

[0114] S80: The large-scale multi-physical field coupling imaging network based on the reinforcement distillation architecture extracts coupling factors based on data feature extraction and optimizes the distillation strategy to integrate small-scale sample data features in the large-scale sample physical field imaging process to obtain a large-scale multi-physical field coupling imaging model.

[0115] S90: The local measurement point position full-time stress, local measurement point position full-time strain, local measurement point position full-time micro-current, full-time and space fracture, local measurement point position full-time resistivity, local measurement point position full-time potential, and full-time and space wave velocity of the test set 2 in the small-scale and large-scale physical field data set are respectively input into the multi-physical field coupling imaging model to obtain the full-time and space physical field (stress field, strain field, fracture field, seepage field, and temperature field) imaging results of the large-scale coal rock sample.

[0116] The imaging method focuses on realizing large-scale coal rock multi-physical field imaging based on detection and monitoring data in the large-scale coal rock sample loading test process through the large-scale coal rock multi-physical field imaging method and system based on iterative reinforcement distillation. Compared with traditional numerical simulation and neural network methods, this large-scale coal rock multi-physical field imaging method based on iterative reinforcement distillation integrates small-scale coal rock sample detection, monitoring, and physical field data features in the large-scale coal rock sample fracture imaging process with the help of reinforcement learning and knowledge distillation, thereby realizing the organic integration of multi-scale coal rock sample data collaborative learning and dynamic distillation optimization, significantly improving the physical field imaging spatial resolution of large-scale coal rock samples, and constructing a new paradigm for coal rock physical field imaging based on the reinforcement distillation architecture.

[0117] The small-scale and large-scale physical feature extraction model in the imaging method is a selective state space neural network based on a data mechanism joint guidance architecture. The network simultaneously uses evolutionary mechanism equations and data-driven equations as constraints, which not only ensures that the output of the guidance model follows the physical law, but also fully excavates the nonlinear features in the multi-source experimental data, thereby exhibiting excellent physical feasibility and robustness under the distribution of experimental data. The selective state space architecture neural network greatly reduces the dependence on computing resources through fine division and selection of state variables and observation variables, enabling the model to efficiently use the data collected during the laboratory coal rock sample loading process and the data obtained from numerical model simulation, and to extract high-credibility physical field imaging data features for the large-scale multi-physical field coupling imaging network.

[0118] The imaging method realizes the fusion of small-scale sample data characteristics in the large-scale sample physical field imaging process through knowledge distillation and reinforcement learning, realizes the deep fusion of measured data and numerical simulation results, and ensures that the large-scale fracture imaging result can truly reflect the physical field distribution characteristics of the coal rock sample. Meanwhile, the equation optimization network based on reinforcement learning calculates the imaging coupling factors between multiple physical fields through the distillation equation and the strategy module, realizes the collaborative innovation of structured knowledge distillation and dynamic coupling strategy optimization, refines the physical field imaging mechanism knowledge of samples of different scales and deeply fuses it into the distillation strategy of reinforcement learning, and effectively improves the imaging precision of large-scale coal rock samples based on detection and monitoring data.

[0119] In one example of the present application, in the step S20, a numerical simulation model is constructed according to the mechanical parameters, constitutive equation and loading boundary condition data recorded in the loading test of the small-scale coal rock sample, and full-time-space physical field data (stress field simulation, strain field simulation data and temperature field simulation data) are obtained by calculating the numerical simulation model, including the following steps:

[0120] Firstly, a complete small-scale coal rock sample model is generated according to the size of the small-scale coal rock sample by using the numerical simulation method, and the model size is 25mm*25mm*50mm;

[0121] Then, the physical constants of the small-scale coal rock sample model are set according to the mechanical parameters and constitutive equation of the small-scale coal rock sample loading test, and the boundary conditions of the small-scale coal rock sample model are set according to the boundary conditions of the small-scale coal rock sample loading test;

[0122] Finally, the small-scale coal rock sample model is calculated to obtain the stress field, strain field and temperature field simulation data in the small-scale coal rock loading test.

[0123] In one example of the present application, in the step S30, the collected detection and monitoring data (measured data such as micro-current, fracture, resistivity, potential, wave speed, stress and strain), limited physical field data (seepage and fracture measured data) and full-time-space physical field data (temperature field simulation data, stress field simulation data and strain field simulation data) are corresponded through the boundary conditions and time data in the small-scale coal rock sample loading test to obtain a small-scale full-physical field data set, including the following steps:

[0124] Firstly, the seepage field and fracture field data (limited physical field data) at the stage time are obtained by processing the seepage and fracture measured data;

[0125] Then, the measured data of micro-current, fracture, resistivity, potential, wave velocity, stress and strain are aligned according to time to obtain aligned detection and monitoring measured data; the finite physical field data and the full-time-space physical field data are aligned according to time and loaded boundary conditions to obtain aligned physical field data.

[0126] Finally, the aligned detection and monitoring measured data and the aligned physical field data are aligned according to time and loaded boundary conditions to obtain a small-scale full-physical field data set.

[0127] In an example of the present application, in the step S50, a numerical simulation model is constructed according to the mechanical parameters, the constitutive equation and the loading boundary condition data recorded in the large-scale coal rock sample loading test, and full-time-space physical field data are obtained by calculating the numerical simulation model, including the following steps:

[0128] First, a complete small-scale coal rock sample model is generated according to the size of the small-scale coal rock sample by using a numerical simulation method, and the model size is 1m*1m*1.5m;

[0129] Then, the physical constants of the large-scale coal rock sample model are set according to the mechanical parameters and the constitutive equation of the large-scale coal rock sample loading test; the boundary conditions of the large-scale coal rock sample model are set according to the boundary conditions of the large-scale coal rock sample loading test;

[0130] Finally, the large-scale coal rock sample model is calculated to obtain simulation data of the fracture field, the seepage field, the temperature field, the stress field and the strain field in the large-scale coal rock loading test.

[0131] In an example of the present application, in the step S60, the collected detection and monitoring data (measured data of micro-current, fracture, resistivity, potential, wave velocity, stress and strain) and the full-time-space physical field data (simulation data of the fracture field, the seepage field, the temperature field, the stress field and the strain field) are corresponded through the boundary conditions and the time data in the large-scale coal rock sample loading test to obtain a large-scale physical field data set, including the following steps:

[0132] First, the measured data of micro-current, fracture, resistivity, potential, wave velocity, stress and strain are aligned according to time to obtain aligned detection and monitoring measured data; the full-time-space physical field data are aligned according to time and loaded boundary conditions to obtain aligned physical field data;

[0133] Finally, the aligned detection and monitoring measured data and the aligned physical field data are aligned according to time and loaded boundary conditions to obtain a large-scale full-physical field data set.

[0134] In one example of the present application, in the step S70, an extraction network based on a data mechanism joint guidance architecture is constructed, and a small-scale physical field feature extraction model is obtained by training and verifying the feature extraction network through a small-scale physical field data set, as shown in Figure 2 The method comprises the following steps:

[0135] S711: The small-scale physical field feature guidance learning module extracts the stress measured data σ n , strain measured data ∈ n , micro-current measured data e n , fracture measured data f n , resistivity measured data ρ n , potential measured data , and wave velocity measured data v n to obtain parameter-enhanced stress features σ fse , strain features ∈ fse , micro-current features e fse , fracture features f fse , resistivity features ρ fse , potential features , and wave velocity features v fse ; the encoded features are spliced into the features σ b , ∈ b , e b , f b , ρ b , , and v b obtained by the first linear feature extraction layer B through the first residual connection layer; and the parameter-enhanced features σf se , ∈f se , ef se , ff se , ρf se , , and vf se are regularized by the first Normalization function to obtain the features σ nn , ∈ nn , e nn , f nn , ρ nn , , and v nn ;

[0136] S712: The regularized features σ nn , ∈ nn , e nn , f nn , ρ nn , , and v nnThe residual connection composed of the second linear feature extraction layer A and the second activation function layer is used for feature extraction to obtain σ a+s a+s a+s a+s a+s a+s The features σ sssm sssm sssm sssm sssm sssm SSSM SSSM SSSM SSSM SSSM SSSM The features σ a a a a a a The higher-dimensional features σ conv+s conv+s conv+s conv+s conv+s conv+s

[0137] S713: The high-dimensional features are calculated by the first selective state space architecture, and are spliced with the features extracted by the first convolutional layer and the first activation function layer to obtain features σ SSSM SSSM SSSM SSSM SSSM SSSM The features extracted by the first linear feature extraction layer B are spliced with the features of the encoded data by the first residual connection layer to obtain features σf b+se b+se b+se b+se b+se b+se ​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​

[0138] S714: Regularize the concatenated features using the second Normalization function to obtain feature σf. n ,∈f n ef n ff n ,ρf n , and vf n The stress characteristic σ is calculated by using the features extracted from the first fully connected layer. m,pred Fracturing characteristics F m,pred , seepage characteristics P m,pred Strain characteristics ∈ m,pred and temperature characteristics T m,pred ;

[0139] S715: Analyzing the characteristic σ of neural networks through data-driven equations and evolutionary mechanism equations. m,pred F m,pred P m,pred ,∈ m,pred and T m,pred The calculations are constrained, and the final result is a small-scale physical field feature extraction model.

[0140] In one example of the present invention, in step S713, the selective state-space architecture is specifically implemented as follows:

[0141] First, a state-space architecture is constructed using continuous state-space equations, the expressions of which are as follows:

[0142] h t =Ah t-1 +Bx t

[0143] y t =Ch t

[0144] Among them, h t-1 Represents the system state; h t Represents the updated system state; matrix A represents the state transition matrix; matrix B represents the input gating matrix; matrix C represents the output mapping matrix; x t The data represents the current state of the input architecture.

[0145] Secondly, the state-space architecture is subjected to discrete data adaptation using discrete state-space equations; the equation for the state transition matrix A is as follows:

[0146]

[0147] In the formula, n is a row index related to an orthogonal basis function in a polynomial space, and k is a column index related to an orthogonal basis function in the polynomial space.

[0148] In an example of the present application, the state transition matrix A is randomly initialized, and a low-rank matrix after matrix decomposition is used, and the low-rank matrix equation is as follows:

[0149]

[0150] Wherein, P and Q are vectors with a length of N; V represents an eigenvector matrix; V * represents a transposed matrix; Lambda represents an eigenmatrix; and * represents a projection operation.

[0151] In an example of the present application, the expression of the discrete state space equation is as follows:

[0152]

[0153] In the formula, h k and y k respectively represent a memory control unit and an output control unit of a current state in a selected state space architecture; h k-1 represents a memory control unit of a previous state in the selected state space architecture; h0 represents an initialized state of the memory control unit; x k represents data of a current state input architecture; and x0 represents data of an initial state input architecture. All are matrices after A, B and C are discretely adapted.

[0154] In an example of the present application, the matrix is obtained by the following derivation formula:

[0155]

[0156] In the formula, Delta represents a discretization nonlinear function for discretizing a continuous state; and I represents a unit matrix.

[0157] In an example of the present application, in the step S715, the calculation of the features sigma m,pred , F m,pred , P m,pred , epsilon m,pred and T m,pred of the neural network by the data-driven equation and the evolution mechanism equation is specifically implemented as follows:

[0158] First, the data-driven equation is used to calculate the features sigma m,pred , F m,pred , P m,pred , epsilon m,predand T m,pred The calculation is constrained, and the expression is as follows:

[0159]

[0160] In the formula, and respectively represent the deviation calculation equations of stress, fracture, seepage, strain and temperature characteristics; m represents the sample quantity; i represents the serial index of the characteristics; σ m,pred , F m,pred , P m,pred , ∈ m,pred and T m,pred respectively represent the predicted stress, fracture, seepage, strain and temperature characteristics; σ m,true , F m,true , P m,true , ∈ m,true and T m,true respectively represent the real stress, fracture, seepage, strain and temperature characteristics.

[0161] Then, the mechanism equation is used to constrain the characteristics F m of the neural network. The calculation is constrained, and the expression is as follows:

[0162]

[0163] In the formula, represent the mechanism constraint equation of the predicted characteristics; M represents the total number of samples; and α represents a weight coefficient.

[0164] In one example of the present application, in the step S70, a physical field imaging guidance network based on a data mechanism joint guidance architecture is constructed, a large-scale physical field imaging guidance network is trained and verified through a large-scale physical field data set, and a large-scale physical field imaging guidance model is obtained, as shown in Figure 3 The method comprises the following steps:

[0165] S721: The large-scale physical field characteristic guidance learning module extracts the parameter-enhanced stress characteristic σ N , the strain characteristic ∈ N , the micro-current characteristic e N , the fracture characteristic f N , the resistivity characteristic ρ N , the potential characteristic and the wave velocity characteristic v N from the stress measured data σ Nfse , the strain measured data ∈ Nfse , the micro-current measured data e Nfse , the fracture measured data f Nfse , the resistivity measured data ρ Nfse, potential features and wave velocity features v Nfse features; encoded features are concatenated to features obtained by the second linear feature extraction layer B Nb , ∈ Nb , e Nb , f Nb , ρ Nb , and v Nb ; features σ Nse , ∈f Nse , ef Nse , ff Nse , ρf Nse , and vf Nse are regularized by the third Normalization function to obtain features σ Nn , ∈ Nn , e Nn , f Nn , ρ Nn , and v Nn ;

[0166] S722: features σ Nn , ∈ Nn , e Nn , f Nn , ρ Nn , and v Nn are extracted by the fourth linear feature extraction layer A and the fourth activation function layer to obtain σ Na+s , ∈ Na+s , e Na+s , f Na+s , ρ Na+s , and v Na+s are concatenated to features σ Nsssm , ∈ Nsssm , e Nsssm , f Nsssm , ρ Nsssm , and v Nsssm to obtain σ NSSSM , ∈ NSSSM , e NSSSM , f NSSSM , ρ NSSSM , and v NSSSM ; features σ Na , ∈ Na , e Na , fNa 、ρ Na 、 and v a Feature extraction is performed through the second convolutional layer and the third activation function layer to obtain higher-dimensional features σ Nconv+s 、∈ Nconv+s 、e Nconv+s 、f Nconv+s 、ρ Nconv+s 、 and v Nconv+s ;

[0167] S723: Gating calculation is performed on the high-dimensional features through the second selective state space architecture, and the features extracted through the second convolutional layer and the third activation function layer are spliced to obtain features σ NSSSM 、∈ NSSSM 、e NSSSM 、f NSSSM 、ρ NSSSM 、 and v NSSSM ; The second linear feature extraction layer B performs feature extraction on the spliced features, and splices the extracted features with the features of the encoded data by the second residual connection layer to obtain features σf Nb+se 、∈f Nb+se 、ef Nb+se 、ff Nb+se 、ρf Nb+se 、 and vf Nb+se ;

[0168] S724: The spliced features are regularized by the fourth Normalization function to obtain features σf Nn 、∈f Nn 、ef Nn 、ff Nn 、ρf Nn 、 and vf Nn ; The extracted features through the second fully connected layer are calculated, and stress features σ M,pred , fracture features F M,pred , seepage features P M,pred , strain features ∈ M,pred and temperature features T M,pred are calculated;

[0169] S725: The features σ M,pred , F M,pred , P M,pred , ∈ M,pred and T M,pred of the neural network are calculated by the data-driven equation and the evolution mechanism equation, and finally the large-scale physical field imaging guide model is obtained.​

[0170] In one example of the present application, in the step S725, the calculation of the features σ M,pred , F M,pred , P M,pred , ∈ M,pred and T M,pred of the neural network by the data-driven equation and the evolution mechanism equation is constrained as follows:

[0171] First, the calculation of the features σ M,pred , F M,pred , P M,pred , ∈ M,pred and T M,pred of the neural network by the data-driven equation is constrained, and the expression is as follows:

[0172]

[0173] In the formula, and represent the deviation calculation equations of the stress, fracture, seepage, strain and temperature features, respectively; M represents the number of samples; i represents the serial index of the features; σ M,pred , F M,pred , P M,pred , ∈ M,pred and T M,pred represent the predicted stress, fracture, seepage, strain and temperature features, respectively; σ M,true , F M,true , P M,true , ∈ M,true and T M,true represent the real stress, fracture, seepage, strain and temperature features, respectively;

[0174] Then, the calculation of the features F m of the neural network by the mechanism equation is constrained, and the expression is as follows:

[0175]

[0176] In the formula, represents the mechanism constraint equation of the predicted features; M represents the total number of samples; and α represents the weight coefficient.

[0177] In one example of the present application, in the step S80, the multi-physical field coupling imaging network based on the reinforced distillation architecture extracts coupling factors based on data features and optimizes the distillation strategy, so as to integrate the data features of small-scale samples in the large-scale sample physical field imaging process, and obtain a large-scale multi-physical field coupling imaging model, as shown in FIG. 8, including the following steps: Figure 4

[0178] ​S81: Multiphysics Coupled Imaging Networks detect and monitor measured data x in the features of small-scale physical field imaging data. i (stress σ) n , strain ∈ n microcurrent e n , rupture f n resistivity ρ n Potential and wave speed v n The small-scale physical field imaging result p(x) is obtained by performing N rounds of distillation on the measured data. i ;Θ p (Stress field σf) m,pred , fracture field Ff m,pred , seepage field Pf m,pred , strain field ∈ f m,pred and temperature field Tf m,pred The distillation equation is expressed as follows:

[0179]

[0180] In the formula, The equation represents N rounds of distillation for imaging small-scale physical fields; m represents the number of samples; x i Represents the input detection and monitoring measured data; Θ g The weight parameters represent the small-scale physical field feature extraction model; p(x) i ;Θ g ) represents the probability distribution of small-scale physical field imaging data features based on the current weight parameters; Θ p The weight parameters represent the multiphysics coupled imaging network; p(x) i ;Θ p ) represents the probability distribution of small-scale physical field imaging results based on the current weight parameters;

[0181] S82: The distillation strategy of the multi-physics coupled imaging network is optimized through a reinforcement learning architecture. The specific implementation process is as follows:

[0182] First, the multiphysics coupled imaging network performs random sampling on the features of small-scale physical field imaging data. The expression for random sampling is as follows:

[0183]

[0184] in, This represents a subset sampled from the features of small-scale physical field imaging data; x b The last sample in the subset; for each sample

[0185] Then, construct the state vector s of the current sample. i, current state vector s i The expression of s i is as follows:

[0186] s d = {f c , f a , f s}

[0187] In the formula, s i represents the state vector of the current sample; f d represents the sample dimension, f d = Dim; f c represents the prediction confidence, f a represents the source-target consistency, f s represents the feature representation, Pool represents the pooling operation, h i represents the hidden state of the neural network, and U represents the trainable embedding matrix of the probability distribution p(x i ; Θ g ) for the current x i .

[0188] Then, the multi-physical field coupling imaging network adjusts and optimizes the coupling strategy between the detection and monitoring measured data and the multi-physical field data through the policy module π φ of the reinforcement learning architecture, and the expression of the policy module is as follows:

[0189] π φ (s i ) = σ (W2·ReLU (W1·s i +b1) +b2)

[0190] In the formula, π φ (s i ) represents the coupling strategy calculation module for the current state vector; σ represents the Sigmoid activation function; W1 and W2 represent the weight parameters of the coupling strategy calculation module for the current state vector; and b1 and b2 represent the bias terms of the coupling strategy calculation module for the current state vector.

[0191] Then, the multi-physical field coupling imaging network adjusts and optimizes the coupling strategy between the detection and monitoring measured data x i (stress σ n , strain ∈ n , micro-current e n , fracture f n , resistivity ρ n , potential , and wave speed v nM rounds of distillation to obtain small-scale physical field imaging results p(x i ; Θ p ) (stress field σf m,pred , fracture field Ff m,pred , seepage field Pf m,pred , strain field ∈f m,pred and temperature field Tf m,pred ), the expression of the distillation equation is as follows:

[0192]

[0193] wherein, represents the equation for M rounds of distillation of small-scale physical field imaging; m represents the number of samples; x i represents the input detection and monitoring measured data; Θ pM-1 represents the multi-physical field coupling imaging network weight parameters of the M-1th round; p(x i ; Θ pM-1 ) represents the probability distribution of the multi-physical field coupling imaging network based on the current weight parameters; Θ pM represents the weight parameters of the multi-physical field coupling imaging network of the Mth round; p(x i ; Θ pM ) represents the probability distribution of the multi-physical field coupling imaging network based on the current weight parameters.

[0194] Then, the weight parameters of the multi-physical field coupling imaging network are updated by gradient, and the gradient update expression is as follows:

[0195]

[0196] wherein, represents the updated weight parameters of the multi-physical field coupling imaging network of the Mth round; η represents the learning rate; represents the gradient of Θ pM .

[0197] Then, the multi-physical field coupling imaging network is coupled by the reinforcement learning architecture to calculate the coupling factor, and the expression is as follows:

[0198]

[0199] wherein, ψ m represents the small-scale multi-physical field coupling factor.

[0200] Then, the multi-physical field coupling imaging network is optimized by the reinforcement learning architecture to optimize the distillation strategy, and the expression is as follows:

[0201]

[0202] ​where Φ * represents all trainable parameters in the original distillation strategy; γ represents the learning rate; B represents the current training sample batch; represents the gradient of all trainable parameters in the original distillation strategy.

[0203] Finally, the small-scale multi-physical field coupling imaging model is obtained.

[0204] S83: The small-scale multi-physical field coupling imaging model is obtained by N rounds of distillation on the large-scale physical field imaging data features and the detected and monitored measured data X I (stress σ N , strain ∈ N , micro-current e N , fracture f N , resistivity ρ N , potential , and wave speed v N ) to obtain the large-scale physical field imaging results P(X I ; Θ P ) (stress field σf M,pred , fracture field Ff M,pred , seepage field Pf M,pred , strain field ∈f M,pred , and temperature field Tf M,pred ). The expression of the distillation equation is as follows:

[0205]

[0206] where, represents the equation for N rounds of distillation of the large-scale physical field imaging; M represents the number of samples; X I represents the input detected and monitored measured data; Θ G represents the weight parameters of the large-scale physical field feature extraction model; P(x I ; Θ G ) represents the probability distribution of the large-scale physical field imaging data features based on the current weight parameters; Θ P represents the weight parameters of the small-scale multi-physical field coupling imaging model; P(x I ; Θ P represents the probability distribution of the large-scale physical field imaging results based on the current weight parameters.

[0207] S84: The distillation strategy of the multi-physical field coupling imaging network is optimized through the reinforcement learning architecture, and the specific implementation process is as follows:

[0208] First, the small-scale multi-physical field coupling imaging model randomly samples in the large-scale physical field imaging data features, and the expression of random sampling is as follows:

[0209]

[0210] wherein, represents a subset sampled from large-scale physical field imaging data features; X B represents the last sample in the subset; for each sample

[0211] Then, the state vector S of the current sample is constructed I , the expression of the current state vector S I is as follows:

[0212] S I = {f D ,f C ,f A ,f S}

[0213] wherein, s i represents the state vector of the current sample; f D represents the sample dimension, f D = Dim; f C represents the prediction confidence, f A represents the source and target consistency, f S represents the feature representation, Pool represents the pooling operation, h I represents the hidden state of the neural network, and U represents the trainable embedding matrix of the probability distribution P(x I ; Θ G ) for the current X I .

[0214] Then, the small-scale multi-physical field coupling imaging model adjusts and optimizes the coupling strategy between the detection, monitoring and measured data and the multi-physical field data through the policy module π φ of the reinforcement learning architecture, and the expression of the policy module is as follows:

[0215] π φ (S I ) = σ (W2·ReLU (W1·S I +b1) +b2)

[0216] wherein, π φ (S I ) represents the coupling strategy calculation module for the current state vector; σ represents the Sigmoid activation function; W1 and W2 represent the weight parameters of the coupling strategy calculation module of the current state vector; and b1 and b2 represent the bias terms of the coupling strategy calculation module of the current state vector.

[0217] Then, the small-scale multiphysics coupled imaging model detects and monitors the measured data X in the features of large-scale physical field imaging data. I (stress σ) N , strain ∈ N Microcurrent e N , rupture f N resistivity ρ N Potential and wave speed v N The large-scale physical field imaging result P(x) is obtained by performing M rounds of distillation on the measured data. I ;Θ P (Stress field σf) M,pred , fracture field Ff M,pred seepage field Pf M,pred , strain field ∈ f M,pred and temperature field Tf M,pred The distillation equation is expressed as follows:

[0218]

[0219] middle, This represents the equation for M-round distillation in large-scale physical field imaging; M represents the number of samples; X I Represents the input detection and monitoring measured data; Θ PM-1 P(X) represents the weight parameters of the small-scale multiphysics coupled imaging model in the (M-1)th round; I ;Θ PM-1 ) represents the probability distribution of a small-scale multiphysics coupled imaging model based on the current weight parameters; Θ PM The weight parameters of the small-scale multiphysics coupled imaging model in the Mth round; P(X I ;Θ PM ) represents the probability distribution of a small-scale multiphysics coupled imaging model based on the current weight parameters.

[0220] Then, the weight parameters of the small-scale multiphysics coupled imaging model are updated using gradients. The gradient update expression is as follows:

[0221]

[0222] in, η represents the weight parameters of the updated small-scale multiphysics coupled imaging model in the Mth round; η represents the learning rate. Represents Θ PM beg The gradient.

[0223] Then, the small-scale multiphysics coupled imaging model calculates the coupling factor through a reinforcement learning architecture, as shown in the following expression:

[0224]

[0225] wherein, ψ M represents large-scale multi-physical field coupling factor.

[0226] Then, the small-scale multi-physical field coupling imaging model is distilled by the reinforcement learning architecture to optimize the strategy, and the expression is as follows:

[0227]

[0228] wherein, Φ * represents all trainable parameters in the updated distilled strategy; Φ represents all trainable parameters in the original distilled strategy; γ represents the learning rate; B represents the current training sample batch; represents the gradient of all trainable parameters in the original distilled strategy.

[0229] Finally, the large-scale multi-physical field coupling imaging model is obtained.

[0230] According to the second aspect of the present application, a large-scale coal rock multi-physical field imaging system based on iterative reinforcement distillation includes:

[0231] The first data acquisition module is configured to arrange stress and optical fiber micro-nano monitoring sensors on the surface and inside of the small-scale coal rock sample, and arrange micro-current monitoring, multi-frequency acoustic wave monitoring, electrode monitoring, ultrasonic phased array detection, and CT detection sensors on the surface of the small-scale coal rock sample. The test machine performs a loading test on the coal rock sample, and records the mechanical parameters, constitutive equation, loading boundary condition data, and local measurement point position full-time stress, local measurement point position full-time strain, local measurement point position full-time micro-current, full-time and space rupture, local measurement point position full-time resistivity, local measurement point position full-time potential, full-time and space wave velocity, phased time full-space seepage, and fracture (finite physical field data) measured data during the test;

[0232] The first numerical simulation model module is configured to construct a numerical simulation model according to the mechanical parameters, constitutive equation, and loading boundary condition data recorded during the loading test of the small-scale coal rock sample, and obtain full-time and space physical field data (stress field simulation, strain field simulation data, and temperature field simulation data) by calculating the numerical simulation model;

[0233] The first data division module is configured to correspond the collected detection and monitoring data (measured data such as micro-current, fracture, resistivity, potential, wave velocity, stress and strain), finite physical field data (measured data of seepage and fracture) and full-time-space physical field data (simulated data of temperature field, stress field and strain field) by the boundary conditions and time data in the loading test process of the small-scale coal rock sample, to obtain a small-scale full physical field data set, and to divide the small-scale full physical field data set into a training set 1, a verification set 1 and a test set 1.

[0234] The second data acquisition module is configured to arrange stress and optical fiber micro-nano monitoring sensors on the surface and inside of the large-scale coal rock sample, and to arrange micro-current monitoring, multi-frequency acoustic wave monitoring, electrode monitoring, CT detection and ultrasonic phased array detection sensors on the surface of the large-scale coal rock sample. The test machine performs a loading test on the coal rock sample, and records mechanical parameters, constitutive equations, loading boundary condition data and local measurement point position full-time stress, local measurement point position full-time strain, local measurement point position full-time micro-current, full-time-space fracture, local measurement point position full-time resistivity, local measurement point position full-time potential and full-time-space wave velocity and other measured data collected by the monitoring sensors during the test;

[0235] The second numerical simulation model module is configured to construct a numerical simulation model according to the mechanical parameters, constitutive equations and loading boundary condition data recorded in the loading test process of the large-scale coal rock sample, and to obtain full-time-space physical field data (simulated data of stress field, strain field, fracture field, seepage field and temperature field) by calculating the numerical simulation model;

[0236] The second data division module is configured to correspond the collected detection and monitoring data (measured data such as micro-current, fracture, resistivity, potential, wave velocity, stress and strain) and full-time-space physical field data (simulated data of fracture field, seepage field, temperature field, stress field and strain field) by the boundary conditions and time data in the loading test process of the large-scale coal rock sample, to obtain a large-scale physical field data set, and to divide the large-scale physical field data set into a training set 2, a verification set 2, a test set 2 and a test set 3.

[0237] The imaging data feature module is configured to construct a feature extraction network based on a data mechanism joint guidance architecture, to train and verify the feature extraction network by the small-scale and large-scale physical field data sets to obtain small-scale and large-scale physical field feature extraction models respectively, and to input the local measurement point position full-time micro-current, full-time-space fracture, local measurement point position full-time resistivity, local measurement point position full-time temperature, full-time-space wave velocity, local measurement point position full-time stress and local measurement point position full-time strain and other measured data of the test set 1 in the small-scale and large-scale physical field data sets into the small-scale and large-scale physical field imaging guidance models respectively to obtain physical field imaging data features.

[0238] The coupling imaging model module is configured to be used for a large-scale multi-physical field coupling imaging network based on a reinforcement distillation architecture to extract coupling factors through data feature extraction and optimize a distillation strategy, so as to realize the fusion of small-scale sample data features in the large-scale sample physical field imaging process and obtain a large-scale multi-physical field coupling imaging model.

[0239] The imaging result module is configured to be used for inputting local measuring point positions, full-time stresses, full-time strains, full-time micro-currents, full-time and space fractures, full-time resistivities, full-time potentials and full-time and space wave velocities and other measured data of a test set 2 in a small-scale and large-scale physical field data set into the multi-physical field coupling imaging model to obtain full-time and space physical field (stress field, strain field, fracture field, seepage field and temperature field) imaging results of a large-scale coal rock sample.

[0240] The imaging system focuses on realizing large-scale coal rock multi-physical field imaging based on detection and monitoring data in a large-scale coal rock sample loading test process through a large-scale coal rock multi-physical field imaging method and system based on iterative reinforcement distillation. Compared with traditional numerical simulation and neural network methods, this large-scale coal rock multi-physical field imaging method based on iterative reinforcement distillation realizes the fusion of small-scale coal rock sample detection, monitoring and physical field data features in the large-scale coal rock sample fracture imaging process by means of reinforcement learning and knowledge distillation, thereby realizing the organic fusion of multi-scale coal rock sample data collaborative learning and dynamic distillation optimization, significantly improving the physical field imaging spatial resolution of the large-scale coal rock sample, and constructing a new paradigm for coal rock physical field imaging based on a reinforcement distillation architecture.

[0241] The small-scale and large-scale physical feature extraction model in the imaging system is a selective state space neural network based on a data mechanism joint guidance architecture. The network simultaneously takes evolutionary mechanism equations and data-driven equations as constraints, which not only ensures that the output of the guidance model follows the physical law, but also fully excavates the nonlinear features in the multi-source experimental data, thereby exhibiting excellent physical feasibility and robustness under the distribution of test data. The selective state space architecture neural network greatly reduces the dependence on computing resources through fine division and screening of state variables and observation variables, so that the model can efficiently use the data collected in the laboratory coal rock sample loading process and the data obtained by numerical model simulation, and extract high-credibility physical field imaging data features for the large-scale multi-physical field coupling imaging network.

[0242] The imaging system realizes the fusion of small-scale sample data characteristics in the large-scale sample physical field imaging process through knowledge distillation and reinforcement learning, realizes the deep fusion of measured data and numerical simulation results, and ensures that the large-scale fracture imaging result can truly reflect the physical field distribution characteristics of the coal rock sample. At the same time, the equation optimization network based on reinforcement learning calculates the imaging coupling factors between multiple physical fields through the distillation equation and the strategy module, realizes the collaborative innovation of structured knowledge distillation and dynamic coupling strategy optimization, refines and deeply fuses the physical field imaging mechanism knowledge of samples of different scales into the distillation strategy of reinforcement learning, and effectively improves the imaging precision of large-scale coal rock samples based on detection and monitoring data.

[0243] It should be noted that the large-scale coal rock multi-physical field imaging system based on iterative reinforcement distillation of the present application can also perform any processing in the large-scale coal rock multi-physical field imaging method based on iterative reinforcement distillation as previously described, and specific details will not be repeated here.

[0244] The exemplary embodiments of the large-scale coal rock multi-physical field imaging method and system based on iterative reinforcement distillation proposed by the present application are described in detail above with reference to preferred embodiments, however, those skilled in the art can understand that various modifications and modifications can be made to the above specific embodiments without departing from the concept of the present application, and various technical features and structures proposed by the present application can be combined without exceeding the protection scope of the present application, and the protection scope of the present application is determined by the appended claims.

Claims

1. A large-scale coal and rock multiphysics imaging method based on iterative enhanced distillation, characterized in that, Includes the following steps: S10: Stress and fiber optic micro / nano monitoring sensors are arranged on the surface and inside of small-scale coal and rock samples. The coal and rock samples are subjected to loading tests by a testing machine. During the test, mechanical parameters, constitutive equations, loading boundary condition data, and measured data collected by the detection and monitoring sensors are recorded. S20: A numerical simulation model is constructed based on the mechanical parameters, constitutive equations, and loading boundary condition data recorded during the loading test of small-scale coal and rock samples. The full-time and spacetime physical field data are obtained by calculating the numerical simulation model. S30: By matching the boundary conditions and time data during the small-scale coal and rock sample loading test, the collected detection and monitoring data, finite physical field data and full-time and space-time physical field data are matched to obtain the small-scale full physical field dataset, and the small-scale full physical field dataset is divided into training set 1, validation set 1 and test set 1. S40: Stress and fiber optic micro / nano monitoring sensors are arranged on the surface and inside of large-scale coal and rock samples. The coal and rock samples are subjected to loading tests by a testing machine. During the test, mechanical parameters, constitutive equations, loading boundary condition data, and measured data collected by the monitoring sensors are recorded. S50: A numerical simulation model is constructed based on the mechanical parameters, constitutive equations, and loading boundary condition data recorded during the large-scale coal and rock sample loading test. The full-time and spacetime physical field data are obtained by calculating the numerical simulation model. S60: By matching the boundary conditions and time data during the large-scale coal and rock sample loading test with the collected detection and monitoring data and the full-time and space-time physical field data, a large-scale physical field dataset is obtained. The large-scale physical field dataset is divided into training set 2, validation set 2, test set 2 and test set 3. S70: Construct a feature extraction network based on a data mechanism joint guidance architecture. Train and validate the feature extraction network using small-scale and large-scale physical field datasets to obtain small-scale and large-scale physical field feature extraction models, respectively. Input the measured data from test set 1 in the small-scale and large-scale physical field datasets into the small-scale and large-scale physical field imaging guidance models to obtain physical field imaging data features. S80: The large-scale multi-physics coupled imaging network based on the enhanced distillation architecture extracts coupling factors from data features and optimizes the distillation strategy to integrate small-scale sample data features into the large-scale sample physical field imaging process, thus obtaining a large-scale multi-physics coupled imaging model. S90: Input the measured data from test set 3 into the multiphysics coupling imaging model to obtain the full-time and spatiotemporal physical field imaging results of the large-scale coal and rock samples.

2. The large-scale coal and rock multiphysics imaging method based on iterative enhanced distillation according to claim 1, characterized in that, In step S30, the collected detection and monitoring data, finite physical field data, and full-time and space-time physical field data are correlated using the boundary conditions and time data during the small-scale coal and rock sample loading test to obtain a small-scale full physical field dataset, including the following steps: First, the measured data of seepage and fractures will be processed to obtain finite physical field data at different time points; Then, the measured data are aligned according to time to obtain aligned detection and monitoring measured data; the finite physics field data and the full-time and spacetime physics field data are aligned according to time and applied boundary conditions to obtain aligned physics field data; Finally, the aligned detection and monitoring data and the aligned physical field data are aligned according to time and loading boundary conditions to obtain a small-scale full physical field dataset.

3. The large-scale coal and rock multiphysics imaging method based on iterative enhanced distillation according to claim 1, characterized in that, In step S50, a numerical simulation model is constructed based on the mechanical parameters, constitutive equations, and loading boundary condition data recorded during the large-scale coal and rock sample loading test. The full-time and spacetime physical field data are obtained by calculating the numerical simulation model, including the following steps: First, a complete small-scale coal and rock sample model is generated based on the size of the small-scale coal and rock sample using numerical simulation methods. Then, the physical constants of the large-scale coal and rock sample model are set according to the mechanical parameters and constitutive equations of the large-scale coal and rock sample loading test; the boundary conditions of the large-scale coal and rock sample model are set according to the boundary conditions of the large-scale coal and rock sample loading test. Finally, a large-scale coal and rock sample model was calculated to obtain simulation data of fracture field, seepage field, temperature field, stress field and strain field in the large-scale coal and rock loading test.

4. The large-scale coal and rock multiphysics imaging method based on iterative enhanced distillation according to claim 1, characterized in that, In step S70, an extraction network based on a data mechanism joint guidance architecture is constructed. The feature extraction network is trained and validated using a small-scale physical field dataset to obtain a small-scale physical field feature extraction model, including the following steps: S711: The small-scale physical field feature-guided learning module uses the first data encoding layer to process the measured stress data σ n , measured strain data ∈ n Microcurrent measured data e n , measured data of rupture f n Measured resistivity data ρ n Measured potential data And measured wave speed data v n The stress characteristic σ enhanced by parameters was obtained from the measured data. fse Strain characteristics ∈ fse Microcurrent characteristics e fse , fracture characteristics f fse Resistivity characteristics ρ fse Potential characteristics and wave speed characteristics v fse Features; the encoded features are concatenated into the first linear feature extraction layer B through the first residual connection layer to obtain feature σ. b ,∈ b e b f b ρ b , and v b In the middle; the feature σf enhanced by the first Normalization function on the parameters. se ,∈f se ef se ff se ,ρf se , and vf se Regularization is performed to obtain the characteristic σ nn ,∈ nn e nn f nn ρ nn , and v nn ; S712: Regularized feature σ nn ,∈ nn e nn f nn ρ nn , and v nn σ is obtained by feature extraction through the residual connection formed by the second linear feature extraction layer A and the second activation function layer. a+s ,∈ a+s e a+s f a+s ρ a+s , and v a+s The feature σ spliced ​​into the output of the first selective state space architecture sssm ,∈ sssm e sssm f sssm ρ sssm , and v sssm σ is obtained from SSSM ,∈ SSSM e SSSM f SSSM ρ SSSM , and v SSSM The feature σ output by the first linear feature extraction layer A a ,∈ a e a f a ρ a , and v a Higher-dimensional features σ are obtained through feature extraction via the first convolutional layer and the first activation function layer. conv+s ,∈ conv+s e conv+s f conv+s ρ conv+s , and v conv+s ; S713: Gated computation of high-dimensional features is performed through the first selective state-space architecture, and then concatenated with the features extracted by the first convolutional layer and the first activation function layer to obtain feature σ. SSSM ,∈ SSSM e SSSM f SSSM ρ SSSM , and v SSSM The first linear feature extraction layer B extracts features from the concatenated features and concatenates the extracted features with the features of the encoded data obtained by the first residual connection layer to obtain feature σf. b+se ,∈f b+se ef b+se ff b+se ,ρf b+se , and vf b+se ; S714: Regularize the concatenated features using the second Normalization function to obtain feature σf. n ,∈f n ef n ff n ,ρf n , and vf n The stress characteristic σ is calculated by using the features extracted from the first fully connected layer. m,pred Fracturing characteristics F m,pred , seepage characteristics P m,pred Strain characteristics ∈ m,pred and temperature characteristics T m,pred ; S715: Analyzing the characteristic σ of neural networks through data-driven equations and evolutionary mechanism equations. m,pred F m,pred P m,pred ,∈ m,pred and T m,pred The calculations are constrained, and the final result is a small-scale physical field feature extraction model.

5. The large-scale coal and rock multiphysics imaging method based on iterative enhanced distillation according to claim 4, characterized in that, In step S713, the selective state-space architecture is specifically implemented as follows: First, a state-space architecture is constructed using continuous state-space equations, the expressions of which are as follows: h t =Ah t-1 +Bx t the t =Ch t Among them, h t-1 Represents the system state; h t Represents the updated system state; matrix A represents the state transition matrix; matrix B represents the input gating matrix; matrix C represents the output mapping matrix; x t The data represents the current state of the input architecture. Secondly, the state-space architecture is subjected to discrete data adaptation using discrete state-space equations; the equation for the state transition matrix A is as follows: In the formula, n is the row index related to the orthogonal basis functions in the polynomial space, and k is the column index related to the orthogonal basis functions in the polynomial space.

6. The large-scale coal and rock multiphysics imaging method based on iterative enhanced distillation according to claim 1, characterized in that, In step S715, the feature σ of the neural network is analyzed through data-driven equations and evolutionary mechanism equations. m,pred F m,pred P m,pred ,∈ m,pred and T m,pred The specific implementation process for calculating constraints is as follows: First, the data-driven equation applies the network characteristics σ of the neural network. m,pred F m,pred P m,pred ,∈ m,pred and T m,pred The calculation is constrained, and the expression is as follows: In the formula, and The deviation calculation equations represent the stress, crack, seepage, strain, and temperature characteristics, respectively; m represents the number of samples; i represents the index of the feature; σ m,pred F m,pred P m,pred ,∈ m,pred and T m,pred These represent the predicted stress, crack, seepage, strain, and temperature characteristics, respectively; σ m,true F m,true P m,true ,∈ m,true and T m,true These represent the actual stress, crack, seepage, strain, and temperature characteristics, respectively. Then, the mechanistic equation describes the characteristics F of the neural network. m The calculation is constrained, and the expression is as follows: In the formula, The mechanistic constraint equation represents the predicted feature; M represents the total number of samples; α represents the weighting coefficient.

7. The large-scale coal and rock multiphysics imaging method based on iterative enhanced distillation according to claim 1, characterized in that, In step S70, a physical field imaging guidance network based on a data mechanism joint guidance architecture is constructed. The physical field imaging guidance network is trained and validated using a large-scale physical field dataset to obtain a large-scale physical field imaging guidance model, including the following steps: S721: The large-scale physical field feature-guided learning module uses the second data encoding layer to process the measured stress data σ N , measured strain data ∈ N Microcurrent measured data e N , measured data of rupture f N Measured resistivity data ρ N Measured potential data And measured wave speed data v N The stress characteristic σ enhanced by parameters was obtained from the measured data. Nfse Strain characteristics ∈ Nfse Microcurrent characteristics e Nfse , fracture characteristics f Nfse Resistivity characteristics ρ Nfse Potential characteristics and wave speed characteristics v Nfse Features; the encoded features are concatenated to the second linear feature extraction layer B through the second residual connection layer to obtain feature σ. Nb ,∈ Nb e Nb f Nb ρ Nb , and v Nb In the middle; the feature σf enhanced by the third Normalization function. Nse ,∈f Nse ef Nse ff Nse ,ρf Nse , and vf Nse Regularization is performed to obtain the characteristic σ Nn ,∈ Nn e Nn f Nn ρ Nn , and v Nn ; S722: Regularized feature σ Nn ,∈ Nn e Nn f Nn ρ Nn , and v Nn σ is obtained by feature extraction through the residual connection formed by the fourth linear feature extraction layer A and the fourth activation function layer. Na+s ,∈ Na+s e Na+s f Na+s ρ Na+s , and v Na+s The feature σ spliced ​​into the output of the second selective state space architecture Nsssm ,∈ Nsssm e Nsssm f Nsssm ρ Nsssm , and v Nsssm σ is obtained from NSSSM ,∈ NSSSM e NSSSM f NSSSM ρ NSSSM , and v NSSSM The feature σ output by the third linear feature extraction layer A Na ,∈ Na e Na f Na ρ Na , and v a Higher-dimensional features σ are obtained through feature extraction via a second convolutional layer and a third activation function layer. Nconv+s ,∈ Nconv+s e Nconv+s f Nconv+s ρ Nconv+s , and v Nconv+s ; S723: High-dimensional features are gated through a second selective state-space architecture and concatenated with features extracted from the second convolutional layer and the third activation function layer to obtain feature σ. NSSSM ,∈ NSSSM e NSSSM f NSSSM ρ NSSSM , and v NSSSM The second linear feature extraction layer B extracts features from the concatenated features and concatenates these features with the features of the encoded data obtained by the second residual connection layer to obtain feature σf. Nb+se ,∈f Nb+se ef Nb+se ff Nb+se ,ρf Nb+se , and vf Nb+se ; S724: Regularize the concatenated features using the fourth Normalization function to obtain feature σf. Nn ,∈f Nn ef Nn ff Nn ,ρf Nn , and vf Nn The stress characteristic σ is calculated by using the features extracted from the second fully connected layer. M,pred Fracturing characteristics F M,pred , seepage characteristics P M,pred Strain characteristics ∈ M,pred and temperature characteristics T M,pred ; S725: Analyzing the characteristic σ of neural networks through data-driven equations and evolutionary mechanism equations. M,pred F M,pred P M,pred ,∈ M,pred and T M,pred The calculations are constrained, ultimately yielding a large-scale physical field imaging guidance model.

8. The large-scale coal and rock multiphysics imaging method based on iterative enhanced distillation according to claim 1, characterized in that, In step S725, the feature σ of the neural network is analyzed through data-driven equations and evolutionary mechanism equations. M,pred F M,pred P M,pred ,∈ M,pred and T M,pred The specific implementation process for calculating constraints is as follows: First, the data-driven equation applies the network characteristics σ of the neural network. M,pred F M,pred P M,pred ,∈ M,pred and T M,pred The calculation is constrained, and the expression is as follows: In the formula, and The deviation calculation equations represent the stress, crack, seepage, strain, and temperature characteristics, respectively; M represents the number of samples; i represents the index of the feature; σ M,pred F M,pred P M,pred ,∈ M,pred and T M,pred These represent the predicted stress, crack, seepage, strain, and temperature characteristics, respectively; σ M,true F M,true P M,true ,∈ M,true and T M,true These represent actual stress, crack, seepage, strain, and temperature characteristics, respectively. Then, the mechanistic equation describes the characteristics F of the neural network. m The calculation is constrained, and the expression is as follows: In the formula, The mechanistic constraint equation represents the predicted feature; M represents the total number of samples; α represents the weighting coefficient.

9. The large-scale coal and rock multiphysics imaging method based on iterative enhanced distillation according to claim 1, characterized in that, In step S80, the multiphysics coupled imaging network based on the enhanced distillation architecture extracts coupling factors from data features and optimizes the distillation strategy to integrate small-scale sample data features into the large-scale sample physical field imaging process, thereby obtaining a large-scale multiphysics coupled imaging model, including the following steps: S81: Multiphysics Coupled Imaging Networks detect and monitor measured data x in the features of small-scale physical field imaging data. i The small-scale physical field imaging results p(x) were obtained by performing N rounds of distillation. i ;Θ p The distillation equation is expressed as follows: In the formula, The equation represents N rounds of distillation for imaging small-scale physical fields; m represents the number of samples; x i Represents the input detection and monitoring measured data; Θ g The weight parameters represent the small-scale physical field feature extraction model; p(x) i ;Θ g ) represents the probability distribution of small-scale physical field imaging data features based on the current weight parameters; Θ p The weight parameters represent the multiphysics coupled imaging network; p(x) i ;Θ p ) represents the probability distribution of small-scale physical field imaging results based on the current weight parameters; S82: By optimizing the distillation strategy of the multi-physics coupled imaging network through reinforcement learning architecture, a small-scale multi-physics coupled imaging model is obtained. S83: Small-scale multiphysics coupled imaging model detects and monitors measured data X in the features of large-scale physical field imaging data. I The large-scale physical field imaging result P(X) was obtained by performing N rounds of distillation. I ;Θ P The distillation equation is expressed as follows: In the formula, The equation represents N rounds of distillation for large-scale physical field imaging; M represents the number of samples; X I Represents the input detection and monitoring measured data; Θ G The weight parameters represent the large-scale physical field feature extraction model; P(x) I ;Θ G ) represents the probability distribution of large-scale physical field imaging data features based on the current weight parameters; Θ P The weight parameters representing the small-scale multiphysics coupled imaging model; P(x I ;Θ P This represents the probability distribution of large-scale physical field imaging results based on the current weight parameters; S84: A large-scale multiphysics coupled imaging model is obtained by optimizing the distillation strategy of the multiphysics coupled imaging network through a reinforcement learning architecture.

10. A large-scale coal and rock multiphysics imaging system based on iterative enhanced distillation, characterized in that, include: The first data acquisition module is configured to deploy stress and fiber optic micro-nano monitoring sensors on the surface and inside of small-scale coal and rock samples, and to conduct loading tests on the coal and rock samples by a testing machine. During the test, mechanical parameters, constitutive equations, loading boundary condition data, and measured data collected by the detection and monitoring sensors are recorded. The first numerical simulation model module is configured to construct a numerical simulation model based on the mechanical parameters, constitutive equations, and loading boundary condition data recorded during the loading test of small-scale coal and rock samples, and to obtain full-time and spacetime physical field data by calculating the numerical simulation model. The first data partitioning module is configured to correspond the collected detection and monitoring data, finite physical field data and full spatiotemporal physical field data with the boundary conditions and time data during the small-scale coal and rock sample loading test to obtain a small-scale full physical field dataset, and to divide the small-scale full physical field dataset into training set 1, validation set 1 and test set 1. The second data acquisition module is configured to deploy stress and fiber optic micro-nano monitoring sensors on the surface and inside of large-scale coal and rock samples. The testing machine performs loading tests on the coal and rock samples and records mechanical parameters, constitutive equations, loading boundary condition data, and measured data acquired by the monitoring sensors during the test. The second numerical simulation model module is configured to construct a numerical simulation model based on the mechanical parameters, constitutive equations, and loading boundary condition data recorded during the large-scale coal and rock sample loading test, and to obtain full-time and spacetime physical field data by calculating the numerical simulation model. The second data partitioning module is configured to match the collected detection and monitoring data with the full-time and space-time physical field data through the boundary conditions and time data during the large-scale coal and rock sample loading test to obtain the large-scale physical field dataset, and to divide the large-scale physical field dataset into training set 2, validation set 2, test set 2 and test set 3. The imaging data feature module is configured to construct a feature extraction network based on a data mechanism joint guidance architecture. The feature extraction network is trained and validated using small-scale and large-scale physical field datasets to obtain small-scale and large-scale physical field feature extraction models, respectively. The measured data from test set 1 in the small-scale and large-scale physical field datasets are input into the small-scale and large-scale physical field imaging guidance models to obtain physical field imaging data features. The coupled imaging model module is configured to use a large-scale multi-physics coupled imaging network based on an enhanced distillation architecture to extract coupling factors from data features and optimize the distillation strategy, thereby integrating small-scale sample data features into the large-scale sample physical field imaging process and obtaining a large-scale multi-physics coupled imaging model. The imaging results module is configured to input the measured data from the centralized test set 2 of the small-scale and large-scale physical field datasets into the multi-physics coupled imaging model to obtain the full-time and spatiotemporal physical field imaging results of the large-scale coal and rock samples.

Images