Coal rock fracture evolution mechanism optimization method and system based on enhanced distillation architecture

Abstract

The invention discloses a coal rock fracture evolution mechanism optimization method and system based on an intensified distillation architecture. The optimization method comprises the following steps: obtaining coal rock staged moment CT detection and full space-time acoustic emission monitoring data; corresponding CT and acoustic emission data according to moments to obtain a fracture data set; training a neural network based on a data mechanism dual-drive architecture through the fracture data set to obtain a guide model; training a neural network based on an evolution mechanism, a distillation equation and SHAP through data features obtained by the test guide model to obtain an expert model; an optimization factor is obtained by testing the expert model; constructing an equation optimization network based on reinforcement learning; the equation optimization network constructs a symbol data set through operational symbols and variable parameters, and optimizes the equation through a fracture data set and an optimization factor; and solving the optimized equation to obtain a coal rock sample fracture field evolution result, and performing deviation calculation according to the evolution result and CT detection data to adjust network parameters.

Description

Technical Field

[0001] This invention relates to the field of coal and rock monitoring technology, and in particular to an optimization method and system for the evolution mechanism of coal and rock fractures based on an enhanced distillation architecture. Background Technology

[0002] During coal mining, factors such as mining activities disrupt the original stress equilibrium of the coal and rock mass. This process typically involves the rapid and unstable propagation of fractures, which is the root cause of dynamic disasters such as rockbursts and coal and gas outbursts. Therefore, in order to gain a deeper understanding of the mechanisms of coal and rock dynamic disasters and to achieve early warning of these disasters, it is necessary to conduct in-depth research on the evolution mechanism of fracture fields and other related aspects.

[0003] While existing rock mechanics theories and methods attempt to elucidate the evolution of fracture fields in coal and rock masses caused by mining activities, this complex process remains a "black box" that is difficult to measure and reveal. Many challenges remain in optimizing the formation mechanism of fracture fields and revealing their evolutionary process. Currently, industrial CT detection and numerical simulation methods have made some progress. These technologies can be used to reconstruct the internal structure of coal and rock masses, especially showing potential in constructing three-dimensional digital models of discontinuous structures and fracture fields. For example, patent CN118688428B discloses a system for analyzing and evaluating the strength degradation mechanism of freeze-thaw fractured rock masses; patent CN117741107A discloses a device and method for visualizing the seepage-dissolution coupling mechanism of rock fractures; and patent CN112435332A discloses a method for microscopic numerical modeling of fractured coal masses based on CT three-dimensional reconstruction. Nevertheless, these methods still have some limitations. They typically rely on high-specification laboratory conditions and cumbersome model reconstruction methods, which limits their widespread application in practical engineering. Furthermore, these methods cannot achieve the analysis of the fracture evolution mechanism, which may pose certain limitations in actual production processes.

[0004] In recent years, experimental methods combining exploration-monitoring technology and numerical simulation have been applied in laboratories. However, current utilization of this exploration-monitoring data mainly focuses on manual analysis of its temporal variation patterns. Furthermore, relying solely on numerical simulation for fracture field evolution mechanism analysis requires extremely high computational costs; even minor modifications to the model can potentially introduce significant negative weights during calculation. Therefore, how to utilize exploration-monitoring data to achieve full-time and spatiotemporal fracture field evolution and overcome the limitations of traditional methods requires further in-depth research. With the continuous advancement of science and technology, breakthroughs in reinforcement learning and distillation methods have attracted attention for their application in exploring fracture field evolution mechanisms. Reinforcement learning, in particular, has significant advantages in exploring equation parameters, enabling large-scale combination of mechanism equations to find optimal solutions in complex combination spaces. Simultaneously, distillation transfers complex guiding model knowledge to expert models, thereby accelerating inference and achieving higher optimization accuracy. Furthermore, sparse regression algorithms automatically eliminate redundant or noisy features through regularization methods, thus improving the accuracy of exploring equations.

[0005] This provides the possibility of optimizing the fracture evolution mechanism equation using the above methods, but currently there is still a lack of specific feasible ideas and implementation methods. Therefore, it is urgent to carry out actual measurement of probe-monitoring signals during the loading process of coal and rock mass, propose a method and system for optimizing the coal and rock fracture evolution mechanism based on the enhanced distillation framework, and realize the optimization of the fracture evolution mechanism equation based on all-time and space-time acoustic emission monitoring data. This will have a significant promoting effect on revealing the "black box" evolution process of coal and rock mass dynamic disasters and realizing the monitoring and early warning of coal and rock dynamic disasters. Summary of the Invention

[0006] This solution addresses the problems and needs raised above by proposing an optimization method and system for the evolution mechanism of coal and rock fractures based on an enhanced distillation framework. The above technical objectives can be achieved by adopting the following technical features, and other technical effects are also brought about.

[0007] One objective of this invention is to propose an optimization method for the coal and rock fracture evolution mechanism based on an enhanced distillation framework, comprising the following steps:

[0008] S10: CT detection sensors and acoustic emission monitoring sensors are deployed on the coal and rock mass, and the coal and rock samples are placed in a testing machine for loading tests. CT detection data and full-time acoustic emission monitoring data are collected at different stages during the loading process of the coal and rock mass through the CT detection sensors and acoustic emission monitoring sensors. Among them, the CT detection data are full-space fracture parameter data at different stages, and the acoustic emission monitoring data are full-time acoustic emission parameter data.

[0009] S20: Correspond the CT detection data and all-time acoustic emission monitoring data at different moments during the loading process of the coal and rock samples according to time to obtain the fracture parameter dataset, and divide the fracture parameter dataset into training set, validation set and test set; and make the operation symbols and variable parameters required for optimizing the mechanism equation into a symbol dataset.

[0010] S30: The neural network with dual-drive architecture based on data mechanism is trained using the training set, and the trained neural network with dual-drive architecture based on data mechanism is validated using the validation set to obtain the bootstrap model;

[0011] S40: Input the test set into the data features obtained by the guiding model, and train a neural network based on the evolutionary mechanism equation, distillation equation and SHAP to obtain an expert model;

[0012] S50: Optimization factors are obtained by inputting the test set into the expert model;

[0013] S60: Construct an equation optimization network based on a reinforcement learning architecture;

[0014] S70: The equation optimization network optimizes the equations of the fracture evolution mechanism by constructing symbolic datasets, fracture datasets, and optimization factors using operators and variable parameters.

[0015] S80: The evolution results of the coal and rock mass fracture field are obtained by solving the fracture evolution mechanism equation. It is then determined whether the accuracy of the coal and rock mass fracture field evolution results meets the requirements. If the accuracy of the coal and rock mass fracture field evolution results does not meet the requirements, the deviation is calculated using the test set CT detection data, and the parameters of the equation optimization network of the reinforcement learning architecture are adjusted until the accuracy of the coal and rock mass fracture field evolution results meets the requirements.

[0016] Furthermore, the coal and rock fracture evolution mechanism optimization method based on the enhanced distillation architecture according to the present invention may also have the following technical features:

[0017] In one example of the present invention, step S30, training and validating a neural network based on a dual-drive architecture using a crack dataset to obtain a guided model includes the following steps:

[0018] S31: The acoustic emission parameter AE is transmitted through the first data encoding layer. p Data extraction yields parameter-enhanced acoustic emission features AEf se The encoded data is concatenated to the feature extraction layer B through a residual connection layer to obtain the feature AE. b In the middle; acoustic emission characteristics AEf enhanced by the first normalization function. se Regularization is performed to obtain the feature AE n ;

[0019] S32: Regularized Feature AE n Feature extraction AE is obtained through the residual connection formed by the second linear feature extraction layer A and the second activation function layer. a+s The feature AE spliced ​​into the output of the first selective state space architecture sssm Get AE SSSM The first linear feature extraction layer A outputs the feature AE. a Higher-dimensional features (AE) are obtained through feature extraction using the first convolutional layer and the first activation function layer. conv+s ;

[0020] S33: 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 AE. 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 AEf. b+se ;

[0021] S34: Regularize the concatenated features using the second Normalization function to obtain feature AEf. n The features extracted by the first fully connected layer are used for calculation to obtain feature F. f,pred ;

[0022] S35: Analyzing the characteristics F of neural networks through data-driven equations and evolutionary mechanism equations. f,pred The calculations are constrained, and the resulting guided model is obtained.

[0023] In one example of the present invention, in step S35, the feature F of the neural network is processed through data-driven equations and evolutionary mechanism equations. f,pred Calculation constraints:

[0024] First, the data-driven equation describes the characteristics F of the neural network. f,pred The calculation is constrained, and the expression is as follows:

[0025]

[0026] In the formula, The equation representing the deviation calculation; M represents the number of samples; i represents the index of the crack parameter; F f,pred F represents the data characteristics that need to be predicted. f,true Represents real data characteristics;

[0027] Then, the evolutionary mechanism equation applies to the characteristics F of the neural network. MThe calculation is constrained, and the expression is as follows:

[0028]

[0029] In the formula, The constraint equations for predicting fracture parameters represent the following: M represents the total number of acoustic emission sensors; α represents the weighting coefficients of the acoustic emission parameters; AE p,i F represents the acoustic emission parameters monitored by the i-th acoustic emission sensor; f,pred This represents the data characteristics that need to be predicted.

[0030] In one example of the present invention, in step S40, training a neural network based on evolutionary mechanisms, distillation equations, and SHAP using data features obtained from the test-guided model to obtain an expert model includes the following steps:

[0031] S41: The acoustic emission parameter ae is transmitted through the second data encoding layer. p Data extraction yields parameter-enhanced acoustic emission features (AEF). se The encoded data is concatenated to the feature extraction layer B through the second residual connection layer to obtain the feature ae. b In the middle; acoustic emission features aef enhanced by parameter enhancement through a third normalization function. se Regularization is performed to obtain the feature ae n ;

[0032] S42: Regularized feature ae n Feature extraction is performed through the residual connection formed by the fourth linear feature extraction layer A and the fourth activation function layer to obtain ae. a+s The feature ae spliced ​​into the output of the second selective state space architecture sssm ae SSSM The feature ae output by the third linear feature extraction layer A a Higher-dimensional features are obtained through feature extraction using the second convolutional layer and the third activation function layer. conv+s ;

[0033] S43: Gated computation of high-dimensional features is performed through a second selective state-space architecture, and then concatenated with the features extracted by the second convolutional layer and the third activation function layer to obtain feature ae. SSSM 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 aef. b+se ;

[0034] S44: Regularize the concatenated features using the fourth Normalization function to obtain the feature aef. n; The features extracted by the second fully connected layer are calculated to obtain the feature f M,pred ;

[0035] S45: The feature f of the neural network is constrained by the distillation equation M,pred to obtain the data feature f through calculation f , and finally an expert model is obtained.

[0036] In an example of the present invention, in the step S45, the feature f of the neural network by the distillation equation M,pred is constrained by calculation to obtain the data feature f f , and the specific implementation process is as follows:

[0037] The data features of the evolution mechanism equation are extracted into the data-driven equation through the distillation equation, and the specific expression is as follows:

[0038]

[0039]

[0040] In the formula, β and γ respectively represent the training weights of the evolution mechanism equation and the data-driven equation; represents the evolution mechanism equation; represents the data-driven equation; 0% < epoch ≤ 50% represents the first half of the training process, and 50% < epoch ≤ 100% represents the second half of the training process; represents the distillation equation.

[0041] In an example of the present invention, in the step S50, an optimization factor is obtained by inputting the test set into the expert model, and the specific implementation process is as follows:

[0042] An optimization factor is obtained through the optimization factor generation equation based on SHAP. Among them, the expression of the optimization factor is as follows:

[0043]

[0044] In the formula, ψ i represents the optimization factor; M represents the number of input data features; represents the set composed of all permutations and combinations of M features, represents the set of feature combinations that are all before the feature k in the permutation R, f x (S) = E(f(X)|X s = x S ) represents the conditional expectation given the known feature subset S; i represents the optimization factor serial number.

[0045] In one example of the present invention, in step S60, an equation optimization network based on a reinforcement learning architecture is constructed, and the specific implementation process is as follows:

[0046] S61: The equation optimization network calculates the reward value of the initial mechanism equation through a reward function, the expression of which is as follows:

[0047]

[0048] In the formula, R represents the reward function; β represents the hyperparameter, which controls the influence of the complexity of the mechanism equation on the optimization factor; This represents the fitting error for all observations.

[0049] S62: The equation optimization network generates a generation strategy through the risk-preference policy gradient function optimization mechanism. The expression for the risk-preference policy gradient function is as follows:

[0050] J risk (θ;∈)=Ε τ～p(τ|θ) [R(τ)|R(τ)≥R ∈ (θ)]

[0051]

[0052] Among them, J risk (θ;∈) represents the risk preference strategy gradient function; θ represents the network parameters of the mechanism equation generation network; ∈ represents the risk preference parameters, used to select the proportion of high-reward advanced mechanism equations; N represents the number of advanced mechanism equations in the current batch; R(τ (i) ) represents the advanced mechanism equation τ (i) The reward value is calculated by the reward function; The threshold representing the reward value (1-∈) in the current batch's advancement mechanism equation; 1 {·} p(τ) represents an exponential function, taking the value 1 if the current condition is true, and 0 otherwise; (i) |θ) represents the network parameters θ that generate the mechanism equation, which in turn generate the advanced mechanism equation τ. (i) The probability, T represents the length of the mechanistic equation; logp(τ) (i) |θ) represents the logarithm of the probability; Represents logp(τ) (i) The partial derivative of |θ) with respect to the parameter θ.

[0053] S63: The optimization factor is input into the equation optimization network. The equation optimization network optimizes the generation strategy of the mechanism equation generation network through the reward function and the risk preference policy gradient function. The expression of the generation strategy of the advanced mechanism equation generation network is as follows:

[0054]

[0055] Where, θ * The updated mechanism equations generate network parameters; η represents the learning rate.

[0056] In one example of the present invention, in step S70, the equation optimization network optimizes the equation by constructing a symbol dataset, a gap dataset, and an optimization factor using operators and variable parameters. The specific implementation process is as follows:

[0057] In step S70, the equation optimization network optimizes the equation by constructing a symbol dataset, a gap dataset, and an optimization factor using operators and variable parameters. The specific implementation process is as follows:

[0058] S71: Transmit the operator O through the third data encoding layer m and variable parameter S n Extracting the enhanced operator feature of Of e and variable parameter features Sf e ;

[0059] S72: The encoded data is concatenated to the third feature extraction layer B through the third residual connection layer to obtain feature O. b and S b middle;

[0060] S73: The characteristics of the operator Of are determined by the fifth Normalization function. e and variable parameter features Sfe e Regularization yields feature O n and S n ;

[0061] S74: Regularized Feature O n and S n Feature O is obtained by performing feature extraction through the residual connection formed by the sixth linear feature extraction layer A and the sixth activation function layer. a+s and S a+s Feature O spliced ​​into the output of the third selective state space architecture sssm and S sssm O SSSM and S SSSM ;

[0062] S75: Feature O output by the fifth linear feature extraction layer A a and S a Higher-dimensional features O are obtained through feature extraction via the third convolutional layer and the fifth activation function layer. conv+s and S conv+s ;

[0063] S76: High-dimensional features are gated through a third selective state-space architecture and concatenated with features extracted from the third convolutional layer and the fifth activation function layer to obtain feature O. SSSM and S SSSM ;

[0064] S77: The third 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 third residual connection layer to obtain feature Of. b+se and Sf b+se ;

[0065] S78: Regularize the concatenated features using the sixth Normalization function to obtain the feature Of. n and Sf n ;

[0066] S79: The features extracted by the third fully connected layer are used to calculate the initial mechanism equation; the crack dataset is used to calculate the coefficients of the initial mechanism equation through a sparse regression algorithm, and then the advanced mechanism equation is obtained.

[0067] In one example of the present invention, in step S79, the coefficients of the initial mechanistic equation are calculated using a sparse regression algorithm, thereby obtaining the advanced mechanistic equation. The expression of the sparse regression algorithm is as follows:

[0068]

[0069] In the formula, These represent the optimal coefficients in the initial mechanism equation; The coefficient vector is represented by Θ(F,AE); Θ(F,AE) represents the initial equation library, which converts the observed F(AE,t) and its derivative information into numerical features, thereby realizing the coefficient vector. F can be obtained by performing matrix multiplication. t The prediction results; F t The time derivative in the representative mechanism equation Numerical approximation; objective function The value represents the fitting error between the prediction and the true value; λ represents the regularization parameter, used to suppress the coefficient vector. This represents the regularization term, which ensures the simplicity of the initial mechanism equations; Represents the squared norm, used to measure the coefficient vector. Size.

[0070] Another objective of this invention is to propose an optimization system for the coal and rock fracture evolution mechanism based on an enhanced distillation architecture, comprising:

[0071] The data acquisition module is configured to deploy CT detection sensors and acoustic emission monitoring sensors on the coal and rock mass, and place the coal and rock sample in the testing machine for loading tests. The CT detection sensors and acoustic emission monitoring sensors acquire CT detection data and full-time and space-time acoustic emission monitoring data at different stages during the loading process of the coal and rock mass. Among them, the CT detection data is the full-space fracture parameter data at different stages, and the acoustic emission monitoring data is the full-time and space-time acoustic emission parameter data.

[0072] The symbolic data module is configured to map the CT detection data and all-time acoustic emission monitoring data at different stages during the loading process of coal and rock samples according to time to obtain a fracture parameter dataset, and divide the fracture parameter dataset into a training set, a validation set, and a test set; and to create a symbolic dataset of the operators and variable parameters required for optimizing the mechanism equation.

[0073] The bootstrap model module is configured to train a neural network with a dual-drive architecture based on a training set, and to validate the trained neural network with a dual-drive architecture based on a validation set to obtain a bootstrap model.

[0074] The expert model module is configured to input the test set into the data features obtained by the guide model, and train a neural network based on the evolutionary mechanism equation, distillation equation and SHAP to obtain the expert model from the data features;

[0075] The optimization factor module is configured to obtain optimization factors by inputting the test set into the expert model.

[0076] Optimize the network module and configure it for building an equation optimization network based on a reinforcement learning architecture;

[0077] The equation optimization module is configured as an equation optimization network to optimize the equations of fracture evolution mechanism by constructing symbol datasets, fracture datasets, and optimization factors through operators and variable parameters.

[0078] The parameter adjustment module is configured to obtain the evolution results of the coal and rock mass fracture field by solving the fracture evolution mechanism equation, determine whether the accuracy of the coal and rock mass fracture field evolution results meets the requirements, and when the accuracy of the coal and rock mass fracture field evolution results does not meet the requirements, it calculates the deviation using CT detection data from the test set and adjusts the parameters of the equation optimization network of the reinforcement learning architecture until the accuracy of the coal and rock mass fracture field evolution results meets the requirements.

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

[0080] 1. This invention focuses on optimizing the coal and rock fracture evolution mechanism equations through a method and system based on a reinforced distillation architecture. Compared to traditional numerical simulation and neural network methods, this method utilizes knowledge distillation and SHAP to extract equation optimization factors from fracture evolution characteristics. Simultaneously, it leverages the reward function of reinforcement learning to dynamically fine-tune the fracture evolution mechanism equation coefficients based on feedback from full-space fracture parameters and full-temporal acoustic emission parameters at various time points. This achieves coal and rock fracture evolution mechanism optimization based on experimental data, improving the accuracy of fracture parameter prediction.

[0081] 2. The guided model in this invention is a selective state-space neural network based on a dual-driven data mechanism. The network uses both evolutionary mechanism equations and data-driven equations as constraints, ensuring that the output of the guided model follows physical laws while fully exploiting the nonlinear characteristics of multi-source experimental data. This demonstrates excellent physical feasibility and robustness under various experimental data distributions. Simultaneously, the selective state-space structure significantly reduces computational resource dependence through fine-grained partitioning and filtering of state and observation variables. This allows the model to efficiently utilize acoustic emission and fracture parameter data collected during the loading of laboratory coal and rock samples, providing expert models with highly reliable data features.

[0082] 3. This invention achieves the transfer of fundamental physical laws from the evolution mechanism equation to the optimized fracture evolution mechanism through distillation equations and reinforcement learning, ensuring that the optimized fracture evolution mechanism follows fundamental physical laws. Simultaneously, the reinforcement learning-based equation optimization network adjusts the optimization strategy of the evolution mechanism equation through reward functions, risk-preference policy gradient functions, and optimization factors extracted by SHAP. This achieves synergistic innovation between structured knowledge distillation and dynamic policy optimization, extracting and deeply integrating domain physics knowledge into the reinforcement learning framework, effectively improving the interpretability and accuracy of the fracture evolution mechanism model.

[0083] The preferred embodiments of the invention will be described in more detail below with reference to the accompanying drawings, so as to facilitate an understanding of the features and advantages of the invention. Attached Figure Description

[0084] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. The drawings are merely illustrative of some embodiments of the present invention and are not intended to limit the scope of the present invention to all embodiments.

[0085] Figure 1 This is a flowchart of an optimization method for the coal and rock fracture evolution mechanism based on an enhanced distillation architecture, according to an embodiment of the present invention.

[0086] Figure 2This is a schematic diagram of a neural network structure based on a data mechanism dual-drive architecture according to an embodiment of the present invention;

[0087] Figure 3 This is a schematic diagram of the result of the distillation neural network according to an embodiment of the present invention;

[0088] Figure 4 To optimize the network structure diagram of the equation according to an embodiment of the present invention;

[0089] Figure 5 This is a schematic diagram of the arrangement of coal and rock samples in a testing machine according to an embodiment of the present invention.

[0090] List of reference numerals in the attached diagram:

[0091] Testing machine 100;

[0092] Loading device 110;

[0093] Acoustic emission sensor 120;

[0094] Coupler 130;

[0095] Preamplifier 140;

[0096] X-ray detector 150;

[0097] X-ray source 160;

[0098] 200 coal and rock samples;

[0099] Lateral direction X;

[0100] The vertical direction is Y. Detailed Implementation

[0101] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The same reference numerals in the drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0102] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, “an” or “a” and similar terms do not necessarily indicate a quantity limitation. Terms such as “comprising” or “including” mean that the element or object preceding the word encompasses the element or object listed following the word and its equivalents, without excluding other elements or objects. Terms such as “connected” or “linked” are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as “upper,” “lower,” “left,” and “right” are used only to indicate relative positional relationships; these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0103] According to a first aspect of the present invention, a method for optimizing the evolution mechanism of coal and rock fractures based on an enhanced distillation framework is provided, such as... Figure 1 As shown. It includes the following steps:

[0104] S10: CT detection sensors and acoustic emission monitoring sensors are deployed on the coal and rock mass, and the coal and rock sample is placed in a testing machine for loading tests. CT detection data and full-time acoustic emission monitoring data are collected at various stages during the loading process using the CT detection sensors and acoustic emission monitoring sensors. The CT detection data represents the full-space fracture parameter data at each stage, and the acoustic emission monitoring data represents the full-time acoustic emission parameter data. (The text repeats itself here, so the translation will only include the first instance.) Figure 5As shown, the testing machine 100 includes: a loading device 110, an acoustic emission sensor 120, a coupler 130, a preamplifier 140, an X-ray detector 150, an X-ray source 160, and a test specimen 170; wherein, the loading device 110, acoustic emission sensor 120, coupler 130, preamplifier 140, X-ray detector 150, and X-ray source 160 are all installed in the test specimen 170, and the coal and rock sample 200 is disposed in the loading device 110. Acoustic emission sensors 120 are installed on both sides of the coal and rock sample 200 in the transverse direction X, wherein the acoustic emission sensors 120 are electrically connected to the coupler 130 and the preamplifier 140, and X-ray detectors 150 and X-ray sources 160 are respectively disposed on both sides of the coal and rock sample 200 in the longitudinal direction Y, wherein the transverse direction X and the longitudinal direction Y are perpendicular to each other. The working process is as follows: The loading device 110 loads the coal and rock sample 200. During this process, the X-ray source 160 emits X-rays and the X-ray detector 150 receives the X-rays, thereby obtaining CT detection data during the fracturing process of the coal and rock sample 200. At the same time, the acoustic emission sensor 120 obtains all-time and all-space acoustic emission monitoring data.

[0105] S20: Correspond the CT detection data and all-time acoustic emission monitoring data at different moments during the loading process of the coal and rock samples according to time to obtain the fracture parameter dataset, and divide the fracture parameter dataset into training set, validation set and test set; and make the operation symbols and variable parameters required for optimizing the mechanism equation into a symbol dataset.

[0106] S30: The neural network with dual-drive architecture based on data mechanism is trained using the training set, and the trained neural network with dual-drive architecture based on data mechanism is validated using the validation set to obtain the bootstrap model;

[0107] S40: Input the test set into the data features obtained by the guiding model, and train a neural network based on the evolutionary mechanism equation, distillation equation and SHAP to obtain an expert model;

[0108] S50: Optimization factors are obtained by inputting the test set into the expert model;

[0109] S60: Construct an equation optimization network based on a reinforcement learning architecture;

[0110] S70: The equation optimization network optimizes the equations of the fracture evolution mechanism by constructing symbolic datasets, fracture datasets, and optimization factors using operators and variable parameters.

[0111] S80: The evolution result of the coal-rock mass fracture field is obtained by solving the fracture evolution mechanism equation. The accuracy of the evolution result is then assessed. If the accuracy meets the requirements, the calculation ends. If the accuracy does not meet the requirements, bias calculations are performed using CT detection data from the test set, and the parameters of the reinforcement learning architecture's equation optimization network are adjusted until the accuracy of the coal-rock mass fracture field evolution result meets the requirements. In other words, the evolution result of the coal-rock mass fracture field is obtained by solving the optimized equation, and the network parameters are adjusted by calculating the bias between the evolution result and CT detection data.

[0112] This optimization method focuses on optimizing the coal and rock fracture evolution mechanism equations through a reinforcement distillation-based approach and system. Compared to traditional numerical simulation and neural network methods, this reinforcement distillation-based optimization method extracts equation optimization factors from fracture evolution characteristics using knowledge distillation and SHAP. Simultaneously, it utilizes the reward function of reinforcement learning to dynamically fine-tune the fracture evolution mechanism equation coefficients based on feedback from full-space fracture parameters and full-temporal acoustic emission parameters at various time points. This achieves coal and rock fracture evolution mechanism optimization based on experimental data, improving the accuracy of fracture parameter prediction.

[0113] The guided model in this optimization method is a selective state-space neural network based on a dual-driven data mechanism. The network uses both evolutionary mechanism equations and data-driven equations as constraints, ensuring that the output of the guided model follows physical laws while fully exploiting the nonlinear characteristics of multi-source experimental data. This demonstrates excellent physical feasibility and robustness under various experimental data distributions. Furthermore, the selective state-space structure significantly reduces computational resource dependence through fine-grained partitioning and filtering of state and observation variables. This allows the model to efficiently utilize acoustic emission and fracture parameter data collected during the loading of laboratory coal and rock samples, providing expert models with highly reliable data features.

[0114] This optimization method achieves the transfer of fundamental physical laws from the evolution mechanism equation to the optimized fracture evolution mechanism through equation distillation and reinforcement learning, ensuring that the optimized fracture evolution mechanism follows fundamental physical laws. Simultaneously, the reinforcement learning-based equation optimization network adjusts the optimization strategy of the evolution mechanism equation through the reward function, risk-preference policy gradient function, and optimization factors extracted by SHAP. This achieves synergistic innovation between structured knowledge distillation and dynamic policy optimization, extracting and deeply integrating domain physics knowledge into the reinforcement learning framework, effectively improving the interpretability and accuracy of the fracture evolution mechanism model.

[0115] In one example of the present invention, step S20, which involves creating a symbol dataset from the operators of the mechanistic equations, includes the following steps:

[0116] First, organize the operators needed for optimizing the mechanism equations;

[0117] Then, the operators are created as a symbolic dataset with a binary tree structure. The symbolic dataset contains the following:

[0118]

[0119] In the formula, OS represents a symbolic dataset; + represents the addition operator; - represents the addition operator; * represents the addition operator; / represents the division operator; ^2 represents the square operator; ^3 represents the cube operator; Represents the first-order partial derivative operator; Represents the second-order partial derivative operator; represents the third-order partial derivative operator; AE represents acoustic emission variable parameter data; F represents fracture variable parameter data.

[0120] In one example of the present invention, in step S30, as Figure 2 As shown, the process of training and validating a guided model using a data mechanism-based dual-drive architecture neural network through a crack dataset includes the following steps:

[0121] S31: The acoustic emission parameter AE is transmitted through the first data encoding layer. p Data extraction yields parameter-enhanced acoustic emission features AEf se The encoded data is concatenated to the feature extraction layer B through a residual connection layer to obtain the feature AE. b In the middle; acoustic emission characteristics AEf enhanced by the first normalization function. se Regularization is performed to obtain the feature AE n ;

[0122] S32: Regularized Feature AE n Feature extraction AE is obtained through the residual connection formed by the second linear feature extraction layer A and the second activation function layer. a+s The feature AE spliced ​​into the output of the first selective state space architecture sssm Get AE SSSM The first linear feature extraction layer A outputs the feature AE. a Higher-dimensional features (AE) are obtained through feature extraction using the first convolutional layer and the first activation function layer. conv+s ;

[0123] S33: 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 AE. SSSMThe 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 AEf. b+se ;

[0124] S34: Regularize the concatenated features using the second Normalization function to obtain feature AEf. n The features extracted by the first fully connected layer are used for calculation to obtain feature F. f,pred ;

[0125] S35: Analyzing the characteristics F of neural networks through data-driven equations and evolutionary mechanism equations. f,pred The calculations are constrained, and the resulting guided model is obtained.

[0126] In one example of the present invention, in step S33, the selective state-space architecture is specifically implemented as follows:

[0127] First, a state-space architecture is constructed using continuous state-space equations, the expressions of which are as follows:

[0128] h t =Ah t-1 +Bx t

[0129] y t =Ch t

[0130] 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.

[0131] 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:

[0132]

[0133] 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.

[0134] In one example of the present invention, the method further includes: randomly initializing the state transition matrix A and using a low-rank matrix obtained from matrix decomposition, the equation of which is as follows:

[0135]

[0136] Where P and Q are both vectors of length N; V represents the eigenvector matrix; V * Λ represents the transpose matrix; Λ represents the eigenvalue matrix; * represents the projection operation.

[0137] In one example of the present invention, the expression of the discrete state-space equation is as follows:

[0138]

[0139]

[0140] In the formula, h k and y k These represent the memory control unit and the output control unit, respectively, for selecting the current state in the state-space architecture; h k-1 h0 represents the memory control unit that selects the previous state in the state space architecture; h0 represents the initialization state of the memory control unit; x k x0 represents the data of the current state input architecture; x0 represents the data of the initial state input architecture. All are matrices after discrete adaptation of A, B, and C.

[0141] In one example of the invention, the matrix It is obtained from the following derivation formula:

[0142]

[0143] In the formula, Δ represents the discretized nonlinear function, which discretizes the continuous state; Ι represents the identity matrix.

[0144] In one example of the present invention, in step S35, the feature F of the neural network is processed through data-driven equations and evolutionary mechanism equations. f,pred Calculation constraints:

[0145] First, the data-driven equation describes the characteristics F of the neural network. f,pred The calculation is constrained, and the expression is as follows:

[0146]

[0147] In the formula, The equation representing the deviation calculation; M represents the number of samples; i represents the index of the crack parameter; F f,pred F represents the data characteristics that need to be predicted. f,true Represents real data characteristics;

[0148] Then, the evolutionary mechanism equation applies to the characteristics F of the neural network. M The calculation is constrained, and the expression is as follows:

[0149]

[0150] In the formula, The constraint equations for predicting fracture parameters represent the following: M represents the total number of acoustic emission sensors; α represents the weighting coefficients of the acoustic emission parameters; AE p,i h represents the acoustic emission parameters monitored by the i-th acoustic emission sensor. f,pred This represents the data characteristics that need to be predicted.

[0151] In one example of the present invention, in step S40, as Figure 3 As shown, training an expert model using data features obtained through testing and guided modeling, based on evolutionary mechanisms, distillation equations, and SHAP, involves the following steps:

[0152] S41: The acoustic emission parameter ae is transmitted through the second data encoding layer. p Data extraction yields parameter-enhanced acoustic emission features (AEF). se The encoded data is concatenated to the feature extraction layer B through the second residual connection layer to obtain the feature ae. b In the middle; acoustic emission features aef enhanced by parameter enhancement through a third normalization function. se Regularization is performed to obtain the feature ae n ;

[0153] S42: Regularized feature ae n Feature extraction is performed through the residual connection formed by the fourth linear feature extraction layer A and the fourth activation function layer to obtain ae. a+s The feature ae spliced ​​into the output of the second selective state space architecture sssm ae SSSM The feature ae output by the third linear feature extraction layer A a Higher-dimensional features are obtained through feature extraction using the second convolutional layer and the third activation function layer. conv+s ;

[0154] S43: Gated computation of high-dimensional features is performed through a second selective state-space architecture, and then concatenated with the features extracted by the second convolutional layer and the third activation function layer to obtain feature ae. SSSM 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 aef. b+se ;

[0155] S44: Regularize the concatenated features using the fourth Normalization function to obtain the feature aef.n ; The features extracted by the second fully connected layer are calculated to obtain the feature f M,pred ;

[0156] S45: The feature f of the neural network is constrained through the distillation equation M,pred The calculation is constrained to obtain the data feature f f , and finally the expert model is obtained.

[0157] In an example of the present invention, in the step S45, the feature f of the neural network through the distillation equation M,pred The calculation is constrained to obtain the data feature f f , and the specific implementation process is as follows:

[0158] The data features of the evolution mechanism equation are extracted into the data-driven equation through the distillation equation, and the specific expression is as follows:

[0159]

[0160] In the formula, β and γ respectively represent the training weights of the evolution mechanism equation and the data-driven equation; represents the evolution mechanism equation; represents the data-driven equation; 0% < epoch ≤ 50% represents the first half of the training process, 50% < epoch ≤ 100% represents the second half of the training process; represents the distillation equation.

[0161] In an example of the present invention, in the step S50, the optimization factor is obtained by inputting the test set into the expert model, and the specific implementation process is as follows:

[0162] The optimization factor is obtained through the optimization factor generation equation based on SHAP, and the expression of the optimization factor is as follows:

[0163]

[0164] In the formula, ψ i represents the optimization factor; M represents the number of input data features; represents the set composed of all permutations and combinations of M features, represents the set of feature combinations before feature k in all permutations R, f x (S)=E(f(X)|X s =x S ) represents the conditional expectation given the known feature subset S; i represents the optimization factor serial number.

[0165] In an example of the present invention, in the step S70, as Figure 4As shown, the equation optimization network optimizes the equation by constructing a symbolic dataset, a gap dataset, and an optimization factor using operators and variable parameters. The specific implementation process is as follows:

[0166] In step S70, the equation optimization network optimizes the equation by constructing a symbol dataset, a gap dataset, and an optimization factor using operators and variable parameters. The specific implementation process is as follows:

[0167] S71: Transmit the operator O through the third data encoding layer m and variable parameter S n Extracting the enhanced operator feature of Of e and variable parameter features Sf e ;

[0168] S72: The encoded data is concatenated to the third feature extraction layer B through the third residual connection layer to obtain feature O. b and S b middle;

[0169] S73: The characteristics of the operator Of are determined by the fifth Normalization function. e and variable parameter features Sfe e Regularization yields feature O n and S n ;

[0170] S74: Regularized Feature O n and S n Feature O is obtained by performing feature extraction through the residual connection formed by the sixth linear feature extraction layer A and the sixth activation function layer. a+s and S a+s Feature O spliced ​​into the output of the third selective state space architecture sssm and S sssm O SSSM and S SSSM ;

[0171] S75: Feature O output by the fifth linear feature extraction layer A a and S a Higher-dimensional features O are obtained through feature extraction via the third convolutional layer and the fifth activation function layer. conv+s and S conv+s ;

[0172] S76: High-dimensional features are gated through a third selective state-space architecture and concatenated with features extracted from the third convolutional layer and the fifth activation function layer to obtain feature O. SSSM and S SSSM ;

[0173] S77: The third 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 third residual connection layer to obtain feature Of. b+se and Sf b+se ;

[0174] S78: Regularize the concatenated features using the sixth Normalization function to obtain the feature Of. n and Sf n ;

[0175] S79: The features extracted by the third fully connected layer are used to calculate the initial mechanism equation; the crack dataset is used to calculate the coefficients of the initial mechanism equation through a sparse regression algorithm, and then the advanced mechanism equation is obtained.

[0176] In one example of the present invention, in step S79, the coefficients of the initial mechanistic equation are calculated using a sparse regression algorithm, thereby obtaining the advanced mechanistic equation. The expression of the sparse regression algorithm is as follows:

[0177]

[0178] In the formula, These represent the optimal coefficients in the initial mechanism equation; The coefficient vector is represented by Θ(F,AE); Θ(F,AE) represents the initial equation library, which converts the observed F(AE,t) and its derivative information into numerical features, thereby realizing the coefficient vector. F can be obtained by performing matrix multiplication. t The prediction results; F t The time derivative in the representative mechanism equation Numerical approximation; objective function The value represents the fitting error between the prediction and the true value; λ represents the regularization parameter, used to suppress the coefficient vector. This represents the regularization term, which ensures the simplicity of the initial mechanism equations; Represents the squared norm, used to measure the coefficient vector. Size.

[0179] In one example of the present invention, in step S70, the equation optimization network optimizes the equation using optimization factors, and the specific implementation process is as follows:

[0180] S701: The equation optimization network calculates the reward value of the initial mechanism equation through a reward function, the expression of which is as follows:

[0181]

[0182] In the formula, R represents the reward function; β represents the hyperparameter, which controls the influence of the complexity of the mechanism equation on the optimization factor; This represents the fitting error for all observations.

[0183] S702: The equation optimization network generates a generation strategy through the risk-preference policy gradient function optimization mechanism. The expression for the risk-preference policy gradient function is as follows:

[0184] J risk (θ;∈)=Ε τ～p(τ|θ) [R(τ)|R(τ)≥R ∈ (θ)]

[0185]

[0186] Among them, J risk (θ;∈) represents the risk preference strategy gradient function; θ represents the network parameters of the mechanism equation generation network; ∈ represents the risk preference parameters, used to select the proportion of high-reward advanced mechanism equations; N represents the number of advanced mechanism equations in the current batch; R(τ (i) ) represents the advanced mechanism equation τ (i) The reward value is calculated by the reward function; The threshold representing the reward value (1-∈) in the current batch's advancement mechanism equation; 1 {·} p(τ) represents an exponential function, taking the value 1 if the current condition is true, and 0 otherwise; (i) |θ) represents the network parameters θ that generate the mechanism equation, which in turn generate the advanced mechanism equation τ. (i) The probability, T represents the length of the mechanistic equation; log p(τ) (i) |θ) represents the logarithm of the probability; Represents log p(τ) (i) The partial derivative of |θ) with respect to the parameter θ.

[0187] S703: The optimization factor is input into the equation optimization network. The equation optimization network optimizes the generation strategy of the mechanism equation generation network through the reward function and the risk preference policy gradient function. The expression of the generation strategy of the advanced mechanism equation generation network is as follows:

[0188]

[0189] Where, θ * The updated mechanism equations generate network parameters; η represents the learning rate.

[0190] In one example of the present invention, in step S80, the deviation calculation formula is as follows:

[0191]

[0192] In the formula, RMSE represents the deviation calculation equation; M represents the number of samples; i represents the index of the crack parameter; F M,pred F represents the predicted value of the Mth sample; M,true The value represents the true value of the Mth sample.

[0193] According to a second aspect of the present invention, a coal and rock fracture evolution mechanism optimization system based on an enhanced distillation architecture includes:

[0194] The data acquisition module is configured to deploy CT detection sensors and acoustic emission monitoring sensors on the coal and rock mass, and place the coal and rock sample in the testing machine for loading tests. The CT detection sensors and acoustic emission monitoring sensors acquire CT detection data and full-time and space-time acoustic emission monitoring data at different stages during the loading process of the coal and rock mass. Among them, the CT detection data is the full-space fracture parameter data at different stages, and the acoustic emission monitoring data is the full-time and space-time acoustic emission parameter data.

[0195] The symbolic data module is configured to map the CT detection data and all-time acoustic emission monitoring data at different stages during the loading process of coal and rock samples according to time to obtain a fracture parameter dataset, and divide the fracture parameter dataset into a training set, a validation set, and a test set; and to create a symbolic dataset of the operators and variable parameters required for optimizing the mechanism equation.

[0196] The bootstrap model module is configured to train a neural network with a dual-drive architecture based on a training set, and to validate the trained neural network with a dual-drive architecture based on a validation set to obtain a bootstrap model.

[0197] The expert model module is configured to input the test set into the data features obtained by the guide model, and train a neural network based on the evolutionary mechanism equation, distillation equation and SHAP to obtain the expert model from the data features;

[0198] The optimization factor module is configured to obtain optimization factors by inputting the test set into the expert model.

[0199] Optimize the network module and configure it for building an equation optimization network based on a reinforcement learning architecture;

[0200] The equation optimization module is configured as an equation optimization network to optimize the equations of fracture evolution mechanism by constructing symbol datasets, fracture datasets, and optimization factors through operators and variable parameters.

[0201] The parameter adjustment module is configured to obtain the evolution results of the coal and rock mass fracture field by solving the fracture evolution mechanism equation, and to determine whether the accuracy of the coal and rock mass fracture field evolution results meets the requirements. When the accuracy of the coal and rock mass fracture field evolution results meets the requirements, the calculation ends; when the accuracy of the coal and rock mass fracture field evolution results does not meet the requirements, the deviation is calculated using CT detection data from the test set, and the parameters of the equation optimization network of the reinforcement learning architecture are adjusted until the accuracy of the coal and rock mass fracture field evolution results meets the requirements.

[0202] This optimization system focuses on optimizing the equations governing coal and rock fracture evolution through a reinforcement distillation-based approach. Compared to traditional numerical simulation and neural network methods, this reinforcement distillation-based optimization method extracts equation optimization factors from fracture evolution characteristics using knowledge distillation and SHAP. Simultaneously, it leverages the reward function of reinforcement learning to dynamically fine-tune the coefficients of the fracture evolution mechanism equations based on feedback from full-space fracture parameters and full-temporal acoustic emission parameters at various time points. This achieves optimization of the coal and rock fracture evolution mechanism based on experimental data, improving the accuracy of fracture parameter prediction.

[0203] The guiding model in this optimization system is a selective state-space neural network based on a dual-driven data mechanism. The network uses both evolutionary mechanism equations and data-driven equations as constraints, ensuring that the output of the guiding model follows physical laws while fully exploiting the nonlinear characteristics of multi-source experimental data. This demonstrates excellent physical feasibility and robustness under various experimental data distributions. Furthermore, the selective state-space structure significantly reduces computational resource dependence through fine-grained partitioning and filtering of state and observation variables. This allows the model to efficiently utilize acoustic emission and fracture parameter data collected during the loading of laboratory coal and rock samples, providing expert models with highly reliable data features.

[0204] This optimization system achieves the transfer of fundamental physical laws from the evolution mechanism equation to the optimized fracture evolution mechanism through equation distillation and reinforcement learning, ensuring that the optimized fracture evolution mechanism follows fundamental physical laws. Simultaneously, the reinforcement learning-based equation optimization network adjusts the optimization strategy of the evolution mechanism equation through the reward function, risk-preference policy gradient function, and optimization factors extracted by SHAP. This achieves synergistic innovation between structured knowledge distillation and dynamic policy optimization, extracting and deeply integrating domain physics knowledge into the reinforcement learning framework, effectively improving the interpretability and accuracy of the fracture evolution mechanism model.

[0205] It should be noted that the coal and rock fracture evolution mechanism optimization system based on the enhanced distillation architecture of the present invention can also perform any of the processing described in the previously described coal and rock fracture evolution mechanism optimization method based on the enhanced distillation architecture, and the specific details are not repeated here.

[0206] The foregoing description, with reference to preferred embodiments, details the exemplary implementation of the coal and rock fracture evolution mechanism optimization method and system based on enhanced distillation architecture proposed in this invention. However, those skilled in the art will understand that various modifications and alterations can be made to the above specific embodiments without departing from the concept of this invention, and various combinations can be made to the various technical features and structures proposed in this invention without exceeding the protection scope of this invention, which is determined by the appended claims.

Claims

1. A method for optimizing the coal and rock fracture evolution mechanism based on an enhanced distillation framework, characterized in that, Includes the following steps: S10: CT detection sensors and acoustic emission monitoring sensors are deployed on the coal and rock mass, and the coal and rock samples are placed in a testing machine for loading tests. CT detection data and full-time acoustic emission monitoring data are collected at different stages during the loading process of the coal and rock mass through the CT detection sensors and acoustic emission monitoring sensors. Among them, the CT detection data are full-space fracture parameter data at different stages, and the acoustic emission monitoring data are full-time acoustic emission parameter data. S20: Correspond the CT detection data and all-time acoustic emission monitoring data at different moments during the loading process of the coal and rock samples according to time to obtain the fracture parameter dataset, and divide the fracture parameter dataset into training set, validation set and test set; and make the operation symbols and variable parameters required for optimizing the mechanism equation into a symbol dataset. S30: The neural network with dual-drive architecture based on data mechanism is trained using the training set, and the trained neural network with dual-drive architecture based on data mechanism is validated using the validation set to obtain the bootstrap model; S40: Input the test set into the data features obtained by the guiding model, and train a neural network based on the evolutionary mechanism equation, distillation equation and SHAP to obtain an expert model; S50: Optimization factors are obtained by inputting the test set into the expert model; S60: Construct an equation optimization network based on a reinforcement learning architecture; S70: The equation optimization network optimizes the equations of the fracture evolution mechanism by constructing symbolic datasets, fracture datasets, and optimization factors using operators and variable parameters. S80: The evolution results of the coal and rock mass fracture field are obtained by solving the fracture evolution mechanism equation. It is then determined whether the accuracy of the coal and rock mass fracture field evolution results meets the requirements. If the accuracy of the coal and rock mass fracture field evolution results does not meet the requirements, the deviation is calculated using the test set CT detection data, and the parameters of the equation optimization network of the reinforcement learning architecture are adjusted until the accuracy of the coal and rock mass fracture field evolution results meets the requirements. The network parameters were adjusted by calculating the deviation based on the evolution results and CT detection data.

2. The method for optimizing the coal and rock fracture evolution mechanism based on the enhanced distillation architecture according to claim 1, characterized in that, In step S30, training the data mechanism dual-driven architecture neural network with the training set and validating the trained data mechanism dual-driven architecture neural network with the validation set to obtain the guided model includes the following steps: S31: The acoustic emission parameter AE is transmitted through the first data encoding layer. p Data extraction yields parameter-enhanced acoustic emission features AEf se The encoded data is concatenated to the feature extraction layer B through a residual connection layer to obtain the feature AE. b In the middle; acoustic emission characteristics AEf enhanced by the first normalization function. se Regularization is performed to obtain the feature AE n ; S32: Regularized Feature AE n Feature extraction AE is obtained through the residual connection formed by the second linear feature extraction layer A and the second activation function layer. a+s The feature AE spliced ​​into the output of the first selective state space architecture sssm Get AE SSSM The first linear feature extraction layer A outputs the feature AE. a Higher-dimensional features (AE) are obtained through feature extraction using the first convolutional layer and the first activation function layer. conv+s ; S33: 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 AE. 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 AEf. b+se ; S34: Regularize the concatenated features using the second Normalization function to obtain feature AEf. n The features extracted by the first fully connected layer are used for calculation to obtain feature F. f,pred ; S35: Analyzing the characteristics F of neural networks through data-driven equations and evolutionary mechanism equations. f,pred The calculations are constrained, and the resulting guided model is obtained.

3. The method for optimizing the coal and rock fracture evolution mechanism based on the enhanced distillation architecture according to claim 2, characterized in that, In step S35, the feature F of the neural network is analyzed through data-driven equations and evolutionary mechanism equations. f,pred Calculation constraints: First, the data-driven equation describes the characteristics F of the neural network. f,pred The calculation is constrained, and the expression is as follows: In the formula, The equation representing the deviation calculation; M represents the number of samples; i represents the index of the crack parameter; F f,pred F represents the data characteristics that need to be predicted. f,true Represents real data characteristics; Then, the evolutionary mechanism equation applies to the characteristics F of the neural network. M The calculation is constrained, and the expression is as follows: In the formula, Constraint equations representing the prediction mechanism of fracture parameters; M represents the total number of acoustic emission sensors; α represents the weighting coefficient of the acoustic emission parameters; AE p,i F represents the acoustic emission parameters monitored by the i-th acoustic emission sensor; f,pred This represents the data characteristics that need to be predicted.

4. The method for optimizing the coal and rock fracture evolution mechanism based on the enhanced distillation architecture according to claim 1, characterized in that, In step S40, training an expert model using data features obtained from the test-guided model to use a neural network based on evolutionary mechanisms, distillation equations, and SHAP includes the following steps: S41: The acoustic emission parameter ae is transmitted through the second data encoding layer. p Data extraction yields parameter-enhanced acoustic emission features (AEF). se The encoded data is concatenated to the feature extraction layer B through the second residual connection layer to obtain the feature ae. b In the middle; acoustic emission features aef enhanced by parameter enhancement through a third normalization function. se Regularization is performed to obtain the feature ae n ; S42: Regularized feature ae n Feature extraction is performed through the residual connection formed by the fourth linear feature extraction layer A and the fourth activation function layer to obtain ae. a+s The feature ae spliced ​​into the output of the second selective state space architecture sssm ae SSSM The feature ae output by the third linear feature extraction layer A a Higher-dimensional features are obtained through feature extraction using the second convolutional layer and the third activation function layer. conv+s ; S43: Gated computation of high-dimensional features is performed through a second selective state-space architecture, and then concatenated with the features extracted by the second convolutional layer and the third activation function layer to obtain feature ae. SSSM 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 aef. b+se ; S44: Regularize the concatenated features using the fourth Normalization function to obtain the feature aef. n The features extracted through the second fully connected layer are used for calculation to obtain feature f. M,pred ; S45: Distillation equation for the features f of the neural network M,pred The data features f are obtained by performing calculations and applying constraints. f Ultimately, an expert model was obtained.

5. The method for optimizing the coal and rock fracture evolution mechanism based on the enhanced distillation architecture according to claim 4, characterized in that, In step S45, the feature f of the neural network is processed by the distillation equation. M,pred The data features f are obtained by performing calculations and applying constraints. f The specific implementation process is as follows: The data features of the evolution mechanism equation are extracted into the data-driven equation through the distillation equation. The specific expression is as follows: where β and γ represent the training weights of the evolution mechanism equation and the data-driven equation, respectively; represents the evolution mechanism equation; represents the data-driven equation; 0% < epoch ≤ 50% represents the first half of the training process, and 50% < epoch ≤ 100% represents the second half of the training process; represents the distillation equation.

6. The method for optimizing the coal and rock fracture evolution mechanism based on the enhanced distillation architecture according to claim 1, characterized in that, In step S50, the optimization factor is obtained by inputting the test set into the expert model. The specific implementation process is as follows: The optimization factor is obtained through the SHAP-based optimization factor generation equation, where the optimization factor expression is as follows: In the formula, ψ i The value represents the optimization factor; M represents the number of features in the input data. This represents the set of all permutations and combinations of M features. f represents the set of all feature combinations preceding feature k in permutation R. x (S)=E(f(X)|X s =x S ) represents the conditional expectation given the feature subset S; i represents the optimization factor index.

7. The method for optimizing the coal and rock fracture evolution mechanism based on the enhanced distillation architecture according to claim 1, characterized in that, In step S60, an equation optimization network based on a reinforcement learning architecture is constructed. The specific implementation process is as follows: S61: The equation optimization network calculates the reward value of the initial mechanism equation through a reward function, the expression of which is as follows: In the formula, R represents the reward function; β represents a hyperparameter that controls the influence of the complexity of the mechanism equation on the optimization factor; This represents the fitting error for all observations. S62: The equation optimization network generates a generation strategy through the risk-preference policy gradient function optimization mechanism. The expression for the risk-preference policy gradient function is as follows: J risk (θ;∈)=E τ～p(τ|θ) [R(τ)|R(τ)≥R ∈ (i)] Among them, J risk (θ;∈) represents the risk preference strategy gradient function; θ represents the network parameters of the mechanism equation generation network; ∈ represents the risk preference parameters, used to select the proportion of high-reward advanced mechanism equations; N represents the number of advanced mechanism equations in the current batch; R(τ (i) ) represents the advanced mechanism equation τ (i) The reward value is calculated by the reward function; The threshold representing the reward value (1-∈) in the current batch's advancement mechanism equation; 1 {·} p(τ) represents an exponential function, taking the value 1 if the current condition is true, and 0 otherwise; (i) |θ) represents the network parameters θ that generate the mechanism equation, which in turn generate the advanced mechanism equation τ. (i) The probability, T represents the length of the mechanistic equation; log p(τ) (i) |θ) represents the logarithm of the probability; Represents log p(τ) (i) The partial derivative of |θ) with respect to the parameter θ. S63: The optimization factor is input into the equation optimization network. The equation optimization network optimizes the generation strategy of the mechanism equation generation network through the reward function and the risk preference policy gradient function. The expression of the generation strategy of the advanced mechanism equation generation network is as follows: Where, θ * The updated mechanism equations generate network parameters; η represents the learning rate.

8. The method for optimizing the coal and rock fracture evolution mechanism based on the enhanced distillation architecture according to claim 1, characterized in that, In step S70, the equation optimization network optimizes the equation by constructing a symbol dataset, a gap dataset, and an optimization factor using operators and variable parameters. The specific implementation process is as follows: S71: Transmit the operator O through the third data encoding layer m and variable parameter S n Extracting the enhanced operator feature of Of e and variable parameter features Sf e ; S72: The encoded data is concatenated to the third feature extraction layer B through the third residual connection layer to obtain feature O. b and S b middle; S73: The characteristics of the operator Of are determined by the fifth Normalization function. e and variable parameter features Sf e Regularization yields feature O n and S n ; S74: Regularized Feature O n and S n Feature O is obtained by performing feature extraction through the residual connection formed by the sixth linear feature extraction layer A and the sixth activation function layer. a+s and S a+s Feature O spliced ​​into the output of the third selective state space architecture sssm and S sssm O SSSM and S SSSM ; S75: Feature O output by the fifth linear feature extraction layer A a and S a Higher-dimensional features O are obtained through feature extraction via the third convolutional layer and the fifth activation function layer. conv+s and S conv+s ; S76: High-dimensional features are gated through a third selective state-space architecture and concatenated with features extracted from the third convolutional layer and the fifth activation function layer to obtain feature O. SSSM and S SSSM ; S77: The third 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 third residual connection layer to obtain feature Of. b+se and Sf b+se ; S78: Regularize the concatenated features using the sixth Normalization function to obtain the feature Of. n and Sf n ; S79: The features extracted by the third fully connected layer are used for calculation to obtain the initial mechanism equation; The coefficients of the initial mechanistic equations were calculated using a sparse regression algorithm on the crack dataset, and then the advanced mechanistic equations were obtained.

9. The method for optimizing the coal and rock fracture evolution mechanism based on the enhanced distillation architecture according to claim 8, characterized in that, In step S79, the sparse regression algorithm expression is as follows: In the formula, These represent the optimal coefficients in the initial mechanism equation; υ represents the coefficient vector; Θ(F,AE) represents the initial equation library, which converts the observed F(AE,t) and its derivative information into numerical features, thus enabling the matrix multiplication of the coefficient vector υ to obtain F. t The prediction results; F t The time derivative in the representative mechanism equation Numerical approximation; objective function The value represents the fitting error between the prediction and the true value; λ represents the regularization parameter, used to suppress the coefficient vector. This represents the regularization term, which ensures the simplicity of the initial mechanism equations; Represents the squared norm, used to measure the coefficient vector. Size.

10. A coal and rock fracture evolution mechanism optimization system based on an enhanced distillation architecture, characterized in that, include: The data acquisition module is configured to deploy CT detection sensors and acoustic emission monitoring sensors on the coal and rock mass, and place the coal and rock sample in the testing machine for loading tests. The CT detection sensors and acoustic emission monitoring sensors acquire CT detection data and full-time and space-time acoustic emission monitoring data at different stages during the loading process of the coal and rock mass. Among them, the CT detection data is the full-space fracture parameter data at different stages, and the acoustic emission monitoring data is the full-time and space-time acoustic emission parameter data. The symbolic data module is configured to map the CT detection data and all-time acoustic emission monitoring data at different stages during the loading process of coal and rock samples according to time to obtain a fracture parameter dataset, and divide the fracture parameter dataset into a training set, a validation set, and a test set; and to create a symbolic dataset of the operators and variable parameters required for optimizing the mechanism equation. The bootstrap model module is configured to train a neural network with a dual-drive architecture based on a training set, and to validate the trained neural network with a dual-drive architecture based on a validation set to obtain a bootstrap model. The expert model module is configured to input the test set into the data features obtained by the guide model, and train a neural network based on the evolutionary mechanism equation, distillation equation and SHAP to obtain the expert model from the data features; The optimization factor module is configured to obtain optimization factors by inputting the test set into the expert model. Optimize the network module and configure it for building an equation optimization network based on a reinforcement learning architecture; The equation optimization module is configured as an equation optimization network to optimize the equations of fracture evolution mechanism by constructing symbol datasets, fracture datasets, and optimization factors through operators and variable parameters. The parameter adjustment module is configured to obtain the evolution results of the coal and rock mass fracture field by solving the fracture evolution mechanism equation, determine whether the accuracy of the coal and rock mass fracture field evolution results meets the requirements, and when the accuracy of the coal and rock mass fracture field evolution results does not meet the requirements, it calculates the deviation using CT detection data from the test set and adjusts the parameters of the equation optimization network of the reinforcement learning architecture until the accuracy of the coal and rock mass fracture field evolution results meets the requirements.

Images