A method and system for inverting the full-time-space fracture field of coal and rock using interpretable graph neural networks

Abstract

This invention discloses a method and system for inverting the full-temporal and spatial fracture field of coal and rock using an interpretable graph neural network. The inversion method includes: obtaining CT detection and full-temporal and spatial acoustic emission monitoring information at different stages from a loading and fracturing test machine, and constructing a dataset; training a graph neural network to extract data features, obtaining a real-time quantitative inversion model for full-temporal and spatial fracture parameters; inputting the acoustic emission scalar parameter data from the test set into the model to obtain the inversion results for full-temporal and spatial fracture parameters; and obtaining the influence factors of the acoustic emission scalar parameters on the inversion results through a Bayesian interpreter, revealing the inversion mechanism model. The inversion model can realize the inversion of the full-temporal and spatial fracture field using full-temporal and spatial acoustic emission monitoring data during the loading process, and can invert fracture parameters in real time and accurately. Simultaneously, the mechanism model for fracture parameter inversion is revealed.

Description

Technical Field

[0001] This invention relates to the field of coal and rock monitoring technology, and in particular to a method and system for inverting the all-temporal fracture field of coal and rock using an interpretable graph neural network. Background Technology

[0002] With the increasing depth of coal mining, the severity of underground disasters is becoming increasingly prominent, posing a significant threat to safe production and efficient mining in coal mines. Studies have shown that during coal mining, excavation activities disrupt the stress equilibrium of the original coal and rock mass, leading to a redistribution of the stress field and abrupt structural changes. This process typically involves the rapid and unstable propagation of fractures, inducing large-scale fracturing (fracture) disasters, which are 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 achieve early warning of these disasters, it is necessary to conduct in-depth research on the distribution and evolution characteristics of fracture fields.

[0003] Although existing rock mechanics theories and methods attempt to explain the evolution of fracture fields in coal and rock masses induced by mining activities, this complex process remains a difficult-to-observe and understand "black box." Challenges persist in achieving real-time monitoring of fracture fields, accurate quantitative inversion, exploring their formation mechanisms, and revealing their development and evolution. Currently, industrial X-ray, CT scanning, and numerical simulation methods have made some progress. These technologies can be used to reconstruct coal and rock mass structures, especially showing potential in constructing three-dimensional digital models of discontinuous structures and fracture fields. For example, patent CN108072467A discloses a method for measuring the internal stress field of discontinuous structures; patent CN114169182A discloses a method and equipment for reconstructing ellipsoidal models of rock mass surface fractures; and patent CN114187423A discloses a method, electronic equipment, and storage medium for reconstructing surrounding rock fractures in three-dimensional simulation experiments. However, these methods have some limitations. First, they often rely on high-level laboratory conditions and complex model reconstruction methods, which limits their widespread application in practical engineering. Secondly, the prospects for widespread application of fracture field reproduction and inversion are still unclear, requiring further research and improvement. Furthermore, these methods cannot provide real-time, dynamic spatiotemporal evolution characteristics of fracture field distribution, which may impose certain limitations in practical production processes.

[0004] In recent years, experimental methods combining detection and monitoring technologies with numerical simulation have been applied in laboratories. However, current utilization of this detection and monitoring information mainly focuses on analyzing its temporal variation patterns. Furthermore, relying solely on numerical simulation for fracture field inversion requires high computational costs, and even minute geometric changes can potentially create bottlenecks during the iteration process. Therefore, how to utilize detection and monitoring information to achieve full-time and spatiotemporal fracture field inversion and overcome the limitations of numerical simulation methods still requires further research. With the continuous advancement and development of science and technology, breakthroughs in deep learning methods have attracted attention to intelligent algorithms in fracture field inversion. In particular, the use of graph neural networks and Bayesian networks can effectively extract features from different types of data. This provides the possibility of using deep learning methods to achieve accurate inversion, real-time reproduction, and mechanism explanation of fracture parameters and evolution processes in loaded coal and rock masses, but currently, specific feasible ideas and implementation methods are still lacking. Therefore, it is urgent to conduct actual measurements of detection and monitoring signals during the loading process of coal and rock masses and establish the relationships between acoustic emission waveforms, characteristic parameters, event locations, full-space fracture localization, and fracture parameter data. Meanwhile, the proposed method and system for inverting the all-time-space fracture field of coal and rock using interpretable graph neural networks has a significant role in revealing the "black box" evolution process of dynamic disasters in coal and rock masses and in achieving early warning of potential disasters. Summary of the Invention

[0005] This solution addresses the problems and needs raised above by proposing an interpretable graph neural network-based method and system for inverting the all-time-space fracture field of coal and rock. Due to the adoption of the following technical features, it can achieve the above-mentioned technical objectives and bring about several other technical benefits.

[0006] One objective of this invention is to propose an interpretable graph neural network-based method for inverting the full-time-space fracture field of coal and rock, comprising the following steps:

[0007] S10: CT detection sensors and acoustic emission monitoring sensors are deployed on the coal and rock mass, and the coal and rock mass is placed in a testing machine for load-bearing fracturing tests. The CT detection sensors and acoustic emission monitoring sensors collect CT detection information and full-time and space-time acoustic emission monitoring information at different stages during the load-bearing fracturing process of the coal and rock mass. Among them, the CT detection information includes full-space fracture parameter information at different stages, and the acoustic emission monitoring information includes full-time and space-time waveform information, full-time and space-time event information, and full-time and space-time scalar parameter information.

[0008] S20: The full-space fracture parameter information and full-time acoustic emission monitoring information at each stage of the coal and rock mass fracturing process are mapped according to time to obtain the fracture parameter dataset; the fracture parameter dataset is divided into training set, validation set and test set according to the corresponding proportions.

[0009] S30: The graph neural network extracts features from the full-space fracture parameter data and corresponding acoustic emission monitoring data at different times in the training and validation sets of the fracture parameter dataset, and trains a real-time quantitative inversion model of the full-space fracture parameters; inputs the full-space waveform data and scalar parameter data in the test set into the inversion model to obtain the inversion results of the full-space fracture parameters, that is, the fracture parameters of the full-space fracture field during the loading process; and obtains the evolution process of the full-space fracture field parameters of the coal and rock mass (30) under load.

[0010] S40: Input the scalar parameter data of the all-time and all-space acoustic emission monitoring data in the test set of the fracture parameter dataset and the all-time and all-space fracture parameter inversion results into a graph neural network interpreter based on a Bayesian network to obtain the influence factor between the acoustic emission scalar parameters and the fracture parameter inversion results, and then obtain a real-time quantitative inversion mechanism model for fracture parameters based on acoustic emission scalar parameters.

[0011] Furthermore, the coal and rock full-time-space fracture field inversion method based on the interpretable graph neural network according to the present invention may also have the following technical features:

[0012] In one example of the present invention, in step S30, the graph neural network extracts features from the full-space fracture parameters and corresponding acoustic emission monitoring information data at different times in the training and validation sets of the fracture parameter dataset, including the following steps:

[0013] The acoustic emission data is encoded into acoustic emission graph structure data through an encoding layer;

[0014] The fusion layer fuses the waveforms and scalar parameters in the input acoustic emission map structure data;

[0015] The feature extraction layer extracts features from the scalar parameter data fusion result, extracts node features, and obtains new acoustic emission waveforms and scalar parameter features.

[0016] Through a learnable shared linear transformation parameterized by the weight matrix W, combined with the weight vector Parameterized self-attention mechanisms extract higher-dimensional features;

[0017] The feature similarity of acoustic emission scalar parameters at different times is calculated by normalizing all choices at time j using the Softmax function.

[0018] By calculating the coefficients, the corresponding linear combination of features is obtained, and the final output features of the scalar parameters of the acoustic emission position at different times are obtained.

[0019] Multiple self-attention mechanisms are connected by an averaging method to stabilize the computation process, resulting in the final features extracted by the multi-head graph attention mechanism.

[0020] The Layer Normalization function is used to normalize the features extracted by the multi-head graph attention mechanism to improve training stability and speed.

[0021] By performing a linear transformation on the features extracted by the Layer Normalization function through a fully connected layer, higher-dimensional features are extracted.

[0022] The Readout function is used to perform regression calculations on the features extracted from the fully connected layer, and the regression parameters for each graph structure are calculated.

[0023] In one example of this invention, the expression for feature extraction from the result of scalar parameter data fusion via a feature extraction layer is as follows:

[0024]

[0025] Where h represents the acoustic emission scalar parameter in the acoustic emission graph structure of the input graph attention neural network; F represents the number of acoustic emission feature parameters; h ′ The acoustic emission scalar parameter F represents the new graph structure output by the feature extraction layer. ′ The number of acoustic emission scalar parameters representing different potential bases.

[0026] In one example of this invention, the expression for the higher-dimensional features extracted by the self-attention mechanism is as follows:

[0027] e ij =a(W·h i ,W·h j )

[0028] j∈N i

[0029]

[0030] Among them, e ij N represents the feature similarity between time i and time j of the acoustic emission waveform; i represents the neighborhood of time j; 'a' represents the self-attention mechanism, which consists of a single-layer feedforward neural network. W represents the weight vector; h represents the weight matrix; i and h j These represent the acoustic emission scalar parameters at different times in the acoustic emission graph structure data.

[0031] In one example of the present invention, in step S30, the graph neural network extracts features from the phased time full-space fracture parameter data and corresponding time acoustic emission data in the training and validation sets of the fracture parameter dataset, including:

[0032] The mean squared error (MSE) loss function is used to calculate the squared difference between the true and inverted values ​​of the full-space fracture parameters. The model is then optimized based on the calculated values. The formula for the MSE loss function is as follows:

[0033]

[0034] Where MSE represents the result of the mean squared error loss function calculation; Represents the true values ​​of fracture parameters across the entire space; This represents the inversion value of the fracture parameters in the entire space.

[0035] In one example of the present invention, in step S40, the acoustic emission pattern structure data G of all test sets is input into Φ. I The model outputs T I The calculation formula is as follows:

[0036] φ I G→T I

[0037] Where, φ I Represents the full-time-space quantitative inversion model of fractures; G represents all acoustic emission map structural data; T I The inversion space represents the quantitative inversion model of fractures across all time and space.

[0038] In one example of the present invention, step S40 specifically includes the following:

[0039] S401: The acoustic emission data in the test set is encoded into acoustic emission graph structure data G and input into the graph neural network interpreter. The data generation module performs activation operations on the acoustic emission feature v at each time step in G. The data generation module controls the activation probability of the acoustic emission feature in the graph structure through parameter p.

[0040] S402: Filter the acoustic emission scalar parameter variable D using the variable filtering module. I The variable data in the module is extracted to accelerate the mechanism extraction efficiency of the interpreter; the variable selection module extracts important variables I and MB through Markov layers. P (I) represents the minimum set of variables obtained through screening, denoted as U(I), where I is independent of other variables in P; simultaneously, at the neighborhood time... Set a random variable s in the middle v Perfect mapping of distribution B * , Provided separately with Same statistical information;

[0041] S403: The interpreter uses the BIC function to access the acoustic emission scalar parameter variable D. I Learn to explain Bayesian networks in Chinese;

[0042] S404: The full-time-space fracture parameter inversion mechanism model is obtained based on the influence factors output by the interpreter. The calculation formula is as follows:

[0043] M I ∶T I =m1×M1+m2×M2+…+m n ×M n

[0044] Among them, M n M represents the acoustic emission scalar parameter in the all-time-space fracture parameter inversion mechanism; I Represents a full-time and spacetime fracture parameter inversion mechanism model; T I The inversion space representing the full-time-space quantitative inversion model of fractures; m n The influence factor of acoustic emission scalar parameters representing the inversion mechanism model of fracture parameters in all time and space.

[0045] In one example of the present invention, step S401 includes the following steps:

[0046] Set random variable s v To indicate whether the acoustic emission feature v is activated, set s = {s v}, thus obtaining the activated acoustic emission map structure data G(s);

[0047] Input G(s) into Φ I The model outputs Φ I (G(s)), Φ is calculated through Ι(·) I (G(s)) and Φ I The differences between (G) are used to obtain the target variable I of the real-time quantitative inversion model of fracture parameters in all time and space based on G(s);

[0048] Set the acoustic emission scalar parameter variable at each time step to Δv = {s} v}, denoted as D I ;

[0049] The neighborhood time of the acoustic emission scalar parameter at each moment. Let V be the value of D after n samples. I The dimensions are all n×(|V|+1), and the calculation formula is as follows:

[0050] p∈(0,1)

[0051] I = {Ι(Φ I (G(s)))};Φ I :g→T I

[0052] Δv={s v};v∈V

[0053] G(s); G(s∈g

[0054] Where, parameter p represents the activation probability of features in the graph structure; Ι(·) represents the calculation of Φ I (G(s)) and Φ I (G) is a function that indicates whether there are significant differences; g represents all acoustic emission map structure data in the dataset.

[0055] In one example of the present invention, in step S403, the interpreter uses the BIC function D I In Chinese, we learn to explain Bayesian networks, and the calculation formula is as follows:

[0056]

[0057] in, θ represents the objective function for calculating the influence factor of each acoustic emission feature; BIC represents the function for calculating the influence factor; B represents a Bayesian network; θ B Represents the parameters of a Bayesian network; represent and D I The log-likelihood between [U(I)]; Dim[B] represents the dimension of the Bayesian network; m n The influence factor of acoustic emission scalar parameters representing the inversion mechanism model of fracture parameters in all time and space; The parameter value representing the maximum likelihood estimate is the log-likelihood maximization. ε represents the inversion mechanism of the spatiotemporal fracture parameters obtained by the interpreter; V(B) represents the interpretable domain; M represents the set of random variables in the Bayesian network; and M represents the number of variables in B.

[0058] Another objective of this invention is to propose an interpretable graph neural network-based all-time-space fracture field inversion system for coal and rock, comprising:

[0059] The information 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 mass in a testing machine for load-bearing fracturing tests. The CT detection sensors and acoustic emission monitoring sensors acquire CT detection information and full-time and space-time acoustic emission monitoring information at different stages during the load-bearing fracturing process of the coal and rock mass. Among them, the CT detection information includes full-space fracture parameter information at different stages, and the acoustic emission monitoring information includes full-time and space-time waveform information, full-time and space-time event information, and full-time and space-time scalar parameter information.

[0060] The dataset partitioning module is configured to map the full-space fracture parameter information and full-time acoustic emission monitoring information at different stages during the loading and fracturing process of coal and rock mass according to time, thereby obtaining a fracture parameter dataset; the fracture parameter dataset is then divided into training set, validation set and test set according to the corresponding proportions.

[0061] The model construction and inversion module is configured to extract features from the full-space fracture parameter data and corresponding acoustic emission monitoring data at different times in the training and validation sets of the fracture parameter dataset through graph neural networks, and train to obtain a real-time quantitative inversion model of full-space fracture parameters; input the full-space waveform data and scalar parameter data in the test set into the inversion model to obtain the inversion results of full-space fracture parameters, that is, the fracture parameters of the full-space fracture field during loading; obtain the evolution process of the full-space fracture field parameters of the coal and rock mass (30) under load;

[0062] The inversion mechanism model module is configured to input the scalar parameter data of the all-time and all-space acoustic emission monitoring data in the test set of the fracture parameter dataset and the all-time and all-space fracture parameter inversion results into a Bayesian network-based graph neural network interpreter to obtain the influence factor between the acoustic emission scalar parameters and the fracture parameter inversion results, and then obtain a real-time quantitative inversion mechanism model for fracture parameters based on the acoustic emission scalar parameters.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] 1. This invention focuses on using deep learning methods to invert the full-temporal and spatial fracture field of loaded coal and rock masses through probe-monitoring information from load-bearing fracture experiments. It combines graph neural networks and Bayesian networks to reveal the mechanism model for quantitative inversion of full-temporal and spatial fracture parameters, achieving high-precision inversion and full-temporal and spatial reproduction of the fracture field and its evolution in loaded coal and rock masses. Compared to traditional numerical simulation methods, this interpretable graph neural network-based full-temporal and spatial fracture field inversion method and system for coal and rock masses has higher computational efficiency and better adaptability, effectively overcoming the limitations of traditional methods.

[0065] 2. The real-time quantitative inversion model of all-temporal fracture parameters in this invention is obtained by joint feature extraction of fracture parameters through a graph neural network, proposing a method for feature extraction of fracture parameters based on acoustic emission information during the loading process of coal and rock mass. This method allows for the inversion of fracture parameters using acoustic emission information obtained in the laboratory during the loading process of coal and rock mass, laying a foundation for subsequent all-temporal fracture field inversion and obtaining the evolution process of all-temporal fracture field parameters during coal and rock mass loading.

[0066] 3. The real-time quantitative inversion mechanism interpreter for all-space fracture parameters in this invention uses a Bayesian network to sample, extract data, and extract structural features from the inversion results of all-space fracture parameters, acoustic emission scalar parameters, and the model, thereby obtaining the influencing factors in the real-time quantitative inversion model for all-space fracture parameters. The model inversion mechanism interpreter can directly obtain the all-space fracture parameter inversion mechanism model during the loading process of coal and rock mass through acoustic emission data and the prediction results of each model, thus revealing the black box process of fracture parameter inversion during the loading process of coal and rock mass.

[0067] 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

[0068] 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.

[0069] Figure 1 This is a flowchart of the coal and rock full-time-space fracture field inversion method using an interpretable graph neural network according to an embodiment of the present invention;

[0070] Figure 2 This is a schematic diagram of the layout structure of the coal and rock mass exploration-monitoring system according to an embodiment of the present invention;

[0071] Figure 3 This is a schematic diagram of the real-time quantitative inversion model structure for the evolution process of the all-time-space fracture field according to an embodiment of the present invention;

[0072] Figure 4 A flowchart illustrating the localization and inversion mechanism of the interpreter according to an embodiment of the present invention.

[0073] List of reference numerals in the attached diagram:

[0074] CT detection sensor 10;

[0075] Acoustic emission monitoring sensor 20;

[0076] X-ray detection sensor 21;

[0077] X-ray source 22;

[0078] Coal and rock mass 30;

[0079] Testing machine 40;

[0080] Coupling device 50. Detailed Implementation

[0081] 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.

[0082] 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.

[0083] According to a first aspect of the present invention, an interpretable graph neural network-based method for inverting the all-temporal fracture field of coal and rock is provided, such as... Figures 1 to 4 As shown, it includes the following steps:

[0084] S10: A CT detection sensor 10 and an acoustic emission monitoring sensor 20 are arranged on the coal and rock mass 30, and the coal and rock mass 30 is placed in a testing machine 40 for a load-bearing fracturing test. The CT detection sensor 10 and the acoustic emission monitoring sensor 20 collect CT detection information and full-time acoustic emission monitoring information at various stages of the load-bearing fracturing process of the coal and rock mass 30. The CT detection information includes full-space fracture parameter information at each stage, and the acoustic emission monitoring information includes full-time waveform information, full-time event information, and full-time scalar parameter information; for example, such as... Figure 2As shown, the testing machine 40 is placed in a sealed box, and the coal and rock mass 30 is placed in the testing machine 40. The acoustic emission monitoring sensor 20 is installed inside the coal and rock mass 30, while the CT detection sensor 10 includes an X-ray detection sensor 21 and an X-ray source 22 and is located on the outside of the coal and rock mass 30. The X-ray detection sensor 21 and the X-ray source 22 are respectively placed on both sides of the coal and rock mass 30. The coupling device 50 is also installed inside the coal and rock mass 30 and is set opposite to the acoustic emission monitoring sensor 20.

[0085] S20: The full-space fracture parameter information and full-time acoustic emission monitoring information of the coal and rock mass 30 during the loading and fracturing process are mapped according to time to obtain the fracture parameter dataset; the fracture parameter dataset is divided into training set, validation set and test set according to the corresponding proportions;

[0086] S30: A graph neural network extracts features from the full-space fracture parameter data and corresponding acoustic emission monitoring data at different times in the training and validation sets of the fracture parameter dataset, and trains a real-time quantitative inversion model of full-space fracture parameters. The full-space waveform data and scalar parameter data in the test set are input into the inversion model to obtain the inversion results of full-space fracture parameters, that is, the fracture parameters of each fracture in the full-space fracture field during loading. The evolution process of the full-space fracture field parameters of the coal and rock mass under load is obtained.

[0087] S40: Input the scalar parameter data of the all-time and all-space acoustic emission monitoring data in the test set of the fracture parameter dataset and the all-time and all-space fracture parameter inversion results into a graph neural network interpreter based on a Bayesian network to obtain the influence factor between the acoustic emission scalar parameters and the fracture parameter inversion results, and then obtain a real-time quantitative inversion mechanism model for fracture parameters based on acoustic emission scalar parameters.

[0088] This inversion method focuses on using deep learning to invert the full-temporal and spatial fracture field of the loaded coal-rock mass 30 based on exploration-monitoring information from the loading and fracturing experiment. It combines graph neural networks and Bayesian networks to reveal the mechanism model for quantitative inversion of full-temporal and spatial fracture parameters, achieving high-precision inversion and full-temporal and spatial reproduction of the fracture field and its evolution in the loaded coal-rock mass 30. Compared to traditional numerical simulation methods, this interpretable graph neural network-based full-temporal and spatial fracture field inversion method and system for coal-rock mass has higher computational efficiency and better adaptability, effectively overcoming the limitations of traditional methods.

[0089] The real-time quantitative inversion model of fracture parameters in this inversion method is obtained by joint feature extraction of fracture parameters through graph neural networks. A method for feature extraction of fracture parameters based on acoustic emission information during the loading process of coal-rock mass 30 is proposed. This method allows for the inversion of fracture parameters using acoustic emission information obtained from the laboratory during the loading process of coal-rock mass 30, laying a foundation for subsequent inversion of the full-time fracture field and obtaining the evolution process of the full-time fracture field parameters of coal-rock mass 30 under loading.

[0090] The interpreter for the real-time quantitative inversion mechanism of fracture parameters in this inversion method uses a Bayesian network to sample, extract data, and extract structural features from the inversion results of fracture parameters, acoustic emission scalar parameters, and the model, thereby obtaining the influencing factors in the real-time quantitative inversion model of fracture parameters in the entire space-time. The model inversion mechanism interpreter can directly obtain the inversion mechanism model of fracture parameters in the entire space-time during the loading process of coal-rock mass 30 through acoustic emission data and the prediction results of each model, thus revealing the black box process of fracture parameter inversion in coal-rock mass 30 during the loading process.

[0091] In one example of the present invention, such as Figure 2 As shown, in step S10, before placing the coal and rock mass 30 in the testing machine 40 for load-bearing fracture test, the following steps are also included: connecting the acoustic emission probe to the coupling device 50, and placing the probe of the coupling device 50 on the surface of the coal and rock mass 30 after applying coupling agent; setting the basic parameters of acoustic emission and the monitoring information to be collected; setting the basic parameters of the CT detector and the detection information to be collected; setting the loading mode of the testing machine 40 and the data to be collected.

[0092] In one example of the present invention, such as Figure 3 As shown, in step S30, the graph neural network extracts features from the full-space fracture parameters and corresponding acoustic emission monitoring information data at different times in the training and validation sets of the fracture parameter dataset, including the following steps:

[0093] The acoustic emission data is encoded into acoustic emission graph structure data through an encoding layer;

[0094] The fusion layer fuses the waveforms and scalar parameters in the input acoustic emission map structure data;

[0095] The feature extraction layer extracts features from the scalar parameter data fusion result, extracts node features, and obtains new acoustic emission waveforms and scalar parameter features.

[0096] Through a learnable shared linear transformation parameterized by the weight matrix W, combined with the weight vector Parameterized self-attention mechanisms extract higher-dimensional features;

[0097] The feature similarity of acoustic emission scalar parameters at different times is calculated by normalizing all choices at time j using the Softmax function.

[0098] By calculating the coefficients, the corresponding linear combination of features is obtained, and the final output features of the scalar parameters of the acoustic emission position at different times are obtained.

[0099] Multiple self-attention mechanisms are connected by an averaging method to stabilize the computation process, resulting in the final features extracted by the multi-head graph attention mechanism.

[0100] The Layer Normalization function is used to normalize the features extracted by the multi-head graph attention mechanism to improve training stability and speed.

[0101] By performing a linear transformation on the features extracted by the Layer Normalization function through a fully connected layer, higher-dimensional features are extracted.

[0102] The Readout function is used to perform regression calculations on the features extracted from the fully connected layer, and the regression parameters for each graph structure are calculated.

[0103] In one example of this invention, the expression for feature extraction from the result of scalar parameter data fusion via a feature extraction layer is as follows:

[0104]

[0105] Where h represents the acoustic emission scalar parameter in the acoustic emission graph structure of the input graph attention neural network; F represents the number of acoustic emission feature parameters; h ′ The acoustic emission scalar parameter F represents the new graph structure output by the feature extraction layer. ′ The number of acoustic emission scalar parameters representing different potential bases.

[0106] In one example of this invention, the expression for the higher-dimensional features extracted by the self-attention mechanism is as follows:

[0107] e ij =a(W·h i ,W·h j )

[0108] j∈N i

[0109]

[0110] Among them, e ij N represents the feature similarity between time i and time j of the acoustic emission waveform; i represents the neighborhood of time j; 'a' represents the self-attention mechanism, which consists of a single-layer feedforward neural network. W represents the weight vector; h represents the weight matrix; i and h j These represent the acoustic emission scalar parameters at different times in the acoustic emission graph structure data.

[0111] In one example of the present invention, in step S30, the graph neural network extracts features from the phased time full-space fracture parameter data and corresponding time acoustic emission data in the training and validation sets of the fracture parameter dataset, including:

[0112] The mean squared error (MSE) loss function is used to calculate the squared difference between the true and inverted values ​​of the full-space fracture parameters. The model is then optimized based on the calculated values. The formula for the MSE loss function is as follows:

[0113]

[0114] Where MSE represents the result of the mean squared error loss function calculation; Represents the true values ​​of fracture parameters across the entire space; This represents the inversion value of the fracture parameters in the entire space.

[0115] In one example of the present invention, in step S40, the acoustic emission pattern structure data G of all test sets is input into Φ. I The model outputs T I The calculation formula is as follows:

[0116] Φ I G→T I

[0117] Where, Φ I Represents the full-time-space quantitative inversion model of fractures; G represents all acoustic emission map structural data; T I The inversion space represents the quantitative inversion model of fractures across all time and space.

[0118] In one example of the present invention, such as Figure 4 As shown, step S40 specifically includes the following:

[0119] S401: The acoustic emission data in the test set is encoded into acoustic emission graph structure data G and input into the graph neural network interpreter. The data generation module performs activation operations on the acoustic emission feature v at each time step in G. The data generation module controls the activation probability of the acoustic emission feature in the graph structure through parameter p.

[0120] S402: Filter the acoustic emission scalar parameter variable D using the variable filtering module. I The variable data is extracted to accelerate the interpreter's mechanism extraction efficiency. The variable selection module extracts important variables I and MB through Markov layers. P(I) represents the minimum set of variables obtained through screening, denoted as U(I), where I is independent of other variables in P. Simultaneously, at the neighborhood time... Set a random variable s in the middle v Perfect mapping of distribution B * , Provided separately with Same statistical information;

[0121] S403: The interpreter uses the BIC function to access the acoustic emission scalar parameter variable D. I Learn to explain Bayesian networks in Chinese;

[0122] S404: The full-time-space fracture parameter inversion mechanism model is obtained based on the influence factors output by the interpreter. The calculation formula is as follows:

[0123] M I ∶T I =m1×M1+m2×M2+…+m n ×M n

[0124] Among them, M n M represents the acoustic emission scalar parameter in the all-time-space fracture parameter inversion mechanism; I Represents a full-time and spacetime fracture parameter inversion mechanism model; T i The inversion space representing the full-time-space quantitative inversion model of fractures; m n The influence factor of acoustic emission scalar parameters representing the inversion mechanism model of fracture parameters in all time and space.

[0125] In one example of the present invention, step S401 includes the following steps:

[0126] Set random variable s v To indicate whether the acoustic emission feature v is activated, set s = {s v}, obtain the activated acoustic emission pattern structure data G(s), and input G(s) into φ I The model outputs φ I (G(s)), Φ is calculated through Ι(·) I (G(s)) and Φ I The differences between G(s) are used to obtain the objective variable I of the real-time quantitative inversion model of fracture parameters in all time and space based on G(s). The acoustic emission scalar parameter variable at each time step is set as Δv={s v}, denoted as D I To ensure that the acoustic emission scalar parameters at each time step have an impact on the inversion results, the neighboring time steps of the acoustic emission scalar parameters at each time step are used as a reference. Let V be the value of D after n samples. I The dimensions are all n×(|V|+1), and the calculation formula is as follows:

[0127] p∈(0,1)

[0128] I = {Ι(Φ I (G(s)))};Φ I :g→T I

[0129] Δv={s v};v∈V

[0130] G(s); G(s∈g

[0131] Where, parameter p represents the activation probability of features in the graph structure; Ι(·) represents the calculation of Φ I (G(s)) and Φ I (G) is a function that indicates whether there are significant differences; g represents all acoustic emission map structure data in the dataset.

[0132] In one example of the present invention, in step S403, the interpreter uses the BIC function D I In Chinese, we learn to explain Bayesian networks, and the calculation formula is as follows:

[0133]

[0134] in, θ represents the objective function for calculating the influence factor of each acoustic emission feature; BIC represents the function for calculating the influence factor; B represents a Bayesian network; θ B Represents the parameters of a Bayesian network; represent and D I The log-likelihood between [U(I)]; Dim[B] represents the dimension of the Bayesian network; m n The influence factor of acoustic emission scalar parameters representing the inversion mechanism model of fracture parameters in all time and space; The parameter value representing the maximum likelihood estimate is the log-likelihood maximization. ε represents the inversion mechanism of the spatiotemporal split parameters obtained by the interpreter; V(B) represents the interpretable domain; M represents the set of random variables in the Bayesian network; and M represents the number of variables in B.

[0135] According to a second aspect of the present invention, a coal and rock full-time-space fracture field inversion system with interpretable graph neural networks includes:

[0136] The information acquisition module is configured to deploy CT detection sensors 10 and acoustic emission monitoring sensors 20 on the coal and rock mass 30, and place the coal and rock mass 30 in the testing machine 40 for a load-bearing fracturing test. The CT detection sensors 10 and acoustic emission monitoring sensors 20 acquire CT detection information and full-time and space-time acoustic emission monitoring information at different stages during the load-bearing fracturing process of the coal and rock mass 30. The CT detection information includes full-space fracture parameter information at different stages, and the acoustic emission monitoring information includes full-time and space-time waveform information, full-time and space-time event information, and full-time and space-time scalar parameter information.

[0137] The dataset partitioning module is configured to map the full-space fracture parameter information and full-time acoustic emission monitoring information of the coal-rock mass (30) during the loading and fracturing process according to time, and obtain the fracture parameter dataset; the fracture parameter dataset is divided into training set, validation set and test set according to the corresponding proportions;

[0138] The model construction and inversion module is configured to extract features from the full-space fracture parameter data and corresponding acoustic emission monitoring data at different times in the training and validation sets of the fracture parameter dataset through graph neural networks, and train to obtain a real-time quantitative inversion model of full-space fracture parameters; input the full-space waveform data and scalar parameter data in the test set into the inversion model to obtain the inversion results of full-space fracture parameters, that is, the fracture parameters of the full-space fracture field during loading; obtain the evolution process of the full-space fracture field parameters of the coal and rock mass (30) under load;

[0139] The inversion mechanism model module is configured to input the scalar parameter data of the all-time and all-space acoustic emission monitoring data in the test set of the fracture parameter dataset and the all-time and all-space fracture parameter inversion results into a Bayesian network-based graph neural network interpreter to obtain the influence factor between the acoustic emission scalar parameters and the fracture parameter inversion results, and then obtain a real-time quantitative inversion mechanism model for fracture parameters based on the acoustic emission scalar parameters.

[0140] This inversion system focuses on using deep learning methods to invert the full-temporal and spatial fracture field of loaded coal-rock mass 30 based on probe-monitoring information from the loading and fracturing experiment. It combines graph neural networks and Bayesian networks to reveal the mechanism model for quantitative inversion of full-temporal and spatial fracture parameters, achieving high-precision inversion and full-temporal and spatial reproduction of the fracture field and its evolution in loaded coal-rock mass 30. Compared to traditional numerical simulation methods, this interpretable graph neural network-based full-temporal and spatial fracture field inversion method and system for coal-rock mass has higher computational efficiency and better adaptability, effectively overcoming the limitations of traditional methods.

[0141] The real-time quantitative inversion model of fracture parameters in this inversion system is obtained by joint feature extraction of fracture parameters through graph neural networks. A method for feature extraction of fracture parameters based on acoustic emission information during the loading process of coal-rock mass 30 is proposed. This method allows for the inversion of fracture parameters using acoustic emission information obtained from the laboratory during the loading process of coal-rock mass 30, laying a foundation for subsequent inversion of the full-time fracture field and obtaining the evolution process of the full-time fracture field parameters of coal-rock mass 30 under loading.

[0142] The real-time quantitative inversion mechanism interpreter for all-temporal fracture parameters in this inversion system uses a Bayesian network to sample, extract data, and extract structural features from the inversion results of all-space fracture parameters, acoustic emission scalar parameters, and the model, thereby obtaining the influencing factors in the real-time quantitative inversion model for all-temporal fracture parameters. The model inversion mechanism interpreter can directly obtain the all-temporal fracture parameter inversion mechanism model for coal-rock mass 30 during loading from acoustic emission data and the prediction results of each model, thus revealing the "black box" process of fracture parameter inversion in coal-rock mass 30 during loading.

[0143] The foregoing description, with reference to preferred embodiments, details the exemplary implementation of the coal and rock full-time-space fracture field inversion method and system with interpretable graph neural networks 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 inverting the full-time-space fracture field of coal and rock using interpretable graph neural networks, characterized in that, Includes the following steps: S10: CT detection sensor (10) and acoustic emission monitoring sensor (20) are arranged on the coal and rock mass (30), and the coal and rock mass (30) is placed in the testing machine (40) for load-bearing fracturing test. The CT detection information and the full-time and space-time acoustic emission monitoring information of the coal and rock mass (30) during the load-bearing fracturing process are collected by the CT detection sensor (10) and the acoustic emission monitoring sensor (20). Among them, the CT detection information includes the full-space fracture parameter information at each stage, and the acoustic emission monitoring information includes the full-time and space-time waveform information, the full-time and space-time event information and the full-time and space-time scalar parameter information. S20: The full-space fracture parameter information and full-time acoustic emission monitoring information of the coal and rock mass (30) during the loading and fracturing process are matched according to time to obtain the fracture parameter dataset; the fracture parameter dataset is divided into training set, validation set and test set according to the corresponding proportions. S30: The graph neural network extracts features from the full-space fracture parameter data and corresponding acoustic emission monitoring data at different times in the training and validation sets of the fracture parameter dataset, and trains a real-time quantitative inversion model of the full-space fracture parameters; inputs the full-space waveform data and scalar parameter data in the test set into the inversion model to obtain the inversion results of the full-space fracture parameters, that is, the fracture parameters of the full-space fracture field during the loading process; and obtains the evolution process of the full-space fracture field parameters of the coal and rock mass (30) under load. S40: Input the scalar parameter data of the full-time and spatiotemporal acoustic emission monitoring data and the full-time and spatiotemporal fracture parameter inversion results from the test set of the fracture parameter dataset into a graph neural network interpreter based on a Bayesian network to obtain the influence factors between the acoustic emission scalar parameters and the fracture parameter inversion results, and thus obtain a real-time quantitative inversion mechanism model for fracture parameters based on acoustic emission scalar parameters; wherein, step S40 specifically includes the following: S401: Encode the acoustic emission data in the test set into acoustic emission graph structure data G and input it into the graph neural network interpreter. The data generation module performs an activation operation on the acoustic emission feature v at each time step in G. The data generation module controls the activation probability of the acoustic emission feature in the graph structure through the parameter p; S402: The acoustic emission scalar parameter variable D is processed by the variable filtering module. I The variable data in the module is extracted to accelerate the mechanism extraction efficiency of the interpreter; the variable selection module extracts important variables I and MB through Markov layers. P (I) represents the minimum set of variables obtained through screening, denoted as U(I), where I is independent of other variables in P; simultaneously, at the neighborhood time... Set a random variable s in the middle v Perfect mapping of distribution B * , Provided separately with Same statistical information; S403: The interpreter uses the BIC function to access the acoustic emission scalar parameter variable D. I In the learning process, a Bayesian network is used to interpret the model; S404: Based on the influence factors output by the interpreter, the full-time-space fracture parameter inversion mechanism model is obtained, and the calculation formula is as follows: M I ∶T I =m1×M1+m2×M2+…+m n ×M n Among them, M n M represents the acoustic emission scalar parameter in the all-time-space fracture parameter inversion mechanism; I Represents a full-time and spacetime fracture parameter inversion mechanism model; T I The inversion space representing the full-time-space quantitative inversion model of fractures; m n The influence factor of acoustic emission scalar parameters representing the inversion mechanism model of fracture parameters in all time and space.

2. The method for inverting the full-time-space fracture field of coal and rock using interpretable graph neural networks according to claim 1, characterized in that, In step S30, the graph neural network extracts features from the full-space fracture parameters and corresponding acoustic emission monitoring information data at different times in the training and validation sets of the fracture parameter dataset, including the following steps: The acoustic emission data is encoded into acoustic emission graph structure data through an encoding layer; The fusion layer fuses the waveforms and scalar parameters in the input acoustic emission map structure data; The feature extraction layer extracts features from the scalar parameter data fusion result, extracts node features, and obtains new acoustic emission waveforms and scalar parameter features. Through a learnable shared linear transformation parameterized by the weight matrix W, combined with the weight vector Parameterized self-attention mechanisms extract higher-dimensional features; The feature similarity of acoustic emission scalar parameters at different times is calculated by normalizing all choices at time j using the Softmax function. By calculating the coefficients, the corresponding linear combination of features is obtained, and the final output features of the scalar parameters of the acoustic emission position at different times are obtained. Multiple self-attention mechanisms are connected by an averaging method to stabilize the computation process, resulting in the final features extracted by the multi-head graph attention mechanism. The Layer Normalization function is used to normalize the features extracted by the multi-head graph attention mechanism to improve training stability and speed. By performing a linear transformation on the features extracted by the Layer Normalization function through a fully connected layer, higher-dimensional features are extracted. The Readout function is used to perform regression calculations on the features extracted from the fully connected layer, and the regression parameters for each graph structure are calculated.

3. The method for inverting the coal and rock fracture field in all time and space using an interpretable graph neural network according to claim 2, characterized in that, The expression for feature extraction from the scalar parameter data fusion result through the feature extraction layer is as follows: Where h represents the acoustic emission scalar parameter in the acoustic emission graph structure of the input graph attention neural network; F represents the number of acoustic emission scalar parameters; h ′ The acoustic emission scalar parameter F represents the new graph structure output by the feature extraction layer. ′ The number of acoustic emission scalar parameters representing different potential bases.

4. The method for inverting the full-time-space fracture field of coal and rock using interpretable graph neural networks according to claim 2, characterized in that, The expression for the higher-dimensional features extracted by the self-attention mechanism is as follows: have been ij =a(W·h i ,W·h j ) j∈N i Among them, e ij N represents the feature similarity between time i and time j of the acoustic emission waveform; i represents the neighborhood of time j; 'a' represents the self-attention mechanism, which consists of a single-layer feedforward neural network. W represents the weight vector; h represents the weight matrix; i and h j These represent the acoustic emission scalar parameters at different times in the acoustic emission graph structure data.

5. The method for inverting the full-time-space fracture field of coal and rock using interpretable graph neural networks according to claim 1, characterized in that, In step S30, the graph neural network extracts features from the phased time-based full-space fracture parameter data and corresponding time-based acoustic emission data in the training and validation sets of the fracture parameter dataset, including: The mean squared error (MSE) loss function is used to calculate the squared difference between the true and inverted values ​​of the full-space fracture parameters. The model is then optimized based on the calculated values. The formula for the MSE loss function is as follows: Where MSE represents the result of the mean squared error loss function calculation; Represents the true values ​​of fracture parameters across the entire space; This represents the inversion value of the fracture parameters in the entire space.

6. The method for inverting the full-time-space fracture field of coal and rock using interpretable graph neural networks according to claim 1, characterized in that, In step S40, the acoustic emission pattern structure data G of all test sets is input into Φ. I The model outputs T I The calculation formula is as follows: F I :G→T I Where, Φ I Represents the full-time-space quantitative inversion model of fractures; G represents all acoustic emission map structural data; T I The inversion space represents the quantitative inversion model of fractures across all time and space.

7. The method for inverting the full-time-space fracture field of coal and rock using interpretable graph neural networks according to claim 1, characterized in that, Step S401 includes the following steps: Set random variable s v To indicate whether the acoustic emission feature v is activated, set s = {s v }, thus obtaining the activated acoustic emission map structure data G(s); Input G(s) into Φ I The model outputs Φ I (G(s)), Φ is calculated through Ι(·) I (G(s)) and Φ I The differences between (G) are used to obtain the target variable I of the real-time quantitative inversion model of fracture parameters in all time and space based on G(s); Set the acoustic emission scalar parameter variable at each time step to Δv = {s} v }, denoted as D I ; The neighborhood time of the acoustic emission scalar parameter at each moment. Let V be the value of D after n samples. I The dimensions are all n×(|V|+1), and the calculation formula is as follows: p∈(0，1) I={Ι(Φ I (G(s)))};Φ I :g→T I Δv={s v };v∈V G(s); G(s∈g Where, parameter p represents the activation probability of features in the graph structure; Ι(·) represents the calculation of Φ I (G(s)) and Φ I (G) is a function that indicates whether there are significant differences; g represents all acoustic emission map structure data in the dataset.

8. The method for inverting the coal and rock fracture field in all time and space using an interpretable graph neural network according to claim 1, characterized in that, In step S403, the interpreter uses the BIC function D I In Chinese, we learn to explain Bayesian networks, and the calculation formula is as follows: in, θ represents the objective function for calculating the influence factor of each acoustic emission feature; BIC represents the function for calculating the influence factor; B represents a Bayesian network; θ B Represents the parameters of a Bayesian network; represent and D I The log-likelihood between [U(I)]; Dim[B] represents the dimension of the Bayesian network; m n The influence factor of acoustic emission scalar parameters representing the inversion mechanism model of fracture parameters in all time and space; The parameter value representing the log-likelihood maximization; ε represents the inversion mechanism of the spatiotemporal fracture parameters obtained by the interpreter; V(B) represents the interpretable domain; M represents the set of random variables in the Bayesian network; and M represents the number of variables in B.

9. A coal and rock fracture field inversion system based on interpretable graph neural networks, characterized in that, include: The information acquisition module is configured to place CT detection sensors (10) and acoustic emission monitoring sensors (20) on the coal and rock mass (30), and place the coal and rock mass (30) in the testing machine (40) for a load-bearing fracture test. The CT detection sensor (10) and acoustic emission monitoring sensor (20) acquire the CT detection information and the full-time and space-time acoustic emission monitoring information of the coal and rock mass (30) during the load-bearing fracture process. The CT detection information includes the full-space fracture parameter information at each stage, and the acoustic emission monitoring information includes the full-time and space-time waveform information, the full-time and space-time event information, and the full-time and space-time scalar parameter information. The dataset partitioning module is configured to map the full-space fracture parameter information and full-time acoustic emission monitoring information of the coal-rock mass (30) during the loading and fracturing process according to time, and obtain the fracture parameter dataset; the fracture parameter dataset is divided into training set, validation set and test set according to the corresponding proportions; The model construction and inversion module is configured to extract features from the full-space fracture parameter data and corresponding acoustic emission monitoring data at different times in the training and validation sets of the fracture parameter dataset through graph neural networks, and train to obtain a real-time quantitative inversion model of full-space fracture parameters; input the full-space waveform data and scalar parameter data in the test set into the inversion model to obtain the inversion results of full-space fracture parameters, that is, the fracture parameters of the full-space fracture field during loading; obtain the evolution process of the full-space fracture field parameters of the coal and rock mass (30) under load; The inversion mechanism model module is configured to input the scalar parameter data and the inversion results of the full-time and spatiotemporal acoustic emission monitoring data from the test set of the fracture parameter dataset into a Bayesian network-based graph neural network interpreter. This yields the influence factors between the acoustic emission scalar parameters and the fracture parameter inversion results, ultimately resulting in a real-time quantitative inversion mechanism model for fracture parameters based on the acoustic emission scalar parameters. Specifically, this includes: encoding the acoustic emission data from the test set into acoustic emission graph structure data G and inputting it into the graph neural network interpreter; the data generation module performing activation operations on the acoustic emission feature v at each time step in G; and controlling the activation probability of the acoustic emission features in the graph structure through the parameter p; and using the variable selection module to select the acoustic emission scalar parameter variable D. I The variable data in the module is extracted to accelerate the mechanism extraction efficiency of the interpreter; the variable selection module extracts important variables I and MB through Markov layers. P (I) represents the minimum set of variables obtained through screening, denoted as U(I), where I is independent of other variables in P; simultaneously, at the neighborhood time... Set a random variable s in the middle v Perfect mapping of distribution B * , Provided separately with The same statistical information; the interpreter uses the BIC function to access the acoustic emission scalar parameter variable D. I The Bayesian network is used to interpret the learning process; based on the influence factors output by the interpreter, the all-temporal fracture parameter inversion mechanism model is obtained, and the calculation formula is as follows: M I ∶T I =m1×M1+m2×M2+…+m n ×M n Among them, M n M represents the acoustic emission scalar parameter in the all-time-space fracture parameter inversion mechanism; I Represents a full-time and spacetime fracture parameter inversion mechanism model; T I The inversion space representing the full-time-space quantitative inversion model of fractures; m n The influence factor of acoustic emission scalar parameters representing the inversion mechanism model of fracture parameters in all time and space.

Images