Inversion method and device for full-time and space stress field of loaded coal and rock based on deep learning

Abstract

The invention discloses a method and device for inverting the full-time and space stress field of loaded coal and rock based on deep learning. The inversion method comprises the following steps: collecting the obtained exploration and monitoring data, stress data and boundary condition information; dividing the measured data into a training set, a verification set and a test set according to the corresponding time; training to obtain an information transcription model; inputting the full-time exploration and monitoring data in the test set into the information transcription model to obtain the full-time local stress value; establishing a numerical simulation model according to the loading boundary conditions of the coal and rock load fracture test, and obtaining the coal and rock stress field data through the numerical simulation method; dividing the full-time local stress value and the full-space stress field data into a training set, a verification set and a test set; training to obtain the full-time and space stress field inversion model; verifying the inversion result, adjusting the model parameters according to the verification result, and obtaining the full-time and space stress field inversion model to obtain the full-time and space stress field in the coal and rock load fracture test process through the full-time local stress value.

Description

Technical Field

[0001] The present invention relates to the technical field of coal rock monitoring, and in particular to a method and device for inverting the full-time and space stress field of loaded coal rock based on deep learning. Background Art

[0002] Studies have shown that during coal mining, mining activities will destroy the stress balance state of the original coal-rock mass, leading to the redistribution of the stress field. This process will lead to structural mutations of the coal-rock mass, which is the root cause of dynamic disasters such as rock burst, coal and gas outbursts, etc. Therefore, in order to deeply understand the mechanism of coal-rock dynamic disasters and achieve early warning of disasters, it is necessary to carry out in-depth research on the distribution and evolution characteristics of the stress field.

[0003] Although existing rock mechanics theories and methods attempt to explain the evolution of coal rock stress fields caused by mining activities, this complex process is still regarded as a "black box" that is difficult to observe and understand, and there are still challenges in achieving real-time monitoring and accurate inversion of stress fields. At present, advanced imaging technologies such as industrial X-rays, CT scanning, and focused ion beam electron microscopy scanning, as well as numerical simulation methods, have made certain progress. These technologies can be used to reconstruct coal rock structures, especially in constructing three-dimensional digital models of discontinuous structures and stress fields. For example, patent CN108072467A discloses a method for measuring the internal stress field of a discontinuous structure, patent CN106326524A discloses a method for numerical simulation of stress fields in non-uniform formations, and patent CN115935702A discloses a method for determining the direction of action of tectonic stress fields based on numerical simulation technology. However, these methods have some limitations. First, they often rely on high laboratory conditions and complex model reconstruction methods, which limits their wide application in actual engineering. Secondly, the prospects for the promotion and application of stress field reproduction and inversion are still unclear, and further research and improvement are needed. In addition, these methods are also unable to provide real-time dynamic spatiotemporal evolution characteristics of stress field distribution, which may cause certain limitations in the actual production process.

[0004] In recent years, some experimental methods combining detection-monitoring technology with numerical simulation have been applied in laboratories. However, the current use of these detection and monitoring information is still mainly focused on analyzing its temporal variation law. At the same time, relying solely on numerical simulation for stress field inversion requires high computational costs, and even small geometric changes are likely to cause bottlenecks in the iteration process. Therefore, how to use detection and monitoring information to achieve full-time and space stress field inversion and make up for the limitations of numerical simulation methods still needs further research. With the continuous advancement and development of science and technology, breakthroughs in deep learning methods have attracted people's attention to intelligent algorithms in stress field inversion. In particular, the use of diffusion and time series prediction neural networks can effectively extract the features of different types of data. This provides the possibility of using deep learning methods to achieve accurate inversion and real-time reproduction of the stress field and its evolution of loaded coal and rock mass, but there is still a lack of specific feasible ideas and implementation methods.

[0005] Therefore, it is urgent to carry out actual measurement of detection and monitoring signals during the loading process of coal and rock masses, and to establish the conversion relationship between detection-monitoring information and stress values, and between local stress values ​​and full-space stress values. At the same time, a method and device for inversion of the full-time and space stress field of loaded coal and rock based on deep learning is proposed, which has a significant driving effect on revealing the "black box" evolution process of coal and rock dynamic disasters and realizing early warning of potential disasters. Summary of the invention

[0006] In response to the problems and needs raised above, this solution proposes a method and device for inversion of the full-time and space stress field of loaded coal and rock based on deep learning. Due to the adoption of the following technical features, the above technical objectives can be achieved and many other technical effects can be brought about.

[0007] An object of the present invention is to propose a method for inverting the full-time and space stress field of loaded coal and rock based on deep learning, comprising the following steps:

[0008] S10: Arranging a detection-monitoring sensor and a stress sensor in the coal-rock mass, and placing the coal-rock mass in a testing machine, and using the testing machine to perform a load fracture test on the coal-rock mass, and recording the collected detection-monitoring data, stress data and boundary condition information during the test; wherein the full-time velocity field, fracture field, internal strain and microcurrent data are obtained by the detection-monitoring sensor, and the full-time local stress value data at the detection-monitoring position is obtained by the stress sensor;

[0009] S20: dividing the data measured by the detection-monitoring sensors and stress sensors at all times into a training set, a validation set, and a test set according to the corresponding time; using the encoder in the sequence transcription neural network to extract features from the detection-monitoring data in the training set and the validation set, and using the decoder in the sequence transcription neural network to extract features from the local stress data output at the previous moment, and training to obtain an information transcription model;

[0010] S30: input the full-time detection-monitoring data in the test set into the information transcription model to obtain the full-time local stress value, and verify the transcription result of the information transcription model by the stress value at the detection-monitoring sensor position measured by the stress sensor, and adjust the information transcription model parameters according to the verification result. The obtained information transcription model transcribes the full-time detection-monitoring data into the full-time local stress value at the detection-monitoring sensor position;

[0011] S40: According to the loading boundary conditions of the coal rock load fracture test, a numerical simulation model is established in combination with the mechanical parameters and constitutive equations of the coal rock mass. The full-space stress field data during the coal rock load fracture test is obtained through the numerical simulation method, and the numerical simulation method is optimized through the local stress value obtained by the information transcription model;

[0012] S50: the local stress values ​​at all times obtained by the information transcription model and the stress field data of the entire space obtained by the numerical simulation are divided into a training set, a validation set and a test set; the diffusion neural network extracts features of the stress field data at the current time in the training set and the validation set through a diffusion process, extracts features of the local stress value data at the current time through an inverse diffusion process, and calculates the feature similarity between the extracted local stress values ​​and the stress field of the entire space, and trains to obtain an inversion model of the stress field of the entire space and time;

[0013] S60: The full-time local stress value obtained by transcribing the information transcription model is input into the full-time and space stress field inversion model to invert the full-time and space stress field, and the inversion result is verified by the stress value measured by the stress sensor in the coal rock load fracture test, and the model parameters are adjusted according to the verification result. The obtained full-time and space stress field inversion model obtains the full-time and space stress field during the coal rock load fracture test through the full-time local stress value of the coal rock.

[0014] In addition, the full-time and space stress field inversion method of loaded coal and rock based on deep learning according to the present invention can also have the following technical features:

[0015] In an example of the present invention, in step S20, feature extraction is performed on the detection-monitoring data in the training set and the validation set using an encoder in a sequence transcription neural network, including the following steps:

[0016] First, read t through the input layer nThe detection and monitoring data at each moment is converted into a tensor format, the data is position-encoded, and the Q, K, and V auxiliary matrices of the multi-head self-attention mechanism are used to calculate the position of the data. n The detection-monitoring data at each moment is used for feature extraction, and the formula is as follows:

[0017]

[0018] Among them, MHSA stands for Multi-Head Self-Attention; Q stands for the query module, which is responsible for querying the relationship between it and the queried features; K stands for the queried module, which is responsible for accepting the query of the Q module; V module represents the actual feature information calculated; Z n-1 and Z n represents the detection-monitoring data features extracted by the self-attention mechanism; X n-1 and X n Represents the tensor format of the current input - monitoring data; softmax represents the activation function; d k Represents the dimension size of Q and K matrices;

[0019] Then, the input layer and output layer features of the multi-head self-attention mechanism are concatenated through residual connection, and the concatenated features are layer normalized. The formula is as follows:

[0020]

[0021] Among them, LN represents the Layer-Normalization layer normalization operation; α and β represent the parameter vectors of scaling and translation respectively; u L and σ L Represent the mean and variance respectively; ∈ represents a constant 10 -5 , to prevent the denominator from being 0; L n-1 and L n The features obtained by the normalization operation of the representative layer; X n-1 +Z n-1 and X n +Z n Represents the feature concatenation result obtained by residual connection of features extracted by input detection-monitoring and multi-head self-attention mechanism;

[0022] Next, the features obtained by the layer normalization operation are input into the feedforward neural network, and a linear transformation is performed through the feedforward neural network to map the input features from the low-dimensional space to the high-dimensional space. The formula is as follows:

[0023] OUT n-1 =FNN(L n-1 )=max(0,L n-1W1+b1)W2+b2

[0024] OUT n =FNN(L n )=max(0,L n W1+b1)W2+b2

[0025] Among them, FNN stands for Feed Forward Neural Network; OUT n-1 and OUT n Represents the characteristics of the feedforward neural network output; W1 and W2 represent weight parameters; b1 and b2 represent bias parameters;

[0026] Finally, the K and V auxiliary matrix features of the input layer and output layer of the feedforward neural network are concatenated through residual connection, and the concatenated features are subjected to layer normalization. The formula is as follows:

[0027]

[0028] Among them, LN n-1 and LN n The features obtained by the layer normalization operation; L n-1 +OUT n-1 and L n +OUT n Represents the feature concatenation result obtained by concatenating the input features and the output features of the feedforward neural network through residual connection.

[0029] In one example of the present invention, in step S20, the decoder in the sequence transcription neural network is used to extract features from the local stress data output at the previous moment, including the following steps:

[0030] First, read t through the input layer n-1 The local stress value data output at each moment is converted into a tensor format, the data is position-encoded, and the input local stress value data is masked through the masked multi-head self-attention mechanism. The Q, K, and V auxiliary matrices are used to calculate the local stress value data. n The local stress value data at the moment is used for feature extraction, and the formula is as follows:

[0031]

[0032] Among them, z n-1 and z n represents the local stress value data feature extracted by the multi-head self-attention mechanism; x n-1 and x n Represents the local stress value data in tensor format input at the current moment;

[0033] Secondly, the Q auxiliary matrix features of the input layer and output layer of the multi-head self-attention mechanism are concatenated through residual connection, and the concatenated features are layer normalized. The formula is as follows:

[0034]

[0035] in, and The features obtained by the normalization operation of the representative layer; x n-1 +z n-1 and x n +z n Represents the feature concatenation result obtained by residual connection of the input local stress value and the features extracted by the multi-head self-attention mechanism;

[0036] Again, the multi-head self-attention mechanism is used to extract features from the features calculated by the K and V auxiliary matrices in the encoder and the Q auxiliary matrix features after layer normalization. The formula is as follows:

[0037]

[0038] Among them, y n-1 and n Represents the data features extracted by the multi-head self-attention mechanism; and and and They represent the Q auxiliary matrix features after the normalization operation of the input, and the K and V auxiliary matrix features after the normalization operation of the encoder output;

[0039] Then, the features of the input layer and output layer of the multi-head self-attention mechanism are concatenated through residual connection, and the concatenated features are layer normalized. The formula is as follows:

[0040]

[0041] Among them, ln n-1 and ln n The features obtained by the normalization operation of the representative layer;

[0042] Next, the features obtained by the layer normalization operation are input into the feedforward neural network, and a linear transformation is performed through the feedforward neural network to map the input features from the low-dimensional space to the high-dimensional space. The formula is as follows:

[0043] out n-1 =FNN(ln n-1 )=max(0,ln n-1 W1+b1)W2+b2

[0044] out n =FNN(ln n )=max(0,ln n W1+b1)W2+b2

[0045] Among them, out n-1 and out n Features representing the output of a feedforward neural network;

[0046] Again, the features of the input layer and the output layer of the feedforward neural network are concatenated through residual connection, and the concatenated features are subjected to layer normalization. The formula is as follows:

[0047]

[0048] Among them, Y n-1 and Y n The local stress characteristics obtained by the normalization operation of the representative layer;

[0049] Finally, Y is quantized by Softmax. n-1 and Y n The features are processed to obtain the local stress value with the highest probability of accuracy as the result output;

[0050]

[0051] Among them, σ n-1 and σ n Represents the local stress value results of the final output.

[0052] In one example of the present invention, in step S40, the full-space stress field data during the coal-rock load fracture test is obtained by a numerical simulation method, and the numerical simulation method is optimized by the local stress value obtained by the information transcription model, including the following steps:

[0053] Add material properties and set boundary conditions to the numerical simulation model according to the coal rock mass parameters of the load fracture test;

[0054] Add solid mechanics physics and set the material model to linear elastic material-soil plasticity;

[0055] Meshing of numerical simulation models;

[0056] The numerical simulation model is calculated to obtain the full-time and space stress field of the coal-rock mass during the load-breaking test;

[0057] The numerical simulation results are optimized by using the local stress values ​​obtained through the information transcription model.

[0058] In one example of the present invention, adding material properties and setting boundary conditions to the numerical simulation model according to the coal rock mass parameters of the load fracture test includes the following steps:

[0059] First, set the material properties details: Young's modulus, Poisson's ratio, density, cohesion and internal friction angle;

[0060] Then, the boundary conditions of the coal rock mass are set using the Solid Mechanics-Boundary Load and Fixed Constraint in the COMSOL component. The equations are as follows:

[0061] S·n=F A

[0062] F A =-pn

[0063] u=0

[0064] Where S represents the stress tensor; n represents the unit normal vector; F A represents the external force acting on the surface; p represents the normal stress; and u represents the displacement.

[0065] In one example of the present invention, adding a solid mechanics physical field and setting the material model to linear elastic material-soil plasticity includes the following steps:

[0066] The solid mechanics-linear elastic material in the COMSOL component is used to model the coal rock mass (1000), and its equation is as follows:

[0067]

[0068] s=s inel +s el

[0069] ∈ el =∈-∈ inel

[0070] ∈ inel =∈0+∈ ext +∈ th +∈ hs +∈ pl +∈ cr +∈ vp +∈ ve

[0071] s el =c:∈ el

[0072] s inel =s0+s ext +s q

[0073]

[0074] c=c(E,v)

[0075] in, represents the divergence of the tensor; s represents the stress tensor; F v represents body force; s represents the total stress tensor; s inel represents the inelastic stress tensor; s el represents the elastic stress tensor; ∈ represents the total strain; ∈ inel represents elastic strain; ∈ el represents the inelastic strain; ∈0 represents the initial strain; ∈ ext represents the strain caused by external load; ∈ th represents thermal strain; ∈ hs Represents historical strain; ∈ pl represents plastic strain; ∈ cr represents creep strain; ∈ vp represents plastic strain; ∈ ve represents the dissolution strain; c represents the elastic tensor; s0 represents the initial stress; s ext represents the stress caused by external load; s q represents the inelastic mechanism stress; u represents the displacement field; E represents the elastic modulus; v represents the Poisson's ratio;

[0076] Using soil plasticity in COMSOL components, the yield criterion uses the Drucker-Prager criterion to match the Mohr-Coulomb criterion to model the coal rock mass, and its equation is as follows:

[0077]

[0078] in, represents the plastic strain rate; λ represents the plastic multiplier; Q represents the plastic potential function; S represents the short-term shear stress; F(σ,σ ys ) represents the function of elastic-plastic conditions; σ represents stress; σ ys represents the yield stress; λF=0 represents the complementary relaxation condition; F cone represents the plasticity criterion; J2 represents the De Voia stress invariant, the intensity of the stress state; α represents the material strength parameter; I1 represents the average intensity of the stress state; k represents a constant; φ represents the internal friction angle of the material.

[0079] In one example of the present invention, in step S50, feature extraction is performed on the local stress value data at the current moment through the reverse diffusion process, and the feature similarity between the extracted local stress value and the full-space stress field is calculated, including the following steps:

[0080] First, the diffusion neural network processes the full-space stress field data at the current moment into noise data that conforms to the Gaussian distribution through the forward Markov diffusion process. The formula of the Gaussian distribution is as follows:

[0081]

[0082] Among them, N(x) represents Gaussian distribution; x represents random variable; exp represents e x ; μ represents mathematical expectation; σ represents standard deviation;

[0083] Among them, the formula of the forward Markov diffusion process is as follows:

[0084]

[0085] Among them, q represents the distribution of the full-space stress field data at different times; β t represents the diffusion rate at time t; z t represents the noisy variant of z0 at time t; z t-1 represents the noisy variant of z0 at time t-1; N represents Gaussian distribution; T represents the full time step of the Markov diffusion process;

[0086] Then, the feature similarity between the local stress value at the current moment and the full-space stress field data is calculated through the inverse process with U-Net as the backbone. The stress field inversion model obtained after training can obtain the full-time and space stress field through the inversion of the local stress value at all times. The formula is as follows:

[0087]

[0088] z=ε(x)

[0089]

[0090] Among them, L SDM represents the mechanism of stress field inversion; x represents the full-space stress field of the current moment of the input diffusion neural network; ε represents the encoder, which can encode the input x into z; y represents the local stress value at the current moment; ∈ represents the added Gaussian noise; N(0,1) represents the Gaussian distribution; t represents the current moment; ∈ θ represents the U-Net neural network based on time conditions; τ θ represents an encoder that processes the local stress value at the current moment; D represents a decoder that can decode the obtained full-space stress field corresponding to the local stress value at the current moment from ε(x); Represents the final result of the full-space stress field at the current moment.

[0091] In one example of the present invention, the diffusion neural network uses an attention mechanism to enhance the U-Net neural network backbone, and the formula is as follows:

[0092]

[0093] in, Represents ∈ θ The intermediate process feature representation; τ θ (y) represents the encoding result of the encoder for the local stress value at the current moment; and They represent the learnable projection matrices respectively.

[0094] In an example of the present invention, in step S60, the inversion result is verified by the stress value measured by the stress sensor in the coal rock load fracture experiment, and the model parameters are adjusted according to the verification result, including the following steps:

[0095] The error between the stress value measured by the stress sensor in the coal-rock mass load fracture experiment and the stress field result obtained by the stress field inversion model is calculated by MAE; if the calculated error is large, the model parameters are adjusted; among them, the error calculation formula between the stress value measured by the stress sensor in the coal-rock mass load fracture experiment and the stress field result obtained by the stress field inversion model is as follows:

[0096]

[0097] Among them, MAE stands for Mean Absolute Error; y i and They respectively represent the stress value measured by the stress sensor in the coal rock load fracture experiment and the inversion results of the stress field inversion model.

[0098] Another object of the present invention is to provide a device for inverting the full-time and space stress field of loaded coal and rock based on deep learning, comprising:

[0099] An information acquisition unit is configured to arrange a detection-monitoring sensor and a stress sensor in the coal-rock mass, and place the coal-rock mass in a testing machine, and the testing machine performs a load-breaking test on the coal-rock mass, and records the collected detection-monitoring data, stress data and boundary condition information during the test; wherein the full-time velocity field, fracture field, internal strain and microcurrent data are obtained by the detection-monitoring sensor, and the local stress value data at the full-time detection-monitoring position is obtained by the stress sensor;

[0100] The transcription model unit is configured to divide the data measured by the detection-monitoring sensor and the stress sensor at all times into a training set, a validation set, and a test set according to the corresponding time; use the encoder in the sequence transcription neural network to extract features from the detection-monitoring data in the training set and the validation set, and use the decoder in the sequence transcription neural network to extract features from the local stress data output at the previous moment, and train to obtain an information transcription model;

[0101] A parameter adjustment unit is configured to input the full-time detection-monitoring data in the test set into the information transcription model to obtain the full-time local stress value, and verify the transcription result of the information transcription model by the stress value at the position of the detection-monitoring sensor measured by the stress sensor, and adjust the parameters of the information transcription model according to the verification result, so that the obtained information transcription model transcribes the full-time detection-monitoring data into the full-time local stress value at the position of the detection-monitoring sensor;

[0102] The simulation model unit is configured to establish a numerical simulation model according to the loading boundary conditions of the coal-rock load fracture test, in combination with the mechanical parameters and constitutive equations of the coal-rock mass, obtain the full-space stress field data during the coal-rock load fracture test by means of a numerical simulation method, and optimize the numerical simulation method by means of the local stress value obtained by the information transcription model;

[0103] The inversion model unit is configured to divide the local stress values ​​at all times obtained by the information transcription model and the stress field data of all spaces obtained by the numerical simulation into a training set, a validation set and a test set; the diffusion neural network extracts features of the stress field data at the current time in the training set and the validation set through a diffusion process, extracts features of the local stress value data at the current time through an inverse diffusion process, and calculates the feature similarity between the extracted local stress value and the stress field of all spaces, thereby training to obtain an inversion model of the stress field of all spaces;

[0104] The inversion calculation unit is configured to input the full-time local stress value obtained by transcribed from the information transcription model into the full-time and space stress field inversion model to invert the full-time and space stress field, and verify the inversion result through the stress value measured by the stress sensor in the coal rock load fracture experiment, adjust the model parameters according to the verification result, and obtain the full-time and space stress field inversion model to obtain the full-time and space stress field during the coal rock load fracture test through the full-time local stress value of the coal rock.

[0105] Compared with the prior art, the present invention has the following beneficial effects:

[0106] 1. The present invention focuses on the deep learning method to invert the full-time and space stress field of the loaded coal-rock mass through the coal-rock mass load fracture experiment combined with numerical simulation, combining sequence transcription and diffusion neural network to achieve high-precision inversion and full-time and space reproduction of the stress field and its evolution of the loaded coal-rock mass. Compared with the traditional numerical simulation method, this deep learning-based full-time and space stress field inversion method and device for loaded coal-rock has higher computational efficiency and better adaptability, and can effectively make up for the limitations of traditional methods.

[0107] 2. The information transcription model in the present invention is obtained by performing joint feature extraction of detection-monitoring information and local stress values ​​at all times through a sequence transcription neural network, and a method for converting detection-monitoring information and local stress values ​​at all times during the loading process of coal and rock masses is proposed. This includes the establishment of a quantitative association between the detection and monitoring data obtained in laboratory experiments and stress field parameters. Through this conversion relationship, the detection and monitoring information obtained in the laboratory can be converted into the local stress values ​​of the actual coal and rock masses, laying a certain foundation for the subsequent stress field inversion.

[0108] 3. The stress field inversion model in the present invention is obtained by extracting the joint features of the local stress value at all times and the stress field in all spaces through a diffusion neural network, and a full-time and full-space stress field inversion method based on the local stress value at all times during the loading process of coal and rock mass is proposed, including the feature similarity between the full-time local stress value obtained by the information transcription model and the full-space stress field obtained by numerical simulation. The stress field inversion model obtained after training combined with the information transcription model can directly realize the full-time and full-space stress field inversion during the loading process of coal and rock mass through the detection-monitoring information, revealing the black box process of stress field inversion of coal and rock mass during the loading process.

[0109] The best embodiment for carrying out the present invention will be described in more detail below with reference to the accompanying drawings so that the features and advantages of the present invention can be easily understood. BRIEF DESCRIPTION OF THE DRAWINGS

[0110] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings of the embodiments of the present invention, wherein the drawings are only used to illustrate some embodiments of the present invention, but not to limit all embodiments of the present invention thereto.

[0111] Figure 1 It is a flow chart of a method and device for inversion of full-time and space stress field of loaded coal and rock and prediction of boundary conditions based on deep learning according to an embodiment of the present invention;

[0112] Figure 2 It is a schematic diagram of the structure of the arrangement of the detection device on the coal rock mass according to an embodiment of the present invention;

[0113] Figure 3A schematic diagram of the structure of a sequence transcription model according to an embodiment of the present invention;

[0114] Figure 4 Schematic diagram of the stress field inversion model structure according to an embodiment of the present invention.

[0115] List of reference numerals:

[0116] Coal and rock mass 1000;

[0117] Detection-monitoring sensor 100;

[0118] Ultrasonic phased array detection sensor 10;

[0119] Multi-frequency acoustic wave detection sensor 20;

[0120] Optical fiber micro-nano monitoring sensor 30;

[0121] Microcurrent monitoring sensor 40;

[0122] Stress sensor 200. DETAILED DESCRIPTION

[0123] In order to make the purpose, technical solution and advantages of the technical solution of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described in conjunction with the drawings of specific embodiments of the present invention. The same figure marks in the drawings represent the same parts. It should be noted that the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0124] Unless otherwise defined, the technical terms or scientific terms used herein shall have the usual meanings understood by persons of ordinary skill in the field to which the present invention belongs. The words "first", "second" and similar words used in the patent application specification and claims of the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one" or "one" do not necessarily indicate a quantity limitation. Words such as "include" or "comprise" mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0125] According to the first aspect of the present invention, a method for inverting the full-time and space stress field of loaded coal and rock based on deep learning is as follows: Figure 1 As shown, the following steps are included:

[0126] S10: Arrange a detection-monitoring sensor and a stress sensor 200 in the coal-rock mass, and place the coal-rock mass in a testing machine, and use the testing machine to perform a load fracture test on the coal-rock mass, and record the collected detection-monitoring data, stress data and boundary condition information during the test; wherein, the full-time velocity field, fracture field, internal strain and microcurrent data are obtained by the detection-monitoring sensor, and the full-time local stress value data at the detection-monitoring position is obtained by the stress sensor 200;

[0127] S20: dividing the data measured by the detection-monitoring sensor and the stress sensor 200 at all times into a training set, a validation set, and a test set according to the corresponding time; using the encoder in the sequence transcription neural network to extract features from the detection-monitoring data in the training set and the validation set, and using the decoder in the sequence transcription neural network to extract features from the local stress data output at the previous moment, and training to obtain an information transcription model;

[0128] S30: input the full-time detection-monitoring data in the test set into the information transcription model to obtain the full-time local stress value, and verify the transcription result of the information transcription model by the stress value at the detection-monitoring sensor position measured by the stress sensor 200, and adjust the information transcription model parameters according to the verification result. The obtained information transcription model transcribes the full-time detection-monitoring data into the full-time local stress value at the detection-monitoring sensor position;

[0129] S40: According to the loading boundary conditions of the coal rock load fracture test, a numerical simulation model is established in combination with the mechanical parameters and constitutive equations of the coal rock mass. The full-space stress field data during the coal rock load fracture test is obtained through the numerical simulation method, and the numerical simulation method is optimized through the local stress value obtained by the information transcription model;

[0130] S50: the local stress values ​​at all times obtained by the information transcription model and the stress field data of the entire space obtained by the numerical simulation are divided into a training set, a validation set and a test set; the diffusion neural network extracts features of the stress field data at the current time in the training set and the validation set through a diffusion process, extracts features of the local stress value data at the current time through an inverse diffusion process, and calculates the feature similarity between the extracted local stress values ​​and the stress field of the entire space, and trains to obtain an inversion model of the stress field of the entire space and time;

[0131] S60: The full-time local stress value obtained by transcribe the information transcription model is input into the full-time and space stress field inversion model to invert the full-time and space stress field, and the inversion result is verified by the stress value measured by the stress sensor 200 in the coal rock load fracture test, and the model parameters are adjusted according to the verification result. The obtained full-time and space stress field inversion model obtains the full-time and space stress field during the coal rock load fracture test through the full-time local stress value of the coal rock.

[0132] The present invention focuses on the deep learning method to invert the full-time and space stress field of the loaded coal-rock mass through the coal-rock mass load fracture experiment combined with numerical simulation, and combines sequence transcription and diffusion neural network to achieve high-precision inversion and full-time and space reproduction of the stress field and its evolution of the loaded coal-rock mass. Compared with the traditional numerical simulation method, this deep learning-based full-time and space stress field inversion method and device for loaded coal and rock has higher computational efficiency and better adaptability, and can effectively make up for the limitations of traditional methods.

[0133] The information transcription model in the present invention is obtained by performing joint feature extraction of detection-monitoring information and local stress values ​​at all times through a sequence transcription neural network, and a method for converting detection-monitoring information and local stress values ​​at all times during the loading process of coal and rock masses is proposed. This includes the establishment of a quantitative association between the detection and monitoring data obtained in laboratory tests and stress field parameters. Through this conversion relationship, the detection and monitoring information obtained in the laboratory can be converted into the local stress values ​​of the actual coal and rock masses, laying a certain foundation for the subsequent stress field inversion.

[0134] The stress field inversion model in the present invention is obtained by extracting the joint features of the local stress value at all times and the stress field in all spaces through a diffusion neural network, and a full-time and space stress field inversion method based on the local stress value at all times during the loading process of coal and rock mass is proposed, including the feature similarity between the full-time local stress value obtained by the extraction information transcription model and the full-space stress field obtained by numerical simulation. The stress field inversion model obtained after training combined with the information transcription model can directly realize the full-time and space stress field inversion during the loading process of coal and rock mass through the detection-monitoring information, revealing the black box process of stress field inversion during the loading process of coal and rock mass.

[0135] In one example of the present invention, Figure 2 As shown, the detection-monitoring sensor 100 includes:

[0136] Ultrasonic phased array detection sensor 10, multi-frequency acoustic wave detection sensor 20, optical fiber micro-nano monitoring sensor 30 and micro-current monitoring sensor 40, wherein the ultrasonic phased array sensor is arranged on the top of the coal rock mass 1000; the multi-frequency acoustic wave detection sensor 20 and the micro-current monitoring sensor 40 are arranged on the side of the coal rock mass 1000; the optical fiber micro-nano monitoring sensor 30 and the stress sensor 200 are arranged inside the coal rock mass 1000.

[0137] In an example of the present invention, in step S10, before the testing machine performs a load fracture test on the coal rock mass, the following steps are also included:

[0138] Set the working frequency of the ultrasonic phased array detection sensor 10 and the ultrasonic shear wave velocity; set the working frequency of the multi-frequency acoustic wave detection sensor 20 according to the corresponding waveform channel; the optical fiber micro-nano instrument has multiple channels, each channel is connected in parallel with multiple sensors, and its scanning frequency is set; the micro-current instrument has multiple channels, each channel is connected in parallel with multiple sensors, and its sampling rate is set; set the sampling frequency and monitoring position of the stress sensor 200; set the loading method of the testing machine and the data to be collected.

[0139] In an example of the present invention, in step S20, the division of the training set, the validation set and the test set in the sequence transcription neural network includes the following steps:

[0140] The coal rock load fracture test process from t0 to t n The full-time detection-monitoring and local stress value data measured by the stress sensor 200 collected are grouped according to the corresponding time, so that the detection-monitoring and local stress value data at each time correspond to each other. The grouped data are divided into a training set, a validation set and a test set according to the proportion. The training set and the validation set data are used to train the sequence transcription neural network to obtain the information transcription model, and the test set is used to test the information transcription model obtained by training.

[0141] In one example of the present invention, in step S20, the encoder in the sequence transcription neural network is used to extract features from the detection-monitoring data in the training set and the validation set, such as Figure 3 As shown, the following steps are included:

[0142] First, read t through the input layer n The detection and monitoring data at each moment is converted into a tensor format, the data is position-encoded, and the Q, K, and V auxiliary matrices of the multi-head self-attention mechanism are used to calculate the position of the data. n The detection-monitoring data at each moment is used for feature extraction, and the formula is as follows:

[0143]

[0144] Among them, MHSA stands for Multi-Head Self-Attention; Q stands for the query module, which is responsible for querying the relationship between it and the queried features; K stands for the queried module, which is responsible for accepting the query of the Q module; V module represents the actual feature information calculated; Z n-1 and Z n represents the detection-monitoring data features extracted by the self-attention mechanism; X n-1 and X n Represents the tensor format of the current input - monitoring data; softmax represents the activation function; d k Represents the dimension size of Q and K matrices; Avoid when the feature gap between Q and K is too large, and the calculation result of softmax is close to 1;

[0145] Then, the input layer and output layer features of the multi-head self-attention mechanism are concatenated through residual connection, and the concatenated features are layer normalized. The formula is as follows:

[0146]

[0147] Among them, LN represents the Layer-Normalization layer normalization operation; α and β represent the parameter vectors of scaling and translation respectively; u L and σ L Represent the mean and variance respectively; ∈ represents a constant 10 -5 , to prevent the denominator from being 0; L n-1 and L n The features obtained by the normalization operation of the representative layer; X n-1 +Z n-1 and X n +Z n Represents the feature concatenation result obtained by residual connection of features extracted by input detection-monitoring and multi-head self-attention mechanism;

[0148] Next, the features obtained by the layer normalization operation are input into the feedforward neural network, and a linear transformation is performed through the feedforward neural network to map the input features from the low-dimensional space to the high-dimensional space. The formula is as follows:

[0149] OUT n-1 =FNN(L n-1 )=max(0,L n-1 W1+b1)W2+b2

[0150] OUT n =FNN(L n )=max(0,L n W1+b1)W2+b2

[0151] Among them, FNN stands for Feed Forward Neural Network; OUT n-1 and OUT n Represents the characteristics of the feedforward neural network output; W1 and W2 represent weight parameters; b1 and b2 represent bias parameters;

[0152] Finally, the K and V auxiliary matrix features of the input layer and output layer of the feedforward neural network are concatenated through residual connection, and the concatenated features are subjected to layer normalization. The formula is as follows:

[0153]

[0154]

[0155] Among them, LN n-1 and LN n The features obtained by the layer normalization operation; L n-1 +OUT n-1 and L n +OUT n Represents the feature concatenation result obtained by concatenating the input features and the output features of the feedforward neural network through residual connection.

[0156] In one example of the present invention, in step S20, as Figure 3 As shown, the decoder in the sequence transcription neural network is used to extract features from the local stress data output at the previous moment, including the following steps:

[0157] First, read t through the input layer n-1 The local stress value data output at each moment is converted into a tensor format, the data is position-encoded, and the input local stress value data is masked through the masked multi-head self-attention mechanism. The Q, K, and V auxiliary matrices are used to calculate the local stress value data. n The local stress value data at the moment is used for feature extraction, and the formula is as follows:

[0158]

[0159] Among them, z n-1 and z n represents the local stress value data feature extracted by the multi-head self-attention mechanism; x n-1 and x n Represents the local stress value data in tensor format input at the current moment;

[0160] Secondly, the Q auxiliary matrix features of the input layer and output layer of the multi-head self-attention mechanism are concatenated through residual connection, and the concatenated features are layer normalized. The formula is as follows:

[0161]

[0162] in, and The features obtained by the normalization operation of the representative layer; x n-1 +z n-1 and x n +z n Represents the feature concatenation result obtained by residual connection of the input local stress value and the features extracted by the multi-head self-attention mechanism;

[0163] Again, the multi-head self-attention mechanism is used to extract features from the features calculated by the K and V auxiliary matrices in the encoder and the Q auxiliary matrix features after layer normalization. The formula is as follows:

[0164]

[0165] Among them, y n-1 and n Represents the data features extracted by the multi-head self-attention mechanism; and and and They represent the Q auxiliary matrix features after the normalization operation of the input, and the K and V auxiliary matrix features after the normalization operation of the encoder output;

[0166] Then, the features of the input layer and output layer of the multi-head self-attention mechanism are concatenated through residual connection, and the concatenated features are layer normalized. The formula is as follows:

[0167]

[0168] Among them, ln n-1 and ln n The features obtained by the normalization operation of the representative layer;

[0169] Next, the features obtained by the layer normalization operation are input into the feedforward neural network, and a linear transformation is performed through the feedforward neural network to map the input features from the low-dimensional space to the high-dimensional space. The formula is as follows:

[0170] out n-1 =FNN(ln n-1 )=max(0,ln n-1 W1+b1)W2+b2

[0171] out n =FNN(ln n )=max(0,ln n W1+b1)W2+b2

[0172] Among them, out n-1 and out n Features representing the output of a feedforward neural network;

[0173] Again, the features of the input layer and output layer of the feedforward neural network are concatenated through residual connection, and the concatenated features are subjected to layer normalization. The formula is as follows:

[0174]

[0175] Among them, y n-1 and n The local stress characteristics obtained by the normalization operation of the representative layer;

[0176] Finally, Y is quantized by Softmax. n-1 and Y n The features are processed to obtain the local stress value with the highest probability of accuracy as the result output;

[0177]

[0178] Among them, σ n-1 and σ n Represents the local stress value results of the final output.

[0179] It should be noted that, regarding the sequence transcription model, the source of the sequence to neural network is time series translation. Chinese and English are time series data, which can be understood as different characters corresponding to each moment. Similarly, the detection-monitoring data and stress value data obtained in the laboratory are also time series data, and can also be translated by this model; among them, the neural network for RNN training can be LSTM, GRU and BI-LSTM; this method is similar to the neural network based on the Transformer architecture.

[0180] In an example of the present invention, in step S30, the transcription result of the information transcription model is verified by the stress value at the position of the detection-monitoring sensor 100 measured by the stress sensor 200, and the model parameters are adjusted according to the verification result, including the following steps:

[0181] The error between the transcription result of the information transcription model and the stress value measured by the stress sensor 200 is calculated by MSE; if the calculated error is large, the model parameters are adjusted; wherein the calculation formula for the error between the transcription result of the information transcription model and the stress value measured by the stress sensor 200 is as follows:

[0182]

[0183] Where MSE stands for Mean Squared Error; N is the length of the prediction sequence; y σp and σt They respectively represent the transcription result of the information transcription model and the stress value measured by the stress sensor 200.

[0184] In one example of the present invention, in step S40, the full-space stress field data during the coal-rock load fracture test is obtained by a numerical simulation method, and the numerical simulation method is optimized by the local stress value obtained by the information transcription model, including the following steps:

[0185] Add material properties and set boundary conditions to the numerical simulation model according to 1000 parameters of coal rock mass under load fracture test;

[0186] Add solid mechanics physics and set the material model to linear elastic material-soil plasticity;

[0187] Meshing of numerical simulation models;

[0188] The numerical simulation model was calculated to obtain the 1000 full-time and space stress field of the coal-rock mass during the load-breaking test;

[0189] The numerical simulation results are optimized by using the local stress values ​​obtained through the information transcription model.

[0190] In one example of the present invention, adding material properties and setting boundary conditions to the numerical simulation model according to the coal rock mass 1000 parameters of the load fracture test includes the following steps:

[0191] First, set the material properties details: Young's modulus, Poisson's ratio, density, cohesion and internal friction angle;

[0192] Then, the solid mechanics-boundary load and fixed constraint in the COMSOL component are used to set the coal rock mass 1000 boundary condition, and its equation is as follows:

[0193] S·n=F A

[0194] F A =-pn

[0195] u=0

[0196] Where S represents the stress tensor; n represents the unit normal vector; F A represents the external force acting on the surface; p represents the normal stress; and u represents the displacement.

[0197] In one example of the present invention, adding a solid mechanics physical field and setting the material model to linear elastic material-soil plasticity includes the following steps:

[0198] The solid mechanics-linear elastic material in the COMSOL component is used to model the coal rock mass 1000, and its equation is as follows:

[0199]

[0200] s=s inel +s el

[0201] ∈ el =∈-∈ inel

[0202] ∈ inel =∈0+∈ ext +∈ th +∈ hs +∈ pl +∈ cr +∈ vp +∈ ve

[0203] s el =c:∈ el

[0204] s inel =s0+s ext +s q

[0205]

[0206] c=c(E,v)

[0207] in, represents the divergence of the tensor; s represents the stress tensor; F v represents body force; s represents the total stress tensor; s inel represents the inelastic stress tensor; s el represents the elastic stress tensor; ∈ represents the total strain; ∈ inel represents elastic strain; ∈ el represents the inelastic strain; ∈0 represents the initial strain; ∈ ext represents the strain caused by external load; ∈ th represents thermal strain; ∈ hs Represents historical strain; ∈ plrepresents plastic strain; ∈ cr represents creep strain; ∈ vp represents plastic strain; ∈ ve represents the dissolution strain; c represents the elastic tensor; s0 represents the initial stress; s ext represents the stress caused by external load; s q represents the specific inelastic mechanism stress; u represents the displacement field; E represents the elastic modulus; v represents the Poisson's ratio;

[0208] Using soil plasticity in COMSOL components, the yield criterion uses the Drucker-Prager criterion to match the Mohr-Coulomb criterion to model the coal rock mass, and its equation is as follows:

[0209]

[0210] in, represents the plastic strain rate; λ represents the plastic multiplier; Q represents the plastic potential function; S represents the short-term shear stress; F(σ,σ ys ) represents the function of elastic-plastic conditions; σ represents stress; σ ys represents the yield stress; λF=0 represents the complementary relaxation condition; F cone represents the plasticity criterion; J2 represents the De Voia stress invariant, the intensity of the stress state; α represents the material strength parameter; I1 represents the average intensity of the stress state; k represents a constant; φ represents the internal friction angle of the material.

[0211] In one example of the present invention, the error between the information transcription model and the local stress value obtained by numerical simulation is calculated by RMSE; if the calculated error is large, the numerical simulation parameters are adjusted; wherein the calculation formula for calculating the error between the information transcription model and the local stress value obtained by numerical simulation by RMSE is as follows:

[0212]

[0213] Where RMSE stands for Root Mean Squared Error; N is the length of the prediction sequence; y mod,i and sim,i Represent the local stress values ​​obtained by the information transcription model and numerical simulation, respectively.

[0214] In one example of the present invention, in step S50, the full-time local stress values ​​obtained by model transcription and the full-space stress field data obtained by numerical simulation are divided into a training set, a validation set and a test set, including the following steps:

[0215] The local stress values ​​at all times obtained by the information transcription model and the full-space stress field data obtained by numerical simulation are grouped according to the corresponding time, so that the local stress value at each time corresponds to the 1000 stress field data of the coal and rock mass. The grouped data are divided into training set, validation set and test set according to the proportion. The training set and validation set data are used to train the diffusion neural network to obtain the stress field inversion model, and the test set is used to test the trained stress field inversion model.

[0216] In one example of the present invention, in step S50, as Figure 4 As shown, the feature extraction of the local stress value data at the current moment is performed through the reverse diffusion process, and the feature similarity between the extracted local stress value and the full-space stress field is calculated, including the following steps:

[0217] First, the diffusion neural network processes the full-space stress field data at the current moment into noise data that conforms to the Gaussian distribution through the forward Markov diffusion process. The formula of the Gaussian distribution is as follows:

[0218]

[0219] Among them, N(x) represents Gaussian distribution; x represents random variable; exp represents e x ; μ represents mathematical expectation; σ represents standard deviation;

[0220] Among them, the formula of the forward Markov diffusion process is as follows:

[0221]

[0222] Among them, q represents the distribution of the full-space stress field data at different times; β t represents the diffusion rate at time t; z t represents the noisy variant of z0 at time t; z t-1 represents the noisy variant of z0 at time t-1; N represents Gaussian distribution; T represents the full time step of the Markov diffusion process;

[0223] Then, the feature similarity between the local stress value at the current moment and the full-space stress field data is calculated through the inverse process with U-Net as the backbone. The stress field inversion model obtained after training can obtain the full-time and space stress field through the inversion of the local stress value at all times. The formula is as follows:

[0224]

[0225] z=ε(x)

[0226]

[0227] Among them, L SDMrepresents the mechanism of stress field inversion; x represents the full-space stress field of the current moment of the input diffusion neural network; ε represents the encoder, which can encode the input x into z; y represents the local stress value at the current moment; ∈ represents the added Gaussian noise; N(0,1) represents the Gaussian distribution; t represents the current moment; ∈ θ represents the U-Net neural network based on time conditions; τ θ represents an encoder that processes the local stress value at the current moment; D represents a decoder that can decode the obtained full-space stress field corresponding to the local stress value at the current moment from ε(x); Represents the final result of the full-space stress field at the current moment.

[0228] In one example of the present invention, the diffusion neural network uses an attention mechanism to enhance the U-Net neural network backbone, and the formula is as follows:

[0229]

[0230] in, Represents ∈ θ The intermediate process feature representation; τ θ (y) represents the encoding result of the encoder for the local stress value at the current moment; and They represent the learnable projection matrices respectively.

[0231] In an example of the present invention, in step S60, the inversion result is verified by the stress value measured by the stress sensor 200 in the load fracture experiment of the coal rock mass 1000, and the model parameters are adjusted according to the verification result, including the following steps:

[0232] The error between the stress value measured by the stress sensor 200 in the coal-rock mass 1000 load fracture experiment and the stress field result obtained by the stress field inversion model is calculated by MAE; if the calculated error is large, the model parameters are adjusted; wherein, the error calculation formula between the stress value measured by the stress sensor 200 in the coal-rock mass 1000 load fracture experiment and the stress field result obtained by the stress field inversion model is as follows:

[0233]

[0234] Among them, MAE stands for Mean Absolute Error; y i and They respectively represent the stress value measured by the stress sensor 200 in the coal rock mass 1000 load fracture experiment and the inversion result of the stress field inversion model.

[0235] It should be noted that regarding the diffusion model, the diffusion idea of ​​generative modeling was actually proposed in 2015. However, it was not until 2019 that Stanford University (Song et al.) and 2020 that Google Brain improved this method, which triggered a new trend of generative models. At present, including OpenAI's GLIDE and DALL-E 2, Heidelberg University's LatentDiffusion and Google Brain's ImageGen, are all based on the diffusion model and can obtain high-quality generation effects. The principle of the diffusion model is to first add noise to the image in a forward direction. The step of adding noise is divided into 1000 steps until the image is added as randomly distributed Gaussian noise. Then, through the reverse process, let the diffusion model predict the Gaussian noise added in each step, delete the predicted Gaussian noise added by the forward process, and you can get the original image.

[0236] This method uses a guided diffusion model, and the model can generate the corresponding image part according to the prompt. That is, the local stress field is known, and the model generates the corresponding stress field according to the guidance of the known part. The diffusion model can be DDPM, Diffusion, Stable Diffusion (Guided Diffusion), for example, similar to this method and Stable Diffusion (Guided Diffusion).

[0237] In short, this method first needs to build a sequence transcription neural network and a diffusion neural network structure, and train the two neural networks to obtain the corresponding information transcription model and stress field inversion model.

[0238] According to the second aspect of the present invention, a device for inverting the full-time and space stress field of loaded coal and rock based on deep learning comprises:

[0239] The information acquisition unit is configured to arrange a detection-monitoring sensor 100 and a stress sensor 200 in the coal-rock body 1000, and place the coal-rock body 1000 in a testing machine, and the testing machine performs a load fracture test on the coal-rock body 1000, and records the collected detection-monitoring data, stress data and boundary condition information during the test; wherein the full-time velocity field, fracture field, internal strain and microcurrent data are obtained by the detection-monitoring sensor 100, and the full-time local stress value data at the detection-monitoring position is obtained by the stress sensor 200;

[0240] The transcription model unit is configured to divide the data measured by the detection-monitoring sensor 100 and the stress sensor 200 at all times into a training set, a validation set and a test set according to the corresponding time; use the encoder in the sequence transcription neural network to extract features from the detection-monitoring data in the training set and the validation set, and use the decoder in the sequence transcription neural network to extract features from the local stress data output at the previous moment, and train to obtain an information transcription model;

[0241] A parameter adjustment unit is configured to input the full-time detection-monitoring data in the test set into the information transcription model to obtain the full-time local stress value, and verify the transcription result of the information transcription model by the stress value at the position of the detection-monitoring sensor 100 measured by the stress sensor 200, and adjust the parameters of the information transcription model according to the verification result, so that the obtained information transcription model transcribes the full-time detection-monitoring data into the full-time local stress value at the position of the detection-monitoring sensor 100;

[0242] A simulation model unit is configured to establish a numerical simulation model according to the loading boundary conditions of the coal-rock load fracture test and in combination with the mechanical parameters and constitutive equation of the coal-rock mass 1000, obtain the full-space stress field data during the coal-rock load fracture test by a numerical simulation method, and optimize the numerical simulation method by using the local stress value obtained by the information transcription model;

[0243] The inversion model unit is configured to divide the local stress values ​​at all times obtained by the information transcription model and the stress field data of all spaces obtained by the numerical simulation into a training set, a validation set and a test set; the diffusion neural network extracts features of the stress field data at the current time in the training set and the validation set through a diffusion process, extracts features of the local stress value data at the current time through an inverse diffusion process, and calculates the feature similarity between the extracted local stress value and the stress field of all spaces, thereby training to obtain an inversion model of the stress field of all spaces;

[0244] The inversion calculation unit is configured to input the full-time local stress value obtained by transcribed from the information transcription model into the full-time and space stress field inversion model to invert the full-time and space stress field, and verify the inversion result through the stress value measured by the stress sensor 200 in the coal rock mass 1000 load fracture test, adjust the model parameters according to the verification result, and obtain the full-time and space stress field inversion model to obtain the full-time and space stress field during the coal rock load fracture test through the full-time local stress value of the coal rock mass 1000.

[0245] The present invention focuses on the deep learning method to invert the full-time and space stress field of the loaded coal rock mass 1000 through the coal rock mass 1000 load fracture experiment combined with numerical simulation, and combines sequence transcription and diffusion neural network to achieve high-precision inversion and full-time and space reproduction of the stress field and its evolution of the loaded coal rock mass 1000. Compared with the traditional numerical simulation method, this deep learning-based full-time and space stress field inversion method and device for loaded coal rock has higher computational efficiency and better adaptability, and can effectively make up for the limitations of traditional methods.

[0246] The information transcription model in the present invention is obtained by performing joint feature extraction of the detection-monitoring information and the local stress value at all times through a sequence transcription neural network, and a conversion method between the detection monitoring information and the local stress value at all times during the loading process of the coal rock mass 1000 is proposed. This includes the establishment of a quantitative association between the detection monitoring data obtained in the laboratory test and the stress field parameters. Through this conversion relationship, the detection monitoring information obtained in the laboratory can be converted into the local stress value of the actual coal rock mass 1000, which lays a certain foundation for the subsequent stress field inversion.

[0247] The stress field inversion model in the present invention is obtained by extracting the joint features of the local stress value at all times and the stress field in all spaces through a diffusion neural network, and a full-time and space stress field inversion method based on the local stress value at all times during the loading process of the coal rock mass 1000 is proposed, including the feature similarity between the local stress value at all times obtained by the extraction information transcription model and the full-time stress field obtained by numerical simulation. The stress field inversion model obtained after training combined with the information transcription model can directly realize the full-time and space stress field inversion of the coal rock mass 1000 during the loading process through the detection-monitoring information, revealing the black box process of the stress field inversion of the coal rock mass 1000 during the loading process.

[0248] The above describes in detail the exemplary implementation of the method and device for inversion of the full-time and space stress field of loaded coal and rock based on deep learning proposed in the present invention with reference to the preferred embodiments. However, those skilled in the art can understand that, without departing from the concept of the present invention, various modifications and variations can be made to the above-mentioned specific embodiments, and various technical features and structures proposed in the present invention can be combined in various ways without exceeding the protection scope of the present invention. The protection scope of the present invention is determined by the attached claims.

Claims

1. A method for inversion of full-time and space stress field of loaded coal and rock based on deep learning, characterized in that: The steps include: S10: arranging a detection-monitoring sensor (100) and a stress sensor (200) in the coal-rock mass (1000), and placing the coal-rock mass (1000) in a testing machine, and using the testing machine to perform a load-breaking test on the coal-rock mass (1000), and recording the collected detection-monitoring data, stress data, and boundary condition information during the test; wherein the detection-monitoring sensor (100) is used to obtain full-time velocity field, fracture field, internal strain, and microcurrent data, and the stress sensor (200) is used to obtain full-time local stress value data at the detection-monitoring position; S20: dividing the data measured by the detection-monitoring sensor (100) and the stress sensor (200) at all times into a training set, a validation set and a test set according to the corresponding time; using an encoder in a sequence transcription neural network to extract features from the detection-monitoring data in the training set and the validation set, and using a decoder in the sequence transcription neural network to extract features from the local stress data output at the previous time, and training to obtain an information transcription model; S30: inputting the full-time detection-monitoring data in the test set into the information transcription model to obtain the full-time local stress value, and verifying the transcription result of the information transcription model through the stress value at the position of the detection-monitoring sensor (100) measured by the stress sensor (200), and adjusting the parameters of the information transcription model according to the verification result, so that the obtained information transcription model transcribes the full-time detection-monitoring data into the full-time local stress value at the position of the detection-monitoring sensor (100); S40: According to the loading boundary conditions of the coal rock load fracture test, a numerical simulation model is established in combination with the mechanical parameters and constitutive equation of the coal rock mass (1000). The full-space stress field data during the coal rock load fracture test is obtained through the numerical simulation method, and the numerical simulation method is optimized through the local stress value obtained by the information transcription model; S50: the local stress values ​​at all times obtained by the information transcription model and the stress field data of the entire space obtained by the numerical simulation are divided into a training set, a validation set and a test set; the diffusion neural network extracts features of the stress field data at the current time in the training set and the validation set through a diffusion process, extracts features of the local stress value data at the current time through an inverse diffusion process, and calculates the feature similarity between the extracted local stress values ​​and the stress field of the entire space, and trains to obtain an inversion model of the stress field of the entire space and time; S60: The local stress values ​​at all times obtained by the information transcription model are input into the full-time and space stress field inversion model to obtain the full-time and space stress field, and the inversion results are verified by the stress values ​​measured by the stress sensor (200) in the coal rock mass (1000) load fracture test, and the model parameters are adjusted according to the verification results. The obtained full-time and space stress field inversion model obtains the full-time and space stress field during the coal rock load fracture test through the full-time local stress values ​​of the coal rock mass (1000).

2. The method for inversion of full-time and space stress field of loaded coal and rock based on deep learning according to claim 1 is characterized in that: In step S20, feature extraction is performed on the detection-monitoring data in the training set and the validation set using the encoder in the sequence transcription neural network, including the following steps: First, read t through the input layer n The detection and monitoring data at each moment is converted into a tensor format, the data is position-encoded, and the Q, K, and V auxiliary matrices of the multi-head self-attention mechanism are used to calculate the position of the data. n The detection-monitoring data at each moment is used for feature extraction, and the formula is as follows: Among them, MHSA stands for Multi-Head Self-Attention; Q stands for the query module, which is responsible for querying the relationship between it and the queried features; K stands for the queried module, which is responsible for accepting the query of the Q module; V module represents the actual feature information calculated; Z n-1 and Z n represents the detection-monitoring data features extracted by the self-attention mechanism; X n-1 and X n Represents the tensor format of the current input - monitoring data; softmax represents the activation function; d k Represents the dimension size of Q and K matrices; Then, the input layer and output layer features of the multi-head self-attention mechanism are concatenated through residual connection, and the concatenated features are layer normalized. The formula is as follows: Among them, LN represents the Layer-Normalization layer normalization operation; α and β represent the parameter vectors of scaling and translation respectively; u L and σ L Represent the mean and variance respectively; ∈ represents a constant 10 -5 , to prevent the denominator from being 0; L n-1 and L n The features obtained by the normalization operation of the representative layer; X n-1 +Z n-1 and X n +Z n Represents the feature concatenation result obtained by residual connection of features extracted by input detection-monitoring and multi-head self-attention mechanism; Next, the features obtained by the layer normalization operation are input into the feedforward neural network, and a linear transformation is performed through the feedforward neural network to map the input features from the low-dimensional space to the high-dimensional space. The formula is as follows: OUT n-1 =FNN(L n-1 )=max(0,L n-1 W1+b1)W2+b2 OUT n =FNN(L n )=max(0,L n W1+b1)W2+b2 Among them, FNN stands for Feed Forward Neural Network; OUT n-1 and OUT n Represents the characteristics of the feedforward neural network output; W1 and W2 represent weight parameters; b1 and b2 represent bias parameters; Finally, the K and V auxiliary matrix features of the input layer and output layer of the feedforward neural network are concatenated through residual connection, and the concatenated features are subjected to layer normalization. The formula is as follows: Among them, LN n-1 and LN n The features obtained by the layer normalization operation; L n-1 +OUT n-1 and L n +OUT n Represents the feature concatenation result obtained by concatenating the input features and the output features of the feedforward neural network through residual connection.

3. The method for inversion of full-time and space stress field of loaded coal and rock based on deep learning according to claim 1 is characterized in that: In the step S20, the decoder in the sequence transcription neural network is used to extract features from the local stress data output at the previous moment, including the following steps: First, read t through the input layer n-1 The local stress value data output at each moment is converted into a tensor format, the data is position-encoded, and the input local stress value data is masked through the masked multi-head self-attention mechanism. The Q, K, and V auxiliary matrices are used to calculate the local stress value data. n The local stress value data at the moment is used for feature extraction, and the formula is as follows: Among them, z n-1 and z n represents the local stress value data feature extracted by the multi-head self-attention mechanism; x n-1 and x n Represents the local stress value data in tensor format input at the current moment; Secondly, the Q auxiliary matrix features of the input layer and output layer of the multi-head self-attention mechanism are concatenated through residual connection, and the concatenated features are layer normalized. The formula is as follows: in, and The features obtained by the normalization operation of the representative layer; x n-1 +z n-1 and x n +z n Represents the feature concatenation result obtained by residual connection of the input local stress value and the features extracted by the multi-head self-attention mechanism; Again, the multi-head self-attention mechanism is used to extract features from the features calculated by the K and V auxiliary matrices in the encoder and the Q auxiliary matrix features after layer normalization. The formula is as follows: Among them, y n-1 and n Represents the data features extracted by the multi-head self-attention mechanism; and and and They represent the Q auxiliary matrix features after the normalization operation of the input, and the K and V auxiliary matrix features after the normalization operation of the encoder output; Then, the features of the input layer and output layer of the multi-head self-attention mechanism are concatenated through residual connection, and the concatenated features are layer normalized. The formula is as follows: Among them, ln n-1 and ln n The features obtained by the normalization operation of the representative layer; Next, the features obtained by the layer normalization operation are input into the feedforward neural network, and a linear transformation is performed through the feedforward neural network to map the input features from the low-dimensional space to the high-dimensional space. The formula is as follows: out n-1 =FNN(ln n-1 )=max(0,ln n-1 W1+b1)W2+b2 out n =FNN(ln n )=max(0,ln n W1+b1)W2+b2 Among them, out n-1 and out n Features representing the output of a feedforward neural network; Again, the features of the input layer and output layer of the feedforward neural network are concatenated through residual connection, and the concatenated features are subjected to layer normalization. The formula is as follows: Among them, Y n-1 and Y n The local stress characteristics obtained by the normalization operation of the representative layer; Finally, Y is quantized by Softmax. n-1 and Y n The features are processed to obtain the local stress value with the highest probability of accuracy as the result output; Among them, σ n-1 and σ n Represents the local stress value results of the final output.

4. The method for inversion of full-time and space stress field of loaded coal and rock based on deep learning according to claim 1 is characterized in that: In the step S40, the full-space stress field data during the coal-rock load fracture test is obtained by a numerical simulation method, and the numerical simulation method is optimized by the local stress value obtained by the information transcription model, including the following steps: Add material properties and set boundary conditions to the numerical simulation model according to the coal rock mass (1000) parameters of the load-breaking test; Add solid mechanics physics and set the material model to linear elastic material-soil plasticity; Meshing of numerical simulation models; The numerical simulation model is calculated to obtain the full-time and space stress field of the coal-rock mass (1000) during the load-breaking test; The numerical simulation results are optimized by using the local stress values ​​obtained through the information transcription model.

5. The method for inversion of full-time and space stress field of loaded coal and rock based on deep learning according to claim 4 is characterized in that: Adding material properties and setting boundary conditions to the numerical simulation model according to the coal-rock mass (1000) parameters of the load-breaking test includes the following steps: First, set the material properties details: Young's modulus, Poisson's ratio, density, cohesion and internal friction angle; Then, the boundary condition of the coal rock mass (1000) is set using the Solid Mechanics-Boundary Load and Fixed Constraint in the COMSOL component, and its equation is as follows: S·n=F A F A =-pn u＝0 Where S represents the stress tensor; n represents the unit normal vector; F A represents the external force acting on the surface; p represents the normal stress; and u represents the displacement.

6. The method for inversion of full-time and space stress field of loaded coal and rock based on deep learning according to claim 5 is characterized in that: Adding solid mechanics physics and setting the material model to linear elastic material-soil plasticity includes the following steps: The solid mechanics-linear elastic material in the COMSOL component is used to model the coal rock mass (1000), and its equation is as follows: s=s inel +s el ∈ el ＝∈-∈ inel ∈ inel ＝∈0+∈ ext +∈ th +∈ hs +∈ pl +∈ cr +∈ vp +∈ ve s el =c∶∈ el s inel =s0+s ext +s q c=c(E,v) in, represents the divergence of the tensor; s represents the stress tensor; F v represents body force; s represents the total stress tensor; s inel represents the inelastic stress tensor; s el represents the elastic stress tensor; ∈ represents the total strain; ∈ inel represents elastic strain; ∈ el represents the inelastic strain; ∈0 represents the initial strain; ∈ ext represents the strain caused by external load; ∈ th represents thermal strain; ∈ hs Represents historical strain; ∈ pl represents plastic strain; ∈ cr represents creep strain; ∈ vp represents plastic strain; ∈ ve represents the dissolution strain; c represents the elastic tensor; s0 represents the initial stress; s ext represents the stress caused by external load; s q represents the inelastic mechanism stress; u represents the displacement field; E represents the elastic modulus; v represents the Poisson's ratio; Using soil plasticity in COMSOL components, the yield criterion uses the Drucker-Prager criterion to match the Mohr-Coulomb criterion to model the coal rock mass, and its equation is as follows: in, represents the plastic strain rate; λ represents the plastic multiplier; Q represents the plastic potential function; S represents the short-term shear stress; F(σ,σ ys ) represents the function of elastic-plastic conditions; σ represents stress; σ ys represents the yield stress; λF=0 represents the complementary relaxation condition; F cone represents the plasticity criterion; J2 represents the De Voia stress invariant, the intensity of the stress state; α represents the material strength parameter; I1 represents the average intensity of the stress state; k represents a constant; φ represents the internal friction angle of the material.

7. The method for inversion of full-time and space stress field of loaded coal and rock based on deep learning according to claim 1 is characterized in that: In step S50, feature extraction is performed on the local stress value data at the current moment through the reverse diffusion process, and the feature similarity between the extracted local stress value and the full-space stress field is calculated, including the following steps: First, the diffusion neural network processes the full-space stress field data at the current moment into noise data that conforms to the Gaussian distribution through the forward Markov diffusion process. The formula of the Gaussian distribution is as follows: Among them, N(x) represents Gaussian distribution; x represents random variable; exp represents e x ; μ represents mathematical expectation; σ represents standard deviation; Among them, the formula of the forward Markov diffusion process is as follows: Among them, q represents the distribution of the full-space stress field data at different times; β t represents the diffusion rate at time t; z t represents the noisy variant of z0 at time t; z t-1 represents the noisy variant of z0 at time t-1; N represents Gaussian distribution; T represents the full time step of the Markov diffusion process; Then, the feature similarity between the local stress value at the current moment and the full-space stress field data is calculated through the reverse process with U-Net as the backbone. The stress field inversion model obtained after training can obtain the full-time and space stress field through the inversion of the local stress value at all times. The formula is as follows: z=ε(x) Among them, L SDM represents the mechanism of stress field inversion; x represents the full-space stress field of the current moment of the input diffusion neural network; ε represents the encoder, which can encode the input x into z; y represents the local stress value at the current moment; ∈ represents the added Gaussian noise; N(0,1) represents the Gaussian distribution; t represents the current moment; ∈ θ represents the U-Net neural network based on time conditions; τ θ represents an encoder that processes the local stress value at the current moment; D represents a decoder that can decode the obtained full-space stress field corresponding to the local stress value at the current moment from ε(x); Represents the final result of the full-space stress field at the current moment.

8. The method for inversion of full-time and space stress field of loaded coal and rock based on deep learning according to claim 7 is characterized in that: The diffusion neural network uses the attention mechanism to enhance the U-Net neural network backbone. The formula is as follows: in, Represents ∈ θ The intermediate process feature representation; τ θ (y) represents the encoding result of the encoder for the local stress value at the current moment; and They represent the learnable projection matrices respectively.

9. The method for inversion of full-time and space stress field of loaded coal and rock based on deep learning according to claim 1 is characterized in that: In the step S60, the inversion result is verified by the stress value measured by the stress sensor (200) in the load fracture experiment of the coal rock mass (1000), and the model parameters are adjusted according to the verification result, including the following steps: The error between the stress value measured by the stress sensor (200) in the load fracture experiment of the coal rock mass (1000) and the stress field result obtained by the stress field inversion model is calculated by MAE; if the calculated error is large, the model parameters are adjusted; wherein the error calculation formula between the stress value measured by the stress sensor (200) in the load fracture experiment of the coal rock mass (1000) calculated by MAE and the stress field result obtained by the stress field inversion model is as follows: Among them, MAE stands for Mean Absolute Error; y i and They respectively represent the stress value measured by the stress sensor (200) in the load fracture test of the coal rock mass (1000) and the inversion result of the stress field inversion model.

10. A device for inversion of full-time and space stress field of loaded coal and rock based on deep learning, characterized in that: include: The information acquisition unit is configured to arrange a detection-monitoring sensor (100) and a stress sensor (200) in a coal-rock mass (1000), and place the coal-rock mass (1000) in a testing machine, so that the testing machine performs a load-breaking test on the coal-rock mass (1000), and records the collected detection-monitoring data, stress data, and boundary condition information during the test; wherein the detection-monitoring sensor (100) is used to obtain full-time velocity field, fracture field, internal strain, and microcurrent data, and the stress sensor (200) is used to obtain full-time local stress value data at the detection-monitoring position; The transcription model unit is configured to divide the data measured by the detection-monitoring sensor (100) and the stress sensor (200) at all times into a training set, a verification set and a test set according to the corresponding time; use the encoder in the sequence transcription neural network to extract features from the detection-monitoring data in the training set and the verification set, and use the decoder in the sequence transcription neural network to extract features from the local stress data output at the previous time, and train to obtain an information transcription model; A parameter adjustment unit is configured to input the full-time detection-monitoring data in the test set into an information transcription model to obtain a full-time local stress value, and to verify the transcription result of the information transcription model using the stress value at the position of the detection-monitoring sensor (100) measured by the stress sensor (200), and to adjust the parameters of the information transcription model according to the verification result, so that the obtained information transcription model transcribes the full-time detection-monitoring data into the full-time local stress value at the position of the detection-monitoring sensor (100); The simulation model unit is configured to establish a numerical simulation model according to the loading boundary conditions of the coal rock load fracture test, in combination with the mechanical parameters and constitutive equation of the coal rock mass (1000), obtain the full-space stress field data during the coal rock load fracture test by the numerical simulation method, and optimize the numerical simulation method by the local stress value obtained by the information transcription model; The inversion model unit is configured to divide the local stress values ​​at all times obtained by the information transcription model and the stress field data of all spaces obtained by the numerical simulation into a training set, a validation set and a test set; the diffusion neural network extracts features of the stress field data at the current time in the training set and the validation set through a diffusion process, extracts features of the local stress value data at the current time through an inverse diffusion process, and calculates the feature similarity between the extracted local stress value and the stress field of all spaces, thereby training to obtain an inversion model of the stress field of all spaces; The inversion calculation unit is configured to input the local stress value at all times obtained by transcription of the information transcription model into the full-time and space stress field inversion model to invert and obtain the full-time and space stress field, and to verify the inversion result through the stress value measured by the stress sensor (200) in the coal rock mass (1000) load fracture test, and to adjust the model parameters according to the verification result. The obtained full-time and space stress field inversion model obtains the full-time and space stress field in the coal rock load fracture test process through the local stress value at all times of the coal rock mass (1000).

Images