Coal rock mass stress field inversion method and system driven by mechanical mechanism

Abstract

The invention discloses a coal-rock mass stress field inversion method and system driven by a mechanical mechanism. The inversion method comprises the following steps: obtaining detection monitoring data; constructing a neural network model of a multi-layer perceptron through a loss function, taking multi-point stress monitoring number stress as an initial value condition, combining a boundary condition and a mechanical equation, and performing iterative solution to obtain global stress field distribution data; constructing an acoustic emission signal as input, global stress field distribution as output, a long-short-term memory network as a main structure, and combining a full connection layer to perform feature extraction and mapping; constructing a loss function with a self-training weight according to the properties of the coal and rock mass, the mechanical balance and the constitutive model; and optimizing the deep neural network according to a stochastic gradient descent method and an adaptive learning rate, and training the deep neural network by using back propagation according to a constructed loss function with a self-training weight to obtain an inversion model.

Description

Technical Field

[0001] The present invention relates to the technical field of coal rock monitoring, and in particular to a method and system for inverting a coal rock mass stress field driven by a mechanical mechanism. Background Art

[0002] As coal mining continues to increase in depth, the stress environment of deep coal and rock masses becomes increasingly complex. Influenced by multiple factors, including tectonic stress and mining disturbances, these factors can easily lead to dynamic hazards such as rock bursts and coal and gas outbursts. Stress monitoring and analysis can provide data support for rock burst mechanism research, disaster warning, and accident emergency rescue, helping coal mines transition from "passive emergency response" to "active prevention," significantly improving coal mine safety and protecting miners' lives. Therefore, achieving real-time monitoring of stress fields is of great significance to the safe production of coal mining and deep geotechnical engineering.

[0003] Based on the research methods of rock mechanics, many scholars have proposed different theories and methods for interpreting and monitoring the evolution of stress fields in coal and rock masses, thereby effectively preventing and controlling dynamic disasters. However, due to the unique scale of rock masses, both macroscopic and microscopic research methods have certain limitations. Patent CN115754013A proposes a method for detecting stress fields using acoustic emission technology, but this is limited to surface stress in coal and rock masses and does not provide a feasible method for internal stress. Overall, traditional methods have not yet provided a good solution for the problem of global stress field inversion in coal and rock masses.

[0004] The development of deep learning has provided a new approach to solving “black box” problems. Faced with complex problems that are difficult to solve with traditional means such as nonlinearity and multivariables, the powerful automatic fitting capabilities of deep neural networks can be used to uncover the hidden relationships between different parameters. However, deep learning training requires a large amount of high-quality data. Conducting a large number of related experiments in the laboratory to obtain real data is relatively time-consuming and costly, making it difficult to accept. Therefore, how to make full use of prior information to achieve efficient training of deep neural networks when data is limited is a very valuable issue. Physical information neural networks provide an effective solution to this problem. Physical information neural networks make full use of the physical laws followed by mechanical models, and use the form of partial differential equations to pre-“prescribe” the distribution of the predicted stress field, so that high-precision prediction results can be obtained when training data is limited. Therefore, the coal rock stress field inversion method based on mechanical mechanisms and deep neural networks proposed in the present invention is of great significance for efficiently realizing coal rock stress field inversion and achieving effective monitoring of dynamic disasters. Summary of the Invention

[0005] In response to the problems and needs raised above, this solution proposes a coal rock stress field inversion method and system driven by mechanical mechanisms. Due to the adoption of the following technical features, it can achieve the above technical objectives and bring about many other technical effects.

[0006] An object of the present invention is to propose a method for inverting the stress field of coal and rock masses driven by mechanical mechanisms, comprising the following steps:

[0007] S10: Arranging acoustic emission signal sensors and stress sensors in the coal rock mass, and performing a true triaxial compression test on the coal rock mass to obtain detection and monitoring data; wherein the detection and monitoring data includes: acoustic emission signal data and multi-point stress monitoring data;

[0008] S20: A multi-layer perceptron neural network model is constructed using a pre-designed loss function. Using multi-point stress monitoring data as initial conditions, combined with boundary conditions and mechanical equations that conform to coal and rock mass characteristics, the loss function is minimized through automatic differentiation and stochastic gradient descent methods. The neural network model then obtains global stress field distribution data. The coal and rock mass acoustic emission signal data and global stress field distribution data are divided into training and test sets.

[0009] S30: constructing a deep neural network model with acoustic emission signal data as input and global stress field distribution data as output, wherein the deep neural network model uses a long short-term memory network in combination with a fully connected layer for feature extraction and mapping;

[0010] S40: Based on the properties of the coal and rock mass, the mechanical equilibrium equation, and the constitutive model, a loss function with self-training weights is constructed;

[0011] S50: Inputting the training set into the deep neural network model for training, wherein the weights and biases of the deep neural network are adjusted according to the constructed self-training weight loss function and using a stochastic gradient descent method and an adaptive learning rate optimization method;

[0012] S60: Input the test set into the trained deep neural network to test the effect and obtain the stress field inversion model.

[0013] In addition, the coal rock mass stress field inversion method driven by the mechanical mechanism of the present invention may also have the following technical features:

[0014] In one example of the present invention, in step S20, the loss function is expressed as:

[0015]

[0016] in,

[0017]

[0018] Where N r is the number of sampling points of the physical equation residual;

[0019]

[0020] Where u0 is the value given by the initial condition;

[0021]

[0022] Where u BC is the value given by the boundary conditions.

[0023] In one example of the present invention, minimizing the loss function by automatic differentiation and stochastic gradient descent method specifically includes the following steps:

[0024] S21: The target solution stress field distribution u is calculated by the neural network model Approximately, the expression is as follows:

[0025]

[0026] Where, Represents the set of model parameters including PINN weights;

[0027] S22: Finding the optimal parameters by solving optimization problems To find an approximate solution Among them, finding the optimal parameters The expression is as follows:

[0028]

[0029] S23: During the training process, the backpropagation algorithm is used to calculate the gradient of the loss function, which is expressed as:

[0030]

[0031] Where, is the total loss function; u i is the i-th component in the stress expression;

[0032] S24: Update parameters using gradient descent To minimize Its expression is as follows:

[0033]

[0034] Where, is the neural network parameter of the tth iteration; is the specific parameter in the neural network; μ is the learning rate when the parameter is updated.

[0035] In one example of the present invention, in step S30, the deep neural network includes a long short-term memory network and a fully connected layer;

[0036] Among them, the calculation formula of long short-term memory network is as follows:

[0037] i t =σ(W ii x t +b ii +W hi h t-1 +b hi )

[0038] f t =σ(W if x t +b if +W hf h t-1 +b hf )

[0039] g t =tanh(W ig x t +b ig +W hg h t-1 +b hg )

[0040] o t =σ(W io x t +b io +W ho h t-1 +b ho )

[0041] c t =f t ⊙c t-1 +i t ⊙g t

[0042] h t =o t ⊙tanh(c t )

[0043] The calculation formula of the fully connected layer is as follows:

[0044]

[0045] Where h t is the hidden state at time t, c t is the cell state at time t, x t is the input at time t, i t 、ft 、g t and o t They are input gate, forget gate, cell gate and output gate respectively, σ is the Sigmoid function, and ⊙ is the Hadamard product.

[0046] In one example of the present invention, in step S40, the loss function with self-training weights is expressed as:

[0047]

[0048] Where λ1, λ2, and λ3 are weights that are automatically trained during model training. λ1+λ2+λ3=1, and λ1>0, λ2>0, and λ3>0.

[0049] in,

[0050]

[0051] Where, Give the stress prediction value of point i for the model, is the accurate stress value of point i obtained by numerical simulation;

[0052]

[0053] Where, each term is the partial derivative of the predicted value with respect to time t and the x, y, and z directions, and N(u) is the stress equation obtained based on the mechanical constitutive model and the equilibrium equation; is the residual term with a pre-given structure;

[0054]

[0055] Where, and are the predicted value and actual value given based on the actual boundary conditions, respectively.

[0056] In an example of the present invention, step S50 specifically includes the following:

[0057] S51: Based on the adaptive learning rate optimization method, historical gradient information is introduced when updating parameters to make training more stable, and different learning rates are applied to different parameters;

[0058] S52: Use backpropagation according to the designed loss function, that is, calculate the gradient of the loss function relative to the network weights, use the chain rule to calculate the gradient, backpropagate the error layer by layer, and update the parameters of each layer to reduce the loss.

[0059] In one example of the present invention, in step S52, the calculation expression for gradient calculation using the chain rule is:

[0060]

[0061]

[0062] Where L is the loss function; W is the parameter matrix in the neural network; is the predicted value; z L is the output result of the Lth fully connected layer; W L is the neural network parameter matrix of the Lth layer; b L is the bias parameter matrix of the neural network of the Lth layer; are all updated parameters of the Lth layer; θ L are all the pre-updated parameters of the L-th layer; η is the learning rate in parameter update.

[0063] Another object of the present invention is to propose a mechanical mechanism driven coal rock stress field inversion system, comprising the following steps:

[0064] a data acquisition module configured to arrange acoustic emission signal sensors and stress sensors in the coal and rock mass, perform a true triaxial compression test on the coal and rock mass, and obtain detection and monitoring data; wherein the detection and monitoring data includes: acoustic emission signal data and multi-point stress monitoring data;

[0065] A data partitioning module is configured to construct a neural network model of a multi-layer perceptron using a pre-designed loss function, using the stress at the monitoring point as an initial condition, combining boundary conditions and mechanical equations that conform to the characteristics of the coal and rock mass, and minimizing the loss function through automatic differentiation and stochastic gradient descent methods, so that the neural network model obtains global stress field distribution data, and divides the acoustic emission signal data of the coal and rock mass and the global stress field distribution data into a training set and a test set;

[0066] A deep neural network module is configured to construct a deep neural network model that uses acoustic emission signal data as input and global stress field distribution data as output, wherein the deep neural network model uses a long short-term memory network in combination with a fully connected layer for feature extraction and mapping;

[0067] A loss function construction module is configured to construct a loss function with self-training weights based on the properties of the coal and rock mass itself, the mechanical equilibrium equation, and the constitutive model;

[0068] A training model module is configured to input the training set into the deep neural network model for training, wherein the weights and biases of the deep neural network are adjusted according to the constructed self-training weight loss function and using a stochastic gradient descent method and an adaptive learning rate optimization method;

[0069] The inversion model module is configured to input the test set into the trained deep neural network to test the effect and obtain the stress field inversion model.

[0070] In one example of the present invention, the loss function is expressed as:

[0071]

[0072] in,

[0073]

[0074] Where N r is the number of sampling points of the physical equation residual;

[0075]

[0076] Where u0 is the value given by the initial condition;

[0077]

[0078] Where u BC is the value given by the boundary conditions.

[0079] In one example of the present invention, the data partitioning module includes:

[0080] The stress field distribution unit is configured to convert the target solution stress field distribution u calculated by the neural network model into Approximately, the expression is as follows:

[0081]

[0082] Where, Represents the set of model parameters including PINN weights;

[0083] Optimal parameter unit, configured to find the optimal parameters by solving the optimization problem To find an approximate solution Among them, finding the optimal parameters The expression is as follows:

[0084]

[0085] The loss function gradient unit is configured to calculate the gradient of the loss function using the backpropagation algorithm during training. Its expression is:

[0086]

[0087] Where, is the total loss function; u i is the i-th component in the stress expression;

[0088] Update parameter unit, configured to update parameters using gradient descent method To minimize Its expression is as follows:

[0089]

[0090] Where, is the neural network parameter of the tth iteration; is the specific parameter in the neural network; μ is the learning rate when the parameter is updated.

[0091] The present invention has the following beneficial effects compared to the prior art:

[0092] 1. The present invention uses prior physical information to construct a neural network, thereby improving the interpretability of the neural network and making the present invention more stable and reliable. Physical information (such as elastic mechanics theory, stress-strain relationship, wave equation, etc.) provides clear mathematical constraints in the network structure, so that the model prediction results conform to physical laws, reducing the unreasonable predictions that may occur in traditional data-driven methods. Due to the introduction of physical constraints, the model can still maintain good generalization capabilities when facing complex coal and rock environments (such as inhomogeneous media, anisotropic materials), and will not become instable or overfitting due to data bias or noise. Physically guided neural networks are significantly superior to black box neural networks in terms of interpretability, which improves the credibility and engineering adaptability of the model in mine safety monitoring.

[0093] 2. The neural network constructed by this invention, by incorporating physical knowledge, effectively reduces data size requirements and improves model training efficiency, saving significant experimental time and financial costs. Traditional deep learning methods typically require large amounts of labeled data for training. However, this invention utilizes a physical loss term as a supplementary constraint, maintaining high prediction accuracy even with small sample sizes and significantly reducing reliance on experimental data.

[0094] 3. The present invention establishes an effective way to predict the global stress field through acoustic emission signals. Traditional acoustic emission technology is mainly used for micro-fracture positioning or local stress assessment, while the present invention innovatively uses physical information neural networks to establish a mapping relationship between acoustic emission signals and the global stress field, achieving a breakthrough in inferring the overall stress state of coal and rock masses from local acoustic emission information. The present invention can efficiently reconstruct the stress field based on limited observation data, overcoming the problem that traditional methods are difficult to accurately solve the global stress field under complex geological conditions. This method can provide more comprehensive stress field information for coal mine safety monitoring, help accurately identify potential disaster areas, and improve the scientificity and reliability of mine disaster warnings.

[0095] Hereinafter, the best embodiment of the present invention will be described in more detail 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

[0096] 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. The drawings are only used to illustrate some embodiments of the present invention, but not to limit all embodiments of the present invention thereto.

[0097] Figure 1 This is a flow chart of a method for inverting a stress field of a coal rock mass driven by a mechanical mechanism according to an embodiment of the present invention;

[0098] Figure 2 Logic diagram of a coal rock mass stress field inversion method driven by a mechanical mechanism according to an embodiment of the present invention

[0099] Figure 3 A schematic diagram of the structure of a deep neural network according to an embodiment of the present invention;

[0100] Figure 4 2 is a schematic diagram of a long short-term memory network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0101] 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 below 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.

[0102] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning understood by persons of ordinary skill in the field to which the invention belongs. The words "first", "second" and similar terms 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 "a" or "an" do not necessarily indicate a quantity limitation. Words such as "include" or "comprising" mean that the elements or objects preceding the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connected" 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 object being described changes, the relative positional relationship may also change accordingly.

[0103] According to the first aspect of the present invention, a mechanical mechanism driven coal rock stress field inversion method is as follows: Figure 1 、 Figure 2 As shown, the following steps are included:

[0104] S10: Arrange acoustic emission signal sensors and stress sensors in the coal rock mass, conduct true triaxial compression tests on the coal rock mass and obtain detection and monitoring data; wherein the detection and monitoring data include: acoustic emission signal data and multi-point stress monitoring data; specifically, acoustic emission signal sensors and stress sensors are used, the acoustic emission signal sensors are arranged on the surface of the coal rock mass, and the stress sensors are arranged inside the coal rock mass, and both sensors are evenly distributed on the surface or inside the coal rock mass in a grid manner.

[0105] S20: A multi-layer perceptron neural network model is constructed using a pre-designed loss function. Using multi-point stress monitoring data as initial conditions, combined with boundary conditions and mechanical equations that conform to coal and rock mass characteristics, the loss function is minimized through automatic differentiation and stochastic gradient descent methods. The neural network model then obtains global stress field distribution data. The coal and rock mass acoustic emission signal data and global stress field distribution data are divided into training and test sets.

[0106] The neural network model consists of an input layer, a hidden layer, and an output layer. Automatic differentiation is used to optimize parameters, and the neural network is used as a partial differential equation solver. This physical information neural network uses the stress at the monitoring point as the initial condition. Combined with pre-defined boundary conditions and mechanical equations that conform to the characteristics of the coal and rock samples, it uses automatic differentiation as an iterative method to gradually solve the global stress value.

[0107] S30: constructing a deep neural network model with acoustic emission signal data as input and global stress field distribution data as output, wherein the deep neural network model uses a long short-term memory network and combines a fully connected layer to perform feature extraction and mapping, that is, uses a long short-term memory network as the main structure and combines a fully connected layer to perform feature extraction and mapping;

[0108] First, the collected acoustic emission signals are preprocessed, including noise reduction, feature extraction, and normalization, to improve data quality and model stability. Subsequently, a neural network architecture is designed: the input layer receives the processed signal features, the LSTM layer extracts temporal dependencies, and further maps them to the stress field distribution space through a fully connected layer. Finally, the output layer generates the stress field prediction results.

[0109] Among them, the collected acoustic emission signals are normalized, where μ and σ are the mean and standard deviation respectively:

[0110]

[0111] S40: Based on the coal and rock mass properties, mechanical equilibrium equations, and constitutive models, a loss function with self-training weights is constructed. The loss function consists of three parts: the mean square error between the stress predicted by the neural network and the known true stress field; the residual of the coal and rock mass stress equilibrium equation; and the mean square error between the predicted stress boundary values ​​and the boundary conditions. The weights of these three components in the loss function are automatically learned during the neural network training process.

[0112] S50: Inputting the training set into the deep neural network model for training, wherein the weights and biases of the deep neural network are adjusted according to the constructed self-training weight loss function and using a stochastic gradient descent method and an adaptive learning rate optimization method;

[0113] S60: Input the test set into the trained deep neural network to test the effect and obtain the stress field inversion model. According to the stochastic gradient descent method, by calculating the gradient of the loss function, adjust the weights and biases of the neural network to minimize the loss. And the adaptive learning rate optimization method introduces historical gradient information when updating parameters to make the training more stable, and applies different learning rates to different parameters. According to the designed loss function, backpropagation is used, that is, the gradient of the loss function relative to the network weights and biases is calculated, and the gradient calculation is performed using the chain rule. The error is backpropagated layer by layer, and the parameters of each layer are updated to reduce the loss. The deep neural network is trained. When the loss function converges or the training reaches the maximum number of iterations, the training ends and the final model is obtained.

[0114] This inversion method uses prior physical information to construct a neural network, thereby improving the interpretability of the neural network and making the present invention more stable and reliable. Physical information (such as elastic mechanics theory, stress-strain relationship, wave equation, etc.) provides clear mathematical constraints in the network structure, so that the model prediction results conform to physical laws, reducing the unreasonable predictions that may occur in traditional data-driven methods. Due to the introduction of physical constraints, the model can still maintain good generalization capabilities when facing complex coal and rock environments (such as inhomogeneous media, anisotropic materials), and will not become instable or overfitting due to data bias or noise. Physically guided neural networks are significantly superior to black box neural networks in terms of interpretability, which improves the credibility and engineering adaptability of the model in mine safety monitoring.

[0115] The neural network constructed using this inversion method, by incorporating physical knowledge, effectively reduces data size requirements and improves model training efficiency, saving significant experimental time and financial costs. Traditional deep learning methods typically require large amounts of labeled data for training, but this method utilizes a physical loss term as an auxiliary constraint, maintaining high prediction accuracy even with small sample sizes and significantly reducing reliance on experimental data.

[0116] This inversion method establishes an effective way to predict the global stress field through acoustic emission signals. Traditional acoustic emission technology is mainly used for micro-fracture positioning or local stress assessment, while the present invention innovatively uses physical information neural networks to establish a mapping relationship between acoustic emission signals and the global stress field, achieving a breakthrough in inferring the overall stress state of coal and rock masses from local acoustic emission information. The present invention can efficiently reconstruct the stress field based on limited observation data, overcoming the problem that traditional methods have difficulty in accurately solving the global stress field under complex geological conditions. This method can provide more comprehensive stress field information for coal mine safety monitoring, help accurately identify potential disaster areas, and improve the scientific nature and reliability of mine disaster warnings.

[0117] In one example of the present invention, in step S20, the loss function is expressed as:

[0118]

[0119] in,

[0120]

[0121] Where N r is the number of sampling points of the physical equation residual;

[0122]

[0123] Where u0 is the value given by the initial condition;

[0124]

[0125] Where u BC is the value given by the boundary conditions.

[0126] In one example of the present invention, minimizing the loss function by automatic differentiation and stochastic gradient descent method specifically includes the following steps:

[0127] S21: The target solution stress field distribution u is calculated by the neural network model Approximately, the expression is as follows:

[0128]

[0129] Where, Represents the set of model parameters including PINN weights;

[0130] S22: Finding the optimal parameters by solving optimization problems To find an approximate solution The parameters obtained by solving the optimization problem Decision, which means Satisfy the pre-given partial differential equations, constitutive equations, and boundary conditions; among them, find the optimal parameters The expression is as follows:

[0131]

[0132] S23: During the training process, the backpropagation algorithm is used to calculate the gradient of the loss function, which is expressed as:

[0133]

[0134] Where, is the total loss function; u i is the i-th component in the stress expression;

[0135] S24: Update parameters using gradient descent To minimize Its expression is as follows:

[0136]

[0137] Where, is the neural network parameter of the tth iteration; is the specific parameter in the neural network; μ is the learning rate when the parameter is updated.

[0138] In one example of the present invention, in step S30, the acoustic emission signals within a time interval are first treated as a time series. Acoustic emission data from multiple monitoring points are collected and arranged according to time steps to construct an acoustic emission matrix. The rows of this matrix represent the acoustic emission signals from different monitoring points, and the columns represent the signal evolution over time steps, enabling a comprehensive characterization of the dynamic response characteristics of the coal and rock mass within the time interval. To enhance data validity, the acoustic emission signals undergo necessary preprocessing and normalization to ensure a consistent distribution of the input data, improving the training stability and generalization capabilities of the model.

[0139] Stress values ​​are recorded as a three-dimensional tensor to represent the stress distribution at different spatial locations. The dimensions of this tensor are set to correspond to the spatial distribution of the coal and rock mass. Subsequently, the preprocessed acoustic emission matrix and stress field tensor are used as the input and target output of a deep neural network for supervised learning training.

[0140] In one example of the present invention, in step S30, the neural network architecture design includes an input layer that receives the acoustic emission matrix and inputs it into a long short-term memory (LSTM) network. The LSTM layer uses its unique gating mechanism to extract the temporal dependencies of the acoustic emission signals and identify the stress evolution characteristics of the coal and rock mass at different time steps. The multi-layer LSTM structure further extracts high-order temporal features and stores long-term dependency information through hidden states. The output of the LSTM layer then undergoes dimensionality reduction and feature mapping in a fully connected layer, ultimately generating a predicted stress field tensor at the output layer.

[0141] In one example of the present invention, in step S30, the deep neural network model includes a long short-term memory network and a fully connected layer; for example, Figure 3 As shown, the deep neural network model includes multiple long short-term memory networks connected in series and two ends of the multiple long short-term memory networks connected in series;

[0142] Among them, long short-term memory networks such as Figure 4 As shown, the calculation formula is as follows:

[0143] i t =σ(W ii x t +b ii +W hi h t-1 +b hi )

[0144] f t =σ(W if x t +b if +W hf h t-1 +bhf )

[0145] g t =tanh(W ig x t +b ig +W hg h t-1 +b hg )

[0146] o t =σ(W io x t +b io +W ho h t-1 +b ho )

[0147] c t =f t ⊙c t-1 +i t ⊙g t

[0148] h t =o t ⊙tanh(c t )

[0149] Specifically, we use an LSTM neural network with a three-gate structure: a forget gate, an input gate, and an output gate. This gate structure controls the removal and retention of features. We apply the LSTM to the input time series. For each element in the input sequence, each layer computes the above function.

[0150] The calculation formula of the fully connected layer is as follows:

[0151]

[0152] Where h t is the hidden state at time t, c t is the cell state at time t, x t is the input at time t, i t 、f t 、g t and o t They are input gate, forget gate, cell gate and output gate respectively, σ is the Sigmoid function, and ⊙ is the Hadamard product.

[0153]

[0154] c t Continue to propagate in the LSTM structure and play the role of long-term memory; h t-1As the output of the previous LSTM, it is also passed as input to the next moment, and is combined with x t Through the three gates together, acting on c t .

[0155] In one example of the present invention, in step S40, the loss function with self-training weights is expressed as:

[0156]

[0157] Where λ1, λ2, and λ3 are weights that are automatically trained during model training. λ1+λ2+λ3=1, and λ1>0, λ2>0, and λ3>0.

[0158] in,

[0159]

[0160] Where, Give the stress prediction value of point i for the model, is the accurate stress value of point i obtained by numerical simulation;

[0161]

[0162] Where, each term is the partial derivative of the predicted value with respect to time t and the x, y, and z directions, and N(u) is the stress equation obtained based on the mechanical constitutive model and the equilibrium equation; is the residual term with a pre-given structure;

[0163]

[0164] Where, and are the predicted value and actual value given based on the actual boundary conditions, respectively.

[0165] In an example of the present invention, step S50 specifically includes the following:

[0166] S51: Based on the adaptive learning rate optimization method, historical gradient information is introduced when updating parameters to make training more stable, and different learning rates are applied to different parameters;

[0167] S52: Use backpropagation according to the designed loss function, that is, calculate the gradient of the loss function relative to the network weights, use the chain rule to calculate the gradient, backpropagate the error layer by layer, and update the parameters of each layer to reduce the loss.

[0168] In one example of the present invention, in step S52, the calculation expression for gradient calculation using the chain rule is:

[0169]

[0170] Where L is the loss function; W is the parameter matrix in the neural network; is the predicted value; z L is the output result of the Lth fully connected layer; W L is the neural network parameter matrix of the Lth layer; b L is the bias parameter matrix of the neural network of the Lth layer; are all updated parameters of the Lth layer; θ L are all the pre-updated parameters of the L-th layer; η is the learning rate in parameter update.

[0171] According to the second aspect of the present invention, a mechanical mechanism-driven coal rock mass stress field inversion system comprises the following steps:

[0172] a data acquisition module configured to arrange acoustic emission signal sensors and stress sensors in the coal and rock mass, perform a true triaxial compression test on the coal and rock mass, and obtain detection and monitoring data; wherein the detection and monitoring data includes: acoustic emission signal data and multi-point stress monitoring data;

[0173] A data partitioning module is configured to construct a neural network model of a multi-layer perceptron using a pre-designed loss function, using multi-point stress monitoring data as initial conditions, combining boundary conditions and mechanical equations that conform to coal and rock mass characteristics, and minimizing the loss function through automatic differentiation and stochastic gradient descent methods, so that the neural network model obtains global stress field distribution data, and divides the coal and rock mass acoustic emission signal data and the global stress field distribution data into a training set and a test set;

[0174] A deep neural network module is configured to construct a deep neural network model that uses acoustic emission signal data as input and global stress field distribution data as output, wherein the deep neural network model uses a long short-term memory network in combination with a fully connected layer for feature extraction and mapping;

[0175] A loss function construction module is configured to construct a loss function with self-training weights based on the properties of the coal and rock mass itself, the mechanical equilibrium equation, and the constitutive model;

[0176] A training model module is configured to input the training set into the deep neural network model for training, wherein the weights and biases of the deep neural network are adjusted according to the constructed self-training weight loss function and using a stochastic gradient descent method and an adaptive learning rate optimization method;

[0177] The inversion model module is configured to input the test set into the trained deep neural network to test the effect and obtain the stress field inversion model.

[0178] The inversion system uses prior physical information to construct a neural network, thereby improving the interpretability of the neural network and making the present invention more stable and reliable. Physical information (such as elastic mechanics theory, stress-strain relationship, wave equation, etc.) provides clear mathematical constraints in the network structure, so that the model prediction results conform to physical laws, reducing the unreasonable predictions that may occur in traditional data-driven methods. Due to the introduction of physical constraints, the model can still maintain good generalization capabilities when facing complex coal and rock environments (such as inhomogeneous media, anisotropic materials), and will not become instable or overfitting due to data bias or noise. Physically guided neural networks are significantly superior to black box neural networks in terms of interpretability, which improves the credibility and engineering adaptability of the model in mine safety monitoring.

[0179] The neural network constructed by this inversion system, by incorporating physical knowledge, effectively reduces data size requirements and improves model training efficiency, saving significant experimental time and financial costs. Traditional deep learning methods typically require large amounts of labeled data for training, but this invention utilizes physical loss terms as auxiliary constraints, maintaining high prediction accuracy even with small sample sizes and significantly reducing reliance on experimental data.

[0180] This inversion system establishes an effective way to predict the global stress field through acoustic emission signals. Traditional acoustic emission technology is mainly used for micro-fracture positioning or local stress assessment, while the present invention innovatively uses physical information neural networks to establish a mapping relationship between acoustic emission signals and the global stress field, achieving a breakthrough in inferring the overall stress state of coal and rock masses from local acoustic emission information. The present invention can efficiently reconstruct the stress field based on limited observation data, overcoming the problem that traditional methods have difficulty in accurately solving the global stress field under complex geological conditions. This method can provide more comprehensive stress field information for coal mine safety monitoring, help accurately identify potential disaster areas, and improve the scientific nature and reliability of mine disaster warnings.

[0181] In one example of the present invention, the loss function is expressed as:

[0182]

[0183] in,

[0184]

[0185] Where N r is the number of sampling points of the physical equation residual;

[0186]

[0187] Where u0 is the value given by the initial condition;

[0188]

[0189] Where u BC is the value given by the boundary conditions.

[0190] In one example of the present invention, the data partitioning module includes:

[0191] The stress field distribution unit is configured to convert the target solution stress field distribution u calculated by the neural network model into Approximately, the expression is as follows:

[0192]

[0193] Where, Represents the set of model parameters including PINN weights;

[0194] Optimal parameter unit, configured to find the optimal parameters by solving the optimization problem To find an approximate solution The parameters obtained by solving the optimization problem Decision, which means Satisfy the pre-given partial differential equations, constitutive equations, and boundary conditions; among them, find the optimal parameters The expression is as follows:

[0195]

[0196] The loss function gradient unit is configured to calculate the gradient of the loss function using the backpropagation algorithm during training. Its expression is:

[0197]

[0198] Where, is the total loss function; u i is the i-th component in the stress expression;

[0199] Update parameter unit, configured to update parameters using gradient descent method To minimize Its expression is as follows:

[0200]

[0201] Where, is the neural network parameter of the tth iteration; is the specific parameter in the neural network; μ is the learning rate when the parameter is updated.

[0202] It should be noted that the mechanical mechanism driven coal rock stress field inversion system of the present invention can also perform any processing in the mechanical mechanism driven coal rock stress field inversion method described previously, and the specific details are not repeated here.

[0203] The exemplary implementation scheme of the coal rock stress field inversion method and system driven by the mechanical mechanism proposed in the present invention is described in detail above with reference to the preferred embodiments. However, it can be understood by those skilled in the art 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 scope of protection of the present invention, which is determined by the appended claims.

Claims

1. A method for inversion of coal rock stress field driven by mechanical mechanism, characterized in that: The steps include: S10: Arranging acoustic emission signal sensors and stress sensors in the coal rock mass, and performing a true triaxial compression test on the coal rock mass to obtain detection and monitoring data; wherein the detection and monitoring data includes: acoustic emission signal data and multi-point stress monitoring data; S20: A multi-layer perceptron neural network model is constructed using a pre-designed loss function. Using multi-point stress monitoring data as initial conditions, combined with boundary conditions and mechanical equations that conform to coal and rock mass characteristics, the loss function is minimized through automatic differentiation and stochastic gradient descent methods. The neural network model then obtains global stress field distribution data. The coal and rock mass acoustic emission signal data and global stress field distribution data are divided into training and test sets. S30: constructing a deep neural network model with acoustic emission signal data as input and global stress field distribution data as output, wherein the deep neural network model uses a long short-term memory network in combination with a fully connected layer for feature extraction and mapping; S40: Based on the properties of the coal and rock mass, the mechanical equilibrium equation, and the constitutive model, a loss function with self-training weights is constructed; S50: Inputting the training set into the deep neural network model for training, wherein the weights and biases of the deep neural network are adjusted according to the constructed self-training weight loss function and using a stochastic gradient descent method and an adaptive learning rate optimization method; S60: Input the test set into the trained deep neural network to test the effect and obtain the stress field inversion model.

2. The method for inversion of coal rock mass stress field driven by mechanical mechanism according to claim 1, characterized in that: In step S20, the loss function is expressed as: in, Where N r is the number of sampling points of the physical equation residual; Where u0 is the value given by the initial condition; Where u BC is the value given by the boundary conditions.

3. The method for inversion of coal rock mass stress field driven by mechanical mechanism according to claim 1, characterized in that: The method of minimizing the loss function by automatic differentiation and stochastic gradient descent specifically includes the following steps: S21: The target solution stress field distribution u is calculated by the neural network model Approximately, the expression is as follows: Where, Represents the set of model parameters including PINN weights; S22: Finding the optimal parameters by solving optimization problems To find an approximate solution Among them, finding the optimal parameters The expression is as follows: S23: During the training process, the backpropagation algorithm is used to calculate the gradient of the loss function, which is expressed as: Where, is the total loss function; u i is the i-th component in the stress expression; S24: Update parameters using gradient descent To minimize Its expression is as follows: Where, is the neural network parameter of the tth iteration; is the specific parameter in the neural network; μ is the learning rate when the parameter is updated.

4. The method for inversion of coal rock mass stress field driven by mechanical mechanism according to claim 1, characterized in that: In step S30, the deep neural network includes a long short-term memory network and a fully connected layer; Among them, the calculation formula of long short-term memory network is as follows: i t =σ(W ii x t +b ii +W hi h t-1 +b hi ) f t =σ(W if x t +b if +W hf h t-1 +b hf ) g t =tanh(W ig x t +b ig +W hg h t-1 +b hg ) o t =σ(W io x t +b io +W ho h t-1 +b ho ) c t =f t ⊙c t-1 +i t ⊙g t h t =o t ⊙tanh(c t ) The calculation formula of the fully connected layer is as follows: Where h t is the hidden state at time t, c t is the cell state at time t, x t is the input at time t, i t 、f t 、g t and o t They are input gate, forget gate, cell gate and output gate respectively, σ is the Sigmoid function, and ⊙ is the Hadamard product.

5. The method for inversion of coal rock mass stress field driven by mechanical mechanism according to claim 1, characterized in that: In step S40, the loss function with self-training weights is expressed as: Where λ1, λ2, and λ3 are weights that are automatically trained during model training. λ1+λ2+λ3=1, and λ1>0, λ2>0, and λ3>0. in, Where, Give the stress prediction value of point i for the model, is the accurate stress value of point i obtained by numerical simulation; Where, each term is the partial derivative of the predicted value with respect to time t and the x, y, and z directions, and N(u) is the stress equation obtained based on the mechanical constitutive model and the equilibrium equation; is the residual term with a pre-given structure; Where, and are the predicted value and actual value given based on the actual boundary conditions, respectively.

6. The method for inversion of coal rock mass stress field driven by mechanical mechanism according to claim 1, characterized in that: The step S50 specifically includes the following: S51: Based on the adaptive learning rate optimization method, historical gradient information is introduced when updating parameters to make training more stable, and different learning rates are applied to different parameters; S52: Use backpropagation according to the designed loss function, that is, calculate the gradient of the loss function relative to the network weights, use the chain rule to calculate the gradient, backpropagate the error layer by layer, and update the parameters of each layer to reduce the loss.

7. The method for inversion of coal rock mass stress field driven by mechanical mechanism according to claim 6, characterized in that: In step S52, the calculation expression for gradient calculation using the chain rule is: Where L is the loss function; W is the parameter matrix in the neural network; is the predicted value; z L is the output result of the Lth fully connected layer; W L is the neural network parameter matrix of the Lth layer; b l is the bias parameter matrix of the neural network of the Lth layer; are all updated parameters of the Lth layer; θ L are all pre-updated parameters of the Lth layer; η is the learning rate in parameter update.

8. A mechanical mechanism driven coal rock stress field inversion system, characterized by: The steps include: a data acquisition module configured to arrange acoustic emission signal sensors and stress sensors in the coal and rock mass, perform a true triaxial compression test on the coal and rock mass, and obtain detection and monitoring data; wherein the detection and monitoring data includes: acoustic emission signal data and multi-point stress monitoring data; A data partitioning module is configured to construct a neural network model of a multi-layer perceptron using a pre-designed loss function, using multi-point stress monitoring data as initial conditions, combining boundary conditions and mechanical equations that conform to coal and rock mass characteristics, and minimizing the loss function through automatic differentiation and stochastic gradient descent methods, so that the neural network model obtains global stress field distribution data, and divides the coal and rock mass acoustic emission signal data and the global stress field distribution data into a training set and a test set; A deep neural network module is configured to construct a deep neural network model that uses acoustic emission signal data as input and global stress field distribution data as output, wherein the deep neural network model uses a long short-term memory network in combination with a fully connected layer for feature extraction and mapping; A loss function construction module is configured to construct a loss function with self-training weights based on the properties of the coal and rock mass itself, the mechanical equilibrium equation, and the constitutive model; A training model module is configured to input the training set into the deep neural network model for training, wherein the weights and biases of the deep neural network are adjusted according to the constructed self-training weight loss function and using a stochastic gradient descent method and an adaptive learning rate optimization method; The inversion model module is configured to input the test set into the trained deep neural network to test the effect and obtain the stress field inversion model.

9. The mechanical mechanism driven coal rock stress field inversion system according to claim 8, characterized in that: The expression of the loss function is: in, Where N r is the number of sampling points of the physical equation residual; Where u0 is the value given by the initial condition; Where u BC is the value given by the boundary conditions.

10. The mechanical mechanism driven coal rock stress field inversion system according to claim 8, characterized in that: The data partitioning module includes: The stress field distribution unit is configured to convert the target solution stress field distribution u calculated by the neural network model into Approximately, the expression is as follows: Where, Represents the set of model parameters including PINN weights; Optimal parameter unit, configured to find the optimal parameters by solving the optimization problem To find an approximate solution Among them, finding the optimal parameters The expression is as follows: The loss function gradient unit is configured to calculate the gradient of the loss function using the backpropagation algorithm during training. Its expression is: Where, is the total loss function; u i is the i-th component in the stress expression; Update parameter unit, configured to update parameters using gradient descent method To minimize Its expression is as follows: Where, is the neural network parameter of the tth iteration; is the specific parameter in the neural network; μ is the learning rate when the parameter is updated.

Images