Coal roadway composite roof stability intelligent identification method based on machine learning

Abstract

The invention discloses a coal roadway composite roof stability intelligent identification method based on machine learning, and the method comprises the following steps: S1, obtaining a sample data set which comprises roadway comprehensive anchor rod support quality, a roof rock stratum inclination angle, roadway surrounding rock stress, roof strength, coal seam burial depth and roof stability classification labels; and S2, training the neural network model by using the roadway comprehensive anchor rod support quality, the roof rock stratum dip angle, the roadway surrounding rock stress, the roof strength, the coal seam burial depth and the roof stability classification label to obtain a trained roof stability prediction model. According to the method, a roof stability prediction model based on parameters such as roadway comprehensive anchor rod supporting quality, a roof rock stratum dip angle, roadway surrounding rock stress, roof strength and coal seam burial depth is established for the first time; and real-time prediction of the stability of the roof can be completed based on real-time data returned by a field sensor, the prediction precision is high, and the prediction complexity is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent and safe mining of coal mines, and specifically relates to an intelligent identification method for the stability of composite roof of a coal lane based on machine learning. Background Art

[0002] Composite roof refers to a roof structure composed of multiple composite rock layers with different physical and mechanical properties. This roof structure is common in sedimentary rock formations, such as shale, sandstone, etc. Unstable roof may cause roof collapse or rock sliding, directly endangering the lives of construction workers. This disaster may cause mine accidents and affect the construction progress. Construction workers need to carry out additional support and reinforcement, resulting in extended project time. This not only increases the cost of the project, but may also cause delays in the project schedule, seriously affecting the normal production and operation of the mine and reducing production efficiency. The deformation and collapse of the rock mass will affect the mining and transportation of the ore body, reducing the output and economic benefits of the mine.

[0003] Real-time monitoring of roof stability can provide immediate early warning feedback, allowing coal mine operations to quickly identify and respond to potential safety risks. Through continuous monitoring, roof deformation, cracks and other problems can be discovered in a timely manner, thereby preventing catastrophic accidents, optimizing maintenance and reinforcement measures, and improving overall safety and production efficiency. At the same time, the accumulation of monitoring data helps to identify long-term trends and patterns, supporting more accurate predictions and decisions.

[0004] Machine learning has significant advantages in solving classification problems such as coal mine roof stability because it can identify potential roof stability problems. Machine learning models can process high-dimensional data and capture subtle nonlinear relationships in roof structures, thereby improving the accuracy and reliability of classification. In addition, these models have the ability to process large-scale data and can adapt to the characteristics of different mining areas through transfer learning, thereby optimizing stability prediction and risk assessment.

[0005] Due to the high uncertainty of underground geotechnical engineering, the geological conditions of each mining area are not completely consistent, resulting in different geological conditions suitable for each mining area. Therefore, it is difficult to meet the classification accuracy by using only a single network model for training. In addition, the traditional neural network model is prone to fall into local minima and cannot find the global optimal solution. Through targeted optimization of the neural network model, it is helpful to further improve the classification accuracy of the stability of the composite roof of the coal roadway.

[0006] Therefore, the present invention proposes an intelligent identification method for the stability of composite roof of coal roadway based on machine learning. Summary of the invention

[0007] The purpose of the present invention is to propose an intelligent identification method for the stability of coal roadway composite roof based on machine learning, so as to solve the problem that the neural network model in the prior art is easy to fall into the local minimum and cannot find the global optimal solution, and further improve the classification accuracy of the model.

[0008] In order to solve the above technical problems, the present invention adopts the following technical solution: a method for intelligently identifying the stability of composite roof of coal roadway based on machine learning, comprising the following steps:

[0009] S1. Obtain a sample data set, wherein the sample data set includes the quality of comprehensive bolt support of the roadway, the inclination of the roof rock layer, the stress of the roadway surrounding rock, the strength of the roof, the buried depth of the coal seam, and the classification label of the roof stability;

[0010] S2. Use the comprehensive anchor support quality of the tunnel, the inclination of the roof rock layer, the stress of the tunnel surrounding rock, the strength of the roof, the buried depth of the coal seam and the classification labels of the roof stability to train the neural network model and obtain a trained roof stability prediction model.

[0011] Furthermore, the specific process of obtaining the sample data set in step S1 is as follows:

[0012] In the field of mine safety, the stability prediction of the roof is the key to controlling the safe production of coal mines; the roof stability prediction model is trained using the sample data set to determine the roof instability mode;

[0013] The comprehensive anchor support quality of the tunnel refers to the anchoring quality of multiple anchors in the plane of the comprehensive tunnel, reflecting the overall anchoring quality of the tunnel section; according to the anchor support design drawing of the composite roof tunnel and JGJ / T182-2009 "Technical Specifications for Nondestructive Testing of Anchor Anchoring Quality", the evaluation indicators of a single anchor are the anchor length and anchor density; the qualified standard for anchor length is that the shorter length shall not exceed 5% of the designed anchor length; the length of a single anchor and the anchor density are obtained by the anchor nondestructive testing instrument. The comprehensive anchor support quality of the tunnel is:

[0014]

[0015] Where, Φ is the comprehensive anchor support quality of the tunnel, d i is the distance between the anchor and the center of the tunnel, l i is the actual anchoring length of the anchor rod, l' i Design anchorage length for anchor rods, is the actual anchoring quality of the anchor rod; the roof rock layer inclination, tunnel surrounding rock stress, roof strength, and coal seam burial depth are obtained from relevant coal mine underground geological exploration data; the roof stability classification label is quantified as the tunnel roof height, specifically Class I roof, Class II roof, Class III roof, and Class IV roof;

[0016] It is crucial to normalize the data of the neural network. Normalization can accelerate model convergence, improve numerical stability, eliminate feature bias, and improve the generalization ability of the model. Normalization also ensures that the input values ​​adapt to the effective range of the activation function, thereby improving training efficiency. Scaling the data to the range of [0,1] avoids data with large eigenvalues ​​that may cause numerical instability.

[0017]

[0018] In the formula, H a , H A are the values ​​before and after normalization, H max is the maximum value of each group of data, H min is the minimum value of each group of data.

[0019] Furthermore, the specific process of obtaining the sample data set in step S2 is as follows: the sample data set is divided into a training data set, a validation set, and a test set in a ratio of 14:3:3, the appropriate network architecture is determined, the number of network layers, the number of neurons in each layer, and the activation function are defined, the appropriate loss function and optimization algorithm are selected, the training set is iterated multiple times, forward propagation, loss calculation, and back propagation are performed in each training round, and the performance of the final roof stability prediction model is evaluated on the test set.

[0020] Furthermore, the process of training the roof stability prediction model in step S2 specifically includes the following steps:

[0021] Step 1: Use the sample data set to divide it into a training sample data set and a test sample data set. The training sample data set is used to train the neural network model and optimize the parameters to determine the roof stability prediction model; the test sample data set is used to evaluate the performance of the trained model on unseen data, help determine the accuracy and generalization ability of the model, detect whether the model is overfitting on the training data, and ensure the reliability of the model in practical applications; determine the number of input layer nodes according to the high-dimensional data at the input end, and determine the hidden layer hyperparameters of the network topology structure according to the Kolmogorov theorem. The single hidden layer structure can map the nonlinear input-output relationship with arbitrary precision. The number of hidden layer nodes is calculated by the following formula, and then refined by trial and error method to minimize the prediction error; the Kolmogorov theorem equation is:

[0022]

[0023] Where a is the number of neurons in the hidden layer, m is the number of neurons in the input layer, n is the number of neurons in the output layer, and b is a constant between [1,10];

[0024] Step 2: According to the historical sample database, determine that the number of neurons in the input layer of the network is equal to the number of features of the input data, which can be represented by a vector as follows: The output layer can be represented by a vector M = [m1, m2, m3, m4] T ; W (1) Indicates that the components are The weight matrix between the input layer neurons and the hidden layer neurons, W (2) Indicates that the components are The weight matrix between the hidden layer neurons and the output layer neurons, B (1) , B (2) Represent the bias terms of the hidden layer and the output layer respectively; in deep networks, if the weights are not initialized properly, the signal may have gradient vanishing or gradient explosion during forward propagation; Kaiming initialization helps keep the variance of the output of each layer consistent by setting a suitable variance, thus avoiding these problems. Therefore, Kaiming normal distribution is used to calculate the W of the neural network. (1) , W (2) Initialize, where Where N(0,σ 2 ) means the mean is μ and the variance is σ 2 The normal distribution of , U(-c,c) represents the uniform distribution in the range [-c, c], n input represents the number of neurons in the input layer, m input Represents the number of neurons in the hidden layer;

[0025] Step 3: In order to explore the possible parameter space in the early stage of model training, adjust the differences between individuals in the population to improve the convergence speed of the group, while maintaining the diversity of the group, and solve combinatorial optimization problems, nonlinear programming problems and other complex optimization algorithm problems.

[0026] Furthermore, the basic structure of the optimization algorithm described in step 3 mainly includes mutation operation, crossover operation and selection operation. First, assuming that the dimension of each individual in the initialization population is G, the individuals contained in the population are: X i =(χ i,1 , χ i,2 ,…,χ i,G ), i∈NP; where NP is the number of individuals in the population, and each individual Contains D variables, so the randomly generated initialization population is as follows:

[0027] Where: and are the upper and lower limits of each variable, respectively, and rand(0,1) represents a random number in the range of 0 to 1; the sequence generated by the Tent chaotic map has strong randomness and uncertainty, and can generate widely distributed and diverse initial individuals; this diversity helps to avoid the premature convergence of the optimization algorithm of the present invention in the search space, making it easier to explore the global optimal solution, and the chaotic sequence is generated in a way that each iteration contains certain regularity and structure, which can help the algorithm to search in an effective direction at an early stage and improve the convergence speed. The new initialization population obtained by performing the Tent chaotic map is as follows:

[0028]

[0029] Where: Indicates that a new initialization population is generated through the Tent chaotic map;

[0030] In order to prevent the population from converging to the local optimal solution too early, new gene combinations are introduced to increase the diversity of solutions, help the algorithm explore a wider solution space, discover potential high-quality solutions, and improve overall performance. Three different individuals are randomly selected from the population, and the vector difference between two of them is superimposed on the third individual to obtain a new variant individual. The mutation operation is as follows:

[0031] R i,G+1 = x r1,G+1 +F(X r2,G+1 -X r3,G+1 )

[0032] In the formula: r1, r2, r3 represent random numbers randomly selected from the population [0, NP], F is the variation factor, and its value is a real constant in the range of [0, 2]. i With the variant individual R i,G+1 Perform crossover operation to obtain new individuals The crossover operation formula is as follows:

[0033]

[0034] Where: rand j represents the random number corresponding to the jth variable; CR is the crossover factor; rand(i) represents a random integer in the range [1, NP]; after calculating the crossover operation, the individual S is obtained i,G+1 The fitness of the G-generation individual χ i,G The fitness of the next generation population is compared with that of the next generation population, and the one closer to the true value is selected as the next generation population individual Z i,G+1 , the formula for the selection operation is as follows:

[0035]

[0036] In the formula, the fitness function f is the objective function of the optimization problem. In this paper, the real value Y is selected i and the predicted value Y i 0 The root mean square error is used as the fitness function, and its calculation formula is as follows:

[0037]

[0038] Cyclic crossover, mutation, and selection operations until the fitness function meets the conditions or reaches the maximum number of iterations; the fitness function meets the conditions, that is, the fitness |f'-f|<ξ, indicating that the fitness curve gradually decreases and tends to be stable, and the change of the fitness value between consecutive generations tends to zero;

[0039] Where ξ represents a very small positive number, generally 0.1; f' represents the fitness value obtained by the next cycle of f; if the fitness function meets the conditions, it proves that the performance of the model on the training set or the validation set tends to be stable, and the current parameter configuration has basically found the optimal or suboptimal solution, and further iterations will not significantly improve the fitness value, then the iteration is stopped and the trained network parameters are returned; if the fitness function does not meet the conditions before completing the set number of iterations, it means that the objective function may have high complexity or nonlinearity, resulting in the optimization process requiring more iterations to find a better solution, then the number of iterations is increased to 150% of the original number of iterations, and the iterative operation is continued; if the fitness function meets the conditions before completing the set number of iterations, then the iteration is stopped and the training is returned The obtained network parameters; if the fitness function still does not meet the conditions before completing the set number of secondary iterations, consider adjusting the hyperparameters of the optimization algorithm: if the fitness change fluctuates greatly, reduce the mutation factor F to 85% of the original size to improve stability; if the optimization process lacks diversity, increase the mutation factor F to 125% of the original size to make the search more exploratory; if it is found that the diversity of new individuals is insufficient, try to increase the crossover factor CR to 125% of the original size to introduce more genetic information; if the fitness of the new individual is poor and the convergence speed is slow, consider reducing the crossover factor CR to 85% of the original size to rely more on the existing better individuals; verify the effects of different parameter settings through multiple experiments, record the fitness changes and convergence speed, and find the optimal parameter combination;

[0040] The iterated weights and thresholds are assigned to the neural network to train the neural network. The forward propagation process of the network can be expressed as:

[0041]

[0042] In the formula, a (1) represents the output of the hidden layer of the neural network, a(2) Represents the prediction result of the neural network; Rule() means passing the result of the linear transformation of the input layer to the Rule activation function for processing, and Softmax() means passing the result of the linear transformation of the hidden layer to the Softmax activation function for processing; the predicted value a is obtained (2) Finally, the cross entropy loss function S is used to calculate the error between the predicted value and the true value:

[0043]

[0044] In the formula, n is the number of categories; the gradient is calculated by the back propagation algorithm, and the partial derivative of the loss function for each parameter is calculated layer by layer using the chain rule. The objective function gradient at time t is calculated, the biased first-order moment vector and second-order moment vector are updated, and the bias-corrected first-order moment vector and second-order data vector are calculated. The weights and thresholds are updated. The specific formula is as follows:

[0045]

[0046] In the formula, α is the custom learning rate, β1 is the first-order decay coefficient, β2 is the second-order decay coefficient, t is the time step, and θ t represents the model parameters at time step t, m t represents the first-order moment vector, v t represents the second-order moment vector, m′ t represents the corrected first-order moment vector, v′ t represents the corrected second-order moment vector, ò represents a small constant used for numerical stability, usually ò=10 -8 ;

[0047] At the end of each round, ensure that the forward propagation and parameter update of all training data have been completed, and before verification, record the weights and biases of the current model; use the verification data set independent of training, which is not used to update the model parameters during the training process, input the data of the verification set into the current model, perform forward propagation, obtain the predicted output of the model, use the loss function to calculate the error between the model output and the true label on the verification set, and calculate the proportion of correctly predicted samples to the total samples; the forward propagation and reverse propagation steps are continuously iterated until the maximum number of training times is reached or the conditions are met;

[0048] The satisfying condition means that the loss value of the cross entropy loss function in the forward propagation process of the monitoring data set |S′-S|<e, which means that the training process of the model is close to completion, and the weights and biases of the model are continuously updated through the optimization algorithm during the training process, and finally reach a stable state and no longer change significantly. The model has found a global optimal solution, indicating that the training process is successful; the weights and biases at this time are assigned to the model, and the model can be used to predict the instance data; if the fitness function does not meet the condition before completing the set number of iterations, it means that the objective function may have high complexity or nonlinearity , resulting in the optimization process requiring more iterations to find a better solution, the number of iterations is increased to 135% of the original number of iterations, and the iterative operation is continued; if the fitness function meets the conditions before completing the set number of iterations, the iteration is stopped and the network parameters obtained by training are returned; if the fitness function still does not meet the conditions before completing the set second number of iterations, consider adjusting the hyperparameters of the optimization algorithm: if the loss value fluctuates greatly during the training process, consider adjusting the learning rate. If the loss fluctuates greatly during the training process, it means that the learning rate is set too high. Reduce the learning rate to 10% of the original learning rate to make the model The update is smoother; if the loss curve decreases slowly, increase the learning rate to 10 times the original learning rate; if the model convergence is unstable, consider adjusting the first-order attenuation coefficient β1 and increase it to 110% of the original first-order attenuation coefficient to make the first-order moment estimate smoother and reduce the sensitivity to noise, thereby improving the convergence stability; if the gradient explodes or disappears during model training, consider adjusting the second-order attenuation coefficient β2. If the model converges slowly and the loss function does not change significantly, increase the second-order attenuation coefficient to 101% of the original coefficient to help the optimizer better adapt to the complex loss surface; if the model shows signs of overfitting, the training When the training loss continues to decrease but the validation loss increases, reduce the large second-order attenuation coefficient to 99% of the original coefficient, so that the model can respond more flexibly to new data and increase its generalization ability; each time a hyperparameter is changed, observe its impact on model performance, keep detailed experimental records, including hyperparameter settings, training time, and model performance indicators, and conduct a comprehensive analysis based on the loss curve, training and validation set performance to find the most appropriate parameter combination; if it meets the requirements, stop iterative updates, assign the updated parameters to the neural network, and test the performance of the neural network with the test sample data, and finally obtain a mature neural network parameter model.

[0049] By adopting the above technical solution, compared with the prior art, the present invention has the following beneficial effects:

[0050] The present invention establishes for the first time a roof stability prediction model based on parameters such as the comprehensive anchor support quality of the tunnel, the inclination of the roof rock layer, the tunnel surrounding rock stress, the roof strength, and the coal seam burial depth; and can complete real-time prediction of the roof stability based on real-time data transmitted back by on-site sensors, with high prediction accuracy and reduced prediction complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of this article or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of this article. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0052] Figure 1 It is a flow chart of the intelligent identification method of roof stability of the present invention;

[0053] Figure 2 Flowchart for training roof stability prediction model;

[0054] Figure 3 It is a schematic diagram of the network structure;

[0055] Figure 4 Schematic diagrams used for roof stability prediction model runs;

[0056] Figure 5 Predict interfaces for roof stability prediction models;

[0057] Figure 6 Return interface for roof stability prediction model results;

[0058] Figure 7 This is the design drawing of bolt support for composite roof roadway;

[0059] Figure 8 Schematic diagram of nondestructive testing of a single anchor rod. DETAILED DESCRIPTION

[0060] The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and examples.

[0061] An intelligent identification method for composite roof stability of coal roadway based on machine learning, such as Figure 1 As shown, the following steps are included:

[0062] S1. Obtain a sample data set, wherein the sample data set includes the quality of comprehensive bolt support of the roadway, the inclination of the roof rock layer, the stress of the roadway surrounding rock, the strength of the roof, the buried depth of the coal seam, and the classification label of the roof stability;

[0063] S2. Use the comprehensive anchor support quality of the tunnel, the inclination of the roof rock layer, the stress of the tunnel surrounding rock, the strength of the roof, the buried depth of the coal seam and the classification labels of the roof stability to train the neural network model and obtain a trained roof stability prediction model.

[0064] Furthermore, the specific process of obtaining the sample data set in step S1 is as follows:

[0065] In the field of mine safety, the stability prediction of the roof is the key to controlling the safe production of coal mines; the roof stability prediction model is trained using the sample data set to determine the roof instability mode;

[0066] The comprehensive anchor support quality of the tunnel refers to the anchoring quality of multiple anchors in the comprehensive tunnel plane, reflecting the overall anchoring quality of the tunnel section; according to the design diagram of the composite roof tunnel anchor support, Figure 7 As shown in JGJ / T182-2009 "Technical Specifications for Nondestructive Testing of Anchor Bolt Anchoring Quality", the evaluation indicators of a single anchor bolt are the anchor bolt length and the anchor bolt anchoring density; the qualified standard for anchor bolt length is that the shorter length shall not exceed 5% of the anchor bolt design length; the length of a single anchor bolt and the anchor bolt anchoring density are obtained by the anchor bolt nondestructive testing instrument, such as Figure 8 The quality of the comprehensive anchor support of the tunnel is:

[0067]

[0068] Where, Φ is the comprehensive anchor support quality of the tunnel, d i is the distance between the anchor and the center of the tunnel, l i is the actual anchoring length of the anchor rod, l' i Design anchorage length for anchor rods, is the actual anchoring quality of the anchor rod; the roof rock layer inclination, tunnel surrounding rock stress, roof strength, and coal seam burial depth are obtained from relevant coal mine underground geological exploration data; the roof stability classification label is quantified as the tunnel roof height, specifically Class I roof, Class II roof, Class III roof, and Class IV roof;

[0069] It is crucial to normalize the data of the neural network. Normalization can accelerate model convergence, improve numerical stability, eliminate feature bias, and improve the generalization ability of the model. Normalization also ensures that the input values ​​adapt to the effective range of the activation function, thereby improving training efficiency. Scaling the data to the range of [0,1] avoids data with large eigenvalues ​​that may cause numerical instability.

[0070]

[0071] In the formula, H a , H A are the values ​​before and after normalization, H maxis the maximum value of each group of data, H min is the minimum value of each group of data.

[0072] The specific process of obtaining the sample data set in step S2 is as follows: divide the sample data set into a training data set, a validation set, and a test set in a ratio of 14:3:3, determine the appropriate network architecture, define the number of network layers, the number of neurons in each layer, and the activation function, select the appropriate loss function and optimization algorithm, iterate the training set multiple times, perform forward propagation, calculate loss, and back propagation in each training round, and evaluate the performance of the final roof stability prediction model on the test set.

[0073] like Figure 2 As shown, the process of training the roof stability prediction model in step S2 specifically includes the following steps:

[0074] Step 1: Use the sample data set to divide it into a training sample data set and a test sample data set. The training sample data set is used to train the neural network model and optimize the parameters to determine the roof stability prediction model; the test sample data set is used to evaluate the performance of the trained model on unseen data, help determine the accuracy and generalization ability of the model, detect whether the model is overfitting on the training data, and ensure the reliability of the model in practical applications; determine the number of input layer nodes according to the high-dimensional data at the input end, and determine the hidden layer hyperparameters of the network topology structure according to the Kolmogorov theorem. The single hidden layer structure can map the nonlinear input-output relationship with arbitrary precision. The number of hidden layer nodes is calculated by the following formula, and then refined by trial and error method to minimize the prediction error; the Kolmogorov theorem equation is:

[0075]

[0076] Where a is the number of neurons in the hidden layer, m is the number of neurons in the input layer, n is the number of neurons in the output layer, and b is a constant between [1,10];

[0077] Step 2: According to the historical sample database, determine that the number of neurons in the input layer of the network is equal to the number of features of the input data, which can be represented by a vector as N = [n1, n2, ..., nn input ] T ; The output layer can be represented by a vector M = [m1, m2, m3, m4] T ; W (1) Indicates that the components are The weight matrix between the input layer neurons and the hidden layer neurons, W (2) Indicates that the components are The weight matrix between the hidden layer neurons and the output layer neurons, B (1) , B (2)Represent the bias terms of the hidden layer and the output layer respectively; in deep networks, if the weights are not initialized properly, the signal may have gradient vanishing or gradient explosion during forward propagation; Kaiming initialization helps keep the variance of the output of each layer consistent by setting a suitable variance, thus avoiding these problems. Therefore, Kaiming normal distribution is used to calculate the W of the neural network. (1) , W (2) Initialize, where Where N(0,σ 2 ) means the mean is μ and the variance is σ 2 The normal distribution of , U(-c,c) represents the uniform distribution in the range [-c, c], n input represents the number of neurons in the input layer, m input Represents the number of neurons in the hidden layer. The specific network structure diagram is as follows Figure 3 As shown;

[0078] Step 3: In order to explore the possible parameter space in the early stage of model training, adjust the differences between individuals in the population to improve the convergence speed of the group, while maintaining the diversity of the group, and solve combinatorial optimization problems, nonlinear programming problems and other complex optimization algorithm problems.

[0079] Furthermore, the basic structure of the optimization algorithm described in step 3 mainly includes mutation operation, crossover operation and selection operation. First, assuming that the dimension of each individual in the initialization population is G, the individuals contained in the population are: X i =(χ i,1 , χ i,2 ,…,χ i,G ), i∈NP; where NP is the number of individuals in the population, and each individual Contains D variables, so the randomly generated initialization population is as follows:

[0080] Where: and are the upper and lower limits of each variable, respectively, and rand(0,1) represents a random number in the range of 0 to 1; the sequence generated by the Tent chaotic map has strong randomness and uncertainty, and can generate widely distributed and diverse initial individuals; this diversity helps to avoid the premature convergence of the optimization algorithm of the present invention in the search space, making it easier to explore the global optimal solution, and the chaotic sequence is generated in a way that each iteration contains certain regularity and structure, which can help the algorithm to search in an effective direction at an early stage and improve the convergence speed. The new initialization population obtained by performing the Tent chaotic map is as follows:

[0081]

[0082] Where: Indicates that a new initialization population is generated through the Tent chaotic map;

[0083] In order to prevent the population from converging to the local optimal solution too early, new gene combinations are introduced to increase the diversity of solutions, help the algorithm explore a wider solution space, discover potential high-quality solutions, and improve overall performance. Three different individuals are randomly selected from the population, and the vector difference between two of them is superimposed on the third individual to obtain a new variant individual. The mutation operation is as follows:

[0084] R i,G+1 = x r1,G+1 +F(χ r2,G+1 -x r3,G+1 )

[0085] In the formula: r1, r2, r3 represent random numbers randomly selected from the population [0, NP], F is the variation factor, and its value is a real constant in the range of [0, 2]. i With the variant individual R i,G+1 Perform crossover operation to obtain new individuals The crossover operation formula is as follows:

[0086]

[0087] Where: rand j represents the random number corresponding to the jth variable; CR is the crossover factor; rand(i) represents a random integer in the range [1, NP]; after calculating the crossover operation, the individual S is obtained i,G+1 The fitness of the G-generation individual χ i,G The fitness of the next generation population is compared with that of the next generation population, and the one closer to the true value is selected as the next generation population individual Z i,G+1 , the formula for the selection operation is as follows:

[0088]

[0089] In the formula, the fitness function f is the objective function of the optimization problem. In this paper, the real value Y is selected i and the predicted value Y i 0 The root mean square error is used as the fitness function, and its calculation formula is as follows:

[0090]

[0091] Cyclic crossover, mutation, and selection operations until the fitness function meets the condition or reaches the maximum number of iterations; the fitness function meets the condition that the fitness |f'-f| < ξ, indicating that the fitness curve gradually decreases and tends to be stable, and the change of the fitness value between consecutive generations tends to zero;

[0092] Where ξ represents a very small positive number, generally 0.1; f' represents the fitness value obtained by the next cycle of f; if the fitness function meets the conditions, it proves that the performance of the model on the training set or the validation set tends to be stable, and the current parameter configuration has basically found the optimal or suboptimal solution, and further iterations will not significantly improve the fitness value, then the iteration is stopped and the trained network parameters are returned; if the fitness function does not meet the conditions before completing the set number of iterations, it means that the objective function may have high complexity or nonlinearity, resulting in the optimization process requiring more iterations to find a better solution, then the number of iterations is increased to 150% of the original number of iterations, and the iterative operation is continued; if the fitness function meets the conditions before completing the set number of iterations, then the iteration is stopped and the training is returned The obtained network parameters; if the fitness function still does not meet the conditions before completing the set number of secondary iterations, consider adjusting the hyperparameters of the optimization algorithm: if the fitness change fluctuates greatly, reduce the mutation factor F to 85% of the original size to improve stability; if the optimization process lacks diversity, increase the mutation factor F to 125% of the original size to make the search more exploratory; if it is found that the diversity of new individuals is insufficient, try to increase the crossover factor CR to 125% of the original size to introduce more genetic information; if the fitness of the new individual is poor and the convergence speed is slow, consider reducing the crossover factor CR to 85% of the original size to rely more on the existing better individuals; verify the effects of different parameter settings through multiple experiments, record the fitness changes and convergence speed, and find the optimal parameter combination;

[0093] The iterated weights and thresholds are assigned to the neural network to train the neural network. The forward propagation process of the network can be expressed as:

[0094]

[0095] In the formula, a (1) represents the output of the hidden layer of the neural network, a (2) Represents the prediction result of the neural network; Rule() means passing the result of the linear transformation of the input layer to the Rule activation function for processing, and Softmax() means passing the result of the linear transformation of the hidden layer to the Softmax activation function for processing; the predicted value a is obtained (2) Finally, the cross entropy loss function S is used to calculate the error between the predicted value and the true value:

[0096]

[0097] In the formula, n is the number of categories; the gradient is calculated by the back propagation algorithm, and the partial derivative of the loss function for each parameter is calculated layer by layer using the chain rule. The objective function gradient at time t is calculated, the biased first-order moment vector and second-order moment vector are updated, and the bias-corrected first-order moment vector and second-order data vector are calculated. The weights and thresholds are updated. The specific formula is as follows:

[0098]

[0099] In the formula, α is the custom learning rate, β1 is the first-order decay coefficient, β2 is the second-order decay coefficient, t is the time step, and θ t represents the model parameters at time step t, m t represents the first-order moment vector, v t represents the second-order moment vector, m' t represents the corrected first-order moment vector, v' t represents the corrected second-order moment vector, ò represents a small constant used for numerical stability, usually ò=10 -8 ;

[0100] At the end of each round, ensure that the forward propagation and parameter update of all training data have been completed, and before verification, record the weights and biases of the current model; use the verification data set independent of training, which is not used to update the model parameters during the training process, input the data of the verification set into the current model, perform forward propagation, obtain the predicted output of the model, use the loss function to calculate the error between the model output and the true label on the verification set, and calculate the proportion of correctly predicted samples to the total samples; the forward propagation and reverse propagation steps are continuously iterated until the maximum number of training times is reached or the conditions are met;

[0101] The satisfying condition mentioned above means that the loss value of the cross entropy loss function in the forward propagation process of the monitoring data set |S'-S|<e, which means that the training process of the model is close to completion, and the weights and biases of the model are continuously updated through the optimization algorithm during the training process, and finally reach a stable state and no longer change significantly. The model has found a global optimal solution, indicating that the training process is successful; the weights and biases at this time are assigned to the model, and the model can be used to predict the instance data; if the fitness function does not meet the condition before completing the set number of iterations, it means that the objective function may have high complexity or nonlinearity , resulting in the optimization process requiring more iterations to find a better solution, the number of iterations is increased to 135% of the original number of iterations, and the iterative operation is continued; if the fitness function meets the conditions before completing the set number of iterations, the iteration is stopped and the network parameters obtained by training are returned; if the fitness function still does not meet the conditions before completing the set second number of iterations, consider adjusting the hyperparameters of the optimization algorithm: if the loss value fluctuates greatly during the training process, consider adjusting the learning rate. If the loss fluctuates greatly during the training process, it means that the learning rate is set too high. Reduce the learning rate to 10% of the original learning rate to make the model The update is smoother; if the loss curve decreases slowly, increase the learning rate to 10 times the original learning rate; if the model convergence is unstable, consider adjusting the first-order attenuation coefficient β1 and increase it to 110% of the original first-order attenuation coefficient to make the first-order moment estimate smoother and reduce the sensitivity to noise, thereby improving the convergence stability; if the gradient explodes or disappears during model training, consider adjusting the second-order attenuation coefficient β2. If the model converges slowly and the loss function does not change significantly, increase the second-order attenuation coefficient to 101% of the original coefficient to help the optimizer better adapt to the complex loss surface; if the model shows signs of overfitting, the training When the training loss continues to decrease but the validation loss increases, reduce the large second-order attenuation coefficient to 99% of the original coefficient, so that the model can respond more flexibly to new data and increase its generalization ability; each time a hyperparameter is changed, observe its impact on model performance, keep detailed experimental records, including hyperparameter settings, training time, and model performance indicators, and conduct a comprehensive analysis based on the loss curve, training and validation set performance to find the most appropriate parameter combination; if it meets the requirements, stop iterative updates, assign the updated parameters to the neural network, and test the performance of the neural network with the test sample data, and finally obtain a mature neural network parameter model.

[0102] Figure 4 As shown, the operation logic of the roof stability prediction model in the present invention specifically includes the following steps:

[0103] Step 1: Obtain parameters such as the comprehensive anchor support quality of the roadway of the tested roof, the inclination of the roof rock layer, the roadway surrounding rock stress, the roof strength, the coal seam burial depth and the roof stability classification label. In this step, the comprehensive anchor support quality of the roadway refers to the anchoring quality of multiple anchors in the plane of the comprehensive roadway, reflecting the overall anchoring quality of the roadway section. The design diagram of the composite roof roadway anchor support is shown in Figure 1. Figure 7 As shown. According to JGJ / T182-2009 "Technical Specifications for Nondestructive Testing of Anchor Bolt Anchoring Quality", the evaluation indicators of a single anchor bolt are the anchor bolt length and anchor bolt anchoring density. The qualified standard for anchor bolt length is that the shorter length shall not exceed 5% of the anchor bolt design length; the length of a single anchor bolt and the anchor bolt anchoring density are obtained by the anchor bolt nondestructive testing instrument, such as Figure 8 The quality of the comprehensive anchor support of the tunnel is:

[0104]

[0105] Where, Φ is the comprehensive anchor support quality of the tunnel, d i is the distance between the anchor and the center of the tunnel, l i is the actual anchoring length of the anchor rod, l' i Design anchorage length for anchor rods, The actual anchoring quality of the anchor rod. The roof rock layer inclination, tunnel surrounding rock stress, roof strength, and coal seam burial depth are obtained from relevant coal mine underground geological exploration data. The roof stability classification label is quantified as the tunnel roof height, specifically Class I roof, Class II roof, Class III roof, and Class IV roof.

[0106] Step 2: Run the client application based on PyQt5 in the present invention, enter the user name and password as prompted to enter the user interface, and enter the parameters such as the tunnel comprehensive anchor support quality of the top plate to be tested, the inclination of the top plate rock layer, the tunnel surrounding rock stress, the roof strength, the coal seam burial depth, etc. into the user interface as prompted by the client user interface. After the input is completed, click the "Prediction" button on the interface. After the prediction is completed, the result will be displayed in the pop-up window. If the input format is incorrect or the number of features does not match, the application will display an error message in the result label. After completion of use, you can directly close the application window and the program will exit normally. Figure 5 , Figure 6 shown.

[0107] The above embodiments illustrate the basic principles and features of the present invention, but the above only illustrates the preferred embodiments of the present invention and is not limited to the embodiments. Under the inspiration of this patent, a person skilled in the art can make many forms of deformation and improvement without departing from the scope of protection of the present invention and the claims, which are all within the protection scope of the present invention. Therefore, the patent and protection scope of the present invention shall be subject to the attached claims.

Claims

1. A method for intelligent identification of composite roof stability of coal roadway based on machine learning, characterized in that: The steps include: S1. Obtain a sample data set, wherein the sample data set includes the quality of comprehensive bolt support of the roadway, the inclination of the roof rock layer, the stress of the roadway surrounding rock, the strength of the roof, the buried depth of the coal seam, and the classification label of the roof stability; S2. Use the comprehensive anchor support quality of the tunnel, the inclination of the roof rock layer, the stress of the tunnel surrounding rock, the strength of the roof, the buried depth of the coal seam and the classification labels of the roof stability to train the neural network model and obtain a trained roof stability prediction model.

2. The method for intelligently identifying the stability of composite roof of coal roadway based on machine learning according to claim 1, characterized in that: The specific process of obtaining the sample data set in step S1 is as follows: In the field of mine safety, tunnel roof disasters are the most frequent and fatal disasters among all types of mining disasters in my country. Roof stability prediction is the key to controlling coal mine safety production. The sample data set is used to train the roof stability prediction model to determine the roof instability mode. The comprehensive anchor support quality of the tunnel refers to the anchoring quality of multiple anchors in the plane of the comprehensive tunnel, reflecting the overall anchoring quality of the tunnel section; according to the anchor support design drawing of the composite roof tunnel and JGJ / T182-2009 "Technical Specifications for Nondestructive Testing of Anchor Anchoring Quality", the evaluation indicators of a single anchor are the anchor length and anchor density; the qualified standard for anchor length is that the shorter length shall not exceed 5% of the designed anchor length; the length of a single anchor and the anchor density are obtained by the anchor nondestructive testing instrument. The comprehensive anchor support quality of the tunnel is: Where, Φ is the comprehensive anchor support quality of the tunnel, d i is the distance between the anchor and the center of the tunnel, l i is the actual anchoring length of the anchor rod, l i 'Design anchorage length for anchor rod, is the actual anchoring quality of the anchor rod; the roof rock layer inclination, tunnel surrounding rock stress, roof strength, and coal seam burial depth are obtained from relevant coal mine underground geological exploration data; the roof stability classification label is quantified as the tunnel roof height, specifically Class I roof, Class II roof, Class III roof, and Class IV roof; It is crucial to normalize the data of the neural network. Normalization can accelerate model convergence, improve numerical stability, eliminate feature bias, and improve the generalization ability of the model. Normalization also ensures that the input values ​​adapt to the effective range of the activation function, thereby improving training efficiency. Scaling the data to the range of [0,1] avoids data with large eigenvalues ​​that may cause numerical instability. In the formula, H a , H A are the values ​​before and after normalization, H max is the maximum value of each group of data, H min is the minimum value of each group of data.

3. The method for intelligently identifying the stability of composite roof of coal roadway based on machine learning according to claim 1 or 2, characterized in that: The specific process of obtaining the sample data set in step S2 is as follows: divide the sample data set into a training data set, a validation set, and a test set in a ratio of 14:3:3, determine the appropriate network architecture, define the number of network layers, the number of neurons in each layer, and the activation function, select the appropriate loss function and optimization algorithm, iterate the training set multiple times, perform forward propagation, calculate loss, and back propagation in each training round, and evaluate the performance of the final roof stability prediction model on the test set.

4. The method for intelligently identifying the stability of composite roof of coal roadway based on machine learning according to claim 3 is characterized by: The process of training the roof stability prediction model in step S2 specifically includes the following steps: Step 1: The sample data set is divided into a training sample data set and a test sample data set. The training sample data set is used to train the neural network model and optimize the parameters to determine the roof stability prediction model. The test sample data set is used to evaluate the performance of the trained model on unseen data, help determine the accuracy and generalization ability of the model, detect whether the model is overfitting on the training data, and ensure the reliability of the model in practical applications; The number of input layer nodes is determined according to the high-dimensional data at the input end, and the hidden layer hyperparameters of the network topology are determined according to the Kolmogorov theorem. The single hidden layer structure can map the nonlinear input-output relationship with arbitrary precision. The number of hidden layer nodes is calculated by the following formula, and then refined by trial and error method to minimize the prediction error. The Kolmogorov theorem equation is: Where a is the number of neurons in the hidden layer, m is the number of neurons in the input layer, n is the number of neurons in the output layer, and b is a constant between [1,10]; Step 2: According to the historical sample database, determine that the number of neurons in the input layer of the network is equal to the number of features of the input data, which can be represented by a vector as follows: The output layer can be represented by a vector M = [m1, m2, m3, m4] T ; W (1) Indicates that the components are The weight matrix between the input layer neurons and the hidden layer neurons, W (2) Indicates that the components are The weight matrix between the hidden layer neurons and the output layer neurons, B (1) , B (2) Represent the bias terms of the hidden layer and the output layer respectively; in deep networks, if the weights are not initialized properly, the signal may have gradient vanishing or gradient explosion during forward propagation; Kaiming initialization helps keep the variance of the output of each layer consistent by setting a suitable variance, thus avoiding these problems. Therefore, Kaiming normal distribution is used to calculate the W of the neural network. (1) , W (2) Initialize, where Where N(0,σ 2 ) means the mean is μ and the variance is σ 2 The normal distribution of , U(-c,c) represents the uniform distribution in the range [-c, c], n input represents the number of neurons in the input layer, m input Represents the number of neurons in the hidden layer; Step 3: In order to explore the possible parameter space in the early stage of model training, adjust the differences between individuals in the population to improve the convergence speed of the group, while maintaining the diversity of the group, and solve combinatorial optimization problems, nonlinear programming problems and other complex optimization algorithm problems.

5. The method for intelligently identifying the stability of composite roof of coal roadway based on machine learning according to claim 4 is characterized by: The basic structure of the optimization algorithm described in step 3 mainly includes mutation operation, crossover operation and selection operation. First, assume that the dimension of each individual in the initialization population is G, then the individuals contained in the population are: X i =(χ i,1 , χ i,2 ,…,χ i,G ), i∈NP; where NP is the number of individuals in the population, and each individual Contains D variables, so the randomly generated initialization population is as follows: Where: and are the upper and lower limits of each variable, respectively, and rand(0,1) represents a random number in the range of 0 to 1; the sequence generated by the Tent chaotic map has strong randomness and uncertainty, and can generate widely distributed and diverse initial individuals; this diversity helps to avoid the premature convergence of the optimization algorithm of the present invention in the search space, making it easier to explore the global optimal solution, and the chaotic sequence is generated in a way that each iteration contains certain regularity and structure, which can help the algorithm to search in an effective direction at an early stage and improve the convergence speed. The new initialization population obtained by performing the Tent chaotic map is as follows: Where: Indicates that a new initialization population is generated through the Tent chaotic map; In order to prevent the population from converging to the local optimal solution too early, new gene combinations are introduced to increase the diversity of solutions, help the algorithm explore a wider solution space, discover potential high-quality solutions, and improve overall performance. Three different individuals are randomly selected from the population, and the vector difference between two of them is superimposed on the third individual to obtain a new variant individual. The mutation operation is as follows: R i,G+1 =x r1,G+1 +F(x r2,G+1 -x r3,G+1 ) In the formula: r1, r2, r3 represent random numbers randomly selected from the population [0, NP], F is the variation factor, and its value is a real constant in the range of [0, 2]. i With the variant individual R i,G+1 Perform crossover operation to obtain new individuals The crossover operation formula is as follows: Where: rand j represents the random number corresponding to the jth variable; CR is the crossover factor; rand(i) represents a random integer in the range [1, NP]; after calculating the crossover operation, the individual S is obtained i,G+1 The fitness of the G-generation individual X i,G The fitness of the next generation population is compared with that of the next generation population, and the one closer to the true value is selected as the next generation population individual Z i,G+1 , the formula for selecting the operation is as follows: In the formula, the fitness function f is the objective function of the optimization problem. In this paper, the real value Y is selected i and the predicted value Y i 0 The root mean square error is used as the fitness function, and its calculation formula is as follows: Cyclic crossover, mutation, and selection operations until the fitness function meets the conditions or reaches the maximum number of iterations; the fitness function meets the conditions, that is, the fitness |f'-f|<ξ, indicating that the fitness curve gradually decreases and tends to be stable, and the change of the fitness value between consecutive generations tends to zero; Where ξ represents a very small positive number, generally 0.1; f' represents the fitness value obtained by the next cycle of f; if the fitness function meets the conditions, it proves that the performance of the model on the training set or the validation set tends to be stable, and the current parameter configuration has basically found the optimal or suboptimal solution, and further iterations will not significantly improve the fitness value, then the iteration is stopped and the trained network parameters are returned; if the fitness function does not meet the conditions before completing the set number of iterations, it means that the objective function may have high complexity or nonlinearity, resulting in the optimization process requiring more iterations to find a better solution, then the number of iterations is increased to 150% of the original number of iterations, and the iterative operation is continued; if the fitness function meets the conditions before completing the set number of iterations, then the iteration is stopped and the training is returned The obtained network parameters; if the fitness function still does not meet the conditions before completing the set number of secondary iterations, consider adjusting the hyperparameters of the optimization algorithm: if the fitness change fluctuates greatly, reduce the mutation factor F to 85% of the original size to improve stability; if the optimization process lacks diversity, increase the mutation factor F to 125% of the original size to make the search more exploratory; if it is found that the diversity of new individuals is insufficient, try to increase the crossover factor CR to 125% of the original size to introduce more genetic information; if the fitness of the new individual is poor and the convergence speed is slow, consider reducing the crossover factor CR to 85% of the original size to rely more on the existing better individuals; verify the effects of different parameter settings through multiple experiments, record the fitness changes and convergence speed, and find the optimal parameter combination; The iterated weights and thresholds are assigned to the neural network to train the neural network. The forward propagation process of the network can be expressed as: In the formula, a (1) represents the output of the hidden layer of the neural network, a (2) Represents the prediction result of the neural network; Rule() means passing the result of the linear transformation of the input layer to the Rule activation function for processing, and Softmax() means passing the result of the linear transformation of the hidden layer to the Softmax activation function for processing; the predicted value a is obtained (2) Finally, the cross entropy loss function S is used to calculate the error between the predicted value and the true value: In the formula, n is the number of categories; the gradient is calculated by the back propagation algorithm, and the partial derivative of the loss function for each parameter is calculated layer by layer using the chain rule. The objective function gradient at time t is calculated, the biased first-order moment vector and second-order moment vector are updated, and the bias-corrected first-order moment vector and second-order data vector are calculated. The weights and thresholds are updated. The specific formula is as follows: In the formula, α is the custom learning rate, β1 is the first-order decay coefficient, β2 is the second-order decay coefficient, t is the time step, and θ t represents the model parameters at time step t, m t represents the first-order moment vector, v t represents the second-order moment vector, m t ' represents the corrected first-order moment vector, v′ t represents the corrected second-order moment vector, ò represents a small constant used for numerical stability, usually ò=10 -8 ; At the end of each round, ensure that the forward propagation and parameter update of all training data have been completed, and before verification, record the weights and biases of the current model; use the verification data set independent of training, which is not used to update the model parameters during the training process, input the data of the verification set into the current model, perform forward propagation, obtain the predicted output of the model, use the loss function to calculate the error between the model output and the true label on the verification set, and calculate the proportion of correctly predicted samples to the total samples; the forward propagation and reverse propagation steps are continuously iterated until the maximum number of training times is reached or the conditions are met; The satisfying condition mentioned above means that the loss value of the cross entropy loss function in the forward propagation process of the monitoring data set |S'-S|<e, which means that the training process of the model is close to completion, and the weights and biases of the model are continuously updated through the optimization algorithm during the training process, and finally reach a stable state and no longer change significantly. The model has found a global optimal solution, indicating that the training process is successful; the weights and biases at this time are assigned to the model, and the model can be used to predict the instance data; if the fitness function does not meet the condition before completing the set number of iterations, it means that the objective function may have high complexity or nonlinearity , resulting in the optimization process requiring more iterations to find a better solution, the number of iterations is increased to 135% of the original number of iterations, and the iterative operation is continued; if the fitness function meets the conditions before completing the set number of iterations, the iteration is stopped and the network parameters obtained by training are returned; if the fitness function still does not meet the conditions before completing the set second number of iterations, consider adjusting the hyperparameters of the optimization algorithm: if the loss value fluctuates greatly during the training process, consider adjusting the learning rate. If the loss fluctuates greatly during the training process, it means that the learning rate is set too high. Reduce the learning rate to 10% of the original learning rate to make the model The update is smoother; if the loss curve decreases slowly, increase the learning rate to 10 times the original learning rate; if the model convergence is unstable, consider adjusting the first-order attenuation coefficient β1 and increase it to 110% of the original first-order attenuation coefficient to make the first-order moment estimate smoother and reduce the sensitivity to noise, thereby improving the convergence stability; if the gradient explodes or disappears during model training, consider adjusting the second-order attenuation coefficient β2. If the model converges slowly and the loss function does not change significantly, increase the second-order attenuation coefficient to 101% of the original coefficient to help the optimizer better adapt to the complex loss surface; if the model shows signs of overfitting, the training When the training loss continues to decrease but the validation loss increases, reduce the large second-order attenuation coefficient to 99% of the original coefficient, so that the model can respond more flexibly to new data and increase its generalization ability; each time a hyperparameter is changed, observe its impact on model performance, keep detailed experimental records, including hyperparameter settings, training time, and model performance indicators, and conduct a comprehensive analysis based on the loss curve, training and validation set performance to find the most appropriate parameter combination; if it meets the requirements, stop iterative updates, assign the updated parameters to the neural network, and test the performance of the neural network with the test sample data, and finally obtain a mature neural network parameter model.

Images