A PSO-BO-DNN-based intelligent prediction method for roadway excavation floor failure depth

Abstract

This invention discloses an intelligent prediction method for the depth of floor failure in roadway excavation based on PSO-BO-DNN, which addresses the problems of low prediction accuracy and poor generalization ability caused by the randomness of initial weights and subjective setting of hyperparameters in traditional deep neural networks. First, Z-score standardization and grey relational analysis are performed on historical data such as mining depth, mining thickness, roadway length, and the proportion of hard rock in the floor. A basic DNN model with three hidden layers is constructed, and Bayesian optimization is used to globally optimize hyperparameters such as the number of neurons and learning rate to form the optimal DNN model. Next, particle swarm optimization is used to globally optimize the initial weight matrix of the optimal DNN to obtain the PSO-BO-DNN prediction model. Finally, the roadway parameters to be predicted are input, and the floor failure depth value is output. This invention achieves dual automatic optimization of hyperparameters and initial weights, with a model determination coefficient R² exceeding 0.8, significantly improving prediction accuracy and generalization ability. It can be used for risk assessment of water hazards to the floor in coal mine roadway excavation and support design.

Description

Technical Field

[0001] This invention relates to the field of mine safety technology, and in particular to an intelligent prediction method for the depth of damage to the tunnel floor based on PSO-BO-DNN. Background Technology

[0002] Accurate prediction of the depth of floor failure during coal mine roadway excavation is of great significance for floor water hazard prevention, roadway support design, and ensuring safe mine production. Floor failure is a complex nonlinear behavior under the combined influence of geological and mining conditions, and its influencing factors include coal seam depth, coal thickness, coal seam dip angle, mining method, roadway length, and the hard rock ratio coefficient of the floor strata.

[0003] Currently, the methods for predicting the depth of base plate failure mainly fall into the following three categories: (1) Empirical formula method: Statistical regression model is established based on engineering measured data. It is simple in form but has limited accuracy and is difficult to consider the nonlinear coupling relationship of multiple factors. (2) Numerical simulation methods: such as finite element method (FEM) and discrete element method (DEM), although they can simulate complex stress and strain fields, they are time-consuming to calculate, require high computing resources, and have strong subjectivity in parameter selection; (3) Machine learning methods: In recent years, intelligent algorithms such as support vector machine (SVM), random forest (RF), and standard deep neural network (DNN) have been introduced into the prediction of base plate destruction depth. However, the standard DNN model has obvious defects: ① The initial weights are randomly generated, which is easy to get trapped in local optima; ② The hyperparameters (network structure, learning rate, regularization coefficient, etc.) depend on human experience to set, which makes it difficult to achieve the global optimal configuration; ③ The model has insufficient generalization ability and is prone to underfitting or overfitting.

[0004] Therefore, there is an urgent need for a method to predict the depth of base plate damage that can automatically optimize the initial weights and hyperparameters of the DNN and improve prediction accuracy and generalization ability. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the purpose of this invention is to provide an intelligent prediction method for the depth of damage to the tunnel floor based on PSO-BO-DNN, which can solve the technical problems of low prediction accuracy and poor generalization ability caused by the randomness of initial weights and the subjectivity of hyperparameter settings in traditional DNN models.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: This invention provides an intelligent prediction method for the failure depth of the tunnel floor based on PSO-BO-DNN. The predicted failure depth is used to guide the design of tunnel support, the determination of the grouting reinforcement range of the floor, and the risk assessment of confined water hazards. The method specifically includes the following steps: S1: Data Acquisition and Preprocessing: Collect historical measured data of coal mine roadway excavation, including mining depth, mining thickness, roadway length, hard rock ratio coefficient of the floor, and measured value of floor failure depth as the prediction target; preprocess the historical measured data to form a model training sample set; S2: Constructing the PSO-BO-DNN prediction model: S2.1: Construct a basic DNN model, which includes an input layer, at least one hidden layer, and an output layer; preferably, there are 3 hidden layers. S2.2: The Bayesian optimization algorithm is used to globally optimize the hyperparameters of the basic DNN model to obtain the optimal DNN model with the optimal combination of hyperparameters. S2.3: The initial weight matrix of the optimal DNN model is globally optimized using the particle swarm optimization algorithm. The mean square error between the prediction results and the measured results on the validation set is used as the fitness function. The PSO-BO-DNN prediction model is obtained by iteratively searching for the globally optimal initial weights through particle swarm optimization. Furthermore, the parameters of the particle swarm optimization algorithm are set as follows: number of particles 40, number of iterations 100, acceleration factors c1 and c2 are positive constants, and momentum term coefficient w is dynamically adjusted.

[0007] S3: Model training and validation: The PSO-BO-DNN prediction model is trained using the model training sample set, and the prediction performance of the PSO-BO-DNN prediction model is evaluated using the validation set in the model training sample set. S4: Prediction of floor failure depth: Input the influencing factor data of the roadway to be predicted into the trained PSO-BO-DNN prediction model, and output the corresponding predicted value of floor failure depth. The influencing factor data includes: the mining depth, mining thickness, roadway length and floor hard rock ratio coefficient of the predicted roadway.

[0008] Preferably, in step S1, the preprocessing of historical measured data includes: using the Z-score standardization method to perform dimensionless processing on the historical measured data, as shown in the formula: In the formula, X is the original sample sequence, Y is the standardized sample sequence, σ is the sample mean, and σ(X) is the sample standard deviation.

[0009] Preferably, in step S1, the Z-score standardization method is used to perform dimensionless processing on the historical measured data to eliminate the differences in the dimensions of each indicator; before model training, the grey relational analysis method is used to verify the correlation between the data of each influencing factor and the depth of damage to the base plate, and the factors input into the model are selected based on the correlation to ensure the rationality of the selected main control factors. Specifically, let the depth of floor failure be the reference sequence X0, and let the mining depth, mining thickness, roadway length and floor hard rock ratio coefficient constitute the comparison sequence X1, X2, X3 and X4. Calculate the grey relational degree. When the resolution coefficient ρ in the grey relational degree analysis is 0.5, and the correlation degree between each influencing factor and the depth of floor failure is greater than 0.75, it is determined that the input factor selection is reasonable.

[0010] Preferably, in step S2.2, the Bayesian optimization algorithm uses a Gaussian process as the prior function and the probability of improvement (POI) as the acquisition function to construct a surrogate model of the objective function. The number of iterations is set to 25, the number of initial sampling points is set to 10, and the next set of hyperparameters to be evaluated is selected based on the acquisition function. The optimal combination of hyperparameters is obtained through iterative optimization.

[0011] Preferably, in step S2.2, the hyperparameters include the number of neurons in each Dense layer, the dropout ratio in the Dropout layer, the training batch size, the learning rate, and the L2 regularization coefficient. The fitness function is the mean squared error between the prediction results and the actual results on the validation set. The optimal combination of hyperparameters is iteratively searched to form the BO-DNN model.

[0012] Furthermore, the optimization range of hyperparameters is as follows: number of neurons in the first Dense layer [32,256], number of neurons in the second Dense layer [16,128], number of neurons in the third Dense layer [8,64], learning rate [0.00005,0.005], dropout ratio [0.1,0.4], training batch size [8,32], and L2 regularization coefficient [0.00001,0.001].

[0013] Preferably, in step S3, the model training sample set is divided into a training set and a validation set in an 8:2 ratio; the Adam algorithm is used to dynamically optimize the learning rate.

[0014] Preferably, in step S3, the coefficient of determination R² is used to evaluate the prediction accuracy of the model. When the R² of both the training set and the validation set reaches a preset threshold of 0.8 or above, the model training is considered complete. The formula for calculating the coefficient of determination R² is as follows: In the formula, y iis the true value of the base plate damage depth, is the predicted value of the base plate damage depth, and is the average value of the true values ​​of the base plate damage depth; when the R² of both the training set and the validation set reaches 0.8 or above, the model training is considered complete.

[0015] Preferably, in step S2.1, a Dropout layer is set after each hidden layer to prevent overfitting, and L2 regularization is applied to the weight matrix and bias term of the Dense layer. The activation function is the LeakyReLU function. The input layer receives four feature parameters: mining depth, mining thickness, roadway length, and floor hard rock ratio coefficient; the output layer outputs the predicted value of floor failure depth.

[0016] Beneficial effects: (1) Dual optimization mechanism improves model performance: Particle swarm optimization (PSO) and Bayesian optimization (BO) are introduced into deep neural networks for the first time. The initial weights and hyperparameters of the DNN are optimized in two ways. The PSO algorithm solves the local optimum problem caused by the randomness of the initial weights, and the BO algorithm solves the problem of low efficiency in hyperparameter optimization. The two work together to significantly improve the model prediction accuracy.

[0017] (2) Strong generalization ability: Through the dual regularization mechanism of Dropout layer and L2 regularization, combined with LeakyReLU activation function, gradient vanishing, dead neurons and overfitting are effectively avoided. The model can achieve R² above 0.8 on both training set and validation set, showing good generalization ability.

[0018] (3) High degree of automation: Compared with the traditional method of relying on expert experience to set network parameters, the present invention realizes automatic optimization of network structure and initial weights, reduces manual intervention and improves modeling efficiency.

[0019] (4) Strong adaptability: The model, trained using measured data from previous roadway excavation in coal mines, can adapt to the prediction of floor damage depth in coal mines with similar geological conditions, providing rapid and accurate technical support for water hazard risk assessment before roadway excavation. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart of the PSO-BO-DNN model construction provided by this invention; Figure 2This is a diagram of the PSO-BO-DNN base plate failure depth prediction model architecture provided by the present invention. Figure 3 The Bayesian optimization curve provided by this invention; Figure 4 Fitness curve of the PSO algorithm provided by this invention; Figure 5 Comparison chart of training set prediction results provided by this invention; Figure 6 A comparison chart of the prediction results for the validation set provided by this invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Taking the 1016 mining area of ​​Chaili Coal Mine of Shandong Energy Zaozhuang Mining Group as an example, the intelligent prediction method for roadway excavation floor failure depth based on PSO-BO-DNN of the present invention is used for prediction. The specific steps are as follows: Step 1: Sample data collection and preprocessing; (1) Sample selection; The study area is the Chaili Coal Mine, located in the Tengxian Coalfield of Zaozhuang City, Shandong Province, part of the North China Coalfield System. The measured data selected for this model mainly comes from measured data on the floor failure depth during tunnel excavation in Shandong Province. The geological conditions of these coal mines are quite similar to those of the Chaili Coal Mine, making the comparative analysis and prediction results more meaningful.

[0024] The main factors influencing the depth of floor failure during roadway excavation include coal seam depth, coal thickness, coal seam dip angle, mining method, roadway length, and the proportion of hard rock in the floor strata. From these numerous influencing factors, four indicators—mining depth, mining thickness, roadway length, and the proportion of hard rock in the floor strata—were selected as the main controlling factors for the prediction model. A total of 56 measured floor failure depth samples from roadway excavation in Shandong coal mines were selected as the model training sample set.

[0025] (2) Sample preprocessing; To eliminate differences in the various metrics, Z-score standardization is used for normalization. For a sequence X with n samples, the standardization formula is: ; In the formula, is the sample mean. σ(X) The standard deviation is the sample standard deviation. The standardized sample sequence is obtained by calculating the above formula. Y, Y for Y ={ y 1 ,y 2 ,…,y n}

[0026] (3) Sample correlation analysis; Forty sets of historical measured data were randomly selected as the calculation sample, and the grey relational analysis was performed using the grey relational analysis theory. The depth of base failure was taken as the reference sequence. X 0, a comparative series consisting of mining depth, mining thickness, roadway length, and the proportion coefficient of hard rock in the floor. X 1. X 2. X 3. X 4. Establish the analysis sequence; perform dimensionless processing on the reference and comparison sequences, and then calculate the grey relational degree according to the following formula: In the formula, r For relevance; ξ The correlation coefficient; n The number of samples; i Number of sub-factors; k For the first k Group samples; |x 0( k ) -x i ( k ) | For sequence X 0 and X i In the k The absolute value of a point; ρ The resolution coefficient has a value range of (0,1), and in this embodiment, the value is 0.5.

[0027] The correlation between the four factors affecting crack zone development and the dependent variable, the depth of base failure, was calculated using the above formula. The factors were then ranked from highest to lowest correlation, and the results are shown in Table 1. Table 1 Results of Grey Relational Analysis The data in the table show that the correlation between the four independent variables and the depth of base plate failure is above 0.75, indicating a good overall correlation. This provides a basis for the rationality of the input factors in the base plate failure depth prediction model.

[0028] Step 2: Construct the PSO-BO-DNN prediction model (1) Model architecture design; The PSO-BO-DNN model for predicting the depth of substrate damage is based on a fundamental DNN model. Bayesian optimization of the fundamental DNN model's hyperparameters is used to obtain the optimal DNN model, and then the PSO optimization method is used to optimize the initial weights of the optimal DNN model to obtain the PSO-BO-DNN model. The fundamental DNN model has three hidden layers (dense layers). To prevent overfitting, a dropout layer is added after each hidden layer. The dropout layer randomly removes a certain proportion of neurons and performs L2 regularization on the coefficient matrix and bias term of each dense layer. The activation function used is LeakyReLU, which effectively avoids gradient vanishing and dead neuron problems compared to traditional ReLU and Sigmoid functions. The main model architecture is as follows: Figure 2 As shown.

[0029] (2) Bayesian optimization is used to adjust hyperparameters and output the optimal DNN model; Bayesian optimization is employed to globally optimize the hyperparameters of the basic DNN model. Bayesian optimization is a probability distribution-based search strategy that uses a prior function to represent the distribution assumptions and a sampling function to determine the sampling location. For parameter selection, a Gaussian process is used as the prior function to reflect the distribution assumptions of the optimization function; the sampling function uses the probability boosting (POI) function, selecting and evaluating the next set of parameters from the posterior distribution. The optimization parameters and their ranges are shown in Table 2. Table 2 Model hyperparameters and optimization range (3) Particle swarm optimization to optimize the initial weights of the DNN model; The initial weights of a DNN model significantly affect the accuracy of its predictions. After obtaining the optimal DNN model optimized by the Boolean algorithm, the PSO algorithm is used to optimize the initial weight matrix of the optimal DNN model. The principle of the PSO algorithm is: each particle represents a point in a 3D space, using... x i ={ x i1 , x i2 , … , x in} indicates the first i The particle, the first i The individual optimal value of each particle is expressed as: pbest i ={ p i1 , p i2 , … , p in}, the set of all optimal values ​​is represented as gbest ={ g 1, g 2, … , g n The particle weight update formula is: In the formula, k Representing the k The next iteration; i = 1 , 2 , … , m ; m The number of particles in the particle swarm; d = 1 , 2 , … , n ; n Let be the dimension of the solution vector; c 1 and c 2 represents the acceleration factor, which consists of two positive constants; rand 1 and rand 2 are two independent random numbers between [0,1]; w i This is the momentum coefficient; adjusting its value can change the strength of the search capability.

[0030] The fitness function is the mean squared error between the prediction results and the measured results of the optimal DNN model trained with particle weights as the initial weight matrix on the validation set. When the convergence condition is met (the optimal DNN model corresponding to all optimal values ​​has R² above 0.8 on both the training set and the validation set), the iteration ends. The current set of all optimal values ​​is output as the optimal initial weight matrix of the DNN, and the model trained with the optimal initial weight matrix of the DNN is output as the final PSO-BO-DNN model.

[0031] The PSO-BO-DNN model construction process is as follows: Figure 1 As shown. Figure 1 The flowchart section needs to correspond to the content of the instruction manual, and the flowchart should include a judgment on whether the convergence condition is met.

[0032] Step 3: Model training and testing; The training samples were divided into training and validation sets in an 8:2 ratio for training the base plate damage depth prediction model. The Bayesian optimization algorithm used had an initial sampling number of 10 and 25 iterations. The optimization process curve is shown below. Figure 3 As shown.

[0033] according to Figure 3The optimal value is reached on the 9th iteration, and the hyperparameter values ​​at this time are shown in Table 3: Table 3 Optimal Hyperparameter Values ​​for the Model Based on the obtained hyperparameters, an optimal DNN model was constructed. The initial number of particles in the PSO optimization algorithm was set to 40, the number of iterations to 100, and the optimization parameters to be the initial weight matrix of the optimal DNN model. The optimal DNN model was trained using training and validation sets. The mean squared error (RMSE) between the prediction results and the measured results on the validation set was used as the fitness function. The PSO optimization fitness curve is shown below. Figure 4 As shown.

[0034] according to Figure 4 The fitness reached its minimum in the 94th round, with a minimum fitness of 0.19289. At this point, the prediction results of the model on the training and validation sets were as follows: Figure 5 and Figure 6 As shown.

[0035] The coefficient of determination (R²) is used to evaluate the model's predictive accuracy, and its formula is as follows: In the formula, y i This represents the true depth of damage to the base plate. This is the predicted value of the base plate failure depth. This represents the average value of the true depth of damage to the base plate.

[0036] The calculated R² values ​​of the model on the training and validation sets are 0.864239432 and 0.818196727, respectively. The trained PSO-BO-DNN model has an R² value of over 0.8 on both the training and validation sets, indicating that the model has a good fit, high prediction accuracy, and good generalization ability, without overfitting.

[0037] Step 4: Prediction of the depth of base plate failure; Data on mining height, coal seam depth, and hard rock ratio of the coal seam floor were collected for the 1016 mining area of ​​Chaili Coal Mine. Based on the mining area design, the roadway length is 200m. The data was input into a trained PSO-BO-DNN floor failure depth prediction model, and the predicted floor failure depth was output.

[0038] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for intelligent prediction of the failure depth of the tunnel floor based on PSO-BO-DNN, characterized in that, Includes the following steps: S1: Data Acquisition and Preprocessing: Collect historical measured data of coal mine roadway excavation, including mining depth, mining thickness, roadway length, hard rock ratio coefficient of the floor, and measured value of floor failure depth as the prediction target; preprocess the historical measured data to form a model training sample set; S2: Constructing the PSO-BO-DNN prediction model: S2.1: Construct a basic DNN model, which includes an input layer, at least one hidden layer, and an output layer; S2.2: The Bayesian optimization algorithm is used to globally optimize the hyperparameters of the basic DNN model to obtain the optimal DNN model with the optimal combination of hyperparameters. S2.3: The initial weight matrix of the optimal DNN model is globally optimized using the particle swarm optimization algorithm to obtain the PSO-BO-DNN prediction model; S3: Model training and validation: The PSO-BO-DNN prediction model is trained using the model training sample set, and the prediction performance of the PSO-BO-DNN prediction model is evaluated using the validation set in the model training sample set. S4: Prediction of floor failure depth: Input the influencing factor data of the roadway to be predicted into the trained PSO-BO-DNN prediction model, and output the corresponding predicted value of floor failure depth. The influencing factor data includes the mining depth, mining thickness, roadway length and floor hard rock ratio coefficient of the predicted roadway.

2. The intelligent prediction method for roadway excavation floor failure depth based on PSO-BO-DNN according to claim 1, characterized in that, Step S1 involves preprocessing the historical measured data, including: using the Z-score standardization method to perform dimensionless processing on the historical measured data, with the following formula: ； In the formula, X is the original sample sequence, Y is the standardized sample sequence, σ is the sample mean, and σ(X) is the sample standard deviation.

3. The intelligent prediction method for the depth of roadway excavation floor failure based on PSO-BO-DNN according to claim 1 or 2, characterized in that, Step S1 also includes: before model training, using grey relational analysis to verify the correlation between the data of each influencing factor and the depth of floor failure, and selecting the factors to be input into the model based on the correlation degree; specifically, let the depth of floor failure be the reference sequence X0, and let the mining depth, mining thickness, roadway length and the hard rock ratio coefficient of the floor form the comparison sequence X1, X2, X3 and X4, calculate the grey relational degree, and when the correlation degree is greater than the preset value, it is determined that the input factors are selected reasonably.

4. The intelligent prediction method for the failure depth of the tunnel floor based on PSO-BO-DNN according to claim 1, characterized in that, In step S2.2, the Bayesian optimization algorithm constructs a surrogate model of the objective function using a Gaussian process as the prior function and the probability of improvement (POI) as the acquisition function. Based on the acquisition function, it selects the next set of hyperparameters to be evaluated and obtains the optimal combination of hyperparameters through iterative optimization.

5. The intelligent prediction method for the failure depth of the tunnel floor based on PSO-BO-DNN according to claim 1, characterized in that, In step S2.2, the hyperparameters include the number of neurons in each Dense layer, the dropout ratio in the Dropout layer, the training batch size, the learning rate, and the L2 regularization coefficient.

6. The intelligent prediction method for the depth of roadway excavation floor failure based on PSO-BO-DNN according to claim 1, characterized in that, In step S3, the model training sample set is divided into a training set and a validation set in an 8:2 ratio; the Adam algorithm is used to dynamically optimize the learning rate.

7. The intelligent prediction method for the failure depth of the tunnel floor based on PSO-BO-DNN according to claim 1, characterized in that, In step S3, the coefficient of determination R² is used to evaluate the prediction accuracy of the model. When the R² of both the training set and the validation set reaches the preset threshold, the model training is considered complete. The formula for calculating the coefficient of determination R² is as follows: ； In the formula, y i is the true value of the base plate damage depth, is the predicted value of the base plate damage depth, and is the average value of the true values ​​of the base plate damage depth; when the R² of both the training set and the validation set reaches a preset value or above, the model training is considered complete.

8. The intelligent prediction method for roadway excavation floor failure depth based on PSO-BO-DNN according to claim 1, characterized in that, In step S2.1, a Dropout layer is set after each hidden layer, and L2 regularization is applied to the weight matrix and bias term of the Dense layer. The activation function is LeakyReLU.

Images