Artificial intelligence-based method for identifying water inrush point in mine and parameters of simulation model

Abstract

The present invention relates to the technical field of coal mine water inrush disaster prevention and control, and relates to an artificial intelligence-based method for identifying a water inrush point in a mine and parameters of a simulation model, comprising the following steps: step 1: establishing a numerical model on the basis of observation data, and determining prior information of parameters to be identified including position coordinates of a water inrush point; step 2: on the basis of the numerical model and the prior information of the parameters, generating a training sample dataset and a test sample dataset for an alternative model; step 3: constructing and training an alternative model neural network; step 4: testing the accuracy of the alternative model; and step 5: executing a simulated annealing algorithm to identify the position of the water inrush point and parameters of a simulation model. By using the method provided by the present invention, the problem in accurate positioning of water inrush points in underground coal mines can be solved, which is of great significance in carrying out scientific disaster management and rescue efforts in a timely manner.

Description

A method for identifying mine water inrush points and simulation model parameters based on artificial intelligence Technical Field

[0001] The present invention belongs to the technical field of coal mine water inrush disaster prevention and control, and in particular relates to a method for identifying mine water inrush locations and model parameters based on artificial intelligence. Background Art

[0002] As coal mining depths continue to increase, hydrogeological conditions become increasingly complex, and coal mine water inrush disasters are a frequent occurrence, causing serious casualties and property losses. Groundwater contained in aquifers is the primary source of mine water inrush. Coal mining activities induce rock strata destruction, forming water channels. This allows aquifer water to surge rapidly into the mine tunnels, leading to serious consequences such as mine flooding, personnel entrapment, and damage to production equipment. To effectively carry out rescue operations for coal mine water inrush disasters and prevent secondary accidents, it is necessary to quickly determine the specific location of the mine water inrush point and to clearly define the key model parameters involved in water inrush simulation and prediction. This allows for accurate analysis of the scope of underground inundation, which is of great guiding significance for the development of rescue and prevention plans.

[0003] Currently, underground workers rely primarily on reporting the location of a water inrush to the dispatch center via underground telephones. However, these reports often provide a regional location without precise coordinates. If a water inrush damages underground communication equipment or traps personnel, preventing them from contacting surface personnel, the water inrush point will be lost. Existing research proposes methods for identifying the source of water inrush by collecting water samples after a water inrush, analyzing the hydrochemical data, and using algorithms such as cluster analysis or support vector machines based on the water quality characteristics of different sources to determine the aquifer from which the water inrush originated. However, these methods cannot pinpoint the coordinates of the water inrush point. Furthermore, the values ​​of the numerical model parameters used in water inrush simulations and predictions directly determine the reliability of the prediction results. However, in most cases, many parameters of groundwater numerical models cannot be directly obtained through existing measurement methods.

[0004] Inversion simulation is a technical approach that uses observational data available in the system to reversely identify model parameters that are difficult to obtain directly, based on the construction of a forward numerical model. Currently, the main approach to implementing inversion simulation is to transform the research problem into an optimization problem and solve it through optimization methods. In this technical field, based on nonlinear optimization theory, the research problem can be transformed into a parameter optimization problem. Artificial intelligence methods combining deep learning and simulated annealing algorithms can be used to achieve efficient and accurate solutions to the optimization problem, thereby simultaneously identifying the location of mine water inrush points and simulation model parameters.

[0005] Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention provides a method for identifying mine water inrush points and simulation model parameters based on artificial intelligence. This method can determine the spatial position coordinates of underground water inrush points and solve the problem of rapid and synchronous identification of underground water inrush point positions and simulation model parameters.

[0007] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a method for identifying water bursting points and simulation model parameters in mines based on artificial intelligence. The overall idea is to establish a groundwater numerical model for the water bursting aquifer, and then use the water level drop change data obtained from the groundwater level observation well of the aquifer known in the early stage as the decision variables together with other unknown parameters such as permeability and water inflow, and establish a nonlinear optimization model. Then, the synchronous identification of other unknown parameters of the model such as the water bursting point position, permeability and water inflow is achieved by solving the nonlinear optimization model based on the simulated annealing algorithm. In this process, in order to ensure the computational efficiency of the optimization process, an alternative model modeling method based on deep learning will also be adopted. The specific implementation of this technical solution includes the following steps:

[0008] Step 1: Establish a numerical model of groundwater flow in the mine water inrush aquifer. Set the water level simulation output point in the model according to the location and observation time of the actual hydrological observation hole. Set the horizontal coordinate x and vertical coordinate y of the water inrush point, the water inrush amount Q and other unknown model parameters (p1, ..., p n , n represents the number of unknown parameters of other models) as decision variables, expressed as: m = [X, Y, Q, p1, ..., p n At the same time, according to the previous data collection, the value range of each decision variable in m is determined, and its upper and lower limits are expressed as: U =[X U ,Y U ,Q U ,p 1U ,…,p nU ] and m L =[X L ,Y L ,Q L ,p 1L ,…,p nL ].

[0009] Step 2: According to the upper limit m of the decision variable determined in step 1 U and the lower limit m L , using the Latin hypercube sampling method, two sets of parameter data sets are randomly sampled as the input parameters of the training sample data set and the input parameters of the test sample dataset Using the groundwater flow numerical model in step 1, calculate M one by one Train and M TestThe water level simulation results at the observation time of the hydrological observation hole corresponding to each model parameter sample in the , and all observation data are stored as vector data format y i Finally, the model response results in the training data set and the test sample data set can be obtained, which are expressed as and In the following description, the training sample data sets are represented as D Train ={M Train ,Y Train} and D Test ={M Test ,Y Test}.

[0010] Step 3: Build a deep convolutional neural network model (DNN), with the input layer and output layer being the model parameter vector m i and the model response vector y i , the model is expressed as Based on the constraints of L1 norm, a deep convolutional neural network is constructed to realize the loss function of the alternative model prediction. Then, with the goal of minimizing the loss function, the θ is updated. DNN To complete the DNN training. At this point, the trained F DNN (m i ,θ DNN ) as an alternative model to the numerical groundwater flow model in step 1.

[0011] Step 4: The test sample dataset D obtained in step 2 Test The input parameter M in Test Substitute into the alternative model F item by item DNN (m i ,θ DNN ), and obtain the corresponding prediction results According to F DNN (m i ,θ DNN ) The convergence loss function value L of the training and the value of Y Test and The calculated coefficient of certainty R2 value determines whether the prediction accuracy of the alternative model meets the requirements. If it does, proceed to step 5; otherwise, return to step 2 to increase the number of samples in the training sample data set.

[0012] Step 5: Substitute the alternative model F obtained in step 4 DNN (m i ,θ DNN) as the equality constraint, the upper and lower limits of the decision variable m in step 1 are used as inequality constraints, and combined with the least squares constraint conditions, a nonlinear optimization inversion model is constructed, which serves as the overall constraint conditions for each decision variable in step 1. Then, the simulated annealing algorithm is used to optimize and solve the decision variable m, and the optimal solution of the decision variable m that meets the constraint conditions of the nonlinear optimization inversion model constructed in this step is found, thereby finally obtaining the position coordinates X and Y of the water inrush point, as well as other key simulation prediction parameters Q and p i .

[0013] The groundwater flow numerical model of the mine water inrush aquifer in step 1 is established using the groundwater numerical simulation program TOUGHREACT. The water inrush volume Q is generalized to a constant value, and the n other unknown model parameters represent the permeability values ​​of n permeability parameter partitions within the simulation area.

[0014] The upper limit m of the decision variable value determined in step 1 in step 2 U and the lower limit m L ,Sampling using the Latin hypercube sampling method follows the sampling principle of uniform distribution.

[0015] n in step 2 Train and n Test The value of n Train >n Test ; and n Test The setting is based on the principle of ≥50.

[0016] The DNN model in step 3 is an improvement on the ResNet-18 deep residual two-dimensional convolutional neural network. First, the input decision variable vector data is mapped using a fully connected neural network to output a 6400-dimensional vector. This vector is then reshaped into an 80×80 rectangular data structure as the input to the ResNet-18. The output layer of the DNN is set to a vector with the same dimension as the observation data y.

[0017] The expression of the loss function for constructing a deep convolutional neural network based on the L1 norm constraint in step 3 to achieve alternative model prediction is as follows:

[0018] Where: θ DNN Represents the weight parameter of the deep neural network; m i and y i Respectively represent the model parameters and model output of the i-th group of samples in the training sample data set; N represents the total number of samples in the training sample data set; w d Represents the regularization term in the neural network training process, which is used to prevent overfitting in training.

[0019] In step 3, the goal is to minimize the loss function (1) DNN The updating process is implemented through the deep learning framework pytorch.

[0020] The coefficient of certainty R in step 4 2 The calculation formula is as follows:

[0021] Where, Represents all y Train(i) The mean of .

[0022] The smaller the value of L in step 4, the greater the 2 The closer the value is to 1, the more likely the alternative model F is. DNN (m i ,θ DNN ) is more accurate. In this technology, the thresholds (L0 and ); then judge L≤L0, and Whether the accuracy requirements are met.

[0023] The basic form of the inverse nonlinear optimization model in step 5 is as follows:

[0024] Where: F represents the objective function based on the least squares constraint; y obs Represents the observation data vector; y obs [i] represents the i-th variable element in the observation data vector; m L and m U Respectively represent the upper limit vector and lower limit vector of the model parameter vector m; N obs Indicates the number of observations.

[0025] The simulated annealing algorithm in step 5 is implemented through the following steps:

[0026] Step 5-1: Set the hyperparameters of the simulated degradation algorithm, the initial iteration temperature T0, and the initial solution m of the decision variable m i ;

[0027] Step 5-2: In m i Randomly generate a new solution m near j , specifically by i Multiply by a random perturbation coefficient e(m j =e×m i ). Among them, e is based on the Gaussian distribution N~(1,σ 2 ) is a randomly generated number with the same dimension as m. σ defaults to 0.01 and can be adjusted in different application scenarios within a range of 0 to 0.1.

[0028] Step 5-3: Set m i and m j Substitute them into formula (3) respectively to calculate m i and m j The corresponding inversion optimization objective function value: F i and F j .

[0029] Step 5-4: Determine if F i ≥F j If it holds, then the current solution m i Update to m j Otherwise, calculate m according to the following formula i Update to m j Probability of:

[0030] Where: a represents the attenuation coefficient in the simulated annealing algorithm, which is 0.99 here; t represents the current time, and represents the current number of loop iterations.

[0031] In a specific computer program, the probability selection in formula (4) is determined by generating a random number rand(x) between 0 and 1. When rand(x)≤P(m i →m j ), then m i Update to m j ; otherwise, no update.

[0032] Step 5-5: Repeat steps 5-2 to 5-4 under the current temperature conditions until the preset number of simulated annealing inner loop iterations is reached. Then update the temperature and time respectively: t = t + 1 and T t =a t T0, then proceed to the next step.

[0033] Step 5-6: Return to step 5-2 and calculate the T value according to the T value obtained in step 5-5. t and t are updated until the preset number of outer loop iterations are completed.

[0034] Beneficial Effects: Based on groundwater inversion theory, this paper proposes an artificial intelligence-based method for identifying mine water inrush points and simulation model parameters. This method combines a deep learning-based surrogate model approach with a simulated annealing-based parameter optimization strategy. This method leverages on-site water level monitoring data to simultaneously identify the specific locations of water inrush points and key simulation model parameters, which are difficult to obtain directly. This method provides technical support for mine water inrush disaster rescue and accurate simulation of the disaster's impact area. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The present invention will be further described below with reference to the accompanying drawings and implementation examples:

[0036] FIG1 is a flow chart of a method for identifying a mine water inrush point and simulation model parameters based on artificial intelligence according to the present invention;

[0037] FIG2 is a schematic diagram of a simulation area model according to an embodiment of the present invention;

[0038] FIG3 is a comparison diagram of the water level difference simulation results of the correction model of the present invention and the actual observation results;

[0039] FIG4 is a schematic diagram of a DNN model structure generated based on the ResNet-18 model result of the present invention. DETAILED DESCRIPTION

[0040] The method of the present invention is described in detail below with reference to specific examples.

[0041] As shown in FIG1 , the method for identifying water inrush points and simulation model parameters in a mine based on artificial intelligence of the present invention comprises the following steps:

[0042] Step 1: Based on the basic hydrogeological data of the study mining area, a groundwater flow numerical model of the mine water inrush aquifer in the study area is established. The groundwater flow numerical model of the mine water inrush aquifer is equipped with hydrological observation holes to study the changes in water level drop in the study area. The hydrological observation holes are defined as water level observation points. Based on the mine excavation engineering plan, the coordinate range of the water inrush points, the water inrush volume, and the priori interval value range of other unknown model parameters are preliminarily determined in the numerical model of the coal seam water inrush aquifer.

[0043] The unknown abscissa x and ordinate y of the water inrush point, the water inrush amount Q and n unknown parameters are taken as decision variables, which can be expressed as follows: the total number of decision variables m = [X, Y, Q, p1,…, p n ],p1,…,p n Represents n unknown parameters in the groundwater model, including permeability parameters and boundary condition parameters of different partitions; at the same time, based on the collected preliminary data, the value range of each decision variable in m is determined, and the upper and lower limits of the decision variables are expressed as: m U =[X U ,Y U ,Q U ,p 1U ,…,p nU ] and m L =[X L ,Y L ,Q L ,p 1L ,…,p nL ];

[0044] Step 2: According to the upper limit m of the determined decision variable U and the lower limit m L , using the Latin hypercube sampling method, the groundwater flow numerical model is randomly sampled to obtain two sets of parameter data sets, which are used as the input parameters of the training sample data set and the input parameters of the test sample dataset n Train Represents the training sample dataset M Train Sample, n Test Represents the test sample dataset M Test samples;

[0045] Calculate M one by one using the groundwater flow numerical model Train and M Test The water level simulation results at the observation time of the hydrological observation hole corresponding to each model parameter sample in the , and all observation data are stored as vector data format y i ; Finally, the model response results of the training data set can be obtained And the model response results of the test sample dataset The training sample data sets are represented as D Train ={M Train ,Y Train} and D Test ={M Test ,Y Test};

[0046] The upper limit m of the decision variable value U and the lower limit m L , using Latin hypercube sampling method to perform sampling, following the sampling principle of uniform distribution; n Train The number of samples n Train >n Test The number of samples, and satisfy n Test ≥50.

[0047] Step 3: Construct a deep convolutional neural network model DNN. The input layer and output layer of DNN are the groundwater numerical model parameter vector m i and the model response vector y i , the DNN model is represented as θ DNN Represents the weight parameters of the deep neural network; based on the constraints of the L1 norm, a deep convolutional neural network is constructed to realize the loss function of the alternative model prediction of the groundwater numerical model, and then the error back propagation algorithm is used to update θ with the goal of minimizing the loss function. DNN To complete the DNN training; at this time, the trained DNN model F DNN (m i ,θDNN ) as an alternative model to the groundwater flow numerical model in step 1;

[0048] The calculation formula for the loss function of constructing a deep convolutional neural network to achieve alternative model prediction under the constraint of L1 norm is as follows:

[0049] Where: θ DNN Represents the weight parameter of the deep neural network; m i and y i Respectively represent the model parameters and model output of the i-th group of samples in the training sample data set; N represents the total number of samples in the training sample data set; w d Represents the regularization term in the neural network training process, which is used to prevent overfitting in training.

[0050] As shown in Figure 4, a schematic diagram of the DNN model structure shows how to establish an alternative model based on the results of the ResNet-18 model: The DNN model is improved based on the deep residual two-dimensional convolutional neural network of ResNet-18. First, the vector data of the input decision variable is mapped into a 6400-dimensional vector using a fully connected neural network, and then reshaped into an 80×80 rectangular data structure as the input layer of ResNet-18. The output layer is a vector with a dimension consistent with the observation data y. The rest are the original structure of ResNet-18.

[0051] Step 4: The test sample dataset D obtained in step 2 Test The input parameter M in Test Substitute into the alternative model F item by item DNN (m i ,θ DNN ), and obtain the corresponding prediction results According to F DNN (m i ,θ DNN ) The convergence loss function value L of the training and the value of Y Test and The calculated coefficient of certainty R 2 The value determines whether the prediction accuracy of the alternative model meets the requirements. If it does, proceed to step 5; otherwise, return to step 2 to increase the number of samples in the training sample data set; the smaller the value of the convergence loss function value L, and the higher the certainty coefficient R 2 The closer the value is to 1, the more likely the alternative model F is. DNN (m i ,θ DNN ) the higher the prediction accuracy. In this technology, the threshold value L0 of the loss function and the threshold value of the certainty coefficient are pre-set. Then judge L≤L0, and Whether the accuracy requirements are met.

[0052] Determination coefficient R 2 The calculation formula is as follows:

[0053] Where, Represents all y Train(i) The mean of .

[0054] Step 5: Substitute the alternative model F that meets the accuracy requirements in step 4 DNN (m i ,θ DNN ) as an equality constraint, and set the upper limit m of the decision variable population m in step 1 U and the lower limit m L As inequality constraints, combined with the least squares constraints, a nonlinear optimization inversion model is constructed, which is used as the decision variable population m=[X,Y,Q,p1,…,p n ] constraints, and then use the simulated annealing algorithm to optimize and solve the overall decision variable m, and find the optimal solution of the overall decision variable m that meets the constraints of the nonlinear optimization inversion model constructed in this step, so as to finally obtain the position coordinates X and Y of the water inrush point, as well as other simulation prediction key parameters Q and p1,…,p n .

[0055] The basic form of the nonlinear optimization inversion model is as follows:

[0056] Where: F represents the objective function based on the least squares constraint; y obs Represents the observation data vector; y obs [i] represents the i-th variable element in the observation data vector; m L and m U Respectively represent the upper limit vector and lower limit vector of the model parameter vector m; N obs Indicates the number of observations.

[0057] The implementation of the simulated annealing algorithm is as follows:

[0058] Step 5-1: Set the hyperparameters of the simulated annealing algorithm, the initial iteration temperature T0, and the initial solution m of the decision variable m. i ;

[0059] Step 5-2: In m i Randomly generate a new solution m near j , specifically by i Multiply by a random perturbation coefficient e(m j =e×m i), where e is based on the Gaussian distribution N~(1,σ 2 ) Randomly generated random number with the same dimension as m. The default value of σ is 0.01 and can be adjusted in different application scenarios within a range of 0 to 0.1.

[0060] Step 5-3: Set m i and m j Substitute them into formula (3) to calculate and get m i and m j The corresponding inversion optimization objective function value: F i and F j ;

[0061] Step 5-4: Determine if F i ≥F j If it holds, then the current solution m i Update to m j Otherwise, calculate m according to the following formula i Update to m j Probability of:

[0062] Where: a represents the attenuation coefficient in the simulated annealing algorithm, which is 0.99; t represents the current time, which represents the number of current loop iterations; T0 represents the temperature at the initial iteration time, which is 100 by default; the probability selection in formula (4) is judged by generating a random number rand(x) between 0 and 1. When rand(x)≤P(m i →m j ), then m i Update to m j ; Otherwise, do not update;

[0063] Step 5-5: Repeat steps 5-2 to 5-4 under the current temperature conditions until the preset number of simulated annealing inner loop iterations is reached; then update the temperature and time respectively: t = t + 1 and T t =a t T0;

[0064] Step 5-6: Return to step 5-2 and calculate the T value according to the T value obtained in step 5-5. t and t are updated until the preset number of outer loop iterations are completed.

[0065] Example

[0066] Construct a coal mine water inrush scenario. Now that the specific water inrush aquifer has been identified, the specific water inrush location needs to be further determined. A two-dimensional groundwater flow model is obtained using TOUGHREACT modeling. The model range is 10,000m×10,000m, with the east and west boundaries assumed to be equal constant head boundaries and the north and south boundaries as zero flow boundaries. There are two known water inrush points in the area, with water inflows of 72m at point I1 and 72m at point I2. 3 / h and 12.54m 3 / h. Assume that a water inrush disaster occurs at a certain working face, but the location of the water inrush point is unknown, the water inrush time is when the model runs for 360 days, and the water inrush volume is 720m 3 / h (point I3). There are 10 known water level change observation wells in the study area (#1~#10). During the TOUGHREACT numerical calculation process, the entire area in the model is divided into 80×80 discrete grids. Among them, the middle 3000m×3000m range is refined using a 60×60 grid. Assume that the permeability parameter partitions in the model are determined to be three. According to the water inrush scenario, there are 6 parameters that need to be identified, namely the horizontal coordinate X of the water inrush point location, the vertical coordinate Y of the water inrush location, the water inrush volume (Q) and the permeability of the three partitions (k1, k2, k3). In this case, k1, k2 and k3 correspond to other model parameters except X, Y and Q, corresponding to p1~p3 in step 1. The specific information of the above specific model is shown in Figure 2.

[0067] To test the feasibility of this patented technology, a two-year simulation was conducted to obtain bimonthly water level change observations from 10 observation points. This data was then used to generate the actual water level observations for this hypothetical case, adding a Gaussian-distributed noise perturbation of N(1, 0.01). Based on this observational data, this patented technology was then used to invert and identify six unknown model parameters, including the location of the water inrush point.

[0068] The prior interval value ranges of the six parameters introduced in step 1 are shown in Table 1.

[0069] The number of samples in the training sample dataset and the test sample dataset in step 2 are 300 and 50 respectively.

[0070] Alternative Model Prediction Accuracy Metrics from Step 4: Loss Function and R 2 The values ​​are 0.0066 and 0.9918 respectively. In order to further improve the accuracy of the alternative model, return to step 2 and increase the training sample data set to 500. The alternative model is retrained, and the loss function and R 2 The values ​​are increased to 0.0040 and 0.9968 respectively. At this point, the prediction accuracy has met the requirements and the subsequent steps are performed.

[0071] The key parameters in the simulated annealing algorithm implementation process in step 5 are set as follows:

[0072] T0 in step 5-1 = 100;

[0073] The temperature decay constant a in steps 5-4 and 5-5 is 0.99;

[0074] The number of inner loops in step 5-5 is 150;

[0075] The number of times of the outer loop in steps 5-6 is 300.

[0076] The inversion identification results of the six identification parameters obtained by this patented technology and their relative errors to the true values ​​are shown in Table 1.

[0077] Table 1 True values, prior intervals, inversion estimates and relative errors of the parameters to be inverted

[0078] As can be seen from the table, the relative errors of the X and Y coordinates of the water inrush point are both within 0.04. The error range of the X coordinate value has been reduced from the original 850m (4875-5725m) to approximately 20m (5625m-5604.44m); the error range of the Y coordinate value has been reduced from the original 150m (4875-5025m) to within 10m (4975m-4966.11m).

[0079] In addition, except for the k2 identification result with a slightly larger relative error (0.295), the other relative errors are all within 0.05. -13 m 2 Compared with the actual value of 5.097×10 -13 m 2 Figure 3 shows a comparison of the water level difference simulation results from the correction model and the actual observation data, corresponding to measurement points #1-#10 in Figure 2 (the dots in the figure represent the actual observation data, and the lines represent the simulation curves after model correction). The fit between the correction model simulation results and the observations in Figure 3 shows that the two-year observation data at the 10 observation points closely matches the corrected simulation curves. Based on the above information, the values ​​of Q, k1, k2, and k3 obtained by the inversion solution are all close to their true values, and the numerical model is well fitted and corrected, making their inversion results reliable.

[0080] It can be seen that the artificial intelligence-based mine water inrush point and simulation model parameter identification method provided by this patent can determine the specific location coordinates of the water inrush point, realize the precise positioning of the water inrush point underground, and can simultaneously identify the water inrush volume and permeability parameter values, providing key information for the prevention and control of water inrush disasters.

Claims

1. A method for identifying mine water inrush points and simulation model parameters based on artificial intelligence, characterized in that The steps are as follows: Step 1: Collect the preliminary data of the mine, establish a numerical model of groundwater flow in the water inrush aquifer of the mine, and set the water level simulation results in the numerical model of groundwater flow to output the water inrush points according to the positions and observation times of the actual hydrogeological observation wells; The abscissa x and ordinate y of the water inrush point, the water inrush volume Q, and n unknown parameters are taken as decision variables, expressed as: the overall decision variable m = [X, Y, Q, p1, …, p n , p1, …, p n represent n unknown parameters in the groundwater model, including permeability parameters and boundary condition parameters in different zones; at the same time, according to the collected preliminary data, the value ranges of each decision variable in m are determined, and the upper and lower limits of the decision variable values are respectively expressed as: m U = [X U , Y U , Q U , p 1U , …, p nU and m L = [X L , Y L , Q L , p 1L , …, p nL ; Step 2: According to the determined upper limit m of the decision variable value U and the lower limit m L , using the Latin hypercube sampling method, randomly sample the groundwater flow numerical model to obtain two groups of parameter data sets, and use the two groups of parameter data sets as the input parameters of the training sample data set and the input parameters of the test sample dataset n Train represents a sample in the training sample data set M Train , and n Test represents a sample in the test sample data set M Test ; Use the groundwater flow numerical model to calculate M one by one Train and M Test The simulated water level results at the observation time at the location of each hydrological observation well corresponding to the model parameter samples in and M are calculated, and all the observation data are stored in the vector data format y i ; Finally, the model response results of the training dataset can be obtained and the model response results of the test sample dataset Let the training sample data sets be represented as D Train ={M Train , Y Train} and D Test ={M Test , Y Test}; Step 3: Construct a deep convolutional neural network model DNN. The input layer and output layer of the DNN are the groundwater numerical model parameter vector m i and the model response vector y i respectively. The DNN model is expressed as θ DNN denotes the weight parameters of the deep neural network; construct a loss function for predicting the surrogate model of the groundwater numerical model by building a deep convolutional neural network based on the L1-norm constraint condition, and then, with the goal of minimizing the loss function, update θ through the error backpropagation algorithm DNN to complete the training of the DNN; at this time, the trained DNN model F DNN (m i , θ DNN ) is used as the surrogate model of the groundwater flow numerical model in step 1; Step 4: Substitute the input parameter M in the test sample data set D obtained in Step 2 Test into the surrogate model F Test one by one DNN (m i , θ DNN ) to obtain the corresponding prediction results According to F DNN (m i , θ DNN ) the convergence loss function value L trained and according to Y Test and Calculated certainty coefficient R 2 Determine whether the prediction accuracy of the value judgment substitution model meets the requirements. If it meets the requirements, continue to execute step 5; otherwise, return to step 2 to increase the number of samples in the training sample dataset. Step 5: Take the alternative model F DNN (m i , θ DNN ) that meets the requirements in Step 4 as an equality constraint, and take the upper limit m U and the lower limit m L of the value of the overall decision variable m in Step 1 as inequality constraints. Combine the least squares constraint conditions to construct a non-linear optimization inversion model, and use this as the constraint condition for each overall decision variable m = [X, Y, Q, p1, …, p n in Step 1. Then, the simulated annealing algorithm is used to optimize and solve the overall decision variable m, and the optimal solution of the overall decision variable m under the constraint conditions of the non-linear optimization inversion model constructed in this step is found, so as to finally obtain the position coordinates X and Y of the water inrush point, as well as other key simulation prediction parameters Q and p1, …, p n .

2. The method for identifying mine water inrush points and simulation model parameters based on artificial intelligence according to claim 1, wherein The numerical model of the coal seam water inrush aquifer in Step 1 is established by using the groundwater numerical simulation software TOUGHREACT.

3. A method for identifying mine water inrush points and simulation model parameters based on artificial intelligence according to claim 1, characterized in that The upper limit m of the decision variable value determined in step 1 in step 2 U and the lower limit m L , sampling is performed using the Latin hypercube sampling method, following the sampling principle of uniform distribution; n Train The number of samples n Train > n Test The number of samples, and satisfying n Test ≥ 50 4. The method for identifying mine water inrush points and simulation model parameters based on artificial intelligence according to claim 1, wherein The DNN model in Step 3 is improved based on the depth residual two-dimensional convolutional neural network of ResNet-18; First, the vector data of the input decision variables are mapped by a fully connected neural network to output a vector of 6400 dimensions, and then it is reshaped into a rectangular data structure of 80×80 as the input of ResNet-18, and the output layer is a vector with the same dimension as the observed data y.

5. A method for identifying mine water inrush points and simulation model parameters based on artificial intelligence according to claim 1, characterized in that, The calculation formula of the loss function for implementing surrogate model prediction by constructing a deep convolutional neural network based on the L1-norm constraint condition in step 3 is as follows: where: θ DNN represents the weight parameters of the deep neural network; m i and y i respectively represent the model parameters and model outputs of the i-th group of samples in the training sample dataset; N represents the total number of samples in the training sample dataset; w d represents the regularization term in the neural network training process, which is used to prevent overfitting during training.

6. The method for identifying mine water inrush points and simulation model parameters based on artificial intelligence according to claim 5, wherein , in step 3, θ is updated with the objective of minimizing the loss function represented by formula (1). DNN This process is implemented through the deep learning framework pytorch.

7. A method for identifying mine water inrush points and simulation model parameters based on artificial intelligence according to claim 1, characterized in that , the certainty coefficient R in the said step 4 2 is calculated as follows: In the formula, Denote the mean of all y Train(i) .

8. A method for identifying mine water inrush points and simulation model parameters based on artificial intelligence according to claim 7, characterized in that , the smaller the value of the convergence loss function value L in step 4, and the closer the value of the certainty coefficient R 2 is to 1, it indicates that the surrogate model F DNN (m i , θ DNN ) has higher prediction accuracy. In this technique, a threshold L0 of the loss function and a threshold of the certainty coefficient are preset in advance , and then by judging whether L≤L0 and meet the accuracy requirements.

9. The method for identifying mine water inrush points and simulation model parameters based on artificial intelligence according to claim 1, characterized in that , The basic form of the non-linear optimization inversion model in step 5 is as follows: where: F represents the objective function based on the least squares constraint; y obs represents the observed data vector; y obs [i] represents the i-th variable element in the observed data vector; m L and m U respectively represent the upper bound vector and the lower bound vector of the model parameter vector m; N obs represents the number of observed data.

10. The method for identifying the water inrush point and simulation model parameters in a mine based on artificial intelligence according to claim 9, wherein , The implementation of the simulated annealing algorithm in Step 5 is carried out through the following steps: Step 5-1: Set the initial iteration temperature T0 of the simulated annealing algorithm and the initial solution m of the decision variable m i ; Step 5-2: Randomly generate a new solution m i near m j , specifically by multiplying m i by a randomly perturbed coefficient e (m j = e × m i ), where e is a random number with the same dimension as m randomly generated according to the Gaussian distribution N~(1, σ 2 ), the default value of σ is 0.01, which can be adjusted in different application scenarios, and the adjustment range is 0 to 0.1; Step 5-3: Substitute m i and m j into formula (3) for calculation respectively to obtain the inversion optimization objective function values corresponding to m i and m j : F i and F j ; Step 5-4: Determine that if F i ≥F j holds, then update the current solution m i to m j ; otherwise, calculate the probability of updating m i to m j according to the following formula: Where: a represents the attenuation coefficient in the simulated annealing algorithm, with a value of 0.99; t represents the current time, representing the current number of loop iterations; T0 represents the temperature at the initial iteration moment, with a default value of 100; the probability selection in formula (4) is judged by generating a random number rand(x) between 0 and 1. When rand(x) ≤ P(m i →m j ), then m i is updated to m j ; otherwise, it is not updated; Step 5-5: Under the current temperature condition, repeat Steps 5-2 to 5-4 until the preset number of inner-loop iterations of simulated annealing is reached; then update the temperature and time respectively: t = t + 1 and T t = a t T0; Step 5-6: Return to Step 5-2, and update according to T t and t obtained in Step 5-5 until the preset number of outer loop iterations is completed.

Images