A method for reducing rock burst dynamic disaster induced by tunnel blasting

Abstract

The application discloses a method, system and storage medium for reducing rock burst dynamic disasters induced by tunnel blasting. The application is based on working conditions and specification data, adopts machine learning and layered execution meta-heuristic optimization algorithm to find a target solution for reducing rock burst dynamic disasters induced by tunnel blasting. On the basis of selecting a machine learning model with a combination of geological variables and blasting variables as prediction variables, model tuning is divided into two layers: the first layer optimizes model hyperparameters. The optimal hyperparameter combination is iteratively searched within the hyperparameter search space to obtain the evolved optimal prediction model, which is used to guide the design of the tunnel blasting scheme. The second layer optimizes the blasting design variables. The prediction model is used to guide the selection of blasting design variables, and the same algorithm logic as the first layer optimization is used, and an improved preferred solution is also included. The application provides a new solution for the field of tunnel construction, can improve construction efficiency, reduce risks, and ensure the quality and sustainability of tunnel projects.

Description

Technical Field

[0001] This invention relates to an engineering scheme design technology, and in particular to a design technology for reducing induced rockburst dynamic disasters in tunnel blasting schemes, belonging to the fields of electrical mathematical data processing and tunnel blasting engineering technology. Background Technology

[0002] Drilling and blasting is the most widely used rock-breaking method in rock tunnel engineering. It involves drilling holes at designated blasting sites on the tunnel face and filling them with explosives. The shock waves generated by the multi-site explosions and the energy released by the explosives break the rock mass within the tunnel area. However, the utilization rate of explosive energy in drilling and blasting is limited; more than 50% of the explosive energy is lost during propagation. This lost energy causes impact damage to the surrounding environment, becoming a source of safety hazards associated with drilling and blasting. In deep-buried tunnel engineering, the adverse vibrations caused by the loss of explosive energy destabilize the surrounding hard and brittle rock mass, easily inducing a chain reaction of deep underground rockburst dynamic disasters. Therefore, in tunnel engineering projects, efforts to reduce or even eliminate reliance on experience, scientifically predict blasting vibrations, and use key blasting parameters as critical control points for designing and optimizing blasting schemes, along with the optimal selection of key control point combinations using modern and even intelligent data processing methods, to reduce the associated safety hazards caused by the impact damage of blasting vibration waves to the surrounding rock mass, have always been important directions and goals for the improvement of tunnel drilling and blasting engineering technology. In summary, the seismic waves generated by tunnel blasting are a major hazard factor affecting the surrounding rock mass and potentially leading to rockbursts. Their impact is extremely widespread and the consequences severe. Therefore, implementing appropriate control measures and designing blasting parameters are crucial to reducing these harmful vibrations.

[0003] CN2023100658227 discloses an intelligent design method for tunnel blasting schemes. This invention obtains blasting design parameters based on tunnel engineering geological survey reports and advanced geological predictions of the tunnel face. Based on blasting design parameter set 01, tunnel engineering blasting theory and computer deep learning technology are used to obtain drill-and-blast tunnel blasting scheme 01. Based on blasting design parameter set 02, computer deep learning technology is used to establish a sample library of existing drill-and-blast tunnel blasting schemes that have been built or are under construction. Inputting blasting design parameters and tunnel engineering requirements, a suitable drill-and-blast tunnel blasting scheme 02 is obtained. Based on the obtained drill-and-blast tunnel blasting scheme group, a final drill-and-blast tunnel blasting scheme is selected manually according to the principles of economy, efficiency, safety, and simplicity. The main shortcomings of this method are: firstly, the model training process does not involve specialized optimization calculations for adjusting the optimal hyperparameter combination, thus the overall performance of machine learning has room for improvement; secondly, the use of a neural network model cannot obtain the ranking results of the importance of sample features. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a method to reduce rockburst dynamic disasters induced by tunnel blasting. This method can improve the overall performance of machine training in tunnel blasting scheme design using machine learning through hyperparameter tuning optimization.

[0005] To achieve the above objectives, the present invention first provides a method for reducing rockburst dynamic disasters induced by tunnel blasting, the technical solution of which is as follows.

[0006] A method for reducing rockburst dynamic hazards induced by tunnel blasting, characterized by: designing a safe blasting target scheme for tunnels using blasting condition data;

[0007] Step S100: Investigate the geological information of the tunnel face at the current blasting site. Based on the tunnel blasting engineering specifications, determine the combination of geological condition variables Xg and the combination of blasting design variables Xd at the current tunnel face. Xd has M blasting design variables xd. i Each has its own range of values ​​[xd] i1 ,xd i2 ], constituting the design space Λ Xd ;

[0008] Step S200: Using existing tunnel blasting case data, construct a machine learning sample set SS. SS takes each tunnel blasting case as a sample, the set of blasting design variable Xd and geological condition variable combination Xg as the sample feature X, and the rock burst dynamic evaluation variable Y as the sample label; it is divided into a training set and a test set.

[0009] Step S300: Determine the machine learning prediction model (MLM) with X as input and Y as output.

[0010] Determine the proposed hyperparameter combination HP for tuning in MLM. HP has N hyperparameter terms hp j ,

[0011] Each has a range of values ​​[hp] j1 ,hp j2 ], constituting HP's search space Λ HP ;

[0012] Step S400: On the training set, in Λ HP A feasible solution for HP is randomly generated within the internal domain, and the hyperparameter combination HP is completed using the metaheuristic optimization algorithm model IPA in Λ. HP Internal optimization and training generate a tunnel blasting scheme design prediction model. This tunnel blasting scheme design prediction model is based on HP's performance in Λ... HP A model that minimizes the internal generalization error;

[0013] Step S500: On the training set, in Λ XdThe feasible solution of the blasting scheme is designed by randomly generating internal blasting schemes. Based on the optimal MLM, the rockburst dynamic evaluation Y value of each feasible solution is predicted. The optimal solution of the combination of blasting design variables Xd is obtained by using the metaheuristic optimization algorithm model IPA, that is, the target blasting scheme of the tunnel is obtained.

[0014] The present invention's method for reducing rockburst dynamic disasters induced by tunnel blasting involves using machine learning to construct a prediction model (MLM) for tunnel blasting schemes, with geological condition variables and blasting scheme design variables as prediction variables. This MLM serves as the initial prediction model. Based on this, the metaheuristic algorithm (IPA) is introduced to intelligently optimize the initial model. The core concept of optimization is two-layer execution: The first layer optimizes some hyperparameters (HP) of the initial model. The optimal hyperparameter combination after iteration is found within the hyperparameter search space, and the initial prediction model is used to evolve into the optimal prediction model, i.e., a prediction model that can guide tunnel blasting scheme design. The second layer optimizes the blasting design variables in the blasting scheme. The predicted variables of the trained MLM include two categories: Xg and Xd. Since Xg is a measured value and cannot be selected in the design scheme (it can also be understood that its upper and lower boundaries are equal to the measured value), it does not participate in the second-layer optimization process to save computational power. Therefore, the second-layer optimization is performed within the design space Λ of Xd. Xd It will be carried out inside. In Λ Xd Internally, a predictive model guides the selection of blasting design variables. The same algorithmic logic as the first-level optimization scheme is used to find the optimal combination of blasting design variables after iteration, which is the tunnel blasting target scheme of this technical solution. The technical advantage of using the same algorithmic logic for both levels of optimization is that: firstly, both problems are driven by the same approach—finding the maximum or minimum value of a function. Secondly, using the same algorithmic logic leverages the powerful capabilities of the so-called greedy algorithm, which can yield a perfect solution. Furthermore, it greatly simplifies the operation.

[0015] The above-mentioned method for reducing rockburst dynamic disasters induced by tunnel blasting involves, during the first layer of optimization, using a fitness value L to evaluate the quality of each feasible solution, and then employing a greedy algorithm to find the optimal solution, thereby obtaining the HP in Λ HP A model that minimizes the internal generalization error. The specific fitness function is expressed according to Equation 1.

[0016]

[0017] In the formula, L-HP represents the fitness value of an individual in a feasible solution, o i - Original Y-value data for the case study, p i - MLM prediction of Y values, K - number of predictor variables in the training set.

[0018] The following example, using the first-level optimization, illustrates the concept of the same metaheuristic optimization algorithm model, IPA, used in both levels of optimization.

[0019] First, in Λ HP Internal random adjustment of HP within each HP i The system generates a certain number of feasible HP solutions, called HP individuals, and all HP individuals constitute the HP population. The system parameters are then initialized.

[0020] Secondly, initial roles are assigned to the HP population, with different roles corresponding to different tags. The first HP individual in the HP population is assigned the Ld tag, the individual closest to the Ld tag is assigned the Fd tag, and the remaining individuals are assigned the FL tag. The position of the Fd tag is denoted as I. * (a), and all other marked positions are denoted as I(a);

[0021] Next, for the current HP population, calculate the fitness value L for each HP individual, and obtain the HP individual with the smallest L value L. min Check L min Individual labeling is achieved by either exchanging labels with individuals labeled with Fd or retaining their own Fd labels to ensure L min Individuals are assigned Fd tags, and I is updated. * (a);

[0022] Next, the Ld-labeled individual first updates its position I(a) according to the update strategy, which is expressed by equations 2 to 5.

[0023] B c =|C·I * (a)-I(a)| Formula 2

[0024] I(a+1)=I * (a)-A·B c Formula 3

[0025] A = 2d·sd Equation 4

[0026] C = 2·s Equation 5

[0027] Where a – the current iteration number; I * (a) – The current optimal feasible solution, corresponding to the current optimal position of HP; I(a) and I(a+1) – The positions of the current and next iterations, respectively; B c - Movement step size; C, A - coefficients; d - Ld marks the individual position and updates the path parameters, which decreases linearly from 2 to 0 according to the number of iterations; s - a random number between (0,1].

[0028] Furthermore, any Fl-marked individual updates its position I(a) according to the update strategy, which is expressed by equations 6 to 8.

[0029] I(a+1)=I * (a)+B c ·ebr ·cos(2πr) Equation 6

[0030] B c =|C·randI(a)-I(a)| Equation 7

[0031] I(a+1)=randI(a)-A·B c Formula 8

[0032] In the formula, b - Fl marks the path parameter for updating the individual's location, with a value of 1; r - a random variable in the range [-1, 1]; e - a natural constant; randI(a) - a randomly selected location;

[0033] Once all Ld-marked and Fl-marked individuals have completed their position updates, the updated HP population is obtained;

[0034] Finally, for the current HP population, a new round of Fd, Ld, and Fl label position updates is started until the preset iteration conditions are met to complete the training and obtain the optimal MLM.

[0035] To enhance the search benefits of the second-layer (for blasting design variables) optimization, the second-layer optimization of the method for reducing rockburst dynamic disasters induced by tunnel blasting described above in this invention can be optimized as follows.

[0036] Optimization 1: The update path parameter b for the Fl-marked individuals is further restricted, changing from a fixed movement method (i.e., a constant value) to one that varies according to the iteration number a. In the early stages of iteration, a large search range is performed, while in the later stages, the search range is gradually narrowed to avoid getting trapped in local optima. This is specifically expressed by Equations 9 and 10.

[0037]

[0038]

[0039] In the formula, c∈[5,20]; a max - Maximum number of iterations, k - k-th design variable xd i u and l- are the k-th design variable xd respectively. k Upper limit of value xd k2 With lower limit xd k1 .

[0040] Optimization 2: Starting from the algorithm's a = 1 iteration, in each I... * (a) After the update, in order to improve I * (a) Nearby search gains, increasing the domain perturbation expressed by Equations 11 and 12, will I * (a) Updated to I g (a),

[0041]

[0042]

[0043] In the formula, I g (a) – The position of the Fd marker Xd after the domain perturbation; rand – A random number between 0 and 1; f(·) – The tunnel blasting scheme design prediction model obtained in step S400; rand(-r,r) – The current design variable xd i The range of values ​​for [xd] i1 ,xd i2 Random numbers within ]

[0044] The above design method, when determining which hyperparameters to combine for optimization in the first layer of tuning of the initial model, generally considers improving the model's generalization ability. Greedy algorithms perform exceptionally well in the optimization process, making the model's predictive power stronger than traditional models. The second layer of tuning, based on the obtained predictive model, uses the predictive model in conjunction with the greedy algorithm to find the optimal blasting design parameters. Unlike traditional methods, relying solely on the predictive model to adjust blasting design parameters is highly inefficient, while the problem of blasting parameter optimization perfectly aligns with the characteristics of greedy algorithm optimization. Therefore, using a greedy algorithm can optimize blasting design parameters very efficiently. Furthermore, the addition of a step that controls the search range based on the number of iterations allows for a very efficient and rapid finding of the optimal combination of blasting parameters.

[0045] The present invention also provides a system for reducing rockburst dynamic disasters induced by tunnel blasting.

[0046] A system for reducing rockburst dynamic hazards induced by tunnel blasting, which utilizes blasting condition data to design a target scheme for reducing rockburst dynamic hazards induced by tunnel blasting, is characterized by comprising:

[0047] The design space acquisition module is configured as follows: Based on the current geological information of the tunnel face and in conjunction with tunnel blasting engineering specifications, determine the combination of geological condition variables Xg and the combination of blasting design variables Xd, where Xd has M blasting design variables xd. i Each has its own range of values ​​[xd] i1 ,xd i2 ], constituting the design space Λ Xd ;

[0048] The sample set construction module is configured as follows: using existing tunnel blasting case data, a machine learning sample set SS is constructed. SS takes each tunnel blasting case as a sample, the set of blasting design variable Xd and geological condition variable combination Xg as the sample feature X, and the rockburst dynamic evaluation variable Y as the sample label; it is divided into a training set and a test set.

[0049] The prediction model training module is configured as follows: It determines the machine learning prediction model MLM with X as input and Y as output, and determines the proposed hyperparameter combination HP in the MLM, where HP has N hyperparameter terms. j Each has its own range of values ​​[hp] j1 ,hp j2 ], constituting HP's search space Λ HP On the training set, in Λ HP A feasible solution for HP is randomly generated within the internal domain, and the hyperparameter combination HP is completed using the metaheuristic optimization algorithm model IPA in Λ. HP Internal optimization and training generate a tunnel blasting scheme design prediction model. This tunnel blasting scheme design prediction model is based on HP's performance in Λ... HP A model that minimizes the internal generalization error;

[0050] The optimized design output module is configured as follows: on the training set, in Λ Xd The feasible solution of the blasting scheme is designed by randomly generating internal blasting schemes. Based on the optimal MLM, the rockburst dynamic evaluation Y value of each feasible solution is predicted. The optimal solution of the combination of blasting design variables Xd is obtained by using the metaheuristic optimization algorithm model IPA, that is, the target blasting scheme of the tunnel is obtained.

[0051] The present invention also provides the following computer-readable storage medium.

[0052] A computer-readable storage medium having a computer program stored thereon, characterized in that: when the computer program is executed by a processor, it implements the steps of the method for reducing rockburst dynamic disasters induced by tunnel blasting as described in any one of claims 1 to 6.

[0053] Specifically, in the above-mentioned technical solutions of the present invention, the current geological condition variable Xg of the tunnel face includes the tunnel burial depth and uniaxial compressive strength, the tunnel blasting parameter Xd includes the charge decoupling coefficient, the spacing between peripheral holes, the minimum resistance line of peripheral holes, the relative distance and the charge concentration of peripheral holes, and the rockburst dynamic evaluation variable Y is the peak value of the blasting vibration particle velocity PPV.

[0054] Specifically, the machine learning prediction model (MLM) selected in the above-mentioned technical solutions of the present invention can be a decision tree classification model, and the hyperparameter term HP includes the number of decision trees n and the maximum depth of the decision trees d. max The minimum number of samples s required to segment internal nodes min .

[0055] Compared with existing technologies, the beneficial effects of this invention are as follows: The technical solution of this invention for reducing rockburst dynamic disasters induced by tunnel blasting is based on tunnel blasting conditions and standard data. It employs machine learning and hierarchical execution, utilizing a metaheuristic optimization algorithm model (IPA) to design an optimal target scheme that can reduce rockburst dynamic disasters induced by tunnel blasting. Optimization, according to general logic, does involve obtaining the optimal prediction model and then predicting the blasting design parameters for each blasting operation. If the over-excavation result obtained from the prediction is too large, the blasting design parameters are adjusted until a satisfactory effect is achieved. However, this is inefficient. The essence of blasting design parameter optimization is to find a set of blasting design parameters that minimizes the rockburst dynamic disaster induced by blasting, which conforms to the characteristics of a greedy algorithm. This allows for the efficient finding of the optimal blasting design parameters. Furthermore, by incorporating a search range variation step that changes with the iteration steps into the greedy algorithm, the algorithm avoids getting trapped in local optima during the iteration process, thus obtaining the best set of blasting design parameters. This invention enables tunnel construction teams to accurately select the most suitable blasting parameters, effectively reducing adverse rock mass vibrations caused by blasting, and thus lowering the risk of rockburst dynamic disasters induced by blasting in deep-buried tunnels. Furthermore, this system avoids potential safety hazards, improving the construction safety of deep-buried tunnels. Therefore, this improved tunnel blasting parameter adjustment and optimization system provides an innovative solution for the tunnel construction field, aiming to improve construction efficiency, reduce unnecessary risks, and ensure the quality and sustainability of tunnel projects. Attached Figure Description

[0056] Figure 1 It is the curve showing the change of the loss function (i.e., the fitness function) during data training in the HP tuning process.

[0057] Figure 2 It is the fitting curve between the predicted value of the constructed tunnel safety blasting scheme design prediction model and the actual value of the case.

[0058] Figure 3 This is a schematic diagram of the blasting parameter layout for the tunnel safety blasting target design scheme. Detailed Implementation

[0059] The preferred embodiments of the present invention will now be further described with reference to the accompanying drawings.

[0060] Example 1

[0061] like Figures 1-3 As shown, a safe blasting scheme for tunnels is designed using the method of this invention.

[0062] 1. The design space Λ constituting the safe blasting scheme for tunnels Xd

[0063] Investigate the current geological information of the tunnel face for blasting, and in conjunction with tunnel blasting engineering specifications, determine the combination of geological condition variables Xg, including two variables: tunnel burial depth and uniaxial compressive strength; determine the combination of blasting design variables Xd, where Xd has M = 5 blasting design variables xd. i These are: charge decoupling coefficient, spacing between peripheral holes, minimum resistance line of peripheral holes, relative distance, and charge concentration of peripheral holes. The parameters for each xd are determined according to the "Technical Specifications for Highway Tunnel Construction" based on the tunnel rock type. i The range of values ​​for [xd] i1 ,xd i2 The design space Λ that constitutes the tunnel safety blasting scheme Xd .

[0064] Table 1. Blasting Design Variables xd i Range of values

[0065]

[0066] 2. Construct the sample set SS

[0067] Using existing tunnel blasting case data, a machine learning sample set SS is constructed. SS uses each tunnel blasting case as a sample, the set of blasting design variables Xd and geological condition variables Xg as sample features X, and the rockburst dynamic evaluation variable Y as the sample label. In this example, the evaluation variable Y is the peak particle velocity (PPV) of blasting vibration. SS is divided into training and testing sets.

[0068] In this embodiment, the SS sample set consists of 400 groups, which are divided into training and testing sets in a 7:3 ratio, that is, 70% of the data is used for training and 30% is used to test the training effect of the model.

[0069] 3. Construct the hyperparameter search space Λ HP

[0070] The machine learning prediction model (MLM) is determined to be a decision tree classification model. The hyperparameter combination HP is to be optimized within the MLM, and HP has N = 3 hyperparameter terms. j These are the number of decision trees n and the maximum depth of the decision trees d, respectively. max The minimum number of samples s required to segment internal nodes min . Each hp j Each has a range of values ​​[hp] j1 ,hp j2 ], constituting HP's search space Λ HP .

[0071] Table 2 Hyperparameter hp j Range of values

[0072] <![CDATA[[hp j1 ,hp j2 ]]]> [1,500] [1,30] [2,50]

[0073] 4. Training a prediction model for tunnel blasting scheme design

[0074] Using the metaheuristic optimization algorithm model IPA to find Λ HP The optimal HP of the inner MLM is used to obtain the prediction model for the design of safe blasting schemes in tunnels.

[0075] 4.1 Initialize system parameters

[0076] Set the HP population size and iteration termination condition; in Λ HP The HP value is randomly adjusted to generate a certain number of feasible HP solutions, called HP individuals, which constitute the HP population. Each HP individual contains N hyperparameter terms hpj, located at I = (I1, I2, ..., I...). N ), initialize system parameters.

[0077] In this example, the hyperparameters n and d in MLM max s min These correspond to I1, I2, and I3 respectively. HP combinations are randomly generated according to the population size.

[0078] Obtain the initial HP population.

[0079] 4.2 Assigning Initial Role Markers

[0080] For the initial HP population, the first HP individual is assigned the Ld tag, the individual closest to the Ld tag is assigned the Fd tag, and the remaining individuals are assigned the Fl tag; the position of the Fd tag is denoted as I. * (a), and all other marked positions are denoted as I(a).

[0081] This also applies to the hyperparameters n and d in MLM. max s min The population that makes up the group.

[0082] 4.3 Update HP Population I * (a)

[0083] For the current HP population, calculate the fitness value L of each HP individual based on MLM and according to Equation 1, and obtain the HP individual with the minimum L value L. min If L min Individuals with the Fd tag are retained; otherwise, the L tag is swapped. min Individual and Fd mark the individual's mark, obtain the updated I * (a).

[0084] Throughout the computation process, the hyperparameters n and d of the MLP are continuously searched. max s minThe combination of these parameters, with each iteration updating the hyperparameter combination, allows the computation of a specific set of n, d parameters within the population during a single iteration. max s min The combination minimizes the value of L, at which point n and d are... max s min The combination is the current I * (a).

[0085] 4.4 Update HP population Ld tag I(a)

[0086] For the individual marked Ld, its position I(a) is updated according to Equations 2 to 5. Ld is the first group of n and d. max s min combination

[0087] 4.5 Update HP population Fl marker I(a)

[0088] For any Fl-marked individual, update its position I(a) according to Equations 6 to 8.

[0089] Fl-labeled individuals are those other than the first group n, d max s min All remaining HP individuals outside of the group.

[0090] Once the Fl marker I(a) is updated, the updated HP population is obtained.

[0091] 4.6 Iterative Evolution

[0092] The training process is completed by iterating through steps 4.3 to 4.5 until the preset iteration conditions are met. The generalization ability of the iteratively evolved model is then evaluated on the test set to obtain the optimal MLM, i.e., the tunnel safety blasting scheme design prediction model.

[0093] Figure 1 It is the curve showing the change of the loss function (i.e., the fitness function) during data training in the HP tuning process.

[0094] Figure 2 It is the fitting curve between the predicted value of the constructed tunnel safety blasting scheme design prediction model and the actual value of the case.

[0095] 5. Generate a safe blasting target plan for the tunnel.

[0096] Based on the prediction model, the metaheuristic optimization algorithm model IPA is used to find Λ. Xd The optimal solution of the prediction model for safe blasting scheme design in internal tunnels.

[0097] 5.1 Initialize system parameters

[0098] Set the population size of Xd and the iteration termination condition; in Λ XdRandomly adjust each Xd within Xd. i The values ​​are selected to generate a certain number of feasible solutions Xd, called Xd individuals, which constitute the Xd population. Each Xd contains M = 5 design variables xd. i The position is I = (I1, I2, ..., I5), and the system parameters are initialized.

[0099] The initial Xd population is obtained.

[0100] 5.2 Assigning Initial Role Markers

[0101] The operation is the same as described in section 4.2 above.

[0102] 5.3 Update Xd Population I * (a)

[0103] For the current Xd population, the rockburst dynamic evaluation Y value of each feasible solution is predicted based on the optimal MLM, and the minimum Y value of individual Xd is obtained. min Similarly, the method of exchanging labels with individuals marked with Fd or retaining its own Fd label is used to ensure Y min Individuals are assigned Fd tags and then updated I. * (a).

[0104] 5.4 Update the Ld marker I(a) of the Xd population.

[0105] The operation is the same as described in section 4.4 above.

[0106] Version 5.5 updates the Fl marker I(a) for the Xd population.

[0107] For any Fl-marked individual, its position I(a) is updated according to Equations 6 to 8, where parameter b is determined according to Equations 9 and 10.

[0108] The variation of parameter 'b' with the number of iterations allows for a wider search range for blasting design parameter combinations in the early stages, followed by a narrower search range in the later stages. This transformation is more conducive to finding the optimal blasting design parameter combination. Furthermore, each design parameter has its own range, and different degrees of variation are determined for each parameter's range.

[0109] Once the Fl marker I(a) is updated, the updated Xd population is obtained.

[0110] 5.6 Iterative Evolution

[0111] The training is completed after iterating through steps 5.3 to 5.5 until the preset iteration conditions are met. As a preferred implementation, this example starts from the first iteration of the algorithm (a = 1), and in each iteration... * (a) After the update, add the domain perturbation expressed in Equations 11 and 12, and I * (a) Updated to I g(a) Then perform Ld tag I(a) update and Fl tag I(a) update.

[0112] The final design scheme for safe blasting of tunnels was obtained.

[0113] Figure 3 This is a schematic diagram of the blasting parameter layout for the tunnel safety blasting target design scheme.

Claims

1. A method for reducing rockburst dynamic hazards induced by tunnel blasting, comprising designing a target scheme to reduce rockburst dynamic hazards induced by tunnel blasting using blasting condition data, characterized in that: Step S100: Investigate the geological information of the tunnel face at the current blasting site. Based on the tunnel blasting engineering specifications, determine the combination of geological condition variables Xg and the combination of blasting design variables Xd at the current tunnel face. Xd has M blasting design variables xd. i Each has its own range of values ​​[xd] i1 , xd i2 ], constituting the design space Λ Xd ; Step S200: Using existing tunnel blasting case data, construct a machine learning sample set SS. SS takes each tunnel blasting case as a sample, the set of blasting design variable Xd and geological condition variable combination Xg as the sample feature X, and the rock burst dynamic evaluation variable Y as the sample label; it is divided into a training set and a test set. Step S300: Determine the machine learning prediction model MLM with X as input and Y as output, and determine the proposed hyperparameter combination HP in the MLM. HP has N hyperparameter terms hp j Each has its own range of values ​​[hp] j1 hp j2 ], constituting HP's search space Λ HP ; Step S400: On the training set, in Λ HP A feasible solution for HP is randomly generated within the internal domain, and the hyperparameter combination HP is completed using the metaheuristic optimization algorithm model IPA in Λ. HP Internal optimization and training generate a tunnel blasting scheme design prediction model. This tunnel blasting scheme design prediction model is based on HP's performance in Λ... HP A model that minimizes the internal generalization error; Step S400 includes step S410, setting the HP population size and the iteration termination condition, in Λ HP The system randomly adjusts the HP value to generate a certain number of feasible HP solutions, called HP individuals, which constitute the HP population. Each HP individual contains N hyperparameter terms hp. j The position is Initialize system parameters; Step S420: For the initial HP population, assign the Ld tag to the first HP individual, assign the Fd tag to the individual closest to the Ld tag, and assign the Fl tag to the remaining individuals. The position of the Fd tag is recorded as... The remaining marked positions are all recorded as ; Step S430: For the current HP population, calculate the fitness value L for each HP individual, and obtain the HP individual with the smallest L value L. min ;L is ensured by either exchanging tags with individuals marked with Fd or retaining its own Fd tag. min Individual assignments obtain Fd tags, then update Execute step S440; Step S440: For the Ld-marked individual, update its position according to Equations 2 to 5. , Formula 2 Formula 3 Formula 4 Formula 5 in, - Current iteration number; - The current best feasible solution corresponds to the current best position of HP; and - These represent the positions of the current iteration and the next iteration, respectively; - Movement step size; C, A - coefficients; d - Ld marks the individual position and updates the path parameters, which decreases linearly from 2 to 0 according to the number of iterations; s - a random number between (0,1]. Step S450: For any Fl-marked individual, update its position according to Equations 6 to 8. , Formula 6 Formula 7 Formula 8 In the formula, b-Fl marks the individual location update path parameter, with a value of 1; - A random quantity in the range [-1, 1]; e - the natural constant; - Randomly selected location; Obtain updated HP population; Step S460: Iterate through steps S430 to S450 until the preset iteration conditions are met to complete the training; obtain the optimal MLM. Step S500: On the training set, in Λ Xd The feasible solution of the blasting scheme is designed by randomly generating internal blasting schemes. Based on the optimal MLM, the rockburst dynamic evaluation Y value of each feasible solution is predicted. The optimal solution of the combination of blasting design variables Xd is obtained by using the metaheuristic optimization algorithm model IPA, that is, the target blasting scheme of the tunnel is obtained.

2. The method according to claim 1, characterized in that: The step S400 is, in Λ HP Internal random adjustment of HP within each HP i A certain number of feasible HP solutions are generated, called HP individuals. All HP individuals constitute the HP population. Based on MLM, the quality of each HP individual is evaluated using a fitness value L. The HP individual with the smallest L value is taken as the current optimal solution individual in the HP population. The positions of the remaining HP individuals are updated based on the position of the current optimal solution individual to obtain the updated HP population. A greedy algorithm is used to find the current optimal solution individual, and the HP population is updated again. This process is iterated until the preset iteration conditions are met, and the training is completed. The optimal MLM is obtained. The fitness function is expressed according to Equation 1. Formula 1 In the formula, L-HP represents the fitness value of an individual in a feasible solution. - Original Y-value data for the case study - MLM prediction of Y values, K - number of predictor variables in the training set; On the test set, the optimal generalization ability of MLM was evaluated, and a tunnel blasting scheme design prediction model was obtained.

3. The method according to claim 2, characterized in that: The step S500 is, in Λ Xd Randomly adjust each xd within Xd. i The process involves selecting values ​​to generate a certain number of feasible solutions for Xd, referred to as Xd individuals, which constitute the Xd population. Based on MLM, the quality of each Xd individual is evaluated using the Y value, and the Xd individual with the smallest Y value is taken as the current optimal solution individual in the Xd population. The positions of the remaining Xd individuals are updated based on the position of the current optimal solution individual to obtain the updated Xd population. A greedy algorithm is then used to find the current optimal solution individual, and the Xd population is updated again. This iterative process continues until the preset iteration conditions are met, completing the training. The optimal solution for the combination of explosive design variables Xd is obtained.

4. The method according to claim 1, characterized in that: Step S500 includes, Step S510: Set the population size of Xd and the iteration termination condition, in Λ Xd Randomly adjust each xd within Xd. i The values ​​are selected to generate a certain number of feasible solutions Xd, called Xd individuals, which constitute the Xd population. Each Xd contains M design variables xd. i The position is Initialize system parameters; Step S520: For the initial Xd population, assign the Ld tag to the first Xd individual, assign the Fd tag to the individual closest to the Ld tag, and assign the Fl tag to the remaining individuals. The position of the Fd tag is recorded as... The positions of the remaining individuals are recorded as ; Step S530: Using the predicted Y value as the evaluation index for each Xd individual, the optimal solution for the combination of blasting design variables Xd is obtained using the same algorithm logic as steps S430 to S460; where, For the location update calculation of any Fl-marked Xd individual, the parameter b is determined according to Equations 9 and 10. Formula 9 Formula 10 In the formula, ; - Maximum number of iterations -No. Design variables xd i ; and - These are the first Design variables xd k Upper limit of value xd k2 With lower limit xd k1 .

5. The method according to claim 4, characterized in that: Starting from the a=1th iteration, each time... After updating, verify and update according to Equations 11 and 12. , Formula 11 Formula 12 In the formula, - Examine the updated Fd marker Xd individual location; - A random number between 0 and 1; - The tunnel blasting scheme design prediction model obtained in step S400; - Current design variable xd i The range of values ​​for [xd] i1 , xd i2 Random numbers within ] 6. A system for reducing rockburst dynamic hazards induced by tunnel blasting, characterized by designing a target scheme to reduce rockburst dynamic hazards induced by tunnel blasting using blasting condition data, wherein... The system for reducing rockburst dynamic hazards induced by tunnel blasting is used to implement the method for reducing rockburst dynamic hazards induced by tunnel blasting as described in any one of claims 1 to 5. The design space acquisition module is configured as follows: Based on the current geological information of the tunnel face and in conjunction with tunnel blasting engineering specifications, determine the combination of geological condition variables Xg and the combination of blasting design variables Xd, where Xd has M blasting design variables xd. i Each has its own range of values ​​[xd] i1 , xd i2 ], constituting the design space Λ Xd ; The sample set construction module is configured as follows: using existing tunnel blasting case data, a machine learning sample set SS is constructed. SS takes each tunnel blasting case as a sample, the set of blasting design variable Xd and geological condition variable combination Xg as the sample feature X, and the rockburst dynamic evaluation variable Y as the sample label; it is divided into a training set and a test set. The prediction model training module is configured as follows: It determines the machine learning prediction model MLM with X as input and Y as output, and determines the proposed hyperparameter combination HP in the MLM, where HP has N hyperparameter terms. j Each has its own range of values ​​[hp] j1 hp j2 ], constituting HP's search space Λ HP On the training set, in Λ HP A feasible solution for HP is randomly generated within the internal domain, and the hyperparameter combination HP is completed using the metaheuristic optimization algorithm model IPA in Λ. HP Internal optimization and training generate a tunnel blasting scheme design prediction model. This tunnel blasting scheme design prediction model is based on HP's performance in Λ... HP A model that minimizes the internal generalization error; The optimized design output module is configured as follows: on the training set, in Λ Xd The feasible solution of the blasting scheme is designed by randomly generating internal blasting schemes. Based on the optimal MLM, the rockburst dynamic evaluation Y value of each feasible solution is predicted. The optimal solution of the combination of blasting design variables Xd is obtained by using the metaheuristic optimization algorithm model IPA, that is, the target blasting scheme of the tunnel is obtained.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the method for reducing rockburst dynamic disasters induced by tunnel blasting as described in any one of claims 1 to 5.

8. The method for reducing rockburst dynamic hazards induced by tunnel blasting according to any one of claims 1 to 5, or the system for reducing rockburst dynamic hazards induced by tunnel blasting according to claim 6, or the computer-readable storage medium according to claim 7, characterized in that: The current geological condition variable Xg at the tunnel face includes the tunnel burial depth and uniaxial compressive strength. The tunnel blasting parameter Xd includes the charge decoupling coefficient, the spacing between peripheral holes, the minimum resistance line of the peripheral holes, the relative distance, and the charge concentration of the peripheral holes. The rockburst dynamic evaluation variable Y is the peak particle velocity (PPV) of the blasting vibration particles.

9. The method for reducing rockburst dynamic hazards induced by tunnel blasting according to any one of claims 1 to 5, or the system for reducing rockburst dynamic hazards induced by tunnel blasting according to claim 6, or the computer-readable storage medium according to claim 7, characterized in that: The machine learning prediction model MLM is a decision tree classification model, and the hyperparameter HP includes the number of decision trees. Maximum depth of decision tree Minimum number of samples required to segment internal nodes .

Images