A neural network-based method for predicting blasting vibrations

Abstract

This invention discloses a method for predicting blasting vibrations based on neural networks, comprising: acquiring historical blasting data during blasting and preprocessing the blasting data to obtain training data; using a neural network as a blasting vibration prediction model, and based on the training data, training the blasting vibration prediction model using a multi-step optimization algorithm to obtain a trained blasting vibration prediction model; acquiring real-time blasting data during blasting, using this real-time blasting data as input to the trained blasting vibration prediction model, and using the trained blasting vibration prediction model to predict blasting vibrations to obtain blasting vibration prediction results. This invention not only eliminates the need for prediction personnel to possess specialized knowledge but also improves prediction efficiency. The multi-step optimization algorithm used to update the neural network not only achieves rapid local convergence but also avoids getting trapped in local optima, thereby achieving accurate training and enabling the trained neural network to better predict blasting vibrations.

Description

Technical Field

[0001] This invention belongs to the field of engineering blasting, specifically relating to a method for predicting blasting vibrations based on neural networks. Background Technology

[0002] When explosives detonate, humans utilize their chemical energy to convert it into mechanical work, enabling them to perform tasks that are impossible or difficult for humans or machines to accomplish. Engineering blasting is a typical example of using explosives to perform work. However, while explosives detonate in rock, they also produce harmful effects such as blasting vibration, air shock waves, noise, occasional flyrock, and toxic gases, which are currently difficult to avoid. Among the harmful effects of blasting, blasting vibration has a wide range of impact and poses a significant threat. Blasting vibration can cause landslides on slopes around mines, roof collapses in tunnel engineering, cracks and collapses in surrounding buildings, and endanger the safety of nearby residential buildings and urban pipelines during urban demolition blasting. Improper prevention and control can lead to huge property losses and casualties, not only causing engineering blasting failure but also resulting in unpredictable legal liabilities. Therefore, effectively predicting blasting vibration and using the predicted data as a guide to optimize blasting design parameters is crucial to controlling blasting vibration within a safe range, reducing its negative impacts, and improving the economic efficiency of blasting projects.

[0003] In existing technologies, empirical models are often used for manual prediction. However, manual prediction requires the predictors to have relatively specialized knowledge, and the prediction efficiency is relatively low. Summary of the Invention

[0004] In view of the above-mentioned shortcomings in the prior art, the present invention provides a blasting vibration prediction method based on neural networks, which solves the problems in the prior art.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a method for predicting blasting vibration based on neural networks, comprising:

[0006] Historical blasting data was acquired and preprocessed to obtain training data.

[0007] Using a neural network as the blasting vibration prediction model, based on the training data, and employing a multi-step optimization algorithm to train the blasting vibration prediction model, a trained blasting vibration prediction model is obtained.

[0008] Real-time blasting data is acquired during blasting and used as input to a trained blasting vibration prediction model. The trained blasting vibration prediction model is then used to predict blasting vibrations, resulting in blasting vibration prediction results.

[0009] In one possible real-time mode, the historical blasting data includes conditional parameters and result parameters. The conditional parameters include the distance between blasting measuring points, the time difference between holes, the number of blast holes, the hole spacing, the row spacing, the maximum hole depth, the total charge, and the maximum charge per section. The result parameters include the peak velocity of the target particle in the XYZ directions and the dominant vibration frequency.

[0010] In one possible real-time approach, the blasting data is preprocessed to obtain training data, including: normalizing the distance between measuring points, the time difference between holes, the number of blast holes, the hole spacing, the row spacing, the maximum hole depth, the total charge, and the maximum charge in a single section, as input data;

[0011] The peak velocities and dominant vibration frequencies of the target particle in the X, Y, and Z directions are normalized and used as the desired output data.

[0012] The input data and the expected output data are used together as training data.

[0013] In one possible real-time approach, using a neural network as a blasting vibration prediction model includes: employing a BP neural network as a blasting vibration prediction model.

[0014] In one possible real-time approach, based on the training data, a multi-step optimization algorithm is used to train the blasting vibration prediction model to obtain a trained blasting vibration prediction model, including:

[0015] Set the minimum number of training rounds Smin, the minimum error threshold q, and the current training round t = 1;

[0016] Construct the error function of the blasting vibration prediction model;

[0017] The weights of the blasting vibration prediction model are randomly initialized N times to obtain N weight vectors. Each weight vector is treated as an individual to obtain a population.

[0018] The update method is obtained using a roulette wheel algorithm, and the update method includes self-update and guided update.

[0019] If the update method is self-update, then each individual in the population is self-updated according to the error function to obtain the initially updated population; if the update method is guided update, then each individual in the population is guided update according to the error function to obtain the initially updated population.

[0020] A random walk is performed on each individual in the population after the initial update, and the individual after the walk is determined according to the error function to obtain the population after the second update;

[0021] Determine whether the current training round number t is greater than or equal to the minimum training round number Smin. If so, obtain the error function value corresponding to each individual in the population after the second update. Otherwise, increment the count value of t by one and proceed to the next round of training.

[0022] Determine if there exists an individual whose error function value is less than the minimum error threshold q. If so, use the individual with the smallest error function value as the weight parameter of the blasting vibration prediction model to obtain the trained blasting vibration prediction model. Otherwise, increment the count value of t by one and proceed to the next round of training.

[0023] In one possible real-time approach, the error function E for constructing the blasting vibration prediction model is:

[0024]

[0025]

[0026]

[0027] Where, d nk y represents the actual output value of the k-th neuron in the output layer. nk Let $\begin{pmatrix} \ ... k w represents the intermediate coefficient. jk h represents the weight between the j-th neuron in the hidden layer and the k-th neuron in the output layer. nj θ represents the output of the j-th neuron in the hidden layer of the blasting vibration prediction model. k This represents the first threshold corresponding to the k-th neuron in the output layer.

[0028] In one possible real-time approach, the roulette wheel algorithm is used to obtain the update method, which includes: setting the probability of self-update to 0.5, generating a random number r in (0, 1), determining whether the random number r is less than 0.5, and if so, performing self-update; otherwise, performing guided update.

[0029] In one possible real-time approach, each individual in the population is self-updated based on an error function, including:

[0030] Based on the training function and the error function, obtain the error function value corresponding to each individual in the population, and select the individual with the smallest error function value as the current best individual;

[0031] Based on the current best individual, each individual is self-updated as follows:

[0032]

[0033] in, W represents the current optimal individual; i t W represents the i-th individual in the population during the t-th round of training; i t+1 Represents the updated individual W i t e represents the natural constant, b represents a constant, l represents a random number uniformly distributed in [0, 1], and π represents pi.

[0034] In one possible real-time approach, guided updates are performed on each individual in the population based on an error function, including:

[0035] Based on the training function and the error function, obtain the error function value corresponding to each individual in the population, and select the individual with the smallest error function value as the current best individual;

[0036] Generate a random vector A, the dimension of which is the same as the dimension of the weight vector, and all dimensions of the random vector A are random numbers uniformly distributed in (-a, a).

[0037] Obtain the modulus of random vector A, and determine whether the modulus of random vector A is less than 1. If it is, perform the first guiding update on the others based on the best individual; otherwise, randomly select an individual and perform the second guiding update on the other individuals based on the randomly selected individual.

[0038] The first guide is updated to:

[0039]

[0040] in, W represents the current optimal individual; i t W represents the i-th individual in the swarm during the t-th round of training; i t+1 Represents the updated individual W i t C represents a random number uniformly distributed between [0, 2].

[0041] The second bootloader is updated as follows:

[0042]

[0043] in, This refers to an individual selected randomly.

[0044] In one possible real-time approach, a random walk is performed on each individual in the initially updated population, and the individuals after the walk are determined based on an error function, resulting in a second-updated population, including:

[0045] Randomly select an individual to obtain the target individual;

[0046] Update the p-th dimension weights of the target individual:

[0047]

[0048] in, Let w represent the updated p-th dimension weight. p Let p represent the p-th dimension weight before the update, π represent pi, p = 1, 2, ..., p', p' represent the total dimension of the target individual, and Step represent the walk step size;

[0049] Determine whether the error function value corresponding to the target individual before the p-th dimension weight update is less than the error function value corresponding to the target individual before the update. If so, accept the update; otherwise, reject the update.

[0050] The weights of all dimensions of the target individual are updated to obtain the updated target individual;

[0051] The target individuals are reselected in the population for updating, and this process is repeated until all individuals have been updated, resulting in a population that has undergone a second update.

[0052] The beneficial effects of this invention are as follows:

[0053] This invention provides a method for predicting blasting vibrations based on neural networks. It not only eliminates the need for personnel with specialized knowledge in prediction, but also improves prediction efficiency by using neural networks for prediction. Furthermore, the use of a multi-step optimization algorithm to update the neural network not only achieves rapid local convergence but also avoids getting trapped in local optima, thereby achieving accurate training. This allows the trained neural network to better predict blasting vibrations. Attached Figure Description

[0054] Figure 1 A flowchart of a neural network-based method for predicting blasting vibrations, provided as an embodiment of the present invention. Detailed Implementation

[0055] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0056] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0057] like Figure 1 As shown, a method for predicting blasting vibrations based on neural networks includes:

[0058] S11. Obtain historical blasting data during the blasting process, and preprocess the blasting data to obtain training data.

[0059] Historical blasting data refers to parameters that affect the vibration of the target point during the blasting process. This historical blasting data can be pre-stored data or data generated through human-computer interaction.

[0060] Preprocessing of blasting data can include normalization. However, if the blasting data contains text, the text can be converted to numbers before normalization. For example, if each character corresponds to a number, then when a character appears, it is converted to a number.

[0061] S12. Using a neural network as the blasting vibration prediction model, based on the training data, and employing a multi-step optimization algorithm to train the blasting vibration prediction model, a trained blasting vibration prediction model is obtained.

[0062] The neural network can be a BP (BackPropagation) neural network, and then a multi-step optimization algorithm is used to train the neural network so that it can predict vibrations more accurately.

[0063] S13. Obtain real-time blasting data during blasting, use this real-time blasting data as input to the trained blasting vibration prediction model, and use the trained blasting vibration prediction model to predict blasting vibration to obtain blasting vibration prediction results.

[0064] Once trained, the blasting vibration prediction model has the ability to predict vibration based on parameters affecting the vibration of the target point. Therefore, it can collect parameters that affect the vibration of the target point during the current blasting and predict the vibration of the target point based on these parameters. This improves the efficiency of prediction and reduces the professional knowledge requirements of the prediction personnel through deep learning algorithms.

[0065] In one possible real-time mode, the historical blasting data includes conditional parameters and result parameters. The conditional parameters include the blasting measurement point distance (unit: R), inter-hole differential time (unit: ms), number of blast holes (unit: holes), hole spacing (unit: m), row spacing (unit: m), maximum hole depth (unit: m), total charge (unit: kg), and maximum charge per section (unit: kg). The result parameters include the peak velocity of the target particle in the XYZ directions (unit: cm / s) and the dominant vibration frequency (unit: Hz).

[0066] It is worth noting that the peak velocities of the target particle in the X, Y, and Z directions can be selected as the vibration result, or only the dominant vibration frequency can be used. When collecting training data, the corresponding expected output can be obtained by collecting only one type of data, thus achieving prediction of a single data point.

[0067] In one possible real-time approach, the blasting data is preprocessed to obtain training data, including normalizing the distance between measuring points, the inter-hole differential time, the number of boreholes, the hole spacing, the row spacing, the maximum hole depth, the total charge, and the maximum charge per section, as input data. It is worth noting that the conditional parameters primarily affect the vibration parameters of the target point; in addition to the conditional parameters described in the embodiments of this application, other parameters can also be selected for prediction.

[0068] The peak velocities and dominant vibration frequencies of the target particle in the X, Y, and Z directions are normalized and used as the desired output data.

[0069] The input data and the expected output data are used together as training data.

[0070] By transforming the blasting data into the input form of a BP neural network, training or prediction can be performed based on the blasting data. Furthermore, data normalization can reduce the amount of data processing and improve training speed.

[0071] In one possible real-time approach, using a neural network as a blasting vibration prediction model includes: employing a BP neural network as a blasting vibration prediction model.

[0072] It is worth noting that, in addition to using a BP neural network as the blasting vibration prediction model, other neural networks can also be used as the blasting vibration prediction model, and trained using the weight training algorithm provided in this embodiment. This embodiment only provides the weight algorithm; for the training or setting of other parameters, conventional methods or existing technologies can be used for training and setting.

[0073] In one possible real-time approach, based on the training data, a multi-step optimization algorithm is used to train the blasting vibration prediction model to obtain a trained blasting vibration prediction model, including:

[0074] Set the minimum number of training rounds Smin, the minimum error threshold q, and the current training round number t=1.

[0075] Construct the error function of the blasting vibration prediction model.

[0076] The weights of the blasting vibration prediction model are randomly initialized N times, resulting in N weight vectors. Each weight vector is treated as an individual, forming a population. In other words, all the weight parameters of the blasting vibration prediction model are combined into a weight vector, and training this weight vector is equivalent to training the blasting vibration prediction model.

[0077] The update method is obtained using a roulette wheel algorithm, and the update method includes self-update and guided update.

[0078] If the update method is self-update, then each individual in the population is self-updated according to the error function to obtain the initially updated population. If the update method is guided update, then each individual in the population is guided to update according to the error function to obtain the initially updated population.

[0079] A random walk is performed on each individual in the population after the initial update, and the individual after the walk is determined according to the error function, thus obtaining the population after the second update.

[0080] Determine if the current training round number t is greater than or equal to the minimum training round number Smin. If so, obtain the error function value corresponding to each individual in the population after the second update. Otherwise, increment the count value of t by one and proceed to the next round of training.

[0081] Determine if there exists an individual whose error function value is less than the minimum error threshold q. If so, use the individual with the smallest error function value as the weight parameter of the blasting vibration prediction model to obtain the trained blasting vibration prediction model. Otherwise, increment the count value of t by one and proceed to the next round of training.

[0082] In one possible real-time approach, the error function E for constructing the blasting vibration prediction model is:

[0083]

[0084]

[0085]

[0086] Where, d nk y represents the actual output value of the k-th neuron in the output layer. nkLet $\begin{pmatrix} \ ... k w represents the intermediate coefficient. jk h represents the weight between the j-th neuron in the hidden layer and the k-th neuron in the output layer. nj θ represents the output of the j-th neuron in the hidden layer of the blasting vibration prediction model. k This represents the first threshold corresponding to the k-th neuron in the output layer.

[0087] It is worth noting that, in addition to the error function shown in this embodiment, other error functions of BP neural networks can also be used.

[0088] In one possible real-time approach, the roulette wheel algorithm is used to obtain the update method, which includes: setting the probability of self-update to 0.5, generating a random number r in (0, 1), determining whether the random number r is less than 0.5, and if so, performing self-update; otherwise, performing guided update.

[0089] In one possible real-time approach, each individual in the population is self-updated based on an error function, including:

[0090] Based on the training function and the error function, obtain the error function value corresponding to each individual in the population, and select the individual with the smallest error function value as the current optimal individual.

[0091] When obtaining the error function value, multiple training data can be input into the blasting vibration prediction model to obtain multiple E values. Then, the obtained multiple E values ​​are used as the error function value, thereby reducing the impact of single data error.

[0092] Based on the current best individual, each individual is self-updated as follows:

[0093]

[0094] in, This represents the current optimal individual. W i t Let W represent the i-th individual in the population during the t-th round of training. i t+1 Represents the updated individual W i t e represents the natural constant, b represents a constant, l represents a random number uniformly distributed in [0, 1], and π represents pi.

[0095] By enabling each individual to update its own position, the individual can continuously update its position, thereby finding the optimal solution.

[0096] Optionally, when performing self-updates for each individual, the error function value before and after the update can be obtained. If the error function value increases, the update is rejected; if the error function value decreases, the update is accepted, thereby reducing useless updates.

[0097] In one possible real-time approach, guided updates are performed on each individual in the population based on an error function, including:

[0098] Based on the training function and the error function, obtain the error function value corresponding to each individual in the population, and select the individual with the smallest error function value as the current optimal individual.

[0099] Generate a random vector A, wherein the dimension of the random vector A is the same as the dimension of the weight vector, and all dimensions of the random vector A are random numbers uniformly distributed in (-a, a).

[0100] Optionally, besides determining training completion upon meeting preset conditions, a maximum number of training epochs can be set. Training can be considered complete after completing the maximum number of training epochs. In this case, the initial value of 'a' can be set to 2, linearly decreasing to 0 as the training epochs progress. Each decrease can be 2 / maximum number of training epochs. If no maximum number of epochs is set, each decrease can be set to a relatively small number to ensure that iterations can proceed normally.

[0101] Obtain the modulus of random vector A, and determine whether the modulus of random vector A is less than 1. If it is, perform the first guiding update on the others based on the best individual; otherwise, randomly select an individual and perform the second guiding update on the others based on the randomly selected individual.

[0102] The first guide is updated to:

[0103]

[0104] in, This represents the current optimal individual. W i t Let W represent the i-th individual in the group during the t-th round of training. i t+1 Represents the updated individual W i t C represents a random number uniformly distributed between [0, 2].

[0105] The second bootloader is updated as follows:

[0106]

[0107] in, This refers to an individual selected randomly.

[0108] Guided updates direct individuals toward the optimal individual or randomly selected individuals to explore different search distances, thereby moving towards the optimal solution. Optionally, during guided updates, the error function values ​​before and after the update can be obtained. If the error function value increases, the update is rejected; if the error function value decreases, the update is accepted, thus reducing useless updates.

[0109] In one possible real-time approach, a random walk is performed on each individual in the initially updated population, and the individuals after the walk are determined based on an error function, resulting in a second-updated population, including:

[0110] Randomly select an individual to obtain the target individual.

[0111] Update the p-th dimension weights of the target individual:

[0112]

[0113] in, Let w represent the updated p-th dimension weight. p Let p represent the weight of the p-th dimension before the update, π represent pi, p = 1, 2, ..., p', p' represent the total dimension of the target individual, and Step represent the walk step size.

[0114] By performing random walks, a larger search space is generated, thus avoiding the problem of getting stuck in local optima during training.

[0115] Determine whether the error function value corresponding to the target individual before the p-th dimension weight update is less than the error function value corresponding to the target individual before the update. If so, accept the update; otherwise, reject the update.

[0116] The weights of all dimensions of the target individual are updated to obtain the updated target individual.

[0117] The target individuals are reselected in the population for updating, and this process is repeated until all individuals have been updated, resulting in a population that has undergone a second update.

[0118] The present invention provides a method for predicting blasting vibrations based on neural networks. This method not only eliminates the need for personnel with professional knowledge to perform predictions, but also improves prediction efficiency by using neural networks. Furthermore, the use of a multi-step optimization algorithm to update the neural network not only achieves rapid local convergence but also avoids getting trapped in local optima, thereby enabling accurate training and allowing the trained neural network to better predict blasting vibrations.

[0119] It is worth noting that any method utilizing the inventive concept should be within the scope of protection of this application. Other embodiments of this application will readily conceive of those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0120] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for predicting blasting vibrations based on neural networks, characterized in that, include: Historical blasting data was acquired and preprocessed to obtain training data. Using a neural network as the blasting vibration prediction model, based on the training data, and employing a multi-step optimization algorithm to train the blasting vibration prediction model, a trained blasting vibration prediction model is obtained. Acquire real-time blasting data during blasting, use this real-time blasting data as input to the trained blasting vibration prediction model, and use the trained blasting vibration prediction model to predict blasting vibration to obtain blasting vibration prediction results. Based on the training data, a multi-step optimization algorithm is used to train the blasting vibration prediction model, resulting in a trained blasting vibration prediction model, including: Set the minimum number of training rounds Smin, the minimum error threshold q, and the current training round t=1; Construct the error function of the blasting vibration prediction model; The weights of the blasting vibration prediction model are randomly initialized N times to obtain N weight vectors. Each weight vector is treated as an individual to obtain a population. The update method is obtained using a roulette wheel algorithm, and the update method includes self-update and guided update. If the update method is self-update, then each individual in the population is self-updated according to the error function to obtain the initially updated population; if the update method is guided update, then each individual in the population is guided update according to the error function to obtain the initially updated population. A random walk is performed on each individual in the population after the initial update, and the individual after the walk is determined according to the error function to obtain the population after the second update; Determine whether the current training round number t is greater than or equal to the minimum training round number Smin. If so, obtain the error function value corresponding to each individual in the population after the second update. Otherwise, increment the count value of t by one and proceed to the next round of training. Determine if there is an individual whose error function value is less than the minimum error threshold q. If so, use the individual with the smallest error function value as the weight parameter of the blasting vibration prediction model to obtain the trained blasting vibration prediction model. Otherwise, increment the count value of t by one and proceed to the next round of training. The population is self-updated based on the error function, including: Based on the training function and the error function, obtain the error function value corresponding to each individual in the population, and select the individual with the smallest error function value as the current best individual; Based on the current best individual, each individual is self-updated as follows: ； in, This represents the current optimal individual; This represents the i-th individual in the population during the t-th round of training; Indicates the updated individual , Let represent the natural constant, b represent a constant, and l represent a random number uniformly distributed in [0, 1]. Represents pi; Guided updates are performed on each individual in the population based on the error function, including: Based on the training function and the error function, obtain the error function value corresponding to each individual in the population, and select the individual with the smallest error function value as the current best individual; Generate a random vector A, wherein the dimension of the random vector A is the same as the dimension of the weight vector, and all dimensions of the random vector A are random numbers uniformly distributed in (-a, a); Obtain the modulus of random vector A, and determine whether the modulus of random vector A is less than 1. If it is, perform the first guiding update on the others based on the best individual; otherwise, randomly select an individual and perform the second guiding update on the other individuals based on the randomly selected individual. The first guide is updated to: ； in, This represents the current optimal individual; This represents the i-th individual in the group during the t-th round of training; Indicates the updated individual , This represents a random number that is uniformly distributed between [0, 2]. The second bootloader is updated as follows: ； in, This refers to an individual selected randomly.

2. The blasting vibration prediction method based on neural networks according to claim 1, characterized in that, The historical blasting data includes conditional parameters and result parameters. The conditional parameters include the distance between blasting measuring points, the time difference between holes, the number of blast holes, the hole spacing, the row spacing, the maximum hole depth, the total charge, and the maximum charge per section. The result parameters include the peak velocity of the target particle in the XYZ directions and the dominant vibration frequency.

3. The blasting vibration prediction method based on neural networks according to claim 2, characterized in that, The blasting data is preprocessed to obtain training data, including: normalizing the distance between measuring points, the time difference between holes, the number of blast holes, the hole spacing, the row spacing, the maximum hole depth, the total charge, and the maximum charge in one section, which are then used as input data; The peak velocities and dominant vibration frequencies of the target particle in the X, Y, and Z directions are normalized and used as the desired output data. The input data and the expected output data are used together as training data.

4. The blasting vibration prediction method based on neural networks according to claim 3, characterized in that, Using neural networks as blasting vibration prediction models includes: using BP neural networks as blasting vibration prediction models.

5. The blasting vibration prediction method based on neural networks according to claim 1, characterized in that, Error function for constructing a blasting vibration prediction model for: ； in, This represents the actual output value of the k-th neuron in the output layer. Let represent the expected output value of the k-th neuron in the output layer, where k = 1, 2, ..., M, and M represents the total number of neurons in the output layer of the blasting vibration prediction model. Represents the natural constant. This represents the steepness factor, j=1,2,…,L, where L represents the total number of neurons in the hidden layer. Indicates the intermediate coefficient. This represents the weight between the j-th neuron in the hidden layer and the k-th neuron in the output layer. This represents the output of the j-th neuron in the hidden layer of the blasting vibration prediction model. This represents the first threshold corresponding to the k-th neuron in the output layer.

6. The blasting vibration prediction method based on neural networks according to claim 1, characterized in that, The update method is obtained by using the roulette wheel algorithm, which includes: setting the probability of self-update to 0.5, generating a random number r in (0,1), and determining whether the random number r is less than 0.

5. If it is, self-update is performed; otherwise, guided update is performed.

7. The blasting vibration prediction method based on neural networks according to claim 1, characterized in that, A random walk is performed on each individual in the population after the initial update, and the individual after the walk is determined according to the error function, resulting in the population after the second update, including: Randomly select an individual to obtain the target individual; Update the p-th dimension weights of the target individual: ； in, This represents the updated p-th dimension weight. This represents the weight of the p-th dimension before the update. Represents pi (π). , This represents the total dimension of the target individual. Indicates the walking stride length; Determine whether the error function value corresponding to the target individual before the p-th dimension weight update is less than the error function value corresponding to the target individual before the update. If so, accept the update; otherwise, reject the update. The weights of all dimensions of the target individual are updated to obtain the updated target individual; The target individuals are reselected in the population for updating, and this process is repeated until all individuals have been updated, resulting in a population that has undergone a second update.

Images