Blasting vibration peak velocity prediction method fusing parameter uncertainty and data driving optimization

Abstract

The invention provides a blasting vibration peak velocity prediction method fusing parameter uncertainty and data-driven optimization, which comprises the following steps of: firstly, acquiring data such as blasting parameters, lithologic indexes, geological conditions and actually measured vibration peak velocity (PPV), and establishing a basic database; a prior probability model is constructed, a joint uncertainty model is established in combination with probability disturbance and a fuzzy triangular number, a multi-dimensional disturbance sample is generated through a joint central value and a joint standard deviation and is fused with original data, and an extended database is formed. Feature analysis is carried out on the fused data, and input variables which have obvious influence on the vibration peak velocity (PPV) are screened; and constructing a BP neural network model based on the screened features, carrying out global optimization on a network weight and a threshold by adopting a PSO algorithm, and then carrying out local fine tuning by utilizing Adam. Finally, through training and verification, prediction of the vibration peak velocity (PPV) is realized, and model precision is evaluated through RMSE, MAE, MAPE, Rand other indexes. The influence of rock and soil parameter uncertainty on prediction precision can be effectively processed, and the reliability and applicability of blasting vibration prediction are improved.

Description

A method for predicting peak velocity of blasting vibration by integrating parameter uncertainty and data-driven optimization Technical Field

[0001] This invention belongs to the field of blasting engineering and geotechnical dynamics prediction technology, specifically involving a method for predicting peak velocity of blasting vibration by integrating parameter uncertainty and data-driven optimization. Background Technology

[0002] Blasting vibration prediction is a crucial aspect of blasting safety control and environmental protection, with its core lying in the accurate estimation of peak blasting vibration velocity. Currently, the prediction methods widely used in engineering are mostly based on traditional empirical formulas, such as the Sadovski formula and the Ambraseys-Hendron model. While these models are simple in structure and easy to use, their prediction results often show significant deviations when faced with complex and variable geological conditions, making it difficult to meet the requirements of high-precision and high-reliability blasting design.

[0003] The limitations of traditional models mainly lie in their insufficient consideration of the inherent uncertainties in soil and rock parameters. In actual engineering projects, key parameters such as P-wave velocity, ground stress, and density of soil and rock masses are influenced by various factors, including stratigraphic structure, the degree of joint and fracture development, and water content, exhibiting significant spatial variability. Traditional methods typically simplify these parameters to fixed constants, ignoring their probabilistic distribution characteristics. This leads to poor model adaptability under complex geological conditions and reduced reliability of prediction results. Furthermore, blasting engineering sites have limited monitoring points, and the data exhibits sparse and locally correlated characteristics, making it difficult to guarantee the fitting accuracy of models based on limited samples. On the other hand, parameters such as vibration attenuation coefficients and propagation exponents in empirical formulas often rely on traditional optimization methods such as least squares. These methods are prone to getting trapped in local optima and struggle to achieve global optimization in the parameter space, further limiting the predictive performance of the model.

[0004] To overcome the aforementioned shortcomings, existing research has attempted to introduce more complex machine learning models for improvement. Patents CN118228592A and CN118332416B employ improved RNN and SVM models, respectively, to predict peak velocities of blasting vibrations. RNNs are more suitable for typical time-series data, but their adaptability to event-driven, discontinuous sequence features like blasting vibrations is limited, and their training is susceptible to gradient vanishing, resulting in poor stability. While SVMs possess good generalization ability, their kernel function selection is limited when dealing with strongly coupled nonlinear features of multiple variables, leading to low optimization efficiency and difficulty in obtaining the global optimum.

[0005] Another patent, CN117252236A, uses a DBN-LSTM-BWOA combined model to improve the prediction performance of blasting vibrations. However, this method also relies on a deep network structure and optimization algorithm, making it sensitive to data quality and insufficient in handling fluctuations in input parameters and the diversity of on-site geological conditions. Furthermore, none of the above methods fully consider the inherent uncertainties of blasting site parameters, such as P-wave velocity, differences in joint and fracture development, and measurement errors. This results in significant deficiencies in the adaptability and robustness of the models under complex working conditions, limiting the reliability of engineering applications. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting peak velocity of blasting vibration that integrates parameter uncertainty and data-driven optimization, aiming to solve the problems in existing peak velocity prediction methods, such as insufficient consideration of the random fluctuation of soil and rock parameters, difficulty in global optimization of neural network model parameters, and low reliability of prediction results.

[0007] To achieve the aforementioned technical features, the present invention aims to provide a method for predicting peak velocity (PPV) of blasting vibration by integrating parameter uncertainty and data-driven optimization, comprising the following steps: Step 1, collecting blasting parameters and geological parameters, preprocessing the original parameters to form a basic database for training; Step 2, quantitatively characterizing the natural fluctuations and cognitive uncertainties of key soil and rock parameters, establishing prior probability models for key parameters, and constructing joint uncertainty perturbation samples using joint uncertainty modeling; Step 3, fusing the joint uncertainty perturbation samples obtained in Step 2 with the basic database from Step 1 to construct an extended database; based on the extended database, using principal component analysis or correlation analysis, extracting key input features that significantly affect PPV, and removing redundant variables; Step 4, using the selected key input features as input and PPV as output, constructing a prediction model for PPV of blasting vibration using a feedforward BP neural network; Step 5, encoding the weights and thresholds of the BP neural network as particle positions, and using a particle swarm optimization (PSO) algorithm for global search optimization to obtain initial parameters with better performance; Step 6, model training and performance verification.

[0008] Preferably, in step one, the blasting parameters and geological parameters include multiple parameters such as explosive charge, blast center distance, charge structure parameters, lithological index, longitudinal wave velocity, ground stress, and field-measured peak velocity (PPV); the preprocessing includes removing duplicate values, handling missing values, noise filtering, and standardization of the original data to ensure data quality.

[0009] Preferably, in step two, the specific process of joint uncertainty modeling is as follows: 1) Probabilistic perturbation modeling: ;in, These are the original input parameters; Parameters after modeling probabilistic perturbations; For the random error of the parameters; The mean is 0 and the variance is The normal distribution; 1) Standard deviation of the original data; 2) Fuzzy trigonometric number modeling: Where a is the minimum possible value, m is the most likely value, and b is the maximum possible value; 3) Calculate the fuzzy mean value; : Furthermore, the fuzzy mean is used as the center value of the normal distribution to achieve a fusion model of fuzzy and random elements.

[0010] Preferably, in step two, the detailed steps for generating the joint uncertainty perturbation sample are as follows: 1) First, determine the originally acquired input geological parameters; 2) Obtain the fuzzy center value according to step two. 3) Construct a joint central value, and use a trade-off approach to merge the two into a single joint central value: ;in: As weight, These are joint central values ​​used to generate perturbation samples. 4) Construct the joint standard deviation; natural standard deviation. : ;in: The standard deviation of the original data; the standard deviation of the fuzzy interval. : Joint standard deviation Combine using the square root of the sum of squares: ; Incorporate data fluctuations and ambiguity intervals into the final disturbance amplitude; 5) Generate joint uncertainty samples; For each parameter, use the joint central value The mean is expressed as the joint standard deviation. Normal perturbation is applied to the standard deviation: ;in, The mean is 0 and the variance is The normal distribution; For joint standard deviation; For the random error of the parameters; For the generated perturbation samples; multiple perturbation samples are generated in this way: 6) Generate multidimensional joint samples; input parameters are: ;in, For the maximum single-shot dose, For the distance between the centers, For longitudinal wave velocity, For geostress, Given the elevation difference, n sets of disturbance samples are generated for each parameter following the steps described above. These samples are then combined to obtain the expanded sample set. ; Finally, an n-fold expanded dataset is formed; 7) It is merged with the original sample to form a joint uncertainty perturbation sample.

[0011] Preferably, in step four, the feedforward BP neural network consists of an input layer, several hidden layers, and an output layer. The number of nodes in the input layer corresponds to the selected blasting influencing factors. The size of the hidden layers is determined by cross-validation to ensure the model's expressive power and generalization performance. The output layer is set to a single node to provide a predicted value of peak velocity (PPV). The model training process is based on the error backpropagation mechanism. By iteratively correcting the network connection weights and bias parameters, the loss function gradually converges to a minimum, thereby achieving effective learning of the input-output mapping relationship.

[0012] Preferably, step five specifically includes: introducing a particle swarm optimization algorithm to perform a global search for the initial weights and thresholds of the BP neural network; finding the optimal initial values ​​in the search space through a position update and velocity adjustment mechanism to minimize the final training loss of the BP model; and introducing an early stopping strategy to prevent overfitting and improve prediction accuracy and stability.

[0013] Preferably, the initial weights and thresholds of the neural network are mapped to particle position vectors, and the fitness function is defined as mean squared error. ;in, This represents the true peak vibration velocity. The peak velocity is the predicted value; the particle swarm update rule is: the update formula for the d-th dimension velocity and position of particle i is: ; Where k is the current iteration number, d = 1, 2, ..., D; i = 1, 2, ..., n; Let c1 be the velocity of particle i in the kth iteration; c1 and c2 are acceleration constants, used to adjust the maximum learning step size. This represents the position of particle i in the k-th iteration. This represents the optimal position experienced by the i-th particle. Inertia factor; The best position experienced by the population; r1 and r2 are random numbers in the interval (0,1), and inertia weights. Dynamically adjusted to: Once PSO converges, Adam is used to fine-tune the optimal weights locally to obtain the final optimized model.

[0014] in The inertial weight decreases linearly with iteration. By adding dynamic inertial weight and adaptive learning factor, PSO can achieve faster convergence speed while searching for the global optimum.

[0015] Preferably, the PSO parameters are set as follows: particle number N p =30~50, maximum number of iterations Learning factor The formula for the linearly decreasing inertia weight is: ;in, The inertia weights are dynamically varying with the number of iterations t. This represents the number of iterations.

[0016] Preferably, step six specifically includes: training the BP neural network optimized in step five on the original data and the uncertainty-corrected data respectively, and finally outputting the predicted value based on the training results, and comparing the performance using the root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²) as evaluation indicators; wherein RMSE directly reflects the average deviation, MAE is not sensitive to outliers and is more robust, and R² measures the goodness of fit of the model, and the closer the value is to 1, the better the prediction effect.

[0017] The present invention has the following advantages: 1. It takes into account the uncertainty of blasting site parameters, and the prediction results are more reliable.

[0018] By jointly modeling the random disturbances and fuzzy uncertainties of soil and rock parameters, the true fluctuation characteristics of factors such as rock mass properties, in-situ stress, and blasting environment are fully reflected, avoiding the systematic bias caused by the assumption of parameter determinism in traditional methods.

[0019] 2. Dynamic correction and data augmentation of blasting parameters were achieved, improving the model's generalization ability.

[0020] A joint uncertainty modeling method combining probabilistic perturbation and fuzzy trigonometric numbers is adopted. A joint distribution model is constructed based on the natural fluctuations and cognitive uncertainties of soil and rock parameters. Monte Carlo sampling is used to generate a corrected sample set containing physical laws, which significantly expands the amount of training data, solves the problem of sparse original data, and improves the generalization ability and adaptability of the model under different blasting areas and complex geological conditions.

[0021] 3. By integrating particle swarm optimization (PSO) with backpropagation (BP) neural networks, the global search and convergence performance of the model is enhanced. PSO performs global optimization of the initial weights and biases of the network, which can effectively prevent the BP algorithm from getting stuck in local optima; at the same time, the introduction of an inertial weight dynamic adjustment mechanism makes the optimization process more stable and the model training speed faster.

[0022] 4. Outputs prediction results including confidence intervals and has uncertainty quantification function.

[0023] This invention performs uncertainty propagation on posterior samples during the prediction phase, outputting the mean, standard deviation, and 95% confidence interval of peak velocities. This makes the prediction results not only meaningful as point values, but also provides a reliable basis for judging the risk range in blasting safety design.

[0024] 5. Prediction accuracy is significantly improved, and the verification results are stable and reliable.

[0025] In practical engineering data validation, the RMSE, MAE, and MAPE of the method of this invention are all superior to traditional empirical formulas and BP models that do not consider uncertainties. Furthermore, its confidence interval coverage is close to the theoretical 95%, demonstrating the accuracy and robustness of the prediction results. Attached Figure Description

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] Figure 1 is a flowchart of the particle swarm optimization algorithm.

[0028] Figure 2 is a flowchart of the model.

[0029] Figure 3 is a flowchart of the uncertainty modeling principle.

[0030] Figure 4 is a schematic diagram of modeling uncertainty in geotechnical parameters.

[0031] Figure 5 shows a comparison of the model effects. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0033] Example 1: This invention aims to improve traditional prediction models. The core improvement lies in quantifying the spatial variability of key soil and rock parameters and testing errors as uncertainty factors, and introducing advanced intelligent optimization algorithms. Specifically, this invention integrates uncertainty analysis of soil and rock parameters with particle swarm optimization algorithms. By performing probabilistic perturbation modeling and intelligent optimization of key geological and blasting parameters, a peak velocity prediction model for blasting vibration is constructed that includes both physical mechanism constraints and probabilistic descriptive capabilities. This method can effectively characterize the spatial fluctuation of soil and rock parameters, and utilizes the global search advantage of particle swarm optimization to avoid the problem of traditional optimization strategies easily getting trapped in local optima. This achieves high-precision and robust prediction of peak velocity, providing more reliable technical support for the safety design and environmental impact assessment of blasting projects.

[0034] This invention discloses a method for predicting peak velocity (PV) of blasting vibration by integrating parameter uncertainty and data-driven optimization. First, blasting parameters and geological parameters are collected to establish a prior probability model of the input variables. Then, the natural fluctuations and cognitive uncertainties of key soil and rock parameters are quantitatively characterized, and an extended uncertainty sample set is constructed based on a probabilistic perturbation model and fuzzy triangular number theory. The extended sample is then fused with the original data, and effective input parameters for PV prediction are selected through feature sensitivity analysis. Based on this, a backpropagation (BP) neural network prediction model is constructed, and a particle swarm optimization algorithm is used to globally optimize the initial weights and thresholds of the network to improve the convergence efficiency and optimization capability of the network training. Finally, the model is trained and validated, outputting the PV prediction results and corresponding performance evaluation indicators. Through these steps, this invention can effectively enhance the stability and accuracy of blasting vibration PV prediction, and improve the robustness and engineering applicability of the model under complex geological conditions.

[0035] Example 2: A method for predicting peak velocity (PPV) of blasting vibration by integrating parameter uncertainty and data-driven optimization, comprising the following steps: Step 1, collecting blasting parameters and geological parameters, preprocessing the original parameters to form a basic database for training; Step 2, quantitatively characterizing the natural fluctuations and cognitive uncertainties of key soil and rock parameters, establishing prior probability models for key parameters, and constructing joint uncertainty perturbation samples using joint uncertainty modeling; Step 3, fusing the joint uncertainty perturbation samples obtained in Step 2 with the basic database in Step 1 to construct an extended database; based on the extended database, using principal component analysis or correlation analysis, extracting key input features that significantly affect PPV, and removing redundant variables; Step 4, using the selected key input features as input and PPV as output, constructing a prediction model for PPV of blasting vibration using a feedforward BP neural network; Step 5, encoding the weights and thresholds of the BP neural network as particle positions, and using a particle swarm optimization (PSO) algorithm for global search optimization to obtain initial parameters with better performance; Step 6, model training and performance verification.

[0036] Preferably, in step one, the blasting parameters and geological parameters include multiple parameters such as explosive charge, blast center distance, charge structure parameters, lithological index, longitudinal wave velocity, ground stress, and field-measured peak velocity (PPV); the preprocessing includes removing duplicate values, handling missing values, noise filtering, and standardization of the original data to ensure data quality.

[0037] Preferably, in step two, the specific process of joint uncertainty modeling is as follows: 1) Probabilistic perturbation modeling: ;in, These are samples of the original input parameters; The parameters to be included with probability perturbations; For the random error of the parameters; The mean is 0 and the variance is The normal distribution; 1) Standard deviation of the original sample; 2) Fuzzy triangular number modeling: Where a is the minimum possible value, m is the most likely value, and b is the maximum possible value; 3) Calculate the fuzzy mean value; : Furthermore, the fuzzy mean is used as the center value of the normal distribution to achieve a fusion model of fuzzy and random elements.

[0038] Preferably, in step two, the detailed steps for generating the joint uncertainty perturbation sample are as follows: 1) First, determine the originally acquired input geological parameters; 2) Obtain the fuzzy center value according to step two. 3) Construct a joint central value, and use a trade-off approach to merge the two into a single joint central value: ;in: As weight, 4) Construct the joint central value to generate perturbed samples; 5) Construct the joint standard deviation; natural standard deviation : ;in: The original sample standard deviation; the fuzzy interval standard deviation. : ;in: Standard deviation for fuzzy intervals; joint standard deviation Combine using the square root of the sum of squares: ; Incorporate data fluctuations and ambiguity intervals into the final disturbance amplitude; 5) Generate joint uncertainty samples; For each parameter, use the joint central value The mean is expressed as the joint standard deviation. Normal perturbation is applied to the standard deviation: ;in, The mean is 0 and the variance is The normal distribution; For joint standard deviation; For the random error of the parameters; For the generated perturbation samples; multiple perturbation samples are generated in this way: 6) Generate multidimensional joint samples; input parameters are: ;in, For the maximum single-shot dose, For the distance between the centers, For longitudinal wave velocity, For geostress, Given the elevation difference, n sets of disturbance samples are generated for each parameter following the steps described above. These samples are then combined to obtain the expanded sample set. ; Finally, an n-fold expanded dataset is formed; 7) It is merged with the original sample to form a joint uncertainty perturbation sample.

[0039] Preferably, in step four, the feedforward BP neural network consists of an input layer, several hidden layers, and an output layer. The number of nodes in the input layer corresponds to the selected blasting influencing factors. The size of the hidden layers is determined by cross-validation to ensure the model's expressive power and generalization performance. The output layer is set to a single node to provide a predicted value of peak velocity (PPV). The model training process is based on the error backpropagation mechanism. By iteratively correcting the network connection weights and bias parameters, the loss function gradually converges to a minimum, thereby achieving effective learning of the input-output mapping relationship.

[0040] Preferably, step five specifically includes: introducing a particle swarm optimization algorithm to perform a global search for the initial weights and thresholds of the BP neural network; finding the optimal initial values ​​in the search space through a position update and velocity adjustment mechanism to minimize the final training loss of the BP model; and introducing an early stopping strategy to prevent overfitting and improve prediction accuracy and stability.

[0041] Preferably, the initial weights and thresholds of the neural network are mapped to particle position vectors, and the fitness function is defined as mean squared error. ;in, This represents the true peak vibration velocity. The peak velocity is the predicted value; the particle swarm update rule is: the update formula for the d-th dimension velocity and position of particle i is: ; Where k is the current iteration number, d = 1, 2, ..., D; i = 1, 2, ..., n; Let c1 be the velocity of particle i in the kth iteration; c1 and c2 are acceleration constants, used to adjust the maximum learning step size. This represents the position of particle i in the k-th iteration. Let be the best position experienced by the i-th particle; Inertia factor; The best position experienced by the population; r1 and r2 are random numbers in the interval (0,1), with inertia weights. Dynamically adjusted to: Once PSO converges, Adam is used to fine-tune the optimal weights locally to obtain the final optimized model.

[0042] in The inertial weight decreases linearly with iteration. By adding dynamic inertial weight and adaptive learning factor, PSO can achieve faster convergence speed while searching for the global optimum.

[0043] Preferably, the PSO parameters are set as follows: particle number N p =30~50, maximum number of iterations Learning factor The formula for the linearly decreasing inertia weight is: ;in, Let be the inertia weight that changes dynamically with the number of iterations t. This represents the number of iterations.

[0044] Preferably, step six specifically includes: training the BP neural network optimized in step five on the original data and the uncertainty-corrected data respectively, and finally outputting the predicted value based on the training results, and comparing the performance using the root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²) as evaluation indicators; wherein RMSE directly reflects the average deviation, MAE is not sensitive to outliers and is more robust, and R² measures the goodness of fit of the model, and the closer the value is to 1, the better the prediction effect.

[0045] Example 3: To verify the applicability and reliability of the method of the present invention, an example is given for a blasting project of a hydropower station.

[0046] (1) Project Overview: The lithology of the study area is mainly columnar basalt with well-developed jointed and fractured structures. Faults and weathering zones of varying degrees also exist in some local locations. Blasting operations adopted a step-by-step excavation method, with multiple vibration monitoring points set up. A total of 107 sets of measured blasting vibration data were collected on site, including blasting parameters and vibration response.

[0047] (2) Data Acquisition and Parameter Setting: Six training parameters were mainly collected, including detonation distance (R), maximum single-explosive charge (Q), and longitudinal wave velocity (C). p ), geostress ( It has five input parameters: elevation difference (H) and one output parameter: peak vibration velocity.

[0048] Table 1 Original Parameters of a Certain Project

[0049] Step 1: Raw data collection.

[0050] Step 2: Use the samples in the case study to model the uncertainty of geotechnical parameters. Considering the two sources of uncertainty, fuzziness and randomness, new geotechnical parameters are obtained by jointly correcting the two uncertainties and updating them to the database.

[0051] Here, we take a set of the above data to demonstrate uncertainty modeling: the fluctuation range of the longitudinal wave velocity in the collected data sample is ±150m / s.

[0052] Original value: Natural fluctuation standard deviation (empirical value): Fuzzy trigonometric numbers (based on experience and geological stratification): ; Calculate the fuzzy mean Joint central value (with weight α=0.5); ; Standard deviation of fuzzy intervals: Joint standard deviation: Generate perturbation samples: Uncertainty modeling of geostress: Given a large error in stress testing, a fuzzy model is used: Original values: Fuzzy trigonometric numbers: Natural fluctuation standard deviation (empirical): Fuzzy mean: Joint Center: ; Standard deviation of fuzzy intervals: Joint standard deviation: Perturbation samples (3 examples): ; ; Step 3: Construct joint uncertainty samples; perform uncertainty modeling on the original collected data to obtain approximately 4950 samples.

[0053] Table 2 shows the data obtained from uncertainty modeling.

[0054] Step 4: Perform principal component analysis on the updated database to determine the input parameters (considering that the actual collected parameters are generally more than those in this case or are different from those in this case). The purpose is to reduce data dimensionality, determine that the input parameters and output parameters are highly correlated, and then train the model to improve prediction efficiency and enhance model stability.

[0055] Step 5: Based on the above analysis, establish a neural network prediction model with a 5-12-1 three-layer network structure. The hidden layers use the Sigmoid activation function, the optimizer uses Adam, the output layer is a linear function, and the loss function is the mean squared error (MSE), with the goal of minimizing prediction bias.

[0056] Step Six: Add the particle optimization algorithm to the neural network model. This algorithm adopts a global optimization approach and achieves a balance between convergence speed and search breadth by introducing dynamic inertia weights and adaptive learning factors. Adding the particle optimization algorithm means finding the parameters that output the optimal result within the searchable range. At the same time, an early stopping mechanism is added to prevent overfitting.

[0057] Step 7: Train the model using both the data before and after replacing the geotechnical parameters, and output the results. Compare the results obtained.

[0058] By comparing the model results trained using the original data with those trained using data after uncertainty modeling, it is evident that the method of this invention significantly improves prediction accuracy. Specifically, when trained using the original data, the model's MAPE, RMSE, and MAE on the test set were 0.3952, 1.6631, and 1.2259, respectively, with a determination coefficient R² of 0.7599. However, after introducing uncertainty modeling to perturb and expand the blasting parameters and retraining, the MAPE, RMSE, and MAE decreased to 0.1519, 0.1396, and 0.1301, respectively, and the R² increased to 0.9291. This demonstrates that uncertainty modeling effectively addresses the problems of sparse original samples, high noise, and poor stability, enabling the model to acquire more comprehensive feature learning capabilities, thereby significantly improving the accuracy and robustness of blasting vibration peak velocity prediction.

[0059] Although the preferred embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many specific modifications under the guidance of the present invention without departing from the spirit of the invention and the scope of protection of the claims, and these modifications all fall within the scope of protection of the present invention.

Claims

1. A method for predicting peak velocity of blasting vibration by integrating parameter uncertainty and data-driven optimization, characterized in that, Includes the following steps: Step 1: Collect blasting and geological parameters, preprocess the raw parameters to form a basic database for training. Step 2: Quantitatively characterize the natural fluctuations and cognitive uncertainties of key soil and rock parameters, establish prior probability models for key parameters, and construct joint uncertainty perturbation samples using joint uncertainty modeling. Step 3: Merge the joint uncertainty perturbation samples obtained in Step 2 with the basic database from Step 1 to construct an extended database. Based on the extended database, use principal component analysis or correlation analysis to extract key input features that significantly affect peak velocity (PPV) and remove redundant variables. Step 4: Using the selected key input features as input and PPV as output, construct a prediction model for blasting peak velocity using a feedforward BP neural network. Step 5: Encode the weights and thresholds of the BP neural network as particle positions and use a particle swarm optimization (PSO) algorithm for global search optimization to obtain better initial parameters. Step 6: Model training and performance verification.

2. The method for predicting peak velocity of blasting vibration by fusing parameter uncertainty and data-driven optimization according to claim 1, characterized in that, In step one, the blasting parameters and geological parameters include blasting charge, blast center distance, charge structure parameters, lithological indicators, longitudinal wave velocity, ground stress, and field-measured peak velocity (PPV) multidimensional parameters; the preprocessing includes removing duplicate values, handling missing values, noise filtering, and standardization of the raw data to ensure data quality.

3. The method for predicting peak velocity of blasting vibration by fusing parameter uncertainty and data-driven optimization according to claim 1, characterized in that, In step two, the specific process of joint uncertainty modeling is as follows: 1) Probabilistic perturbation modeling: ;in, These are the original input parameters; Parameters after modeling probabilistic perturbations; For the random error of the parameters; The mean is 0 and the variance is The normal distribution; 1) Standard deviation of the original data; 2) Fuzzy trigonometric modeling: Where a is the minimum possible value, m is the most likely value, and b is the maximum possible value; 3) Calculate the fuzzy mean value; ： The fuzzy mean is used as the center value of the normal distribution to achieve a fusion model of fuzzy and randomness.

4. The method for predicting peak velocity of blasting vibration by fusing parameter uncertainty and data-driven optimization according to claim 1, characterized in that, In step two, the detailed steps for generating the joint uncertainty perturbation sample are as follows: 1) First, determine the original input geological parameters; 2) Obtain the fuzzy center value according to step two. 3) Construct a joint central value, and use a trade-off approach to merge the two into a single joint central value: ;in: As weight, These are joint central values ​​used to generate perturbation samples. 4) Construct the joint standard deviation; natural standard deviation. ： ;in: Standard deviation of the original data; standard deviation of the fuzzy interval ： Joint standard deviation Combine using the square root of the sum of squares: ; Incorporate data fluctuations and ambiguity intervals into the final disturbance amplitude; 5) Generate joint uncertainty samples; For each parameter, use the joint central value The mean is expressed as the joint standard deviation. Normal perturbation is applied to the standard deviation: in, The mean is 0 and the variance is The normal distribution; For joint standard deviation; For the random error of the parameters; For the generated perturbation samples; multiple perturbation samples are generated in this way: 6) Generate multidimensional joint samples; input parameters are: ;in, For the maximum single-shot dose, For the distance between the centers, For longitudinal wave velocity, For geostress, Given the elevation difference, n sets of disturbance samples are generated for each parameter following the steps described above. These samples are then combined to obtain the expanded sample set. ; Finally, an n-fold expanded dataset is formed; 7) It is merged with the original sample to form a joint uncertainty perturbation sample.

5. The method for predicting peak velocity of blasting vibration by fusing parameter uncertainty and data-driven optimization according to claim 1, characterized in that, In step four, the feedforward BP neural network consists of an input layer, several hidden layers, and an output layer. The number of nodes in the input layer corresponds to the selected blasting influencing factors. The size of the hidden layers is determined by cross-validation to ensure the model's expressive power and generalization performance. The output layer is set to a single node to provide the predicted value of peak vibration volume (PPV). The model training process is based on the error backpropagation mechanism. By iteratively correcting the network connection weights and bias parameters, the loss function gradually converges to a minimum, thereby achieving effective learning of the input-output mapping relationship.

6. The method for predicting peak velocity of blasting vibration by fusing parameter uncertainty and data-driven optimization according to claim 1, characterized in that, Step five specifically includes: introducing a particle swarm optimization algorithm to perform a global search for the initial weights and thresholds of the BP neural network; finding the optimal initial values ​​in the search space through a position update and velocity adjustment mechanism to minimize the final training loss of the BP model; and introducing an early stopping strategy to prevent overfitting and improve prediction accuracy and stability.

7. The method for predicting peak velocity of blasting vibration by fusing parameter uncertainty and data-driven optimization according to claim 6, characterized in that, The initial weights and threshold of the neural network are mapped to particle position vectors, and the fitness function is defined as mean squared error. ;in, This represents the true peak vibration velocity. The peak velocity is the predicted value; the particle swarm update rule is: the update formula for the d-th dimension velocity and position of particle i is: ； Where k is the current iteration number, d = 1, 2, ..., D; i = 1, 2, ..., n; Let c1 be the velocity of particle i in the kth iteration; c1 and c2 are acceleration constants, used to adjust the maximum learning step size. This represents the position of particle i in the k-th iteration; This represents the optimal position experienced by the i-th particle. Inertia factor; The best position experienced by the population; r1 and r2 are random numbers in the interval (0,1), and inertia weights. Dynamically adjusted to: Once PSO converges, Adam is used to fine-tune the optimal weights locally to obtain the final optimized model; where The inertial weight decreases linearly with iteration. By adding dynamic inertial weight and adaptive learning factor, PSO can achieve faster convergence speed while searching for the global optimum.

8. The method for predicting peak velocity of blasting vibration by fusing parameter uncertainty and data-driven optimization according to claim 1, characterized in that, Set the PSO parameters as follows: Particle number N p =30~50, maximum number of iterations Learning factor The formula for the linearly decreasing inertia weight is: ;in, The inertia weights are dynamically varying with the number of iterations t. This represents the number of iterations.

9. The method for predicting peak velocity of blasting vibration by fusing parameter uncertainty and data-driven optimization according to claim 1, characterized in that, Step six specifically includes: training the optimized BP neural network from step five on both the original data and the uncertainty-corrected data, and finally outputting the predicted value based on the training results. The performance is compared using the root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²) as evaluation metrics. Among these, RMSE directly reflects the average deviation, MAE is not sensitive to outliers and is more robust, and R² measures the goodness of fit of the model, with a value closer to 1 indicating better prediction performance.

Images