Engineering blasting vibration peak velocity prediction method based on AI deep learning

By constructing a dynamic fusion mechanism between the initial physical prediction model and the AI ​​deep learning model, and combining the particle swarm optimization algorithm and the core mapping prediction method, the problems of low prediction accuracy and poor adaptability in the existing technology are solved, and high-precision and adaptive peak velocity prediction of blasting vibration is achieved, supporting intelligent management of large-scale engineering blasting.

CN123115665APending Publication Date: 2026-11-17SHENZHEN TIEKE INSPECTION & TESTING ENG CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611297453.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-25
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing peak velocity prediction methods for blasting vibrations struggle to balance physical mechanisms with complex nonlinear data characteristics, resulting in low prediction accuracy, poor adaptability, and a lack of effective utilization of the similarity and correlation of regional blasting conditions.

Method used

A dynamic fusion mechanism is constructed between the initial physical prediction model and the AI ​​deep learning model. The deep learning fusion strategy is triggered by the prediction accuracy evaluation mechanism. The fusion weight is optimized by combining the particle swarm optimization algorithm. The core mapping prediction method is adopted in similar blasting condition groups to realize a multi-level prediction system.

Benefits of technology

It significantly improves the prediction accuracy of peak vibration velocity under complex geological and blasting conditions, enhances the adaptability and robustness of the prediction system, and improves prediction accuracy and generalization ability, providing reliable technical support for the intelligent and refined safety management of blasting vibration in large-scale engineering projects.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115665A_ABST
    Figure CN123115665A_ABST
Patent Text Reader

Abstract

This invention relates to the field of engineering blasting technology, and more particularly to a method for predicting peak velocity of vibration in engineering blasting based on AI deep learning. This method constructs an initial physical prediction model based on basic engineering data and performs predictions. The accuracy of the prediction is evaluated by comparing it with actual monitoring data, and a fusion mechanism is triggered when the accuracy is insufficient. A data-driven model is constructed using AI deep learning algorithms, and this model is weighted and fused with the physical model to form a fused prediction model. The optimal prediction model is dynamically selected through weight analysis. The blasting area is divided into several sub-regions. When the proportion of fused models used within a region exceeds a threshold, a clustering algorithm is used to select similar working condition groups. Within each group, a core mapping method is used to achieve group collaborative prediction. This method effectively integrates the advantages of physical mechanisms and data-driven approaches, significantly improving prediction accuracy. Simultaneously, the regional collaborative strategy improves prediction efficiency, making it suitable for vibration safety control in complex blasting environments.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of engineering blasting technology, and in particular to a method for predicting peak velocity of vibration in engineering blasting based on AI deep learning. Background Technology

[0002] With the continuous advancement of infrastructure construction and mineral resource development, engineering blasting is widely used as an important means of rock breaking and earthwork excavation. However, the vibrations generated during blasting may adversely affect surrounding buildings, underground pipelines, rock mass stability, and residents' lives, and in severe cases, may even cause structural damage or safety accidents. Therefore, accurately predicting the peak velocity (PPV) of blasting vibration is of great significance for assessing the safety impact of blasting, optimizing blasting design, and implementing effective vibration control.

[0003] Traditional blasting vibration prediction mainly relies on empirical formulas, such as the Sadovsky formula. These methods establish physical relationship models based on finite parameters such as the detonation center distance and charge amount, offering the advantage of computational simplicity. However, under complex geological conditions and variable blasting environments, they often fail to fully reflect the actual vibration propagation laws, resulting in limited prediction accuracy. In recent years, with the development of artificial intelligence technology, some studies have begun to introduce machine learning methods, such as support vector machines and random forests, to improve prediction capabilities by learning from historical monitoring data. However, these methods still have limitations in handling nonlinear and time-series blasting vibration data, and most models operate independently, failing to effectively combine the advantages of physical mechanisms and data-driven approaches. Furthermore, existing prediction methods often model individual blasting points independently, lacking the ability to explore the similarities and correlations of blasting conditions across regions, making it difficult to achieve multi-point collaborative prediction and model transfer applications. Especially in large mining areas or linear projects, where blasting areas are wide and geological conditions vary greatly, there is an urgent need for an intelligent prediction method that can adaptively adjust prediction strategies, integrate physical mechanisms and deep learning capabilities, and utilize group blasting characteristics to improve overall prediction accuracy.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a method for predicting peak velocity of blasting vibration based on AI deep learning. This method aims to solve the technical problems of existing methods for predicting peak velocity of blasting vibration, which are difficult to take into account both physical mechanisms and complex nonlinear data characteristics, and generally suffer from low prediction accuracy, poor adaptability, and lack of effective utilization of the similarity and correlation of regional blasting conditions.

[0006] To achieve the above objectives, this invention provides a method for predicting peak velocity of engineering blasting vibration based on AI deep learning, the method comprising: Obtain basic data for engineering blasting projects and construct an initial physical prediction model for blasting vibrations; The peak velocity of blasting vibration is predicted using an initial physical prediction model, and the accuracy of the prediction is evaluated by comparing the blasting parameter sequence obtained from actual monitoring with the corresponding actual peak velocity sequence. If the accuracy is low, the AI ​​deep learning prediction method is triggered. If the fusion AI deep learning prediction method is triggered, a deep learning prediction model is constructed by combining the actual blasting parameter sequence and the actual peak vibration velocity sequence through the AI ​​deep learning algorithm. The deep learning prediction model and the initial physical prediction model are weighted and fused to construct the fusion prediction model for engineering blasting. The prediction model used for the current blasting condition is determined according to the weight. The engineering blasting area is divided into several blasting sub-regions. It is determined whether the blasting sub-regions trigger the similarity set prediction method. If not, the peak vibration velocity of each blasting point is predicted according to the prediction model adopted. If it is triggered, similar blasting condition groups are selected by clustering algorithm. Within the similar blasting condition groups, the core mapping prediction method is used to predict the peak vibration velocity of blasting.

[0007] Optionally, the method for evaluating the prediction accuracy is as follows: Set the prediction period, and the interval between adjacent prediction periods is the model evaluation and update period; For a single or multiple blasting events within each prediction period, the actual blasting parameter sequence of the blasting event is obtained, input into the initial physical prediction model, and the predicted peak vibration velocity is obtained; the actual peak vibration velocity generated by the blasting event is collected in real time. Calculate the average absolute percentage error between the predicted peak vibration velocity and the actual peak vibration velocity. If the average absolute percentage error is greater than or equal to the preset error threshold, the prediction accuracy of the initial physical prediction model is judged to be low.

[0008] Optionally, the prediction model used for the current blasting condition is determined as follows: The prediction model includes a fusion prediction model and an initial physics prediction model; For each blasting condition that triggers the fusion AI deep learning prediction method, a weight evaluation flag value is set and initialized to 0. The weight of the initial physical prediction model in the fusion prediction model is compared with the preset weight threshold. If the weight is greater than or equal to the weight threshold, the weight evaluation flag value is incremented once. If the weight is less than the weight threshold, the weight evaluation flag value is reset to 0. If the weight evaluation flag value is less than the stability flag value, the current blasting condition adopts the fusion prediction model; otherwise, the current blasting condition ends the fusion AI deep learning prediction method and adopts the initial physical prediction model.

[0009] Optionally, the fusion prediction model is obtained as follows: The deep learning prediction model and the initial physical prediction model are obtained and weighted and fused to construct a fused prediction model, wherein the sum of the weights of the deep learning prediction model and the initial physical prediction model is 1. Using a particle swarm optimization algorithm, a set of initial physical prediction model weights is randomly generated. The historical blasting parameter sequence is obtained and input into the fusion prediction model. The mean absolute percentage error between the prediction result and the corresponding actual vibration peak velocity is calculated. By updating the position and adjusting the velocity of the particles, the particle swarm is continuously optimized. The initial physical prediction model weights that minimize the mean absolute percentage error are gradually selected. The weights of the initial physical prediction model are determined, and the fusion prediction model is obtained.

[0010] Optionally, the deep learning prediction model is obtained as follows: Determine whether a deep learning prediction model exists for the current blasting condition; if not, obtain all actual blasting parameter sequences and corresponding actual peak velocity sequences for this type of blasting condition up to the current moment, integrate them according to the time series to obtain the total blasting parameter sequence and the total peak velocity sequence, obtain the training dataset, construct an LSTM model based on the AI ​​deep learning algorithm and train it using the training dataset, and obtain the trained LSTM model as the deep learning prediction model for this condition; If they exist, obtain the latest actual blasting parameter sequence and the corresponding actual vibration peak velocity sequence that were not included in the training dataset to supplement the training dataset, perform incremental training on the deep learning prediction model, and update the deep learning prediction model for this working condition.

[0011] Optionally, the method for determining whether the similarity set prediction method is triggered is as follows: The engineering blasting area is divided into several blasting sub-areas based on geological conditions, blasting zone range, and blasting type. For each blasting sub-area, the number of blasting conditions using the fusion prediction model is obtained, and the ratio of the number of blasting conditions to the total number of blasting conditions in the blasting sub-area is calculated. If the ratio is greater than the preset ratio standard, the similarity set prediction method is triggered in the blasting sub-area.

[0012] Optionally, the similar blasting condition group is obtained as follows: Divide all blasting conditions within the blasting sub-region into two clusters, calculate the centers of the two clusters and the Euclidean distance between the cluster centers, and compare them with the set condition judgment threshold. If the number of clusters is greater than or equal to the set working condition judgment threshold, each cluster is further divided into two clusters, and the Euclidean distance between the centers of the two clusters is calculated and compared again. If the number of clusters is greater than the preset number standard, the re-clustering and comparison are stopped. If the value is less than the set condition judgment threshold, the two clusters are marked as similar clusters. The normalized vibration characteristic sequence of each blasting condition in the cluster and the Euclidean distance between the cluster center are obtained and calculated. The blasting conditions that are less than the condition judgment threshold are classified into the same similar blasting condition group.

[0013] Optionally, the division into two clusters is as follows: The actual peak vibration velocity sequences and main influencing parameters of several recent blasting operations were obtained and normalized to obtain normalized vibration characteristic sequences. The K-means clustering algorithm was used to cluster all blasting conditions in the blasting sub-region based on the normalized vibration characteristic sequences, dividing all blasting conditions into two clusters.

[0014] Optionally, the core mapping prediction method includes: Obtain the Euclidean distance between the normalized vibration characteristic sequence of each blasting condition within a group of similar blasting conditions and the cluster center; The blasting condition with the smallest Euclidean distance is taken as the core prediction condition of the similar blasting condition group, and the blasting conditions that are not core prediction conditions are taken as the mapping prediction conditions of the similar blasting condition group. For the core predicted condition, its specific blasting parameters are obtained and input into the fusion prediction model of the core predicted condition to obtain the predicted peak vibration velocity of the core predicted condition. Based on the proportional relationship or mapping function between the actual peak vibration velocities of the core predicted condition and the mapped predicted condition under similar historical blasting conditions, the predicted peak vibration velocity of the core predicted condition is mapped to obtain the predicted peak vibration velocity of the mapped predicted condition.

[0015] Furthermore, to achieve the above objectives, the present invention also provides an AI deep learning-based engineering blasting vibration peak velocity prediction device, the device comprising: a memory, a processor, and an AI deep learning-based engineering blasting vibration peak velocity prediction program stored in the memory and executable on the processor, the AI ​​deep learning-based engineering blasting vibration peak velocity prediction program being configured to implement the steps of the AI ​​deep learning-based engineering blasting vibration peak velocity prediction method as described above.

[0016] This invention provides a method for predicting peak velocity of vibration in engineering blasting based on AI deep learning. The method leverages the advantages of the physical model's strong interpretability and the deep learning model's superior nonlinear fitting ability by constructing a dynamic fusion mechanism between an initial physical prediction model and an AI deep learning model, significantly improving the prediction accuracy of peak velocity under complex geological and blasting conditions. By introducing a prediction accuracy evaluation mechanism, intelligent model switching is achieved, automatically triggering a data-driven deep learning fusion strategy when the model's prediction deviation is large, enhancing the predictive system's adaptability and robustness. Furthermore, a weighted fusion strategy combined with particle swarm optimization algorithm is employed for optimization. By incorporating weights, the fusion prediction model can dynamically balance the influence of physical laws and data characteristics. Furthermore, the project area is divided into multiple blasting sub-regions, and clustering analysis and core mapping prediction methods are introduced within eligible sub-regions. This effectively uncovers the correlations between similar blasting conditions, achieving knowledge transfer and efficient prediction based on group characteristics, and avoiding resource waste and data redundancy caused by modeling each blast point individually. The overall method forms a multi-level prediction system guided by physics, driven by data, and coordinated by regions. This not only improves prediction accuracy and generalization ability but also provides reliable technical support for the intelligent and refined safety management of blasting vibrations in large-scale engineering projects. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an embodiment of the method for predicting peak velocity of engineering blasting vibration based on AI deep learning according to the present invention.

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0020] Reference Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the AI ​​deep learning-based method for predicting peak velocity of engineering blasting vibration according to the present invention.

[0021] In one embodiment, the AI ​​deep learning-based method for predicting peak velocity of engineering blasting vibration includes: Step S100: Obtain basic data for the engineering blasting project and construct an initial physical prediction model for blasting vibration.

[0022] The basic data for engineering blasting projects can be a set of structured or semi-structured information describing the basic conditions of the blasting project, which can be used to support the construction of the initial physical prediction model and subsequent deep learning training. In this embodiment, the basic data for engineering blasting projects may include, but is not limited to, one or more of the following: blasting design parameters, geological survey data, and historical monitoring records. The initial physical prediction model can be an analytical model built based on physical mechanisms (such as the Sadovsky formula) to predict the peak velocity of blasting vibration, which can be used to provide preliminary PPV prediction results with physical interpretability as a guiding basis for the fusion system. Furthermore, the initial physical prediction model can be described in context as follows: it is calculated by substituting the basic data of the engineering blasting project (such as the distance from the detonation center, the charge amount, etc.) into empirical physical formulas. For example, the initial physical prediction model may include, but is not limited to, one or more of the following: the Sadovsky-type model, the modified distance attenuation model, and the multi-parameter empirical regression model.

[0023] Obtaining basic data for engineering blasting projects can be achieved by collecting structured data from sources such as engineering design documents, geological reports, and monitoring systems. This provides the necessary input for building the initial physical prediction model. Constructing the initial physical prediction model for blasting vibrations involves substituting the basic data into empirical physical formulas (such as the Sadovsky formula) to establish a PPV prediction expression, thereby generating a benchmark prediction model with physical interpretability.

[0024] Step S200: The initial physical prediction model is used to predict the peak velocity of blasting vibration, and the accuracy of the prediction is evaluated by comparing the blasting parameter sequence obtained from actual monitoring with the corresponding actual peak velocity sequence. If the accuracy is low, the AI ​​deep learning prediction method is triggered.

[0025] The actual blasting parameter sequence can be a time series of measurable or recordable engineering operations and environmental variables for each blasting event, which can be used as model input features for prediction and evaluation. In an exemplary embodiment, the actual blasting parameter sequence may include, but is not limited to, one or more of the following: charge amount time series, hole spacing series, rock mass wave velocity series, etc. The actual peak velocity (PV) sequence can be a sequence of measured PPV values ​​corresponding to each blasting event obtained through on-site sensor monitoring, which can be used as model output labels for training, evaluation, and fusion weight optimization. For example, the actual PV sequence may include, but is not limited to, one or more of the following: vertical PPV sequence, horizontal radial PPV sequence, synthetic PPV sequence, etc.

[0026] The prediction accuracy assessment mechanism can be a judgment logic used to quantify the deviation between the output of the initial physical prediction model and the actual PPV, and can be used to determine whether to trigger the fusion AI deep learning prediction method. Furthermore, the acquisition method of the prediction accuracy assessment mechanism can be explained in context: calculating the error index (such as MAE, RMSE) between the predicted value and the measured value, and comparing it with a threshold. In a specific embodiment, the prediction accuracy assessment mechanism can rely on the actual blasting parameter sequence and the actual peak velocity sequence, and control the initiation of the fusion prediction process. Predicting the peak velocity of blasting vibration using the initial physical prediction model can involve inputting the current blasting parameters into the initial physical prediction model and outputting the PPV prediction value, thereby achieving the technical effect of providing preliminary prediction results for subsequent evaluation.

[0027] The accuracy of predictions is evaluated by comparing the blasting parameter sequences obtained from actual monitoring with the corresponding actual peak velocity sequences. This can involve calculating the error between the predicted value and the measured PPV and comparing it with a preset threshold. This allows for determining whether the current physical model is applicable and deciding whether to activate the fusion mechanism. Triggering the fusion AI deep learning prediction method can be achieved by activating the deep learning modeling process when the prediction accuracy evaluation result is below a threshold, thereby introducing data-driven capabilities to cope with complex working conditions.

[0028] Step S300: If the fusion AI deep learning prediction method is triggered, a deep learning prediction model is constructed by combining the actual blasting parameter sequence and the actual vibration peak velocity sequence through the AI ​​deep learning algorithm. The deep learning prediction model and the initial physical prediction model are weighted and fused to construct a fusion prediction model for engineering blasting. The prediction model used for the current blasting condition is determined according to the weight.

[0029] The AI ​​deep learning prediction model can be a data-driven prediction model trained on historical blasting monitoring data based on a neural network architecture. It can be used to capture the nonlinear and temporal characteristics of vibration propagation under complex geological and blasting conditions, improving fitting accuracy. In this embodiment, the acquisition method of the AI ​​deep learning prediction model can be explained in context: using the actual blasting parameter sequence and the actual peak velocity sequence as input-output pairs, a nonlinear mapping relationship is generated through deep learning algorithm training. For example, the AI ​​deep learning prediction model can include, but is not limited to, one or more of LSTM network models, Transformer temporal models, and graph neural network models. The AI ​​deep learning algorithm can be a neural network training method used to construct the deep learning prediction model, which can be used to automatically learn the complex mapping relationship between blasting parameters and PPV from data. In a specific embodiment, the AI ​​deep learning algorithm can include, but is not limited to, one or more of backpropagation algorithms, self-attention mechanisms, and temporal convolution algorithms.

[0030] The fusion prediction model can be a composite prediction structure composed of an initial physics prediction model and an AI deep learning prediction model through a weighted fusion strategy. It can be used to dynamically balance physical constraints and data feature learning, improving prediction robustness and generalization ability. Furthermore, the acquisition method of the fusion prediction model can be explained in context: the outputs of the two sub-models are linearly or non-linearly combined according to the optimized fusion weights. In an exemplary embodiment, the fusion prediction model can receive the outputs of the initial physics prediction model and the AI ​​deep learning prediction model, be controlled by the fusion weights, and provide the final basis for predicting the explosion point PPV.

[0031] Fusion weights can be coefficients used to adjust the contribution ratio of the initial physical prediction model and the AI ​​deep learning prediction model in the fused prediction. They can be used to dynamically reflect the relative importance of physical laws and data characteristics under the current operating conditions. For example, fusion weights can include, but are not limited to, one or more of static preset weights, operating condition adaptive weights, and error feedback weights. Weighted fusion strategies can be mathematical rules for weighted combination of the outputs of the physical model and the deep learning model. They can be used to achieve complementary advantages between the two types of models, forming a fused prediction result. In a specific embodiment, weighted fusion strategies can include, but are not limited to, one or more of linear weighted fusion, gated weighted fusion, and confidence-weighted fusion. Particle swarm optimization (PSO) can be a numerical optimization algorithm based on swarm intelligence, which can be used to automatically optimize fusion weights so that the fused prediction model achieves optimal performance under the current operating conditions. Furthermore, the PSO algorithm can be described in context as obtaining its optimal weights by simulating bird flock foraging behavior to search for the optimal fusion weights in the solution space.

[0032] By constructing a deep learning prediction model using AI deep learning algorithms combined with actual blasting parameter sequences and actual peak velocity (PV) sequences, a neural network can be trained using historical monitoring data to establish a nonlinear mapping from parameters to PV. Furthermore, this operation can be achieved by using an LSTM network to process the time-series blasting parameter sequences, or by using a graph neural network to model the vibration propagation relationship between spatially related blast points, thereby improving the fitting ability for nonlinear and highly time-series data. Weighted fusion of the deep learning prediction model and the initial physical prediction model can be performed by weighted summation or nonlinear combination of the outputs of the two models according to the fusion weights. Further, this operation can be achieved through linear weighting (i.e., the fused prediction value equals the physical model prediction value multiplied by the weights plus the deep learning model prediction value multiplied by the complementary weights), or through gated fusion (i.e., dynamically generating weights through a small neural network), thereby achieving the technical effect of integrating physical laws and data characteristics and improving prediction stability.

[0033] Constructing a fusion prediction model for engineering blasting can be achieved by encapsulating the weighted fusion structure into a unified prediction interface, thus forming a deployable composite prediction system. The prediction model used for the current blasting condition is determined based on the weights; if the fusion weights favor a particular model, its output is prioritized as the final prediction, achieving adaptive model selection. The fusion weights are optimized using a particle swarm optimization algorithm, with the prediction error as the fitness function, iteratively searching for the optimal weight combination. Furthermore, this operation can be achieved by fine-tuning the weights online after each new blasting event or by periodically re-optimizing the weights in batches using recent monitoring data, enabling the fusion model to dynamically adapt to current geological and blasting conditions.

[0034] Step S400: Divide the engineering blasting area into several blasting sub-regions, determine whether the blasting sub-regions trigger the similarity set prediction method, if not triggered, predict the peak vibration velocity of each blasting point according to the prediction model adopted, if triggered, filter similar blasting condition groups through clustering algorithm, and predict the peak vibration velocity of blasting within the similar blasting condition groups using the core mapping prediction method.

[0035] The engineering blasting area can be the geographical scope of the blasting operation, serving as the original object for regional division and supporting sub-region modeling. For example, the engineering blasting area can include, but is not limited to, one or more of open-pit mine areas, tunnel sections, and slope excavation zones. A blasting sub-region can be a local blasting unit divided from the entire engineering blasting area according to spatial or geological conditions, supporting regional modeling and similar working condition mining, avoiding the lack of adaptability caused by globally unified modeling. In a specific embodiment, the acquisition method of the blasting sub-region can be described in context: regional segmentation based on Geographic Information System (GIS) data, lithological zoning, or construction planning. For example, the blasting sub-region can include, but is not limited to, one or more of lithologically consistent sub-regions, elevation gradient sub-regions, and construction stage sub-regions. The similarity set prediction method can be a group prediction strategy based on working condition similarity, activated within blasting sub-regions that meet specific conditions, and can be used to activate clustering analysis and core mapping prediction processes. In an exemplary embodiment, the similarity set prediction method can be triggered by the blasting sub-region and invoke the clustering algorithm and core mapping prediction method.

[0036] Similar blasting condition groups can be sets of blast points with similar blasting parameters and geological response characteristics identified by clustering algorithms within the same blasting sub-region. These groups can provide a data foundation for knowledge transfer and reduce the need for independent modeling. Furthermore, the acquisition method of similar blasting condition groups can be explained in context: unsupervised clustering is performed based on blasting parameter sequences (such as charge structure, perforation parameters, and medium wave velocity) to form clusters. In a specific embodiment, similar blasting condition groups can be generated by a clustering algorithm and used as input objects for the core mapping prediction method. The clustering algorithm can be an unsupervised learning method used to group blasting points according to parameter similarity, and can be used to identify blasting conditions with common characteristics to form similar blasting condition groups. For example, the clustering algorithm can include, but is not limited to, one or more of K-means clustering, DBSCAN clustering, and spectral clustering. In the similarity set prediction, the clustering features are selected from three dimensions: charge parameters, geometric parameters, and geological parameters. Specifically, these include the maximum single-segment charge, total charge, detonation center distance, elevation difference, rock mass integrity coefficient, and rock mass wave velocity. These parameters are normalized and mapped to specific intervals to form feature vectors for clustering.

[0037] Core mapping prediction methods can be used to predict the PPV of other samples within a group of similar blasting conditions by establishing a mapping relationship based on representative samples (core points). This can be used to achieve efficient prediction and knowledge reuse within the group, reducing computational overhead. In a specific embodiment, the core mapping prediction method can be described in context as follows: select typical blasting points within the group as cores, construct a local mapping model of their parameters and PPV, and use it to predict other blasting points in the same group. For example, core mapping prediction methods can include, but are not limited to, one or more of K-nearest neighbor core mapping, prototype network mapping, and local linear embedding mapping. Dividing the engineering blasting area into several blasting sub-regions can be done by dividing the overall area according to geological zoning, construction stages, or spatial grids. Furthermore, this operation can be achieved by GIS spatial segmentation based on lithological layers or by dividing time-space sub-regions according to blasting design stages, thereby achieving the technical effect of supporting regional differential modeling and similarity mining. Specifically, the condition closest to the cluster center is first defined as the core prediction condition, and the rest are mapping prediction conditions. Secondly, based on historical data, the historical average proportionality coefficient of the mapped condition relative to the core condition is calculated; that is, the average ratio of the actual peak vibration velocities of the two conditions in multiple historical blasting operations. Finally, when making new predictions, the fusion model is run only on the core prediction condition to obtain predicted values. Then, using the aforementioned historical average proportionality coefficient, the peak vibration velocity of the mapped prediction condition is calculated through direct multiplication. If the variance of the calculated proportionality coefficient is too large, it indicates that the linear relationship between the two is unstable. In this case, the mapped condition will automatically revert to calculating using the fusion prediction model independently.

[0038] Determining whether a blasting sub-region triggers a similarity set prediction method can be achieved by checking whether conditions such as the number of blast points and parameter diversity within the sub-region meet clustering requirements. This allows for a decision on whether to activate a group prediction strategy. Filtering similar blasting condition groups using clustering algorithms can be done by clustering the parameter vectors of blast points within the sub-region to form multiple condition groups. Furthermore, this operation can be achieved by using K-means to cluster standardized blasting parameters or by using DBSCAN to identify densely connected clusters of similar conditions, thereby discovering potential commonalities in blasting conditions and providing a foundation for knowledge transfer.

[0039] Predicting peak velocities (PVs) of blasting vibrations within similar blasting condition groups can be achieved by selecting core samples in each group and constructing a local mapping model to predict the PPVs of other blasting points within the same group. Furthermore, this operation can be implemented by establishing a linear local model with the average blasting point within the group as the core, or by using historical high-precision monitoring points within the group as prototypes for KNN interpolation prediction. This achieves the technical effects of reducing redundant modeling, improving prediction efficiency, and increasing data utilization.

[0040] Taking bench blasting in a large open-pit copper mine as an example, the AI ​​deep learning-based peak velocity (PV) prediction method for engineering blasting in this embodiment can be implemented in an open-pit copper mine where the engineering blasting area is divided into multiple blasting sub-regions with consistent lithology. Within a granite sub-region, the system first constructs an initial physical prediction model based on the Sadovsky formula and performs PPV prediction on three recent blasts. After comparing the prediction with measured vibration data, it is found that the prediction error exceeds a threshold, and the system automatically triggers a fusion AI deep learning prediction method. Using the parameter sequence (including hole depth, charge amount, and differential time) and corresponding PPV sequences from the past 30 blasts, an LSTM deep learning model is trained. Subsequently, the fusion weights of the physical model and the LSTM model are determined to be 0.4:0.6 using a particle swarm optimization algorithm, and a fusion prediction model is constructed. Simultaneously, there are 50 holes to be blasted within this sub-region. The system uses K-means clustering to divide them into four similar blasting condition groups, and selects a core blast point in each group to establish a local mapping model. Finally, the core mapping prediction method was used to quickly output the PPV for the remaining burst points in the group, avoiding running the fusion model for each of the 50 points one by one, which significantly improved the prediction efficiency while maintaining high accuracy.

[0041] In one embodiment, the prediction accuracy is evaluated as follows: Set the prediction period, and the interval between adjacent prediction periods is the model evaluation and update period; For a single or multiple blasting events within each prediction period, the actual blasting parameter sequence of the blasting event is obtained, input into the initial physical prediction model, and the predicted peak vibration velocity is obtained; the actual peak vibration velocity generated by the blasting event is collected in real time. Calculate the average absolute percentage error between the predicted peak vibration velocity and the actual peak vibration velocity. If the average absolute percentage error is greater than or equal to the preset error threshold, the prediction accuracy of the initial physical prediction model is judged to be low.

[0042] The prediction period can be a time window for concentrated execution of one or more blasting event predictions and assessments. It can be used to define the data range and timeliness boundaries of model evaluation and support periodic performance verification. In this embodiment, the prediction period is set by the system according to the project progress or monitoring rhythm, with a model evaluation update period between adjacent prediction periods. Further, the prediction period can include, but is not limited to, one or more of the following: single blasting event evaluation period, multi-blasting round evaluation period, and construction shift evaluation period. The model evaluation update period can be a fixed time interval between adjacent prediction periods, representing the cycle of model performance evaluation and possible updates. It can be used to control the operating rhythm of the evaluation mechanism and avoid system oscillations caused by frequent model switching. For example, the model evaluation update period is preset according to the engineering blasting frequency, geological stability, and system response requirements. In a specific embodiment, the model evaluation update period can be a daily update period, a shift-level update period, or a blasting round-level update period.

[0043] The prediction accuracy evaluation mechanism can be an automatic judgment logic based on periodic prediction periods and quantified error indicators to determine whether the performance of the initial physical prediction model meets the requirements. It can provide an objective, dynamic, and quantifiable decision-making basis for whether to trigger the fusion AI deep learning prediction method. In an exemplary embodiment, the prediction accuracy evaluation mechanism uses the actual blasting parameter sequence to drive the initial physical prediction model to generate predicted peak vibration velocities within each prediction period, calculates the mean absolute percentage error with the synchronously acquired actual peak vibration velocities, and then compares it with a preset error threshold to determine the model's effectiveness. Furthermore, the prediction accuracy evaluation mechanism relies on the prediction period, model evaluation update period, actual blasting parameter sequence, actual peak vibration velocity, mean absolute percentage error, and error threshold, and controls the initiation of the fusion process.

[0044] Setting a prediction period can be done by defining a time window within which input and output data from several blasting events are collected for model evaluation. Furthermore, the prediction period can be set as a 24-hour period based on the daily blasting plan, or as a 1-2 hour period based on each round of micro-delay blasting operations. This limits the time range of the evaluation data, ensuring that the evaluation results reflect recent conditions. Within each prediction period, the actual blasting parameter sequence is obtained for one or more blasting events. This can be achieved by extracting the complete parameter vectors of each blasting event within that period from the blasting design system or field records, providing realistic input conditions for the initial physical prediction model.

[0045] The actual blasting parameter sequence is input into the initial physical prediction model to obtain the predicted peak velocity (PV). This can be achieved by calling an existing initial physical prediction model, substituting the current blasting parameters, and calculating the predicted PPV value, thereby generating a benchmark prediction result that can be compared with the measured value. The actual PV generated by the blasting event is acquired in real time. This can be done by synchronously recording the measured PPV value generated by the blast using vibration sensors deployed at monitoring points, providing the true value label required for model evaluation. The mean absolute percentage error (MASE) is a standardized statistical indicator that measures the relative deviation between the predicted and actual PV values. It can be used to provide a dimensionless, cross-condition comparable model accuracy measure, suitable for unified evaluation under different PPV scenarios. In this embodiment, the MASE is obtained by averaging the absolute value of the difference between the actual and predicted values ​​for all blasting events within the prediction period, divided by the actual value. For example, the MASE threshold is set to 15%. When the error value of the initial physical model reaches or exceeds 15%, the system determines its accuracy is low and triggers the prediction process integrating AI deep learning. For example, the mean absolute percentage error (MAPE) can be one or more of the following, including but not limited to single-point MAPE, weighted MAPE, and sliding window MAPE. The initial physical prediction model can employ the classic Sadovsky formula, which is based on the geometric attenuation and medium damping effect of blast vibration propagation, and its specific expression is as follows: ; in: The predicted peak velocity of the blasting vibration; This is the maximum charge amount per segment; The straight-line distance from the measuring point to the explosion center; Site coefficients related to the blasting site and blasting method; This is the attenuation index related to geological conditions. The parameter can be determined by collecting historical monitoring data (no fewer than 20 sets of valid data) from the early stages of the engineering blasting project when constructing the initial physical prediction model, linearizing it by taking the logarithm of both sides of the above formula, and then fitting the result using least squares regression analysis. and The specific values ​​can be set initially, referring to empirical values ​​for similar rock conditions (such as medium-hard rock). , It is dynamically corrected as data accumulates.

[0046] The mean absolute percentage error (MAPE) between the predicted peak vibration velocity and the actual peak vibration velocity can be calculated by taking the arithmetic mean of the relative errors of all blasting events within the prediction period, thus generating a standardized model performance index and eliminating the influence of dimensions. The preset error threshold can be the upper limit of the MAPE used to determine whether the initial physical prediction model has failed, and can be used as a boundary to trigger the fusion of AI deep learning prediction methods. In one specific embodiment, the preset error threshold is pre-set based on the engineering safety level, historical model performance, or regulatory requirements. Furthermore, the preset error threshold may include, but is not limited to, one or more of the following: safety-sensitive threshold, efficiency-priority threshold, and adaptive dynamic threshold.

[0047] Determining whether the mean absolute percentage error (MAPE) is greater than or equal to a preset error threshold can be achieved by comparing the calculated MAPE with the system's preset threshold, thereby automatically assessing model performance. If the error is greater than or equal to the threshold, the prediction accuracy of the initial physical prediction model is considered low. This can be achieved by outputting a "low prediction accuracy" signal when MAPE ≥ the error threshold, triggering subsequent fusion processes to complete the closed loop from evaluation to decision-making and initiating the AI ​​deep learning fusion strategy.

[0048] For example, in the scenario of blasting for highway tunnel excavation in mountainous areas, the AI ​​deep learning-based peak velocity prediction method for engineering blasting vibration in this embodiment can be as follows: In a tunnel excavation project, the system sets each prediction period to 8 hours, and the model evaluation update period is also 8 hours. Three cut-out blasts are carried out within a certain prediction period. The system automatically acquires the actual blasting parameter sequence (including hole depth, charge amount, rock mass wave velocity, etc.) for these three blasts, inputs it into the initial physical prediction model, and obtains predicted PPVs of 2.1 cm / s, 2.8 cm / s, and 3.0 cm / s, respectively. Simultaneously, the vibration monitoring system collects the corresponding actual PPVs in real time as 1.5 cm / s, 2.0 cm / s, and 2.2 cm / s. The system calculates MAPE = (|1.5−2.1| / 1.5+|2.0−2.8| / 2.0+|2.2−3.0| / 2.2) / 3≈36.7%. The preset error threshold is 30%. Since 36.7% > 30%, the system determines that the initial physical prediction model has low prediction accuracy and automatically triggers the fusion AI deep learning prediction method to start the LSTM model training and fusion weight optimization process, providing higher accuracy prediction capability for the next prediction period.

[0049] In one embodiment, the prediction model used to determine the current blasting condition is as follows: The prediction model includes a fusion prediction model and an initial physics prediction model; For each blasting condition that triggers the fusion AI deep learning prediction method, a weight evaluation flag value is set and initialized to 0. The weight of the initial physical prediction model in the fusion prediction model is compared with the preset weight threshold. If the weight is greater than or equal to the weight threshold, the weight evaluation flag value is incremented once. If the weight is less than the weight threshold, the weight evaluation flag value is reset to 0. If the weight evaluation flag value is less than the stability flag value, the current blasting condition adopts the fusion prediction model; otherwise, the current blasting condition ends the fusion AI deep learning prediction method and adopts the initial physical prediction model.

[0050] The weights of the initial physical prediction model can be fusion coefficients assigned to the output of the fusion prediction model, reflecting the contribution of physical laws to the prediction results under the current operating conditions. In this embodiment, the weights of the initial physical prediction model can be dynamically generated by a particle swarm optimization algorithm and used as the basis for updating the weight evaluation flag value. For example, the weights of the initial physical prediction model can include, but are not limited to, one or more of the following: physical dominance threshold, fusion retention boundary, and model switching sensitivity parameter. The preset weight threshold can be a numerical boundary used to determine whether the weights of the initial physical prediction model are dominant, and can be used as a criterion for determining whether the weight evaluation flag value is auto-incrementing. For example, if the weight threshold is set to 0.8, and the physical model weights are stable and greater than or equal to 0.8 after optimization, it indicates that the physical model performs dominantly. In an exemplary embodiment, the preset weight threshold can include, but is not limited to, the physical dominance threshold, fusion retention boundary, and model switching sensitivity parameter.

[0051] The blasting conditions that trigger the fusion AI deep learning prediction method can be specific engineering blasting events that, according to the prediction accuracy assessment mechanism, are deemed to have insufficient physical model accuracy and require the activation of the fusion strategy. These events can be used as binding objects for weight evaluation flag values, ensuring that each independent condition has independent state tracking. Furthermore, the blasting conditions that trigger the fusion AI deep learning prediction method can include, but are not limited to, one or more of the following: highly nonlinear conditions, geologically abrupt changes, and complex porous micro-differential conditions. A weight evaluation flag value is set for each blasting condition that triggers the fusion AI deep learning prediction method. This can be achieved by assigning an independent weight evaluation flag value variable when the blasting condition first enters the fusion prediction process. Furthermore, this operation can be implemented through a memory variable allocation mechanism, thereby achieving condition-level state isolation and avoiding evaluation interference between different engineering blasting events.

[0052] The initial value of the weight evaluation flag is 0, which can be initialized to an integer 0 when the flag is created. Furthermore, this operation can be implemented through program initialization statements, ensuring that the evaluation starts from a state without prior knowledge and guaranteeing the fairness of the judgment logic. The weight evaluation flag can be an integer counter used to record the number of times the initial physical prediction model weight is continuously higher than the threshold under the current blasting condition. It can be used to quantify the continuity of the dominant physical trend and serve as a basis for determining whether to exit the fusion prediction mode. In a specific embodiment, the weight evaluation flag can be one or more of the following, including but not limited to a continuous achievement counter, a physical credibility accumulator, and a fusion exit trigger counter. Obtaining the weight of the initial physical prediction model in the fusion prediction model can be done by reading the fusion weight corresponding to the physical model from the parameter structure of the current fusion prediction model. Furthermore, this operation can be implemented through the model parameter parsing interface, thus providing real-time basis for subsequent comparison with the threshold.

[0053] The initial weights of the physical prediction model are compared with a preset weight threshold. This can be achieved by performing a numerical judgment: if the weight is greater than or equal to the threshold, it is considered physically dominant; otherwise, it is considered data-dominated. Furthermore, this operation can be implemented using conditional statements to identify whether the current operating condition tends to be accurately described by the physical model. If the weight is greater than or equal to the weight threshold, the weight evaluation flag is incremented, which can be done by incrementing an integer variable (e.g., flag = flag + 1). This operation can also be implemented using an arithmetic increment command to accumulate the consecutive occurrences of the physically dominant state. If the weight is less than the weight threshold, the weight evaluation flag is reset to 0. This operation can also be implemented using an assignment statement to interrupt continuous recording and prevent accidental high weights from causing misjudgments.

[0054] The stability flag value can be a preset integer threshold used to determine whether the weight evaluation flag value has reached a stable state of the regressible physical model. It can be used to control the triggering condition for switching back from the fused prediction model to the initial physical prediction model, preventing frequent switching. For example, if the stability flag value is set to 5, the system determines the operating condition is stable when the weight threshold is met a cumulative total of 5 times, ending the fusion calculation and switching back to the initial physical model to reduce resource consumption. In an exemplary embodiment, the stability flag value can be one or more of the following, including but not limited to the operating condition stability determination threshold, the fusion exit confirmation threshold, and the number of physical regression confirmation steps. Determining whether the weight evaluation flag value is less than the stability flag value can be done by comparing the current flag value with the preset stability flag value. Furthermore, this operation can be implemented through numerical comparison logic, thereby determining whether to maintain the fused prediction or switch back to the physical model.

[0055] If the weight evaluation flag value is less than the stability flag value, the current blasting condition uses a fusion prediction model, which can continue to use the output of the fusion prediction model as the PPV prediction result. Furthermore, this operation can be achieved through model call path selection, thus maintaining high-precision prediction capability even before the physical laws have stabilized. If the weight evaluation flag value is greater than or equal to the stability flag value, the fusion AI deep learning prediction method ends and the initial physical prediction model is used. This can be achieved by stopping the deep learning component calls and directly using the initial physical prediction model for subsequent predictions. Furthermore, this operation can be achieved by disabling the neural network inference module and enabling the analytical physics calculation module, thereby reducing computational overhead and improving system efficiency and interpretability after confirming the reliability of the physical model.

[0056] Taking cyclic blasting in tunnel excavation as an example, the peak velocity prediction method for engineering blasting vibration based on AI deep learning in this embodiment can be used in the process of tunnel excavation in a mountainous area. In the first five blasts, due to the fractured and jointed surrounding rock, the initial physical prediction model has a large error, triggering the system to fuse the AI ​​deep learning prediction method. After each blast, the system obtains the physical weights in the fused model: the first blast is 0.3 (<0.6 threshold), and the flag value is reset to 0; the second blast is 0.7 (≥0.6), and the flag value is 1; the third blast is 0.65, and the flag value is 2; the fourth blast is 0.55, and the flag value is reset to 0; the fifth blast is 0.7, and the flag value is 1. The stable flag value is set to 3, so the system continues to use the fused model. From the sixth blast onwards, the surrounding rock tends to be homogeneous and intact, and the physical weights are 0.8, 0.75, and 0.82 for three consecutive times, with the flag values ​​increasing to 1, 2, and 3 respectively. When the flag value is greater than or equal to the stable flag value (3), the system automatically exits the fusion mode, and the subsequent blasting directly adopts the initial physical prediction model, saving GPU computing resources while maintaining prediction reliability.

[0057] In one embodiment, the fusion prediction model is obtained as follows: The deep learning prediction model and the initial physical prediction model are obtained and weighted and fused to construct a fused prediction model, wherein the sum of the weights of the deep learning prediction model and the initial physical prediction model is 1. The sum of weights being 1 can be a normalization constraint applied to the weights of the two sub-models in the fusion prediction model. In this embodiment, a sum of weights being 1 ensures the numerical stability and interpretability of the fusion result, avoiding weight amplification or cancellation effects. For example, a sum of weights being 1 can include, but is not limited to, one or more of convex combination constraints, probabilistic interpretation constraints, and linear interpolation constraints. Setting the sum of weights of the deep learning prediction model and the initial physical prediction model to 1 can be achieved by forcibly satisfying in the weighted fusion formula that the sum of the physical model weights and the deep learning model weights equals 1, where the deep learning model weights are obtained by subtracting the physical model weights from 1. Furthermore, this operation can be achieved by directly substituting the constraint relationship, thereby making the fusion process a convex combination, ensuring the output lies between the predicted values ​​of the two models, and improving numerical robustness. In this embodiment, the deep learning prediction model employs a Long Short-Term Memory (LSTM) network, aiming to capture the complex time-series dependencies between blasting parameters and vibration response. In terms of network structure configuration, the input layer receives a normalized seven-dimensional feature vector, covering the maximum segment charge, total charge, borehole depth, borehole spacing, row spacing, rock mass wave velocity, and blast center distance. The hidden layer uses a two-layer LSTM structure, with the first layer containing 64 units and the second layer containing 32 units. The Tanh activation function is used, and a dropout layer with a dropout rate of 0.2 is added after each layer to prevent overfitting. The output layer is a single-node fully connected layer, using a linear activation function to output the predicted peak vibration velocity. For the training strategy, the mean squared error is used as the loss function, the Adam optimizer is selected, and the initial learning rate is set to 0.001. The training data is divided into training and validation sets in an 8:2 ratio according to time sequence, with a maximum of 100 iterations and a batch size of 16. If the validation set loss does not decrease within 10 consecutive iterations, training is terminated early. For incremental training, newly acquired blasting data is added to the training set, and the model is fine-tuned and updated using a smaller learning rate while maintaining the network structure.

[0058] Using a particle swarm optimization algorithm, a set of initial physical prediction model weights is randomly generated. The historical blasting parameter sequence is obtained and input into the fusion prediction model. The mean absolute percentage error between the prediction result and the corresponding actual vibration peak velocity is calculated. By updating the position and adjusting the velocity of the particles, the particle swarm is continuously optimized. The initial physical prediction model weights that minimize the mean absolute percentage error are gradually selected. The weights of the initial physical prediction model are determined, and the fusion prediction model is obtained.

[0059] The particle swarm for the weights of the initial physics prediction model can be a group in which each particle encodes a candidate initial physics prediction model weight value during particle swarm optimization. In an exemplary embodiment, the particle swarm for the weights of the initial physics prediction model can be used to explore the optimal fusion weight space through a swarm parallel search strategy. Further, the particle swarm for the weights of the initial physics prediction model can be obtained by randomly initializing the positions of multiple particles in the interval [0, 1], with each position corresponding to a physics model weight, and the deep learning model weight obtained by subtracting this value from 1. Exemplarily, the particle swarm for the weights of the initial physics prediction model can include, but is not limited to, one or more of the following: uniformly initialized particle swarm, Gaussian perturbation particle swarm, hierarchical sampling particle swarm, etc.

[0060] The particle's position can be its coordinates in the optimization space, specifically referring to the weight value of the initial physical prediction model it represents. In one specific embodiment, the particle's position can be used to construct a fusion prediction model; its quality is evaluated using the mean absolute percentage error. Furthermore, the particle's position can be obtained by adjusting it in each iteration according to the particle swarm optimization criterion update rule based on the individual optimal and global optimal positions. The particle's velocity can be a vector controlling the particle's movement direction and step size in the search space, and is a scalar in this scheme. In this embodiment, the particle's velocity can be used to guide the particle to converge towards a better weight region, avoiding getting trapped in local minima. Further, the particle's velocity can be obtained by iteratively adjusting it according to the particle swarm optimization criterion velocity update formula, combined with inertial weights, cognitive factors, and social factors. For example, the particle's velocity can include, but is not limited to, one or more of constant inertial velocity, adaptive decay velocity, and chaotic perturbation velocity.

[0061] Mean Absolute Percentage Error (MAPE) can be a statistical metric used to measure the relative error between predicted and actual values, calculated as the average of the absolute percentage errors at each point. In an exemplary embodiment, MAPE can be used as the fitness function of a particle swarm optimization algorithm to quantify the prediction performance under a fusion weight configuration. Furthermore, MAPE can be obtained by calculating |(actual value - predicted value) / actual value| point-by-point on the output of the fusion prediction model on a historical blasting parameter sequence and the corresponding actual peak vibration velocity, and then averaging the results. Exemplarily, MAPE can include, but is not limited to, one or more of symmetric MAPE, weighted MAPE, and piecewise MAPE.

[0062] A particle swarm optimization algorithm is used to randomly generate a set of initial physical prediction model weights. This can be achieved by randomly sampling N values ​​within the range [0, 1] as initial particle positions, with each position corresponding to a physical model weight. Furthermore, this operation can be implemented using uniformly distributed random initialization or non-uniform sampling within a high-confidence interval (e.g., 0.4–0.8) based on prior knowledge, thereby constructing diverse initial solution sets and improving global search capabilities. Historical explosion parameter sequences are input into the fusion prediction model. This can be done by constructing a temporary fusion model for the weights corresponding to each particle and using the historical explosion parameter sequences as input for batch prediction. This operation generates a prediction result set for error evaluation. Specifically, the particle swarm optimization algorithm is used to determine the initial physical prediction model weights, while the weights of the deep learning model are obtained by subtracting the physical model weights from 1. The operating parameters of the particle swarm optimization algorithm are set as follows: the population size is 30 particles; the search space for the physical model weights is limited to 0 to 1; velocity and position updates adopt a linear decreasing strategy, with the inertia weight decreasing from 0.9 to 0.4; both the individual cognitive factor and the social cognitive factor are set to 1.5; the maximum velocity is limited to 20% of the search space range, i.e., 0.2. The algorithm terminates when the maximum number of iterations reaches 50, or when the fitness value improves by less than 0.001 for 5 consecutive iterations. During the optimization process, the mean absolute percentage error is used as the fitness function to evaluate the predictive performance of the fusion model under different weight configurations.

[0063] The mean absolute percentage error between the prediction results of the fusion prediction model and the corresponding actual peak velocity can be calculated by taking the arithmetic mean of the absolute percentage errors for each historical sample, and using this as the fitness value of the current particle. Furthermore, this operation can be achieved by directly calculating using the standard MAPE formula or by smoothing the zero or near-zero true values ​​(e.g., adding ε) before calculation. This provides an optimization objective with weak differentiability but clear engineering significance, suitable for positive real number prediction tasks such as peak velocity prediction in engineering vibrations.

[0064] Updating particle position and velocity based on mean absolute percentage error (MASE) can be achieved by adjusting the current particle's velocity and position according to the standard formula for particle swarm optimization, utilizing the individual historical best position and the swarm's global best position. This operation drives particles to migrate towards regions with low MSE, gradually approaching the optimal fusion weights. Continuously optimizing the particle swarm to progressively select the weights of the initial physical prediction model that minimizes the MSE can be achieved by repeatedly executing a prediction-evaluation-update loop until convergence conditions are met (such as the maximum number of iterations or an error change threshold). This operation enables data-driven adaptive optimization of the fusion weights, avoiding manual parameter tuning. Determining the optimal initial physical prediction model weights yields the final fusion prediction model. This can be achieved by selecting the position value corresponding to the particle with the smallest MSE as the final physical weight and fixing this weight to construct the fusion model for deployment. This operation outputs a fusion prediction structure with optimal empirical performance, supporting subsequent operational condition predictions.

[0065] Taking the vibration of a hydropower station dam foundation excavation project as an example, the AI ​​deep learning-based peak velocity prediction method for engineering blasting vibration in this embodiment can be implemented in a large hydropower station dam foundation excavation project. The system collects historical blasting parameter sequences (including hole depth, energy input, rock mass P-wave velocity, etc.) and corresponding measured PPVs from the past 60 blasts. To construct a fusion prediction model, the system initiates a particle swarm optimization process: initializing 30 particles, each representing a physical model weight (e.g., 0.25, 0.63, etc.), and automatically setting the deep learning weight to 1 minus this value. For each particle, the system constructs a temporary fusion model and predicts the PPV of all 60 blasts, calculating the MAPE. The first round of evaluation shows that particle 12 (weight = 0.58) has the lowest MAPE (8.2%). Subsequently, the positions and velocities of all particles are updated according to the PSO rule. After 15 iterations, particle 7 converges to a weight of 0.67, and the MAPE drops to 6.1%, with no significant improvement for 5 consecutive rounds. The system locks this weight and constructs the final fusion prediction model. In subsequent blasting operations, the model performed stably in hard rock sections and maintained high accuracy in fault zones, verifying the effectiveness of PSO optimized weights.

[0066] In one embodiment, the deep learning prediction model is obtained as follows: Determine whether a deep learning prediction model exists for the current blasting condition; if not, obtain all actual blasting parameter sequences and corresponding actual peak velocity sequences for this type of blasting condition up to the current moment, integrate them according to the time series to obtain the total blasting parameter sequence and the total peak velocity sequence, obtain the training dataset, construct an LSTM model based on the AI ​​deep learning algorithm and train it using the training dataset, and obtain the trained LSTM model as the deep learning prediction model for this condition; If they exist, obtain the latest actual blasting parameter sequence and the corresponding actual vibration peak velocity sequence that were not included in the training dataset to supplement the training dataset, perform incremental training on the deep learning prediction model, and update the deep learning prediction model for this working condition.

[0067] This type of blasting condition can be a category of rock breaking operation events with similar physical response characteristics, categorized based on geological conditions, rock breaking operation design, or construction stage. It can be used as the binding granularity for deep learning prediction models, enabling model customization according to specific scenarios. Furthermore, this type of blasting condition can include, but is not limited to, one or more of the following: hard rock bench blasting, weak interlayer blasting, and blasting for protection of adjacent structures. Determining whether a deep learning prediction model exists for the current blasting condition can be done by querying the model registry or storage system to check if an LSTM model has already been created for this type of blasting condition. Further, determining whether a deep learning prediction model exists for the current blasting condition can be achieved by comparing the condition identifier with existing model metadata, thereby deciding whether to perform full training or incremental update.

[0068] If not found, obtain all actual blasting parameter sequences and corresponding actual peak velocity (PV) sequences for that type of blasting condition up to the current moment. This can be achieved by filtering all historical rock breaking operation events belonging to the same condition type from the monitoring database. Furthermore, this operation can be performed jointly through time range filtering and condition label matching, thus providing complete training material for the initial modeling. The total blasting parameter sequence can be a unified input sequence set formed by concatenating all actual blasting parameter sequences for a certain type of blasting condition up to the current moment in chronological order. This can be used as the input feature source for LSTM model training, preserving the complete temporal context. In an exemplary embodiment, the total blasting parameter sequence can be one or more of the following, including but not limited to multidimensional parameter tensor sequences, standardized parameter sequences, and sequences with masked missing value processing. The total peak velocity (PV) sequence can be a set of measured PPV values ​​corresponding one-to-one with the total blasting parameter sequence, organized in the same chronological order, and can be used as the target label for LSTM model training. For example, the total PV sequence can be one or more of the following, including but not limited to multi-directional PPV sequences, filtered PPV sequences, and frequency-band PPV sequences.

[0069] The total blasting parameter sequence and total peak velocity (PV) sequence are obtained by integrating them according to the time series. This can be done by sorting the sequences according to the blasting time and organizing the parameter vectors and PPV values ​​of each blast into ordered lists or tensors. Further, this operation can be achieved by constructing a three-dimensional tensor in (N, T, D) format, where N is the number of rock breaking operations, T is the time series length of a single parameter, and D is the parameter dimension; or by padding or truncating sequences of unequal length to achieve a uniform length, thus forming a standard time series input format acceptable to LSTM. The training dataset can be a supervised learning sample set consisting of all historical actual blasting parameter sequences and their corresponding actual PV sequences under the same type of blasting condition, which can be used to support the initial training or incremental updates of the LSTM model. In a specific embodiment, the training dataset can include, but is not limited to, one or more of the following: a full historical training set, a sliding window training set, and a key event augmentation training set.

[0070] Obtaining the training dataset can be achieved by pairing the integrated total blasting parameter sequence as input X and the total peak velocity sequence as label Y to form a supervised sample set. Further, this operation completes the data structure encapsulation by establishing an input-output mapping relationship, thereby completing the data preparation before training. The LSTM model can be a recurrent neural network with a gating mechanism, specifically designed for modeling and predicting long-term dependencies in time-series data, and can be used to effectively capture the nonlinear temporal coupling characteristics between parameters and vibration response during blasting. In this embodiment, the LSTM model controls the information flow through input gates, forget gates, and output gates, learning to dynamically map to PPV in the time dimension of the blasting parameter sequence. For example, the LSTM model can include, but is not limited to, one or more of stacked LSTM, bidirectional LSTM, and attention-enhanced LSTM.

[0071] Building an LSTM model based on AI deep learning algorithms can involve instantiating the LSTM network structure, including configuring hyperparameters such as the number of layers, hidden units, and activation functions. Furthermore, this operation can employ a single-layer LSTM followed by a fully connected output layer, or use a bidirectional LSTM to capture contextual dependencies, thereby establishing a neural network skeleton with temporal modeling capabilities. Training the LSTM model using a training dataset can involve minimizing the loss function between the predicted PPV and the actual PPV through backpropagation and an optimizer (such as Adam). Further, this operation iteratively optimizes the network weights to make the output approximate the measured value, allowing the model to learn the nonlinear mapping relationship between the burst parameters and PPV.

[0072] The trained LSTM model is used as the deep learning prediction model for this work condition. This can be achieved by saving the model weights and structure after training convergence and registering them in the work condition-model mapping table. Further, this operation prepares for deployment by persisting the model state, thus completing the initial model deployment for this type of work condition. If a deep learning prediction model exists, the latest actual blasting parameter sequence and corresponding actual vibration peak velocity sequence that are not included in the training dataset are obtained. This can be achieved by comparing data timestamps with the model's last training time and extracting newly added monitoring records. Further, this operation uses an incremental data identification mechanism to filter out new samples that did not participate in training, thereby identifying new knowledge that can be used for incremental learning.

[0073] The latest actual blasting parameter sequence can be a newly generated record of actual blasting parameters since the last model training that has not yet been used for training. It can be used as a new sample source for incremental training, ensuring the timeliness of the model. The corresponding actual peak velocity (PV) sequence (not included in the training dataset) can be measured PPV values ​​collected synchronously with the latest actual blasting parameter sequence but not yet used in model training. It can be used to provide the supervision signals required for incremental training. Supplementing the training dataset with the latest data can be done by appending new samples to the end of the original training set or by constructing a separate mini-batch incremental set. Furthermore, this operation expands the training data coverage by expanding the sample pool to cover new working condition features.

[0074] Incremental training can be a retraining method that uses newly added data to fine-tune local parameters on the basis of an existing trained deep learning model. This avoids the high computational cost of training from scratch and allows the model to continuously adapt to changes in the environment. In this embodiment, incremental training updates the model weights only with new data by freezing some network layers or using a smaller learning rate. For example, incremental training can include, but is not limited to, one or more of online fine-tuning, experience replay incremental training, and elastic weight consolidation incremental training. Incremental training of the deep learning prediction model can involve loading existing model weights and performing several rounds of fine-tuning training using only new data. Furthermore, this operation can use a smaller learning rate (e.g., 1e-4) for 1–3 epochs of fine-tuning, or use knowledge distillation to constrain the magnitude of output distribution changes, thereby adapting to new operating conditions while retaining historical fitting capabilities. Updating the deep learning prediction model for the operating condition can involve replacing the original model file or version number and updating the model metadata. Furthermore, this operation uses a version management mechanism to track model evolution, thereby enabling continuous model evolution and adaptation to the environment.

[0075] Taking blasting in different rock zones of an open-pit iron mine as an example, the AI ​​deep learning-based method for predicting peak velocity (PV) vibration in engineering blasting in this embodiment can be as follows: An open-pit iron mine is divided into two blasting conditions: a magnetite main mining area and a surrounding rock stripping area. When the system first performs blasting in the magnetite area, it detects that there is no corresponding deep learning model. Therefore, it collects the parameter sequences (including time-series features such as borehole parameters, fracturing agent arrangement structure, and rock mass wave velocity) and measured PPV values ​​from the previous 20 similar blasting operations, integrates them into a training dataset, constructs a 3-layer LSTM model, completes training, and generates an initial deep learning prediction model. After each new rock fracturing operation, the system automatically checks whether there is any new monitoring data that has not been included in the training. After the 25th rock fracturing operation, the system extracts the newly added data and performs one round of incremental training (learning rate 1e-4) on the original LSTM model to update the model. In the surrounding rock stripping area, due to the loose lithology and rapid vibration attenuation, the system builds a separate LSTM model for it to avoid performance degradation caused by mixing with the main mining area. This mechanism enables high-precision predictions for both types of working conditions, and the model automatically adapts to slight changes in lithology as mining progresses.

[0076] In one embodiment, the method for determining whether to trigger the similarity set prediction method is as follows: The engineering blasting area is divided into several blasting sub-areas based on geological conditions, blasting zone range, and blasting type.

[0077] This operation can involve multidimensionally dividing the overall area based on comprehensive lithological zoning, spatial geometric boundaries, and blasting methods (such as bench blasting and cut-out blasting). Furthermore, this operation can be achieved by overlaying GIS lithological layers with blasting design CAD drawings to generate sub-region boundaries, or by using clustering algorithms to perform unsupervised region division based on the combined features of "geological + blasting parameters," thereby forming local units with high internal working condition consistency and providing reasonable spatial granularity for subsequent similarity mining.

[0078] For each blasting sub-region, obtain the number of blasting conditions within the blasting sub-region that use the fusion prediction model.

[0079] The number of blasting events within a specific blasting sub-region that employ the fusion prediction model can be considered as the total number of blasting events within that sub-region that have been predicted and actually used the fusion prediction model (rather than just the initial physical model). This number reflects the accumulation of high-quality prediction samples within the sub-region and is a key indicator for determining whether a foundation for group modeling is available. In this embodiment, the number of blasting events within a blasting sub-region that employ the fusion prediction model can be obtained by recording the prediction model type used for each blast and counting the number of times the fusion model is invoked within the sub-region.

[0080] Calculate the ratio of the number of blasting conditions to the total number of blasting conditions within the blasting sub-region.

[0081] The total number of blasting conditions within a blasting sub-region can be the total number of all blasting events implemented or recorded within that specific sub-region. This number can be used as the denominator to calculate the application ratio of the fusion model, characterizing the data completeness of the sub-region. For example, the total number of blasting conditions within a blasting sub-region can be obtained from blasting logs or construction management systems, aggregated by spatial affiliation. The ratio can be the ratio of the number of blasting conditions using the fusion prediction model within the blasting sub-region to the total number of blasting conditions. This ratio can be used to quantify the coverage density of high-quality prediction samples within the sub-region, serving as a direct basis for whether to trigger the similarity set prediction method. In an exemplary embodiment, the ratio is calculated in real-time through division: the ratio equals the number of fused conditions divided by the total number of conditions.

[0082] If the ratio is greater than the preset ratio standard, the similarity set prediction method is triggered in the explosion sub-region.

[0083] The preset proportion standard can be a threshold proportion used to determine whether there are enough samples for the fusion model application within a blasting sub-region to activate the similarity set prediction method. It can serve as a quantitative criterion for triggering the similarity set prediction method, ensuring that the group modeling strategy is only activated in sub-regions with sufficient data and reliable prediction quality. In one specific embodiment, the preset proportion standard is pre-set based on engineering experience, data density requirements, or system resource constraints, representing the minimum acceptable proportion of scenarios using the fusion prediction model out of the total scenarios. Furthermore, the preset proportion standard can include, but is not limited to, one or more of the following: high confidence proportion standard (e.g., ≥70%), medium coverage proportion standard (e.g., ≥50%), and low threshold exploration proportion standard (e.g., ≥30%).

[0084] Determining whether the ratio is greater than a preset ratio standard can be achieved by comparing the calculated ratio with a system-preset ratio threshold, enabling automatic decision-making on whether to activate higher-order group prediction strategies. If the ratio is greater than the preset ratio standard, a similarity set prediction method is triggered within the blasting sub-region. This can be achieved by activating cluster analysis and core mapping prediction processes when conditions are met, thereby ensuring that the knowledge transfer mechanism is initiated only when data is sufficient and predictions are reliable, improving resource utilization efficiency and prediction stability.

[0085] For example, in the scenario of a multi-stage blasting zone in a large open-pit iron mine, the AI ​​deep learning-based peak velocity (PV) prediction method for engineering blasting in this embodiment can be as follows: A certain open-pit iron mine divides the entire mining area into 12 blasting sub-regions according to lithology (gneiss / granite), bench elevation (one bench every 15 meters), and blasting type (pre-splitting blasting / main blasting). In one of the granite main blasting sub-regions, a total of 40 blasts were completed, of which 30 used the fusion prediction model due to errors exceeding the limit in the initial physical model. The system calculated the ratio to be 30 / 40 = 75%. The preset ratio standard is 70%. Since 75% > 70%, the system determines that the sub-region has sufficient high-quality samples and automatically triggers the similarity set prediction method. Subsequently, cluster analysis is performed on the 40 blast points in the sub-region to identify 3 similar blasting condition groups. Within each group, the core mapping prediction method is used to quickly predict the PPV of new blast points, avoiding repeated calls to the complete fusion model and significantly improving prediction efficiency.

[0086] In one embodiment, the similar blasting condition group is obtained as follows: All blasting conditions within the blasting sub-region are divided into two clusters.

[0087] This operation can involve using a binary clustering algorithm to initially partition the feature vectors of all blasting conditions within a sub-region. Furthermore, this operation can be achieved by using k-means with k=2 to initially cluster the normalized vibration feature sequence, or by using principal component analysis projection followed by median segmentation to achieve rapid binary partitioning. This allows the initiation of an adaptive hierarchical clustering process, providing the foundation for subsequent recursive subdivision or merging.

[0088] Calculate the two cluster centers and the Euclidean distance between them, and compare them with the set working condition judgment threshold.

[0089] The cluster center can be the center point representing the mean of the characteristic vectors of all blasting conditions within a cluster. It can be used as a benchmark to measure the consistency within a cluster and the differences between clusters, in order to determine whether to continue splitting or merging. In this embodiment, the cluster center is obtained by calculating the arithmetic mean of the normalized vibration characteristic sequences of all blasting conditions within the cluster. For example, the cluster center can be one or more of the following, including but not limited to parameter space cluster center, temporal characteristic cluster center, and weighted characteristic cluster center.

[0090] The working condition judgment threshold can be an upper limit of Euclidean distance used to determine whether two clusters belong to similar working conditions. It can be used to control the clustering granularity, determine when to stop splitting and mark clusters as similar, for example, a working condition judgment threshold set to 0.5 is used to determine whether two clusters are highly similar; a preset proportion standard is set to 60%, triggering the similarity set prediction method when the proportion of working conditions using the fusion model in a sub-region exceeds 60%. In an exemplary embodiment, the working condition judgment threshold is preset based on the maximum reasonable feature difference of similar working conditions in historical blasting data, or determined through cross-validation optimization. Furthermore, the working condition judgment threshold can include, but is not limited to, a high-sensitivity threshold, a medium-discrimination threshold, and a lenient merging threshold. This operation can involve separately calculating the feature mean vectors of the two clusters and calculating their L2 norm distance. Further, this operation performs a numerical comparison: if the distance ≥ the threshold, the difference between clusters is considered significant; otherwise, they are considered similar, thereby quantifying the degree of separation between the two clusters in the feature space and realizing conditional branch control of the clustering path.

[0091] If the number of clusters is greater than or equal to the set working condition judgment threshold, each cluster is further divided into two clusters, and the Euclidean distance between the centers of the two clusters is calculated and compared again. If the number of clusters is greater than the preset number standard, the re-clustering and comparison are stopped.

[0092] The preset quantity standard can be an upper limit on the maximum number of clusters allowed to be generated, which can be used to prevent excessive subdivision leading to model fragmentation. In a specific embodiment, the preset quantity standard is pre-set based on the total number of blasting conditions in the sub-region, computational resource limitations, or engineering management granularity. For example, the preset quantity standard may include, but is not limited to, one or more of the following: a fixed upper limit standard (e.g., ≤8), a proportional upper limit standard (e.g., ≤20% of the total number of conditions), and an adaptive upper limit standard (dynamically adjusted based on the profile coefficient). This operation may involve further dividing each cluster into two clusters if the Euclidean distance is greater than or equal to the condition judgment threshold. Further, this operation generates finer-grained sub-clusters by recursively applying binary clustering to each current cluster, thereby enabling refined subdivision in areas with significant differences in conditions and improving model resolution. This operation may also involve calculating the Euclidean distance between the centers of the two newly generated clusters and comparing them again. Further, this operation supports multi-level adaptive clustering by repeating the distance calculation and threshold comparison process for each newly generated sub-cluster, thereby refining the condition grouping layer by layer. This operation may also involve stopping re-clustering and comparison if the number of clusters exceeds the preset quantity standard. Furthermore, this operation monitors the current total number of clusters in real time, and terminates recursive splitting once the upper limit is exceeded, thereby preventing overfitting and wasting computational resources, and ensuring the engineering practicality of the clustering results.

[0093] If the value is less than the set condition judgment threshold, the two clusters are marked as similar clusters. The normalized vibration characteristic sequence of each blasting condition in the cluster and the Euclidean distance between the cluster center are obtained and calculated. The blasting conditions that are less than the condition judgment threshold are classified into the same similar blasting condition group.

[0094] In this context, a similar cluster can be a pair of clusters whose Euclidean distance between their cluster centers is less than a threshold for determining the working condition, and is thus marked as having similar working conditions. In an exemplary embodiment, during the recursive clustering process, similar clusters are merged and marked as a single similar cluster unit when the distance between the centers of two sub-clusters is less than the threshold. Furthermore, similar clusters can serve as the basic unit for constructing a group of similar blasting working conditions, ensuring a high degree of consistency in vibration response characteristics within the group. The normalized vibration feature sequence can be a dimensionless, comparable feature vector sequence obtained by standardizing the original vibration signal or PPV-related features. It can be used to eliminate dimensional and amplitude differences caused by different blasting scales, sensor locations, or rock damping, making different working conditions comparable in a unified feature space. In this embodiment, the normalized vibration feature sequence normalizes the peak velocity time history, spectral characteristics, or derived indices of each blasting event according to the maximum value, mean, or standard deviation. For example, the normalized vibration feature sequence can include, but is not limited to, one or more of the following: PPV time history normalized sequence, frequency domain energy normalized sequence, and multi-channel synthetic normalized sequence.

[0095] This operation can be implemented by marking two clusters as similar clusters if the Euclidean distance is less than a set condition judgment threshold. Further, this operation merges cluster pairs that meet the similarity condition into a single logical unit and labels them "similar," thereby identifying sets of conditions with similar vibration response patterns and providing candidates for subsequent group construction. This operation can also involve acquiring and calculating the Euclidean distance between the normalized vibration feature sequence of each blasting condition within a cluster and the cluster center. Further, this operation calculates the L2 distance from the normalized feature vector of each blasting condition within a similar cluster to the cluster center, thereby assessing the deviation of a single condition from the typical pattern of the group. This operation can also group blasting conditions whose Euclidean distance to the cluster center is less than the condition judgment threshold into the same similar blasting condition group. Further, this operation can use a single global condition judgment threshold for filtering, or dynamically adjust the local threshold based on the standard deviation within the cluster to adapt to the dispersion of different clusters, thereby ensuring high consistency in vibration characteristics among group members and providing a high-quality, low-noise sample set for subsequent core mapping prediction.

[0096] For example, in the scenario of fine blasting in the excavation area of ​​a hydropower station dam foundation, the AI ​​deep learning-based method for predicting peak velocity (PV) vibration in engineering blasting can be as follows: In a granite blasting sub-region of the excavation area of ​​a hydropower station dam abutment, the system collects normalized vibration feature sequences of 25 blasts (including the rising slope of the normalized PPV time history, the proportion of the dominant frequency energy, etc.). First, these are divided into two initial clusters. The Euclidean distance between the cluster centers is calculated to be 1.8, which is greater than the preset working condition judgment threshold of 1.5. Therefore, each cluster is further divided into two sub-clusters. After the second split, four clusters are obtained. The center distance between clusters A1 and A2 is 1.6 (>1.5), so further splitting is performed. The distance between B1 and B2 is 1.2 (<1.5), and they are marked as similar clusters. At this point, the total number of clusters has reached 5, which is within the preset number standard of 8, so the process continues. Finally, six clusters are formed, one pair of similar clusters containing eight blasting conditions. The system calculates the distance from each of the eight blasting conditions to the cluster center, removes two outliers with a distance greater than 1.5, and groups the remaining six into the same similar blasting condition group. Subsequently, a core mapping prediction method is applied to this group, significantly improving the PPV prediction accuracy of new blasting points.

[0097] In one embodiment, the division into two clusters is as follows: The actual peak vibration velocity sequences and main influencing parameters of several recent blasting operations were obtained and normalized to obtain normalized vibration characteristic sequences. The K-means clustering algorithm was used to cluster all blasting conditions in the blasting sub-region based on the normalized vibration characteristic sequences, dividing all blasting conditions into two clusters.

[0098] The actual peak velocity (PV) sequences of recent blasting events can be time-series data reflecting the real dynamic process of blasting vibration, providing measured dynamic characteristics of blasting vibration as a fundamental component for constructing high-dimensional features. In an exemplary embodiment, this sequence may include one or more of the following: a single-point three-dimensional PPV time sequence, a multi-point synchronous PPV sequence, or a frequency-domain converted PPV spectrum sequence. The main influencing parameters can be engineering and geological variables that play a dominant role in the propagation of blasting vibration, used to supplement the physical background information of the PPV sequence and enhance the completeness of feature expression. For example, the main influencing parameters may include one or more of the following: charge structure parameters, propagation path parameters, or medium dynamics parameters.

[0099] Normalization is a mathematical transformation process that converts feature variables of different dimensions and magnitudes to a unified numerical range. It eliminates clustering bias caused by differences in units or scale, ensuring the effectiveness of distance metrics. Furthermore, normalization can be achieved through linear or nonlinear mapping, such as scaling each feature dimension to the [0, 1] interval using Min-Max, or employing Z-score standardization to make the mean of each dimension 0 and the standard deviation 1. The normalized vibration feature sequence can be a unified feature vector sequence formed by fusing the actual peak velocity sequence and the main influencing parameters after normalization. This sequence is used to construct a comparable, dimensionless multi-source fusion feature representation, supporting subsequent clustering analysis in a unified feature space. In a specific embodiment, the normalized vibration feature sequence can be one or more of the following: a time-parameter joint normalized sequence, a principal component compression normalized sequence, or a piecewise energy normalized sequence.

[0100] K-means clustering can be an iterative partitioning clustering method based on distance minimization. It divides data into a predetermined number of clusters to achieve initial binary partitioning within blasting sub-regions, separating working condition groups with significantly different vibration response modes. Furthermore, the K-means clustering algorithm can employ standard K-means, K-means++ initialization variants, or Mini-batch K-means. Obtaining the actual peak velocity (PV) sequences and key influencing parameters from several recent blasts can be achieved by extracting PPV time-series data and corresponding engineering parameters from vibration monitoring systems and blasting design databases. This can be further achieved by extracting sensor records and design logs from all blasts performed within the past 7 days, or by retrospectively analyzing blasting logs and monitoring data from the corresponding time period according to the model evaluation update period, thus constructing a complete feature input set containing both dynamic response and static conditions. Normalizing the actual PV sequences and key influencing parameters can be performed by merging the PPV sequence-derived features and key influencing parameters into a feature vector, and then independently normalizing each dimension. Furthermore, this operation can be achieved by scaling each feature dimension to the [0, 1] interval using Min-Max or by using Z-score normalization, thereby eliminating dimensional interference and enabling K-means clustering to effectively reflect the similarity of working conditions based on geometric distance.

[0101] The normalized vibration feature sequence is obtained, which can be a set of normalized high-dimensional feature vectors, with each vector corresponding to a blasting condition. This forms a standard input format suitable for clustering algorithms. The K-means clustering algorithm is used to cluster all blasting conditions within the blasting sub-region based on the normalized vibration feature sequence. K can be set to 2, using the normalized vibration feature sequence as input, and the K-means algorithm is run to complete the initial partitioning. Further, this operation can be improved by using K-means++ to optimize the initial center selection to enhance convergence stability, or by combining the elbow rule to verify the rationality of K=2. This allows for the automatic identification of two dominant vibration modes within the sub-region, providing a starting point for identifying similar blasting condition groups. All blasting conditions are divided into two clusters, which can be done by assigning all blasting conditions within the sub-region to two initial clusters based on the cluster labels output by the K-means algorithm. This completes the initial condition grouping and supports subsequent adaptive subdivision or merging decisions based on inter-cluster distance.

[0102] For example, in the scenario of blasting near a building complex in an urban subway tunnel, the AI ​​deep learning-based peak velocity (PPV) prediction method for engineering blasting in this embodiment can be as follows: Within a blasting sub-region where a subway tunnel passes under a densely built-up area, the system acquires the actual PPV time history (sampling rate 1kHz) of the most recent 10 blasts and the main influencing parameters (including single-hole charge, minimum resistance line, overburden thickness, and stratum shear wave velocity). The PPV time history is extracted into three indicators: peak value, 90% decay time, and dominant frequency energy percentage, which are concatenated with four engineering parameters to form a 7-dimensional vector. After Min-Max normalization, the vector is input into a K-means clustering algorithm (K=2). The clustering results divide the 10 blasts into two clusters: one cluster corresponds to shallow-buried soft soil sections (fast PPV decay, low high-frequency components), and the other cluster corresponds to deep-buried hard rock sections (long PPV duration, abundant high frequencies). This binary division result is used to determine whether further subdivision or merging into similar clusters is needed, ultimately supporting the efficient operation of the core mapping prediction method within highly similar groups.

[0103] In one embodiment, the core mapping prediction method includes: Obtain the Euclidean distance between the normalized vibration characteristic sequence of each blasting condition within a group of similar blasting conditions and the cluster center.

[0104] The core mapping prediction method can be a technique for predicting PPV migration within a group of similar blasting conditions by identifying a core prediction condition and utilizing its historical vibration response relationship with the mapped prediction condition. This approach can achieve efficient and low-overhead PPV prediction within the group, avoiding the repeated running of complex models for each blast point. In this embodiment, the core mapping prediction method can determine the core condition based on the distance between the normalized vibration feature sequence and the cluster center, call its fusion prediction model to obtain a high-precision PPV, and then derive the PPV of other conditions through scaling relationships or mapping functions. Furthermore, the core mapping prediction method can include, but is not limited to, one or more of the following: scaling mapping, nonlinear regression mapping, and local interpolation mapping.

[0105] Obtaining the Euclidean distance between the normalized vibration characteristic sequence of each blasting condition within a similar blasting condition group and the cluster center can be achieved by calculating the L2 norm of the normalized vibration characteristic vector of each condition within the group and the vector of the cluster center. Furthermore, this operation can be implemented by constructing feature vectors from the normalized vibration time-frequency characteristics (such as PPV rise time and dominant frequency energy percentage) and calculating their Euclidean distance to the cluster center obtained from clustering. This allows for the quantification of the deviation of each condition from the typical pattern of the group, providing a basis for selecting the core.

[0106] The blasting condition with the smallest Euclidean distance is taken as the core prediction condition of the similar blasting condition group, and the blasting conditions that are not core prediction conditions are taken as the mapping prediction conditions of the similar blasting condition group.

[0107] The core prediction case can be the representative blasting case among similar blasting case groups whose normalized vibration characteristic sequence has the smallest Euclidean distance to the cluster center. It can serve as a high-quality prediction anchor point, undertaking the complete computational task of the fusion model and providing a prediction benchmark for other cases. In an exemplary embodiment, the core prediction case can be obtained by calculating the distances from all cases within the group to the cluster center and selecting the smallest. Exemplarily, the core prediction case can be input into its dedicated fusion prediction model, and its prediction results are used to drive the generation of PPV for the mapped prediction cases.

[0108] The mapping prediction condition can be any blasting condition in a similar blasting condition group other than the core prediction condition. It can be used to reuse the prediction results of the core condition through a mapping function or proportional relationship to achieve fast PPV estimation. In one specific embodiment, once the mapping prediction condition is determined by the core prediction condition, the remaining members are automatically classified into this category. Furthermore, the mapping prediction condition can be mapped based on the historical actual PPV and the current predicted PPV of the core prediction condition. Using the blasting condition with the smallest Euclidean distance as the core prediction condition of the similar blasting condition group can be achieved by traversing all distance values ​​and selecting the condition corresponding to the smallest distance as the core. Further, this operation can be implemented through sorting or extreme value search algorithms, thereby ensuring that the core condition is most representative and improving mapping reliability. Using blasting conditions that are not core prediction conditions as mapping prediction conditions in the similar blasting condition group can be achieved by marking the remaining conditions in the group as objects to be mapped. Further, this operation can be implemented through set difference operations or Boolean mask filtering, thereby clearly defining the division of labor and distinguishing between conditions requiring precise modeling and those that can be transferred for prediction.

[0109] For the core predicted condition, its specific blasting parameters are obtained and input into the fusion prediction model of the core predicted condition to obtain the predicted peak vibration velocity of the core predicted condition. Based on the proportional relationship or mapping function between the actual peak vibration velocities of the core predicted condition and the mapped predicted condition under similar historical blasting conditions, the predicted peak vibration velocity of the core predicted condition is mapped to obtain the predicted peak vibration velocity of the mapped predicted condition.

[0110] The fusion prediction model for the core prediction condition can be a PPV prediction model specifically built or invoked for that core condition, formed by a weighted fusion of a physical model and a deep learning model. This model can be used to generate high-precision core PPV prediction values ​​as a mapping starting point. In one specific embodiment, the fusion prediction model for the core prediction condition can receive the specific blasting parameters of the core condition and output the predicted peak vibration velocity for mapping.

[0111] The actual peak vibration velocity (PV) under historically similar blasting conditions can be the measured PPV values ​​of the core predicted working condition and the mapped predicted working condition under similar geological and blasting parameters in past blasting events. This value can be used to construct the proportional relationship between the core working condition and the mapped working condition or to train the mapping function. For example, the actual PV under historically similar blasting conditions can include, but is not limited to, one or more of the following: vertical historical PPV pairs, horizontal historical PPV pairs, and synthetic vector historical PPV pairs.

[0112] The proportional relationship can be a scalar ratio between the actual peak velocity (PPV) of the core predicted working condition and the mapped predicted working condition under historical conditions. It can be used to provide the simplest mapping method and is suitable for linear response regions. In an exemplary embodiment, the proportional relationship can be obtained by calculating the historical PPV_map / PPV_core to obtain a fixed or weighted average proportional coefficient. Further, the proportional relationship can include, but is not limited to, a fixed proportional, a working condition weighted proportional, a dynamic sliding window proportional, etc. The mapping function can be a function describing the mathematical relationship between the actual peak velocity of vibration of the core predicted working condition and the mapped predicted working condition under similar historical conditions. It can be used to convert the predicted PPV of the core working condition into the predicted PPV of the mapped working condition, preserving the geological-blasting coupling effect. In this embodiment, the mapping function can be based on historical monitoring data and fit the correspondence between the two PPVs through regression, interpolation, or machine learning methods. Exemplary mapping functions can include, but are not limited to, linear proportional functions, multinomial regression functions, neural network mapping functions, etc.

[0113] Obtaining specific blasting parameters for the core prediction condition can be achieved by extracting parameters such as charge quantity, hole spacing, and rock mass wave velocity from the blasting design database. Furthermore, this operation can be implemented through a structured query language or API call, providing the necessary input for the fusion prediction model. Inputting the specific blasting parameters into the fusion prediction model for the core prediction condition can be done by calling the already constructed fusion prediction model interface and passing in the parameter vector. Further, this operation can be executed through the model inference engine, thus initiating the high-precision PPV prediction process. Obtaining the predicted peak velocity of vibration for the core prediction condition can be achieved by receiving the PPV value output by the fusion model. Further, this operation can be completed through the model output parsing module, thereby obtaining the baseline prediction value for subsequent mapping.

[0114] Based on the proportional relationship or mapping function between the actual peak vibration velocities (PVs) of the core predicted working condition and the mapped predicted working condition under historical similar blasting conditions, this can be achieved by querying a historical database, extracting past PPV pairs under similar conditions, and fitting a ratio or function. Furthermore, this operation can be implemented by directly using historical PPV ratios as fixed proportional coefficients or by using locally weighted regression to fit a nonlinear mapping function, thereby establishing a physically consistent mapping bridge and ensuring the rationality of the transfer. Mapping can be performed separately based on the predicted PV of the core predicted working condition, by substituting the core PPV into the proportional relationship or mapping function to calculate the PPV for each mapped working condition. Furthermore, this operation can be implemented using PPV_map = k × PPV_core (k is the historical ratio) or PPV_map = f(PPV_core) (f is a trained small MLP), thereby enabling an extension from accurate single-point prediction to efficient multi-point transfer. The predicted PV of the mapped predicted working condition can be obtained by outputting the PPV result after mapping calculation. Furthermore, this operation can be completed through a batch post-processing module, thereby completing the PPV prediction for the entire similar group and significantly reducing computational costs.

[0115] Taking the protective blasting of buildings adjacent to urban subway tunnels as an example, the AI ​​deep learning-based engineering blasting vibration peak velocity prediction method in this embodiment can identify a similar blasting condition group containing 5 blasting conditions within a blasting sub-region of a subway tunnel through adaptive clustering. The Euclidean distance between the normalized vibration characteristic sequence (including PPV rise time and dominant frequency energy ratio) of each condition and the cluster center is calculated. Condition 3 is found to have the smallest distance (0.12) and is selected as the core prediction condition. The system retrieves its specific blasting parameters (charge weight 8kg, hole depth 3m, rock wave velocity 4200m / s) and inputs them into its dedicated fusion prediction model, obtaining a predicted PPV of 1.8cm / s. Simultaneously, the system queries the historical database and finds that the measured PPV ratios of this core condition and the other 4 mapped conditions in similar rock strata in the past are 0.92, 1.05, 0.88, and 1.10, respectively. Based on this, the predicted PPVs for the four mapped working conditions were directly calculated to be 1.66, 1.89, 1.58, and 1.98 cm / s, respectively. This eliminates the need to run the fusion model separately for each working condition, saving 75% of computational resources. Furthermore, the error between the predicted results and the subsequent measured results is less than 8%.

[0116] Furthermore, to achieve the above objectives, the present invention also provides an AI deep learning-based engineering blasting vibration peak velocity prediction device, the device comprising: a memory, a processor, and an AI deep learning-based engineering blasting vibration peak velocity prediction program stored in the memory and executable on the processor, the AI ​​deep learning-based engineering blasting vibration peak velocity prediction program being configured to implement the steps of the AI ​​deep learning-based engineering blasting vibration peak velocity prediction method as described above.

[0117] Other embodiments or specific implementations of the AI ​​deep learning-based peak velocity prediction device for engineering blasting vibration described in this invention can be found in the above-described method embodiments, and will not be repeated here.

[0118] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for predicting peak velocity of vibration in engineering blasting based on AI deep learning, characterized in that, The method includes: Obtain basic data for engineering blasting projects and construct an initial physical prediction model for blasting vibrations; The peak velocity of blasting vibration is predicted using an initial physical prediction model, and the accuracy of the prediction is evaluated by comparing the blasting parameter sequence obtained from actual monitoring with the corresponding actual peak velocity sequence. If the accuracy is low, the AI ​​deep learning prediction method is triggered. If the fusion AI deep learning prediction method is triggered, a deep learning prediction model is constructed by combining the actual blasting parameter sequence and the actual peak vibration velocity sequence through the AI ​​deep learning algorithm. The deep learning prediction model and the initial physical prediction model are weighted and fused to construct the fusion prediction model for engineering blasting. The prediction model used for the current blasting condition is determined according to the weight. The engineering blasting area is divided into several blasting sub-regions. It is determined whether the blasting sub-regions trigger the similarity set prediction method. If not, the peak vibration velocity of each blasting point is predicted according to the prediction model adopted. If it is triggered, similar blasting condition groups are selected by clustering algorithm. Within the similar blasting condition groups, the core mapping prediction method is used to predict the peak vibration velocity of blasting. The method for evaluating prediction accuracy is as follows: Set the prediction period, and the interval between adjacent prediction periods is the model evaluation and update period; For a single or multiple blasting events within each prediction period, the actual blasting parameter sequence of the blasting event is obtained, input into the initial physical prediction model, and the predicted peak vibration velocity is obtained; the actual peak vibration velocity generated by the blasting event is collected in real time. Calculate the average absolute percentage error between the predicted peak vibration velocity and the actual peak vibration velocity. If the average absolute percentage error is greater than or equal to the preset error threshold, the prediction accuracy of the initial physical prediction model is judged to be low.

2. The method for predicting peak velocity of engineering blasting vibration based on AI deep learning as described in claim 1, characterized in that, The judgment method of the prediction model used for the current blasting condition is as follows: The prediction model includes a fusion prediction model and an initial physics prediction model; For each blasting condition that triggers the fusion AI deep learning prediction method, a weight evaluation flag value is set and initialized to 0. The weight of the initial physical prediction model in the fusion prediction model is compared with the preset weight threshold. If the weight is greater than or equal to the weight threshold, the weight evaluation flag value is incremented once. If the weight is less than the weight threshold, the weight evaluation flag value is reset to 0. If the weight evaluation flag value is less than the stability flag value, the current blasting condition adopts the fusion prediction model; otherwise, the current blasting condition ends the fusion AI deep learning prediction method and adopts the initial physical prediction model.

3. The method for predicting peak velocity of engineering blasting vibration based on AI deep learning as described in claim 2, characterized in that, The fusion prediction model is obtained as follows: The deep learning prediction model and the initial physical prediction model are obtained and weighted and fused to construct a fused prediction model, wherein the sum of the weights of the deep learning prediction model and the initial physical prediction model is 1. Using a particle swarm optimization algorithm, a set of initial physical prediction model weights is randomly generated. The historical blasting parameter sequence is obtained and input into the fusion prediction model. The mean absolute percentage error between the prediction result and the corresponding actual vibration peak velocity is calculated. By updating the position and adjusting the velocity of the particles, the particle swarm is continuously optimized. The initial physical prediction model weights that minimize the mean absolute percentage error are gradually selected. The weights of the initial physical prediction model are determined, and the fusion prediction model is obtained.

4. The method for predicting peak velocity of engineering blasting vibration based on AI deep learning as described in claim 3, characterized in that, The deep learning prediction model is obtained as follows: Determine whether a deep learning prediction model exists for the current blasting condition; if not, obtain all actual blasting parameter sequences and corresponding actual peak velocity sequences for this type of blasting condition up to the current moment, integrate them according to the time series to obtain the total blasting parameter sequence and the total peak velocity sequence, obtain the training dataset, construct an LSTM model based on the AI ​​deep learning algorithm and train it using the training dataset, and obtain the trained LSTM model as the deep learning prediction model for this condition; If they exist, obtain the latest actual blasting parameter sequence and the corresponding actual vibration peak velocity sequence that were not included in the training dataset to supplement the training dataset, perform incremental training on the deep learning prediction model, and update the deep learning prediction model for this working condition.

5. The method for predicting peak velocity of engineering blasting vibration based on AI deep learning as described in claim 1, characterized in that, The method for determining whether the similarity set prediction method is triggered is as follows: The engineering blasting area is divided into several blasting sub-areas based on geological conditions, blasting zone range, and blasting type. For each blasting sub-area, the number of blasting conditions using the fusion prediction model is obtained, and the ratio of the number of blasting conditions to the total number of blasting conditions in the blasting sub-area is calculated. If the ratio is greater than the preset ratio standard, the similarity set prediction method is triggered in the blasting sub-area.

6. The method for predicting peak velocity of engineering blasting vibration based on AI deep learning as described in claim 5, characterized in that, The method for obtaining the similar blasting condition group is as follows: Divide all blasting conditions within the blasting sub-region into two clusters, calculate the centers of the two clusters and the Euclidean distance between the cluster centers, and compare them with the set condition judgment threshold. If the number of clusters is greater than or equal to the set working condition judgment threshold, each cluster is further divided into two clusters, and the Euclidean distance between the centers of the two clusters is calculated and compared again. If the number of clusters is greater than the preset number standard, the re-clustering and comparison are stopped. If the value is less than the set condition judgment threshold, the two clusters are marked as similar clusters. The normalized vibration characteristic sequence of each blasting condition in the cluster and the Euclidean distance between the cluster center are obtained and calculated. The blasting conditions that are less than the condition judgment threshold are classified into the same similar blasting condition group.

7. The method for predicting peak velocity of engineering blasting vibration based on AI deep learning as described in claim 6, characterized in that, The method of dividing into two clusters is as follows: The actual peak vibration velocity sequences and main influencing parameters of several recent blasting operations were obtained and normalized to obtain normalized vibration characteristic sequences. The K-means clustering algorithm was used to cluster all blasting conditions in the blasting sub-region based on the normalized vibration characteristic sequences, dividing all blasting conditions into two clusters.

8. The method for predicting peak velocity of engineering blasting vibration based on AI deep learning as described in claim 6, characterized in that, The core mapping prediction method includes: Obtain the Euclidean distance between the normalized vibration characteristic sequence of each blasting condition within a group of similar blasting conditions and the cluster center; The blasting condition with the smallest Euclidean distance is taken as the core prediction condition of the similar blasting condition group, and the blasting conditions that are not core prediction conditions are taken as the mapping prediction conditions of the similar blasting condition group. For the core predicted condition, its specific blasting parameters are obtained and input into the fusion prediction model of the core predicted condition to obtain the predicted peak vibration velocity of the core predicted condition. Based on the proportional relationship or mapping function between the actual peak vibration velocities of the core predicted condition and the mapped predicted condition under similar historical blasting conditions, the predicted peak vibration velocity of the core predicted condition is mapped to obtain the predicted peak vibration velocity of the mapped predicted condition.

9. A peak velocity prediction device for engineering blasting vibration based on AI deep learning, characterized in that, The device includes: a memory, a processor, and an AI deep learning-based peak velocity prediction program for engineering blasting vibration stored in the memory and executable on the processor, the AI ​​deep learning-based peak velocity prediction program for engineering blasting vibration being configured to implement the steps of the AI ​​deep learning-based peak velocity prediction method for engineering blasting vibration as described in any one of claims 1 to 8.