A Multi-Indicator Prediction Method for Tunnel Blasting Block Size Based on Zonal Crushing and Crack Radius

By using a multi-index prediction method for tunnel blasting block size based on zonal crushing and crack radius, the problem of obtaining rock strength parameters in tunnel blasting is solved. This method achieves high-precision and interpretable block size prediction, outputs complete gradation curves and non-uniformity coefficients, and supports tunnel blasting design.

CN123116818APending Publication Date: 2026-11-17RAILWAY CONSTR RES INST OF CHINA ACAD OF RAILWAY SCI CO LTD +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611302891.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-26
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for predicting the block size of tunnel blasting suffer from several drawbacks: difficulty in accurately obtaining rock strength parameters, large errors, inability to reflect the fracture range of the rock mass around the blast hole, inability to output complete gradation curves and non-uniformity coefficients, lack of physical constraints in data-driven models, inability to decouple the fracture differences among the three zones, large prediction errors, and poor interpretability.

Method used

A multi-index prediction method for tunnel blasting block size based on zonal crushing and cracking radii is proposed. This method divides the full-section tunnel blast holes into a cutting zone, an auxiliary zone, and a surrounding zone, obtains the crushing zone radius and cracking zone radius of each zone, establishes a semi-empirical prediction formula for single-block size, constructs a physical constraint neural network, and performs zonal weighted synthesis and multi-index output.

Benefits of technology

It achieves high-precision and highly interpretable tunnel blasting block size prediction, and can output complete gradation curves and non-uniformity coefficients. It solves the problems of large prediction errors and poor interpretability of traditional methods, and provides a complete set of blasting design basis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116818A_ABST
    Figure CN123116818A_ABST
Patent Text Reader

Abstract

This invention discloses a multi-index prediction method for tunnel blasting block size based on zoned crushing and crack radius, belonging to the field of tunnel blasting technology. The method divides tunnel blast holes into a cutting zone, an auxiliary zone, and a peripheral zone. The crushing and crack radius of each zone are obtained as core inputs. A semi-empirical mechanism formula for each zone is established and weighted to synthesize an overall gradation curve to generate prior mechanism values. A physically constrained neural network is constructed, whose loss function simultaneously minimizes the data fitting error and the deviation of the network output from the prior mechanism. After training, the overall average block size and uniformity index are output, and the characteristic particle size, non-uniformity coefficient, and complete gradation curve are derived. This invention starts from the physical essence of crushing, precisely characterizes the differences between multiple zones, integrates the advantages of mechanism and data, has high prediction accuracy and reliable extrapolation, and can provide complete block size information at once, meeting the precise control requirements for the resource utilization of tunnel blasting muck.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of tunnel and underground engineering blasting technology, and in particular to a method for intelligent prediction of multiple indicators of blasting block size distribution by using the inversion of the radius of the fractured rock zone around the blast hole and multi-zone weighted synthesis, and by using a physical soft-constraint neural network. Background Technology

[0002] In tunnel drilling and blasting construction, the size distribution of rock blocks after blasting (referred to as block size) is a core indicator for evaluating the blasting effect, directly affecting the muck removal efficiency, secondary crushing costs, and the energy consumption and efficiency of subsequent muck resource processing (crushing, screening, sand making, etc.). Therefore, accurate prediction of blast block size is of great significance for optimizing blasting design and reducing overall construction costs.

[0003] Currently, commonly used methods for predicting blasting block size are mainly divided into two categories: empirical formula methods and data-driven methods. Empirical formula methods, represented by the Kuz-Ram model and its various modifications, are based on parameters such as rock uniaxial compressive strength, explosive consumption, minimum resistance line, and hole spacing, using semi-empirical regression formulas to calculate the average block size X50 and the uniformity index n. Data-driven methods, such as artificial neural networks, random forests, and XGBoost machine learning models, use blasting parameters and rock parameters as inputs and the measured block size as the output for black-box fitting. Furthermore, existing patents mainly focus on blasting parameter optimization methods or blasting block size image recognition devices.

[0004] However, existing technologies still have the following shortcomings: First, the empirical formula method is difficult and prone to errors in obtaining input parameters such as rock strength on-site, and it is difficult to truly reflect the actual fracture range of the rock mass around the blast hole under impact load; Second, the empirical formula method cannot reflect the significant fracture differences caused by different free surface conditions, clamping effects, and charge structures in the cut area, auxiliary area, and surrounding area during full-section tunnel blasting. Treating the entire section as homogeneous rock mass for averaging inevitably produces large systematic deviations, with prediction errors often exceeding 15% in Class III and above surrounding rock; Third, the empirical formula method can only output two statistical features, X50 and n, and cannot provide complete gradation curves, X80, and non-uniformity coefficients for subsequent processing. Fourth, the pure data-driven model lacks any physical mechanism constraints. When the input rock mass parameters or blasting parameters exceed the distribution of the training set, the model is prone to producing prediction results that violate the basic physical laws of rock fragmentation. For example, in the case of a significantly increased radius of the cracked zone (meaning a wider range of crack networks), the model may give an abnormal prediction of smaller block size. Fifth, the pure data-driven model cannot effectively decouple the differences in the three zones of fragmentation, resulting in poor interpretability and making it difficult for field engineers to trust. Sixth, existing published patents do not involve a multi-index prediction technology for block size that uses the radius of the crushed zone and cracked zone around the blast hole as core feature inputs and deeply integrates them with physical soft constraints through partitioned weighted synthesis.

[0005] Therefore, there is an urgent need for a prediction method that starts from the physical nature of rock fragmentation, can accurately reflect the differences between the slotted, auxiliary, and surrounding areas, has both high precision and extrapolation reliability, and can output complete block size information including gradation curves, characteristic particle sizes, and non-uniformity coefficients in one go. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention discloses a multi-index prediction method for tunnel blasting block size based on zonal crushing and crack radius, the technical solution of which is as follows:

[0007] A multi-index prediction method for tunnel blasting block size based on zonal crushing and crack radius, characterized by the following steps: Step 1, obtaining zonal features: dividing the tunnel full-section blast holes into a cutting zone, an auxiliary zone, and a surrounding zone according to function and spatial location, denoted as […]. Obtain the crushing zone radius of each partition. radius of cracked zone and the percentage of rock-broken volume The radius of the crushing zone The radius of the columnar region where the rock around the borehole undergoes pulverization or plastic flow under the action of the blast shock wave; the radius of the cracked zone. This refers to the radius of the region outside the crushing zone where a radial crack network forms under circumferential tensile stress; Step 2, Single-zone mechanism prediction: For each zone, based on rock dynamics and columnar charge theory, using the radius of the crushing zone and the radius of the cracking zone as core inputs, a semi-empirical prediction formula for single-zone block size is established, and the average block size mechanism prediction value for each zone is calculated respectively. And the predicted value of the uniformity index mechanism :

[0008]

[0009]

[0010] in The average borehole depth, This refers to the cross-sectional area of ​​the tunnel excavation. , , , , , , , , Preset constants related to rock properties and explosive type; Step 3, partitioned weighted synthesis: The values ​​of each partition are... and stated Transform them into the Rosin-Rammler distribution function corresponding to each partition. And according to the proportion of rock breaking volume in each zone By performing weighted superposition, the mechanistic synthesis value of the overall gradation curve is obtained. Numerical processing was performed on the synthesis curve of this mechanism to extract the prior average block size of the overall mechanism. and the overall mechanism prior homogeneity index Step 4, Physically Constrained Neural Network Training: Construct a fully connected neural network. Its input layer consists of an 11-dimensional feature vector containing the radius of the crushed zone, the radius of the cracked zone, the proportion of rock-broken volume, the average borehole depth, and the tunnel excavation cross-sectional area for each zone. The output layer is the overall average block size prediction value. and uniformity index prediction value A composite loss function, consisting of a weighted sum of a data fitting loss term and a physical soft constraint loss term, is constructed for network training. The physical soft constraint loss term is used to force the network output to not deviate from the prior average block size and the prior uniformity index of the overall mechanism. Step 5, multi-index output: The corresponding 11-dimensional feature vector of the explosion cycle to be predicted is input into the trained neural network to obtain... and The average block size is calculated and output based on the Rosin-Rammler distribution relationship. Cumulative gradation curve of under-sieve Characteristic particle size and coefficient of uniformity .

[0011] This invention also discloses a multi-index prediction system for tunnel blasting block size based on zonal crushing and crack radius, characterized in that it includes:

[0012] The partition feature acquisition module is used to acquire the radius of the crushed zone, the radius of the cracked zone, and the proportion of rock-broken volume of the blast hole in the full-section excavation zone, auxiliary zone, and surrounding zone of the tunnel.

[0013] The mechanism prior generation module is connected to the partition feature acquisition module and is used to generate the overall mechanism prior average block size and the overall mechanism prior uniformity index based on the crushing zone radius, the cracking zone radius, the rock breaking volume ratio, and the pre-acquired average borehole depth and tunnel excavation cross-sectional area of ​​each partition.

[0014] The physical constraint neural network module, connected to the mechanism prior generation module, is used to train the neural network with the radius of the crushing zone, the radius of the cracking zone, the proportion of rock breaking volume, the average blast hole depth and the tunnel excavation cross-sectional area of ​​each partition as inputs, and with the overall mechanism prior average block size and the overall mechanism prior uniformity index as physical constraints, and output the overall average block size prediction value and uniformity index prediction value.

[0015] The multi-index output module, connected to the physical constraint neural network module, is used to calculate and output the average block size, undersize cumulative gradation curve, characteristic particle size, and non-uniformity coefficient based on the overall average block size prediction value and the uniformity index prediction value, according to the Rosin-Rammler distribution relationship.

[0016] Beneficial effects

[0017] (1) Direct physical characteristics: This invention directly uses the radius of the crushing zone, which characterizes the range of pore wall crushing, as the physical characteristics. and the radius of the cracked area As input features, indirect and difficult-to-measure input parameters such as rock strength and explosive consumption are discarded, and the final effect of explosive-rock interaction is captured from the root, which greatly enhances the physical interpretability of the model.

[0018] (2) Fine depiction of differences in multiple zones: This invention independently obtains the crushing radius of the three zones of slotting, auxiliary and peripheral zones and weights them to synthesize the overall gradation curve, which fully reproduces the essence of non-uniform crushing of the entire tunnel cross section and solves the problem that the traditional single-zone equivalent model cannot describe the distortion of the distribution of large block ratio and the large deviation of the powder ratio estimation.

[0019] (3) Mechanism-data deep fusion: This invention introduces a mechanism prior bias term in the loss function through the architecture of “partition mechanism synthesis prior + physical soft constraint neural network”, which forces the network output to not deviate from the physical reasonable range. It not only explores the high-order nonlinear relationship in the data, but also ensures the prediction trend that conforms to the laws of mechanics under the extreme working condition of scarce data, thus overcoming the extrapolation risk of pure black box model.

[0020] (4) Parallel output of multiple metrics: This invention is based on the network output after training. and The average block size is obtained by calculating it once based on the Rosin-Rammler distribution relationship. Complete gradation curve Characteristic particle size and coefficient of uniformity This provides a complete set of guidelines for selecting loading equipment, setting crusher parameters, and predicting finished product yield.

[0021] (5) Adaptive evolution capability: This invention utilizes an incremental training mechanism and physical constraint weights. The gradual decay strategy allows the model accuracy to continuously evolve as construction data accumulates, maintaining high prediction performance even when surrounding rock conditions change. Attached Figure Description

[0022] Figure 1 This is an overall flowchart of the method of the present invention;

[0023] Figure 2 A schematic diagram of the principle of multi-zone weighted synthesis of the overall gradation curve;

[0024] Figure 3 This is a schematic diagram of the structure of a physical soft-constraint neural network. Detailed Implementation

[0025] Example 1

[0026] See Figure 1 As shown in the figure, this embodiment provides a multi-index prediction method for tunnel blasting block size based on zonal crushing and crack radius. This method involves zoning the full-section tunnel boreholes, establishing semi-empirical prediction formulas for single-zone block size in each zone, weighted synthesis of overall mechanism priors, constructing and training a physically constrained neural network, and finally achieving parallel output of multiple indices, thereby completing the accurate prediction of tunnel blasting block size. This method is particularly suitable for drilling and blasting construction scenarios of full-section or large-section tunnels with complex surrounding rock conditions and significant cross-sectional non-uniformity.

[0027] Step 1: Obtain partition features

[0028] First, the tunnel's full-section blast holes are divided into three zones based on function and spatial location: the cut zone, the auxiliary zone, and the peripheral zone. This zoning is necessary because in full-section tunnel blasting, the cut zone is located near the center of the section. Its blast holes are typically arranged in a straight-hole or oblique-hole cut pattern. During blasting, it faces two free surfaces (the bottom of the hole and the cut cavity). Under strong impact loads, the rock is mainly thrown outwards axially and radially, with fracturing primarily consisting of intense impact crushing and tensile cracking, exhibiting significant clamping effect. The auxiliary zone is located between the cut zone and the peripheral zone. During blasting, a free surface has already been formed, and the clamping effect of the blast holes is weaker than in the cut zone. However, it has the largest number of blast holes and bears the main collapse function, accounting for the largest proportion of rock volume. The peripheral zone is located near the tunnel's cross-section outline. The blast hole spacing is smaller, and the charge is less. During blasting, the goal is to create a smooth outline. Rock fracturing is mainly due to shearing and tensile fracture along the outline surface, with limited crack propagation. The free surface conditions, clamping effects, and charge structures of the three zones mentioned above are significantly different. If the entire cross section is treated as a homogeneous rock mass and averaged out, it will be impossible to accurately depict the actual fracturing effect of each zone. Therefore, it is necessary to model the three zones separately.

[0029] For each partition, obtain the following three core features: crushing zone radius. radius of cracked zone and the percentage of rock-broken volume subscript These correspond to the slotting area, auxiliary area, and surrounding area, respectively.

[0030] Crushing zone radius The crushing zone refers to the radius of the cylindrical region around the borehole where the rock undergoes pulverization or plastic flow under the influence of the blast shock wave. Its physical essence lies in the fact that when the explosive detonates inside the borehole, the detonation products instantly release enormous energy, creating an impact pressure of several GPa on the borehole wall. This pressure far exceeds the dynamic compressive strength of the rock, causing the rock near the borehole wall to undergo inelastic deformation within an extremely short time (microseconds to milliseconds), resulting in pulverization. The size of the crushing zone directly reflects the proportion of shock wave energy dissipated in the pulverization process—a larger radius means more energy is consumed in rock pulverization and plastic flow, correspondingly reducing the energy available for subsequent crack propagation.

[0031] radius of crack zone This refers to the radius of the region outside the crushing zone where a radial crack network forms under circumferential tensile stress. Its physical essence lies in the following: the blast shock wave propagates outward as a compression wave; when the wavefront reaches the free surface or rock interface, it is reflected, forming a tensile wave. Simultaneously, at the boundary of the crushing zone, the tangential circumferential tensile stress exceeds the dynamic tensile strength of the rock, causing radial cracks to initiate and propagate outward from the boundary of the crushing zone. The density and extent of the crack network directly determine the initial size distribution of the rock fragments after the blast—the larger the cracked area, the more developed the crack network, and the finer the rock mass is fragmented. Therefore, and It is a direct physical variable that determines the size of the blasted fragments, reflecting the combined result of explosive energy distribution and rock dynamic response.

[0032] and The following technical approaches can be used to obtain the data: Establish a columnar charge fluid-structure interaction numerical model for typical boreholes in each zone. This model simultaneously considers the interaction between the explosive detonation products (fluid domain) and the rock (solid domain). The Euler-Lagrange coupling algorithm is used to track the propagation process of the blast shock wave in the rock medium and the resulting damage evolution. After outputting the spatial distribution of the damage zone around the borehole, the crushing zone is defined using an equivalent plastic strain threshold—when the effective plastic strain of the rock element exceeds a preset critical value (usually taken as...). When the maximum principal tensile stress exceeds the dynamic tensile strength of the rock element, the element is determined to be in the crushing zone. This threshold is based on the dynamic plastic constitutive relationship of the rock under impact load, reflecting the critical state of irreversible pulverization failure. The cracking zone is defined by the region where the maximum principal tensile stress exceeds the dynamic tensile strength of the rock element. When the maximum principal tensile stress of a rock element exceeds its dynamic tensile strength, the element is considered to have undergone tensile cracking failure. This criterion is based on the maximum tensile stress failure criterion, a classic criterion for brittle fracture of rock under explosive tensile wave action. The results obtained from numerical simulation need to be calibrated online in conjunction with on-site acoustic wave testing or borehole television inversion to ensure that the simulation parameters can truly reflect the actual mechanical response characteristics of the rock mass. This calibration process can use the Bayesian inversion method to assimilate and fuse the simulation output with the measured data, gradually correcting the constitutive model parameters.

[0033] also, and Alternatively, it can be obtained through a pre-trained surrogate model of "blasting parameters - fracture zone radius". This surrogate model takes the hole spacing, resistance line, and charge decoupling coefficient of each zone as input, and outputs the corresponding crushing zone radius and cracking zone radius. This surrogate model can be established using a Kriging model or Gaussian process regression. Its advantage is that once trained, it can calculate the fracture zone radius for any new working condition within milliseconds without having to rerun time-consuming numerical simulations, greatly improving the real-time performance for field use.

[0034] Rock breaking volume percentage The volume of rock borne by each borehole in each section is determined by the ratio of the total volume of the entire cross-section, satisfying the normalization constraint. This proportion is related to the density of blast holes, the depth of blast holes, and the area covered by each zone, reflecting the weight of blasting workload in each zone in the overall blasting effect.

[0035] Step 2: Single-region mechanism prediction

[0036] After obtaining the crushing zone radius of each partition. radius of cracked zone and the proportion of rock breaking volume Next, an independent block size prediction formula needs to be established for each zone. The technical necessity of this step lies in the fact that the fragmentation mechanisms of the three zones are fundamentally different, and they cannot share the same set of averaging parameters; therefore, the average block size of each zone must be established separately. and uniformity index The mechanism prediction formula enables the independent variables in each formula to independently reflect the specific fragmentation pattern of the area.

[0037] For each zone, based on rock dynamics and cylindrical charge theory, a semi-empirical prediction formula for single-block size is established, expressed in terms of the radius of the crushing zone and the radius of the fracture zone:

[0038] (1)

[0039] (2)

[0040] The physical basis and necessity of introducing each factor are explained below.

[0041] about Linear proportional term: average block size The fundamental dimension is length, while the radius of the cracked zone is... The dimension of is also length, therefore it is denoted by in terms of dimensions. Using it as a baseline variable is the most direct choice. From a physical perspective, the average size of the rock blocks after blasting primarily depends on the scale of the crack network that divides the entire rock. A larger crack radius means a longer radial crack propagation distance, more thorough crack penetration between adjacent boreholes and between boreholes and the free surface, resulting in a greater number of blocks, but the average block size actually increases. The physical logic here needs to be correctly understood: a large crack radius does not necessarily mean a small block size, but rather a larger "fractured area" covered by cracks, increasing the number of large blocks formed from the original rock mass. Specifically, with a fixed borehole spacing, an increased crack radius leads to earlier overlap between the cracked areas of each borehole. The block size formed by crack penetration along the borehole centerline is determined by the borehole spacing; a larger crack radius means a longer single crack extension distance, resulting in larger characteristic blocks of rock mass "cut off." Therefore, and The positive correlation between the two phenomena has been statistically verified in numerous blasting engineering practices and forms the cornerstone of the mechanism of this invention.

[0042] about Ratio term (where The physical meaning of this ratio is the ratio of the crushing zone to the cracking zone; it reflects the distribution of explosive energy between crushing dissipation and cracking work. Crushing zone radius. The larger the borehole diameter, the more explosive energy is consumed in the crushing and plastic deformation of rock near the borehole wall. This energy can no longer be used to drive long-distance crack propagation, thus reducing crack density. The opening and extension length of each crack are limited, and the connections between rock blocks are more difficult to completely break, ultimately leading to an increase in average block size. Conversely, if... A smaller ratio indicates that most of the energy is effectively converted into the driving force for crack propagation, resulting in a denser crack network with smaller blocks. Therefore, this ratio term is mechanistically necessary, and The positive value is determined by the principle of energy conservation—there is a competition between the energy consumed in crushing and the driving force for cracking, which is a fundamental law of rock blasting dynamics.

[0043] Correction function for hole depth clamping Hole depth The impact on blasting fragment size is significant, and its physical mechanism lies in the confinement effect. As the borehole depth increases, the constraint on the rock at the bottom of the hole strengthens, limiting the expansion space of the explosive gases and delaying the time it takes for the reflected tensile wave to reach the bottom of the hole. This makes rock fragmentation under impact loads more difficult, resulting in a larger fragment size. The correction function adopts a linear form. ,in As the reference hole depth, This is the clamping effect coefficient. This linear correction is not an arbitrary assumption, but a first-order approximation derived from the one-dimensional wave equation of stress wave propagation along the bore axis in a cylindrical charge: Under cylindrical charge conditions, the attenuation of the axial stress wave amplitude with propagation distance approximately follows an exponential law. Expanding this into a Taylor series and retaining the first-order term yields the linear approximation. When the range of variation is not particularly large (e.g.) It has good approximate accuracy. This ensures that the physical law that the greater the hole depth, the larger the block size is correctly described.

[0044] Regarding the cross-section correction function Tunnel excavation cross-sectional area This reflects the size of the free face and the space available for the free expansion of explosive gases, directly affecting the throwing conditions and the degree of secondary collision fragmentation. A larger cross-section provides more space for the explosive gases to expand, increasing the probability of collisions between rock fragments during throwing and resulting in smaller fragment sizes; conversely, a smaller cross-section leads to stronger clamping effects and larger fragment sizes. This correction function uses a power function form, where... The reference cross-sectional area, This is the cross-sectional influence index. Its power function form originates from the adiabatic expansion equation of explosive gases. , of which the cross-sectional area Approximately proportional to the characteristic scale of the expansion volume, the work capacity of gas expansion is related to the power of the volume, therefore, using a power function correction is theoretically reasonable.

[0045] Regarding the formula for the uniformity index Uniformity Index It is a key parameter in the Rosin-Rammler distribution that characterizes the degree of concentration in the granularity distribution. The larger the value, the steeper the gradation curve and the more uniform the block size distribution. (Using...) As a core independent variable, its physical basis lies in the fact that when the fractured zone is larger than the crushed zone, the crack network is more fully and uniformly distributed in space, the crack density in all directions tends to be consistent, and the size variability of the rock block decreases. Therefore, the uniformity index... Increase. Due to Taking a negative value, the expression is equivalent to and Positive correlation, and The positive correlation is physically self-consistent: the larger the crushing zone, the more energy is dissipated during crushing, the lower the crack density and the more uneven the distribution, the larger the block size and the worse the uniformity; both of these conditions hold true simultaneously. (Sectional correction term) The reason for its introduction is that the larger the cross-section, the less constraint there is on rock fragmentation, the greater the space for cracks to propagate freely, and the homogeneity of the crack network is slightly improved. Taking a positive value reflects this trend.

[0046] constant , , , , , , , , These are pre-defined constants related to rock properties and explosive types. These constants are not purely empirical regression coefficients, but rather characteristic parameters with explicit physical meanings: The comprehensive proportionality coefficient reflects the energy coupling efficiency of explosive and rock. The larger the value, the larger the average block size corresponding to the unit fracture zone radius. The crushing-cracking energy distribution index is used to control the nonlinear influence of crushing energy consumption on particle size. The baseline uniformity coefficient; The crack density uniformity index; This is the index of the influence of cross-sectional dimensions on uniformity. The above constants can be determined in a small number of calibration tests, for example, by measuring the actual block size distribution through 3 to 5 sets of field blasting tests with different charge structures, and then calculating the optimal combination of constants.

[0047] Step 3: Prior to the Partition-Weighted Synthesis Overall Mechanism

[0048] The single-zone mechanism formula calculates the independent average block size and uniformity index of each zone, but in actual engineering, the focus is on the overall block size distribution of the entire tunnel cross-section. Therefore, it is necessary to combine the independent prediction results of each zone into an overall gradation curve. This weighted synthesis process is the key technology of this invention that reflects the essential differences between multiple zones.

[0049] First, the average block size mechanism prediction value of each partition is calculated. And the predicted value of the uniformity index mechanism The distribution is then converted to the Rosin-Rammler distribution function corresponding to each partition. The Rosin-Rammler distribution is chosen over other distribution functions because it is the most widely validated and accepted undersize cumulative distribution model in the field of rock crushing. Its mathematical form is concise and can completely describe the entire gradation curve using only two parameters (characteristic particle size and evenness index). Both parameters have clear physical meanings, facilitating integration with mechanistic formulas. The specific conversion method is as follows:

[0050] (3)

[0051] Characteristic particle size Depend on and Derived from the intrinsic relationship of the Rosin-Rammler distribution:

[0052] (4)

[0053] The derivation of the above transformation relationship is as follows: In the Rosin-Rammler distribution, the cumulative pass rate ,make ,Right now Organized Taking the natural logarithm of both sides, we get Therefore ,and then ,Right now This derivation guarantees that from arrive The mathematical precision of the transformation.

[0054] Get each partition Then, based on the proportion of rock-breaking volume in each zone... By performing weighted superposition, the mechanistic synthesis value of the overall gradation curve of the entire cross section is obtained:

[0055] (5)

[0056] The physical meaning of this weighted superposition is: the proportion of rock-breaking volume in each zone. This reflects the mass share of the total rock debris generated in that zone within all the cave debris. The overall gradation curve is essentially a mass-weighted average of the gradation curves of each zone. This weighting method implicitly assumes a reasonable assumption—that the rock blocks in each zone have the same density (i.e., the physical density of the rock is uniform across the cross-section), and therefore holds true in engineering practice.

[0057] right Within the particle size range Numerical processing is performed internally: This is achieved through linear interpolation. The corresponding particle size is used as the prior average size of the overall mechanism. The reason for choosing a particle size with a throughput of 50% is... The definition is based on the fact that it is an internationally accepted standard for characterizing blasted block size, which facilitates comparison and verification with the prediction results of the Kuz-Ram model, which is widely used in engineering practice.

[0058] Overall mechanism prior homogeneity index Extraction is achieved through the following methods: in pass rate to Within the interval (excluding the numerically unstable regions near 0% and 100%, as these regions are unreliable due to strong statistical fluctuations in sample size), with For dependent variable, Performing least squares linear regression on the independent variable yields the slope. The theoretical basis of this method lies in the Rosin-Rammler distribution. By performing mathematical transformations, we can obtain , that is to x-axis The curve plotted with the vertical axis is a straight line, and its slope is exactly equal to the uniformity index. Therefore, the slope is extracted through linear regression. It is a rigorous and precise mathematical method, not an engineering approximation.

[0059] See Figure 2 As shown, the overall mechanism prior value obtained through the above steps and It incorporates independent physical mechanisms and weighted mixing effects across three zones, providing a high-quality physical benchmark for subsequent data-driven models. The advantage of this prior value lies in its ability to provide block size estimates that conform to the basic physical laws of rock fracturing, even under new working conditions without measured data.

[0060] Step 4: Construction and Training of Physically Constrained Neural Networks

[0061] While prior values ​​of mechanistic mechanisms exhibit good physical consistency and extrapolation stability, their prediction accuracy is limited by the simplification assumptions of semi-empirical formulas and errors in constant values. On the other hand, while purely data-driven neural networks can fit complex high-order nonlinear relationships, they are prone to producing predictions that violate physical laws when training samples are limited or input parameters exceed the sample distribution range. The inventive contribution of this invention lies in deeply integrating the advantages of both approaches to construct a physically soft-constrained neural network.

[0062] Specifically, a fully connected neural network is constructed. The input layer consists of an 11-dimensional feature vector.

[0063] (6)

[0064] The design logic for this 11-dimensional input is: 3 parameters in each of the three zones ( , , There are a total of 9 dimensions, plus global parameters. and It has 2 dimensions in total, for a total of 11 dimensions. The completeness of this feature set is reflected in its ability to independently characterize the fragmentation state of each region (through...). and It can also reflect the relative workload of each district (through...) It can also characterize the overall hole depth and cross-sectional conditions (through...). and There is no missing key physical information.

[0065] The output layer has two neurons, which output the overall average block size prediction value. and uniformity index prediction value The output layer activation function uses the Softplus function. Its advantage lies in: compared to ReLU, Softplus... The surrounding area exhibits smooth transition characteristics, preventing gradient vanishing completely, and the output value remains consistently positive, which conforms to... and All physical constraints are positive real numbers. The hidden layers are designed to have at least three layers; a preferred network structure is 128-64-32, meaning the first hidden layer has 128 neurons, the second hidden layer has 64 neurons, and the third hidden layer has 32 neurons. The decreasing number of neurons layer by layer helps to gradually extract more abstract feature representations from the high-dimensional input space, while also reducing the risk of overfitting. Each hidden layer uses the ReLU activation function. This allows for the introduction of nonlinear expressive power and sparse activation properties.

[0066] During training, a composite loss function containing two terms is constructed to mathematically and precisely express the dual optimization objectives of data fitting and physical consistency:

[0067] (7)

[0068] in:

[0069] The total number of training samples, The sample number;

[0070] The first term is the conventional data fitting loss (mean squared error), which drives the network output to approximate the measured values. and This is the foundation for networks to achieve high accuracy;

[0071] The second term is the physical soft constraint loss, which forces the network output to not significantly deviate from the mechanism prior value generated in step 3. and This is crucial to ensuring that network predictions conform to physical laws;

[0072] These are physical constraint weights, and their value range is... The engineering basis for this value is that the prior confidence level of the mechanism is usually high (based on mature physical theories), but it is not completely accurate. Therefore, [the following is a possible interpretation:] Set to a moderately small value so that the network prioritizes fitting the measured data while moderately referencing prior mechanistic guidance. As the amount of training data gradually increases... The attenuation can be gradually reduced according to a preset attenuation strategy, so that the network can autonomously discover more refined data patterns when it has sufficient data support, while still retaining the prior mechanism as a safety constraint.

[0073] It's called a "soft constraint" rather than a "hard constraint" because the mechanistic prior term in the loss function isn't a mandatory equality constraint, but rather exists as a penalty term—the loss increases when the network output deviates from the mechanistic prior, but this deviation can still occur under the dominance of the data fitting loss, provided that the measured data provides a sufficiently strong signal. The advantage of this soft constraint mechanism is that it neither forcibly suppresses the true patterns reflected in the data like hard constraints, nor does it maintain the physical reasonableness of the prediction when data is scarce.

[0074] The network training employs the Adam optimizer (Adaptive Moment Estimation), which combines the advantages of momentum optimization and adaptive learning rate adjustment strategies. It adaptively adjusts the learning rate for each parameter by calculating the first and second moments of the gradient, exhibiting superior convergence performance when handling non-convex optimization problems and high-dimensional parameter spaces. The initial learning rate is set to... This is the recommended default value for the Adam optimizer in deep learning practice, which balances convergence speed and convergence stability; each After each training round, multiply the learning rate by the decay factor. This exponential decay strategy helps achieve more refined local optimization in the later stages of training. The batch size is set to... That is, the gradient is calculated and the parameters are updated once every 32 samples are randomly selected from the training set. This batch size achieves a reasonable balance between computational efficiency and gradient estimation stability.

[0075] During training, the gradient of the composite loss function is propagated layer by layer through the backpropagation algorithm, and the weights and bias parameters of each layer of the network are updated according to the partial derivative of the loss function with respect to each parameter. Mathematically, for the th... Layer weight matrix and bias vector The update rules are as follows:

[0076] (8)

[0077] in This represents the current learning rate. The calculation of this partial derivative relies on the chain rule, propagating the error of the output layer back to the input layer layer by layer. The criterion for training convergence is: the loss value on the validation set no longer significantly decreases after 50 consecutive training iterations.

[0078] See Figure 3As shown, after the above training is completed, the network has both high accuracy of data fitting and mechanism-guided extrapolation stability. Within the training sample distribution range, the network is dominated by data fitting loss, and the prediction accuracy is better than that of the pure mechanism model. Under new working conditions that exceed the training sample distribution, the predicted value output by the network will be pulled back to the reasonable range indicated by the mechanism prior by the physical soft constraint, and will not produce abnormal results that violate the basic physical laws of rock fracturing.

[0079] Step 5: Multi-indicator output

[0080] The 11-dimensional feature vector of the explosion cycle to be predicted is input into the trained network to obtain the overall average block size prediction value. and uniformity index prediction value Based on this, and utilizing the inherent mathematical relationships of the Rosin-Rammler distribution, the following multiple indices are calculated and output:

[0081] (1) Characteristic particle size

[0082] (9)

[0083] This parameter is a characteristic scaling parameter of the Rosin-Rammler distribution, and its physical meaning is the cumulative screening rate. (Right now The corresponding particle size at time ) is greater than It can more sensitively reflect the overall position of the gradation curve.

[0084] (2) Cumulative gradation curve of undersize material

[0085] (10)

[0086] This function provides arbitrary particle size. The corresponding cumulative pass rate is a complete continuous curve that contains all the information about the gradation, and has richer engineering guidance value than a single statistical indicator.

[0087] (3) 80% pass through particle size

[0088] (11)

[0089] The engineering significance of this indicator is as follows: It is a key design basis for the setting of the discharge port of the jaw crusher and the selection of the screen hole of the vibrating screen in the crushing and screening process. About 80% of the slag can pass through the screen hole of this size to enter the next stage of processing, which directly determines the throughput capacity and energy consumption of the crushing production line.

[0090] (4) Coefficient of non-uniformity

[0091] (12)

[0092] in The particle size corresponding to a 60% throughput is... This represents the particle size corresponding to a 10% pass rate. The uniformity coefficient is a classic indicator in geotechnical engineering for evaluating the quality of particle size distribution. A higher value indicates a more uneven particle size distribution. The smaller the value, the more uniform the particle size. The derivation of this formula is as follows: For the Rosin-Rammler distribution... pass rate The corresponding particle size satisfies ,therefore This ratio depends only on the uniformity index. , with characteristic particle size This is irrelevant; it aligns with the definition of a uniformity index—it only reflects the dispersion of the distribution, not its overall location. The engineering significance of this index lies in: when... When the blasted pieces are very uniform in size, it indicates that the blasted pieces are conducive to the stable operation of subsequent crushing production; when If the particle size distribution is too dispersed, it indicates that the blasting parameters need to be optimized to improve the crushing effect.

[0093] The simultaneous output of these multiple indicators provides comprehensive data support for on-site engineering decisions. A comprehensive level used for rapid assessment of blasting effects; complete gradation curve. Used to accurately match the optimal operating parameters for subsequent crushing and screening processes; It is directly used for the design of the crusher discharge port and the selection of the screen aperture size of the vibrating screen; Used to evaluate the operational stability of a crushing system and to optimize the uniformity of blasting parameters.

[0094] Steps 1 to 5 above constitute a complete "physical mechanism-data driven" deep fusion method for predicting the multi-index of tunnel blasting block size. This method first generates high-quality mechanism priors through independent modeling and weighted synthesis in three regions, and then fuses and trains the mechanism priors with measured data through a physical soft-constraint neural network, ultimately achieving parallel output of multiple indices with both high accuracy and extrapolation stability.

[0095] To further enhance the model's adaptability to changes in surrounding rock conditions, this method can also introduce an adaptive incremental learning mechanism. When the change in the rock mass integrity coefficient (such as RQD value or rock mass wave velocity ratio) at the construction site exceeds a preset threshold (e.g., the rate of change exceeds...), an adaptive incremental learning mechanism can be introduced. When the comprehensive block size evaluation value is lower than the preset level for multiple consecutive cycles (e.g., three consecutive cycles), incremental training of the neural network is automatically triggered. This means retaining existing network parameters as initial values ​​and continuing training for several rounds using newly collected field data, rather than retraining from scratch. During incremental training, the physical constraint weights are gradually reduced according to preset rules based on the amount of new data accumulated since the current changes in surrounding rock conditions. The value, for example, will be increased for every 10 new sets of data. reduce until it decreases to The lower limit value. This mechanism ensures that the model can evolve autonomously as the surrounding rock conditions change, maintain high prediction performance in the long term, and avoid model solidification caused by over-reliance on old data.

[0096] The technical solution provided in this embodiment fully discloses the technical solution of the present invention through steps 1 to 5 and their specific implementation details. It fully embodies the creative technical concept of taking the crushing zone and cracking zone radius as the core input, multi-zone weighted synthesis mechanism prior, physical soft constraint neural network training, and multi-index parallel output, and has significant creativity and industrial applicability.

[0097] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0098] Application Example 1: Full-face blasting of a Class III granite railway tunnel

[0099] A single-track railway tunnel has surrounding rock of Class III medium-coarse-grained granite. The dynamic compressive strength of the rock is... Dynamic tensile strength The tunnel was excavated using a single blasting operation across its entire cross-section, with a cross-sectional area of... Cycle advance 3.5m, average borehole depth The borehole diameter was 45mm, and No. 2 rock emulsion explosive with a charge diameter of 32mm was used. A total of 128 boreholes were arranged across the entire cross-section, including 20 slotting holes, 72 auxiliary holes, and 36 peripheral holes.

[0100] Step (1) Partition feature acquisition:

[0101] Fluid-structure interaction numerical simulations were performed on typical single-hole charge structures in three zones using dynamic finite element software. Combined with acoustic inversion and borehole television observations from three field test boreholes, the radii and volume proportions of representative fractured zones in each zone were obtained as follows:

[0102] ①Slotting area ( ): , , ;

[0103] ②Auxiliary area ( ): , , ;

[0104] ③ Surrounding areas ( ): , , .

[0105] The volume percentage is calculated by multiplying the borehole area of ​​each zone by the cyclic advance.

[0106] Steps (2)-(3) prior calculation of mechanism:

[0107] Pick Substituting into the formula, the values ​​for each region are calculated. and After weighted synthesis, the prior values ​​of the overall mechanism are extracted. , .

[0108] Step (4) Network training:

[0109] Data was continuously collected from 42 blasting cycles at the site, and the measured values ​​for each cycle were extracted using an image analysis system. , And the gradation curve. After removing duds and severely over-mined abnormal samples, 40 sets of valid data were obtained. 30 sets were randomly selected as the training set and 10 sets as the test set.

[0110] The network uses a three-layer structure of 128-64-32, with a Softplus output layer. It employs the Adam optimizer, a batch size of 32, an initial learning rate of 0.001, and decays to 0.8 every 200 epochs for a total of 1000 epochs. Physically constrained weights are used. .

[0111] Step (5) Result Evaluation:

[0112] After training, the prediction results on the 10 test sets were as follows: Root mean square error 0.62 cm, coefficient of determination ; The root mean square error is 0.028. One typical cycle (original scheme) forecast output The gradation curve almost perfectly matches the measured value. This cycle... and Substituting into the formula, we get , This is consistent with the image analysis results.

[0113] Extrapolation capability verification: The radius of the cracked area in the slotted zone of this cycle is manually determined. The depth was increased from 120cm to 150cm (simulating abnormal development of surrounding rock fractures), while other inputs remained unchanged. The physical soft-constraint neural network of this invention was used for prediction. The size increased from 35.0 cm to approximately 48.5 cm, an increase of 38.6%, which aligns with the physical intuition that crack network expansion leads to increased block size; while an unconstrained pure neural network trained with the same data, under the same input, The predicted value, however, decreased from 35.0 cm to 31.4 cm, exhibiting a clear physical paradox. This comparison fully demonstrates the crucial role of the soft constraint mechanism of this invention in extrapolating stability.

[0114] Application Example 2: Bench blasting of Class IV limestone highway tunnels

[0115] A highway tunnel is located in a Class IV surrounding rock formation, consisting of thin-layered limestone with a dynamic compressive strength of 85 MPa and a dynamic tensile strength of 7 MPa. The upper and lower bench method is used. The cross-sectional area of ​​the upper bench is... Circular advance 2.5m, average hole depth The borehole diameter is 42mm. The parameters for the three zones were obtained through simulation and inversion: ; ; .

[0116] Due to the decrease in rock strength, the constant is adjusted to Everything else remains the same. Collect 30 sets of valid data, 20 sets for training and 10 sets for testing. Adjust the network structure to 96-48-24. Test set The root mean square error is 0.51 cm. This embodiment demonstrates that the method of the present invention can be adapted to different lithologies and cross-sections by adjusting a small number of constants, exhibiting good universality.

[0117] The two application examples above, from hard rock to soft rock and from full-section to step method, both verify the effectiveness and reliability of the present invention.

[0118] Example 2

[0119] This embodiment provides a multi-index prediction system for tunnel blasting block size based on zonal crushing and crack radius. This system is used to execute the prediction method described in Embodiment 1 above. This embodiment focuses on describing the system's module composition and its working process.

[0120] The system includes a partition feature acquisition module, a mechanism prior generation module, a physical constraint neural network module, and a multi-index output module. These modules are connected in sequence to form a complete data processing pipeline.

[0121] The zone feature acquisition module is used to obtain the radius of the crushing zone, the radius of the cracking zone, and the percentage of rock-breaking volume of the blast holes in the full-section cut zone, auxiliary zone, and surrounding zone of the tunnel. The data source for this module can be a field numerical simulation database or a pre-trained blasting parameter-crushing zone radius proxy model. When using a proxy model, the module takes the hole spacing, resistance line, and charge decoupling coefficient of each zone as input and automatically calculates the corresponding crushing zone radius and cracking zone radius. The output of this module is three sets of feature parameters. , , (in ), and global parameter average borehole depth and tunnel excavation cross-sectional area Together, they form the basic data of the 11-dimensional feature vector.

[0122] The mechanism prior generation module is connected to the partition feature acquisition module. It generates the overall mechanism prior average block size and overall mechanism prior homogeneity index based on the crushing zone radius, cracking zone radius, rock breaking volume ratio, and pre-acquired average borehole depth and tunnel excavation cross-sectional area of ​​each partition. This module contains three sub-units: a single-partition calculation unit, a weighted synthesis unit, and a prior extraction unit. The single-partition calculation unit receives data from the partition feature acquisition module. , , , The data is processed according to the formula for each partition. and Calculate the average block size mechanism prediction value and the uniformity index mechanism prediction value for each partition, where each constant... , , , , , , , , The system parameters are preset based on rock properties and explosive type. The weighted synthesis unit combines the data from each partition. and Transform them into the Rosin-Rammler distribution function corresponding to each partition. (in ), and according to the proportion of rock breaking volume in each zone By performing weighted superposition, the mechanistic synthesis value of the overall gradation curve is obtained. The prior extraction unit performs numerical processing on the mechanism synthesis curve: the particle size corresponding to a throughput of 50% is obtained through linear interpolation and used as the overall mechanism prior average particle size. With a pass rate ranging from 2% to 98%, For dependent variable, The slope obtained by performing least squares linear regression on the independent variable is determined as the prior homogeneity index of the overall mechanism. At this point, the mechanism prior generation module has completed its function, and will... and Output to the next module.

[0123] The physical constraint neural network module is connected to the mechanism prior generation module, which uses the crushing zone radius, cracking zone radius, rock breaking volume ratio, average borehole depth, and tunnel excavation cross-sectional area of ​​each zone as input (i.e., 11-dimensional feature vectors). The neural network is trained using the overall mechanistic prior average block size and overall mechanistic prior uniformity index from the mechanistic prior generation module as physical constraints. The core of this module lies in constructing a composite loss function containing two terms: the first term is the predicted overall average block size output by the network. and uniformity index prediction value The first term is the mean squared error between the measured value and the actual value; the second term is the mean squared error between the network output and the prior mechanism value. The weighted sum of the two terms constitutes the total loss. This module uses the Adam optimizer to update network parameters during training, and the output layer uses the Softplus activation function to ensure... and The value is always positive. After training, this module can receive any new 11-dimensional input features and output the corresponding values ​​in real time. and .

[0124] The multi-index output module is connected to the physical constraint neural network module and is used to predict the overall average block size based on the output of the physical constraint neural network module. and uniformity index prediction value The average block size is calculated and output based on the Rosin-Rammler distribution relationship. Cumulative gradation curve of under-sieve Characteristic particle size and coefficient of uniformity The specific calculation method is as follows: first, by Calculate the characteristic particle size, and then according to... , and Output various indicators, among which It can be used directly to guide the setting of the crusher discharge port. Used to evaluate the uniformity of block size distribution, a complete gradation curve. It can accurately match the optimal operating parameters for subsequent crushing and screening processes.

[0125] In addition, the system includes an adaptive incremental learning module that monitors changes in surrounding rock conditions in real time. When the change in the rock mass integrity coefficient exceeds a preset threshold, or when the comprehensive block size evaluation value for multiple consecutive cycles is lower than a preset level, this module automatically triggers incremental training of the neural network in the physical constraint neural network module—using the current network parameters as initial values ​​and continuing training with newly acquired field data, rather than retraining from scratch. During incremental training, this module gradually reduces the physical constraint weights in the composite loss function according to preset rules, based on the amount of new data accumulated since the current changes in surrounding rock conditions. The value of allows the model to gradually increase its reliance on the inherent patterns of the data as it gains sufficient new data support, while retaining prior mechanistic knowledge as a safety constraint. This adaptive mechanism ensures that the system can maintain high forecasting performance over the long term.

[0126] The system described in this embodiment corresponds to the method described in Embodiment 1. Each module of the system executes the corresponding steps in the method: the partition feature acquisition module corresponds to step 1, the mechanism prior generation module corresponds to steps 2 and 3, the physical constraint neural network module corresponds to step 4, and the multi-index output module corresponds to step 5. Through modular design, this system organically integrates mechanism prior generation and physical constraint neural network training, achieving a high-precision, high-reliability, and multi-index parallel intelligent prediction function for tunnel blasting block size.

[0127] Many specific details have been set forth in the foregoing description to provide a thorough understanding of the present invention. However, the above description is merely a preferred embodiment of the present invention, and the present invention can be implemented in many other ways different from those described herein. Therefore, the present invention is not limited to the specific embodiments disclosed above. Furthermore, any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention, or modify them into equivalent embodiments, using the methods and techniques disclosed above, without departing from the scope of the present invention. Any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention, without departing from the content of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A multi-index prediction method for tunnel blasting block size based on zonal crushing and crack radius, characterized in that, Includes the following steps: Step 1, Obtaining Zoning Features: Divide the tunnel's full-section blast holes into a cutting zone, an auxiliary zone, and a surrounding zone according to their function and spatial location, denoted as follows: Obtain the crushing zone radius of each partition. radius of cracked zone and the percentage of rock-broken volume The radius of the crushing zone The radius of the columnar region where the rock around the borehole undergoes pulverization or plastic flow under the action of the blast shock wave; the radius of the cracked zone. It refers to the radius of the region outside the crushing zone where a radial crack network forms under circumferential tensile stress; Step 2, Single-zone mechanism prediction: For each zone, based on rock dynamics and cylindrical charge theory, using the radius of the crushing zone and the radius of the cracking zone as core inputs, a semi-empirical prediction formula for single-zone block size is established, and the average block size mechanism prediction value for each zone is calculated. And the predicted value of the uniformity index mechanism The semi-empirical prediction formula for this single-block degree is: ; ; in The average borehole depth, This refers to the cross-sectional area of ​​the tunnel excavation. , , , , , , , , These are preset constants related to rock properties and explosive type; Step 3, Partition Weighted Synthesis: Combine the data from each partition... and stated Transform them into the Rosin-Rammler distribution function corresponding to each partition. And according to the proportion of rock breaking volume in each zone By performing weighted superposition, the mechanistic synthesis value of the overall gradation curve is obtained. Numerical processing was performed on the synthesis curve of this mechanism to extract the prior average block size of the overall mechanism. and the overall mechanism prior homogeneity index ; Step 4, Training the Physically Constrained Neural Network: Construct a fully connected neural network. Its input layer consists of an 11-dimensional feature vector containing the radius of the crushed zone, the radius of the cracked zone, the proportion of rock-broken volume, the average borehole depth, and the tunnel excavation cross-sectional area for each zone. The output layer is the overall average block size prediction value. and uniformity index prediction value A composite loss function consisting of a weighted sum of a data fitting loss term and a physical soft constraint loss term is constructed for network training. The physical soft constraint loss term is used to force the network output not to deviate from the overall mechanism prior average block size and the overall mechanism prior uniformity index. Step 5, Multi-index Output: Input the corresponding 11-dimensional feature vector of the explosion cycle to be predicted into the trained neural network to obtain... and The average block size is calculated and output based on the Rosin-Rammler distribution relationship. Cumulative gradation curve of under-sieve Characteristic particle size and coefficient of uniformity .

2. The method according to claim 1, characterized in that, In step 3, the Rosin-Rammler distribution function corresponding to each partition Determined by the following formula: ; in The overall mechanism prior average block size By synthesizing the curve in the aforementioned mechanism Upper linear interpolation is used to obtain the particle size corresponding to a throughput of 50%; the overall mechanism prior uniformity index By achieving a pass rate between 2% and 98%, For dependent variable, The slope obtained by performing least squares linear regression on the independent variable is determined as follows: .

3. The method according to claim 1, characterized in that, In step 1, the radius of the crushing zone and the radius of the cracked area The following methods were used to obtain the data: establish a columnar charge fluid-structure interaction numerical model for typical boreholes in each zone, output the distribution of damage zones around the boreholes, define the crushing zone with the equivalent plastic strain threshold, define the cracking zone with the area where the maximum principal tensile stress is greater than the dynamic tensile strength of the rock, and calibrate the model by combining on-site acoustic wave testing or borehole television inversion.

4. The method according to claim 1, characterized in that, In step 4, the composite loss function Represented as: ; in The total number of training samples, The sample number. and The first The measured average block size and measured evenness index of each sample. and The first The overall mechanism prior average block size and the overall mechanism prior homogeneity index corresponding to each sample. The weights are physical constraints; the output layer of the neural network uses the Softplus activation function to ensure that the predicted value is always positive, the hidden layer has no less than 3 layers and the number of neurons in each hidden layer decreases sequentially, and the Adam optimizer is used for training.

5. The method according to claim 1, characterized in that, In step 5, the characteristic particle size Calculated using the following formula: ; in The non-uniformity coefficient Calculated using the following formula: ; The cumulative gradation curve under sieve Determined by the following formula: 。 6. The method according to claim 1, characterized in that, In step 1, the radius of the crushing zone and the radius of the cracked area The parameters are obtained through a pre-trained blasting parameter-fracture zone radius proxy model. The proxy model takes the hole spacing, resistance line and charge decoupling coefficient of each zone as input and the corresponding crushing zone radius and cracking zone radius as output.

7. The method according to claim 1, characterized in that, In step 2, the range of values ​​for the preset constant is: , , , , , , , , .

8. A multi-index prediction system for tunnel blasting block size based on zonal crushing and crack radius, characterized in that, include: The partition feature acquisition module is used to acquire the radius of the crushed zone, the radius of the cracked zone, and the proportion of rock-broken volume of the blast hole in the full-section excavation zone, auxiliary zone, and surrounding zone of the tunnel. The mechanism prior generation module is connected to the partition feature acquisition module and is used to generate the overall mechanism prior average block size and the overall mechanism prior uniformity index based on the crushing zone radius, the cracking zone radius, the rock breaking volume ratio, and the pre-acquired average borehole depth and tunnel excavation cross-sectional area of ​​each partition. The physical constraint neural network module, connected to the mechanism prior generation module, is used to train the neural network with the radius of the crushing zone, the radius of the cracking zone, the proportion of rock breaking volume, the average blast hole depth and the tunnel excavation cross-sectional area of ​​each partition as inputs, and with the overall mechanism prior average block size and the overall mechanism prior uniformity index as physical constraints, and output the overall average block size prediction value and uniformity index prediction value. The multi-index output module, connected to the physical constraint neural network module, is used to calculate and output the average block size, undersize cumulative gradation curve, characteristic particle size, and non-uniformity coefficient based on the overall average block size prediction value and the uniformity index prediction value, according to the Rosin-Rammler distribution relationship.

9. The system according to claim 8, characterized in that, The mechanism prior generation module includes: a single-region calculation unit, used for each region according to the single-region block-degree semi-empirical prediction formula: ; and ; Calculate the average block size mechanism prediction value for each partition. And the predicted value of the uniformity index mechanism The weighted synthesis unit is used to combine the values ​​of each partition. and stated The distribution functions are converted into Rosin-Rammler distribution functions corresponding to each partition, and then weighted and superimposed according to the proportion of rock breaking volume in each partition to obtain the mechanism synthesis value of the overall gradation curve; the prior extraction unit is used to perform numerical processing on the mechanism synthesis curve to extract the overall mechanism prior average block size and the overall mechanism prior uniformity index.

10. The system according to claim 8, characterized in that, In the multi-index output module, the feature particle size Calculated using the following formula: ; in , This is the predicted value for the overall average block size. The predicted value of the uniformity index; the non-uniformity coefficient Calculated using the following formula: ; The cumulative gradation curve under sieve Determined by the following formula: ; The system also includes an adaptive incremental learning module, which automatically triggers incremental training of the neural network when the surrounding rock conditions change, and gradually reduces the weight of the physical constraints during the incremental training process.