Tunnel blasting slotting slot cavity range prediction method based on physical information neural network

Abstract

The application discloses a tunnel blasting slotting slot cavity range prediction method and system based on a physical information neural network, and belongs to the technical field of the cross of artificial intelligence and computational mechanics. The method acquires blasting parameters and blast hole arrangement parameters, and equivalently converts a columnar charge into a spherical charge based on volume equivalence; a deep neural network is constructed by taking observation point coordinates and time as input; a physical constraint is constructed and embedded into a loss function in the form of a residual term, a dynamic weight is applied to train the network; the trained neural network is used to predict time series stress and extract peak stress, a damage criterion is combined to determine the damage boundary, and the slot cavity range is output. The application replaces the traditional partial differential equation residual constraint with a physical constraint, improves the prediction accuracy and physical consistency, and reduces the training difficulty and dependence on labeled data.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and computational mechanics, specifically to a method for predicting the range of tunnel blasting slots based on a physical information neural network. Background Technology

[0002] With the continuous expansion of tunnel construction scale in my country, drill-and-blast method has become a widely used primary excavation method. Cut-and-blast blasting, as a key component of drill-and-blast tunnel construction, creates a cavity that provides an important second free surface and compensation space for subsequent blast holes, directly determining the overall blasting effect and blast hole utilization rate. However, current cut-and-blast design still relies heavily on field tests and engineers' experience, resulting in significant subjective influences and insufficient accuracy. Therefore, achieving intelligent and precise prediction of the cavity range in tunnel cut-and-blasting is of great significance.

[0003] Rock blasting is a complex dynamic process involving intense physicochemical reactions under high pressure and high temperature gases. Various empirical formulas for calculating the boundaries of the blasting range have been proposed both domestically and internationally. However, due to factors such as the heterogeneity of the rock mass structure and the stress state of the surrounding rock, theoretical predictions still have significant errors compared to actual conditions. In solving the dynamic equations, traditional numerical methods are computationally intensive and time-consuming when dealing with high-dimensional nonlinear blasting dynamics problems, and heavily rely on specialized computing software, suffering from limitations such as commercial software licensing and strong dependence on the computing environment. In recent years, Physical-Informed Neural Networks (PINN), as an innovative method that integrates physical constraints and data-driven approaches, has shown great potential in solving nonlinear partial differential equations. However, standard PINN is prone to imbalance between boundary loss and equation loss when solving strongly nonlinear blasting dynamics problems, leading to a focus on equation loss during training while neglecting boundary conditions. Furthermore, it requires additional labeled data to achieve high-precision solutions, and the optimization process is prone to getting trapped in local optima.

[0004] In existing technologies, relevant research has established a mechanical model for the cavity range of tunnel cut-out blasting. This model uses a planar cutting method to discretize the observation area, determines the failure range of each plane based on rock failure conditions, and ultimately derives the overall cavity range of the cut-out area. However, solving this mechanical model relies entirely on specialized mathematical software, resulting in high computational resource consumption, long processing time, and reliance on commercial software licenses. Furthermore, some studies have proposed a cavity range prediction method based on continuous learning neural networks. This method gradually approximates the original mechanical model through a progressive learning strategy. However, this method is purely data-driven and does not utilize prior knowledge of the blasting physics process to constrain the neural network's learning process, leading to low sample utilization and limited generalization ability and prediction accuracy. In addition, existing PINN-based blasting prediction methods typically employ traditional partial differential equation residual constraints, which have difficulty converging when dealing with highly nonlinear problems such as blasting stress wave propagation and lack continuous learning capabilities, requiring retraining from scratch when faced with new parameter ranges. Although there have been technical explorations in recent years that combine PINN with continuous learning, the relevant research has mainly focused on computer vision and general physical system modeling, and has not yet involved specific engineering scenarios for predicting the range of tunnel blasting slots.

[0005] In summary, existing techniques for predicting the cavity range in blasting operations are inadequate in terms of prediction accuracy, computational efficiency, physical consistency, and generalization ability. There is an urgent need for an intelligent prediction method for the cavity range that can integrate prior knowledge of blasting physics, reduce dependence on labeled data, possess continuous learning capabilities, and be computationally efficient. Summary of the Invention

[0006] This invention addresses the technical problems in existing technologies, such as insufficient accuracy in predicting the cavity range of tunnel blasting, low efficiency of traditional mechanical models, poor generalization ability of pure data-driven methods, and the tendency of physical information neural networks to become unbalanced between boundary loss and equation loss when solving blasting dynamics equations. It provides a method and system for predicting the cavity range of tunnel blasting based on physical information neural networks.

[0007] The specific details of the invention are as follows: This invention provides a method for predicting the range of tunnel blasting slots based on a physical information neural network, the specific contents of which are as follows: The blasting parameters and borehole layout parameters of the cut area are obtained. Based on the volume equivalence relationship between columnar and spherical explosive charges, each columnar explosive charge is discretized into several equivalent spherical explosive charges and the spatial coordinates of each equivalent spherical explosive charge are determined. A deep neural network is constructed with the spatial and temporal coordinates of the observation points as input and the predicted stress value of the observation points as output. The trained deep neural network is used to predict the temporal stress value of the observation points. The failure boundary is determined based on the comparison between the peak stress of each observation point and the rock failure criterion, thereby generating the range of the cut cavity. The training process of the deep neural network includes: constructing physical constraints, which include attenuation constraints that the stress generated by a single equivalent spherical explosive charge at the observation point satisfies the stress wave attenuation exponentially with the propagation distance, and superposition constraints that the stress generated by multiple equivalent spherical explosive charges at the observation point satisfies the linear superposition of stresses. The physical constraints are embedded in the loss function of the deep neural network in the form of residual terms, and dynamic weights are applied to each constraint term in the loss function to balance the contribution of each constraint term during the training process. After using a trained deep neural network to predict the temporal stress values ​​of the observation points, the peak stress of each observation point is extracted for the determination of the failure boundary.

[0008] Specifically, the attenuation constraint is as follows: the stress generated by a single equivalent spherical explosive charge at the observation point is constrained by the algebraic relationship between the initial impact pressure peak of the borehole wall, the stress wave waveform function, the spatial distance between the observation point and the equivalent spherical explosive charge, the attenuation index of the stress wave in the rock mass, the depth coordinates of the explosive charge center, and the borehole radius; the superposition constraint is as follows: the total stress at the observation point is equal to the algebraic sum of the stresses generated by all equivalent spherical explosive charges in all boreholes in the cut area at that observation point.

[0009] Specifically, the loss function includes attenuation constraint residuals, superposition constraint residuals, initial condition constraint residuals, and boundary condition constraint residuals. The dynamic weights are adaptively adjusted according to the magnitude and convergence trend of each constraint during the training process. The initial condition constraints include zero initial stress at each observation point inside the rock mass before blasting, and the stress at the center of the equivalent spherical explosive charge at the moment of detonation being equal to the peak value of the initial impact pressure of the borehole wall. The boundary condition constraints include normal stress release constraints at the free boundary of the working face, and soft penalty constraints that determine failure when the peak stress at the observation point exceeds the rock failure threshold.

[0010] Specifically, the rock failure criteria are as follows: for the free surface area of ​​the tunnel face, dynamic tensile strength is used as the failure threshold, and for the bottom area of ​​the borehole, dynamic compressive strength is used as the failure threshold; the failure boundary and the range of the slotted cavity are determined by the planar cutting method, the failure state is determined point by point on each observation surface parallel to the tunnel face, the failure range boundary of each observation surface is determined, and the failure range boundary of each observation surface is spatially synthesized along the depth direction to generate the three-dimensional spatial morphology of the slotted cavity.

[0011] Optionally, the deep neural network takes the blasting parameters and spatiotemporal coordinates as input and directly calculates the residuals of each physical constraint in a forward propagation manner, without relying on automatic differentiation to calculate the residuals of the partial differential equations.

[0012] This invention also provides a tunnel blasting cut cavity range prediction system based on a physical information neural network, including a data acquisition module, an equivalent charge module, a neural network module, a stress prediction module, and a cavity generation module. The data acquisition module acquires blasting parameters and borehole layout parameters for the cut area; the equivalent charge module discretizes each cylindrical charge into several equivalent spherical charges based on the volume equivalence relationship between cylindrical and spherical charges, determining the spatial coordinates of each equivalent spherical charge; the neural network module constructs and trains a deep neural network using the spatial and temporal coordinates of the observation points as input and the predicted stress value of the observation points as output. The loss function of the deep neural network embeds a physical constraint residual term, which includes stress wave exponential decay constraint and stress linear superposition constraint, with each constraint term balanced using dynamic weights; the stress prediction module uses the trained deep neural network to predict the temporal stress value of the observation points and extract the peak stress of each observation point; the cavity generation module determines the failure boundary based on the comparison result between the peak stress of each observation point and the rock failure criterion, generating the range of the cut cavity.

[0013] This invention offers the following advantages: Firstly, by embedding the stress wave exponential decay law and the principle of linear superposition into the neural network loss function as algebraic analytical constraints, it replaces the partial differential equation residual constraints that are difficult to converge in traditional physical information neural networks. This allows the network to accurately predict the blasting stress field distribution even under unlabeled or weakly labeled conditions, significantly reducing the dependence on large-scale labeled data. Secondly, by introducing a dynamic weight balancing strategy, the weights are adaptively adjusted according to the magnitude and convergence trend of each constraint term during training, avoiding the problem of boundary conditions being ignored due to the imbalance between boundary loss and equation loss in traditional physical information neural networks. Thirdly, by organically coupling the improved physical information neural network with the plane cutting method and the double-threshold rock failure criterion, a complete prediction chain is realized, from point stress prediction to surface failure determination to the generation of volumetric slots, enabling the generation of the three-dimensional spatial morphology of the slotted cavity. Overall, it improves the prediction accuracy, physical consistency, and computational efficiency of the tunnel blasting cavity range, providing reliable technical support for intelligent tunnel blasting design. Attached Figure Description

[0014] Figure 1 This is an overall technical roadmap of the embodiments of the present invention; Figure 2 This is a schematic diagram illustrating the equivalent division of a columnar medicine packet into a spherical medicine packet in an embodiment of the present invention; Figure 3 This is a schematic diagram of the arrangement of observation surfaces and observation points in an embodiment of the present invention; Figure 4This is a schematic diagram of the damage range of a single borehole at the borehole opening in an embodiment of the present invention, wherein (a) is the damage range when the borehole is perpendicular to the free surface, (b) is the damage range when the borehole is oblique to the free surface, and the dashed circle represents the damage boundary; Figure 5 This is a schematic diagram of the evolution of the damage range at the borehole opening of the double straight hole slotting blast hole under different hole spacing conditions in the embodiments of the present invention, wherein (a) to (f) correspond to six working conditions with hole spacing from large to small respectively; Figure 6 This is a schematic diagram of the damage situation of the double straight hole slot at different depth sections in an embodiment of the present invention, wherein (a) to (c) are the damage range of the free surface area from shallow to deep, and (d) to (f) are the damage range of the bottom area of ​​the borehole from shallow to deep; Figure 7 This is a data visualization example of the tunnel blasting slot cavity range prediction system in an embodiment of the present invention; Figure 8 This is a comparison chart of the model performance evaluation of the embodiments of the present invention and the prior art at different parameter expansion stages; Figure 9 These are diagrams showing the predicted concrete failure range obtained at different cross-sections using embodiments of the present invention. Figure 10 This is a visual schematic diagram of the three-dimensional spatial morphology of the slotted cavity in an embodiment of the present invention; Figure 11 This is a three-dimensional cloud diagram of the slotted cavity in an embodiment of the present invention. Detailed Implementation

[0015] The following embodiments illustrate the present invention in detail. In the description of these embodiments, specific details such as particular system structures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail. Example 1

[0016] refer to Figure 1 This embodiment provides a method for predicting the range of tunnel blasting slots based on physical information neural networks. This method achieves intelligent prediction of the three-dimensional spatial morphology of the slot through three core steps: equivalent transformation of columnar explosive charges, physical constraint embedding neural network training, and spatial synthesis using the planar cutting method.

[0017] (1) Equivalent conversion of columnar medicine pack refer to Figure 2In this embodiment, the Starfield superposition method is used to equate the cylindrical explosive charge to several spherical explosive charges. When dividing the cylindrical explosive charge into a certain number of spherical explosive charges of equal volume, it is necessary to ensure that the center of the equivalent spherical explosive charge coincides with the cylindrical explosive charge. When performing equivalent replacement, it is essential to ensure that the charge amount does not change due to the change in the shape of the explosive charge. Then, the stress wave parameters excited by the entire extended charge at a certain detonation wave propagation velocity are obtained by using empirical calculation formulas for the stress wave parameters of the equivalent spherical explosive charges.

[0018] The radius of the equivalent spherical propellant charge is determined based on the principle that the volumes of the cylindrical and spherical propellant charges are equal. Let the radius of the cylindrical propellant charge be... Then the radius of a single equivalent spherical drug pack The calculation formula is:

[0019] Therefore, the number of equivalent spherical medicine packets can be obtained. This is the ratio of the length of the cylindrical propellant charge to the diameter of the equivalent spherical propellant charge. After discretizing the cylindrical propellant charge into multiple equivalent spherical propellant charges along the axial direction, the propellant charges are distributed at equal intervals along the borehole axis.

[0020] For the slotted area The first gun hole, the... The number of boreholes is discrete as follows: If there are 1 equivalent spherical medicine pack, then the 1st... The first blast hole The spatial coordinates of an equivalent spherical medicine bag are as follows: The detonation sequence of each explosive charge is set according to the actual borehole detonation network.

[0021] (2) Rock failure conditions Based on the differences in failure mechanisms, the calculation of the cavity area is divided into a free surface region and a borehole bottom region, which are solved separately. Failure in the free surface region is mainly dominated by reflected tensile stress perpendicular to the free surface, while failure at the borehole bottom originates from compressive stress generated by the blast load. For any observation point... The total stress under the superimposed action of multiple boreholes and multiple spherical explosive charges The rock at a given point is considered damaged if the following conditions are met:

[0022] in, It represents the dynamic tensile strength of the rock, which is the critical value for failure in the free surface region; This represents the dynamic compressive strength of the rock, which is the critical failure value in the bottom region of the borehole. When the stress exceeds the corresponding critical value, the rock in this region will fail.

[0023] (3) Construction of physical constraints The core innovation of this embodiment lies in replacing the partial differential equation residual constraints in traditional PINN with physical constraints in algebraic analytical form. The constructed physical constraints include the following two categories: refer to Figure 1 The physical constraints on the left are connected to the branch, the first type being the single-charge stress wave exponential attenuation constraint. For the... The first one inside the gun hole An equivalent spherical drug cartridge, which can be observed at any point in space. The generated stress must strictly satisfy the exponential decay law, the expression of which is:

[0024] in, The initial impact pressure peak value of the borehole wall, in Pascals, is calculated as follows:

[0025] In the above formula, For the density of the explosive, For the detonation velocity of the explosive, For the diameter of the explosive charge, The diameter of the borehole.

[0026] The stress wave waveform function is expressed as an exponentially decaying pulse.

[0027] in The waveform attenuation coefficient, The waveform angular frequency, The time required for the stress wave to travel from the center of the explosive charge to the observation point, expressed in seconds, is calculated using the following formula:

[0028] in For the longitudinal wave velocity of the rock, The distance from the center of the medicine pack to the observation point is expressed in meters. The calculation formula is as follows:

[0029] The Z-axis depth coordinate of the center of the medicine pack. The radius of the borehole is... is the attenuation index of stress waves in rock mass.

[0030] The second type is the linear superposition constraint of multiple explosive charges. The total stress at any observation point in space is equal to the linear algebraic sum of the contributions from all equivalent explosive charges, and its expression is:

[0031] in, This represents the total number of blast holes in the cut-out area. For the first Each cylindrical medicine packet is equivalent to the number of spherical medicine packets.

[0032] (4) Initial conditions and boundary conditions In this embodiment, the initial condition constraints include: the initial stress at each observation point inside the rock mass before blasting is zero, that is:

[0033] And the stress at the center of the equivalent spherical explosive charge at the moment of detonation is equal to the peak value of the initial impact pressure on the borehole wall:

[0034] Boundary condition constraints include: the free boundary of the tunnel face, i.e. The system employs a normal stress release constraint at the observation point, ensuring that only tensile reflected waves contribute to the free surface; and a soft penalty constraint that determines failure when the peak stress at the observation point exceeds the rock failure threshold. In the loss function, the ReLU function is used to apply a soft penalty to the portion exceeding the threshold.

[0035] (5) Improve the design of the loss function of PINN refer to Figure 1 In the core section of the middle segment, the deep neural network constructed in this embodiment uses blasting parameters and the spatial coordinates of the observation points. and time coordinates As input, the predicted stress value at the observation point. This is the output. The network's loss function consists of the residual terms of each physical constraint, and its total loss function expression is:

[0036] in, This is the attenuation constraint residual term, used to constrain the exponential attenuation behavior of the stress wave of a single explosive charge with distance; The superposition constraint residual term is used to constrain the linear superposition relationship of stress in multiple drug packs; This is the initial condition constraint residual term, used to constrain the initial stress state before blasting; This is the boundary condition constraint residual term, used to constrain stress release at the boundary of the free surface.

[0037] refer to Figure 1 In the dynamic weighted feedback branch, each weight coefficient The weights of each constraint are adaptively adjusted based on their magnitude and convergence trend during training. Specifically, when the magnitude of a constraint is significantly greater than that of other constraints, its corresponding weight is reduced accordingly; when the convergence speed of a constraint is significantly slower than that of other constraints, its corresponding weight is increased accordingly. This ensures that all physical constraints converge synchronously during training, avoiding the imbalance between boundary loss and equation loss in traditional PINN.

[0038] (6) Neural network training and stress prediction After the network is trained, the improved PINN is used to predict the temporal stress values ​​at each observation point. For any observation point in space... The network outputs the stress value of that point over the entire time series. The peak stress at each observation point is extracted from the time-series stress curve. Its expression is:

[0039] (7) Determination of failure boundary and generation of slot cavity refer to Figure 3 In this embodiment, a planar cutting method is used to determine and synthesize the cavity range. On each observation surface parallel to the tunnel face... An observation point grid is arranged on top, and the observation range is appropriately expanded based on the borehole coordinates. The interval between observation points is set according to the accuracy requirements.

[0040] For each observation point on each observation surface, the extracted peak stress Compare with the rock failure criterion. The rock failure criterion is: for the free surface region of the tunnel face (close to...) (Region) adopts dynamic tensile strength As a critical value for destruction, for the bottom region of the borehole (far from the center) The area uses dynamic compressive strength. This serves as the critical value for failure. When the peak stress at a certain observation point exceeds the corresponding critical value, that point is determined to be a failure point.

[0041] refer to Figure 10 and Figure 11After point-by-point damage assessment, the damage extent boundaries of each observation surface are determined. The damage extent boundaries of all observation surfaces are then spatially synthesized along the depth direction to ultimately generate the three-dimensional spatial morphology of the cut cavity. The cavity boundary predicted by this method is composed of layers of damage extent boundaries from each observation surface stacked along the depth direction, accurately representing the continuous change of the cavity's spatial morphology with depth. The stereo cloud map further uses color mapping to visually reflect the depth distribution of the cavity at different spatial locations. The horizontal and vertical axes correspond to two mutually perpendicular directions within the working face plane, and the color bars indicate the spatial position of the cavity boundary in the depth direction. This allows direct acquisition of key geometric parameters such as the maximum depth of the cavity and the damage extent dimensions at any depth section. This method elevates the traditional cavity extent determination process, which relies on analyzing discrete observation surfaces one by one, to a continuous prediction and visualization of the complete spatial volume. It comprehensively and intuitively reflects the three-dimensional morphology of the cavity and its evolution trend along the depth direction, providing more accurate data support for the evaluation of cut blasting effects and parameter optimization.

[0042] refer to Figure 7 This embodiment also includes a data visualization module, which integrates multiple functions such as monitoring training convergence dynamics with loss function curves, quantifying prediction accuracy with evaluation index curves, reflecting the parameter range expansion process with parameter space expansion diagrams, comparing the degree of agreement between predicted and actual values ​​with stress distribution comparison diagrams, and outputting the final cavity range prediction results with grooving range prediction diagrams.

[0043] refer to Figure 8 In comparative experiments across multiple parameter expansion stages, the relative error of the method in this embodiment remained consistently at a low level, significantly outperforming the CLEMM method and the standard PINN method.

[0044] Example 2 refer to Figure 4 This embodiment uses a single borehole as an example to demonstrate the specific application of the method of the present invention, which is used to determine the damage range at the borehole opening of a single borehole.

[0045] The cylindrical explosive charge within the borehole is discretized into several equivalent spherical charges. Physical constraints are constructed according to the method described in Example 1, and the improved PINN is trained. Based on the spatial relationship between the borehole and the free surface, two cases can be distinguished: perpendicular and oblique. (Reference) Figure 4 In (a), when the borehole is perpendicular to the free surface, the damage area on the free surface is distributed in a circular pattern centered on the borehole opening; Reference Figure 4 In (b), when the borehole is oblique to the free surface, the center of the damage area will shift relative to the borehole opening due to the deflection of the stress wave propagation path. The method in this embodiment, through precise modeling of stress wave attenuation constraints and linear superposition constraints, can accurately predict the damage range and its boundary at the borehole opening in both cases.

[0046] Specifically, for the case of vertical boreholes, on the free surface A grid of observation points is arranged on top, and the peak stress at each observation point exhibits a distribution that gradually decreases radially from the orifice opening, with the peak stress exceeding [a certain value]. The area is the damage range. For oblique boreholes, since the central axis of the explosive charge is not perpendicular to the free surface, the distance distribution from the center of each equivalent spherical explosive charge to each observation point on the free surface is no longer symmetrical, causing the center of the peak stress distribution to shift relative to the borehole opening. This embodiment can accurately capture this shift by calculating the peak stress at each observation point and determining the damage state.

[0047] Example 3 refer to Figure 5 This embodiment uses two straight-hole slotting blast holes as an example to demonstrate the actual effect of the method of the present invention in predicting the evolution of the damage range at the hole opening under different hole spacing conditions.

[0048] refer to Figure 5 (a) to (f) represent six working conditions with hole spacing decreasing from large to small. When the hole spacing is large, the distance between the two boreholes is large, and the stress superposition in their common area does not reach the rock failure threshold, thus forming two independent circular failure zones. As the hole spacing gradually decreases, the superimposed stress in the middle area increases and exceeds the critical value, and the two failure zones gradually connect. When the hole spacing decreases further, the maximum superimposed stress at the middle observation point increases significantly, and the failure areas merge into a unified whole. When the hole spacing decreases further to near the borehole diameter, the superposition effect is enhanced again, and the failure morphology approaches the approximately circular characteristic under the action of a single columnar explosive charge.

[0049] The specific operation steps of this embodiment are as follows: setting different borehole spacings. The blasting parameters were obtained, the equivalent explosive charge was completed, and physical constraints were constructed according to the method described in Example 1; the improved PINN was independently trained for each working condition; the peak stress at each observation point was predicted using the trained network; and the plane cutting method was used on the free surface. The failure state is determined point by point, and the failure range boundaries are plotted under different working conditions. By comparing the prediction results under different hole spacing conditions, the critical hole spacing range for dual-hole continuous failure under this rock mass condition can be determined. (Reference) Figure 5 The hole spacing corresponding to (c) is the critical through hole spacing under this condition.

[0050] Example 4 refer to Figure 6 This embodiment takes a straight hole slotted double borehole as an example to demonstrate the evolution law of the slotted cavity range in the depth direction predicted by the method of the present invention.

[0051] To determine the complete failure range of the double-straight-hole cut, the failure characteristics of the free surface region and the bottom region of the borehole need to be examined separately. (Reference) Figure 6From (a) to (c), the damage range at the free surface is calculated using the free surface rock stress calculation method: due to the small distance between the two holes, the stress waves superimpose to form an approximately circular damage zone in the shallow part. The damage range gradually narrows as the depth from the free surface increases. (Reference) Figure 6 From (d) to (f), the bottom of the borehole is located in a deep region far from the free surface: at the deepest point, the destruction zones of the two boreholes are independent of each other; as they transition upwards, the independent zones gradually connect, forming a dumbbell shape, and the connecting part in the middle continues to widen.

[0052] The specific operation steps of this embodiment are as follows: Set the borehole spacing to be relatively small (the spacing between two damaged zones that can be connected), arrange the observation point grid according to the method described in Embodiment 1, and set the observation surface along the depth direction. Layer by layer, the spacing between adjacent observation surfaces is set according to the required spatial resolution. An observation point grid is arranged on each observation surface, and the peak stress at each observation point on each observation surface is predicted using a trained improved PINN. For observation surfaces in the free surface region ( (Smaller), using dynamic tensile strength As a critical value for failure; for the observation surface in the bottom region of the borehole ( (Larger), using dynamic compressive strength As the critical value for failure. After completing the failure assessment layer by layer and determining the failure range boundary of each observation surface, the failure boundaries of each observation surface are spatially synthesized along the depth direction. By combining the failure morphologies of the upper part (free surface side) and the lower part (bottom of the hole side), the complete three-dimensional spatial morphology of the cavity is finally defined through the intermediate transition region. Reference Figure 6 The transition relationship between the three diagrams of the free surface region and the three diagrams of the hole bottom region can completely reconstruct the spatial geometry of the cavity formed by the double straight hole slotting.

[0053] Example 5 refer to Figure 9 This embodiment demonstrates the verification application of the method of the present invention in concrete model tests, used to verify the feasibility of the intelligent prediction system for the range of the blasting cavity.

[0054] Following the plan for model testing at the test site, concrete test blocks were used to simulate rock mass, and cut-out blast holes were set up and blasting parameters were determined. As described in Example 1, the center coordinates of each equivalent spherical charge were first calculated based on the geometric parameters of the columnar charge, and multiple observation sections parallel to the free surface were set. Through pre-testing, the observation range of each section was appropriately expanded based on the borehole coordinates, and a certain interval was set between observation points to ensure that the observation area could completely cover the blasting influence range.

[0055] The trained, improved PINN network is used to calculate the superimposed stress at each observation point within each observation surface. For each observation point, the network outputs the time-series stress curve, extracts the peak stress, and compares it with the critical value of the dynamic tensile strength of concrete to determine whether failure has occurred at that point. After completing the determination point by point, the failure range boundary of each observation surface is plotted.

[0056] refer to Figure 9 The prediction results at different cross-sections demonstrate that the method in this embodiment can clearly present the morphology of concrete failure range at different depths. Comparison of the prediction results with experimental measurements verifies the effectiveness of the method in terms of prediction accuracy and cavity spatial morphology reconstruction. By adjusting the input parameters, this method can quickly predict the cavity range under different combinations of blasting parameters, providing data support for optimizing on-site blasting schemes.

[0057] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for predicting the range of tunnel blasting cut cavity based on a physical information neural network, characterized in that, Includes the following steps: (1) Obtain blasting parameters and borehole layout parameters, and discretize each columnar explosive charge into an equivalent spherical explosive charge based on volume equivalence and determine its spatial coordinates; (2) Construct a deep neural network with the coordinates of the observation point and time as input and stress as output; (3) Construct a physical constraint consisting of an exponential decay constraint of stress wave of a single drug pack and a linear superposition constraint of stress of multiple drug packs, embed the physical constraint into the loss function as a residual term, and apply dynamic weights to balance the training contribution of each constraint term, and train the deep neural network. (4) The trained deep neural network is used to predict the time-series stress and extract the peak stress. The failure boundary is determined based on the comparison between the peak stress and the rock failure criterion, and the range of the slot cavity is generated.

2. The method according to claim 1, characterized in that, The attenuation constraint is specifically defined as follows: the stress generated by a single equivalent spherical explosive charge at the observation point is constrained by the algebraic relationship between the initial impact pressure peak of the borehole wall, the stress wave waveform function, the spatial distance between the observation point and the equivalent spherical explosive charge, the attenuation index of the stress wave in the rock mass, the depth coordinates of the explosive charge center, and the borehole radius.

3. The method according to claim 1, characterized in that, The superposition constraint is specifically defined as follows: the total stress at the observation point is equal to the algebraic sum of the stresses generated at that observation point by all equivalent spherical explosive charges in all boreholes of the cut area.

4. The method according to claim 1, characterized in that, The loss function includes decay constraint residuals, superposition constraint residuals, initial condition constraint residuals, and boundary condition constraint residuals. The dynamic weights are adaptively adjusted according to the magnitude and convergence trend of each constraint during the training process.

5. The method according to claim 4, characterized in that, The initial condition constraints include: the initial stress at each observation point inside the rock mass before blasting is zero, and the stress at the center of the equivalent spherical explosive charge at the moment of detonation is equal to the peak value of the initial impact pressure of the borehole wall.

6. The method according to claim 4, characterized in that, The boundary condition constraints include: normal stress release constraints at the free boundary of the tunnel face, and soft penalty constraints that determine failure when the peak stress at the observation point exceeds the rock failure threshold.

7. The method according to claim 1, characterized in that, The rock failure criterion is as follows: for the free surface area of ​​the tunnel face, the dynamic tensile strength is used as the critical failure value, and for the bottom area of ​​the borehole, the dynamic compressive strength is used as the critical failure value.

8. The method according to claim 1, characterized in that, The specific steps for determining the damage boundary and generating the cut-out cavity are as follows: using the planar cutting method, the damage state is determined point by point on each observation surface parallel to the working face, the damage range boundary of each observation surface is determined, and the damage range boundary of each observation surface is spatially synthesized along the depth direction to generate the three-dimensional spatial morphology of the cut-out cavity.

9. The method according to claim 1, characterized in that, The deep neural network takes the blasting parameters and spatiotemporal coordinates as input and directly calculates the residuals of each physical constraint using a forward propagation method.

10. A tunnel blasting cut cavity range prediction system based on a physical information neural network, comprising: The data acquisition module is used to acquire blasting parameters and borehole layout parameters in the cut area; The equivalent module for drug packs is used to discretize each cylindrical drug pack into several equivalent spherical drug packs based on the volume equivalence relationship between cylindrical and spherical drug packs, and to determine the spatial coordinates of each equivalent spherical drug pack. Its characteristic is that it further includes: The neural network module is used to construct and train a deep neural network by taking the spatial and temporal coordinates of the observation point as input and the predicted stress value of the observation point as output. The loss function of the deep neural network embeds a physical constraint residual term, which includes stress wave exponential decay constraint and stress linear superposition constraint. Each constraint term is balanced by dynamic weights. The stress prediction module is used to predict the time-series stress values ​​of observation points and extract the peak stress of each observation point using a trained deep neural network. The cavity generation module is used to determine the failure boundary based on the comparison results of the peak stress at each observation point and the rock failure criterion, and to generate the range of the slotted cavity.

Images