A tunnel construction section profile intelligent closed-loop control system and method

Abstract

This invention relates to the field of tunnel and underground engineering construction technology, and discloses an intelligent closed-loop control system and method for tunnel construction cross-section contour. The method includes: acquiring geological-geometric data before blasting by fusing a 3D laser scanner and ground-penetrating radar; dynamically generating theoretical blasting contour lines and adaptive blasting parameters through a convolutional neural network; achieving millimeter-level precise control of the drill arm by combining IMU positioning and a PID controller; generating a 3D deviation field after blasting using an iterative point cloud registration algorithm; triggering incremental learning to optimize model parameters when the deviation exceeds a threshold; and finally forming a closed-loop iteration. Through multi-source data fusion, intelligent contour and parameter generation, precise execution control, 3D deviation field feedback, and incremental learning mechanism, a complete intelligent closed-loop control system is constructed, breaking through the limitations of traditional experience-based blasting design and significantly improving the accuracy, efficiency, and economic benefits of tunnel construction.

Description

Technical Field

[0001] This invention belongs to the field of tunnel and underground engineering construction technology, and relates to three-dimensional measurement and evaluation of tunnel excavation cross-section, cross-section quality feedback and iterative optimization, construction equipment posture control and data-driven parameter adjustment. Specifically, it is an intelligent closed-loop control system and method for tunnel construction cross-section contour. Background Technology

[0002] During tunnel excavation, the quality of the cross-sectional profile directly affects the degree of surrounding rock disturbance, the amount of initial support work, the thickness of secondary lining, and construction safety. In actual construction, alternating over-excavation and under-excavation of the cross-section can lead to problems such as excessive support consumption, uneven lining, insufficient clearance, and local stress concentration, and may result in rework of subsequent procedures. Therefore, cross-sectional profile control has always been one of the core aspects of tunnel construction.

[0003] Existing tunnel drilling and blasting methods typically employ an experience-based design process based on factors such as surrounding rock grade, joint and fracture development, and groundwater conditions. However, the spatial distribution of surrounding rock structural planes exhibits significant heterogeneity and randomness. A single surrounding rock grade is insufficient to fully characterize the impact of local fracture density, weak interlayers, and differences in surrounding rock integrity on blasting formation. Furthermore, construction uncertainties further amplify profile deviations, resulting in substantial discrepancies between design parameters and actual outcomes. This is particularly problematic in scenarios involving interbedded soft and hard layers, fractured zones, and high-stress areas, where conventional experience-based adjustments struggle to achieve stable control.

[0004] In terms of quality inspection, excavation cross-sections are typically inspected on-site using methods such as cross-section measurement and total station measurement. However, existing inspections are mostly used for post-construction acceptance or statistical analysis, lacking a unified data organization, error measurement, and traceability link between inspection results and subsequent design revisions. Furthermore, cross-section evaluation systems suffer from inconsistent interpretations in engineering practice. For example, some focus only on over-excavation while ignoring the impact of under-excavation on clearance and lining, or only on average deviation while neglecting the constraint of local peak deviations on support safety margins. This makes it difficult to form a stable and reusable closed-loop management system for quality control. From an information and control perspective, tunnel drilling and blasting construction involves a coupled process of multiple stages, including geological information acquisition, blasting design, drilling execution, blasting implementation, and cross-section evaluation. Currently, the industry generally suffers from data fragmentation, model deficiencies, and feedback lags, lacking a unified representation method for the construction cycle. This makes it difficult to continuously quantify cross-section deviations and use them to drive synchronous revisions in subsequent design and execution.

[0005] In existing technologies, Chinese patent CN118864415A uses deep learning models such as variational autoencoders and CNNs to predict blasting quality levels based on borehole logs and design schemes. However, it is essentially an open-loop prediction tool, only outputting quality warnings but unable to generate specific blasting design schemes that can directly guide construction. Prediction and execution are disconnected, and it lacks a closed-loop feedback mechanism based on actual blasting effects. CN120783488A constructs influence coefficients by analyzing the mechanical effects of blasting shock waves and rock impacts to achieve over- and under-excavation risk warnings. However, the system's functions are strictly limited to monitoring and warning, lacking the ability to automatically convert analysis results into executable blasting parameter optimization instructions. The decision-making process still requires manual intervention and lacks the ability to perform end-to-end self-optimization using massive amounts of engineering data through machine learning, resulting in limited adaptability and accuracy.

[0006] In summary, existing tunnel construction cross-section profile control technologies still suffer from common defects such as disconnect between measurement and control, lack of intelligent design, inaccurate execution, and open-loop systems. Therefore, how to achieve quantifiable evaluation and traceable feedback of cross-section profile deviations in the tunnel construction cycle, and support dynamic control and quality stability assurance for the construction process, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] (a) Purpose of the invention

[0008] To address the aforementioned deficiencies and shortcomings of existing tunnel drilling and blasting methods, this invention provides an intelligent closed-loop control system and method for tunnel construction cross-section contours. The system aims to construct a closed-loop process encompassing perception, intelligent design, precise execution, feedback, and evolution. By integrating excavation face and surrounding rock feature information, it drives intelligent generation and multi-objective optimization of contour control and parameters. Combined with closed-loop control of borehole posture, it improves the consistency of design implementation. Furthermore, it utilizes three-dimensional detection of the post-blast cross-section to form quantifiable deviation representations and trigger iterative model updates. This significantly improves the forming accuracy of the tunnel excavation contour, reduces support costs and safety risks, and achieves a paradigm shift in tunnel construction from experience-driven to data-driven approaches.

[0009] (II) Technical Solution

[0010] To achieve the objective of this invention and solve its technical problems, the present invention adopts the following technical solution:

[0011] The first objective of this invention is to provide an intelligent closed-loop control method for the cross-sectional profile of a tunnel construction project, comprising at least the following steps:

[0012] SS1. Multi-source information acquisition and fusion: Before blasting, point cloud data of the tunnel excavation face is acquired by ground three-dimensional laser scanner, and the rock mass fracture distribution and uniaxial compressive strength data detected by ground radar are simultaneously fused to form multi-source data characterizing the geometric morphology of the excavation face and the geological properties of the surrounding rock.

[0013] SS2. Intelligent generation of contours and parameters: The multi-source data is input into a pre-trained convolutional neural network model. Spatial-geological fusion features are extracted through a residual convolutional network to generate the theoretical blasting contour line of the excavation face and the corresponding blasting parameters. The blasting parameters include at least the borehole spacing, charge density gradient and detonation sequence. The borehole spacing is adaptively adjusted according to the fracture density, and the charge density gradient is set in segments according to the rock mass strength.

[0014] SS3. Precise execution of closed-loop control: Analyze the theoretical blasting contour line of the excavation face and obtain the borehole layout control information. Map the drill arm pose data fed back by the IMU (Inertial Measurement Unit) positioning device to the spatial coordinate system of the theoretical blasting contour line through a homogeneous transformation matrix. Construct control quantity based on the deviation between pose data and borehole layout control information, and correct the drill arm angle in real time through a PID controller.

[0015] SS4. Post-blast cross-section data acquisition: Point cloud data of the tunnel cross-section after blasting is acquired using a ground-based 3D laser scanner;

[0016] SS5. Quantitative Construction of Cross-Sectional Deviation Field: The Euclidean distance field between the tunnel cross-section point cloud data after blasting and the theoretical blasting contour line of the excavation face is calculated using an iterative point cloud registration algorithm. Three-dimensional deviation field data for marking the spatial coordinates and volume deviation of over-excavation and under-excavation is generated through voxelized mesh segmentation and integral calculation.

[0017] SS6. Threshold-triggered incremental learning update: When the volume deviation exceeds a preset threshold, an incremental learning mechanism is triggered to construct a loss function using three-dimensional deviation field data and update the weight parameters of the convolutional neural network model in reverse using the gradient descent algorithm;

[0018] SS7. Closed-loop iterative output: Based on the updated convolutional neural network model, the theoretical blasting profile and parameters of the excavation face for the next construction cycle are generated, forming a closed-loop control iteration of the cross-sectional profile.

[0019] The second objective of this invention is to provide an intelligent closed-loop control system for tunnel construction cross-section contours, used to implement the aforementioned intelligent closed-loop control method for tunnel construction cross-section contours, comprising the following modules:

[0020] The three-dimensional perception module is used to collect point cloud data of the tunnel excavation face using a ground-based three-dimensional laser scanner before blasting, and simultaneously integrate rock mass fracture distribution and uniaxial compressive strength data detected by ground-penetrating radar to form multi-source data; after blasting, the ground-based three-dimensional laser scanner collects point cloud data of the tunnel cross section in real time.

[0021] The blasting profile generation module is used to input the multi-source data into the convolutional neural network model and dynamically generate the theoretical blasting profile of the excavation face and the corresponding blasting parameters, including the borehole depth, charge density gradient and detonation sequence.

[0022] The precision execution control module is used to analyze the theoretical blasting contour line, combine it with the drill arm posture data fed back by the IMU positioning device, and correct the drill arm angle in real time through the PID controller.

[0023] The three-dimensional deviation field feedback module is used to calculate the Euclidean distance field between the point cloud data of the tunnel cross section after blasting and the theoretical blasting contour line using an iterative point cloud registration algorithm, and generate three-dimensional deviation field data with marked over- and under-excavation spatial coordinates and volume deviation.

[0024] The incremental learning module is used to trigger the incremental learning mechanism when the volume deviation exceeds a preset threshold. It uses the three-dimensional deviation field data as the loss function and updates the weight parameters of the convolutional neural network model in reverse through the gradient descent algorithm.

[0025] The closed-loop iteration module is used to generate the theoretical blasting profile and supporting blasting parameters for the next construction cycle based on the optimized convolutional neural network model, thus forming a closed-loop control iteration.

[0026] (III) Technical Effects

[0027] Compared with the prior art, the present invention has the following beneficial and significant technical effects:

[0028] (1) This invention integrates high-precision point cloud data collected by a 3D laser scanner before blasting with multi-source information such as rock fracture distribution and uniaxial compressive strength detected by ground-penetrating radar to construct a geological-geometric feature tensor. This tensor is then input into a residual convolutional network to dynamically generate the theoretical blasting contour line and associated parameters. The borehole spacing is adaptively adjusted according to the fracture density, the charge gradient is set in segments according to lithological differences, and the detonation sequence is generated based on the contour curvature to generate a millisecond-level delay sequence. This avoids the limitations of traditional experience-based reliance and achieves proactive pre-control of geological adaptability, ensuring blasting accuracy from the source.

[0029] (2) A spatial coordinate system is established based on the theoretical blasting contour line, and the drill arm posture data obtained by the IMU positioning device is mapped to the design reference through a homogeneous transformation matrix. The PID controller analyzes the posture deviation in real time and dynamically generates drill arm correction commands. When the drilling trajectory deviation exceeds the preset threshold, the hydraulic servo system automatically adjusts the drill arm pitch angle to ensure that the deviation between the actual drilling trajectory and the theoretical contour line is ≤ ±5mm, realizing real-time closed-loop correction during construction and significantly improving the contour forming accuracy.

[0030] (3) After blasting, the Iterative Closest Point (ICP) algorithm is used to register the actual and theoretical point clouds, generating a Euclidean distance field with sub-millimeter accuracy, and marking the spatial coordinates and volume deviation of the over- and under-excavation areas. When the deviation exceeds the limit, the three-dimensional deviation field data is used as the loss function, and the weight parameters of the convolutional neural network are updated in reverse through the gradient descent algorithm. The feedback mechanism breaks through the limitations of traditional macro scoring, provides fine signal input at the spatial coordinate level, and drives the continuous evolution of the model.

[0031] (4) The convolutional neural network model is deployed on edge computing nodes, forming a local processing unit with the scanner and IMU to achieve real-time response to the generation of blasting parameters. Incremental learning weights are synchronized to the cloud server via an encrypted link to build a cross-engineering blasting knowledge base. The transfer learning algorithm is used to adapt to new engineering geological conditions, realizing the dynamic evolution and cross-scenario generalization capabilities of the model, thus solving the shortcomings of insufficient adaptability of static models.

[0032] (5) Simultaneously construct a multi-objective optimization function for blasting quality indicators (over-excavation rate, block size distribution) and economic benefit indicators (energy consumption, cost), and use the NSGA-II algorithm to solve for the Pareto optimal solution set. The user interface visualizes the quality-economic trade-off relationship to assist in scientific decision-making, realize the linkage optimization of technical and economic dual objectives in the field of blasting, and improve the overall benefits. Attached Figure Description

[0033] Figure 1 This is a flowchart of the intelligent closed-loop control method for tunnel construction cross-section contour according to the present invention.

[0034] Figure 2 This is a diagram of the architecture of the intelligent closed-loop control system for tunnel construction cross-section contours according to the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be described in more detail below with reference to the accompanying drawings. The described embodiments are some, but not all, embodiments of this invention, and are exemplary and intended to explain the invention, not to limit it. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0036] Example 1: Closed-loop control method

[0037] For a specific example, please refer to Figure 1 to Figure 2The intelligent closed-loop control method for tunnel construction cross-section contour provided in this invention sequentially completes pre-blasting sensing, parameter generation, drilling execution, post-blasting re-measurement, deviation quantification, and incremental update within a single construction cycle, and outputs the contour line and parameters for the next cycle to form iterative control. Specifically, the method mainly includes the following steps in implementation:

[0038] SS1. Multi-source information acquisition and fusion:

[0039] Before blasting, point cloud data of the tunnel excavation face is collected in real time by a ground-based three-dimensional laser scanner. Simultaneously, the distribution of rock fissures and uniaxial compressive strength data detected by ground-penetrating radar are integrated to form multi-source data characterizing the geometric morphology of the excavation face and the geological properties of the surrounding rock.

[0040] In specific implementation, it is preferable to establish a construction reference coordinate system consistent with the tunnel axis before collecting point cloud data of the tunnel excavation face, and to perform time synchronization and spatial calibration of the ground 3D laser scanner, ground-penetrating radar and IMU positioning device; in particular, the extrinsic parameters of each sensor are obtained through a common calibration target, the coordinate transformation relationship between the point cloud coordinate system and the ground-penetrating radar detection coordinate system and the IMU pose coordinate system is determined, and a unified timestamp and coordinate system label are added to the multi-source data to ensure that the subsequent construction of spatial geological fusion features and execution control mapping have a consistent data benchmark.

[0041] Point cloud data of the tunnel excavation face is acquired in real time using a ground-based 3D laser scanner (sampling rate ≥ 500,000 points / second), and simultaneously fused with rock fracture distribution, uniaxial compressive strength, and rock mass integrity index (RQD ≥ 70%) detected by ground-penetrating radar. The point cloud data undergoes noise reduction and topology reconstruction to extract the attitude (dip, dip angle) and contour curvature features of rock joint surfaces; ground-penetrating radar data provides fracture spatial distribution and strength parameters, constructing a 3D geological-geometric feature tensor. This overcomes the limitations of traditional single data sources (such as relying solely on point clouds), achieving multi-dimensional collaborative perception of geology and geometry, and providing high-precision input for dynamic blasting design.

[0042] In front of the tunnel excavation face, a ground-based 3D laser scanner collects high-precision point cloud data in real time at a sampling rate of no less than 500,000 points per second. Simultaneously, ground-penetrating radar is activated to detect the distribution of rock fractures, uniaxial compressive strength, and rock mass integrity index (RQD≥70%). The point cloud data undergoes statistical outlier removal and noise reduction processing to eliminate environmental interference such as dust and water mist. Then, the topological structure is reconstructed using the Delaunay triangulation algorithm to accurately extract the attitude (such as dip and dip angle) and contour curvature characteristics of rock mass joint surfaces (curvature radius accuracy ±0.1m). The ground-penetrating radar data is analyzed using time-domain reflectometry to determine the fracture density (fractures / meter) and rock mass strength grading (RMR≥60), ensuring the reliability of geological parameters.

[0043] Based on the collaborative processing of high-precision point cloud geometric features and geological parameters, the joint orientation, contour curvature and fracture distribution, and uniaxial compressive strength of rock mass are fused through feature cascading to construct a three-dimensional geological-geometric feature tensor. This tensor has dimensions [H×W×C], where H and W represent a spatial resolution of 1 cm / pixel, and C contains both geometric features (3 channels) and geological features (3 channels). The fusion process employs Min-Max normalization to eliminate dimensional differences, achieving scale unification of multi-source data and providing high-fidelity input for blasting design.

[0044] To improve the real-time performance and accuracy of data acquisition, the scanner adopts a multi-view synchronous acquisition method (scanning time per station ≤3 minutes) to avoid blind spots in single-station scanning; the ground-penetrating radar emits electromagnetic waves at a frequency of 10MHz, with a penetration depth ≥20m and a spatial resolution of ±5cm. The spatial positioning accuracy of the fused data reaches ±1mm, meeting the sub-millimeter level blasting design input requirements and significantly outperforming the ±10cm error level of traditional manual surveying.

[0045] The multi-source fusion mechanism of this invention effectively solves the problem of blasting design deviations caused by the lack of geological information. Taking a national highway tunnel project as an example, in weak interlayer areas, by integrating fracture distribution data, the spacing between blast holes was dynamically reduced from the empirical value of 30cm to 20cm, and the over-excavation was reduced by 22%, achieving source control of geological adaptability. Through the collaborative operation of three-dimensional laser scanning and ground-penetrating radar, the data acquisition before blasting not only covers the surface geometry of the rock mass but also deeply depicts the internal structural characteristics, laying a solid foundation for the subsequent dynamic generation of blasting designs.

[0046] SS2. Intelligent generation of contours and parameters:

[0047] Multi-source data is input into a convolutional neural network model, and spatial-geological fusion features are extracted through a residual convolutional network to dynamically generate theoretical blasting contour lines and supporting blasting parameters. The blasting parameters include at least the borehole spacing, charge density gradient, and detonation sequence. The borehole spacing is adaptively adjusted according to the fracture density, and the charge density gradient is set in segments according to the rock mass strength.

[0048] In this invention, the theoretical blasting contour line refers to a three-dimensional contour line of the tunnel after blasting under an ideal state, calculated in real time by the blasting contour generation module based on multi-source data (including tunnel excavation face point cloud data, rock mass fracture distribution, uniaxial compressive strength, etc.) collected before blasting. It represents the precise blasting target expected to be achieved in the current loop. The step of inputting the multi-source data into a convolutional neural network model to dynamically generate the theoretical blasting contour line and associated blasting parameters specifically includes:

[0049] S201. Noise reduction and topology reconstruction are performed on the point cloud data of the tunnel excavation face collected before blasting to extract the attitude and contour curvature features of the rock mass joint surfaces. In this embodiment of the invention, the point cloud data of the tunnel excavation face collected before blasting, P={p1,p2,…,p…}, is processed. n} (p i ∈R 3 Noise reduction processing is performed. First, outlier removal is performed, and outliers are deleted. p i ,if , among which m The average distance of the neighborhood. s Standard deviation, α =1.5 is the threshold coefficient. Next, Delaunay triangulation is performed, and the denoised point cloud is projected onto the 2D tangent plane to generate P'. The triangulation is performed to satisfy the empty circle property. ∈ T , C ( t )∩ P ′= , t It is any two-dimensional triangular element in the triangular mesh, and T is the set of triangular meshes generated after partitioning. C ( t ) is the circumcircle. P ' is the set of points projected onto the two-dimensional tangent plane. Output a three-dimensional triangular mesh M={t1,t2,…,t…} m}, extracting the dip of the joint surfaces of the rock mass α ,inclination β and the Gaussian curvature K of the contour curvature characteristics = k1k2. Here, k1 is the first principal curvature, i.e., the curvature value corresponding to the normal section with the largest curvature at a certain point on the surface, and k2 is the second principal curvature, the curvature value corresponding to the normal section orthogonal to the first principal direction at the same point on the surface. By processing the point cloud data before blasting through noise reduction and topology reconstruction, the attitude (dip, dip angle) and contour curvature features of rock mass joint surfaces are extracted (Gaussian curvature accuracy ±0.01), effectively eliminating the ±5° measurement error caused by dust interference, improving the accuracy of geological feature extraction to ±1.5°, providing high-fidelity input for dynamic blasting design, optimizing the borehole dip angle based on accurate joint attitude, and significantly reducing under-excavation.

[0050] S202. A three-dimensional geological-geometric feature tensor is constructed by integrating real-time rock fracture distribution, uniaxial compressive strength, and rock mass integrity index (RQD) detected by ground-penetrating radar; specifically, this involves fusing ground-penetrating radar data with point cloud features. The geological parameter vector is g=[λ,σ]. c [,RQD], where λ is the fracture density, λ∈R + ;σ c For uniaxial compressive strength, σ c∈[20,120]MPa; RQD∈[0,100]%. The geometric eigenvector is d=[α,β,K], where α ∈[0,360°], β∈[0,90°], K∈R. The construction of the feature tensor includes F=Concat(g,d)∈R. H×W×6 The spatial resolution is H×W=1 cm / pixel, and the number of channels is 6.

[0051] The rock mass parameters detected by ground-penetrating radar (fracture density λ, uniaxial compressive strength σ) c The Rock Mass Integrity Index (RQD) and geometric features extracted from point clouds (joint orientation α / β, Gaussian curvature K) are concatenated to construct a three-dimensional tensor, achieving a synergistic expression of geological attributes and spatial morphology, avoiding the limitations of single point cloud input in existing technologies. The tensor channel dimensions (6 channels) structurally organize geological and geometric features, directly inputting them into a convolutional neural network to generate blasting parameters; the fracture density λ drives dynamic adjustment of the borehole spacing, such as reducing the time interval from λ=28 fractures / m to 18cm, achieving geologically adaptable design.

[0052] S203. Process the feature tensor using a three-layer residual convolutional network to extract spatial-geological fusion features. Input the three-dimensional geological-geometric feature tensor F∈R constructed in step S202. H×W×6 (H×W represents the spatial dimensions, with 6 channels). Fusion features are extracted using a three-layer residual convolutional network, specifically including:

[0053] First-layer feature extraction: X1=ReLU(BN(Conv) 3×3 (F,64))) is set to a 3×3 convolution kernel, outputting 64 channels to capture local fracture patterns and joint topology.

[0054] Construct residual block geological-geometric co-data; enhance geological features and calculate channel attention weights:

[0055] w g =Sigmoid(W g • GAP(X i ))

[0056] GAP stands for Global Average Pooling, W g To enhance learnable parameters, channels for fracture density and rock mass strength are strengthened. Channel weight w g Geological parameters, such as fracture density λ, are activated preferentially.

[0057] The formula for calculating the residual is X2=w g ⊙F res (X1)+X1,F res represents convolution, batch normalization, and ReLU; ⊙ represents channel weighting.

[0058] Finally, spatial features are further refined:

[0059] f fusion = Conv 1×1 ( ReLU ( BN ( Conv 3×3 (X2,128))))

[0060] The final output is a 128-dimensional fused feature, which enables the dynamic generation of borehole spacing and charge gradient, and is then input into the subsequent dynamic parameter generation module.

[0061] This invention fuses geological and geometric features through a three-layer residual convolutional network and introduces a geologically sensitive channel attention mechanism. It prioritizes the enhancement of key geological channels such as fracture density and rock mass strength, and drives precise generation of blasting parameters through fracture channel weight adjustment. Combined with multi-scale feature extraction, it achieves collaborative modeling of local fracture patterns and global contours, reducing feature error by 5% compared to traditional ResNet. The network depth is compressed to three layers, with edge node inference latency ≤15ms, meeting the real-time control requirements of boreholes. This invention overcomes the limitations of existing technologies in equally processing channels, enabling dynamic parameter generation driven by geological data and significantly improving geological adaptability.

[0062] S204. Based on fusion features, the theoretical blasting contour line and supporting blasting parameters are dynamically generated: the borehole spacing is adaptively adjusted according to the fracture density; the charge gradient is set in segments according to lithological differences; the detonation sequence is generated as a millisecond-level delay sequence based on the contour curvature; among which, the parameters generated based on fusion features are specifically as follows:

[0063] The borehole spacing is adaptively adjusted to:

[0064]

[0065] Where it is driven by the fracture density λ, d min =15cm, d max =50cm, k=0.1, λ0=25.

[0066] The charge gradient segmentation is set as follows:

[0067]

[0068] The detonation sequence is generated as follows:

[0069]

[0070] Curvature gradient Drive, t min =20ms, t max =50ms.

[0071] By adaptively adjusting the borehole spacing driven by fracture density, controlling the charge gradient in segments based on lithological strength, and modulating the initiation sequence using the contour curvature gradient, dynamic and precise generation of blasting parameters is achieved. The fracture density λ is controlled by a nonlinear function to smoothly transition the borehole spacing between 15-50 cm, avoiding deviations from empirical values; the curvature gradient... It directly correlates with millisecond-level delay, suppresses stress concentration in contour transition zones, breaks through the limitations of static parameters, and achieves active pre-control under geological constraints.

[0072] S205. Using rock mass integrity data as a priori constraints, the blasting parameters are weighted and corrected using an attention mechanism. The rock mass integrity index (RQD) is used as the constraint correction parameter, and the attention weight is calculated as w = Sigmoid(W a ·RQD+b a W a ,b a These are learnable parameters; i final = w · i init +(1- w )· i safe , i init Let θ be the initial parameter. safe This is a safety threshold parameter.

[0073] Using the rock mass integrity index as a constraint, blasting parameters are corrected through an attention mechanism. Channel weights generated by the Sigmoid function are calculated, and initial parameters are weighted and fused with safety threshold parameters to ensure automatic reduction of charge gradient in low RQD regions, avoiding the risk of excessive rock mass fragmentation. Geological indicators are embedded as prior conditions into the parameter generation process, overcoming the limitation of existing technologies lacking geological constraints.

[0074] Furthermore, by inputting multi-source data into a convolutional neural network model to dynamically generate theoretical blasting contour lines and corresponding blasting parameters, the method also includes:

[0075] S206. Construct a multi-objective optimization function and simultaneously define blasting quality indicators and economic benefit indicators; blasting quality indicators include over-excavation and under-excavation rates, block size distribution uniformity, and contour smoothness; economic benefit indicators include explosive consumption, drilling energy consumption, and support material costs.

[0076] Among them, the over-excavation and under-excavation rates of the blasting quality index quantification function are:

[0077]

[0078] V overFor over-excavation volume, V under For the volume of under-excavation, V design This represents the theoretical blast volume.

[0079] The uniformity of block size distribution is:

[0080]

[0081] d i For the first i Regional average grain size of rock blocks The overall average particle size.

[0082] The smoothness of the contour is:

[0083]

[0084] k j For the first on the outline j Gaussian curvature of sampling points.

[0085] The explosive consumption amount in the economic benefit index quantification function is:

[0086]

[0087] r k For the first k Hole charge density, L k This represents the depth of the borehole.

[0088] Drilling energy consumption is:

[0089] g2( i )= P · t drill

[0090] P For drilling rig power, t drill This represents the total drilling time.

[0091] The cost of support materials is:

[0092] g 3( i )= C concrete · V over + C bolt · N bolt

[0093] C concrete C represents the unit price of concrete. bolt The unit price of anchor bolts N bolt This represents the number of anchor bolts.

[0094] Therefore, the multi-objective optimization function is:

[0095]

[0096] i For the combination of blasting parameters, such as i =[ d hole , r charge , t delay ].

[0097] S207. Based on the generated dynamic blasting parameters, the values ​​of each index are quantified through a three-dimensional parametric model to generate an initial parameter combination set.

[0098] S208. The non-dominated sorting genetic algorithm NSGA-II is used to solve for the Pareto optimal solution set and screen blasting parameter combinations that simultaneously satisfy quality and cost constraints. First, the blasting parameter combinations (borehole spacing, charge gradient, detonation sequence, etc.) are encoded as chromosomes, with strictly limited parameter value ranges, and multiple sets of parameter combinations are randomly generated as the initial population. Then, non-dominated sorting is performed: the objective function value of each parameter combination on blasting quality indicators (over-excavation / under-excavation rate, block size uniformity, contour smoothness) and economic benefit indicators (explosive consumption, drilling energy consumption, support cost) is calculated, and the combinations are sorted hierarchically according to the objective value—if a combination is not inferior to another combination in all indicators and is better in at least one indicator, it is determined to be dominant; undominated solutions are included in the optimal non-dominated solution set, and the rest are stratified according to the dominance relationship. Then, individuals in the same layer are sorted according to the objective function value, and the crowding distance of each individual in the target space, i.e., the sum of the function value differences between adjacent solutions, is calculated to ensure the uniformity of the solution set distribution. Parents are selected based on hierarchical sorting and crowding, and offspring are generated by exchanging gene fragments with a high probability. Diversity is introduced by randomly adjusting the gene values ​​of offspring with a low probability. The parent and offspring populations are merged, and a new population is selected by reordering. This process is iterated until convergence, and finally, all optimal non-dominated solutions are output, forming a set of parameter combinations that balance blasting quality and economic benefits.

[0099] The non-dominated sorting genetic algorithm effectively solves the limitations of single-objective optimization, simultaneously balancing over-mining and under-mining control (technical) and cost control (economic), achieving dual-objective collaborative optimization; the crowding mechanism ensures that parameter combinations cover the entire spectrum of quality-cost trade-offs, and the solution set is evenly distributed; the crossover and mutation probabilities are calibrated with blasting parameter sensitivity, improving convergence efficiency by 40%, which is significantly better than traditional methods.

[0100] S209. Visualize the Pareto front in the user interface to show the quality-economy trade-offs of different parameter combinations: present the non-dominated solution set output by NSGA-II in the form of two-dimensional or three-dimensional scatter plots, with the coordinate axes corresponding to indicators such as over-excavation / under-excavation rate, contour smoothness, explosive consumption, or support cost; at the same time, provide filtering conditions such as surrounding rock grade, fracture density range, and RQD range, and display key blasting parameters and their predicted index values ​​for each solution, support highlighting feasible solutions according to constraint thresholds and generating comparative summaries.

[0101] S210. Receive user interaction commands, lock the target parameter combination and output it to the precision execution control module: Based on the target solution or target range selected by the user, automatically complete parameter consistency verification and safety boundary verification, including at least the upper and lower limits of borehole spacing, the legality of charge density gradient segmentation, the detonation timing delay range and the satisfaction of clearance constraints; then generate a unique version identifier and check code for the combination, form a structured distribution package, including borehole layout control information, parameter values, partition labels and applicable working condition descriptions, and write the distribution results to the log for subsequent deviation traceability and rollback.

[0102] Preferably, step SS2 further includes rasterizing and channelizing the multi-source data to form a multi-channel tensor for input to the convolutional neural network model; wherein, the point cloud data is voxelized or projected rasterized to obtain the geometric height field and normal vector field channels, the fracture density, fracture orientation and uniaxial compressive strength are mapped to the geological attribute channels, and the scale is normalized and missing values ​​are filled in for each channel, so that the convolutional neural network model can extract spatial and geological coupling features at a uniform resolution.

[0103] SS3. Precise execution of closed-loop control:

[0104] The theoretical blasting contour line is analyzed to obtain borehole layout control information. The drill arm pose data fed back by the IMU positioning device is mapped to the spatial coordinate system of the theoretical blasting contour line through a homogeneous transformation matrix. The control quantity is constructed based on the deviation between the pose data and the borehole layout control information, and the drill arm angle is corrected in real time through a PID controller.

[0105] The process involves analyzing the theoretical blasting contour line through a precision execution control module, combining it with the construction machinery posture data fed back by the IMU positioning device, and then using a PID controller to correct the drill arm angle in real time. Specifically, this includes:

[0106] S301. Establish a spatial coordinate system for the theoretical blasting profile, and map the drill arm pose data fed back by the IMU positioning device to this coordinate system through a homogeneous transformation matrix.

[0107] First, a spatial rectangular coordinate system is established based on the tunnel design axis. The origin of the coordinate system is fixed at the center point of the starting section of the current construction cycle, ensuring strict alignment with the blasting design starting point. The Z-axis extends along the tunnel excavation direction, pointing towards the unexcavated area; the X-axis points horizontally towards the normal direction of the tunnel outline; and the Y-axis points vertically upward. The three axes satisfy the spatial relationship of the right-hand rule. This provides a unified spatial reference for the blasting outline, ensuring spatial consistency between the borehole trajectory and the design outline.

[0108] The IMU positioning device provides real-time feedback on the position and attitude data of the drill arm's end effector (including three-dimensional coordinates and pitch, yaw, and roll angles). Data mapping is achieved through a homogeneous transformation matrix: rotation mapping: a rotation matrix is ​​calculated based on the drill arm's pitch, yaw, and roll angles. The rotation sequence is as follows: rotation around the Z-axis (yaw angle), rotation around the Y-axis (pitch angle), and rotation around the X-axis (roll angle), ensuring that the attitude transformation conforms to the actual movement logic of the drill arm. Translation mapping: based on the IMU's physical installation position on the drill arm, its geometric offset (including offset values ​​in the X, Y, and Z directions) from the drill arm's end effector is calculated, and the coordinates are corrected to the actual working position of the drill arm using a translation vector. The mapping process outputs the drill arm's precise pose in the design coordinate system through matrix multiplication, providing input for subsequent deviation calculations.

[0109] S302. Real-time calculation of the pose deviation between the drilling trajectory and the theoretical contour line, generating drill arm angle correction commands through a PID controller. The pose deviation between the drilling trajectory and the theoretical contour line is calculated in real time. The three-dimensional coordinates (X, Y, Z) of the preset target point on the theoretical blasting contour line are compared with the actual drill arm end position (mapped from IMU positioning data to the design coordinate system) to generate position deviation (three-dimensional coordinate difference). Simultaneously, the differences between the actual pitch angle and yaw angle of the drill arm and the design angle are compared to generate angle deviation (pitch angle difference, yaw angle difference). Position deviation reflects spatial offset, and angle deviation reflects the degree of attitude inaccuracy; together, they constitute the pose deviation vector, providing input signals for PID control. The PID controller dynamically generates drill arm angle correction commands based on the pose deviation vector: the outputs of the three stages are superimposed to form the final correction command, controlling the drill arm to adjust the pitch angle and yaw angle in real time.

[0110] S303. When the position deviation exceeds the preset range, the hydraulic actuator is triggered to adjust the drill arm pitch angle: the position deviation modulus and attitude deviation components calculated in step S302 are monitored in real time. When any index exceeds the corresponding threshold, the precision execution control module sends a linkage correction command to the hydraulic servo system. The command includes the target increment of pitch angle and yaw angle, angular velocity limit and acceleration limit, and adaptively adjusts the control gain in combination with the current drill arm load pressure and oil temperature state to suppress overshoot and oscillation. During the correction process, the IMU feedback position is continuously collected and iteratively corrected multiple times until the deviation falls back to the allowable range and meets the clearance constraint and drilling safety attitude constraint.

[0111] S304. Verify the matching degree between the corrected trajectory and the theoretical contour line through closed-loop feedback until the deviation converges to within the allowable threshold. When the position deviation exceeds the preset threshold (e.g., 3 mm) or the angle deviation exceeds the tolerance (e.g., pitch angle deviation > 1 degree), immediately trigger the hydraulic servo system to adjust the drill arm pitch angle. After adjustment, new pose data is fed back through the IMU, the deviation is recalculated and iteratively corrected until the matching degree between the trajectory and the theoretical contour line meets the requirements (e.g., position deviation ≤ 5 mm, angle deviation ≤ 0.5 degrees).

[0112] As a preferred option, when constructing the trajectory tracking control quantity based on the borehole layout control information in step SS3, the pose deviation is decomposed into position deviation and attitude deviation. Among them, the boreholes near the arch crown and arch shoulder area are given a high weight coefficient, and the boreholes near the side wall and invert arch area are given a constraint weight related to the clearance control. The weighted control quantity is then input into the PID controller to form a multi-region differentiated correction command, so that the drill arm angle correction can simultaneously meet the local contour accuracy and the overall cross-sectional clearance constraint.

[0113] SS4. Obtaining post-explosion cross-section data:

[0114] Following the blast, a ground-based 3D laser scanner was used to collect point cloud data of the tunnel cross-section in real time. The scanner employed a station-mounted, multi-view synchronous scanning mode (single-station scanning time ≤ 3 minutes), capturing the 3D spatial coordinates of the post-blast contour in real time at a sampling rate of ≥ 500,000 points / second. To avoid interference from blasting dust, a high-pressure water mist dust suppression system was simultaneously activated, and the scanning path was planned to avoid areas with dust concentrations > 50 mg / m³. 3 The scanning process must cover the entire cross-section without blind spots: A multi-view station is set up every 5 meters along the tunnel axis, with each station rotating and scanning at a 120° overlap rate to ensure the integrity of the point cloud stitching; coordinate system calibration is based on the coordinate system of the pre-blasting scan, and cross-period data spatial alignment is achieved through target ball control points (accuracy ±1mm); data quality control involves real-time filtering of dynamic interference points (such as falling debris), and point cloud density analysis (≥500 points / m²) is used. 2 Verify coverage integrity.

[0115] Through multi-view redundant scanning and coordinate system calibration, the actual contour point cloud positioning accuracy is ±1mm, providing high-fidelity input for the deviation field calculation in step SS5; water mist dust suppression and path planning avoid dust interference, and the effective data acquisition rate in the Fengjian Expressway tunnel application is increased to 98%; the full-section scanning time is ≤15 minutes (100-meter advance), matching the tunnel construction cycle rhythm.

[0116] SS5. Quantitative Construction of Cross-Sectional Deviation Field:

[0117] An iterative point cloud registration algorithm was used to calculate the Euclidean distance field between the post-blast tunnel cross-section point cloud data and the theoretical blasting contour line. Three-dimensional deviation field data was generated through voxelized mesh segmentation and integral calculation to mark the spatial coordinates and volumetric deviations of over- and under-excavation. Specifically, this includes:

[0118] S501. The iterative nearest point (ICP) algorithm is used to register the actual point cloud data collected after blasting with the theoretical blasting contour line to achieve spatial alignment with sub-millimeter accuracy.

[0119] A high-precision registration of the actual point cloud and the theoretical contour line after blasting is achieved using an iterative nearest-point algorithm. First, nearest-point matching is performed: for each point in the actual point cloud, the corresponding point in the theoretical point cloud with the closest Euclidean distance is searched, establishing a point-to-point mapping relationship. Then, a rigid transformation is used to solve the problem: the covariance matrix of the actual and theoretical point sets is calculated, and the optimal rotation matrix and translation vector are obtained through matrix decomposition, aligning the actual point cloud to the theoretical contour line space. Finally, iterative updates are performed: the positions of the actual point cloud are updated using rotation and translation transformations, and the matching-solving-update process is repeated until the mean square error change is less than one ten-thousandth or the number of iterations exceeds 50, ensuring registration convergence.

[0120] To improve adaptability to tunnel blasting scenarios, three optimizations were implemented: accelerated initial alignment by pre-setting the initial transformation matrix based on the coordinate system scanned before blasting, avoiding iterative oscillations caused by random initialization, resulting in a 60% improvement in convergence speed; downsampling by using voxel mesh filtering (2 mm resolution) to compress the point cloud scale while retaining key features such as abrupt changes in contour curvature, reducing computation by 70%; and robust outlier removal by removing point pairs with a matching distance exceeding three standard deviations (such as outliers caused by blasting dust or falling debris), improving anti-interference capability by 90%.

[0121] S502. Calculate the Euclidean distance between each point in the actual point cloud after registration and the corresponding point on the theoretical contour line, and generate a three-dimensional Euclidean distance field.

[0122] S503. Spatial coordinates of over-excavation and under-excavation areas are marked based on Euclidean distance field: a positive distance value marks an over-excavation area; a negative distance value marks an under-excavation area.

[0123] S504. The marked area is divided into voxel grids, and the volume deviation within each voxel grid is calculated by integration to achieve sub-millimeter level precision quantization;

[0124] S505. Output structured 3D deviation field data containing spatial coordinates and volume deviations as input to the incremental learning mechanism.

[0125] This system employs efficient spatial retrieval technology to calculate the spatial distance between each point in the actual point cloud after blasting and the nearest point on the theoretical contour line. Nearest point matching quickly finds the Euclidean distance-closest corresponding point in the theoretical contour point cloud for each point in the actual point cloud. A multi-dimensional spatial tree structure accelerates the retrieval process, with a query time of less than 0.2 seconds for millions of point clouds, representing a 90% improvement in efficiency compared to linear search. Distance quantization calculates the three-dimensional straight-line distance between the actual point and its theoretical counterpart, achieving sub-millimeter accuracy. Positive distance values ​​indicate that the actual point is located outside the theoretical contour line, marked as an over-excavation area; negative distance values ​​indicate that the actual point is located inside the theoretical contour line, marked as an under-excavation area.

[0126] The distance values ​​of discrete point clouds are mapped to a regular 3D mesh space. The tunnel cross-section space is divided into uniform 3D mesh cells, each with a size of 1 mm × 1 mm × 1 mm (resolution), forming a voxelized spatial structure. For each mesh cell, the distance values ​​of all actual points within it are statistically analyzed; the average distance within the cell is calculated as the Euclidean distance field value of that mesh; adjacent cell interpolation is used to smooth the contour edge mesh (such as curvature abrupt change regions) to reduce discretization errors. A four-dimensional tensor with dimensions [height × width × length × 1] is generated to fully record the magnitude and direction of the deviation at each location in space.

[0127] By preprocessing theoretical point clouds using a multidimensional tree structure, millisecond-level nearest point queries are achieved. A bilinear interpolation algorithm is used for the mesh in the contour transition area to eliminate jagged discrete errors at voxel edges and improve the accuracy of curved areas. At the same time, abnormal points (such as blasting dust and gravel interference points) with distance values ​​exceeding three times the standard deviation are automatically identified and removed to ensure data reliability.

[0128] SS6. Threshold-triggered incremental learning update:

[0129] When the volume deviation exceeds a preset threshold, an incremental learning mechanism is triggered. A loss function is constructed using the 3D deviation field data, and the weight parameters of the convolutional neural network model are updated in reverse using a gradient descent algorithm. Specifically, this includes the following sub-steps:

[0130] S601. Normalize the 3D deviation field data to construct a standardized deviation dataset. A standardized deviation dataset is constructed by systematically preprocessing the 3D deviation field data to provide high-quality input for incremental learning. First, the volume deviation of all voxel grids is extracted to form the original deviation vector. The mean and standard deviation of the overall deviation are calculated through statistical analysis, and outliers deviating from the mean by more than three times the standard deviation (such as outliers caused by blasting dust or gravel interference) are removed. Then, a normalization method is selected based on the data distribution characteristics: for uniformly distributed deviations, minimum-maximum normalization is used to linearly compress the deviation values ​​to the 0-1 range; for non-uniformly distributed deviations, Z-score normalization is used to make the data mean 0 and the standard deviation 1, eliminating dimensional differences. Simultaneously, the spatial 3D coordinates of each voxel grid are independently normalized, mapping the actual coordinate values ​​proportionally to the 0-1 range according to the maximum and minimum values.

[0131] Based on the normalized bias and spatial coordinates, a four-dimensional structured dataset is constructed, with each data point containing normalized spatial coordinates and normalized bias. For geologically complex areas (such as fracture-dense zones and weak interlayers), a data augmentation strategy is implemented: samples from key areas are resampled, and their weights are increased by 20%, enhancing the model's adaptability to complex geological conditions.

[0132] Through operations such as anomaly filtering, dimension normalization, spatial compression, and data augmentation, the model avoids oscillation defects caused by unnormalization and eliminates scale differences between spatial location and deviation (e.g., compressing ±10-meter coordinates to [0, 1] and scaling ±5-cubic-meter deviations to [-1, 1]), ensuring the stability of gradient calculation in convolutional neural networks. Simultaneously, outlier filtering improves data robustness. Furthermore, sample balance is enhanced, with key region weights increasing the model's generalization ability by 22%. The final standardized dataset becomes the core input of the incremental learning mechanism, driving the precise evolution of the blasting design model.

[0133] S602. Using the mean squared error (MSE) of the standardized deviation data as the loss function, calculate the prediction bias of the convolutional neural network model.

[0134] Using a standardized deviation dataset as input, a convolutional neural network (CNN) model predicts the deviation between the theoretical blasting profile and the actual profile, quantifying the prediction deviation using mean squared error (MSE) as the loss function. First, a normalized deviation dataset is loaded, eliminating scale differences between spatial coordinates and deviations (e.g., coordinates compressed to the 0-1 range, deviations standardized to a mean of 0 and a standard deviation of 1), ensuring model input stability. The CNN generates the predicted deviation of the theoretical blasting profile based on the input spatial-geological features (e.g., fracture distribution density, rock mass strength grade), extracting features through a three-layer residual convolutional structure, and finally outputting the predicted value through a fully connected layer. Then, the MSE between the predicted and actual deviations is calculated: the average of the squared differences between the predicted and actual values ​​for each sample in a batch of 32 data sets is taken; this loss value reflects the model's prediction accuracy.

[0135] S603. The gradient of the loss function with respect to the model weights is calculated using the Stochastic Gradient Descent (SGD) algorithm, and the weight parameters are updated via backpropagation. This achieves dynamic optimization of the weights in a Convolutional Neural Network (CNN) through the SGD algorithm, enabling deep collaboration with the tunnel blasting closed-loop control system. Using the mean square error (MSE) of the three-dimensional deviation field data as the loss function, geologically relevant weight channels (such as feature layers corresponding to fracture density and rock mass strength) are prioritized for strengthening during backpropagation.

[0136] S604. When the volume deviation of N consecutive construction cycles exceeds the preset threshold, the model optimization is forcibly triggered and the learning rate is reduced; when the deviation of consecutive cycles exceeds the threshold, the learning rate η is reduced from 0.01 to 0.001 to avoid model oscillation.

[0137] S605. The optimized model weights are encrypted and synchronized to the cloud server for cross-project generalization deployment. Edge computing nodes encrypt the incrementally learned optimized model weights using AES-256 encryption (256-bit key length) and transmit them to the cloud server via HTTPS to prevent man-in-the-middle attacks and data theft, with a transmission latency of ≤2 seconds. The cloud server aggregates optimized weights from multiple projects (e.g., a weak rock layer model for a national highway tunnel, a fractured zone model for a highway), constructs a cross-project blasting knowledge base, and stores it as a structured graph database, supporting lithology-parameter association queries. Based on new engineering geological radar data (e.g., rock mass RQD, fracture density), the generalized model in the knowledge base is adapted to the target scenario using a feature projection algorithm: calculating the cosine similarity of geological features between the new project and knowledge base cases; using a domain adversarial network to align feature distributions and minimize the difference between the source and target domains; and outputting the adapted weights to the target project's edge nodes.

[0138] Preferably, in step SS6, volume deviation thresholds or peak deviation thresholds are set for the arch crown, arch shoulder, sidewall, and invert arch areas respectively, and the volume deviation and local peak deviation are considered simultaneously when the threshold is triggered; when the volume deviation of any area exceeds the corresponding threshold or the local peak deviation exceeds the corresponding limit, the incremental learning mechanism is triggered, and the trigger area label is written into the weight term of the loss function.

[0139] SS7. Closed-loop iterative output:

[0140] Based on the updated convolutional neural network model, the theoretical blasting profile and parameters for the next construction cycle are generated, forming a closed-loop control iteration of the cross-sectional profile.

[0141] This invention utilizes a three-dimensional deviation field feedback module and an iterative point cloud registration (ICP) algorithm to calculate the Euclidean distance field between the post-blasting point cloud and the theoretical contour line, generating sub-millimeter-level precision three-dimensional deviation field data to mark the spatial coordinates and volumetric deviation of over- and under-excavation areas. This refined feedback provides spatial positioning-level deviation signals, rather than macroscopic scoring, thereby driving an incremental learning mechanism to perform targeted weight updates on the convolutional neural network model, achieving precise parameter correction. Simultaneously, before blasting, this invention uses an AI-driven blasting contour generation module to dynamically generate the theoretical blasting contour line and associated parameters based on multi-source data, overcoming the limitations of post-blast correction and forming a complete closed loop of proactive pre-control, precise execution, refined feedback, and iterative optimization. This difference ensures the geological adaptability and spatial accuracy of the blasting design from the source, solving the problem of insufficient control caused by coarse feedback signals and design lag in existing technologies.

[0142] To verify the technical effectiveness and feasibility of this invention, it was applied to specific engineering cases, and the engineering performance under different tunnel conditions was compared as shown in Table 1 below:

[0143] Table 1. Comparison of Key Technical Implementation Points and Engineering Effectiveness under Different Tunnel Working Conditions

[0144]

[0145] Example 2: Closed-loop control system

[0146] Based on Embodiment 1 above, Embodiment 2 further provides an intelligent closed-loop control system for tunnel construction, used to implement the intelligent closed-loop control method for tunnel construction cross-sectional contour of the present invention, mainly including the following modules:

[0147] The three-dimensional perception module is used to collect point cloud data of the tunnel excavation face using a ground-based three-dimensional laser scanner before blasting, and simultaneously integrate rock fracture distribution and uniaxial compressive strength data detected by ground-penetrating radar to form multi-source data; after blasting, it collects point cloud data of the tunnel cross-section after blasting using a ground-based three-dimensional laser scanner.

[0148] The blasting profile generation module is used to input the multi-source data into a pre-trained convolutional neural network model to dynamically generate the theoretical blasting profile of the excavation face and the corresponding blasting parameters, including the borehole depth, charge density gradient and detonation sequence.

[0149] The precision execution control module is used to analyze the theoretical blasting contour line, combine it with the drill arm construction machinery posture data fed back by the IMU positioning device, and correct the drill arm angle in real time through the PID controller.

[0150] The three-dimensional deviation field feedback module is used to calculate the Euclidean distance field between the point cloud data of the tunnel cross section after blasting and the theoretical blasting contour line using an iterative point cloud registration algorithm, and generate three-dimensional deviation field data with marked over- and under-excavation spatial coordinates and volume deviation.

[0151] The incremental learning module is used to trigger the incremental learning mechanism when the volume deviation exceeds a preset threshold. It uses the three-dimensional deviation field data as the loss function and updates the weight parameters of the convolutional neural network model in reverse through the gradient descent algorithm.

[0152] The closed-loop iteration module is used to generate the theoretical blasting profile and supporting blasting parameters for the next construction cycle based on the optimized convolutional neural network model, thus forming a closed-loop control iteration.

[0153] The objectives of this invention have been fully and effectively achieved through the above embodiments. Those skilled in the art will understand that this invention includes, but is not limited to, the contents described in the accompanying drawings and the specific embodiments described above. Although the invention has been described with reference to what is currently considered the most practical and preferred embodiments, it should be understood that the invention is not limited to the disclosed embodiments, and any modifications that do not depart from the functional and structural principles of the invention will be included within the scope of the claims.

Claims

1. A tunnel construction section profile intelligent closed-loop control method, characterized in that, At least the following steps are included: SS1. Before blasting, point cloud data of the tunnel excavation face is collected by ground 3D laser scanner, and the rock mass fracture distribution and uniaxial compressive strength data detected by ground radar are simultaneously integrated to form multi-source data characterizing the geometric morphology of the excavation face and the geological properties of the surrounding rock. SS2. Multi-source data is rasterized and channelized to form a multi-channel feature tensor for inputting into a pre-trained convolutional neural network model. This multi-channel feature tensor includes geometric height and normal vector fields obtained from tunnel excavation face point cloud data through voxelization or projective rasterization, as well as geological attribute channels obtained by mapping fracture density, fracture orientation, and uniaxial compressive strength. The multi-channel feature tensor is processed through a residual convolutional network to extract rock mass joint surface attitude and contour curvature feature parameters, and integrates real-time rock mass fracture distribution and uniaxial compressive strength detected by ground-penetrating radar. Spatial-geological fusion characteristics of compressive strength and rock mass integrity index data are used to generate a theoretical blasting profile of the excavation face and supporting blasting parameters. The blasting parameters include at least the borehole spacing, charge density gradient, and detonation sequence. The theoretical blasting profile of the excavation face is a three-dimensional profile of the tunnel after blasting under ideal conditions, calculated in real time based on multi-source data collected before blasting. The borehole spacing is adaptively adjusted according to the fracture density. The charge density gradient is set in segments according to the rock mass strength. The detonation sequence is generated as a millisecond-level delay sequence based on the profile curvature characteristic parameters. SS3. Analyze the theoretical blasting contour line of the excavation face and obtain the borehole layout control information. Map the drill arm pose data fed back by the IMU positioning device to the spatial coordinate system of the theoretical blasting contour line through a homogeneous transformation matrix. Construct trajectory tracking control quantity based on the deviation between the drill arm pose data and the borehole layout control information. When constructing the trajectory tracking control quantity, the pose deviation is decomposed into position deviation and attitude deviation. The position deviation reflects the spatial offset, and the attitude deviation includes the drill arm pitch angle difference and yaw angle difference and reflects the degree of attitude inaccuracy of the drill arm. The position deviation and attitude deviation together constitute the pose deviation vector. The drill arm angle correction command is dynamically generated by the PID controller to correct the drill arm angle in real time. SS4. Collect point cloud data of the tunnel cross-section after blasting using a ground-based 3D laser scanner; SS5. The iterative point cloud registration algorithm is used to calculate the Euclidean distance field between the point cloud data of the tunnel cross section after blasting and the theoretical blasting contour line of the excavation face. The three-dimensional deviation field data for marking the spatial coordinates and volume deviation of over-excavation and under-excavation is generated by voxel grid segmentation and integral calculation. SS6. When the volume deviation exceeds a preset threshold, an incremental learning mechanism is triggered. A loss function is constructed using the three-dimensional deviation field data, and the weight parameters of the convolutional neural network model are updated in reverse using the gradient descent algorithm. SS7. Based on the updated convolutional neural network model, the theoretical blasting profile and parameters of the excavation face for the next construction cycle are generated, forming a closed-loop control iteration of the cross-sectional profile.

2. The intelligent closed-loop control method for tunnel construction cross-sectional profile according to claim 1, characterized in that, Step SS2 includes the following sub-steps when implemented: S201. Denoise, downsample, and reconstruct the point cloud data of the tunnel excavation face collected before blasting, and extract the rock mass joint surface attitude and contour curvature characteristic parameters. S202. By integrating real-time data on rock mass fracture distribution, uniaxial compressive strength, and rock mass integrity index detected by ground-penetrating radar, a three-dimensional geological-geometric feature tensor is constructed. S203. The feature tensors are processed by a residual convolutional network to extract spatial-geological fusion features jointly characterized by data on rock mass joint surface attitude, contour curvature feature parameters, rock mass fracture distribution, uniaxial compressive strength and rock mass integrity index. S204. Based on spatial-geological fusion characteristics, the theoretical blasting contour line of the excavation face and supporting blasting parameters are dynamically generated. The theoretical blasting contour line of the excavation face is a three-dimensional contour line of the tunnel after blasting under ideal conditions, calculated in real time based on multi-source data collected before blasting. The spacing between blast holes is adaptively adjusted according to the fracture density. The charge density gradient is set in segments according to the rock mass strength. The detonation sequence is generated based on the contour curvature to generate a millisecond-level delay sequence. S205. Using the rock mass integrity index as a priori constraint, the blasting parameters are weighted and corrected through an attention mechanism.

3. The intelligent closed-loop control method for tunnel construction section profile according to claim 2, characterized in that, Step SS2, when implemented, also includes: S206. Construct a multi-objective optimization function with blasting parameter combinations as optimization variables and blasting quality indicators and economic benefit indicators as objective function sets. Simultaneously define blasting quality indicators and economic benefit indicators. The blasting parameter combinations include borehole spacing, charge density gradient and detonation sequence. The blasting quality indicators include at least over-excavation and under-excavation rates, block size distribution uniformity and contour smoothness. The economic benefit indicators include at least explosive consumption, drilling energy consumption and support material cost. S207. Based on the generated blasting parameters, calculate each blasting quality index and economic benefit index through a three-dimensional parametric model to generate an initial parameter combination set; S208. The non-dominated sorting genetic algorithm NSGA-II is used to solve the Pareto optimal solution set and select the combination of blasting parameters that simultaneously satisfies the quality and cost constraints.

4. The intelligent closed-loop control method for cross-section profile of tunnel construction according to claim 1, characterized in that, Step SS3 includes the following when implemented: S301. Establish a spatial coordinate system for the theoretical blasting profile, and map the drill arm pose data fed back by the IMU positioning device to this coordinate system through a homogeneous transformation matrix; S302. Calculate the pose deviation between the drilling trajectory and the theoretical contour line in real time, and generate drill arm angle correction instructions through the PID controller; S303. When the drill arm position deviation exceeds the preset range, the actuator is triggered to adjust the drill arm pitch angle; S304. Verify the matching degree between the corrected trajectory and the theoretical blasting profile through closed-loop feedback until the pose deviation converges to within the allowable threshold.

5. The intelligent closed-loop control method for cross-section profile of tunnel construction according to claim 1, characterized in that, Step SS5 includes the following when implemented: S501. The iterative nearest point ICP algorithm is used to register the actual point cloud data of the tunnel cross section collected after blasting with the theoretical blasting contour line, and the iteration is terminated when the registration error converges to a preset threshold. S502. Calculate the Euclidean distance between each point in the actual point cloud of the registered tunnel cross section and the corresponding point of the theoretical blasting contour line, and generate a three-dimensional Euclidean distance field. S503. Based on the Euclidean distance field, mark the spatial coordinates of the over-excavation and under-excavation areas. When the distance value is positive, mark it as an over-excavation area; when the distance value is negative, mark it as an under-excavation area. S504. Divide the marked region into voxel grids and calculate the volume deviation within each voxel grid by integration; S505. Output three-dimensional deviation field data containing spatial coordinates and volume deviation as input for the incremental learning mechanism.

6. The intelligent closed-loop control method for cross-section profile of tunnel construction according to claim 1, characterized in that, Step SS6 includes the following when implemented: S601. Normalize the three-dimensional deviation field data to construct a standardized deviation dataset; S602. Using the mean squared error (MSE) of the standardized deviation data as the loss function, calculate the prediction bias of the convolutional neural network model; S603. Calculate the gradient of the loss function with respect to the model weights using the stochastic gradient descent algorithm, and update the weight parameters through backpropagation; S604. When the volume deviation of multiple consecutive construction cycles exceeds the preset threshold, the model optimization is forcibly triggered and the learning rate is reduced; S605. Encrypt and synchronize the optimized model weight parameters to the cloud server.

7. The intelligent closed-loop control method for tunnel construction cross-sectional profile according to claim 6, characterized in that, The optimized model weight parameters are encrypted and synchronized to the cloud server, specifically including: The convolutional neural network model is deployed on edge computing nodes and forms a localized processing unit with a 3D laser scanner and an IMU positioning device to achieve real-time response to the generation of blasting parameters. The weight parameters during the incremental learning process are synchronized to the cloud server via an encrypted link; The cloud server aggregates the optimized model weights from multiple projects to build a cross-project blasting knowledge base; Based on transfer learning algorithms, the generalized model in the knowledge base is adapted to new engineering geological conditions; The adapted model weight parameters are then sent to the target project edge nodes.

8. The intelligent closed-loop control method for tunnel construction cross-section contour according to claim 1, characterized in that, In step SS1, before collecting point cloud data of the tunnel excavation face, a construction reference coordinate system consistent with the tunnel axis is established, and the ground 3D laser scanner, ground-penetrating radar and IMU positioning device are synchronized in time and calibrated in space. Among them, the external parameters of each sensor are obtained through a common calibration target, the coordinate transformation relationship between the point cloud coordinate system and the ground-penetrating radar detection coordinate system and the IMU pose coordinate system is determined, and a unified timestamp and coordinate system label are added to the multi-source data.

9. The intelligent closed-loop control method for cross-section profile of tunnel construction according to claim 1, characterized in that, In step SS3, when constructing the trajectory tracking control quantity based on the borehole layout control information, the pose deviation is decomposed into position deviation and attitude deviation. Among them, the boreholes near the arch crown and arch shoulder area are given a high weight coefficient, and the boreholes near the sidewall and invert arch area are given constraint weights related to the clearance control. The weighted control quantity is input into the PID controller to form a multi-region differentiated correction command, so that the drill arm angle correction can simultaneously meet the local contour accuracy and the overall cross-sectional clearance constraint.

10. The intelligent closed-loop control method for tunnel construction cross-section profile according to claim 1, characterized in that, In step SS6, volume deviation thresholds or peak deviation thresholds are set for the arch crown, arch shoulder, sidewall, and invert arch areas respectively, and the volume deviation and local peak deviation are considered simultaneously when the threshold is triggered. When the volume deviation of any area exceeds the corresponding threshold or the local peak deviation exceeds the corresponding limit, the incremental learning mechanism is triggered, and the label of the triggered area is written into the weight term of the loss function.

11. A tunnel construction cross-section contour intelligent closed-loop control system, used to implement the tunnel construction cross-section contour intelligent closed-loop control method according to any one of claims 1 to 10, characterized in that, include: The three-dimensional perception module is used to collect point cloud data of the tunnel excavation face using a ground-based three-dimensional laser scanner before blasting, and simultaneously integrate rock mass fracture distribution and uniaxial compressive strength data detected by ground-penetrating radar to form multi-source data; after blasting, the ground-based three-dimensional laser scanner collects point cloud data of the tunnel cross section in real time. The blasting profile generation module is used to input the multi-source data into a pre-trained convolutional neural network model to dynamically generate the theoretical blasting profile of the excavation face and the corresponding blasting parameters, including the borehole spacing, charge density gradient and detonation sequence. The precision execution control module is used to analyze the theoretical blasting contour line, combine it with the drill arm posture data fed back by the IMU positioning device, and correct the drill arm angle in real time through the PID controller. The three-dimensional deviation field feedback module is used to calculate the Euclidean distance field between the point cloud data of the tunnel cross section after blasting and the theoretical blasting contour line using an iterative point cloud registration algorithm, and generate three-dimensional deviation field data with marked over- and under-excavation spatial coordinates and volume deviation. The incremental learning module is used to trigger the incremental learning mechanism when the volume deviation exceeds a preset threshold. It uses the three-dimensional deviation field data as the loss function and updates the weight parameters of the convolutional neural network model in reverse through the gradient descent algorithm. The closed-loop iteration module is used to generate the theoretical blasting profile and supporting blasting parameters for the next construction cycle based on the optimized convolutional neural network model, thus forming a closed-loop control iteration.

Images