Tunnel secondary lining steel bar modularization disassembly control system based on multi-dimensional data

Abstract

The application relates to the technical field of adjusting systems and discloses a tunnel secondary lining steel bar modularization disassembly control system based on multidimensional data. The system acquires three-dimensional point cloud data of a tunnel secondary lining surface and performs space registration on a steel bar design building information model, extracts spatial coordinate deviations of actual steel bar nodes, and constructs a stress topology graph with the actual steel bar nodes as vertices, the steel bar connection relationship as edges and the node stress values as edge weights in combination with stress monitoring data. A control device adopts a graph neural network to perform feature extraction on the stress topology graph, calculates stress transmission contribution degrees to determine stress release zero points, generates modularization cutting boundary instructions according to the spatial clustering results of adjacent stress release zero points, and drives an executing mechanism to perform displacement and cutting actions. The scheme bypasses the mechanical conduction path in the high bearing state, overcomes the residual structural stress redistribution imbalance phenomenon, and avoids local instability collapse.

Description

Technical Field

[0001] This invention relates to the field of control system technology and discloses a modular dismantling control system for tunnel secondary lining reinforcement based on multidimensional data. Background Technology

[0002] Tunnel secondary lining reinforcement dismantling control systems typically rely on two-dimensional construction drawings or pre-stored building information models, pre-setting fixed-size grid cutting paths within the control system. During dismantling operations, the control system directly sends displacement and cutting commands to the cutting equipment's actuators according to the preset fixed grid coordinates. This control method depends entirely on static design drawing data; the control system does not include a detection mechanism for the actual spatial shape of the reinforcement after construction, nor does it possess dynamic calculation and analysis capabilities for the mechanical transmission paths within the secondary lining structure. The generation process of the cutting trajectory is independent of the actual stress state of the current structure.

[0003] Due to construction errors during the casting and forming process of the tunnel secondary lining, the actual spatial position of the formed steel bars deviates from the design building information model, and the internal stress distribution of the concrete is non-uniform. Using the existing technology of issuing cutting commands based on a fixed grid, the cutting action will directly cut off the mechanical transmission path in the steel reinforcement skeleton under high load, causing the residual structure of the secondary lining to experience stress redistribution imbalance at the moment of cutting, which in turn leads to local instability and collapse. Summary of the Invention

[0004] The purpose of this invention is to provide a modular dismantling and control system for tunnel secondary lining reinforcement based on multi-dimensional data, which can effectively solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The modular dismantling control system for tunnel secondary lining reinforcement based on multidimensional data includes a data processing device, a control device, and a cutting execution mechanism. The data processing device acquires three-dimensional laser point cloud data of the tunnel secondary lining surface and the reinforcement design BIM model. It uses an iterative nearest point algorithm to spatially register the three-dimensional laser point cloud data and the reinforcement design BIM model and extracts the spatial coordinate deviation of the actual reinforcement nodes. The data processing device combines the stress monitoring data of the secondary lining concrete to construct a stress topology graph with the actual steel bar nodes as vertices, the steel bar connection relationships as edges, and the node stress values ​​as edge weights. The control device uses a graph neural network to extract features from the stress topology graph, calculates the stress transmission contribution of each edge, and defines the steel bar connection node corresponding to the edge whose stress transmission contribution is lower than a preset threshold as the stress release zero point. The control device generates modular cutting boundary commands based on the spatial clustering results of adjacent stress release zero points, and drives the cutting execution mechanism to perform displacement and cutting actions along the modular cutting boundary according to the modular cutting boundary commands until the modular separation of the secondary lining steel bars is completed.

[0006] Preferably, the process by which the data processing device spatially registers the three-dimensional laser point cloud data with the steel reinforcement design BIM model using an iterative nearest-point algorithm includes: the data processing device performing voxel mesh downsampling on the three-dimensional laser point cloud data to extract an initial key point set, and performing coarse registration between the initial key point set and the standard node set in the steel reinforcement design BIM model to obtain an initial rotation and translation matrix. Based on the initial rotation and translation matrix, the data processing device calculates the nearest neighbor distance from each point in the initial key point set to the standard node set, and constructs an objective function that includes point pair matching error; The optimal rigid body transformation matrix that makes the objective function converge is obtained by using singular value decomposition. The optimal rigid body transformation matrix is ​​then applied to the three-dimensional laser point cloud data, and the spatial coordinate deviation of the actual steel bar node is extracted based on the transformed coordinates.

[0007] Preferably, the process of the data processing device constructing the stress topology map includes: the data processing device receiving the stress monitoring data collected by the strain sensor network deployed inside the secondary lining concrete, and mapping the stress monitoring data to the actual steel bar nodes to obtain the initial stress value of the nodes; For each edge in the rebar connection relationship, the data processing device calculates the gradient difference of the initial stress values ​​of the nodes connecting the two ends, and multiplies the gradient difference by the stiffness coefficient of the rebar physical properties to obtain a standardized stress value that eliminates the influence of cross-sectional size differences. The data processing device assigns the standardized stress value to the corresponding edge as the edge weight, generating the stress topology map that integrates spatial location and mechanical distribution.

[0008] Preferably, the process of the control device extracting features from the stress topology graph using a graph neural network includes: the control device inputting the spatial coordinate deviation of the vertex and the edge weight as input feature vectors into the graph convolutional layer; In the graph convolutional layer, the control device updates the hidden state representation of the current node by aggregating the input feature vectors of neighboring nodes defined by the adjacency matrix. The control device inputs the hidden layer state representation after propagation through multiple graph convolutional layers into a multilayer perceptron and outputs the predicted stress transfer contribution, wherein the loss function of the graph neural network is configured as the mean square error between the predicted stress transfer contribution and the actual contribution calculated by finite element simulation.

[0009] Preferably, the process by which the control device generates modular cutting boundary instructions based on the spatial clustering results of adjacent stress release zero points includes: the control device uses the stress release zero point as the initial cluster center and employs a density-based noise application spatial clustering algorithm to divide the steel bar point cloud in the three-dimensional laser point cloud data into regions to obtain multiple initial cutting blocks. For each initial cut block, the control device extracts the block edge point set and projects the block edge point set onto a plane perpendicular to the tunnel longitudinal axis; The control device performs convex hull calculation and polygon fitting on the block edge point set on the plane, and back-projects the fitted polygon boundary onto the three-dimensional spatial coordinate system to generate the modular cutting boundary command.

[0010] Preferably, the process by which the control device drives the cutting actuator according to the modular cutting boundary command includes: the control device discretizing the three-dimensional spatial trajectory in the modular cutting boundary command to generate a series of equidistant spatial navigation point sequences; For the adjacent spatial navigation points, the control device calculates the corresponding joint angle increment based on the inverse kinematic model of the cutting actuator; The control device uses a cubic spline interpolation algorithm to smooth the joint angle increment and generate speed planning curves for each joint motor. The control device sends servo drive pulses to each joint motor according to the speed planning curve, and controls the end tool of the cutting actuator to perform smooth transition displacement and cutting between adjacent spatial navigation points.

[0011] Preferably, after the data processing device extracts the spatial coordinate deviation of the actual rebar node, the system further includes: the data processing device continuously acquiring updated real-time three-dimensional laser point cloud data during the cutting action of the cutting actuator; For the missing point cloud areas in the real-time three-dimensional laser point cloud data that are destroyed by the cutting action, the data processing device performs virtual point cloud filling according to the corresponding structural completion rules in the steel reinforcement design BIM model. The data processing device performs local registration and fine-tuning of the real-time three-dimensional laser point cloud data after filling it in with the three-dimensional laser point cloud data of the currently uncut part, updates the optimal rigid body transformation matrix based on the fine-tuning result, and corrects the spatial coordinate deviation of the subsequent nodes to be cut based on the updated optimal rigid body transformation matrix.

[0012] Preferably, during the process of the control device updating the hidden state representation of the current node by aggregating the input feature vectors of neighboring nodes defined by the adjacency matrix, the system further includes: the control device introducing a spatial attention mechanism in the graph convolutional layer to calculate the spatial distance vector and the angle vector between the current node and any neighboring node of the current node. The control device inputs the spatial distance vector and the angle vector between the steel bars into the attention weight calculation network and outputs attention coefficients that reflect the differences in spatial structure. The control device performs the aggregation operation after multiplying the attention coefficient with the input feature vector, so that the closer the neighboring node is and the smaller the angle between them, the greater the contribution weight of the hidden state representation of the current node.

[0013] Preferably, during the process of the control device using a density-based noise-based spatial clustering algorithm to divide the steel bar point cloud in the three-dimensional laser point cloud data into regions, the system further includes: the control device acquiring the historical cutting trajectory and historical operation time of the cutting actuator, and calculating the predicted heat-affected zone distribution range of each spatial location in the three-dimensional laser point cloud data based on the historical cutting trajectory and the historical operation time; When performing region division, if the line connecting two adjacent stress release zero points passes through the predicted heat-affected zone distribution range, the control device will block the density connectivity between the two stress release zero points and use the predicted heat-affected zone distribution range as an isolation zone to redivide the initial cutting block.

[0014] Preferably, after the control device redivides the initial cutting block as an isolation zone using the predicted heat-affected zone distribution range, the system further includes: after the cutting actuator completes the cutting and separation action on the first initial cutting block, the control device acquires real-time temperature field data around the first initial cutting block collected by the infrared thermal imager. The control device corrects the predicted thermally affected zone distribution range based on the real-time temperature field data to obtain the dynamic actual thermally affected zone. The control device re-executes the blocking and re-division operation on the boundary of the remaining uncut initial cut block based on the dynamic actual heat-affected zone, and generates subsequent modular cutting boundary instructions based on the re-divided block.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention spatially registers 3D laser point cloud data with a reinforced concrete building information model to extract the spatial coordinate deviation of actual reinforced concrete nodes. It then constructs a stress topology graph with actual reinforced concrete nodes as vertices, reinforced concrete connections as edges, and node stress values ​​as edge weights, based on stress monitoring data. A graph neural network is used to calculate the stress transmission contribution of edges to determine the zero-point of stress release. Modular cutting boundary commands are generated based on the spatial clustering results of adjacent zero-points of stress release. This technique transforms actual spatial deviations and internal stress states into graph structural data for analysis, enabling the cutting boundary commands generated by the control system to bypass the mechanical transmission path under high load conditions. This overcomes the stress redistribution imbalance in the secondary lining residual structure caused by the cutting action severing the load-bearing path, thus preventing local instability and collapse during dismantling operations.

[0016] 2. By using singular value decomposition to solve for the optimal rigid body transformation matrix that makes the objective function converge, the interference of accumulated errors in the registration process on the extraction of node coordinate deviations is reduced. By multiplying the gradient difference of the initial stress values ​​of the nodes by the stiffness coefficient to obtain the standardized stress value as the edge weight, the influence of the difference in the cross-sectional size of the reinforcing bars on the construction of the topology graph is eliminated. A spatial attention mechanism is introduced into the graph convolutional layer, and the attention coefficient is calculated based on the spatial distance vector and the angle vector between the reinforcing bars, which enhances the feature extraction capability of the graph neural network for local spatial structural differences. By calculating and predicting the distribution range of the heat-affected zone by historical cutting trajectory and operation time and blocking the density connectivity between stress release zero points, the interference of cutting heat on subsequent block division is eliminated, and the system stability of multi-block continuous dismantling operation is improved. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall workflow of the present invention; Figure 2 This is a flowchart of the spatial registration and deviation extraction process of the present invention; Figure 3 This is a flowchart illustrating the construction process of the stress topology diagram of the present invention. Figure 4 This is a flowchart of the graph neural network feature extraction and attention calculation process of the present invention; Figure 5 This is a flowchart illustrating the process of generating the cutting boundary and blocking the heat-affected zone in this invention. Figure 6 This is a flowchart of the cutting drive and dynamic update process of the present invention. Detailed Implementation

[0018] Please refer to the attached document. Figure 1This embodiment provides a modular dismantling control system for tunnel secondary lining reinforcement based on multi-dimensional data, including a data processing device, a control device, and a cutting actuator. The data processing device and the control device interact with each other via a wired or wireless communication link. The signal output terminal of the control device is electrically connected to the drive unit of the cutting actuator. The drive commands output by the control device can directly act on the cutting actuator to control its displacement and cutting action.

[0019] The data processing device acquires 3D laser point cloud data of the tunnel secondary lining surface and a BIM model of the reinforcement design. The 3D laser point cloud data acquired by the data processing device is a set of 3D coordinate points covering the entire cross-section of the tunnel secondary lining, including the outline of the secondary lining concrete surface, the ends of exposed reinforcement bars, and the locations of reinforcement connectors. Each coordinate point contains X, Y, and Z axis coordinate values ​​corresponding to its spatial location. The coordinate system adopts a unified independent coordinate system for tunnel engineering, where the X-axis extends along the longitudinal axis of the tunnel, the Y-axis extends horizontally along the tunnel cross-section, and the Z-axis extends vertically along the tunnel cross-section. The reinforcement design BIM model acquired simultaneously by the data processing device is a parametric information model of the tunnel secondary lining reinforcement engineering. The model includes the design 3D coordinates of all standard reinforcement nodes, the start and end node numbers of individual reinforcement bars, the connection topology between reinforcement bars, and the material property parameters of the reinforcement bars.

[0020] Please refer to the attached document. Figure 2 The data processing device spatially registers the 3D laser point cloud data with the rebar design BIM model using an iterative nearest-point algorithm, extracting the spatial coordinate deviation of the actual rebar nodes. The core objective of spatial registration is to map the actually acquired 3D laser point cloud data and the design BIM model to the same spatial coordinate system, eliminating coordinate system offset and rotation errors generated during the acquisition process. The data processing device first extracts the actual rebar node set corresponding to the standard rebar nodes in the rebar design BIM model from the 3D laser point cloud data. This point set extraction is achieved through geometric feature recognition of the point cloud, based on the curvature features of the rebar endpoints and the point cloud density features. The cluster centers of the point cloud that conform to the geometric features of the rebar endpoints are used as the initial coordinates of the actual rebar nodes.

[0021] The data processing device spatially registers the point set formed by the initial coordinates of the extracted actual rebar nodes with the point set formed by the coordinates of the standard rebar nodes in the rebar design BIM model using an iterative nearest-point algorithm. During the registration process, the optimal rigid body transformation matrix is ​​solved, which includes rotation and translation parameters in three-dimensional space. The data processing device applies the solved optimal rigid body transformation matrix to the initial coordinates of all actual rebar nodes to obtain the transformed spatial coordinates of the actual rebar nodes. The difference between the transformed spatial coordinates of the actual rebar nodes and the design coordinates of the corresponding standard rebar nodes is calculated to extract the spatial coordinate deviation of each actual rebar node.

[0022] The spatial coordinate deviation is calculated using the following formula:

[0023] in, These represent the spatial coordinate deviations of the i-th actual rebar node in the X-axis, Y-axis, and Z-axis directions, respectively. These are the X-axis, Y-axis, and Z-axis coordinates of the i-th actual rebar node after transformation by the rigid body transformation matrix; These are the design X-axis, Y-axis, and Z-axis coordinates of the standard rebar node corresponding to the i-th actual rebar node.

[0024] Please refer to the attached document. Figure 3 The data processing device combines stress monitoring data from the secondary lining concrete to construct a stress topology graph with actual rebar nodes as vertices, rebar connection relationships as edges, and node stress values ​​as edge weights. The data processing device receives stress monitoring data collected by a network of strain sensors deployed within the secondary lining concrete. This data includes real-time strain values ​​at each monitoring point within the secondary lining concrete, with each strain monitoring point corresponding to the spatial location of one or more actual rebar nodes. The data processing device converts the collected strain monitoring data into stress values ​​at the corresponding monitoring points using constitutive relations in mechanics of materials. Then, through spatial interpolation, it maps the stress values ​​at the monitoring points to all actual rebar nodes, obtaining the initial stress value at each actual rebar node.

[0025] The data processing device uses all actual rebar nodes as the vertex set of the stress topology graph. The attributes of each vertex include the spatial coordinate deviation and the initial stress value of the node. Based on the rebar connection relationship in the rebar design BIM model and the spatial coordinate matching results of the actual rebar nodes, the data processing device determines the connection relationship between vertices. Each set of two vertices with direct rebar connection forms an edge in the stress topology graph. The data processing device assigns the calculation results related to the initial stress values ​​of the two nodes connected to the rebar corresponding to the edge as the edge weight, and finally generates a stress topology graph that integrates spatial location information and mechanical distribution information.

[0026] In this embodiment, the correspondence between the core components of the stress topology diagram and the physical parameters of the tunnel secondary lining reinforcement is shown in the table below: Table 1. Mapping of Stress Topology Diagram Elements and Physical Parameters of Tunnel Secondary Lining Reinforcement

[0027] This table clarifies the correspondence between the core components of the stress topology diagram and the physical parameters of the tunnel secondary lining reinforcement, ensuring that the construction process of the stress topology diagram is entirely based on the actual structural parameters and stress state of the tunnel secondary lining, and avoiding deviations between the topology diagram structure and the actual reinforcement skeleton. The parameters defined in this table are the basic input parameters for the subsequent graph neural network feature extraction process; the mapping relationship of these parameters directly determines the degree of matching between the feature extraction results and the actual mechanical transmission path.

[0028] The control device employs a graph neural network to extract features from the stress topology graph, calculates the stress transmission contribution of each edge, and defines the rebar connection nodes corresponding to edges with stress transmission contributions below a preset threshold as stress release zero points. The control device receives stress topology graph data output from the data processing device and uses the vertex attributes and edge weights of the stress topology graph as input data to the graph neural network. The graph neural network includes multiple graph convolutional layers and a multilayer perceptron. The graph convolutional layers are used to extract local structural features and global topological features of the stress topology graph, while the multilayer perceptron is used to map the extracted features to the stress transmission contribution of each edge.

[0029] The control device concatenates the spatial coordinate deviation of each vertex with the initial stress value of the node to form the input feature vector of the vertex, and uses the edge weights as the input features of the edges, which are then input into the graph convolutional layer. During the operation of the graph convolutional layer, the features of each vertex are aggregated and updated based on the features of the neighboring nodes defined by the adjacency matrix. After feature propagation through multiple layers of graph convolutional layers, the features of each vertex and edge are fused with the spatial structure information and mechanical distribution information of its local region.

[0030] The feature aggregation and node hidden state update of the graph convolutional layer are performed using the following formula:

[0031] in, For vertex v in the stress topology graph, at the 1st... Hidden state representation of the output of a layer graph convolutional layer; It is a non-linear activation function; Let v be the set of all neighboring nodes of vertex v in the stress topology graph; is the feature aggregation normalization coefficient between vertex v and its neighbor u, and its value is the product of the degree of vertex v and the square root of the degree of vertex u. Let be the trainable weight matrix of the l-th graph convolutional layer; For vertex u at the th Hidden state representation of the output of a layer graph convolutional layer; For the first Trainable bias terms for layer graph convolutional layers.

[0032] The control device inputs the features propagated through multiple graph convolutional layers into a multilayer perceptron. The output layer of the multilayer perceptron is a single-neuron structure, and the output value is the stress transmission contribution of the corresponding edge. The stress transmission contribution ranges from 0 to 1, with a higher value indicating a greater role played by the corresponding rebar in the stress transmission path of the entire secondary lining rebar skeleton. The control device presets a threshold for the stress transmission contribution, which is pre-set according to the safety control requirements of the tunnel secondary lining structure. The control device compares the stress transmission contribution of each edge with the preset threshold. When the stress transmission contribution of an edge is lower than the preset threshold, the two rebar connection nodes corresponding to that edge are defined as stress release zero points.

[0033] During the training of a graph neural network, the loss function uses the following formula:

[0034] in, represents the loss function value of the graph neural network; M represents the total number of edges in the stress topology graph. The contribution of the predicted stress transfer to the m-th edge output by the graph neural network; This represents the actual stress transfer contribution of the m-th edge obtained through finite element simulation.

[0035] The control device generates modular cutting boundary commands based on the spatial clustering results of adjacent stress release zero points, and drives the cutting actuator to perform displacement and cutting actions along the modular cutting boundary according to the modular cutting boundary commands until the modular separation of the secondary lining reinforcement is completed. The control device acquires the three-dimensional spatial coordinates of all stress release zero points and performs spatial clustering processing on all stress release zero points. The core of spatial clustering processing is to divide the stress release zero points that are spatially adjacent and located in the same reinforcement skeleton area into the same cluster. The control device uses the stress release zero points in each cluster as boundary control points to generate a closed three-dimensional spatial cutting boundary, which is the modular cutting boundary. The control device converts the three-dimensional spatial coordinates, cutting feed direction, and cutting speed parameters of the modular cutting boundary into modular cutting boundary commands that can be recognized by the cutting actuator.

[0036] The control device outputs modular cutting boundary commands to the drive unit of the cutting actuator. The drive unit, based on the commands, controls the end-effector of the cutting actuator to move spatially along the modular cutting boundary, simultaneously initiating the cutting action to cut the reinforcing bars corresponding to the modular cutting boundary. Once all reinforcing bars corresponding to the modular cutting boundary have been cut, the secondary lining reinforcing bar module enclosed by that boundary is separated from the surrounding reinforcing bar skeleton. The control device then continues to drive the cutting actuator to execute the cutting action corresponding to the next modular cutting boundary, until all reinforcing bars within the target dismantling area of ​​the tunnel secondary lining have been modularly separated.

[0037] In this embodiment, the spatial coordinate deviation of the actual rebar nodes is obtained by spatial registration of 3D laser point cloud data and the rebar design BIM model, solving the problem of deviation from the actual formed structure caused by relying solely on static design drawings in the prior art. By constructing a stress topology graph that integrates spatial location and mechanical distribution, the mechanical transmission path of the rebar skeleton is transformed into graph structure data that can be extracted for features. The stress transmission contribution of each rebar is calculated by graph neural network, and the stress release zero point that will not affect the overall structural stress is determined. Based on the stress release zero point, a modular cutting boundary is generated, which makes the cutting path bypass the high-load mechanical transmission path and avoids stress redistribution imbalance and local instability and collapse of the secondary lining residual structure during the cutting process.

[0038] In a preferred embodiment, the data processing device performs spatial registration of 3D laser point cloud data with the rebar design BIM model using an iterative nearest-point algorithm, comprising two stages: coarse registration and fine registration. The data processing device first performs voxel mesh downsampling on the acquired 3D laser point cloud data. The size of the voxel mesh is pre-set according to the point cloud density of the 3D laser point cloud data. During the downsampling process, the data processing device divides the 3D space into multiple voxel meshes of equal size. For all point clouds within each voxel mesh, the average coordinates of all point clouds within that mesh are taken as the representative point of that voxel mesh, ultimately generating the downsampled point cloud dataset.

[0039] The data processing device extracts key points from the downsampled point cloud dataset based on the normal vector and curvature features of the point cloud. Points with normal vector change rates and curvature values ​​exceeding preset thresholds are selected as initial key points, forming an initial key point set. The data processing device then performs coarse registration between the initial key point set and the standard node set in the reinforcement design BIM model. This coarse registration employs a rigid body transformation solution based on feature matching. By matching corresponding feature points in the initial key point set and the standard node set, the initial rotation and translation matrix is ​​obtained. This initial rotation and translation matrix includes an initial rotation matrix and an initial translation vector.

[0040] Based on the initial rotation and translation matrix, the data processing device performs a fine registration operation. Specifically, the data processing device applies the initial rotation and translation matrix to the initial keypoint set to obtain the keypoint set after initial transformation. It then calculates the nearest neighbor of each point in the initially transformed keypoint set to the standard node set, constructing corresponding matching point pairs. Based on the matching point pairs, the data processing device constructs an objective function that includes the point pair matching error. The objective function quantifies the overall matching error between the initially transformed keypoint set and the standard node set.

[0041] The objective function is calculated using the following formula:

[0042] in, Let R be the overall error objective function for matching point pairs; R is the rotation matrix in three-dimensional space; t is the translation vector in three-dimensional space; and n is the total number of matching point pairs. Let be the coordinate vector of the i-th key point in the initial key point set; For standard node concentration and The coordinate vector of the corresponding nearest neighbor; It is the Euclidean norm.

[0043] The data processing device uses singular value decomposition (SVD) to find the optimal rigid body transformation matrix that converges the objective function. During the solution process, the device first calculates the centralized coordinates of the matching point pairs, constructs a covariance matrix based on the centralized coordinates, and performs singular value decomposition on the covariance matrix to obtain the decomposed left singular matrix, singular value diagonal matrix, and right singular matrix. The optimal rotation matrix is ​​calculated based on the left and right singular matrices, and the optimal translation vector is calculated based on the optimal rotation matrix and the centralized coordinates. The optimal rotation matrix and the optimal translation vector constitute the optimal rigid body transformation matrix. The data processing device applies the obtained optimal rigid body transformation matrix to the complete 3D laser point cloud data to obtain transformed 3D laser point cloud data. Based on the transformed 3D laser point cloud data, the spatial coordinates of the actual rebar nodes are extracted. The spatial coordinate deviation of the actual rebar nodes is extracted based on the transformed coordinates and the design coordinates of the standard nodes.

[0044] In this embodiment, the parameter iteration convergence of the fine registration process of the iterative nearest point algorithm is shown in the following table: Table 2. Iterative Convergence Comparison Table of Rigid Body Transformation Matrix Parameters During ICP Registration

[0045] This table records the changes in rigid body transformation matrix parameters, objective function value, and mean matching error for each iteration during the iterative nearest-point algorithm for fine registration. In this embodiment, the iteration termination condition is that the change in the objective function value is less than a preset threshold. When the number of iterations reaches 7, the change in the objective function value has met the iteration termination condition, and the registration process converges. Through the parameter iteration process in this table, the solution process of the rigid body transformation matrix can be clearly defined, ensuring that the matching error between the registered point cloud data and the design BIM model is controlled within a preset range, reducing the interference of accumulated registration error on the extraction of node coordinate deviation.

[0046] Furthermore, during the construction of the stress topology map, the data processing device standardizes the edge weights to eliminate the influence of differences in the cross-sectional dimensions and material properties of the reinforcing bars. The data processing device receives stress monitoring data collected by a network of strain sensors deployed within the secondary lining concrete. This data consists of real-time strain values ​​collected by each strain sensor at a set sampling frequency. Based on the constitutive relationship between the concrete and the reinforcing bars, the data processing device converts the collected strain values ​​into stress values ​​at the corresponding monitoring points. The data processing device then uses an inverse distance weighted interpolation algorithm to map the stress values ​​at the monitoring points to all actual reinforcing bar nodes, obtaining the initial stress value for each actual reinforcing bar node.

[0047] For each edge in the rebar connection relationship, the data processing device obtains the initial stress values ​​of the two nodes connected by that edge, calculates the gradient difference between the initial stress values ​​of the two nodes, which is the absolute difference between the initial stress values ​​of the two nodes. The data processing device obtains the physical property parameters of the rebar corresponding to that edge, including the elastic modulus, cross-sectional area, and calculated length of the rebar, and calculates the stiffness coefficient of the rebar based on the physical property parameters. The data processing device multiplies the gradient difference of the initial stress values ​​of the nodes with the stiffness coefficient to obtain a standardized stress value that eliminates the influence of cross-sectional size differences. The standardized stress value is assigned to the corresponding edge as an edge weight, generating a stress topology map that integrates spatial location and mechanical distribution.

[0048] The standardized stress value is calculated using the following formula:

[0049] in, The normalized stress value is the edge connecting node i and node j. Let be the stiffness coefficient of the reinforcement connecting node i and node j; Let be the initial stress value of node i; Let be the initial stress value of node j; This is the absolute value operator.

[0050] In this embodiment, the graph neural network used by the control device is a graph convolutional neural network, and the feature extraction process includes multiple rounds of feature aggregation and feature mapping. The control device normalizes the spatial coordinate deviation of each vertex and the initial stress value of the node in the stress topology graph, and then concatenates them into a fixed-dimensional input feature vector. The edge weights of each edge are normalized and used as the feature parameters of the edges. The control device inputs the vertex input feature vectors and the edge feature parameters into the graph convolutional layers, and the number of graph convolutional layers is preset according to the size of the stress topology graph.

[0051] In the graph convolutional layer, the control device determines the set of neighboring nodes for each vertex based on the adjacency matrix of the stress topology graph. The adjacency matrix is ​​a square matrix corresponding to the number of vertices in the stress topology graph. An element with a value of 1 represents an edge connection between the corresponding two vertices, while an element with a value of 0 represents no edge connection between the corresponding two vertices. The control device updates the hidden state representation of the current node by aggregating the input feature vectors of the neighboring nodes defined by the adjacency matrix. The aggregation process adopts the aforementioned graph convolution operation rules.

[0052] The control device inputs the hidden state representation after propagation through multiple graph convolutional layers into a multilayer perceptron. The multilayer perceptron includes an input layer, hidden layers, and an output layer. The input layer receives the hidden state representation output by the graph convolutional layers, the hidden layers perform nonlinear transformations on the features, and the output layer outputs the predicted stress transfer contribution of each edge. During the training of the graph neural network, the aforementioned mean squared error function is used as the loss function. The trainable parameters of the graph convolutional layers and the multilayer perceptron are updated through the backpropagation algorithm until the loss function value converges to a preset threshold range.

[0053] In a preferred embodiment, please refer to the appendix. Figure 4 The control device introduces a spatial attention mechanism in the graph convolutional layer to dynamically adjust the feature aggregation weights of neighboring nodes. During the process of updating the hidden state representation of the current node by aggregating the input feature vectors of neighboring nodes defined by the adjacency matrix, the control device calculates the spatial distance vector and the reinforcement angle vector between the current node and any neighboring node. The spatial distance vector is a vector composed of the difference in three-dimensional coordinates between the current node and its neighboring nodes, and the reinforcement angle vector is a vector composed of the average angle between the axial vector of the reinforcement connecting the current node and its neighboring nodes and the axial vectors of the other adjacent reinforcements of the current node.

[0054] The control device concatenates the spatial distance vector and the angle vector between the reinforcing bars with the feature vectors of the current node and its neighbors, and inputs this concatenation into the attention weight calculation network. This network is a single-layer fully connected network that outputs attention coefficients reflecting the differences in spatial structure. The control device then multiplies these attention coefficients with the input feature vectors of the neighboring nodes and performs feature aggregation, ensuring that neighboring nodes that are closer in distance and have smaller angles contribute more to the hidden state representation of the current node.

[0055] The attention coefficient is calculated using the following formula:

[0056] in, Let be the attention coefficient of neighbor node u to the current node v in the stress topology graph; It is a normalized exponential function; is a linear rectified activation function with leakage; a is the trainable weight vector of the attention weight calculation network; W is the feature transformation matrix; This is the feature vector of the current node v; The feature vector of neighbor node u; This is a feature vector concatenation operation; This is the spatial distance vector between the current node v and its neighbor node u; Let be the angle vector between the current node v and its neighboring node u.

[0057] Furthermore, after the data processing device extracts the spatial coordinate deviation of the actual rebar nodes, during the cutting action of the cutting actuator, the data processing device continuously acquires updated real-time 3D laser point cloud data. The real-time 3D laser point cloud data is the point cloud data of the current state of the tunnel lining collected by the 3D laser scanning equipment at set time intervals during the cutting operation. For the missing point cloud areas in the real-time 3D laser point cloud data damaged by the cutting action, the data processing device performs virtual point cloud filling according to the corresponding structural completion rules in the rebar design BIM model.

[0058] The data processing device compares real-time 3D laser point cloud data with the transformed initial 3D laser point cloud data to identify areas where the number of point clouds is reduced and the point cloud density is below a preset threshold as missing point cloud areas, and obtains the 3D spatial range of these missing areas. The data processing device extracts the standard steel reinforcement structure model corresponding to the missing point cloud areas from the steel reinforcement design BIM model. Based on the geometric features of the standard steel reinforcement structure model, it generates virtual point cloud data with the same density as the initial 3D laser point cloud data and fills the missing point cloud areas with this virtual point cloud data. The data processing device then performs local registration and fine-tuning between the filled real-time 3D laser point cloud data and the 3D laser point cloud data of the currently uncut portion. This local registration and fine-tuning uses the same iterative nearest-point algorithm as the aforementioned fine registration, performing registration calculations only on the point cloud data of the uncut portion. Based on the fine-tuning results, the optimal rigid body transformation matrix is ​​updated, and the spatial coordinate deviation of subsequent nodes to be cut is corrected based on the updated optimal rigid body transformation matrix.

[0059] In this embodiment, a two-stage ICP registration process involving voxel mesh downsampling and coarse and fine registration is employed, combined with singular value decomposition to solve for the optimal rigid body transformation matrix. This reduces the interference of accumulated errors during registration on node coordinate deviation extraction. Standardized stress values ​​are obtained by multiplying the gradient difference of initial node stress values ​​by the stiffness coefficient, thus eliminating the influence of rebar cross-sectional size differences on topology graph construction. A spatial attention mechanism is introduced into the graph convolutional layer, calculating attention coefficients based on spatial distance vectors and rebar angle vectors, enhancing the graph neural network's ability to extract features from local spatial structural differences. Real-time point cloud completion and local registration fine-tuning during the cutting process correct the impact of cutting actions on subsequent node coordinate deviation extraction, improving the system's control accuracy during continuous cutting operations.

[0060] In a preferred embodiment, please refer to the appendix. Figure 5 The process by which the control device generates modular cutting boundary commands based on the spatial clustering results of adjacent stress release zero points includes three stages: spatial clustering, boundary fitting, and command conversion. Using all stress release zero points as initial cluster centers, the control device employs a density-based noise-based spatial clustering algorithm to divide the rebar point cloud in the 3D laser point cloud data into regions, obtaining multiple initial cutting blocks. The rebar point cloud is a subset of the point cloud corresponding to the rebar entities extracted through geometric feature identification from the 3D laser point cloud data.

[0061] During the execution of the density-based noise spatial clustering algorithm, the control device pre-sets two core parameters: the neighborhood radius and the minimum number of points within the neighborhood. For each stress release zero point, the number of rebar point clouds within its neighborhood radius is counted. When the number is greater than or equal to the minimum number of points within the neighborhood, the stress release zero point is marked as a core point. Based on the density reachability relationship of the core points, the control device divides the rebar point clouds corresponding to core points that have a density reachability relationship into the same cluster. Each cluster corresponds to an initial cutting block. Rebar point clouds not assigned to any cluster are marked as noise points. The rebars corresponding to noise points are non-stressed connected rebars and can be processed in the final stage of the cutting operation.

[0062] The core point is determined using the following formula:

[0063] in, Let be the core point determination function for point p. When the function is true, point p is the core point; D is the rebar point cloud dataset; q is any point in the rebar point cloud dataset. Let be the Euclidean distance between points p and q; The preset neighborhood radius; The minimum number of points within the preset neighborhood; This is a count operation for the number of elements in a set.

[0064] For each initial cutting block, the control device extracts the block edge point set, which is the set of rebar point clouds located at the boundary of the cluster corresponding to the initial cutting block. The control device projects the block edge point set onto a plane perpendicular to the longitudinal axis of the tunnel. During the projection process, the Y-axis and Z-axis coordinates of each point remain unchanged, and the X-axis coordinate is uniformly set to the mean X-axis coordinate of the initial cutting block, thus completing the two-dimensional plane projection.

[0065] On the projected 2D plane, the control device performs convex hull calculation and polygon fitting on the block edge point set. The convex hull calculation uses the Graham scan algorithm to obtain the minimum convex hull polygon containing all block edge point sets. Polygon fitting is then performed on the minimum convex hull polygon to remove redundant vertices, resulting in a smooth closed polygon boundary. The control device then inversely projects the 2D coordinates of the fitted polygon boundary into a 3D coordinate system. During the inverse projection, the X-axis coordinate of each vertex of the polygon boundary is restored to its original X-axis coordinate. Combined with the Y-axis and Z-axis coordinates during projection, a closed modular cutting boundary in 3D space is obtained. Based on the 3D coordinates of the modular cutting boundary and the cutting feed parameters, a modular cutting boundary command is generated.

[0066] In this embodiment, the correspondence between the boundary parameters of each modular cutting block and the cutting trajectory planning parameters is shown in the table below: Table 3 Comparison of Modular Cutting Block Boundary Parameters and Cutting Trajectory Planning

[0067] This table clearly defines the boundary parameters and corresponding cutting trajectory planning parameters for each modular cutting block. The boundary parameters are generated based on the results of spatial clustering and polygon fitting, while the cutting trajectory planning parameters are set based on the operational performance of the cutting actuator and the safety control requirements of the tunnel lining structure. Through the parameter correspondence in this table, the cutting operation time for each modular cutting block can be accurately estimated, providing data support for the progress control of continuous multi-block cutting operations, while ensuring that the cutting trajectory perfectly matches the modular cutting boundaries.

[0068] In this embodiment, please refer to the appendix. Figure 6 The process of the control device driving the cutting actuator according to the modular cutting boundary command includes four stages: trajectory discretization, inverse kinematics solution, velocity planning, and servo drive. The control device discretizes the three-dimensional spatial trajectory in the modular cutting boundary command. During the discretization process, a series of equidistant spatial navigation point sequences are generated along the extension direction of the three-dimensional spatial trajectory. The distance between adjacent spatial navigation points is preset according to the motion control accuracy of the cutting actuator.

[0069] For adjacent navigation points, the control device acquires the Cartesian space coordinates of the end effector tool at the current navigation point and the Cartesian space coordinates of the next navigation point, and calculates the Cartesian space displacement increment between the two navigation points. Based on the inverse kinematic model of the cutting actuator, the control device calculates the joint angle increment corresponding to the Cartesian space displacement increment. The inverse kinematic model is established based on the link parameters of the cutting actuator and reflects the mapping relationship between the Cartesian space displacement of the end effector and the joint angles.

[0070] The joint rotation increment is calculated using the following formula:

[0071] in, This is the rotation increment vector for each joint of the cutting actuator; To cut the Jacobian matrix of the actuator, the Moore-Penrose pseudoinverse is required; This is the displacement increment vector of the end-effector tool in Cartesian space.

[0072] The control device employs a cubic spline interpolation algorithm to smooth the calculated joint angle increments, generating speed planning curves for each joint motor. During the cubic spline interpolation process, the control device uses the node number corresponding to the spatial navigation point as the independent variable and the angle of each joint as the dependent variable to construct a piecewise cubic polynomial function. This ensures that the function values, first derivatives, and second derivatives of adjacent piecewise polynomial functions are continuous at the nodes, generating a smooth curve showing the joint angle changing over time. The first derivative of this joint angle curve is then calculated to obtain the speed planning curves for each joint motor.

[0073] The cubic spline interpolation function uses the following formula:

[0074] in, It is a cubic spline interpolation function; Let be the value of the independent variable corresponding to the i-th spatial navigation point; Let be the value of the independent variable corresponding to the (i+1)th spatial navigation point; These are the constant term, linear coefficient, quadratic coefficient, and cubic coefficient of the i-th cubic polynomial, respectively.

[0075] According to the speed planning curve, the control device sends servo drive pulses to each joint motor. The frequency and number of servo drive pulses correspond to the speed and displacement parameters in the speed planning curve. The control device controls each joint motor of the cutting actuator to run according to the speed planning curve, so that the end tool can smoothly transition between adjacent spatial navigation points. At the same time, the control device starts the cutting motor of the end tool to perform the cutting action, ensuring that the feed speed of the end tool is stable during the cutting process and that the cutting trajectory completely coincides with the modular cutting boundary.

[0076] In a preferred embodiment, the control device employs a density-based noise-based spatial clustering algorithm to divide the rebar point cloud in the 3D laser point cloud data into regions, while introducing an isolation mechanism for the heat-affected zone (HAZ). The control device acquires the historical cutting trajectory and historical operation duration of the cutting actuator. The historical cutting trajectory is the 3D spatial coordinate of the cutting path completed by the cutting actuator in the current dismantling operation, and the historical operation duration is the duration of the cutting operation corresponding to the historical cutting trajectory. Based on the historical cutting trajectory and historical operation duration, the control device calculates the predicted HAZ distribution range for each spatial location in the 3D laser point cloud data. The calculation of the predicted HAZ distribution range is based on a finite element analysis model of heat conduction during cutting. The input parameters are the coordinates of the historical cutting trajectory, the cutting operation duration, and the thermal conductivity coefficients of the concrete and rebar. The output is the temperature field distribution at each spatial location. Regions with temperatures exceeding a preset threshold are defined as the predicted HAZ distribution range.

[0077] When performing area division, the control device determines whether the line connecting two adjacent stress release zero points crosses the predicted heat-affected zone distribution range. If the line connecting two adjacent stress release zero points crosses the predicted heat-affected zone distribution range, the control device blocks the density connectivity between the two stress release zero points, uses the predicted heat-affected zone distribution range as an isolation zone, and redivides the initial cutting blocks to ensure that the boundary of each initial cutting block does not cross the predicted heat-affected zone distribution range.

[0078] Furthermore, after the control device re-divides the initial cutting blocks using the predicted heat-affected zone distribution range as an isolation zone, and the cutting execution mechanism completes the cutting and separation action of the first initial cutting block, the control device acquires the real-time temperature field data around the first initial cutting block collected by the infrared thermal imager. The real-time temperature field data includes the real-time temperature values ​​of each spatial location around the cutting area.

[0079] The control device compares the real-time temperature field data with the predicted temperature field data corresponding to the predicted distribution range of the heat-affected zone, calculates the error between the predicted and actual values, corrects the input parameters of the heat conduction finite element analysis model based on the error, corrects the predicted distribution range of the heat-affected zone, and obtains the dynamic actual heat-affected zone. Based on the dynamic actual heat-affected zone, the control device re-executes the blocking and re-division operation on the boundaries of the remaining uncut initial cutting blocks, adjusts the boundary control points of the remaining initial cutting blocks to ensure that the boundaries of the remaining blocks do not cross the dynamic actual heat-affected zone, and generates subsequent modular cutting boundary commands based on the re-divided blocks to drive the cutting execution mechanism to perform subsequent cutting operations.

[0080] In this embodiment, a density-based noise spatial clustering algorithm is used to divide the rebar point cloud into regions. A modular cutting boundary is generated by combining convex hull calculation and polygon fitting, ensuring a perfect match between the spatial distribution of the cutting boundary and the stress release zero points. A smooth velocity planning curve is generated using inverse kinematics and cubic spline interpolation algorithms, enabling smooth displacement and precise cutting of the cutting actuator's end-effector. The distribution range of the heat-affected zone is predicted by calculating historical cutting trajectories and operation durations, and the density connectivity between stress release zero points is blocked, eliminating interference from cutting heat on subsequent block division. The range of the heat-affected zone is corrected using real-time temperature field data, dynamically adjusting the cutting block boundaries and improving the system stability of continuous multi-block dismantling operations.

Claims

1. A modular dismantling and control system for tunnel secondary lining reinforcement based on multidimensional data, characterized in that, It includes a data processing device, a control device, and a cutting execution mechanism. The data processing device acquires three-dimensional laser point cloud data and a BIM model of the rebar design on the surface of the tunnel secondary lining. It uses an iterative nearest point algorithm to spatially register the three-dimensional laser point cloud data and the BIM model of the rebar design and extracts the spatial coordinate deviation of the actual rebar nodes. The data processing device combines the stress monitoring data of the secondary lining concrete to construct a stress topology graph with the actual steel bar nodes as vertices, the steel bar connection relationships as edges, and the node stress values ​​as edge weights. The control device uses a graph neural network to extract features from the stress topology graph, calculates the stress transmission contribution of each edge, and defines the steel bar connection node corresponding to the edge whose stress transmission contribution is lower than a preset threshold as the stress release zero point. The control device generates modular cutting boundary commands based on the spatial clustering results of adjacent stress release zero points, and drives the cutting execution mechanism to perform displacement and cutting actions along the modular cutting boundary according to the modular cutting boundary commands until the modular separation of the secondary lining steel bars is completed.

2. The modular dismantling and control system for tunnel secondary lining reinforcement based on multidimensional data according to claim 1, characterized in that, The process of spatially registering the three-dimensional laser point cloud data with the steel reinforcement design BIM model by the data processing device through the iterative nearest point algorithm includes: the data processing device performs voxel mesh downsampling on the three-dimensional laser point cloud data to extract an initial key point set, and performs coarse registration between the initial key point set and the standard node set in the steel reinforcement design BIM model to obtain an initial rotation and translation matrix. Based on the initial rotation and translation matrix, the data processing device calculates the nearest neighbor distance from each point in the initial key point set to the standard node set, and constructs an objective function that includes point pair matching error; The optimal rigid body transformation matrix that makes the objective function converge is obtained by using singular value decomposition. The optimal rigid body transformation matrix is ​​then applied to the three-dimensional laser point cloud data, and the spatial coordinate deviation of the actual steel bar node is extracted based on the transformed coordinates.

3. The modular dismantling and control system for tunnel secondary lining reinforcement based on multidimensional data according to claim 1, characterized in that, The process of the data processing device constructing the stress topology map includes: the data processing device receiving the stress monitoring data collected by the strain sensor network deployed inside the secondary lining concrete, and mapping the stress monitoring data to the actual steel bar nodes to obtain the initial stress value of the nodes; For each edge in the rebar connection relationship, the data processing device calculates the gradient difference of the initial stress values ​​of the nodes connecting the two ends, and multiplies the gradient difference by the stiffness coefficient of the rebar physical properties to obtain a standardized stress value that eliminates the influence of cross-sectional size differences. The data processing device assigns the standardized stress value to the corresponding edge as the edge weight, generating the stress topology map that integrates spatial location and mechanical distribution.

4. The modular dismantling and control system for tunnel secondary lining reinforcement based on multidimensional data according to claim 1, characterized in that, The process of the control device using a graph neural network to extract features from the stress topology graph includes: the control device inputting the spatial coordinate deviation of the vertex and the edge weight as input feature vectors into the graph convolutional layer; In the graph convolutional layer, the control device updates the hidden state representation of the current node by aggregating the input feature vectors of neighboring nodes defined by the adjacency matrix. The control device inputs the hidden layer state representation after propagation through multiple graph convolutional layers into a multilayer perceptron and outputs the predicted stress transfer contribution, wherein the loss function of the graph neural network is configured as the mean square error between the predicted stress transfer contribution and the actual contribution calculated by finite element simulation.

5. The modular dismantling and control system for tunnel secondary lining reinforcement based on multidimensional data according to claim 1, characterized in that, The process by which the control device generates modular cutting boundary instructions based on the spatial clustering results of adjacent stress release zero points includes: the control device uses the stress release zero point as the initial cluster center and employs a density-based noise application spatial clustering algorithm to divide the steel bar point cloud in the three-dimensional laser point cloud data into regions to obtain multiple initial cutting blocks. For each initial cut block, the control device extracts the block edge point set and projects the block edge point set onto a plane perpendicular to the tunnel longitudinal axis; The control device performs convex hull calculation and polygon fitting on the block edge point set on the plane, and back-projects the fitted polygon boundary onto the three-dimensional spatial coordinate system to generate the modular cutting boundary command.

6. The modular dismantling and control system for tunnel secondary lining reinforcement based on multidimensional data according to claim 1, characterized in that, The process by which the control device drives the cutting actuator according to the modular cutting boundary command includes: the control device discretizes the three-dimensional spatial trajectory in the modular cutting boundary command to generate a series of equidistant spatial navigation point sequences; For the adjacent spatial navigation points, the control device calculates the corresponding joint angle increment based on the inverse kinematic model of the cutting actuator; The control device uses a cubic spline interpolation algorithm to smooth the joint angle increment and generate speed planning curves for each joint motor. The control device sends servo drive pulses to each joint motor according to the speed planning curve, and controls the end tool of the cutting actuator to perform smooth transition displacement and cutting between adjacent spatial navigation points.

7. The modular dismantling and control system for tunnel secondary lining reinforcement based on multidimensional data according to claim 2, characterized in that, After the data processing device extracts the spatial coordinate deviation of the actual rebar node, the system further includes: the data processing device continuously acquires updated real-time three-dimensional laser point cloud data during the cutting action of the cutting actuator; For the missing point cloud areas in the real-time three-dimensional laser point cloud data that are destroyed by the cutting action, the data processing device performs virtual point cloud filling according to the corresponding structural completion rules in the steel reinforcement design BIM model. The data processing device performs local registration and fine-tuning of the real-time three-dimensional laser point cloud data after filling it in with the three-dimensional laser point cloud data of the currently uncut part, updates the optimal rigid body transformation matrix based on the fine-tuning result, and corrects the spatial coordinate deviation of the subsequent nodes to be cut based on the updated optimal rigid body transformation matrix.

8. The modular dismantling and control system for tunnel secondary lining reinforcement based on multidimensional data according to claim 4, characterized in that, In the process of the control device updating the hidden state representation of the current node by aggregating the input feature vectors of neighboring nodes defined by the adjacency matrix, the system further includes: the control device introducing a spatial attention mechanism in the graph convolutional layer to calculate the spatial distance vector and the angle vector between the current node and any neighboring node of the current node. The control device inputs the spatial distance vector and the angle vector between the steel bars into the attention weight calculation network and outputs attention coefficients that reflect the differences in spatial structure. The control device performs the aggregation operation after multiplying the attention coefficient with the input feature vector, so that the closer the neighboring node is and the smaller the angle between them, the greater the contribution weight of the hidden state representation of the current node.

9. The modular dismantling and control system for tunnel secondary lining reinforcement based on multidimensional data according to claim 5, characterized in that, During the process of the control device using a density-based noise spatial clustering algorithm to divide the steel bar point cloud in the three-dimensional laser point cloud data into regions, the system further includes: the control device acquiring the historical cutting trajectory and historical operation time of the cutting execution mechanism, and calculating the predicted heat-affected zone distribution range of each spatial location in the three-dimensional laser point cloud data based on the historical cutting trajectory and historical operation time; When performing region division, if the line connecting two adjacent stress release zero points passes through the predicted heat-affected zone distribution range, the control device blocks the density connectivity between the two stress release zero points and uses the predicted heat-affected zone distribution range as an isolation zone to redivide the initial cutting block.

10. The modular dismantling and control system for tunnel secondary lining reinforcement based on multidimensional data according to claim 9, characterized in that, After the control device redivides the initial cutting block as an isolation zone using the predicted heat-affected zone distribution range, the system further includes: after the cutting actuator completes the cutting and separation action on the first initial cutting block, the control device acquires real-time temperature field data around the first initial cutting block collected by the infrared thermal imager. The control device corrects the predicted thermally affected zone distribution range based on the real-time temperature field data to obtain the dynamic actual thermally affected zone. The control device re-executes the blocking and re-division operation on the boundary of the remaining uncut initial cut block based on the dynamic actual heat-affected zone, and generates subsequent modular cutting boundary instructions based on the re-divided block.

Images