Tunnel construction safety early warning method and system based on three-dimensional laser scanning

Abstract

This invention discloses a method and system for early warning of tunnel construction safety based on three-dimensional laser scanning. The method includes: acquiring multi-source data of the tunnel construction area; unifying the point cloud to a geodetic coordinate system; obtaining multi-dimensional features of each tunnel section; constructing a spatiotemporal dataset; training a spatiotemporal graph neural network model based on the spatiotemporal dataset; inputting real-time collected data into the trained spatiotemporal graph neural network model to output prediction results of the deformation risk of each section of the tunnel within a specified future time period and issue an alarm. Thus, by predicting future section settlement risk through the spatiotemporal graph neural network model, sufficient time is reserved for risk management; using three-dimensional laser scanning as the core and integrating environmental and construction data improves the accuracy of model prediction; and constructing an integrated system for acquisition, processing, modeling, early warning, and feedback improves management efficiency. The system will feed back the actual effect of each adjustment to the model for continuous learning and improvement.

Description

Technical Field

[0001] This invention relates to the field of tunnel construction technology, and in particular to a method and system for early warning of tunnel construction safety based on three-dimensional laser scanning. Background Technology

[0002] During tunnel construction, complex and variable geological conditions and harsh working environments make it highly susceptible to safety accidents such as rock deformation, collapse, and water inrush. Therefore, effectively predicting and providing early warnings of safety risks during tunnel construction is a crucial issue for ensuring the safe construction of tunnel projects.

[0003] Currently, tunnel construction safety monitoring and early warning technologies mainly include traditional manual monitoring, automated monitoring systems, and data analysis-based early warning methods. Among these, three-dimensional laser scanning technology has been widely used in tunnel construction monitoring due to its high precision, non-contact nature, and full coverage. CN104680579B discloses a tunnel construction information monitoring system based on three-dimensional scanning point clouds. This system uses three-dimensional laser scanning point cloud measurement technology to collect coordinate information and strength information of the tunnel's full-section spatial structure, performs automatic filtering and automatic point cloud data stitching, and generates a two-dimensional planar model and a three-dimensional spatial model of the tunnel. This enables real-time processing, updating, and three-level early warning forecasting of monitoring data and construction information.

[0004] With the development of deep learning technology, methods for assessing and predicting tunnel construction quality by combining 3D point cloud data with intelligent algorithms have gradually emerged. CN118864415A discloses a method for predicting the blasting quality of mechanized tunnel construction based on 3D point cloud and neural network. This method acquires tunnel point cloud data through a 3D laser scanner, constructs a 3D polygonal surface model after preprocessing and point processing, and builds an over-excavation and under-excavation prediction model based on drilling construction logs, blasting design schemes, etc., to predict the blasting quality of the tunnel.

[0005] In the field of tunnel deformation monitoring, CN120141338B proposes a three-dimensional laser point cloud detection method and system for the wide and thick lining zone of a tunnel in the water-bearing area. This method acquires tunnel point cloud data through three-dimensional laser scanning, establishes a high-fidelity tunnel twin model, simulates tunnel deformation through finite element method, and performs tunnel point cloud registration and deformation analysis by combining a neural network model, which significantly improves the accuracy of tunnel deformation analysis and processing.

[0006] To address the identification and early warning of risk areas in tunnel surrounding rock, CN120163430A proposes a safety classification and early warning method for risk areas in tunnel surrounding rock. This method uses three-dimensional laser scanning technology to acquire high-precision tunnel point cloud data, analyzes the spatial deformation distribution of the surrounding rock and its evolution over time, develops a potential risk area identification algorithm, and constructs a hierarchical safety early warning system.

[0007] In addition, CN120808580B discloses a method and system for intelligent early warning of tunnel composite disasters based on propensity digital twins. The system constructs a digital twin model based on the three-dimensional point cloud data and design parameters of the tunnel structure, uses a spatiotemporal graph convolutional network to model the spatiotemporal characteristics of the real-time collected monitoring data, and uses a Bayesian network to model the dependency relationship between different disaster factors and calculate the probability of disaster occurrence.

[0008] However, existing tunnel construction safety early warning technologies still have the following shortcomings:

[0009] Existing technologies largely rely on manual fixed-point monitoring or post-event review, lacking the ability to perceive and predict geological conditions at the tunnel face in real time, making it difficult to identify potential risks in advance. Although 3D laser scanning technology has been applied to tunnel monitoring, a comprehensive real-time monitoring system for the tunnel face and surrounding geological conditions has not yet been established.

[0010] Existing technologies fail to fully integrate high-precision geological data from 3D laser scanning with multi-source information such as construction parameters and environmental monitoring data, resulting in incomplete data coverage and an inability to support comprehensive advanced early warning modeling needs.

[0011] Existing technologies mostly employ traditional statistical methods or simple machine learning models, failing to fully consider the spatiotemporal correlation of tunnel deformation and the dynamic influence of geological conditions. This results in low accuracy of advance predictions and poor timeliness of early warnings. In particular, there is a lack of comprehensive models that can simultaneously learn the propagation patterns of tunnel cross-sectional characteristics in spatial structure and their evolutionary patterns over time.

[0012] In existing technologies, data acquisition, processing, modeling, and early warning processes are often fragmented and independent, failing to form an integrated system solution. This results in insufficient engineering practicality and makes it difficult to meet the actual needs of early warning for tunnel construction safety.

[0013] In summary, existing early warning methods for tunnel construction safety have at least the problems mentioned above. Summary of the Invention

[0014] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for early warning of tunnel construction safety based on three-dimensional laser scanning.

[0015] The objective of this invention is achieved through the following technical solution:

[0016] Firstly, this application discloses a method for early warning of tunnel construction safety based on three-dimensional laser scanning, comprising: S1, acquiring multi-source data of the tunnel construction area, including three-dimensional scanning data, construction parameters, and environmental monitoring data, wherein the three-dimensional scanning data is stored as a point cloud; S2, preprocessing the multi-source data; S3, performing coarse registration of the point cloud based on the coordinates of preset design control points in the tunnel, and then performing fine registration using a three-dimensional point cloud registration algorithm to unify the point cloud to the geodetic coordinate system; S4, extracting geometric features of the point cloud using a semantic segmentation model to obtain multi-dimensional features of each tunnel cross-section; S5, constructing a spatiotemporal dataset, including: using tunnel cross-sections as nodes, The multi-dimensional features of each node, along with the construction parameters and environmental monitoring data, are spliced ​​and fused to obtain the feature vector of the corresponding node. Furthermore, based on spatial proximity and geological similarity, a weighted graph structure representing the relationship between cross sections is established. The node features of the nodes in multiple consecutive time steps are combined with the corresponding weighted graph structure in chronological order to construct a spatiotemporal dataset. S6. A spatiotemporal graph neural network model is trained based on the spatiotemporal dataset. S7. The real-time collected data is input into the trained spatiotemporal graph neural network model to output the prediction results of the deformation risk of each cross section of the tunnel within a specified time period in the future, and a safety warning is issued based on the prediction results.

[0017] Furthermore, in S5, the construction of the spatiotemporal dataset includes: stacking the node features of multiple consecutive time steps into a three-dimensional feature tensor according to a preset time window length; classifying risk levels based on the ratio of the predicted convergence value to the preset allowable value, and labeling each node with a risk level label for a specified future time period for each time window, so as to construct a standardized spatiotemporal dataset containing feature tensors and label tensors.

[0018] Furthermore, the training process of the spatiotemporal graph neural network model includes: dividing the spatiotemporal dataset into a training set, a validation set, and a test set according to a certain ratio; training using a weighted multi-task loss function, wherein the loss function includes a regression loss term for predicting deformable values ​​and a classification loss term for predicting risk levels, wherein the classification loss term assigns different weights to samples of different risk levels, and the weight ratio of the regression loss term to the classification loss term in the total loss is a preset ratio.

[0019] Furthermore, the weighted graph structure representing the relationship between cross sections is established based on spatial proximity and geological similarity, wherein: the spatial proximity rule includes: connecting each node to multiple adjacent cross section nodes, and calculating the connection weight based on the mileage difference between nodes; the geological similarity rule includes: calculating the similarity of spatial geometric features between nodes, establishing a connection between nodes when the similarity exceeds a preset threshold, and using the similarity as the connection weight.

[0020] Furthermore, in S6, the spatiotemporal graph neural network model includes at least one spatiotemporal block, each spatiotemporal block including: a spatial convolutional layer, a temporal gating layer, and a multi-scale temporal convolutional layer; and each spatiotemporal block sets a residual path, achieves feature dimension matching through a 1×1 convolutional kernel, and completes node feature update after activation by the ReLU activation function.

[0021] Furthermore, the spatial convolutional layer performs the following: a linear transformation on the original node features of node i to obtain the transformed node features. And the node features obtained after transforming the original node features of neighbor node j. Based on the node characteristics of node i Node characteristics of neighbor node j The attention coefficient between node i and its neighbor node j is calculated. :

[0022] ;

[0023] Where a is a 128-dimensional attention vector, Indicates transpose. The representation features are concatenated; attention weights are obtained through normalization. :

[0024]

[0025] in, Let exp() represent the attention coefficients between node i and any neighboring node k, and let exp() denote the exponential function. Then, for node i's node features... renew:

[0026] .

[0027] Furthermore, a fixed attention bias weight is applied to the nodes corresponding to the vault region.

[0028] Furthermore, the multi-scale temporal convolutional layer employs parallel dilated convolutional paths with different dilation rates to simultaneously extract short-term, medium-term, and long-term temporal features of node features.

[0029] Furthermore, in S3, the coarse registration of the point cloud based on the coordinates of the preset design control points in the tunnel includes: obtaining the geodetic coordinates of at least three preset control points and their scan coordinates in the point cloud; and, based on the geodetic coordinates and scan coordinates, initially aligning the point cloud to the geodetic coordinate system by solving the rigid transformation matrix.

[0030] Secondly, this application discloses a tunnel construction safety early warning system based on three-dimensional laser scanning, including a processor and a storage device; the processor performs read and write operations on the storage device, and the storage device stores a computer program for implementing the tunnel construction safety early warning method based on three-dimensional laser scanning.

[0031] The beneficial effects of this invention are:

[0032] The spatiotemporal neural network model predicts future cross-sectional settlement risks, allowing sufficient time for risk management; 3D laser scanning is used as the core, integrating environmental and construction data to improve the accuracy of model predictions; an integrated system for data acquisition, processing, modeling, early warning, and feedback is constructed to improve management efficiency. The system will feed back the actual effects of each adjustment to the model for continuous learning and improvement. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the process of a tunnel construction safety early warning method based on three-dimensional laser scanning according to an embodiment of this application. Detailed Implementation

[0034] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] refer to Figure 1 This application discloses a method and system for early warning of tunnel construction safety based on three-dimensional laser scanning.

[0036] The tunnel construction safety early warning method based on three-dimensional laser scanning according to the embodiments of this application includes the following steps:

[0037] S1. Multi-source data acquisition:

[0038] Point cloud data of the working face is acquired by a 3D laser scanner. The constructed section is scanned using a mobile SLAM (Simultaneous Localization and Mapping) system, which automatically generates point clouds of cross-sections with 1-meter spacing and convergence value calculation accuracy ≤ ±0.5mm. Construction parameters such as construction mileage, anchor spacing and length, and shotcrete thickness are obtained. Environmental monitoring data such as water level are collected in real time by a groundwater pressure sensor.

[0039] Specifically, the multi-source data includes the following operations: geological data is obtained by acquiring information such as lithology, joints, and faults at the tunnel face through a 3D laser scanner; deformation data is obtained by scanning the constructed section at a fixed time of 10:00 every day through a mobile SLAM system, and cross-sectional point clouds are extracted at 1-meter intervals along the tunnel axis with a convergence value calculation accuracy of ≤±0.5mm; environmental data is collected in real time through a pre-embedded groundwater pressure sensor.

[0040] S2. Preprocessing of multi-source data:

[0041] The collected point cloud data is processed by denoising, downsampling, and invalid region removal. The basic deformation value is calculated and the basic geological features are manually labeled.

[0042] Specifically, the 3D scan data is denoised using methods such as filtering from the PCL library (point cloud library). The number of neighboring points is set to 50, and the standard deviation is set to 3 to remove isolated noise that deviates from the overall point cloud distribution. This 3-fold standard deviation setting maximizes the balance between denoising effectiveness and feature preservation. Then, the point cloud data is labeled with its basic geological features, and its lithology and the initial number of joint groups are recorded as a geological feature training set.

[0043] S3, Point cloud coordinates are unified:

[0044] The point cloud is coarsely registered based on the coordinates of the pre-set design control points in the tunnel, and then finely registered using the ICP algorithm (three-dimensional point cloud registration algorithm) to unify the point clouds from multiple devices and time periods into the geodetic coordinate system.

[0045] Specifically, the coarse registration of the point cloud based on the coordinates of the pre-set design control points in the tunnel includes: obtaining precise geodetic coordinates (X1,Y1,Z1), (X2,Y2,Z2), and (X3,Y3,Z3) through the pre-set control points in the tunnel and connecting them with the real-time scan coordinates using a total station; matching the pre-set geodetic coordinates of the three control points with their real-time scan coordinates; and substituting them into the rigid transformation equation.

[0046]

[0047] In the formula, R is the rotation matrix and T is the translation vector. For geodetic coordinates, The coordinates are scanned in real time; then, the transformation matrix is ​​solved using the SVD algorithm. This achieves coarse alignment between two point clouds by mapping the coordinates of all points in the source point cloud through the transformation matrix, thus completing coarse registration.

[0048] Then, the ICP algorithm (3D point cloud registration algorithm) is used for fine registration, with a search radius of 5cm and the iteration termination condition being RMSE≤3mm (i.e., root mean square error≤3mm) or the number of iterations≤50.

[0049] S4. Deep Learning Feature Extraction:

[0050] The PointNet++ semantic segmentation model (a deep learning model for processing 3D point cloud data) is trained based on the geological feature training set to extract the geometric features of the point cloud. The model uses the cross-entropy loss function and the Adam optimization algorithm, and iterates for 100 rounds to extract multi-dimensional features of each tunnel section from the 3D scan data (point cloud set).

[0051] For example, the PointNet++ model uses the point cloud of the working face and manually annotated geological features as the training set, divides the training set and the validation set in a 7:3 ratio, and iterates for 50 rounds using the cross-entropy loss function and the Adam optimization algorithm (an adaptive optimization method). Training stops when the recognition accuracy of the validation set is ≥85%, and geological parameters such as joint dip angle and rock mass integrity coefficient are output as the multidimensional features.

[0052] S5. Spatiotemporal dataset construction:

[0053] Using the tunnel cross-section as the basic node, the multi-dimensional features of each node, along with the construction parameters and environmental monitoring data, are spliced ​​and integrated.

[0054] Specifically, each node first stores the data from the construction parameters in S1, such as basic attributes like mileage, geodetic coordinates, scanning timestamps, and construction stages. Then, it integrates the multi-dimensional features output by the PointNet++ model, such as 6-dimensional spatial geometric features like joint dip angle, rock mass integrity coefficient, and cross-sectional area, the deformation rate calculated by the mobile SLAM system (simultaneous localization and mapping system), and environmental parameters like groundwater pressure and anchor density. Finally, it forms the node's feature vector by splicing through channels and normalizes it to the [0,1] interval to eliminate the influence of dimensions.

[0055] Then, based on spatial proximity and geological similarity, a weighted graph structure representing the relationship between cross sections is established. The spatial proximity rule includes connecting each node to multiple adjacent cross section nodes and calculating the connection weight based on the mileage difference between nodes. The geological similarity rule includes calculating the similarity of spatial geometric features between nodes. When the similarity exceeds a preset threshold, a connection between nodes is established, and the similarity is used as the connection weight.

[0056] Specifically, the spatial proximity rule connects each node to nodes of three adjacent cross sections, with the weight calculated as "1 / the difference in mileage between the two nodes"; the geological similarity rule establishes a connection when the cosine similarity of the spatial geometric features between nodes is ≥0.7, and the weight value is equal to the cosine similarity. All edges are organized into an edge index matrix and a corresponding edge weight vector, and stored in a sparse matrix format.

[0057] In some embodiments, the node features of multiple consecutive time steps are stacked into a three-dimensional feature tensor according to a preset time window length. Risk levels are classified based on the ratio of the predicted convergence value to a preset allowable value, and risk level labels are assigned to each node within a specified future time period for each time window, thereby constructing a standardized spatiotemporal dataset containing feature tensors and label tensors. For example, node features of consecutive time steps are stacked into a [10×N×6] feature tensor in 10-day windows, where N is the number of valid nodes, ensuring that the features of each time step are consistent with the scan data of the corresponding date. Risk levels are classified according to "predicted convergence value / design allowable value," with values ​​<30% indicating low risk, 30%-70% indicating medium risk, and >70% indicating high risk. A [7×N×1] label tensor is assigned to each 10-day feature window corresponding to the next 7 days, constructing a standardized spatiotemporal dataset that meets the input requirements of the spatiotemporal graph neural network model.

[0058] S6. Construct a spatiotemporal graph neural network model:

[0059] The spatiotemporal graph neural network model consists of three spatiotemporal blocks connected in sequence. Each spatiotemporal block contains a spatial convolutional layer corresponding to GAT (Graph Attention Network), a temporal gating layer corresponding to GRU (Gated Recurrent Unit), and a multi-scale temporal convolutional layer corresponding to TCN (Temporal Convolutional Network).

[0060] The spatial convolutional layer employs a 4-head attention mechanism, first performing a linear transformation on the features of each node. Where W is a 64-dimensional learnable weight matrix, The original node features of the i-th tunnel section node are transformed to obtain the transformed node features. Calculate the attention coefficient between node i and its neighbor node j. In the formula, is the node feature obtained after linear transformation of the neighbor node j; the specific process will not be repeated. a is a 128-dimensional attention vector. This represents the transpose of 'a'. Indicates feature splicing, It is the activation function. Then, the attention weights are obtained by normalization using the softmax function. ,in Let be the attention coefficient between the target node i and any neighboring node k, and exp() denotes the exponential function; in some preferred embodiments, an additional fixed attention bias of 0.2 is added to the dome node.

[0061] Then, the node features of node i Update to enhance key region features:

[0062] , This is the activation function.

[0063] The time-gating layer uses a 64-dimensional hidden state, which is updated by the gate. : Control the proportion of historical information retained, among which, It is a parameter matrix. It is a bias vector. It is the hidden state of the previous time step. This is the current input. It's an activation function, a reset gate. : Controlling the degree of new information reception, among which, and It resets the parameter matrix and bias vector of the gate; calculates the candidate states. ,in, This indicates the hidden state of the previous time step. With Reset Door By combining these factors to control the extent of their influence, the current state is ultimately obtained. It effectively captures the temporal dependence and mutation characteristics of deformation.

[0064] The multi-scale temporal convolutional layer employs three parallel dilated convolutional paths, each with a kernel size of 3 and dilation rates of 1, 2, and 4, corresponding to receptive fields of 3, 7, and 15, to achieve simultaneous extraction of short-term, medium-term, and long-term temporal features. Causal padding (e.g., padding length = (kernel size - 1) × dilation rate) ensures temporal causality and prevents future data leakage.

[0065] In some preferred embodiments, a residual path is set for each spatiotemporal block, feature dimension matching is achieved through a 1×1 convolution kernel, and node features are updated after ReLU activation, effectively alleviating the gradient vanishing problem.

[0066] The spatiotemporal graph neural network model is trained using the spatiotemporal dataset constructed above as input, and the dataset is divided into training, validation, and test sets in a 7:2:1 ratio. The model employs a weighted multi-task loss function: Huber loss is used for deformable value regression, and weighted cross-entropy is used for risk level classification. The weights for high-risk samples are set to 1.2, medium-risk samples to 0.5, and low-risk samples to 0.3.

[0067] For example, to address the training bias issue caused by the limited number of high-risk samples, the loss weight ratio for the two tasks is set to 0.6:0.4, with a total loss L = 0.6 × L. regression + 0.4×L classification ; among which, L regression For regression loss, L classification For classification loss.

[0068] The training process employs data augmentation strategies, including time-series interpolation, point cloud rotation of ±5°, and ±5% random noise. The number of training iterations is set to 300, the learning rate is set to 0.001, and an early stopping mechanism is incorporated to prevent overfitting and save training time.

[0069] Understandably, three spatiotemporal blocks are sequentially connected to form a complete spatiotemporal graph neural network structure. The output of each spatiotemporal block serves as the input to the next, enabling layer-by-layer feature extraction and transformation. Through the stacking of multiple spatiotemporal blocks, the model can progressively extract high-order spatiotemporal feature representations from the original input data, enhancing its ability to model complex spatiotemporal dynamic processes. Thus, by combining GAT spatial convolutional layers with dilated convolutional layers featuring causal padding, the model effectively captures spatial relationships and temporal evolution patterns in graph-structured data, simultaneously handling spatial dependencies between nodes and temporal evolution patterns of node features, making it suitable for various spatiotemporal prediction tasks.

[0070] S7. Tunnel Construction Safety Warning:

[0071] Real-time collected and preprocessed 3D scanning data, construction parameters, and environmental monitoring data are input into a trained spatiotemporal neural network model. The model simultaneously learns the propagation law of tunnel cross-sectional features in the spatial graph structure and the evolution law in the time series, and outputs the prediction results of the deformation risk of each cross-section of the tunnel in a specified time period in the future. Based on the prediction results, safety warnings are issued to realize the early prediction and prevention of risks in the tunnel construction process.

[0072] The tunnel construction safety early warning system based on three-dimensional laser scanning according to an embodiment of this application includes a processor and a memory; the processor performs read and write operations on the memory, and the memory stores a computer program for implementing the tunnel construction safety early warning method based on three-dimensional laser scanning.

[0073] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for early warning of tunnel construction safety based on three-dimensional laser scanning, characterized in that, include: S1. Acquire multi-source data of the tunnel construction area, including three-dimensional scanning data, construction parameters and environmental monitoring data, wherein the three-dimensional scanning data is stored as a point cloud; S2. Preprocess the multi-source data; S3. Based on the coordinates of the preset design control points in the tunnel, the point cloud is coarsely registered, and then finely registered using a three-dimensional point cloud registration algorithm to unify the point cloud to the geodetic coordinate system. S4. Extract point cloud geometric features using a semantic segmentation model to obtain multi-dimensional features of each tunnel section; S5. Construct a spatiotemporal dataset, including: Using tunnel cross-sections as nodes, the multi-dimensional features of each node, along with construction parameters and environmental monitoring data, are spliced ​​and fused to obtain the feature vector of the corresponding node. Furthermore, based on spatial proximity and geological similarity, a weighted graph structure representing the relationships between cross-sections is established. The node features of the nodes at multiple consecutive time steps are combined with the corresponding weighted graph structure in chronological order to construct a spatiotemporal dataset. The spatial proximity rule includes connecting each node to multiple adjacent cross-section nodes and calculating the connection weight based on the mileage difference between nodes. The geological similarity rule includes calculating the similarity of spatial geometric features between nodes; when the similarity exceeds a preset threshold, a connection is established between nodes, and the similarity is used as the connection weight. S6. Train a spatiotemporal graph neural network model based on the spatiotemporal dataset; S7. Input the real-time collected data into the trained spatiotemporal neural network model to output the prediction results of the deformation risk of each section of the tunnel within a specified time period in the future, and issue a safety warning based on the prediction results.

2. The method for early warning of tunnel construction safety based on three-dimensional laser scanning according to claim 1, characterized in that, In S5, the construction of the spatiotemporal dataset includes: According to the preset time window length, the node features of multiple consecutive time steps are stacked into a three-dimensional feature tensor; the risk level is divided based on the ratio of the predicted convergence value to the preset allowable value, and the risk level label of each node in the future specified time period is labeled for each time window, so as to construct a standardized spatiotemporal dataset containing feature tensors and label tensors.

3. The method for early warning of tunnel construction safety based on three-dimensional laser scanning according to claim 2, characterized in that, The training process of the spatiotemporal graph neural network model includes: The spatiotemporal dataset is divided into a training set, a validation set, and a test set according to a certain ratio; The training is performed using a weighted multi-task loss function, which includes a regression loss term for predicting deformable values ​​and a classification loss term for predicting risk levels. The classification loss term assigns different weights to samples with different risk levels, and the weight ratio of the regression loss term to the classification loss term in the total loss is a preset ratio.

4. The method for early warning of tunnel construction safety based on three-dimensional laser scanning according to claim 1, characterized in that, The weighted graph structure representing the relationship between cross sections is established based on spatial proximity and geological similarity, wherein: The spatial proximity rule includes: connecting each node to multiple adjacent cross-section nodes, and calculating the connection weight based on the mileage difference between nodes; The geological similarity-based rules include: calculating the similarity of spatial geometric features between nodes; establishing connections between nodes when the similarity exceeds a preset threshold; and using the similarity as the connection weight.

5. The method for early warning of tunnel construction safety based on three-dimensional laser scanning according to claim 1, characterized in that, In S6, the spatiotemporal graph neural network model includes at least one spatiotemporal block, and each spatiotemporal block includes: a spatial convolutional layer, a temporal gating layer, and a multi-scale temporal convolutional layer; Furthermore, a residual path is set for each spatiotemporal block, feature dimension matching is achieved through a 1×1 convolution kernel, and node features are updated after activation by the ReLU activation function.

6. The method for early warning of tunnel construction safety based on three-dimensional laser scanning according to claim 5, characterized in that, The spatial convolutional layer performs: A linear transformation is performed on the original node features of node i corresponding to the i-th tunnel section to obtain the transformed node features. And the node features obtained after transforming the original node features of neighbor node j. ; Based on the node characteristics of node i Node characteristics of neighbor node j The attention coefficient between node i and its neighbor node j is calculated. : ； Where a is a 128-dimensional attention vector, This represents the transpose of 'a'. Indicates feature splicing, For activation functions; Attention weights are obtained through normalization. : in, Let be the attention coefficient between node i and any neighbor node k, and exp() denotes the exponential function; Then, the node features of node i Update: ； In the formula This is the activation function.

7. The method for early warning of tunnel construction safety based on three-dimensional laser scanning according to claim 6, characterized in that, A fixed attention bias weight is applied to the nodes corresponding to the vault region.

8. The method for early warning of tunnel construction safety based on three-dimensional laser scanning according to claim 5, characterized in that, The multi-scale temporal convolutional layer employs parallel dilated convolutional paths with different dilation rates to simultaneously extract short-term, medium-term, and long-term temporal features of node features.

9. The method for early warning of tunnel construction safety based on three-dimensional laser scanning according to claim 1, characterized in that, In S3, the coarse registration of the point cloud based on the preset design control point coordinates in the tunnel includes: Obtain the geodetic coordinates of at least three preset control points and their scan coordinates in the point cloud; Based on the geodetic coordinates and scan coordinates, the point cloud is initially aligned to the geodetic coordinate system by solving the rigid transformation matrix.

10. A tunnel construction safety early warning system based on three-dimensional laser scanning, characterized in that, Includes processor and storage; The processor performs read and write operations on the memory, which stores a computer program for implementing the tunnel construction safety early warning method based on three-dimensional laser scanning as described in any one of claims 1-9.

Images