Tunnel back break real-time detection method and system based on edge calculation enabling

Abstract

The invention relates to an edge calculation enabling tunnel back break real-time detection method and system, and belongs to the technical field of tunnel construction intelligent monitoring and quality control, and the method comprises the following steps: 1, collecting data in real time through array multi-mode sensors deployed on a tunnel excavation surface; 2, dynamically constructing a local BIM model at an edge computing node, and carrying out noise reduction processing on the point cloud data by adopting an adaptive voxel filtering algorithm; step 3, carrying out nonlinear registration on the real-time point cloud and the design model based on an improved ICP (Inductively Coupled Plasma) algorithm; 4, extracting back-break features through a lightweight convolutional neural network, and dynamically updating network parameters by adopting a transfer learning framework; step 5, generating an augmented reality visualization report including the back break amount, the spatial distribution and the risk level on the edge side, and transmitting the augmented reality visualization report back to the control center through the low-delay communication module; the method has the beneficial effects that the real-time point cloud and the design model are subjected to nonlinear registration, and the requirement of tunneling deviation is monitored in real time.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring and quality control technology for tunnel construction, and specifically relates to a real-time detection method and system for tunnel over-excavation and under-excavation enabled by edge computing. Background Technology

[0002] In tunnel construction, over-excavation and under-excavation have a significant impact on the project. Over-excavation increases the amount of concrete spraying and backfilling, significantly increasing project costs and extending the construction period; under-excavation requires secondary treatment, affecting support quality, construction progress, and structural safety. Therefore, rapid and accurate detection of over-excavation and under-excavation is a core aspect of tunnel construction quality control.

[0003] Current detection technologies have many limitations:

[0004] Manual measurement methods are inefficient, especially in long tunnels, with sparse detection points and serious data lag, and also pose personal safety hazards.

[0005] Three-dimensional laser scanning equipment is expensive and bulky, data processing is complex and time-consuming, real-time performance is poor, and it is highly dependent on the on-site network.

[0006] Photogrammetry accuracy is greatly affected by lighting and texture, and data processing is also complex and slow, making it difficult to meet the requirements of high precision and real-time operation.

[0007] The core bottlenecks of these existing technologies are mainly reflected in: poor real-time performance, delayed detection results, and inability to provide effective guidance before the next cycle; low efficiency, with the detection process being time-consuming and affecting the construction progress; low safety, requiring personnel to frequently enter the risk area immediately after blasting; high cost, with high costs for purchasing and maintaining high-end equipment; and strong dependence on the network, while the tunnel construction environment is difficult to meet the stable, high-bandwidth network conditions required for large data transmission.

[0008] Edge computing pushes computing, storage, and analysis capabilities down to the network edge, closer to the data source. It features low latency, high bandwidth utilization, local data processing, enhanced privacy, and reliability, making it a perfect fit for the poor network conditions and high real-time requirements of tunnel construction sites.

[0009] Therefore, there is an urgent need for a detection method and system that can quickly acquire three-dimensional information of the tunnel face immediately after tunnel blasting, without requiring personnel to enter the risk area, process and calculate the accurate over- and under-excavation amounts locally in real time, and provide timely feedback to guide construction. Summary of the Invention

[0010] This invention provides a real-time detection method and system for tunnel over- and under-excavation enabled by edge computing. It addresses the problem of the inability to monitor over- and under-excavation in tunnel construction in real time in existing technologies. Based on an improved ICP algorithm, it performs nonlinear registration between real-time point clouds and design models. The registration of point clouds near the tunnel face and key structures is more accurate. The regularization term makes the registration unaffected by noise and complex structures, and is related to construction accuracy. Combined with dynamic weighting of the tunnel boring machine's posture, the registration can be aligned in real time with the construction progress, guiding the tunnel boring machine to adjust its posture and position. It directly serves the need for real-time comparison of design models and monitoring of tunneling deviations in tunnel construction.

[0011] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0012] A real-time detection method for tunnel over-excavation and under-excavation enabled by edge computing includes the following steps:

[0013] Step 1: Real-time acquisition of 3D point cloud data, inertial measurement data, and ambient light intensity data using multimodal sensors deployed on an array on the tunnel excavation surface;

[0014] Step 2: Dynamically construct a local BIM model at the edge computing node, and use an adaptive voxel filtering algorithm to denoise the point cloud data;

[0015] Step 3: Based on the improved ICP algorithm, perform nonlinear registration between the real-time point cloud and the design model; wherein, the nonlinear registration weight is dynamically adjusted by the tunneling machine pose data;

[0016] Step 4: Extract over-mining and under-mining features using a lightweight convolutional neural network, and dynamically update the network parameters using a transfer learning framework;

[0017] Step 5: Generate an augmented reality visualization report on the edge side that includes over- and under-excavation amounts, spatial distribution, and risk levels, and send it back to the control center via a low-latency communication module.

[0018] Optionally, in step 1, the spatial morphology of the excavation face can be accurately restored by the three-dimensional point cloud data. The specific acquisition principle is as follows: a laser beam is emitted or a three-dimensional image is acquired by a stereo vision sensor device. The three-dimensional coordinates (X, Y, Z) of each point on the excavation face are calculated based on the time difference and phase difference of the reflected signal. Finally, a dense set of point clouds is formed, which outlines the contour of the excavation face, the distribution of rock layers, protrusions and depression details like three-dimensional pixels.

[0019] Inertial measurement data can capture the motion state of the excavation face and equipment. Through the inertial measurement unit, the acceleration and angular velocity of the sensor itself can be sensed in real time, and the position, velocity and attitude information can be further calculated.

[0020] Ambient light intensity data can help determine the environmental condition of the excavation face by measuring the light intensity around the excavation face in real time using a light sensor.

[0021] Optionally, in step 2, the dynamic construction of the local BIM model can be performed using the following steps:

[0022] Step a: Point cloud segmentation and feature extraction;

[0023] Step b: Geometric model generation;

[0024] Step c: Dynamic fusion of local and global models;

[0025] Step d: Model lightweighting and visualization.

[0026] Optionally, in step 2, the noise reduction processing of the point cloud data by the adaptive voxel filtering algorithm is performed as voxel-based noise filtering:

[0027] To count the number of points within each voxel, let voxel V be... a，b，c The number of points inside is n a,b,c ;

[0028] Dynamic threshold calculation determines the retention threshold T based on the local voxel number distribution. i :

[0029] in, σ is the average number of neighborhood voxels. n,i Let n be the standard deviation of the number of neighboring voxels, and α and β be the weighting coefficients; the filtering method is as follows: if n a，b，c ≥T i Preserve the points within the voxel and replace all points with the voxel's center point: Where j is the point index within the voxel; p′ a，b，c Let n be the center point of voxels (a, b, c). a，b，c x represents the number of original points within the voxel (a, b, c); j y j and z j These are the three-dimensional coordinates of the j-th primitive point within the voxel, where j ranges from 1 to n. a，b，c traverse all points within the voxel; ∑x j ,∑y j and ∑z j Let x, y, and z be the sum of the x, y, and z coordinates of all points within the voxel; if n a，b，c <T i Mark as noise voxels and delete all points.

[0030] Optionally, in step 3, the improved ICP algorithm performs nonlinear registration between the real-time point cloud and the design model, introducing weights and considering nonlinear errors, with the objective function being:

[0031]

[0032] Where, min T To find the minimum value of the transformation matrix T, we need to find a set of rotations R and translations t, where T = [R|t], R is the rotation matrix, and t is the translation vector.

[0033] To traverse all points in the source point cloud, there are n points in total; q j In the target point cloud, the source point p i The nearest neighbor of the corresponding point after transformation; T·p i For source point p i Coordinates after transformation by transformation matrix T;

[0034] ||q j -T·p i || 2 The squared Euclidean distance between the source and target points after transformation measures the registration error. The smaller the value, the more accurate the alignment. ||·|| is the Euclidean distance norm.

[0035] λ·R(T) is a regularization term that avoids overfitting noise during registration and ensures smooth transformation.

[0036] Optionally, λ is the regularization intensity hyperparameter, where λ ≥ regularization coefficient, balancing point cloud alignment error and transformation smoothness; the larger λ is, the more emphasis is placed on transformation smoothness, which is suitable for scenarios where the tunneling machine moves continuously, avoiding abrupt changes in transformation during iteration;

[0037] R(T) is a regularization function that constrains the model parameter T. R(T) constrains the nonlinear regularization term of the transformation matrix, limiting the transformation difference between adjacent iterations and ensuring that the tunneling machine's pose changes conform to engineering reality.

[0038] Optionally, in step 4, the over-mining and under-mining features extracted for the lightweight convolutional neural network are calculated using the following convolution formula:

[0039] Among them, y i，j After the convolution operation, the pixel value at position (i, j) in the output feature map is generated. The double summation symbol corresponds to the size of the convolution kernel; w k，l x represents the weight parameters at position (k, l) in the convolution kernel; i+k，j+l is the pixel value at position (i+k,j+l) in the input feature map, which is the local region covered by the convolution kernel as it slides across the input; b is the bias term, which provides a learnable offset for the convolution operation.

[0040] Optionally, in step 4, the network parameters are dynamically updated for the transfer learning framework as follows:

[0041] Using the incremental dataset D new The network is updated using mini-batch gradient descent, with the learning rate set to 1 / 10 of the pre-training rate. The parameter update formula is as follows:

[0042]

[0043] Where, θ t+1 The updated parameters for the next iteration; θ t The parameters are the model parameters for the current iteration; η is the learning rate, which controls the step size for parameter updates. L(θ) represents the gradient of the loss function L with respect to the parameter θ, reflecting the direction and magnitude of the parameter's influence on the loss; t ;D new ) is the loss function, which measures the current parameter θ. t Predicting new data D new Error at time; D new This refers to newly collected training data or incremental samples.

[0044] Optionally, in step 5, spatial distribution modeling is performed; based on the preprocessed data, a three-dimensional spatial model of the construction area is constructed, and the over-excavation and under-excavation amounts are bound to specific locations to form spatial distribution data with attributes;

[0045] Define risk level classification rules and label the risk level of each over- or under-excavation area;

[0046] AR visualization report generation: The edge-side AR engine calls the 3D model and over- and under-mining data to overlay the over- and under-mining amount, spatial distribution, and risk level onto the real scene in real time to generate an AR visualization report;

[0047] Low-latency data is transmitted back to the control center; at the edge, key data from the AR visualization report is compressed and transmitted via a low-latency communication module.

[0048] An edge computing-enabled real-time detection system for tunnel over-excavation and under-excavation includes:

[0049] The multimodal sensing module is an array of sensors deployed on the tunnel excavation face;

[0050] The noise reduction module, connected to the multimodal sensing module, is used to dynamically construct local BIM models at edge computing nodes and uses an adaptive voxel filtering algorithm to reduce noise in point cloud data.

[0051] The data processing module, connected to the noise reduction module, is used to perform nonlinear registration of real-time point clouds with the design model based on the improved ICP algorithm.

[0052] The data update module, connected to the data processing module, is used to extract over-mining and under-mining features using a lightweight convolutional neural network and dynamically update network parameters using a transfer learning framework.

[0053] The data control module, connected to the data update module, generates an augmented reality visualization report at the edge, including over- and under-mining amounts, spatial distribution, and risk levels, and transmits it back to the control center via a low-latency communication module.

[0054] The beneficial effects of this invention are:

[0055] 1. This invention is based on an improved ICP algorithm to perform nonlinear registration between real-time point clouds and design models. The registration of point clouds near the tunnel face and key structures is more accurate. The regularization term makes the registration unaffected by noise and complex structures, which is directly related to the construction accuracy. Combined with dynamic weighting of the tunnel boring machine's posture, the registration can be aligned in real time with the construction progress, guiding the tunnel boring machine to adjust its posture and position, directly serving the need for real-time comparison of design models and monitoring of tunneling deviations in tunnel construction.

[0056] 2. Over-excavation and under-excavation features are extracted using a lightweight convolutional neural network, and the network parameters are dynamically updated using a transfer learning framework. New over-excavation and under-excavation data are collected in real time and merged with historical samples to form an incremental dataset. The incremental dataset is used to update the network with small-batch gradient descent, which can dynamically adapt to geological changes and tunneling deviations during tunnel construction, making over-excavation and under-excavation detection increasingly accurate. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a schematic diagram of the system structure of the present invention;

[0059] Figure 2 This is a schematic diagram of the workflow of the present invention. Detailed Implementation

[0060] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0061] Example 1

[0062] like Figure 1As shown, this embodiment provides a real-time detection system for tunnel over-excavation and under-excavation enabled by edge computing, including:

[0063] The multimodal sensing module is an array of sensors deployed on the tunnel excavation face;

[0064] The noise reduction module, connected to the multimodal sensing module, is used to dynamically construct local BIM models at edge computing nodes and uses an adaptive voxel filtering algorithm to reduce noise in point cloud data.

[0065] The data processing module, connected to the noise reduction module, is used to perform nonlinear registration of real-time point clouds with the design model based on the improved ICP algorithm.

[0066] The data update module, connected to the data processing module, is used to extract over-mining and under-mining features using a lightweight convolutional neural network and dynamically update network parameters using a transfer learning framework.

[0067] The data control module, connected to the data update module, generates an augmented reality visualization report at the edge, including over- and under-mining amounts, spatial distribution, and risk levels, and transmits it back to the control center via a low-latency communication module.

[0068] Example 2

[0069] Based on Example 1, such as Figure 2 As shown, this embodiment provides a real-time detection method for tunnel over-excavation and under-excavation enabled by edge computing, including the following steps:

[0070] Step 1: Real-time acquisition of 3D point cloud data, inertial measurement data, and ambient light intensity data using multimodal sensors deployed on an array on the tunnel excavation surface;

[0071] Step 2: Dynamically construct a local BIM model at the edge computing node, and use an adaptive voxel filtering algorithm to denoise the point cloud data;

[0072] Step 3: Based on the improved ICP algorithm, perform nonlinear registration between the real-time point cloud and the design model; wherein, the nonlinear registration weight is dynamically adjusted by the tunneling machine pose data;

[0073] Step 4: Extract over-mining and under-mining features using a lightweight convolutional neural network, and dynamically update the network parameters using a transfer learning framework;

[0074] Step 5: Generate an augmented reality visualization report on the edge side that includes over- and under-excavation amounts, spatial distribution, and risk levels, and send it back to the control center via a low-latency communication module.

[0075] Example 3

[0076] Based on Example 2, in step 1, the spatial morphology of the excavation face can be accurately restored by the three-dimensional point cloud data. The specific acquisition principle is as follows: a laser beam is emitted or a three-dimensional image is acquired by a stereo vision sensor device. The three-dimensional coordinates (X, Y, Z) of each point on the excavation face are calculated based on the time difference and phase difference of the reflected signal. Finally, a dense set of point clouds is formed, which outlines the contour of the excavation face, the distribution of rock layers, protrusions and depression details like three-dimensional pixels.

[0077] By comparing point cloud data at different times, the advance distance and excavation volume of the excavation face can be calculated to determine whether it conforms to the construction plan, thus facilitating real-time monitoring of the excavation progress.

[0078] By observing the texture and shape changes of point clouds, dangerous geological structures such as faults, fissures, and karst caves can be quickly identified, providing early warnings of collapse risks and facilitating the identification of geological anomalies.

[0079] Provides precise spatial coordinate references for tunnel boring machines (TBMs) or blasting operations, ensuring that the excavation path conforms to the design axis and avoiding over-excavation or under-excavation.

[0080] Inertial measurement data can capture the motion state of the excavation face and equipment. Through the inertial measurement unit (IMU, which includes accelerometers and gyroscopes), the acceleration of the sensors themselves (such as vibration during excavation, acceleration of equipment movement) and angular velocity (such as rotation angle, tilt degree) can be sensed in real time. Combined with algorithms, position, velocity and attitude information can be further calculated.

[0081] For example, if the tilt angle and vibration amplitude of the TBM during advancement exceed the safety threshold, an alarm will be triggered in time to avoid equipment damage or tunnel structure deformation, thus monitoring the stability of the excavation equipment.

[0082] During the excavation process, the equipment may experience slight vibrations. Inertial data can be used to correct the positional deviation of the point cloud, improve the accuracy of 3D modeling, and compensate for the dynamic errors of the 3D point cloud.

[0083] The frequency and amplitude of ground vibrations captured by inertial sensors indirectly reflect the stress release of the strata near the excavation face, helping to determine whether there is a risk of collapse and assess changes in strata stress.

[0084] Ambient light intensity data can help determine the environmental condition of the excavation face. Light sensors can measure the light intensity around the excavation face in real time (such as changes in the intensity of natural light and construction lights).

[0085] It can help determine the safety of the construction environment. For example, if the light suddenly dims (which may be due to excessive dust concentration, equipment obstruction, or partial collapse that blocks the light), it can be combined with other data (such as dust sensors) to quickly find the cause.

[0086] It can optimize the quality of 3D point cloud acquisition. Light intensity affects the imaging effect of lidar or vision sensors. Ambient light data can be used to dynamically adjust the sensor's exposure parameters to ensure the clarity of point cloud data.

[0087] Example 4

[0088] Based on Example 2, in step 2, the dynamic construction of the local BIM model adopts the following steps:

[0089] Based on denoised point clouds, local models are constructed hierarchically and dynamically linked with global BIM.

[0090] Step a: Point cloud segmentation and feature extraction; use a region growing algorithm to segment the point cloud (e.g., divide planar / curved regions according to normal vectors and curvature) and identify components (tunnels, columns and beams).

[0091] Extract component feature parameters: bounding box and thickness of planar regions; radius of curvature and fitting equation of curved regions.

[0092] Step b: Geometric model generation; For planar components (such as walls and floors), polygonal patches are generated by point cloud fitting, and three-dimensional solids are constructed by combining thickness parameters (such as extrusion operation).

[0093] For curved surface components (such as tunnel structures), NURBS surface fitting is used to fit point clouds to generate parametric models, and basic attributes (such as material, dimensions, and ID) are assigned to the models and linked to BIM data standards (such as IFC format).

[0094] Step c: Dynamic fusion of local and global models;

[0095] By using the positioning information of edge nodes (such as GPS and SLAM odometry), the local model coordinates are transformed to the global BIM coordinate system.

[0096] The overlapping areas between the local and global models are detected in real time, and the ICP algorithm is used for registration optimization to eliminate accumulated errors.

[0097] When a local model is updated (e.g., component size adjustment, or addition of a component), incremental update data is pushed to the global BIM in the cloud via edge computing nodes to avoid full data transmission.

[0098] Step d: Model lightweighting and visualization; simplify local models (e.g., reduce the number of faces, retain LOD200 accuracy), and adapt to the storage and display resources of edge nodes;

[0099] Lightweight visualization is implemented using WebGL or Three.js, supporting interactive functions such as real-time roaming, component querying, and property viewing.

[0100] The adaptive voxel filtering algorithm is used to denoise point cloud data as follows:

[0101] Point cloud data preprocessing (lightweight processing of edge nodes); Input: Local point cloud data P = {p i | p i = (x i ,y i ,z i Let P = {i = 1, 2, ..., N}, where P represents the set of point cloud data; N is the number of points; p i Represents the i-th 3D point in the point cloud; (x i y i , z i () is point p i The three-dimensional coordinates; i = 1, 2, ..., N indicate the number of points and the index range in the point cloud.

[0102] The i-th 3D point in a coordinate-normalized point cloud This involves transforming the point cloud coordinates to a local coordinate system (e.g., with the edge nodes as the origin), thus reducing computational complexity. Where (x0, y0, z0) are the coordinates of the edge nodes.

[0103] Preliminary statistical noise screening: Calculate the average distance of all points. Mark distance greater than The points are suspected noise (further voxel verification is required).

[0104] The adaptive voxel mesh is divided as follows:

[0105] The initial voxel size is determined by calculating the initial voxel side length v0 based on the average density ρ of the point cloud (the number of points per unit volume): v0 = k·ρ -1 / 3 Where k is an adjustment coefficient (usually taken as 1.0 to 1.5, controlling the average number of points within a voxel).

[0106] Dynamically adjusting the voxel size is done for each local region (e.g., point p). i Calculate the local density p in the neighborhood of the center. i Adjust the voxel side length v i vi = v0·(ρ0 / ρ) i ) 1 / 3 Where ρ0 is the global average density, details are preserved using small voxels in high-density regions and noise is removed using large voxels in low-density regions.

[0107] Voxel index mapping: mapping each point Mapping to the corresponding voxel grid indices (a, b, c): in, and These are the first normalized point, the second normalized point, and the third normalized point, respectively. To round down, a tool that maps each point in the 3D point cloud to a regular voxel grid is used. The voxel index (a, b, c) is obtained by dividing the coordinates by the voxel side length and then rounding down.

[0108] Voxel-based noise filtering is as follows:

[0109] To count the number of points within each voxel, let voxel V be... a，b，c The number of points inside is n a，b，c ;

[0110] Dynamic threshold calculation determines the retention threshold T based on the local voxel number distribution. i :

[0111] in, σ is the average number of neighborhood voxels. n，i α is the standard deviation of the number of neighborhood voxels, and β are weighting coefficients (taken as 1.0 and 0.5).

[0112] The filtering method is as follows: if n a，b，c ≥T i Preserve the points within the voxel and replace all points with the voxel's center point (or centroid) (downsampling): Where j is the point index within the voxel; p′ a，b，c n is the center (or centroid) of voxels (a, b, c). a，b，c x represents the number of original points within the voxel (a, b, c); j y j and z j These are the three-dimensional coordinates of the j-th original point within the voxel (j ranges from 1 to n). a，b，c (traverse all points within the voxel); ∑x j ,∑y j and ∑z j These are the sums of the x, y, and z coordinates of all points within the voxel (summing up the coordinates of all points within the small cube); if n a，b，c <T i Mark as noise voxels and delete all points.

[0113] The noise-reduced point cloud data is fused with the local BIM model to remove isolated points: the filtered point cloud is checked again, and isolated points whose distance from other points exceeds a threshold (e.g., 2v0) are deleted.

[0114] Point cloud stitching: Aligning the denoised local point cloud with the local point cloud of the BIM model that has been constructed from edge nodes (e.g., using the ICP algorithm).

[0115] Dynamically update the BIM model: Update the geometric information of the local BIM model (e.g., surface point sets of walls and components) with the processed point cloud.

[0116] Example 5

[0117] Based on Example 2, in step 3, for the initial registration parameters, let the real-time acquired point cloud be the source point set: P = {p i |i=1,2,...,n}, where n is the number of points, and the point cloud of the BIM model is the target point set: Q={q j |j=1,2,...,m}, where m is the number of points.

[0118] The improved ICP algorithm performs nonlinear registration between the real-time point cloud and the design model, introducing weights and considering nonlinear errors (such as the curvature effect of the tunnel surface). The objective function is:

[0119]

[0120] Where, min T To find the minimum value of the transformation matrix T, we need to find a set of rotations R and translations t, T = [R|t], where R is a 3×3 rotation matrix and t is a 3×1 translation vector, so that the result of the entire function is minimized.

[0121] To traverse all points (n in total) in the source point cloud, calculate the overall error; w i The dynamic weight of the i-th point (e.g., based on the tunneling machine pose, points closer to the tunnel face have higher weights, highlighting key areas) satisfies Make key data more important in registration.

[0122] q j In the target point cloud, the source point p i The nearest neighbor of the corresponding point after transformation (finding the correspondence through KD trees, etc.); T·p i For source point p i The coordinates after transformation by transformation matrix T (position after rotation and translation);

[0123] ||q j -T·p i || 2 The squared Euclidean distance between the source and target points after transformation measures the registration error. The smaller the value, the more accurate the alignment. ||·|| is the Euclidean distance norm.

[0124] λ·R(T) is the regularization term, which avoids overfitting noise during registration and ensures smooth transformation; λ is the regularization strength hyperparameter, used to control the constraint strength (λ=0 means no regularization, which is prone to overfitting);

[0125] λ≥0 regularization coefficient, balancing point cloud alignment error and transformation smoothness (the larger λ is, the more emphasis is placed on transformation smoothness, which is suitable for scenarios where the tunneling machine moves continuously and avoids abrupt changes in transformation during iteration);

[0126] R(t) is a regularization function that constrains the model parameters T. R(T) constrains the nonlinear regularization term of the transformation matrix (e.g., limiting the transformation difference between adjacent iterations to make the tunneling machine's pose change conform to engineering reality, such as constraining the amplitude of rotation and translation based on mechanical kinematics).

[0127] Point cloud registration near the tunnel face and critical structures is more accurate. Regularization terms make registration unaffected by noise (e.g., point cloud noise from ambient light interference) and complex structures (tunnel surfaces, junctions), directly impacting construction accuracy (e.g., tunnel lining thickness, tunneling direction). Combined with dynamic weighting of the tunnel boring machine's posture, registration can align in real-time with the construction progress, guiding the tunnel boring machine to adjust its attitude and position. By integrating engineering priority (weighting), mathematical optimization (minimizing error), and physical rationality (regularization), point cloud registration becomes both accurate and stable, directly serving the need for real-time comparison of design models and monitoring of tunneling deviations in tunnel construction.

[0128] Example 6

[0129] Based on Example 2, in step 4, the extraction of over-mining and under-mining features for the lightweight convolutional neural network is as follows:

[0130] To adapt to the computing power limitations of edge computing nodes, a simplified CNN structure is adopted to receive pre-processed image data (such as 256×256×3 RGB images or depth maps);

[0131] It contains 3-5 convolutional blocks (each convolutional block consists of a 3×3 convolutional layer + batch normalization + ReLU activation function + 2×2 max pooling layer), and extracts multi-scale features by gradually reducing the feature map size (e.g. 256→128→64→32).

[0132] The formula for calculating convolution is:

[0133] Among them, y i，j After the convolution operation, the output pixel value (or feature value) at position (i, j) in the feature map represents the feature result extracted by the convolution kernel at that position;

[0134] The double summation symbol corresponds to the size of the convolution kernel (here it is a 3×3 convolution kernel, because k and l range from 0 to 2, with a total of 3×3=9 weights). If the convolution kernel is 5×5, the summation range will become k=0 to 4 and l=0 to 4.

[0135] w k，lThe weight parameters (learnable filter coefficients) at position (k, l) in the convolution kernel control how the convolution kernel extracts features (e.g., edges, textures), and are continuously optimized through backpropagation during training.

[0136] x i+k，j+l The pixel value (or feature value) at position (i+k, j+l) in the input feature map (or original image) is the local region covered by the convolution kernel as it slides across the input.

[0137] b is the bias term, which provides a learnable offset for convolution operations, similar to b in the linear equation y = kx + b, making the model fitting more flexible.

[0138] The specific operation is that the convolution kernel (w) k，l The 3×3 matrix (composed of the matrix) slides pixel-by-pixel on the input feature map (x), each time covering a local 3×3 region (x) in the input. i+k，j+l ).

[0139] Weights at corresponding positions (w) k，l ) and input pixel value (x i+k,k+l After multiplying each element point by point and summing the results, we add the bias b to obtain the y-coordinate of the corresponding position in the output feature map. i，j .

[0140] After traversing all positions of the input feature map, a complete output feature map is generated, completing one convolution operation.

[0141] In tunnel over-excavation and under-excavation detection, the input x may be a preprocessed point cloud projection image or a grayscale image, and the convolutional kernel learns the weights w. k,l It can automatically extract over- or under-mined features from images (such as edges, textures, and anomaly patterns) and output y. i,j It is an intermediate result after feature extraction, which is ultimately used to determine the over- or under-mining category.

[0142] For the transfer learning framework, dynamically updating network parameters is as follows:

[0143] New over- and under-excavation data are collected in real time (the sample set is updated every 10-50 meters of excavation) and merged with historical samples to form an incremental dataset D. new .

[0144] Using the incremental dataset D new The network is updated using mini-batch gradient descent (SGD), with the learning rate set to 1 / 10 of the pre-training rate (e.g., η = 0.001). The parameter update formula is as follows:

[0145]

[0146] Where, θ t+1The updated parameters for the next iteration, i.e., the parameters optimized after the model learns new data; θ t η represents the model parameters for the current iteration (e.g., the kernel weights and biases of a CNN, or the parameters to be updated in transfer learning); η is the learning rate, which controls the step size for parameter updates. L(θ) represents the gradient of the loss function L with respect to the parameter θ, reflecting the direction and magnitude of the parameter's influence on the loss; t ;D new ) is the loss function, which measures the current parameter θ. t Predicting new data D new Error at time; D new The newly collected training data consists of incremental samples (new point clouds of tunnel over- or under-excavation and labeled over- or under-excavation tags).

[0147] Take the partial derivative of the loss function L with respect to the parameter θ. Obtain the gradient direction:

[0148] If the gradient is positive, it means that the current parameters are making the loss larger, and the parameters need to be adjusted in the opposite direction to reduce the loss.

[0149] If the gradient is negative, it means that the current parameters are reducing the loss, and optimization can continue along the gradient direction.

[0150] The learning rate η controls the adjustment magnitude, and the parameters are updated in the opposite direction of the gradient:

[0151]

[0152] The minus sign indicates gradient descent: it updates the parameters in the direction that reduces the loss (if the gradient is positive, the parameters decrease; if the gradient is negative, the parameters increase).

[0153] η controls the step size. If η is too large, it may skip the optimal parameters, resulting in inaccurate over-excavation and under-excavation detection. If η is too small, training will be slow, and the model will have difficulty adapting to the dynamic changes in tunnel construction.

[0154] After each excavation distance (e.g., 5 meters), new point cloud / image data is collected. new Mark the area with over- or under-excavation labels;

[0155] Using the current model (parameter θ) t Predict this new data, calculate the loss L, and then find the gradient.

[0156] θ is obtained by updating the parameters according to the formula. t+1 This allows the model to store the over- and under-excavation patterns under new geological conditions;

[0157] By repeating this process, the model can dynamically adapt to geological changes and excavation deviations during tunnel construction, making over-excavation and under-excavation detection increasingly accurate.

[0158] Example 7

[0159] Based on Example 2, in step 5, spatial distribution modeling is performed; based on the preprocessed data, a three-dimensional spatial model of the construction area is constructed, and the over-excavation and under-excavation amounts are bound to specific locations to form spatial distribution data with attributes (e.g., over-excavation of 0.5m in a certain area, under-excavation of 0.3m at a certain point).

[0160] Define risk level classification rules (e.g., over-excavation > 1m is "high risk", 0.3-1m is "medium risk", and < 0.3m is "low risk"), and mark the risk level for each over-excavation and under-excavation area.

[0161] AR visualization report generation: The edge-side AR engine calls the 3D model and over- and under-excavation data, and over- and under-excavation amount, spatial distribution (distinguished by color / label, such as red indicating high-risk over-excavation), and risk level are superimposed on the real scene in real time to generate an AR visualization report.

[0162] The data is transmitted back to the control center with low latency. At the edge, key data from the AR visualization report (including real-time video streams, spatial coordinates, and risk level labels) is compressed and transmitted via low-latency communication modules (such as 5G / 6G private networks or industrial Ethernet). The control center receives and decodes the data, reconstructs the AR scene, and enables remote real-time monitoring of the construction status.

[0163] By following the above steps, localized processing on the edge side and remote collaboration with the control center can be achieved, which is suitable for scenarios such as tunnel excavation and foundation pit construction that require real-time monitoring of over-excavation and under-excavation.

[0164] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope described in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A real-time detection method for tunnel over-excavation and under-excavation enabled by edge computing, characterized in that, Includes the following steps: Step 1: Real-time acquisition of 3D point cloud data, inertial measurement data, and ambient light intensity data using multimodal sensors deployed on an array on the tunnel excavation surface; Step 2: Dynamically construct a local BIM model at the edge computing node, and use an adaptive voxel filtering algorithm to denoise the point cloud data; Step 3: Based on the improved ICP algorithm, perform nonlinear registration between the real-time point cloud and the design model; wherein, the nonlinear registration weight is dynamically adjusted by the tunneling machine pose data; Step 4: Extract over-mining and under-mining features using a lightweight convolutional neural network, and dynamically update the network parameters using a transfer learning framework; Step 5: Generate an augmented reality visualization report on the edge side that includes over- and under-excavation amounts, spatial distribution, and risk levels, and send it back to the control center via a low-latency communication module.

2. The edge computing-enabled real-time detection method for tunnel over-excavation and under-excavation according to claim 1, characterized in that, In step 1, the spatial morphology of the excavation face can be accurately restored by the three-dimensional point cloud data. The specific acquisition principle is as follows: a laser beam is emitted or a three-dimensional image is acquired by a stereo vision sensor device. The three-dimensional coordinates (X, Y, Z) of each point on the excavation face are calculated based on the time difference and phase difference of the reflected signal. Finally, a dense set of point clouds is formed, which outlines the contour of the excavation face, the distribution of rock layers, protrusions and depression details like three-dimensional pixels. Inertial measurement data can capture the motion state of the excavation face and equipment. Through the inertial measurement unit, the acceleration and angular velocity of the sensor itself can be sensed in real time, and the position, velocity and attitude information can be further calculated. Ambient light intensity data can help determine the environmental condition of the excavation face by measuring the light intensity around the excavation face in real time using a light sensor.

3. The edge computing-enabled real-time detection method for tunnel over-excavation and under-excavation according to claim 1, characterized in that, In step 2, the dynamic construction of the local BIM model is carried out using the following steps: Step a: Point cloud segmentation and feature extraction; Step b: Geometric model generation; Step c: Dynamic fusion of local and global models; Step d: Model lightweighting and visualization.

4. The edge computing-enabled real-time detection method for tunnel over-excavation and under-excavation according to claim 1, characterized in that, In step 2, the noise reduction process of the point cloud data by the adaptive voxel filtering algorithm is based on voxel-based noise filtering: To count the number of points within each voxel, let voxel V be... a，b，c The number of points inside is n a，b，c ; Dynamic threshold calculation determines the retention threshold T based on the local voxel number distribution. i : in, σ is the average number of neighborhood voxels. n,i Let n be the standard deviation of the number of neighboring voxels, and α and β be the weighting coefficients; the filtering method is as follows: if n a，b，c ≥T i Preserve the points within the voxel and replace all points with the voxel's center point: Where j is the point index within the voxel; p′ a，b，c Let n be the center point of voxels (a, b, c). a，b，c x represents the number of original points within the voxel (a, b, c); j y j and z j These are the three-dimensional coordinates of the j-th primitive point within the voxel, where j ranges from 1 to n. a，b，c traverse all points within the voxel; ∑x j ,∑y j and ∑z j Let x, y, and z be the sum of the x, y, and z coordinates of all points within the voxel; if n a，b，c <T i Mark as noise voxels and delete all points.

5. The edge computing-enabled real-time detection method for tunnel over-excavation and under-excavation according to claim 1, characterized in that, In step 3, the improved ICP algorithm performs nonlinear registration between the real-time point cloud and the design model, introducing weights and considering nonlinear errors. The objective function is: Where, min T To find the minimum value of the transformation matrix T, we need to find a set of rotations R and translations t, where T = [R|t], R is the rotation matrix, and t is the translation vector. To traverse all points in the source point cloud, there are n points in total; q j In the target point cloud, the source point p i The nearest neighbor of the corresponding point after transformation; T·p i For source point p i Coordinates after transformation by transformation matrix T; ||q j -T·p i || 2 The squared Euclidean distance between the source and target points after transformation measures the registration error. The smaller the value, the more accurate the alignment. ||·|| is the Euclidean distance norm. λ·R(T) is a regularization term that avoids overfitting noise during registration and ensures smooth transformation.

6. The edge computing-enabled real-time detection method for tunnel over-excavation and under-excavation according to claim 5, characterized in that, λ is the regularization intensity hyperparameter, λ≥0 is the regularization coefficient, which balances the point cloud alignment error and transformation smoothness; the larger λ is, the more emphasis is placed on transformation smoothness, which is suitable for scenarios where the tunneling machine moves continuously and avoids abrupt changes in transformation during iteration; R(T) is a regularization function that constrains the model parameter T. R(T) constrains the nonlinear regularization term of the transformation matrix, limiting the transformation difference between adjacent iterations and ensuring that the tunneling machine's pose changes conform to engineering reality.

7. The edge computing-enabled real-time detection method for tunnel over-excavation and under-excavation according to claim 1, characterized in that, In step 4, the over-mining and under-mining features extracted by the lightweight convolutional neural network are calculated using the following convolution formula: Among them, y i，j After the convolution operation, the pixel value at position (i, j) in the output feature map is generated. The double summation symbol corresponds to the size of the convolution kernel; w k，l x represents the weight parameters at position (k, l) in the convolution kernel; i+k，j+l is the pixel value at position (i+k, j+l) in the input feature map, which is the local region covered by the convolution kernel as it slides across the input; b is the bias term, which provides a learnable offset for the convolution operation.

8. The edge computing-enabled real-time detection method for tunnel over-excavation and under-excavation according to claim 1, characterized in that, In step 4, the network parameters of the transfer learning framework are dynamically updated as follows: Using the incremental dataset D new The network is updated using mini-batch gradient descent, with the learning rate set to 1 / 10 of the pre-training rate. The parameter update formula is as follows: Where, θ t+1 The updated parameters for the next iteration; θ t The parameters are the model parameters for the current iteration; η is the learning rate, which controls the step size for parameter updates. L(θ) represents the gradient of the loss function L with respect to the parameter θ, reflecting the direction and magnitude of the parameter's influence on the loss; t ;D new ) is the loss function, which measures the current parameter θ. t Predicting new data D new Error at time; D new This refers to newly collected training data or incremental samples.

9. The edge computing-enabled real-time detection method for tunnel over-excavation and under-excavation according to claim 1, characterized in that, In step 5, spatial distribution modeling is performed; based on the preprocessed data, a three-dimensional spatial model of the construction area is constructed, and the over-excavation and under-excavation amounts are bound to specific locations to form spatial distribution data with attributes. Define risk level classification rules and label the risk level of each over- or under-excavation area; AR visualization report generation: The edge-side AR engine calls the 3D model and over- and under-mining data to overlay the over- and under-mining amount, spatial distribution, and risk level onto the real scene in real time to generate an AR visualization report; Low-latency data is transmitted back to the control center; at the edge, key data from the AR visualization report is compressed and transmitted via a low-latency communication module.

10. A real-time detection system for tunnel over-excavation and under-excavation enabled by edge computing, used to execute the real-time detection method for tunnel over-excavation and under-excavation enabled by edge computing according to any one of claims 1-9, characterized in that, include: The multimodal sensing module is an array of sensors deployed on the tunnel excavation face; The noise reduction module, connected to the multimodal sensing module, is used to dynamically construct a local BIM model at the edge computing node and to perform noise reduction processing on the point cloud data using an adaptive voxel filtering algorithm. A data processing module, connected to the noise reduction processing module, is used to perform nonlinear registration of real-time point clouds and design models based on an improved ICP algorithm; The data update module, connected to the data processing module, is used to extract over-mining and under-mining features using a lightweight convolutional neural network and to dynamically update network parameters using a transfer learning framework. The data control module, connected to the data update module, generates an augmented reality visualization report at the edge side, including over- and under-mining amounts, spatial distribution, and risk levels, and transmits it back to the control center via a low-latency communication module.

Images