Unmanned aerial vehicle tunnel inspection method based on three-dimensional modeling

Abstract

The invention discloses an unmanned aerial vehicle tunnel inspection method based on three-dimensional modeling, and the method comprises the steps: obtaining multi-modal data through a mobile collection device, constructing a high-precision three-dimensional model, extracting spatial attitude information, analyzing a deformation trend, recognizing structural defects through combining deep learning, and fusing multi-source data to generate a dynamic risk thermodynamic diagram. And finally, constructing a digital twinborn model to realize virtualized inspection. According to the routing inspection method, comprehensive perception, accurate analysis and visual display of the tunnel structure state are achieved, the intelligent level of tunnel health monitoring and the accuracy of risk assessment are effectively improved, a reliable basis is provided for tunnel operation and maintenance decisions, and the routing inspection method is of great significance in guaranteeing tunnel structure safety and prolonging the service life.

Description

Technical Field

[0001] The present invention belongs to the field of tunnel detection, and in particular relates to a tunnel inspection method using an unmanned aerial vehicle (UAV) based on three-dimensional modeling. Background Art

[0002] As a core component of transportation infrastructure, the safety and stability of tunnels are crucial to ensuring transportation efficiency and public safety. As tunnel construction scales up and their service life increases, regular inspections have become a key means of ensuring the health of tunnel structures. Traditional tunnel inspection methods rely primarily on manual visual inspections or simple two-dimensional image recordings, which suffer from low efficiency, insufficient coverage, and limited data accuracy. Inspection quality is particularly difficult to guarantee in complex environments, such as at night or in severe weather conditions. Furthermore, traditional methods struggle to achieve a comprehensive digital representation of tunnel structures, resulting in insufficient capabilities for early identification and dynamic monitoring of potential defects.

[0003] Given these limitations, the tunnel inspection field urgently needs a technical solution that can achieve all-weather, high-precision, and intelligent inspection. The core challenge lies in balancing comprehensive data collection with modeling accuracy by combining drone technology with 3D modeling. Comprehensive data collection requires drones to operate stably in complex tunnel environments and acquire multidimensional data. However, high-resolution, multi-source data often increases processing complexity, limiting modeling accuracy. This lack of modeling accuracy further hinders the precise identification of tunnel defects, such as cracks and deformations, and the accuracy of defect identification directly determines the inspection system's ability to assess the tunnel's health. These interrelated technical factors together constitute the core challenge of intelligent tunnel inspection.

[0004] Therefore, how to efficiently collect multi-source data through drones under all-weather conditions and achieve high-precision disease identification and dynamic assessment of tunnel status based on 3D modeling technology has become a key issue that needs to be urgently addressed in the drone tunnel inspection method based on 3D modeling. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides a UAV tunnel inspection method based on three-dimensional modeling, comprising:

[0006] Using a mobile acquisition device equipped with a multimodal sensor, data is collected along a preset trajectory on the inner wall of the tunnel to obtain a first multispectral image and a first point cloud data;

[0007] Performing image enhancement processing on the first multispectral image to obtain a second multispectral image; performing noise filtering on the first point cloud data to obtain second point cloud data;

[0008] Splicing the second multispectral images into an expanded image, and generating an initial three-dimensional model by combining the second point cloud data with point cloud registration and surface reconstruction algorithms;

[0009] If the resolution of the initial three-dimensional model is lower than a preset threshold, performing resolution enhancement processing on the second multispectral image to obtain a third multispectral image, performing density enhancement processing on the second point cloud data to obtain third point cloud data, and re-performing surface reconstruction to generate a target three-dimensional model;

[0010] Extracting spatial posture data based on the target three-dimensional model, generating an inclination distribution map and comparing it with design parameters, calculating the deviation amount and generating an inclination change data set;

[0011] Performing deformation analysis on the tilt angle change data set to generate a cross-sectional contour map and a longitudinal deformation map, and quantifying the deformation degree and storing it as a contour data set;

[0012] Input the contour data set into a pre-trained deep learning model to identify structural defect characteristics and generate a disease distribution map, and calculate the evolution trend based on historical data to generate a disease trend data set;

[0013] fusing the disease trend dataset and the inclination change dataset, generating a dynamic risk heat map through multi-parameter weighted analysis, marking high-risk areas and storing them as a risk heat map;

[0014] Based on the risk thermal dataset and the target three-dimensional model, a digital twin model including structural status and risk distribution is constructed to generate a virtual tunnel environment. Based on the digital twin model, VR / AR technology is combined to roam and inspect various parts of the tunnel in the virtual tunnel environment. Abnormal areas are automatically marked and historical comparative analysis is performed to obtain inspection results.

[0015] Preferably, the process of acquiring the first multispectral image and the first point cloud data further includes:

[0016] If it is determined that noise interference exists in the first multispectral image, a denoising algorithm based on wavelet transform is used to process the first multispectral image to obtain a denoised multispectral image;

[0017] Performing data fusion using a point cloud registration algorithm based on the denoised multispectral image and the first point cloud data to generate a first fused data set;

[0018] If there is a spatial positioning deviation in the first fused dataset, the dataset is adjusted using a preset spatial positioning calibration model to obtain a second fused dataset;

[0019] For the second fused data set, the surface material features and geometric structure features are obtained to determine whether there are defects on the inner wall of the tunnel;

[0020] Processing the surface material features and geometric structure features through a preset defect classification model to determine the defect type and location information;

[0021] According to the defect type and location information, condition assessment data of the tunnel inner wall is generated.

[0022] Preferably, performing image enhancement processing on the first multispectral image to obtain a second multispectral image; and performing noise filtering on the first point cloud data to obtain the second point cloud data include:

[0023] Processing the first multispectral image using a preset image enhancement algorithm to obtain a second multispectral image;

[0024] Processing the first point cloud data using a preset noise filtering algorithm to obtain second point cloud data;

[0025] Extracting multispectral features using a convolutional neural network from the second multispectral image to obtain an image feature set;

[0026] extracting point cloud features using a point cloud segmentation algorithm based on the second point cloud data to obtain a point cloud feature set;

[0027] If the dimensions of the image feature set and the point cloud feature set are consistent, the two are combined through a feature fusion algorithm to obtain a fused feature set;

[0028] If the dimensions are inconsistent, the high-dimensional feature set is reduced to obtain a fused feature set with consistent dimensions;

[0029] For the fused feature set, a clustering algorithm is used to group the features to obtain a feature grouping result.

[0030] Preferably, the process of stitching the second multispectral images into an expanded image and generating an initial three-dimensional model by combining the second point cloud data with point cloud registration and surface reconstruction algorithms includes:

[0031] generating an expanded image according to the second multispectral image by an image stitching algorithm;

[0032] Extract texture features from the expanded image, and use a weighted fusion algorithm based on the texture features of the expanded image and the spatial coordinates of the second point cloud data to obtain a fused data set;

[0033] The fused dataset is spatially aligned using a point cloud registration algorithm to generate a registered point cloud;

[0034] If the spatial deviation of the fused dataset exceeds the preset threshold, the iterative closest point algorithm is used to perform point cloud registration to obtain the registered point cloud; if the spatial deviation of the fused dataset is within the threshold, the registered point cloud is directly output;

[0035] Performing surface reconstruction on the registered point cloud using a triangulation algorithm to obtain an initial three-dimensional model;

[0036] According to the curvature distribution of the initial three-dimensional model, a geometric feature extraction algorithm is used to obtain an optimized point cloud;

[0037] If the surface smoothness of the optimized point cloud is lower than a preset threshold, a surface optimization algorithm is used to process it to obtain a final three-dimensional model; if the smoothness meets the requirements, the final three-dimensional model is directly output;

[0038] According to the texture features of the unfolded image, a texture mapping algorithm is used to project the texture information onto the final three-dimensional model to obtain a three-dimensional model with texture.

[0039] Preferably, the process of re-performing surface reconstruction to generate the target three-dimensional model includes:

[0040] If the resolution of the initial three-dimensional model is lower than a preset threshold, acquiring a second multispectral image and a second point cloud data to determine areas of insufficient resolution and insufficient density;

[0041] performing resolution enhancement on the second multispectral image by a super-resolution algorithm to generate a third multispectral image;

[0042] performing density enhancement on the second point cloud data using a point cloud interpolation algorithm to generate third point cloud data;

[0043] Based on the third multispectral image and the third point cloud data, performing data fusion to generate a unified multimodal dataset;

[0044] The multimodal data set is processed through a surface reconstruction algorithm to generate a target three-dimensional model;

[0045] If the surface smoothness of the target three-dimensional model is lower than a preset standard, performing texture optimization on the third multispectral image to generate optimized texture data;

[0046] Based on the optimized texture data and the target 3D model, texture mapping is performed to generate the final target 3D model.

[0047] Preferably, the process of generating the tilt change data set includes:

[0048] Obtain spatial posture data from the final target 3D model, and use the volume geometry algorithm to extract the spatial coordinates and inclination information of each point cloud to obtain a spatial posture dataset;

[0049] For the spatial posture dataset, a kernel density estimation algorithm is used to generate an inclination angle distribution map, standard inclination angle parameters are obtained from a preset design parameter database, and the Euclidean distance algorithm is used to calculate the deviation between the inclination angle distribution map and the standard angle parameters to obtain a deviation dataset;

[0050] If the deviation amount in the deviation data set exceeds a preset threshold, the deviation amount is classified, and a cluster analysis method is used to identify abnormal inclination areas to obtain the abnormal area distribution;

[0051] According to the distribution of the abnormal areas, the spatial posture data of the corresponding areas are extracted to generate a local tilt angle change data set. The statistical characteristics of the tilt angle changes in each area are calculated based on the local tilt angle change data set to generate a tilt angle change feature set.

[0052] Preferably, the process of performing deformation analysis on the tilt angle change dataset, generating a cross-sectional contour map and a longitudinal deformation map, and quantifying the deformation degree and storing the resultant contour map as a contour dataset comprises:

[0053] Removing noise from the tilt angle change data set through preprocessing to obtain a first data set;

[0054] Extract features from the first data set and determine a deformation feature set using a principal component analysis algorithm;

[0055] For the deformation feature set, if the feature value exceeds a preset threshold, deformation analysis is performed to obtain deformation distribution data;

[0056] generating a cross-sectional contour map using an interpolation algorithm based on the deformation distribution data to obtain a first graph;

[0057] Extracting a longitudinal variation trend based on the deformation distribution data, generating a longitudinal deformation graph using a curve fitting algorithm, and obtaining a second graph;

[0058] Calculating a deformation degree index based on the first graph and the second graph to obtain a quantitative result;

[0059] A contour line data set is generated according to the quantification result, the first graph and the second graph, and the data is stored and saved.

[0060] Preferably, the process of inputting the contour dataset into a pre-trained deep learning model, identifying structural defect characteristics and generating a damage distribution map, and calculating the evolution trend in combination with historical data to generate a damage trend dataset includes:

[0061] Standardizing the data format of the contour data set through a preprocessing technology to obtain a standard data set;

[0062] Processing the standard data set through a pre-trained deep learning model to extract structural defect features and obtain a defect feature set;

[0063] Using a convolutional neural network algorithm, the spatial distribution pattern is identified from the defect feature set to generate a disease distribution map;

[0064] Obtain historical data. If the historical data matches the current disease distribution map in spatial coordinates, integrate them through data fusion technology to obtain an integrated data set.

[0065] For the integrated data set, the time series analysis algorithm is used to calculate the disease evolution trend and obtain the disease trend data set;

[0066] If the evolution rate of the trend feature set exceeds a preset threshold, a dynamic trend graph is generated through a visualization tool.

[0067] Preferably, the process of fusing the disease trend dataset and the inclination change dataset, generating a dynamic risk heat map through multi-parameter weighted analysis, marking high-risk areas and storing the result as a risk heat map comprises:

[0068] Based on the disease trend dataset and the inclination change dataset, outliers are removed through data cleaning to obtain a standardized dataset;

[0069] Extracting disease trend characteristics and inclination change characteristics from the standardized data set through multi-parameter analysis to determine a feature set;

[0070] Using a weighted analysis method, the feature set is weighted according to preset parameter weights to obtain a risk score matrix;

[0071] Generate a dynamic risk heat map based on the risk scoring matrix and mark the risk areas;

[0072] If the risk score of the risk area exceeds a preset threshold, it is marked as a high-risk area to obtain a high-risk marked data set; then the high-risk marked data set and the dynamic risk heat map are stored as a risk heat data set;

[0073] The regional division algorithm is used to perform spatial clustering on the risk thermal dataset to determine the final high-risk area distribution.

[0074] Preferably, the process of obtaining the inspection results includes:

[0075] By using the risk thermal data set and the target 3D model, a data fusion algorithm is used to generate a digital twin model that includes the structural status and risk distribution;

[0076] Based on the digital twin model, a virtualized tunnel environment is generated using three-dimensional rendering technology;

[0077] Using VR / AR technology, a roaming inspection path is obtained in a virtual tunnel environment, and visual data of each part of the tunnel is determined;

[0078] For the visualization data of various parts of the tunnel, if the data deviation exceeds the preset threshold, a convolutional neural network algorithm is used to identify the abnormal area and an automatic labeling algorithm is used to obtain the coordinates and labels of the abnormal area to generate the labeled tunnel part data;

[0079] Based on the marked tunnel area data, historical data comparison results are obtained and time series analysis algorithms are used to determine the change trend of the abnormal area;

[0080] Based on the changing trends, the inspection results are generated and the priority ranking of abnormal areas is obtained.

[0081] Compared with the prior art, the present invention has the following advantages and technical effects:

[0082] This method uses mobile data acquisition equipment to acquire multimodal data, construct a high-precision three-dimensional model, extract spatial posture information, and analyze deformation trends. Deep learning is then used to identify structural defects. Multi-source data is then integrated to generate dynamic risk heat maps, ultimately creating a digital twin model for virtualized inspections. This method enables comprehensive perception, precise analysis, and visualization of tunnel structural status, effectively improving the intelligent level of tunnel health monitoring and the accuracy of risk assessment. It provides a reliable basis for decision-making in tunnel operations and maintenance, and is of great significance for ensuring tunnel structural safety and extending its service life. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0084] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention. DETAILED DESCRIPTION

[0085] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0086] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0087] like Figure 1 As shown, this embodiment provides a UAV tunnel inspection method based on three-dimensional modeling, including:

[0088] Using a mobile acquisition device equipped with a multimodal sensor, data is collected along a preset trajectory on the inner wall of the tunnel to obtain a first multispectral image and a first point cloud data;

[0089] Performing image enhancement processing on the first multispectral image to obtain a second multispectral image; performing noise filtering on the first point cloud data to obtain second point cloud data;

[0090] splicing the second multispectral images into an expanded image, and combining the second point cloud data to generate an initial three-dimensional model through point cloud registration and surface reconstruction algorithms;

[0091] If the resolution of the initial three-dimensional model is lower than a preset threshold, the second multispectral image is subjected to resolution enhancement processing to obtain a third multispectral image, the second point cloud data is subjected to density enhancement processing to obtain a third point cloud data, and the surface reconstruction is re-performed to generate the target three-dimensional model;

[0092] Extract spatial posture data based on the target 3D model, generate an inclination distribution map and compare it with the design parameters, calculate the deviation and generate an inclination change data set;

[0093] Perform deformation analysis on the inclination change dataset to generate cross-sectional contour maps and longitudinal deformation maps, quantify the deformation degree and store it as a contour dataset;

[0094] Input the contour dataset into the pre-trained deep learning model to identify structural defect characteristics and generate a damage distribution map. Combined with historical data, the evolution trend is calculated to generate a damage trend dataset.

[0095] The disease trend dataset and the inclination change dataset are integrated to generate a dynamic risk heat map through multi-parameter weighted analysis, marking high-risk areas and storing them as a risk heat map dataset.

[0096] Based on the risk thermal dataset and the target 3D model, a digital twin model that includes structural status and risk distribution is constructed to generate a virtual tunnel environment. Based on the digital twin model, VR / AR technology is combined to roam and inspect various parts of the tunnel in the virtual tunnel environment. Abnormal areas are automatically marked and historical comparative analysis is performed to obtain inspection results.

[0097] Furthermore, the process of acquiring the first multispectral image and the first point cloud data further includes:

[0098] If it is determined that noise interference exists in the first multispectral image, a denoising algorithm based on wavelet transform is used to process the first multispectral image to obtain a denoised multispectral image;

[0099] Performing data fusion using a point cloud registration algorithm based on the denoised multispectral image and the first point cloud data to generate a first fused data set;

[0100] If there is a spatial positioning deviation in the first fused dataset, the dataset is adjusted using a preset spatial positioning calibration model to obtain a second fused dataset;

[0101] For the second fused data set, the surface material features and geometric structure features are obtained to determine whether there are defects on the inner wall of the tunnel;

[0102] The surface material characteristics and geometric structure characteristics are processed through the preset defect classification model to determine the defect type and location information;

[0103] Generate condition assessment data of the tunnel inner wall based on defect type and location information.

[0104] Specifically, the mobile acquisition device is equipped with a multispectral camera and a lidar sensor, and moves along a preset trajectory along the inner wall of the tunnel at a uniform speed of 0.5 meters per second. The multispectral camera uses a channel imaging system with a band range of 450-900 nanometers, capturing one frame of image every 10 centimeters and synchronously recording GPS coordinates and IMU attitude data. The lidar emits 905-nanometer pulsed lasers at a frequency of 10 Hz, with a single-point ranging accuracy of ±2 centimeters. The point cloud is spliced ​​in real time through the SLAM algorithm to generate a three-dimensional model with a point density of 1000 points per square meter.

[0105] The multispectral images use the Retinex algorithm for illumination compensation to eliminate the effects of uneven lighting within the tunnel. Principal component analysis is performed on the 400-500 nm band images to extract corrosion features, and a threshold of 0.35 is set to segment suspected damaged areas. The point cloud data is fitted to a cylindrical model of the tunnel's inner wall using the RANSAC algorithm. The distance residual between each point and the model is calculated, and areas with a residual exceeding 3 cm are marked as structural deformation. Combined with the corrosion areas identified in the multispectral images, a support vector machine classifier is used for multimodal fusion analysis. When the overlap between the corrosion area and the structural deformation area exceeds 0.2 square meters, a Level 3 damage alert is triggered. The system automatically generates an inspection report containing the coordinates, damage type, and severity, and transmits it back to the data center in real time via the 5G network.

[0106] Furthermore, the process of performing image enhancement processing on the first multispectral image to obtain a second multispectral image and performing noise filtering on the first point cloud data to obtain the second point cloud data includes:

[0107] Processing the first multispectral image using a preset image enhancement algorithm to obtain a second multispectral image;

[0108] Processing the first point cloud data using a preset noise filtering algorithm to obtain second point cloud data;

[0109] According to the second multispectral image, a convolutional neural network is used to extract multispectral features to obtain an image feature set;

[0110] According to the second point cloud data, point cloud segmentation algorithm is used to extract point cloud features to obtain a point cloud feature set;

[0111] If the dimensions of the image feature set and the point cloud feature set are consistent, the two are combined through the feature fusion algorithm to obtain the fused feature set;

[0112] If the dimensions are inconsistent, the high-dimensional feature set is reduced to obtain a fused feature set with consistent dimensions;

[0113] For the fused feature set, a clustering algorithm is used to group the features to obtain the feature grouping results.

[0114] Specifically, when performing image enhancement processing on the first multispectral image, this embodiment adopts a histogram equalization algorithm to enhance contrast by adjusting the grayscale distribution of the image.

[0115] For example, for an image containing red, green, and blue bands, we first calculate the grayscale histogram for each band, then perform histogram equalization on each band separately. Assuming the grayscale range of the red band is 0 to 255, after equalization, the grayscale values ​​are more evenly distributed, thereby improving the image's detail. Next, the same processing is performed on the green and blue bands, resulting in a second multispectral image with significantly improved overall contrast and detail clarity.

[0116] When performing noise filtering on the first point cloud data, this embodiment adopts a statistically based outlier removal algorithm. First, the k-nearest neighbor distance of each point in the point cloud is calculated, assuming that the k value is 10, and then the average and standard deviation of these distances are calculated. For each point, if the average value of its k-nearest neighbor distance exceeds a preset threshold (for example, 3 times the standard deviation), the point is regarded as a noise point and removed. In this way, outliers in the point cloud data can be effectively removed, and cleaner and more accurate second point cloud data can be obtained. This process not only improves the quality of the point cloud data, but also provides a reliable foundation for subsequent tasks such as three-dimensional reconstruction or target recognition.

[0117] Furthermore, the second multispectral image is stitched into an expanded image, and the process of generating an initial three-dimensional model by combining the second point cloud data with point cloud registration and surface reconstruction algorithm includes:

[0118] generating an expanded image according to the second multispectral image by an image stitching algorithm;

[0119] Extract texture features from the expanded image, and use a weighted fusion algorithm based on the texture features of the expanded image and the spatial coordinates of the second point cloud data to obtain a fused data set;

[0120] The fused dataset is spatially aligned using a point cloud registration algorithm to generate a registered point cloud;

[0121] If the spatial deviation of the fused dataset exceeds the preset threshold, the iterative closest point algorithm is used to perform point cloud registration to obtain the registered point cloud; if the spatial deviation of the fused dataset is within the threshold, the registered point cloud is directly output;

[0122] The registered point cloud is reconstructed using a triangulation algorithm to obtain an initial three-dimensional model;

[0123] According to the curvature distribution of the initial 3D model, the geometric feature extraction algorithm is used to obtain the optimized point cloud;

[0124] If the surface smoothness of the optimized point cloud is lower than the preset threshold, the surface optimization algorithm is used to process it and obtain the final 3D model; if the smoothness meets the requirements, the final 3D model is directly output;

[0125] According to the texture features of the unfolded image, a texture mapping algorithm is used to project the texture information onto the final three-dimensional model to obtain a textured three-dimensional model.

[0126] Specifically, when stitching the second multispectral image into an expanded image, the multispectral images are first aligned using an image registration algorithm. A feature point-based registration method is used to extract SIFT feature points from the image, calculate the Euclidean distance between the feature points, set the matching threshold to 0.7, screen out matching point pairs, and use the RANSAC algorithm to remove mismatched points. Finally, the multispectral images are stitched into an expanded image using an affine transformation. Combined with the second point cloud data, the ICP (Iterative Closest Point) algorithm is used for point cloud registration. The number of iterations is set to 100 and the convergence threshold is set to 0.001. By minimizing the mean square error between the point clouds, accurate alignment of the point clouds is achieved. In the surface reconstruction stage, the Poisson surface reconstruction algorithm is used, with a depth of 8 and a smoothing coefficient of 0.5. The initial three-dimensional model is generated by solving the Poisson equation. During this process, the point cloud is color mapped using the spectral information of the multispectral image to ensure that the model has realistic spectral characteristics. Through the above steps, the generated three-dimensional model not only has a high-precision geometric structure, but also retains the spectral information of the multispectral image, providing a reliable foundation for subsequent analysis and application.

[0127] Furthermore, the process of re-performing surface reconstruction to generate the target three-dimensional model includes:

[0128] If the resolution of the initial three-dimensional model is lower than a preset threshold, obtaining a second multispectral image and a second point cloud data to determine areas of insufficient resolution and insufficient density;

[0129] performing resolution enhancement on the second multispectral image using a super-resolution algorithm to generate a third multispectral image;

[0130] Density enhancement is performed on the second point cloud data using a point cloud interpolation algorithm to generate third point cloud data;

[0131] Based on the third multispectral image and the third point cloud data, performing data fusion to generate a unified multimodal dataset;

[0132] The multimodal data set is processed through a surface reconstruction algorithm to generate a target three-dimensional model;

[0133] If the surface smoothness of the target three-dimensional model is lower than a preset standard, texture optimization is performed on the third multispectral image to generate optimized texture data;

[0134] Based on the optimized texture data and the target 3D model, texture mapping is performed to generate the final target 3D model.

[0135] For example, when processing the initial 3D model, if the resolution falls below a preset threshold, the second multispectral image and second point cloud data are analyzed to identify areas of insufficient resolution and density. Insufficient resolution typically manifests as blurred image details, while insufficient point cloud density results in holes in the model surface. For resolution enhancement, super-resolution algorithms, based on deep learning models, map low-resolution images to a high-resolution space.

[0136] For example, this embodiment uses a convolutional neural network to input the second multispectral image. The network extracts features through multi-layer convolution, predicts high-frequency details, and generates a third multispectral image with a resolution twice as high.

[0137] Preferably, a multispectral image dataset can be used during training to ensure spectral channel consistency.

[0138] Specifically, point cloud density is increased through a point cloud interpolation algorithm. One possible implementation involves using a radial basis function-based interpolation method. For areas of insufficient density in the second point cloud, adjacent points are selected, with an interpolation radius of 0.05 meters. New points are generated to fill in the gaps, forming the third point cloud.

[0139] It should be noted that the geometric consistency of the point cloud must be maintained during the interpolation process to avoid introducing noise.

[0140] In one embodiment, the data fusion combines the third multispectral image and the third point cloud data into a multimodal dataset.

[0141] For example, spatial mapping aligns the pixels of the multispectral image with the coordinates of the point cloud, setting a pixel-to-point mapping threshold of 0.01 meters to ensure precise correspondence between spectral and geometric information. After fusion, the dataset contains both spatial coordinates and spectral features, providing comprehensive information for subsequent modeling.

[0142] It is understood that the surface reconstruction algorithm can use a stereo microscope-based algorithm to process multimodal data sets.

[0143] For example, a voxel grid resolution of 0.02 meters was set, and surface normals were calculated using a stereo microscope algorithm to generate a 3D model of the target. This method can effectively handle complex geometries and ensure surface continuity.

[0144] Preferably, if the surface smoothness of the target three-dimensional model is insufficient, the visual quality can be improved through texture optimization.

[0145] For example, we perform texture enhancement based on wavelet transform on the third multispectral image, setting the decomposition level to 3, enhancing high-frequency texture details, and generating optimized texture data. This process can improve the image's sharpness and spectral expressiveness.

[0146] In one possible implementation, texture mapping applies optimized texture data to a target three-dimensional model.

[0147] For example, UV mapping technology is used to map the 2D coordinates of the texture image to the 3D surface points of the model. The texture resolution is set to 2048x2048 pixels to ensure that the texture fits the surface without stretching. The final 3D model has both high-resolution geometry and realistic spectral textures.

[0148] Furthermore, the process of generating the tilt change dataset includes:

[0149] Obtain spatial posture data from the final target 3D model, and use the volume geometry algorithm to extract the spatial coordinates and inclination information of each point cloud to obtain a spatial posture dataset;

[0150] For the spatial posture dataset, the kernel density estimation algorithm is used to generate the inclination angle distribution map. The standard inclination angle parameters are obtained from the preset design parameter database. The Euclidean distance algorithm is used to calculate the deviation between the inclination angle distribution map and the standard angle parameters to obtain the deviation dataset.

[0151] If the deviation amount in the deviation data set exceeds the preset threshold, the deviation amount is classified, and the abnormal dip area is identified using the cluster analysis method to obtain the abnormal area distribution;

[0152] According to the distribution of abnormal areas, the spatial posture data of the corresponding areas are extracted to generate a local tilt angle change data set. The statistical characteristics of the tilt angle changes in each area are calculated through the local tilt angle change data set to generate a tilt angle change feature set.

[0153] Specifically, to extract spatial pose data from the target 3D model, the model is first sampled using a point cloud processing algorithm to obtain the 3D coordinates of key points. For example, a Poisson sampling algorithm is used to generate 1,000 sampling points to ensure uniform data density. Next, the principal component analysis (PCA) algorithm is used to calculate the normal vector of each sampling point to determine its spatial pose.

[0154] For example, for a certain sampling point, its normal vector is (0.707, 0.707, 0), indicating that the angle of inclination of the point to the horizontal plane is 45 degrees. Based on these normal vector data, an inclination distribution map is generated. The inclination data is smoothed using the kernel density estimation (KDE) algorithm to generate a continuous inclination distribution curve.

[0155] For example, the inclination distribution map shows that the inclination angle in most areas is concentrated between 30 and 60 degrees, with a peak at around 45 degrees. The inclination distribution map is compared with the design parameters, which specify an inclination angle of 40 degrees. The deviation between the actual inclination angle and the design value is calculated to generate an inclination variation dataset.

[0156] For example, if a certain area has an inclination of 50 degrees and a deviation of 10 degrees, the least squares method is used to fit the deviation to generate an inclination curve, showing how the deviation changes with spatial position. This process allows for a comprehensive analysis of the degree of consistency between the spatial pose of the target 3D model and the design parameters, providing data support for subsequent optimization.

[0157] Furthermore, deformation analysis is performed on the inclination change dataset to generate a cross-sectional contour map and a longitudinal deformation map. The deformation degree is quantified and stored as a contour dataset. The process includes:

[0158] Removing noise from the tilt angle change data set through preprocessing to obtain a first data set;

[0159] Extract features from the first data set and determine a deformation feature set using a principal component analysis algorithm;

[0160] For the deformation feature set, if the feature value exceeds the preset threshold, deformation analysis is performed to obtain deformation distribution data;

[0161] Based on the deformation distribution data, an interpolation algorithm is used to generate a cross-section contour map to obtain a first graph;

[0162] Extracting the longitudinal variation trend based on the deformation distribution data, generating a longitudinal deformation graph through a curve fitting algorithm, and obtaining a second graph;

[0163] Calculating a deformation degree index based on the first graph and the second graph to obtain a quantitative result;

[0164] A contour line data set is generated according to the quantification result, the first graph and the second graph, and the data is stored and saved.

[0165] Furthermore, the contour dataset is input into the pre-trained deep learning model to identify structural defect characteristics and generate a damage distribution map. The process of calculating the evolution trend and generating a damage trend dataset based on historical data includes the following:

[0166] The contour data set is standardized by preprocessing technology to obtain a standard data set;

[0167] The standard data set is processed by pre-trained deep learning models to extract structural defect features and obtain a defect feature set;

[0168] Using convolutional neural network algorithm, we can identify the spatial distribution pattern from the defect feature set and generate a disease distribution map;

[0169] Obtain historical data. If the historical data matches the current disease distribution map in spatial coordinates, integrate them through data fusion technology to obtain an integrated data set.

[0170] For the integrated data set, the time series analysis algorithm is used to calculate the disease evolution trend and obtain the disease trend data set;

[0171] If the evolution rate of the trend feature set exceeds a preset threshold, a dynamic trend graph is generated through a visualization tool.

[0172] For example, when acquiring a contour dataset, high-precision sensors are used to collect deformation data of the tunnel structure. This contour dataset records changes in tunnel surface elevation, reflecting the spatial distribution of structural deformation. This data may come from a laser scanner, achieving millimeter-level accuracy. Preprocessing techniques standardize the data format, requiring a unified coordinate system and units.

[0173] For example, outliers in tunnel data were removed using a box plot method to remove sudden changes in height, such as a point with a height of 10 meters, which far exceeded the average value by 0.5 meters. After cleaning, the data was normalized to a range of 0 to 1 to generate a standard dataset. This process ensures data consistency and facilitates subsequent analysis.

[0174] In one possible implementation, a pre-trained deep learning model processes a standard dataset to extract structural defect features. This model, which can use a ResNet architecture and be pre-trained on a large-scale tunnel deformation dataset, excels at identifying defects such as cracks and settlement. After inputting the standard dataset, the model outputs a defect feature set, including defect type, size, and location.

[0175] For example, a crack was identified in the middle of a tunnel, measuring 0.02 meters wide and 1.5 meters long, with coordinates of x = 100 meters and y = 20 meters. Feature extraction focused on critical structural areas, such as the tunnel roof and sidewalls, ensuring that the features were relevant to safety. This defect feature set provided the foundation for subsequent spatial analysis.

[0176] Specifically, a convolutional neural network algorithm identifies spatial distribution patterns from a set of defect features and generates a defect distribution map. Convolutional neural networks excel at capturing spatial patterns, using a 3x3 convolution kernel to scan the feature set and extract clusters of defects.

[0177] For example, a tunnel with dense cracks in the middle of its span could be identified by the network, which then generates a defect distribution map. The map uses color-coded defect density, with red representing high-density areas, covering the middle 50 meters of the tunnel. This map visually illustrates defect distribution and facilitates the identification of high-risk areas.

[0178] It should be noted that when acquiring historical data and matching spatial coordinates, the historical data may include monitoring records from the past five years in CSV format, recording information such as cracks and settlement. Matching is based on coordinate alignment, with a tolerance of 0.1 meters.

[0179] For example, a current defect distribution map shows a crack at a certain point with coordinates x = 100 meters and y = 20 meters. A matching point is found in the historical data, recording a crack width of 0.01 meters in 2023. Data fusion technology integrates current and historical data to generate a fused dataset containing timestamps and defect change records. This dataset provides comprehensive information for trend analysis.

[0180] Preferably, a time series analysis algorithm is used to calculate the disease evolution trend, and an ARIMA model is used to analyze the fused data set to output a trend feature set. Features include defect expansion rate, periodic changes, etc.

[0181] For example, a crack's width increases from 0.01 to 0.02 meters, with an expansion rate of 0.005 meters per year. Trend feature sets quantify the dynamics of the damage, providing a basis for prediction. Based on these trend feature sets, a damage trend dataset is generated, recording the expansion rate, predicted value, and other information.

[0182] For example, a crack in a certain area is expected to reach a width of 0.025 meters in 2026. This dataset facilitates long-term monitoring.

[0183] In one embodiment, if the evolution rate exceeds a threshold of 0.01 m / year, a dynamic trend graph is generated using a visualization tool. The tool may employ Tableau to generate an interactive graph showing the change in crack width over time.

[0184] For example, a crack rate of 0.015 meters per year is observed in the middle section of a tunnel. A dynamic trend chart is presented as a line graph, with areas exceeding the threshold highlighted in red. This chart visually demonstrates the deterioration trend of the problem, facilitating maintenance planning. Extended analysis results support precise management and optimized resource allocation.

[0185] Furthermore, the disease trend dataset and the inclination change dataset are integrated to generate a dynamic risk heat map through multi-parameter weighted analysis. The high-risk areas are marked and stored as a risk heat map dataset. The process includes:

[0186] Based on the disease trend dataset and the inclination change dataset, outliers were removed through data cleaning to obtain a standardized dataset;

[0187] Through multi-parameter analysis, the disease trend characteristics and inclination change characteristics are extracted from the standardized data set to determine the feature set;

[0188] Using the weighted analysis method, the feature set is weighted according to the preset parameter weights to obtain the risk score matrix;

[0189] Generate a dynamic risk heat map based on the risk scoring matrix and mark the risk areas;

[0190] If the risk score of a risk area exceeds a preset threshold, it is marked as a high-risk area, and a high-risk labeled dataset is obtained; the high-risk labeled dataset and the dynamic risk heat map are then stored as a risk heat dataset;

[0191] The regional division algorithm is used to perform spatial clustering on the risk thermal dataset to determine the final distribution of high-risk areas.

[0192] For example, when acquiring a defect trend dataset and an inclination change dataset, monitoring data of the tunnel structure is typically collected through a sensor network. The defect trend dataset records the time series changes in defects such as cracks and cavities, while the inclination change dataset reflects the dynamic evolution of the structure's inclination angle.

[0193] During the data cleaning phase, outliers need to be removed to ensure data quality.

[0194] In one possible implementation, a statistical anomaly detection method is used, for example, using the mean and standard deviation to set a threshold and eliminating data points that deviate from the mean by 3 times the standard deviation.

[0195] For example, in a tunnel crack area data set, a record showed an outlier of 100 square meters, far exceeding the mean by 5 square meters and with a standard deviation of 2 square meters. This outlier was identified and removed. After cleaning, the data was normalized using z-score scaling to a standard normal distribution, generating a standardized dataset. This process ensures the reliability of subsequent analysis and avoids noise interference.

[0196] Specifically, multi-parameter analysis extracts features from standardized data sets. Defect trend features include the rate of change in defect area and the direction of expansion, while inclination change features include the rate of change in inclination angle and periodic fluctuations.

[0197] In one embodiment, the annual change rate of crack area is extracted as 1 square meter per year, and the rate of change of inclination is extracted as 0.5 degrees per year. During feature extraction, principal component analysis can be used to screen key variables and reduce redundant dimensions. The resulting feature set encompasses the spatiotemporal characteristics of the damage and inclination, providing a basis for risk assessment.

[0198] It should be noted that feature extraction needs to be combined with structural mechanics properties to ensure that the features are directly related to tunnel safety.

[0199] Preferably, the weighted analysis method assigns weights according to parameter importance.

[0200] For example, the weight of the crack propagation rate is set to 0.6, and the weight of the inclination change rate is set to 0.4, reflecting the greater impact of cracks on structural stability. After weighted calculation, a risk score matrix is ​​generated, and the matrix elements represent the risk value of each area.

[0201] In one example, a region with rapid crack growth and significant dip angle changes received a score of 0.85, higher than the average of 0.5. The risk scoring matrix intuitively quantifies potential threats and facilitates subsequent visualization.

[0202] In one possible implementation, a dynamic risk heat map is generated based on a scoring matrix, using a color gradient to represent risk levels, with red indicating high risk and green indicating low risk. The heat map resolution is set to 1 meter, covering the entire tunnel area. Areas with a risk score exceeding a threshold of 0.8 are marked as high-risk areas.

[0203] For example, an area in the middle of a tunnel with a score of 0.9 is labeled as high risk, generating a high-risk labeled dataset. This dataset records the geographic coordinates and score of the risk area, facilitating precise location.

[0204] It is understood that high-risk annotated datasets and heat maps are stored as risk heat datasets, typically in GeoJSON format, compatible with geographic information systems. They are stored with timestamps and metadata to ensure data traceability.

[0205] For example, a dataset records the distribution of high-risk areas in May 2025, and its metadata includes monitoring equipment information. Storage technology must ensure data security and prevent loss.

[0206] For example, the regionalization algorithm uses the DBSCAN clustering method to identify high-risk clusters based on the spatial distribution of the risk thermal dataset. The clustering radius is set to 3 meters, and the minimum sample size is set to 5 points. This method effectively delineates risk areas and provides a basis for precise maintenance. The clustering results intuitively reflect the concentration of high-risk areas, helping to optimize resource allocation.

[0207] In one embodiment, each step, from data cleaning to feature extraction, risk scoring, and visualization, supports the final regional division. Cleaning ensures data quality, feature extraction focuses on key indicators, weighted analysis quantifies risk, heat maps and annotations provide intuitive presentation, and storage and clustering ensure the practicality of the results. This multi-faceted solution collectively aims to accurately assess and manage tunnel safety.

[0208] Furthermore, the process of obtaining the inspection results includes:

[0209] By using the risk thermal data set and the target 3D model, a data fusion algorithm is used to generate a digital twin model that includes the structural status and risk distribution;

[0210] Based on the digital twin model, 3D rendering technology is used to generate a virtual tunnel environment;

[0211] Using VR / AR technology, a roaming inspection path is obtained in a virtual tunnel environment, and visual data of each part of the tunnel is determined;

[0212] For the visualization data of various parts of the tunnel, if the data deviation exceeds the preset threshold, a convolutional neural network algorithm is used to identify the abnormal area and an automatic labeling algorithm is used to obtain the coordinates and labels of the abnormal area to generate the labeled tunnel part data;

[0213] Based on the marked tunnel area data, historical data comparison results are obtained and time series analysis algorithms are used to determine the change trend of the abnormal area;

[0214] Based on the changing trends, the inspection results are generated and the priority ranking of abnormal areas is obtained.

[0215] Specifically, based on the risk thermal dataset and the target 3D model, point cloud data acquisition technology was first used to acquire high-precision 3D point cloud data of the tunnel interior, with a point cloud density of 1,000 points per square meter to ensure model precision. A point cloud registration algorithm was then used to align the multi-station scan data, with a registration error of less than 2 mm, to form a complete 3D model. Next, the risk thermal dataset was fused with the 3D model, and a heatmap interpolation algorithm was used to map risk data such as temperature and humidity onto the model surface, generating a risk distribution map.

[0216] For example, an abnormal temperature area on the wall of a certain section of tunnel is calculated through an interpolation algorithm, and the temperature gradient is 5 degrees Celsius per meter, which is automatically marked as a high-risk area. In the virtualized tunnel environment, combined with VR / AR technology, users can roam in the virtual tunnel. The system renders the three-dimensional model and risk distribution of each part of the tunnel in real time, with a roaming speed of 10 frames per second to ensure a smooth experience. The system automatically detects abnormal areas such as cracks and water seepage through image recognition algorithms. Areas with crack widths exceeding 3 mm will be automatically marked. At the same time, the system calls historical data for comparative analysis. For example, if a crack on the wall of a certain section of tunnel has expanded by 1 mm in the past three months, the system will generate an inspection report, indicating that this area requires special attention. Finally, the inspection results are presented in the form of visual charts, including risk levels, locations of abnormal areas, and historical change trends, providing data support for tunnel maintenance.

[0217] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A UAV tunnel inspection method based on three-dimensional modeling, characterized in that: include: Using a mobile acquisition device equipped with a multimodal sensor, data is collected along a preset trajectory on the inner wall of the tunnel to obtain a first multispectral image and a first point cloud data; performing image enhancement processing on the first multispectral image to obtain a second multispectral image; performing noise filtering on the first point cloud data to obtain second point cloud data; Splicing the second multispectral images into an expanded image, and generating an initial three-dimensional model by combining the second point cloud data with point cloud registration and surface reconstruction algorithms; If the resolution of the initial three-dimensional model is lower than a preset threshold, performing resolution enhancement processing on the second multispectral image to obtain a third multispectral image, performing density enhancement processing on the second point cloud data to obtain third point cloud data, and re-performing surface reconstruction to generate a target three-dimensional model; Extracting spatial posture data based on the target three-dimensional model, generating an inclination distribution map and comparing it with design parameters, calculating the deviation and generating an inclination change data set; Performing deformation analysis on the tilt angle change data set to generate a cross-sectional contour map and a longitudinal deformation map, and quantifying the deformation degree and storing it as a contour data set; Input the contour data set into a pre-trained deep learning model to identify structural defect characteristics and generate a disease distribution map, and calculate the evolution trend based on historical data to generate a disease trend data set; fusing the disease trend dataset and the inclination change dataset, generating a dynamic risk heat map through multi-parameter weighted analysis, marking high-risk areas and storing them as a risk heat map; Based on the risk thermal dataset and the target three-dimensional model, a digital twin model including structural status and risk distribution is constructed to generate a virtual tunnel environment. Based on the digital twin model, VR / AR technology is combined to roam and inspect various parts of the tunnel in the virtual tunnel environment. Abnormal areas are automatically marked and historical comparative analysis is performed to obtain inspection results.

2. The method according to claim 1, characterized in that The process of acquiring the first multispectral image and the first point cloud data further includes: If it is determined that noise interference exists in the first multispectral image, a denoising algorithm based on wavelet transform is used to process the first multispectral image to obtain a denoised multispectral image; Performing data fusion using a point cloud registration algorithm based on the denoised multispectral image and the first point cloud data to generate a first fused data set; If there is a spatial positioning deviation in the first fused dataset, the dataset is adjusted using a preset spatial positioning calibration model to obtain a second fused dataset; For the second fused data set, the surface material features and geometric structure features are obtained to determine whether there are defects on the inner wall of the tunnel; Processing the surface material characteristics and geometric structure characteristics through a preset defect classification model to determine the defect type and location information; According to the defect type and location information, condition assessment data of the tunnel inner wall is generated.

3. The method according to claim 1, characterized in that performing image enhancement processing on the first multispectral image to obtain a second multispectral image; The process of performing noise filtering on the first point cloud data to obtain second point cloud data includes: Processing the first multispectral image using a preset image enhancement algorithm to obtain a second multispectral image; Processing the first point cloud data using a preset noise filtering algorithm to obtain second point cloud data; Extracting multispectral features using a convolutional neural network from the second multispectral image to obtain an image feature set; extracting point cloud features using a point cloud segmentation algorithm based on the second point cloud data to obtain a point cloud feature set; If the dimensions of the image feature set and the point cloud feature set are consistent, the two are combined through a feature fusion algorithm to obtain a fused feature set; If the dimensions are inconsistent, the high-dimensional feature set is reduced to obtain a fused feature set with consistent dimensions; For the fused feature set, a clustering algorithm is used to group the features to obtain a feature grouping result.

4. The method according to claim 1, wherein The process of stitching the second multispectral images into an expanded image and combining the second point cloud data to generate an initial three-dimensional model through point cloud registration and surface reconstruction algorithm includes: generating an expanded image according to the second multispectral image by an image stitching algorithm; Extract texture features from the expanded image, and use a weighted fusion algorithm based on the texture features of the expanded image and the spatial coordinates of the second point cloud data to obtain a fused data set; The fused dataset is spatially aligned using a point cloud registration algorithm to generate a registered point cloud; If the spatial deviation of the fused dataset exceeds the preset threshold, the iterative closest point algorithm is used to perform point cloud registration to obtain the registered point cloud; if the spatial deviation of the fused dataset is within the threshold, the registered point cloud is directly output; Performing surface reconstruction on the registered point cloud using a triangulation algorithm to obtain an initial three-dimensional model; According to the curvature distribution of the initial three-dimensional model, a geometric feature extraction algorithm is used to obtain an optimized point cloud; If the surface smoothness of the optimized point cloud is lower than a preset threshold, a surface optimization algorithm is used to process it to obtain a final three-dimensional model; if the smoothness meets the requirements, the final three-dimensional model is directly output; According to the texture features of the unfolded image, a texture mapping algorithm is used to project the texture information onto the final three-dimensional model to obtain a three-dimensional model with texture.

5. The method according to claim 1, wherein The process of re-performing surface reconstruction to generate the target 3D model includes: If the resolution of the initial three-dimensional model is lower than a preset threshold, acquiring a second multispectral image and a second point cloud data to determine areas of insufficient resolution and insufficient density; performing resolution enhancement on the second multispectral image by a super-resolution algorithm to generate a third multispectral image; performing density enhancement on the second point cloud data using a point cloud interpolation algorithm to generate third point cloud data; Based on the third multispectral image and the third point cloud data, performing data fusion to generate a unified multimodal dataset; The multimodal data set is processed through a surface reconstruction algorithm to generate a target three-dimensional model; If the surface smoothness of the target three-dimensional model is lower than a preset standard, performing texture optimization on the third multispectral image to generate optimized texture data; Based on the optimized texture data and the target 3D model, texture mapping is performed to generate the final target 3D model.

6. The method according to claim 1, characterized in that The process of generating a tilt change dataset includes: Obtain spatial posture data from the final target 3D model, and use the volume geometry algorithm to extract the spatial coordinates and inclination information of each point cloud to obtain a spatial posture dataset; For the spatial posture dataset, a kernel density estimation algorithm is used to generate an inclination angle distribution map, standard inclination angle parameters are obtained from a preset design parameter database, and the Euclidean distance algorithm is used to calculate the deviation between the inclination angle distribution map and the standard angle parameters to obtain a deviation dataset; If the deviation amount in the deviation data set exceeds a preset threshold, the deviation amount is classified, and a cluster analysis method is used to identify abnormal inclination areas to obtain the abnormal area distribution; According to the distribution of the abnormal areas, the spatial posture data of the corresponding areas are extracted to generate a local tilt angle change data set. The statistical characteristics of the tilt angle changes in each area are calculated based on the local tilt angle change data set to generate a tilt angle change feature set.

7. The method according to claim 1, characterized in that The process of performing deformation analysis on the tilt change dataset, generating a cross-sectional contour map and a longitudinal deformation map, and quantifying the deformation degree and storing it as a contour dataset includes: Removing noise from the tilt angle change data set through preprocessing to obtain a first data set; Extract features from the first data set and determine a deformation feature set using a principal component analysis algorithm; For the deformation feature set, if the feature value exceeds a preset threshold, deformation analysis is performed to obtain deformation distribution data; generating a cross-sectional contour map using an interpolation algorithm based on the deformation distribution data to obtain a first graph; Extracting a longitudinal variation trend based on the deformation distribution data, generating a longitudinal deformation graph using a curve fitting algorithm, and obtaining a second graph; Calculating a deformation degree index based on the first graph and the second graph to obtain a quantitative result; A contour line data set is generated according to the quantification result, the first graph and the second graph, and the data is stored and saved.

8. The method according to claim 1, characterized in that The process of inputting the contour dataset into a pre-trained deep learning model, identifying structural defect characteristics and generating a damage distribution map, and calculating the evolution trend in combination with historical data to generate a damage trend dataset includes: Standardizing the data format of the contour data set through a preprocessing technology to obtain a standard data set; Processing the standard data set through a pre-trained deep learning model to extract structural defect features and obtain a defect feature set; Using a convolutional neural network algorithm, the spatial distribution pattern is identified from the defect feature set to generate a disease distribution map; Obtain historical data. If the historical data matches the current disease distribution map in spatial coordinates, integrate them through data fusion technology to obtain an integrated data set. For the integrated data set, the time series analysis algorithm is used to calculate the disease evolution trend and obtain the disease trend data set; If the evolution rate of the trend feature set exceeds a preset threshold, a dynamic trend graph is generated through a visualization tool.

9. The method according to claim 1, characterized in that The process of fusing the disease trend dataset and the inclination change dataset, generating a dynamic risk heat map through multi-parameter weighted analysis, marking high-risk areas, and storing the result as a risk heat map dataset includes: Based on the disease trend dataset and the inclination change dataset, outliers are removed through data cleaning to obtain a standardized dataset; Extracting disease trend characteristics and inclination change characteristics from the standardized data set through multi-parameter analysis to determine a feature set; Using a weighted analysis method, the feature set is weighted according to preset parameter weights to obtain a risk score matrix; Generate a dynamic risk heat map based on the risk scoring matrix and mark the risk areas; If the risk score of the risk area exceeds a preset threshold, it is marked as a high-risk area to obtain a high-risk marked data set; then the high-risk marked data set and the dynamic risk heat map are stored as a risk heat data set; The regional division algorithm is used to perform spatial clustering on the risk thermal dataset to determine the final high-risk area distribution.

10. The method according to claim 1, characterized in that The process of obtaining inspection results includes: By using the risk thermal data set and the target 3D model, a data fusion algorithm is used to generate a digital twin model that includes the structural status and risk distribution; Based on the digital twin model, a virtualized tunnel environment is generated using three-dimensional rendering technology; Using VR / AR technology, a roaming inspection path is obtained in a virtual tunnel environment, and visual data of each part of the tunnel is determined; For the visualization data of various parts of the tunnel, if the data deviation exceeds the preset threshold, a convolutional neural network algorithm is used to identify the abnormal area and an automatic labeling algorithm is used to obtain the coordinates and labels of the abnormal area to generate the labeled tunnel part data; Based on the marked tunnel area data, historical data comparison results are obtained and time series analysis algorithms are used to determine the change trend of the abnormal area; Based on the changing trends, inspection results are generated and the priority ranking of abnormal areas is obtained.

Images