A tunnel structure disease identification method fusing laser point cloud and visible light

Abstract

The application relates to the technical field of tunnel disease identification, and discloses a tunnel structure disease identification method fusing laser point clouds and visible light, which comprises the following steps: constructing a space projection model to obtain a fusion data set; extracting a bidirectional coupling feature vector and screening a structure disease response factor; performing disease grade quantitative evaluation and constructing a three-dimensional space distribution atlas; training a structure deterioration trend prediction model to obtain a future expansion state and a bearing capacity change trend; and calculating a grade transition probability and outputting an early warning report containing positioning and disposal suggestions. The application can realize accurate alignment of multi-modal data by using a registration error compensation operator, deeply excavate deformation and damage correlation, eliminate redundant information, realize dynamic quantification of disease grades and early prediction of deterioration trends, and therefore improves the intelligent level and safety of tunnel operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of tunnel defect identification technology, and more specifically, to a method for identifying tunnel structural defects by integrating laser point clouds and visible light. Background Technology

[0002] With the deepening demand for intelligent operation and maintenance of tunnel structures, especially in areas such as multimodal sensing data spatial fusion, dynamic assessment of structural defects, and prediction of deterioration trends, the need to establish a systematic and complete technical chain is becoming increasingly urgent. Although traditional defect identification and risk assessment methods have made some progress in visible light image semantic segmentation, laser point cloud geometric feature anomaly detection, and static risk scoring based on human experience, they have not yet systematically solved core problems such as spatial projection alignment and bidirectional coupling feature extraction between laser point clouds and visible light images, accurate screening of defect sensing factors based on classifier output confidence, and dynamic quantification of structural defect levels driven by sensing factors. These issues make it difficult to meet the actual needs of high-precision identification and full-cycle risk management of tunnel structural defects in complex service environments.

[0003] Therefore, a method for identifying and assessing tunnel structure defects that integrates multimodal data from laser point clouds and visible light images, deeply mines the bidirectional coupling relationship of cross-modal features, accurately extracts structural defect sensing factors and achieves dynamic quantification of defect levels, and has the comprehensive output capability of collaborative prediction of deterioration trends and graded early warning has become an urgent technical problem to be solved. Summary of the Invention

[0004] This invention provides a method for identifying tunnel structure defects by integrating laser point cloud and visible light, which solves the technical problems in the prior art such as the lack of multimodal data spatial projection alignment and bidirectional coupling feature extraction mechanism, insufficient ability to accurately screen defect sensing factors, and lack of ability to dynamically quantify structural defect levels and collaboratively predict deterioration trends.

[0005] This invention provides a method for identifying tunnel structure defects by fusing laser point clouds and visible light, including: Obtain historical monitoring data of the tunnel structure; Based on the monitoring data, a spatial projection model is constructed to obtain a fused dataset; Based on the fused dataset, a bidirectional coupled feature vector is obtained; The bidirectional coupled feature vector is input into a preset multi-category disease classifier. Based on the confidence distribution of each disease type output by the classifier, the contribution of each dimension feature to the disease discrimination result is calculated, and a ranking vector of feature contribution is generated. Based on the sorting vector and the preset feature accumulation contribution threshold, the smallest feature subset whose cumulative contribution meets the contribution threshold is selected; the feature components in the smallest feature subset are the structural disease sensing factors. Based on the aforementioned structural disease sensing factors, the structural disease level is obtained; Based on the time-series acquisition sequence of the monitoring data and the evolution record of the structural disease level, a structural degradation sample library is constructed, and a structural deterioration trend prediction model is trained and obtained using a multi-task learning strategy. Real-time acquisition of current monitoring data of the tunnel structure; based on the structural defect sensing factors corresponding to the current monitoring data and the structural deterioration trend prediction model, the future expansion state of various defects and the trend of structural bearing capacity change are obtained. Based on the trend of structural bearing capacity change, the probability of structural defect level transition is obtained; When the transition probability exceeds the preset safety margin, a comprehensive early warning report on structural defects, including precise location of the defect, level assessment, and maintenance and treatment suggestions, is output in the form of a graded early warning.

[0006] Furthermore, based on the monitoring data, a spatial projection model is constructed, and an external parameter calibration matrix and a registration error compensation operator are introduced to obtain a fused dataset; Based on the fused dataset, the normal vector deviation, local curvature distribution, and reflection intensity signal-to-noise ratio of the laser point cloud are extracted to obtain a geometric anomaly probability map, which serves as prior information for the spatial structure. Based on the prior information of the spatial structure, a region of interest mask is defined to obtain the texture semantic features of the visible light image; Based on the prior information of the spatial structure and the semantic features of the texture, a two-way coupling analysis model is constructed to obtain the two-way coupling feature vector.

[0007] Further, based on the sorting vector and the preset feature accumulation contribution threshold, the smallest feature subset whose cumulative contribution meets the contribution threshold is selected; the feature components in the smallest feature subset are the structural disease sensing factors, including: The features of each dimension in the sorting vector are sorted in descending order of contribution value to obtain an ordered list of feature importance. The contribution values ​​of each feature are accumulated sequentially according to the ordered list to calculate the real-time cumulative contribution. The real-time cumulative contribution is compared with a preset feature cumulative contribution threshold, and the first feature position that causes the cumulative contribution to reach the threshold is determined as the cutoff point. Extract all features up to and including the cutoff point to form a minimal feature subset; the feature components in the minimal feature subset are the structural disease sensing factors.

[0008] Furthermore, based on the aforementioned structural defect sensing factors, a three-dimensional spatial distribution map of tunnel lining defects is constructed to obtain a dynamic evaluation matrix of structural defects in space. Based on the dynamic evaluation matrix, a multi-factor weighted comprehensive scoring model is constructed to obtain the comprehensive structural damage index, and the safety risk classification threshold is determined according to the safety assessment standard. Based on the structural defect sensing factor and the pre-acquired tunnel lining geometric deformation parameters, the defect expansion trend vector and the structural bearing capacity attenuation rate are obtained. Based on the disease expansion trend vector and the structural bearing capacity attenuation rate, a structural health status degradation curve is constructed to obtain the structural damage deterioration coefficient. The structural damage level is obtained based on the safety risk classification threshold and the structural damage deterioration coefficient.

[0009] Furthermore, based on the aforementioned structural defect sensing factors, a three-dimensional spatial distribution map of tunnel lining defects is constructed to obtain a dynamic evaluation matrix of structural defects in space, including: By associating each structural defect sensing factor with the corresponding three-dimensional point cloud coordinates in the fused dataset, the position information of each sensing factor in the tunnel lining space is obtained. Based on the location information, the three-dimensional spatial clustering of the lining surface disease response area is performed to delineate the spatial boundary of each independent disease area. For each diseased area, calculate the geometric dimensions such as the area, depth, and contour curvature of the diseased area; By combining the spatial location, geometric parameters, and disease type annotations of each disease area, a three-dimensional spatial distribution map of tunnel lining diseases is constructed. Based on the three-dimensional spatial distribution map, the geometric expansion rate and spatial density distribution changes of each disease area under different acquisition time sequences are calculated to generate a dynamic evaluation matrix of structural disease space.

[0010] Furthermore, the characteristic component changes of each structural disease sensing factor under multiple acquisition time series are extracted, and the unit time expansion increment of each disease type is calculated to obtain the disease expansion rate sequence. Based on the disease expansion rate sequence, the spatial expansion direction and expansion magnitude of each disease area are fitted to construct a disease expansion trend vector. Based on the pre-acquired geometric deformation parameters of the tunnel lining, the equivalent stiffness loss of the lining section in each disease-affected area is calculated. The equivalent stiffness loss is then substituted into the structural mechanics model to obtain the structural bearing capacity attenuation rate. Using the disease expansion trend vector as the input variable and the structural bearing capacity attenuation rate as the response variable, a mapping relationship between the two is established in the time-series coordinate system, and a structural health state degradation curve is fitted and constructed. Based on the structural health degradation curve, the slope and cumulative area of ​​the curve at the current moment are calculated to comprehensively characterize the degradation rate and cumulative damage degree of the structural state, and the structural damage deterioration coefficient is obtained.

[0011] Furthermore, based on the safety risk classification threshold and the structural damage deterioration coefficient, the structural defect level is obtained, including: The safety risk classification threshold is divided into several risk level intervals according to the boundary values ​​of continuous intervals, and a mapping relationship between each threshold interval and the structural defect level label is established. Using the total range of the risk level interval as the normalization benchmark, the structural damage deterioration coefficient is normalized to obtain the standardized deterioration index. The standardized degradation index is compared step by step with the upper and lower boundary values ​​of each risk level interval to determine the target risk level interval in which the standardized degradation index falls. When the standardized deterioration index falls near the boundary of an adjacent level, the membership function is introduced to calculate the degree of membership of the standardized deterioration index to the two adjacent risk levels, with the target risk level interval and its adjacent intervals as references. The final classification level is determined by the weighted comprehensive evaluation result. Based on the target risk level range and the final attribution level, and in conjunction with the mapping relationship, the structural disease level is obtained.

[0012] Furthermore, real-time monitoring data of the tunnel structure is acquired; based on the structural defect sensing factors corresponding to the current monitoring data and the structural deterioration trend prediction model, the future expansion state of various defects and the changing trend of structural bearing capacity are obtained, including: According to the preset acquisition cycle, the laser scanning device and the visible light imaging device are driven by synchronous triggering commands to collect laser point cloud data and visible light image data of the current tunnel lining surface in real time, and obtain the current monitoring data. Following the historical monitoring data processing flow, spatial projection, geometric feature extraction, and bidirectional coupling analysis are sequentially performed on the current monitoring data to obtain the bidirectional coupling feature vector corresponding to the current monitoring data. The bidirectional coupled feature vector is input into a preset multi-category disease classifier, and the structural disease sensing factor corresponding to the current monitoring data is extracted based on the gradient weighted feature importance algorithm and the feature cumulative contribution threshold. The structural disease sensing factor is input into the disease expansion prediction branch of the structural deterioration trend prediction model to obtain the future expansion state of various diseases in the prediction time domain. The structural disease sensing factors are synchronously input into the risk level prediction branch of the structural deterioration trend prediction model, and combined with the future expansion status of various diseases, the trend of structural bearing capacity change is obtained.

[0013] Furthermore, the trend of the structural bearing capacity change is differentiated over time to calculate the gradient of the bearing capacity decay rate in the predicted time domain, thereby obtaining the bearing capacity accelerated degradation index. Based on the aforementioned load-bearing capacity accelerated degradation index, and combined with the statistical laws governing the leap in the level of historical structural defects, a probability distribution function for the leap in the level is constructed. Substitute the trend of structural bearing capacity change into the probability distribution function of grade transition, calculate the probability of each structural defect grade transition in the prediction time domain, and obtain the structural defect grade transition probability. The transition probability of the structural disease level is compared with the preset safety margin level by level to determine the disease area where the transition probability exceeds the safety margin and its trigger warning level. Based on the aforementioned trigger warning level, combined with the precise location information of the disease in the three-dimensional spatial distribution map and the conclusion of the structural disease level assessment, maintenance and treatment recommendations are generated. The system integrates the precise location of the disease, the level assessment conclusions, and the maintenance and treatment recommendations in the form of a tiered early warning system, and outputs a comprehensive early warning report on structural diseases.

[0014] In a second aspect, a computer-readable storage medium is characterized in that it is used to store computer-readable instructions, which, when read by a computer, enable the execution of the method for identifying tunnel structure defects by fusing laser point clouds and visible light.

[0015] The beneficial effects of this invention are as follows: By constructing a spatial projection model that introduces an external parameter calibration matrix and a registration error compensation operator, this invention achieves sub-pixel-level alignment between laser point clouds and visible light images in a unified geometric semantic space. Based on this, a bidirectional coupling analysis mechanism is established that guides image segmentation with spatial structure priors and enhances point cloud intensity through texture semantics, effectively improving the collaborative utilization level of cross-modal features. Through a gradient-weighted feature importance algorithm and a cutoff point mechanism driven by a cumulative contribution threshold, structural disease sensing factors are accurately screened based on the classifier output confidence level, systematically eliminating redundant sensing features. Driven by sensing factors, a dynamic evaluation matrix, multi-factor weighted comprehensive scoring, and limited... The system employs a meta-stress analysis approach and introduces a trapezoidal membership function to perform fuzzy comprehensive evaluation of the grade boundaries, achieving dynamic quantification of structural defect grades. Based on a multi-task learning strategy, it simultaneously trains defect expansion prediction branches and risk grade prediction branches, enabling collaborative prediction of defect spatial expansion and structural bearing capacity trends. By fitting historical transition statistical laws with a Weibull distribution to construct a grade transition probability distribution function, it transforms traditional qualitative judgment into a probabilistic graded early warning triggering mechanism. Furthermore, it integrates and outputs precise defect location, grade assessment, and maintenance recommendations, forming a closed-loop management capability from data perception to decision support. This meets the practical needs of high-precision identification and full-cycle risk management of tunnel structural defects in complex service environments. Attached Figure Description

[0016] Figure 1 This is a schematic flowchart of a method for identifying tunnel structure defects by fusing laser point clouds and visible light, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the hardware structure of a computing device provided in an embodiment of the present invention. Detailed Implementation

[0017] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0018] At least one embodiment of the present invention discloses a method for identifying tunnel structure defects by fusing laser point clouds and visible light, comprising: like Figure 1 As shown, a method for identifying tunnel structure defects by fusing laser point clouds and visible light includes the following steps: Step 1: Obtain historical monitoring data of the tunnel structure; based on the monitoring data, construct a spatial projection model to obtain a fused dataset; Step 2: Based on the fused dataset, obtain the bidirectional coupling feature vector; input the bidirectional coupling feature vector into a preset multi-category disease classifier, calculate the contribution of each dimension feature to the disease discrimination result according to the confidence distribution of each disease type output by the classifier, and generate a sorting vector of feature contribution. Step 3: Based on the sorting vector and the preset feature accumulation contribution threshold, filter the smallest feature subset whose cumulative contribution meets the contribution threshold; the feature components in the smallest feature subset are the structural disease sensing factors; based on the structural disease sensing factors, obtain the structural disease level; Step 4: Based on the time-series acquisition sequence of the monitoring data and the evolution record of the structural disease level, construct a structural degradation sample library, and train and obtain a structural deterioration trend prediction model using a multi-task learning strategy. Step 5: Acquire the current monitoring data of the tunnel structure in real time; based on the structural defect sensing factor corresponding to the current monitoring data and the structural deterioration trend prediction model, obtain the future expansion state of various defects and the trend of structural bearing capacity changes; Step 6: Based on the trend of structural bearing capacity change, obtain the probability of structural defect level transition; when the transition probability exceeds the preset safety margin, output a comprehensive structural defect early warning report in the form of graded early warning, which includes precise defect location, level assessment and maintenance treatment suggestions.

[0019] This invention provides a method for identifying tunnel structural defects by integrating laser point cloud and visible light. In practical applications, this method relies on an inspection platform integrating a high-precision laser scanner and a high-resolution industrial camera. Through a full-coverage scan of the tunnel lining surface, it achieves digital perception of the structural condition. First, it acquires historical monitoring data of the tunnel structure, a process involving the simultaneous acquisition of 3D point cloud data streams and visual image sequences. A preset time synchronization trigger ensures that the laser scanner records spatial coordinates while the industrial camera completes exposure imaging of the corresponding area. The acquired historical monitoring data is stored for later retrieval. The 3D point cloud data includes the spatial 3D coordinates of each measurement point, the reflection intensity value, and the system attitude information at the time of scanning. The visual image sequence contains high-resolution RGB color information.

[0020] After acquiring historical monitoring data, a spatial projection model is constructed, and a fused dataset is obtained through coordinate system transformation. First, the relative position parameters of the laser scanning device and the visible light imaging device are acquired, including translation vectors and rotation matrices, to construct an external parameter calibration matrix. To eliminate projection misalignment caused by equipment vibration or installation deviations, a registration error compensation operator is introduced. This operator dynamically corrects the external parameter calibration matrix by calculating the reprojection error between feature points in the point cloud and corresponding pixels in the image. Specifically: Where e is the reprojection error value, p is the pixel coordinates of the corresponding feature point in the image, π() is the perspective projection function that projects the 3D point onto the 2D image plane, T is the extrinsic parameter calibration matrix (including rotation matrix and translation vector), and P is the 3D spatial coordinates of the feature point in the laser point cloud. The Euclidean norm is represented by ; by minimizing e and iteratively correcting T, the reprojection error is controlled within 0.5 pixels, achieving sub-pixel alignment between the laser point cloud coordinate system and the image pixel coordinate system. Based on this, the reflection intensity information of the 3D point cloud is mapped to the corresponding 2D image pixels, generating a fused dataset containing spatial coordinates, reflection intensity, and color information. Each data point in this dataset possesses six-dimensional attributes, namely 3D geometric coordinate information and 3D color information, providing a multimodal data foundation for subsequent feature mining.

[0021] Feature mining is performed based on the fused dataset to obtain bidirectional coupled feature vectors. First, the normal vector deviation, local curvature distribution, and reflection intensity signal-to-noise ratio of the laser point cloud in the fused dataset are extracted. The normal vector deviation is determined by extracting feature vectors from the covariance matrix of the point set in the local neighborhood, and is used to characterize the smoothness of the lining surface; the local curvature distribution reflects the degree of geometric deformation of the structure. A geometric anomaly probability map is constructed using these parameters: when the rate of change of the normal vector at a sampling point exceeds 5° (0.087 rad) or the absolute value of the local curvature exceeds 0.05 m... -1 If the reflection intensity signal-to-noise ratio is below 3dB, the point is marked as a geometric anomaly. The geometric anomaly probability of each point is calculated based on the above three judgment results: ;where P anom(i) Let κ be the geometric anomaly probability of the i-th sampling point, Δni be the rate of change of the normal vector at that point (in degrees), and κ be the geometric anomaly probability. i The absolute value of local curvature (unit: m) -1 ), SNR i Let I be the signal-to-noise ratio of the reflection intensity (in dB), and let I() be the indicator function (value 1 if the condition in parentheses is true, otherwise 0). The mean of the three indicator values ​​is the geometric anomaly probability of the point, and its value range is {0, 1 / 3, 2 / 3, 1}. Generate a geometric anomaly probability map and use this map as prior information of the spatial structure.

[0022] Subsequently, based on the prior information about the spatial structure, a region of interest mask is defined on the visible light image. This mask can pinpoint regions with severe geometric deformation or abnormal reflection intensity. Within the masked region, a dual-branch convolutional network with ResNet as its backbone is used to extract texture semantic features from the visible light image: the backbone network outputs multi-scale convolutional feature maps layer by layer, and the geometric anomaly probability map is upsampled to the same spatial resolution as the feature maps of each layer through bilinear interpolation, and then element-wise multiplied with the feature maps of each layer. ;where F' l For the weighted feature map of the l-th layer, F l This is the convolutional feature map output by the l-th layer of the backbone network, where ⊙ represents element-wise multiplication, U pl () indicates that the bilinear interpolation of the geometric anomaly probability map is upsampled to the level of F. l For operations with the same spatial resolution, Panom is the geometric anomaly probability map calculated above. The geometric anomaly probability value is used as the spatial attention weight to weight the feature response at each location, thereby transforming the geometric prior information into a spatial attention constraint on the intermediate layer features of the image segmentation network. This enables the network to have a higher feature activation response in the geometric anomaly region, realizing the quantitative guidance of spatial structure prior on the edge segmentation of visible light images, and effectively correcting the drift of the disease boundary under low illumination environment or water stain interference. The weighted feature map is decoded by subsequent convolutional layers to output a texture semantic feature map containing crack direction, seepage contour and peeling area edge, and the semantic embedding vector of each pixel position is extracted as the texture semantic descriptor of that position.

[0023] Based on the texture semantic feature map, a semantic enhancement operator is performed on each 3D point cloud sampling point in the fused dataset: according to the 2D image pixel coordinates corresponding to the 3D point in the fused dataset, the semantic embedding vector of the pixel position is retrieved from the texture semantic feature map, the semantic embedding vector is transformed to the same dimension as the point cloud intensity feature through a lightweight mapping network consisting of two fully connected layers, and weighted channel concatenation is performed with the original reflection intensity value of the point. The disease category confidence distribution carried by the semantic embedding vector is used as the recalibration coefficient of each intensity channel, and the semantically corrected enhanced intensity representation is output, thereby distinguishing disease categories with the same geometric shape but different causes in the intensity feature space.

[0024] When constructing the bidirectional coupling analysis model, a two-branch cross-attention fusion architecture is adopted: the geometric feature branch takes the normal vector deviation, local curvature, and semantically enhanced intensity representation of the point cloud as input, and encodes them into geometric feature vectors through a multilayer perceptron (MLP); the semantic feature branch takes the mean pooling result of the texture semantic feature map within the mask of the region of interest as input, and encodes it into semantic feature vectors through a fully connected layer; the feature vectors of the two branches are mutually queried and aggregated through a cross-attention layer, so that the geometric branch feature vector focuses on the image region response consistent with its semantics, and the semantic branch feature vector focuses on the geometric anomaly intensity signal corresponding to its space, thus fully establishing the correspondence between the two types of modal features; finally, the two enhanced feature vectors output by the cross-attention layer are dimensionally concatenated to obtain a bidirectional coupling feature vector representing the dual attributes of structural deformation and surface damage.

[0025] The bidirectional coupled feature vector is input into a pre-defined multi-class disease classifier. The construction steps of the multi-class disease classifier are as follows: using ResNet-50 as the backbone network, the output dimension of the last fully connected layer is adjusted to the number of disease categories; fine-tuning training is performed on a historical monitoring sample database labeled with various diseases such as cracks, water leakage, lining spalling, and voids using a transfer learning strategy, with the cross-entropy loss function as the optimization objective during training, and the initial learning rate set to 1×10⁻⁶. -4 The weights are decayed using a cosine annealing strategy, with at least 100 training rounds. The convergence criterion is a classification accuracy of over 90% on the test set, and the optimal weight parameters are saved for later use.

[0026] The bidirectional coupled feature vector is input into the classifier to perform forward inference. After softmax activation at the output layer, the confidence distribution of each disease category is obtained, i.e., the output probability value corresponding to each category. Then, a gradient-weighted feature importance algorithm is used to calculate the contribution of each dimension feature. The specific calculation process is as follows: For each disease category, backpropagation is performed on the output confidence value of that category with respect to each dimension feature in the input feature vector. The partial derivative of the output confidence value with respect to that feature dimension is calculated. The absolute value of the obtained partial derivative represents the sensitivity of the feature dimension change to the category classification decision. To comprehensively reflect the overall contribution of the feature dimension in the discrimination of all disease categories, the output confidence value of each category is used as a weighting coefficient. The absolute values ​​of the partial derivatives of the feature dimension with respect to all categories are weighted and summed to obtain the weighted gradient value of the confidence of that feature dimension. This transforms the multi-category confidence distribution into a single contribution of each dimension feature. Quantitative metrics: Since the bidirectional coupled feature vector contains geometric features such as normal vector deviation, local curvature, and enhancement intensity from laser point clouds and semantic features such as texture semantic embedding from visible light images, the physical dimensions and numerical scales of the two types of features are significantly different. Directly comparing the absolute values ​​of gradients cannot reflect the relative importance between each dimension. Therefore, the confidence-weighted gradient values ​​of all feature dimensions are subjected to L1 normalization, that is, the weighted gradient value of each dimension is divided by the sum of the weighted gradient values ​​of all dimensions, which is converted into the relative percentage of each dimension's contribution to the classification result. Thus, in a unified dimensionless contribution space, the horizontal comparability of the contribution of different attribute features is achieved, and the ranking vector of the contribution of each dimension feature is output.

[0027] Based on the sorting vector and a preset feature cumulative contribution threshold (95%), the features in each dimension of the sorting vector are sorted in descending order of contribution percentage to construct an ordered list of feature importance. The contribution percentage of each feature is accumulated sequentially according to the ordered list to calculate the real-time cumulative contribution. This value is compared with the 95% feature cumulative contribution threshold to determine the first feature position that causes the cumulative contribution to reach the threshold as the cutoff point. All features up to and including the cutoff point are extracted to form a minimum feature subset. The feature components in this minimum feature subset are the structural disease sensing factors. The above algorithm uses the classifier's own gradient response as an objective evaluation criterion. Through two steps—confidence distribution weighting and cross-modal L1 normalization—it establishes a complete calculation path from the confidence distribution to the contribution values ​​of each dimension, achieving quantitative removal of redundant sensing features and ensuring that subsequent calculations mainly focus on the features most sensitive to disease manifestations.

[0028] Based on the aforementioned structural defect sensing factors, a three-dimensional spatial distribution map of tunnel lining defects is constructed to obtain a dynamic evaluation matrix for the structural defect space. Specifically, each structural defect sensing factor is associated with the corresponding three-dimensional point cloud coordinates in the fused dataset to obtain the precise location information of each sensing factor in the tunnel lining space. Based on this location information, the DBSCAN density clustering algorithm is used to perform three-dimensional spatial clustering of the defect response areas on the lining surface. The neighborhood radius ε ranges from 0.05 to 0.10 m, and the minimum number of neighborhood points MinPts is set to 5 to delineate the spatial boundaries of each independent defect area. For each defect area, geometric dimensional parameters such as area, depth, and contour curvature are calculated. By integrating the spatial location, geometric parameters, and defect type labels of each defect area, a three-dimensional spatial distribution map of tunnel lining defects is constructed. Based on the three-dimensional spatial distribution map, the geometric expansion rate and spatial density distribution changes of each defect area under different acquisition time sequences are calculated to generate a dynamic evaluation matrix for the structural defect space.

[0029] Based on the dynamic evaluation matrix, a multi-factor weighted comprehensive scoring model is constructed to obtain the comprehensive structural damage index, and the safety risk classification threshold is determined according to the safety assessment standards. The construction steps of the multi-factor weighted comprehensive scoring model are as follows: using the disease area, disease depth, geometric expansion rate, and spatial density as scoring indicators, the hazard weight of each indicator is determined by the analytic hierarchy process (AHP), with weights of 0.35, 0.30, 0.20, and 0.15, respectively; each indicator is normalized according to the unified dimensional standard and mapped to the [0,1] interval; the comprehensive structural damage index is calculated by weighted summation: the normalized score of disease area is multiplied by a weight of 0.35, the normalized score of disease depth by a weight of 0.30, the normalized score of geometric expansion rate by a weight of 0.20, and the normalized score of spatial density by a weight of 0.15, and the four results are summed to obtain the comprehensive structural damage index, which ranges from 0 to 1. According to the tunnel structure safety assessment standard, the safety risk classification thresholds are defined as follows: a comprehensive structural damage index between 0 and 0.25 (excluding 0.25) corresponds to Level 1 (intact); between 0.25 and 0.50 (excluding 0.50) corresponds to Level 2 (minor damage); between 0.50 and 0.75 (excluding 0.75) corresponds to Level 3 (moderate damage); and between 0.75 and 1.00 (including both ends) corresponds to Level 4 (severe damage).

[0030] The characteristic component changes of each structural defect sensing factor under multiple acquisition time series are extracted, and the unit time expansion increment of each defect type is calculated to obtain the defect expansion rate sequence. Based on the defect expansion rate sequence, the spatial expansion direction and expansion magnitude of each defect area are fitted to construct a defect expansion trend vector. Based on the pre-acquired tunnel lining geometric deformation parameters, a finite element stress analysis model is constructed: the lining section is discretized with shell elements, and the initial value of the elastic modulus is set according to the reinforcement drawings and concrete strength grade in the tunnel design stage; according to the depth and area of ​​each defect area, the elastic modulus of the corresponding element is reduced in the finite element model. ; where E damE0 is the initial elastic modulus of the lining material under the influence of defects (taken according to the design concrete strength grade), α is the defect reduction factor (reflecting the degree of weakening of the cross-sectional stiffness by defects, with a value ranging from 0.3 to 0.8, determined according to the type and development degree of defects), h is the defect depth (taken as the measured maximum depth in the defect area), and H is the lining design thickness (taken as the design value of the corresponding section in the tunnel structure design document). The equivalent stiffness loss of the lining section in each defect-affected area is calculated. The elastic modulus after damage is equal to the initial elastic modulus multiplied by the reduction factor. The reduction factor is equal to 1 minus the product of the defect reduction factor and the ratio of the defect depth to the lining design thickness. The defect depth and the lining design thickness are both taken as the measured values ​​of each defect area. The disease reduction coefficient is determined based on the disease type label and disease level output by the multi-category disease classifier. The specific value criteria are as follows: For crack-type diseases, when the crack width does not exceed 0.2 mm and the depth does not exceed 20% of the lining thickness, the disease reduction coefficient is 0.30, corresponding to level 1 to level 2 mild diseases; when the crack width is between 0.2 mm and 0.5 mm or the depth is between 20% and 40% of the lining thickness, the disease reduction coefficient is 0.45, corresponding to level 2 to level 3 moderate diseases; when the crack width exceeds 0.5 mm or the depth exceeds 40% of the lining thickness, the disease reduction coefficient is 0.65, corresponding to level 3 to level 4 severe diseases. For water seepage defects, the reduction factor is 0.35 when the seepage is damp or point-like; 0.50 when the seepage is linear and the affected area does not exceed 0.5 m²; and 0.60 when the seepage is large-area gushing or accompanied by crystallization. For spalling defects, the reduction factor is 0.50 when the spalling depth does not exceed the protective layer thickness and the area does not exceed 0.1 m²; 0.65 when the spalling depth exceeds the protective layer thickness or the area exceeds 0.1 m²; and 0.80 when the spalling exposes the reinforcing steel or the area exceeds 0.5 m². For void defects, the reduction factor is uniformly taken as 0.75 to 0.80 regardless of the area size, to conservatively estimate the degree of weakening of the overall stiffness due to localized load-bearing loss in the lining. When multiple types of defects overlap in the same defect area, the maximum value among the reduction coefficients of each type of defect is taken as the comprehensive defect reduction coefficient for that area to ensure that the bearing capacity assessment result is biased towards the safe side. After determining the defect reduction coefficient for each defect area according to the above criteria, the equivalent stiffness loss of the lining section in each defect-affected area is calculated. The equivalent stiffness loss is substituted into the structural mechanics model to obtain the structural bearing capacity attenuation rate. Using the defect expansion trend vector as the input variable and the structural bearing capacity attenuation rate as the response variable, a mapping relationship between the two is established in the time-series coordinate system, and a structural health state degradation curve is fitted and constructed.Based on the structural health degradation curve, the slope and cumulative area of ​​the curve at the current moment are calculated to comprehensively characterize the degradation rate and cumulative damage degree of the structural state, and the structural damage deterioration coefficient is obtained.

[0031] The structural damage level is obtained based on the safety risk classification threshold and the structural damage deterioration coefficient. Specifically, the safety risk classification thresholds are divided into four risk level intervals based on four continuous intervals: [0, 0.25), [0.25, 0.50), [0.50, 0.75), and [0.75, 1.00]. A mapping relationship is established between each threshold interval and the structural damage level labels from level one to level four. Using the total value range [0, 1] of the risk level intervals as the normalization benchmark, the structural damage deterioration coefficient is normalized to obtain a standardized deterioration index. The standardized deterioration index is compared step by step with the upper and lower boundary values ​​of each risk level interval to determine the target risk level interval into which the standardized deterioration index falls. For standardized deterioration indices falling near the boundary of adjacent levels (within ±0.05), the target risk level interval and its adjacent intervals are used as references. A trapezoidal membership function is introduced to calculate the degree of membership of the standardized deterioration index to two adjacent risk levels. The final classification level is determined by a weighted comprehensive evaluation result. Based on the target risk level interval and the final classification level, combined with the mapping relationship, the structural damage level is obtained.

[0032] Based on the time-series acquisition sequence of the historical monitoring data and the evolution record of the structural disease level, a structural degradation sample library is constructed. Historical maintenance records are automatically associated, and areas that have taken reinforcement measures are marked as non-natural degradation samples, so as to correct the model parameters to make them more consistent with the actual physical evolution law of the structure. The construction steps of the structural deterioration trend prediction model are as follows: A bidirectional long short-term memory network (Bi-LSTM) is used as the basic architecture, with a hidden layer dimension of 128 and a stacking layer of 2 layers. Historical sensing factor time series (with a time window length of 12 acquisition cycles) is used as input, and two output branches are set: a disease expansion prediction branch and a risk level prediction branch. A multi-task joint loss function is adopted, consisting of a weighted sum of the disease expansion prediction branch loss and the risk level prediction branch loss: the weight coefficient of the disease expansion prediction branch loss is 0.6, and the weight coefficient of the risk level prediction branch loss is 0.4; the weighted sum of the two is the overall optimization objective. The parameters are updated using the Adam optimizer, with an initial learning rate of 1×10⁻³. The structural deterioration trend prediction model is trained on the validation set with a root mean square error (RMSE) below 0.05 as the convergence criterion.

[0033] During the real-time monitoring phase, the laser scanning device and visible light imaging device are driven by synchronous trigger commands according to a preset collection cycle (no less than once per quarter) to collect laser point cloud data and visible light image data of the current tunnel lining surface in real time, obtaining the current monitoring data. Following the processing flow of historical monitoring data, spatial projection, geometric feature extraction, and bidirectional coupling analysis are sequentially performed on the current monitoring data to obtain the bidirectional coupling feature vector corresponding to the current monitoring data. The bidirectional coupling feature vector is input into a preset multi-category disease classifier, and based on a gradient-weighted feature importance algorithm and a 95% feature cumulative contribution threshold, structural disease sensing factors corresponding to the current monitoring data are extracted. The structural disease sensing factors are input into the disease expansion prediction branch of the structural deterioration trend prediction model to obtain the future expansion state of various diseases in the prediction time domain; simultaneously, the structural disease sensing factors are synchronously input into the risk level prediction branch, and combined with the future expansion state of various diseases, the trend of structural bearing capacity change is obtained.

[0034] The structural bearing capacity variation trend is processed by time-series differentiation to calculate the gradient of bearing capacity decay rate in the prediction time domain, thereby obtaining the bearing capacity accelerated degradation index. The construction steps of the grade transition probability distribution function are as follows: For each adjacent grade pair (i.e., level 1 to level 2, level 2 to level 3, level 3 to level 4), an independent transition probability model is established; the observed values ​​of the bearing capacity accelerated degradation index for each diseased area at each acquisition time series are extracted from the historical monitoring database, and it is marked whether a corresponding grade transition event has occurred between this time series and the next acquisition time series, forming a historical sample set with the bearing capacity accelerated degradation index as the independent variable and whether a transition has occurred as the dependent variable; the Weibull cumulative distribution function is used as the mapping function for the transition probability, and its functional form is: ;where P trans(A) Let P be the probability of grade transition when the bearing capacity acceleration degradation index is A, where A is the bearing capacity acceleration degradation index (i.e., the gradient of the bearing capacity decay rate in the prediction time domain), k is the shape parameter (controlling the growth rate of the probability curve; when k > 1, the curve shows an accelerated upward trend, reflecting the concentration of historical transition events), λ is the scale parameter (characterizing the eigenvalue of A corresponding to a transition probability of approximately 63.2%, having the same dimensions as A), and exp() is an exponential function with the natural constant e as its base; trans(A)The transition probability monotonically increases with increasing A; it equals 1 minus a power function with the natural constant as the base and the opposite exponent of the power of the ratio of the accelerated degradation index to the scale parameter. In other words, the transition probability monotonically increases with increasing accelerated degradation index. The shape parameter controls the growth rate of the probability curve, and the scale parameter represents the characteristic value of the accelerated degradation index corresponding to a transition probability of approximately 63.2%. Using historical sample sets of each adjacent level pair as input, maximum likelihood estimation is performed on the shape and scale parameters of the aforementioned Weibull mapping function. This involves maximizing the joint logarithmic probability of transition events occurring at each observation point in the sample set, solving for the shape and scale parameter values ​​that best fit the historical data, and solidifying the calibrated parameters into a dedicated parameter set for each level pair, forming the level transition probability distribution function for each adjacent level pair. In the real-time application stage, the real-time calculated accelerated degradation index is substituted into the Weibull mapping function of each level pair to directly output the probability of transition in each adjacent level, thus obtaining the level transition probability of structural defects. The probability of structural defects escalating to a higher level is compared step-by-step with a preset safety margin (70%) to identify defect areas where the probability exceeds the 70% safety margin and their trigger warning levels. Based on these trigger warning levels, the precise location information of the defects in the three-dimensional spatial distribution map is retrieved. Combined with the defect level assessment conclusion, a preset maintenance and treatment expert database is used to generate targeted maintenance and treatment recommendations. The maintenance and treatment expert database uses the combination of defect type and structural defect level as the rule retrieval key to store corresponding standardized treatment measures. Its logical rules are constructed as follows: For crack-type defects, if the structural defect level is level two, the recommendation is to regularly monitor crack width changes and establish a crack development ledger; if the level is level three and the crack width is less than 0.2 mm, the recommendation is to use epoxy resin low-pressure injection + surface sealing; if the level is level three and the crack width is not less than 0.2 mm, the recommendation is to use chemical grouting + flexible caulking and sealing; if the level is level four, the recommendation is to use carbon fiber cloth bonding reinforcement or steel plate reinforcement + chemical grouting. For water leakage-related defects, if the level is Level II, the recommended treatment is water diversion and drainage facility inspection; if the level is Level III, the recommended treatment is high-pressure grouting and surface sealing; if the level is Level IV, the recommended treatment is structural waterproofing repair, grouting to stop water leakage, and speed-limited operation control. For lining spalling-related defects, if the level is Level II, the recommended treatment is partial removal of loose layers, interface agent treatment, and mortar repair; if the level is Level III, the recommended treatment is removal of loose layers, shotcrete repair, and surface protective coating; if the level is Level IV, the recommended treatment is temporary support erection and overall lining repair or replacement. For void-related defects, if the level is Level III, the recommended treatment is drilling to detect the void area and backfilling with grout; if the level is Level IV, the recommended treatment is drilling and grouting filling, structural calculation, and, if necessary, removal and reconstruction.When multiple matching rules are triggered simultaneously in the same diseased area, the highest level of treatment intensity among all rules is taken as the final output. The precise location of the disease, the level assessment conclusion, and the maintenance treatment recommendations are integrated in the form of a tiered early warning system to output a comprehensive structural disease early warning report.

[0035] This invention achieves closed-loop management of tunnel defects from discovery to early warning through deep fusion of multimodal data and intelligent prediction of deterioration trends. By using a registration error compensation operator to control reprojection errors within 0.5 pixels, spatial deviations between different sensors are eliminated, ensuring high consistency of acquired features. Minimum feature subset extraction driven by a 95% cumulative contribution threshold ensures computational efficiency when processing massive amounts of monitoring data. A Bi-LSTM-based multi-task learning model not only provides the current defect level but also offers scientific predictions of future evolution trends, providing data support for maintenance decisions. A probabilistic level transition assessment driven by a 70% safety margin avoids the subjectivity of traditional qualitative assessments, significantly improving the intelligence level of tunnel safety operation and maintenance.

[0036] In executing the above method, multi-threaded parallel processing technology enables simultaneous data acquisition and preliminary processing. A distributed storage architecture supports rapid retrieval and access to massive amounts of monitoring data throughout the tunnel's entire lifecycle. The system interfaces with the tunnel management platform through standardized data interfaces, enabling automatic push notifications of early warning reports and closed-loop tracking of maintenance tasks. In complex tunnel environments, such as insufficient lighting or high dust levels, interference signals are filtered out by using a threshold of a reflection intensity signal-to-noise ratio below 3dB, ensuring the robustness of the identification results. The method and system of this invention are not only applicable to highway and railway tunnels but can also be extended to the field of defect monitoring and safety assessment of similar structures such as urban subways and underground utility tunnels.

[0037] For different types of tunnel linings, such as shotcrete or cast-in-place concrete linings, the judgment threshold of the geometric anomaly probability map is automatically adjusted based on the initial roughness of the lining surface (the normal vector change rate threshold is adaptively adjusted within the range of 3°-8°). When dealing with different types of defects, the weight coefficients of the multi-category defect classifier are dynamically adjusted; for example, in areas with frequent water seepage, the extraction weight of humidity-sensing features is increased. When predicting deterioration, changes in the hydrogeological conditions around the tunnel, such as the impact of groundwater level fluctuations on lining pressure, are also considered and input as external variables into the prediction model, further improving the prediction accuracy of the bearing capacity evolution curve. Through this comprehensive and multi-dimensional technical approach, this invention can provide the most accurate decision-making basis for the preventive maintenance of tunnels.

[0038] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principles of this invention are further supplemented below with a specific application scenario.

[0039] Step 1: In the application scenario of regular structural inspection of a highway tunnel, the inspection vehicle integrating a 3D laser scanner and a high-resolution CCD camera is first driven into the tunnel. An acquisition command is sent to the laser scanner via an internal pulse synchronizer, while an external trigger signal is sent to the CCD camera. The laser scanner acquires the spatial 3D coordinates and reflection intensity values ​​of the tunnel lining surface at a frequency of millions of points per second, while the CCD camera simultaneously captures the corresponding color image sequence. These raw data streams are written to the storage device in real time. A pre-calibrated external parameter calibration matrix is ​​invoked, which contains the rotation and translation components between the laser scanner coordinate system and the camera coordinate system. To eliminate the dynamic bias caused by vehicle vibration during driving, a registration error compensation operator is run. This operator identifies feature points such as lining joints and embedded bolts in the point cloud and performs feature matching with the corresponding pixels in the image to calculate the reprojection error, controlling the reprojection error to within 0.5 pixels. By minimizing the reprojection error and dynamically correcting the rotation and translation component parameters, each point in the 3D point cloud is accurately projected onto the corresponding 2D image pixel coordinates, generating a fused dataset containing six-dimensional attributes, thus achieving physical alignment of multimodal data in spatial dimensions.

[0040] Step 2: After obtaining the fused dataset, perform local neighborhood analysis on the point cloud data. Specifically, calculate the covariance matrix of each point within its neighborhood radius, and obtain the local curvature and rate of change of the normal vector through eigenvalue decomposition. When the rate of change of the normal vector in a certain region exceeds 5° (0.087 rad) or the absolute value of the curvature exceeds 0.05 m⁻¹, it is marked as a geometric anomaly point, and a 3D geometric anomaly probability map is generated as prior information of spatial structure. Based on this, the geometric anomaly probability map is upsampled through bilinear interpolation and projected back to the corresponding visible light image pixel coordinate system to generate a region of interest mask. The upsampled geometric anomaly probability map is then used as a spatial attention weight map for later use. Within the image region covered by the mask, a convolutional network with ResNet as the backbone is used to extract multi-scale feature maps layer by layer. The geometric anomaly probability weight map is multiplied element-wise with each layer's feature map, allowing geometric prior information to form a spatial attention constraint on the intermediate layer features of the CNN. This enhances the feature activation response of disease areas such as crack direction, seepage contours, and the edges of peeling areas, while suppressing background interference. The decoded output is a texture semantic feature map containing semantic labels and confidence distributions, and semantic embedding vectors are extracted from each pixel as texture semantic descriptors. For each 3D point in the fused dataset, a semantic embedding vector is retrieved from the texture semantic feature map based on its corresponding 2D pixel coordinates. The semantic embedding vector is transformed to the same dimension as the intensity feature through a two-layer fully connected mapping network. The semantic confidence distribution is used as the recalibration coefficient for each intensity channel, and the original reflection intensity value is weighted and concatenated to output a semantically corrected enhanced intensity representation, enabling diseases with similar geometric shapes but different causes to be distinguished in the intensity feature space. Finally, the geometric feature branch and the semantic feature branch are mutually queried and aggregated through the cross-attention layer. The two output feature vectors are concatenated in dimension to output a bidirectional coupled feature vector that represents the deep attributes of the disease.

[0041] Step 3: Input the extracted bidirectional coupled feature vector into the preset multi-category disease classifier. After receiving the feature input, the classifier performs weight mapping on the features of each dimension in the bottom layer, and after being activated by Softmax, generates a confidence distribution vector for disease types such as cracks, water leakage, and peeling in the output layer, which is the output probability value of each category. Subsequently, a gradient-weighted feature importance algorithm is used to calculate the contribution of each dimension of features: For each disease category, the partial derivative of the output probability of that category with respect to each dimension of the input feature vector is calculated. The absolute value of the partial derivative reflects the sensitivity of that dimension to the classification decision for that category. Using the output probability of each category as a weighting coefficient, the absolute values ​​of the partial derivatives of the same feature dimension on all categories are summed in a weighted manner to obtain the confidence-weighted gradient value of that dimension. Since there are differences in the dimensions and numerical scales between point cloud geometric features (such as curvature, normal vector deviation, etc.) and image semantic features (such as semantic embedding components, etc.) in the feature vector, L1 normalization is performed on the weighted gradient values ​​of all dimensions. That is, the weighted gradient value of each dimension is divided by the sum of the weighted gradient values ​​of all dimensions, which is converted into the relative percentage of the contribution of each dimension, realizing the unified and comparable quantification of the contribution between cross-modal features. All feature dimensions are sorted in descending order of contribution percentage to form an ordered list of feature importance. The contribution percentage is accumulated sequentially from the first element of the list, stopping when the accumulated value first reaches 95%. This position is determined as the cutoff point. All feature components up to the cutoff point are extracted to form the minimum feature subset. The feature components in this minimum feature subset are the structural disease sensing factors, thus eliminating the interference of environmental noise and redundant features on subsequent assessments based on quantitative calculation results.

[0042] Step 4: Obtain structural defect sensing factors and reconstruct the defects in a 3D virtual space using spatial coordinates from the fused dataset. The DBSCAN algorithm (ε=0.08m, MinPts=5) is used to associate discrete sensing factor points as independent defect entities. The volume, surface area, and maximum extension length of each entity's 3D bounding box are calculated to construct a 3D spatial distribution map of tunnel lining defects. Based on this, historical monitoring records for the same mileage section are retrieved to calculate the geometric expansion rate and spatial density distribution changes of each defect area under different acquisition time sequences, generating a dynamic evaluation matrix of the structural defect space. Using defect area (weight 0.35), depth (weight 0.30), expansion rate (weight 0.20), and spatial density (weight 0.15) as scoring indicators, a multi-factor weighted comprehensive scoring model is constructed to calculate the comprehensive structural damage index D. Safety risk classification thresholds are determined according to four intervals: [0,0.25), [0.25,0.50), [0.50,0.75), and [0.75,1.00]. Simultaneously, based on the structural distress induction factor and the geometric deformation parameters of the tunnel lining, the equivalent stiffness loss is calculated using a finite element stress analysis model to obtain the structural bearing capacity attenuation rate. A structural health state degradation curve is then fitted to obtain the structural damage deterioration coefficient. After normalizing the deterioration coefficient, it is compared step by step with the above four intervals. For the exponent near the boundary (±0.05), a trapezoidal membership function is introduced for weighted evaluation to output the current structural distress level.

[0043] Step 5: To achieve full-cycle risk management, a structural degradation sample library with spatiotemporal annotation information is constructed based on the time-series acquisition sequence of historical monitoring data and the evolution record of structural disease levels. A multi-task learning strategy is adopted, using Bi-LSTM as the foundation, with a hidden layer dimension of 128, a loss weight of 0.6 for the disease expansion prediction branch, and a loss weight of 0.4 for the risk level prediction branch. The disease expansion prediction branch and the risk level prediction branch are trained simultaneously to obtain a structural deterioration trend prediction model. During the real-time monitoring phase, dual sensors are driven to synchronously acquire current monitoring data according to a preset acquisition cycle. Structural disease sensing factors corresponding to the current monitoring data are extracted according to the historical data processing flow and input into the two prediction branches to obtain the future expansion state of various diseases and the changing trend of structural bearing capacity. The changing trend of structural bearing capacity is then subjected to temporal differentiation to obtain the accelerated degradation index of bearing capacity. Substituting the bearing capacity accelerated degradation index into the pre-calibrated Weibull mapping function for each adjacent level: This function uses the bearing capacity accelerated degradation index as the only independent variable. The shape parameter and scale parameter have been fixed in the model building stage through maximum likelihood estimation of historical transition samples. The scale parameter corresponds to the characteristic value of the accelerated degradation index when the transition probability in historical data reaches about 63.2%, and the shape parameter reflects the concentration of historical transition events. By directly substituting the real-time bearing capacity accelerated degradation index value into the above function, the probability of transition between each adjacent level at the current moment can be output, the level transition probability distribution function can be constructed, and the level transition probability can be calculated. If the transition probability exceeds the preset safety margin of 70%, the system immediately retrieves the precise location information and severity assessment conclusions of the defects from the 3D spatial distribution map. Using the combination of defect type and structural defect severity level as the search key, it matches the corresponding standardized treatment entry from the maintenance and treatment expert database. For example, if the defect area triggering the warning is classified as Level 3 and the defect type is water leakage, the system searches for and matches a treatment suggestion of high-pressure grouting + surface sealing; if the level is Level 4 and the defect type is cracking, the system matches a treatment suggestion of carbon fiber cloth bonding reinforcement + chemical grouting; if the level is Level 4 and the defect type is voids, the system matches a treatment suggestion of borehole grouting filling + structural verification. When multiple types of defects occur concurrently in the same area, the highest level of treatment intensity among all rules is taken as the final output. After generating maintenance and treatment suggestions, a comprehensive structural defect early warning report is output in the form of a graded early warning and sent to the monitoring terminal via the system bus, completing the closed-loop operation from data perception to decision support.

[0044] like Figure 2 As shown, a computer-readable storage medium is characterized in that it is used to store computer-readable instructions, which, when read by a computer, enable the execution of the method for identifying tunnel structure defects by fusing laser point clouds and visible light.

[0045] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for identifying tunnel structure defects by integrating laser point cloud and visible light, characterized in that, include: Obtain historical monitoring data of the tunnel structure; Based on the monitoring data, a spatial projection model is constructed to obtain a fused dataset; Based on the fused dataset, a bidirectional coupled feature vector is obtained; Based on the monitoring data, a spatial projection model is constructed, and an external parameter calibration matrix and a registration error compensation operator are introduced to obtain a fused dataset. Based on the fused dataset, the normal vector deviation, local curvature distribution, and reflection intensity signal-to-noise ratio of the laser point cloud are extracted to obtain a geometric anomaly probability map, which serves as prior information for the spatial structure. Based on the prior information of the spatial structure, a region of interest mask is defined to obtain the texture semantic features of the visible light image; Based on the prior information of the spatial structure and the semantic features of the texture, a two-way coupling analysis model is constructed to obtain the two-way coupling feature vector. The bidirectional coupled feature vector is input into a preset multi-category disease classifier. Based on the confidence distribution of each disease type output by the classifier, the contribution of each dimension feature to the disease discrimination result is calculated, and a ranking vector of feature contribution is generated. Based on the sorting vector and the preset feature accumulation contribution threshold, the smallest feature subset whose cumulative contribution meets the contribution threshold is selected; the feature components in the smallest feature subset are the structural disease sensing factors. Based on the aforementioned structural disease sensing factors, the structural disease level is obtained; Based on the time-series acquisition sequence of the monitoring data and the evolution record of the structural disease level, a structural degradation sample library is constructed, and a structural deterioration trend prediction model is trained and obtained using a multi-task learning strategy. Real-time acquisition of current monitoring data of the tunnel structure; based on the structural defect sensing factors corresponding to the current monitoring data and the structural deterioration trend prediction model, the future expansion state of various defects and the trend of structural bearing capacity change are obtained. Based on the trend of structural bearing capacity change, the probability of structural defect level transition is obtained; When the transition probability exceeds the preset safety margin, a comprehensive early warning report on structural defects, including precise location of the defect, level assessment, and maintenance and treatment suggestions, is output in the form of a graded early warning.

2. The method for identifying tunnel structure defects by fusing laser point clouds and visible light according to claim 1, characterized in that, Based on the sorting vector and the preset feature accumulation contribution threshold, the smallest feature subset whose cumulative contribution meets the contribution threshold is selected; the feature components in the smallest feature subset are the structural disease sensing factors, including: The features of each dimension in the sorting vector are sorted in descending order of contribution value to obtain an ordered list of feature importance. The contribution values ​​of each feature are accumulated sequentially according to the ordered list to calculate the real-time cumulative contribution. The real-time cumulative contribution is compared with a preset feature cumulative contribution threshold, and the first feature position that causes the cumulative contribution to reach the threshold is determined as the cutoff point. Extract all features up to and including the cutoff point to form a minimal feature subset; the feature components in the minimal feature subset are the structural disease sensing factors.

3. The method for identifying tunnel structure defects by fusing laser point clouds and visible light according to claim 1, characterized in that, Based on the aforementioned structural defect sensing factors, a three-dimensional spatial distribution map of tunnel lining defects is constructed to obtain a dynamic evaluation matrix of structural defects in space. Based on the dynamic evaluation matrix, a multi-factor weighted comprehensive scoring model is constructed to obtain the comprehensive structural damage index, and the safety risk classification threshold is determined according to the safety assessment standard. Based on the structural defect sensing factor and the pre-acquired tunnel lining geometric deformation parameters, the defect expansion trend vector and the structural bearing capacity attenuation rate are obtained. Based on the disease expansion trend vector and the structural bearing capacity attenuation rate, a structural health status degradation curve is constructed to obtain the structural damage deterioration coefficient. The structural damage level is obtained based on the safety risk classification threshold and the structural damage deterioration coefficient.

4. The method for identifying tunnel structure defects by fusing laser point clouds and visible light according to claim 3, characterized in that, Based on the aforementioned structural defect sensing factors, a three-dimensional spatial distribution map of tunnel lining defects is constructed to obtain a dynamic evaluation matrix of structural defects in space, including: By associating each structural defect sensing factor with the corresponding three-dimensional point cloud coordinates in the fused dataset, the position information of each sensing factor in the tunnel lining space is obtained. Based on the location information, the three-dimensional spatial clustering of the lining surface disease response area is performed to delineate the spatial boundary of each independent disease area. For each diseased area, calculate the geometric dimensions such as the area, depth, and contour curvature of the diseased area; By combining the spatial location, geometric parameters, and disease type annotations of each disease area, a three-dimensional spatial distribution map of tunnel lining diseases is constructed. Based on the three-dimensional spatial distribution map, the geometric expansion rate and spatial density distribution changes of each disease area under different acquisition time sequences are calculated to generate a dynamic evaluation matrix of structural disease space.

5. The method for identifying tunnel structural defects by fusing laser point clouds and visible light according to claim 3, characterized in that, Extract the characteristic component changes of each structural disease sensing factor under multiple acquisition time series, calculate the unit time expansion increment of each disease type, and obtain the disease expansion rate sequence. Based on the disease expansion rate sequence, the spatial expansion direction and expansion magnitude of each disease area are fitted to construct a disease expansion trend vector. Based on the pre-acquired geometric deformation parameters of the tunnel lining, the equivalent stiffness loss of the lining section in each disease-affected area is calculated. The equivalent stiffness loss is then substituted into the structural mechanics model to obtain the structural bearing capacity attenuation rate. Using the disease expansion trend vector as the input variable and the structural bearing capacity attenuation rate as the response variable, a mapping relationship between the two is established in the time-series coordinate system, and a structural health state degradation curve is fitted and constructed. Based on the structural health degradation curve, the slope and cumulative area of ​​the curve at the current moment are calculated to comprehensively characterize the degradation rate and cumulative damage degree of the structural state, and the structural damage deterioration coefficient is obtained.

6. The method for identifying tunnel structure defects by fusing laser point clouds and visible light according to claim 3, characterized in that, Based on the safety risk classification threshold and the structural damage deterioration coefficient, the structural defect level is obtained, including: The safety risk classification threshold is divided into several risk level intervals according to the boundary values ​​of continuous intervals, and a mapping relationship between each threshold interval and the structural defect level label is established. Using the total range of the risk level interval as the normalization benchmark, the structural damage deterioration coefficient is normalized to obtain the standardized deterioration index. The standardized degradation index is compared step by step with the upper and lower boundary values ​​of each risk level interval to determine the target risk level interval in which the standardized degradation index falls. When the standardized degradation index falls near the boundary of an adjacent level, the membership function is introduced to calculate the degree of membership of the standardized degradation index to the two adjacent risk levels, with the target risk level interval and its adjacent intervals as references. The final classification level is determined by the weighted comprehensive evaluation result. Based on the target risk level range and the final attribution level, and in conjunction with the mapping relationship, the structural disease level is obtained.

7. The method for identifying tunnel structural defects by fusing laser point clouds and visible light according to claim 4, characterized in that, Real-time acquisition of current monitoring data of the tunnel structure; Based on the structural disease sensing factors corresponding to the current monitoring data and the structural deterioration trend prediction model, the future expansion state of various diseases and the changing trend of structural bearing capacity are obtained, including: According to the preset acquisition cycle, the laser scanning device and the visible light imaging device are driven by synchronous triggering commands to collect laser point cloud data and visible light image data of the current tunnel lining surface in real time, and obtain the current monitoring data. Following the historical monitoring data processing flow, spatial projection, geometric feature extraction, and bidirectional coupling analysis are sequentially performed on the current monitoring data to obtain the bidirectional coupling feature vector corresponding to the current monitoring data. The bidirectional coupled feature vector is input into a preset multi-category disease classifier, and the structural disease sensing factor corresponding to the current monitoring data is extracted based on the gradient weighted feature importance algorithm and the feature cumulative contribution threshold. The structural disease sensing factor is input into the disease expansion prediction branch of the structural deterioration trend prediction model to obtain the future expansion state of various diseases in the prediction time domain. The structural disease sensing factors are synchronously input into the risk level prediction branch of the structural deterioration trend prediction model, and combined with the future expansion status of various diseases, the trend of structural bearing capacity change is obtained.

8. The method for identifying tunnel structural defects by fusing laser point clouds and visible light according to claim 5, characterized in that, The trend of the structural bearing capacity change is differentiated over time, and the gradient of the bearing capacity decay rate in the predicted time domain is calculated to obtain the bearing capacity accelerated degradation index. Based on the aforementioned load-bearing capacity accelerated degradation index, and combined with the statistical laws governing the leap in the level of historical structural defects, a probability distribution function for the leap in the level is constructed. Substitute the trend of structural bearing capacity change into the probability distribution function of grade transition, calculate the probability of each structural defect grade transition in the prediction time domain, and obtain the structural defect grade transition probability. The transition probability of the structural disease level is compared with the preset safety margin level by level to determine the disease area where the transition probability exceeds the safety margin and its trigger warning level. Based on the aforementioned trigger warning level, combined with the precise location information of the disease in the three-dimensional spatial distribution map and the conclusion of the structural disease level assessment, maintenance and treatment recommendations are generated. The system integrates the precise location of the disease, the level assessment conclusions, and the maintenance and treatment recommendations in the form of a tiered early warning system, and outputs a comprehensive early warning report on structural diseases.

9. A computer-readable storage medium, characterized in that, Used to store computer-readable instructions, which, when read by a computer, enable the execution of a tunnel structure defect identification method that integrates laser point cloud and visible light as described in any one of claims 1-8.

Images