Intelligent three-dimensional modeling of subway tunnel structure and disease diagnosis method and system

Abstract

The application discloses a subway tunnel structure intelligent three-dimensional modeling and disease diagnosis method and system, relates to the technical field of intelligent operation and maintenance of urban rail transit infrastructure, adopts master-slave cooperative inspection units to collect laser radar point clouds, high-definition images, infrared thermal images, ultrasonic wave data and environmental sensing data, pre-identifies suspected disease areas after time synchronization and processing, constructs a self-adaptive separable four-parameter model M(beta, theta, delta, epsilon), wherein beta is a structure form parameter set, theta is a structure posture parameter set, delta is a local disease parameter set, and epsilon is an environmental correlation parameter set, and completes model initialization based on a design BIM model and historical data. The LSTM time sequence prediction model of the application realizes 12-month structure state change prediction based on a historical optimal parameter set, generates targeted maintenance suggestions, supports predictive maintenance decision-making, and reduces operation risks and maintenance costs caused by sudden diseases.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for urban rail transit infrastructure, and in particular to a method and system for intelligent three-dimensional modeling and disease diagnosis of subway tunnel structures. Background Technology

[0002] During long-term operation, subway tunnel structures are susceptible to defects such as convergence deformation, cracks, spalling, and leakage due to geological activity, load cycles, and environmental influences, threatening operational safety. Existing monitoring technologies mainly face the following bottlenecks:

[0003] 1. Insufficient model representation capability: Traditional 3D models (such as reconstruction based on laser point clouds) are only geometric shells and fail to separate the inherent structural form of the tunnel from dynamic deformation and defects, making it difficult to accurately quantify the evolution of the structural state.

[0004] 2. Low level of automation and intelligence: Disease identification relies heavily on manual interpretation of two-dimensional images or point clouds, which is inefficient, highly subjective, and cannot achieve correlation analysis between disease and overall deformation.

[0005] 3. Difficulty in fusing multi-source data: Structural response data (such as settlement and convergence) and apparent image data (such as crack images) are independent of each other, lacking a unified model for fusion analysis, making it difficult to form a comprehensive assessment.

[0006] 4. Lack of predictive maintenance capability: Existing methods are unable to establish high-precision parametric digital models, and cannot dynamically predict and warn of structural performance degradation. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a method and system for intelligent 3D modeling and defect diagnosis of subway tunnel structures. The technical solution adopted is as follows:

[0008] A method for intelligent 3D modeling and defect diagnosis of subway tunnel structures includes the following steps:

[0009] Step 1: The master-slave collaborative inspection unit collects lidar point cloud, high-definition images, infrared thermal images, ultrasonic data and environmental sensor data. After time synchronization and processing, suspected disease areas are pre-identified.

[0010] Step 2: Construct an adaptive separable four-parameter model M(β,θ,δ,ε), where β is the set of structural morphology parameters, θ is the set of structural attitude parameters, δ is the set of local defects parameters, and ε is the set of environmental correlation parameters. The model is initialized based on the design BIM model and historical data.

[0011] Step 3: Employ a hybrid alternating optimization algorithm, performing two rounds of coarse and fine tuning iterations to sequentially optimize θ, δ, β, and ε until the overall error between the model and the measured data meets the dual convergence thresholds, outputting the optimal parameter set. ;

[0012] Step 4: Extract three types of indicators from the optimal parameter set: structural deformation, disease quantification, and environmental impact. Use the analytic hierarchy process (AHP) to establish a five-level health status assessment system and trigger graded early warnings.

[0013] Step 5: Based on the LSTM time series prediction model, input the historical optimal parameter set, predict the structural state changes in the next 12 months, and generate predictive maintenance suggestions.

[0014] Optionally, the master-slave collaborative inspection unit in step 1 includes a master unit and a slave unit. The master unit is equipped with a lidar, a high-definition industrial camera, an infrared thermal imager, and an ultrasonic cross-section scanner. The slave unit is equipped with a fiber optic grating sensor and a temperature and humidity sensor. The acquisition frequency is dynamically adjusted according to the tunnel type.

[0015] Optionally, the preprocessing in step 1 includes: removing point cloud noise using an improved bilateral filtering algorithm, achieving image super-resolution reconstruction using the ESRGAN model, completing sparse completion using a Transformer-based point cloud inpainting algorithm, and pre-identifying suspected disease areas and outputting ROI coordinates using a lightweight YOLOv8 model.

[0016] Optionally, in step 2, the structural morphology parameter set β characterizes the inherent geometric and material properties of the tunnel structure, including the reference dimensions of the tunnel cross-section design, the characteristics of the ring joints, the location of the reserved functional holes, and the mechanical parameters of the segment material. Based on the tunnel design drawings, BIM model, and historical health status data, the initial values ​​are optimized through principal component analysis to adapt to construction errors and individual structural differences.

[0017] The structural attitude parameter set θ represents the spatial attitude and deformation state of the tunnel as a whole and in a local area, including the three-dimensional translation of the axis, the convergence value of the cross section and the spatial rotation angle of the ring section. A multi-degree-of-freedom spatial transformation matrix is ​​used to achieve quantitative representation.

[0018] Local disease parameter set For typical tunnel defects, specific parameter sub-models are constructed, each sub-model including the location, geometric dimensions, and quantitative indicators of the severity of the defect;

[0019] Environmental correlation parameter set Environmental factors affecting the structural state of tunnels are characterized, including temperature and humidity inside the tunnel, surrounding soil stress, and train traffic frequency. By establishing a correlation model with the structural attitude parameter set θ, the influence weight of environmental factors on structural deformation is quantified.

[0020] Optionally, the hybrid alternating optimization algorithm described in step 3 uses a coarse-tuning and fine-tuning hierarchical iterative framework as its core framework, specifically including:

[0021] In the coarse adjustment stage, priority is given to optimizing dynamic parameters that have a significant impact on the structural state representation. The set of structural morphology parameters β is fixed, and the set of structural attitude parameters θ and the set of environmental correlation parameters ε are optimized in turn. The goal is to minimize the geometric deviation between the model and the measured 3D point cloud data. In the fine adjustment stage, based on the coarse-adjusted θ and ε, the set of structural morphology parameters β and the set of local disease parameters δ are optimized in turn. The goal is to minimize the feature matching deviation between the model and the diseased areas in the measured appearance image.

[0022] Adaptive optimization algorithms are selected based on the characteristics of different parameter sets. Gradient descent algorithms are used to achieve fast convergence for the structural attitude parameter set θ and the environmental correlation parameter set ε, while high-precision numerical optimization algorithms are used to improve the accuracy of parameter solution for the structural morphology parameter set β and the local disease parameter set δ.

[0023] A dual convergence threshold is set: the first threshold is the overall deviation between the model and the measured data, and the second threshold is the change in the overall deviation during multiple consecutive iterations. When both thresholds meet the preset requirements, the iteration stops and the optimal parameter set is output. .

[0024] Optionally, in step 4, the weights of the three core indicators are determined using the analytic hierarchy process (AHP), with the structural deformation indicator accounting for 40%, the disease quantification indicator accounting for 45%, and the environmental impact indicator accounting for 15%.

[0025] The overall score is calculated by multiplying the weights of each category of indicators by the sum of the average scores of each sub-indicator under the corresponding category of indicators, and then classifying the health status into five levels based on the overall score:

[0026] Scores of 85-100 correspond to Excellent, 70-84 to Good, 60-69 to Pass, 40-59 to Warning, and less than 40 to Danger. The triggering conditions for graded warnings are related to the proportion of core indicators exceeding the standard limits and the rate of disease development.

[0027] Optionally, in step 5, the LSTM time series prediction model takes into account the optimal parameter set of the past 5 periods and outputs the structural attitude parameter θ(t) and the disease parameter δ(t) for the next 3 months, 6 months and 12 months.

[0028] The intelligent 3D modeling and defect diagnosis system for subway tunnel structures is used to realize intelligent 3D modeling and defect diagnosis methods for subway tunnel structures. It adopts a three-level architecture of perception layer, edge computing layer, and cloud application layer, including: perception layer, edge computing layer, and cloud application layer.

[0029] The perception layer is used for real-time acquisition and preliminary screening of multi-source heterogeneous data;

[0030] The edge computing layer is deployed at the inspection robot terminal and the nearest edge node of the tunnel for data preprocessing, suspected area identification and preliminary parameter optimization;

[0031] The cloud application layer includes a model calculation module, a diagnosis and assessment module, a predictive analysis module, a visualization management module, and an early warning push module. The cloud application layer is used for optimal parameter solving, health status assessment, time series prediction, visualization display, hierarchical early warning push, data archiving, and model iteration.

[0032] Optionally, the master-slave collaborative inspection robot is equipped with the master-slave collaborative inspection unit as described in claim 2, which supports adaptive movement and data acquisition along the tunnel track or sidewall, realizes data timestamp alignment of multiple devices through the PTP time synchronization protocol, and completes spatial registration by using the feature point matching algorithm of lidar point cloud and camera image.

[0033] Optionally, the cloud application layer's visualization management module supports the automatic generation of high-precision 3D mesh models based on the optimal parameter set, providing functions such as highlighting diseased areas, scaling, rotating, and sectioning models; the early warning push module supports the push of tiered early warning information through multiple channels such as SMS, APP, and system pop-ups; the system supports user permission tiers, operation log recording, model parameter configuration, and standard update functions.

[0034] In summary, the present invention has at least one of the following beneficial technical effects:

[0035] This invention provides a method and system for intelligent 3D modeling and defect diagnosis of subway tunnel structures. The master-slave collaborative inspection unit, combined with multi-source data preprocessing technology, improves the comprehensiveness of tunnel inspection data collection and the accuracy of preprocessing, enabling efficient pre-identification of suspected defect areas and providing high-quality data support for subsequent modeling and diagnosis.

[0036] The adaptive separable four-parameter model collaboratively represents structural characteristics through a multi-dimensional parameter set. By combining the design BIM model with historical data initialization and principal component analysis optimization, it enhances the model's adaptability to tunnel construction errors and individual differences, and improves the fit of 3D modeling.

[0037] The hybrid alternating optimization algorithm adopts a coarse-fine tuning hierarchical iterative strategy, matching and adapting the optimization algorithm to the characteristics of different parameter sets, achieving rapid convergence and high-precision solution of parameter optimization, and ensuring low-error matching between the model and the measured data.

[0038] The five-level health status assessment system constructed by the analytic hierarchy process (AHP) achieves quantitative and graded assessment of the health status of tunnel structures by scientifically allocating indicator weights and clarifying scoring rules. The graded early warning mechanism can promptly trigger the corresponding level of early warning response, thereby improving the timeliness of structural safety management.

[0039] The LSTM time series prediction model uses historical optimal parameter sets to predict structural state changes over the next 12 months, generating targeted maintenance recommendations, supporting predictive maintenance decisions, and reducing operational risks and maintenance costs caused by sudden defects. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the intelligent 3D modeling and defect diagnosis method for subway tunnel structures according to the present invention. Detailed Implementation

[0041] The present invention will be further described in detail below with reference to the accompanying drawings.

[0042] This invention discloses a method and system for intelligent three-dimensional modeling and disease diagnosis of subway tunnel structures.

[0043] Reference Figure 1 Example 1, Intelligent 3D Modeling and Defect Diagnosis Method for Subway Tunnel Structures, includes the following steps:

[0044] Step 1: The master-slave collaborative inspection unit collects lidar point cloud, high-definition images, infrared thermal images, ultrasonic data and environmental sensor data. After time synchronization and processing, suspected disease areas are pre-identified.

[0045] Step 2: Construct an adaptive separable four-parameter model M(β,θ,δ,ε), where β is the set of structural morphology parameters, θ is the set of structural attitude parameters, δ is the set of local defects parameters, and ε is the set of environmental correlation parameters. The model is initialized based on the design BIM model and historical data.

[0046] Step 3: Employ a hybrid alternating optimization algorithm, performing two rounds of coarse and fine tuning iterations to sequentially optimize θ, δ, β, and ε until the overall error between the model and the measured data meets the dual convergence thresholds, outputting the optimal parameter set. ;

[0047] Step 4: Extract three types of indicators from the optimal parameter set: structural deformation, disease quantification, and environmental impact. Use the analytic hierarchy process (AHP) to establish a five-level health status assessment system and trigger graded early warnings.

[0048] Step 5: Based on the LSTM time series prediction model, input the historical optimal parameter set, predict the structural state changes in the next 12 months, and generate predictive maintenance suggestions.

[0049] Example 2: In step 1, the master-slave collaborative inspection unit includes a master unit and a slave unit. The master unit is equipped with a lidar, a high-definition industrial camera, an infrared thermal imager, and an ultrasonic cross-section scanner. The slave unit is equipped with a fiber optic grating sensor and a temperature and humidity sensor. The acquisition frequency is dynamically adjusted according to the tunnel type.

[0050] Example 3, the preprocessing in step 1 includes: removing point cloud noise using an improved bilateral filtering algorithm, realizing image super-resolution reconstruction using the ESRGAN model, completing sparse completion using a Transformer-based point cloud inpainting algorithm, and pre-identifying suspected disease areas and outputting ROI coordinates using a lightweight YOLOv8 model.

[0051] Example 4: In step 2, the structural morphology parameter set β represents the inherent geometric and material properties of the tunnel structure, including the reference dimensions of the tunnel cross section design, the characteristics of the ring joints, the location of the reserved functional holes, and the mechanical parameters of the segment material. Based on the tunnel design drawings, BIM model, and historical health status data, the initial values ​​are optimized through principal component analysis to adapt to construction errors and individual structural differences.

[0052] The structural attitude parameter set θ represents the spatial attitude and deformation state of the tunnel as a whole and in a local area, including the three-dimensional translation of the axis, the convergence value of the cross section and the spatial rotation angle of the ring section. A multi-degree-of-freedom spatial transformation matrix is ​​used to achieve quantitative representation.

[0053] For typical tunnel defects, the local defect parameter set δ constructs a dedicated parameter sub-model. Each sub-model includes the location, geometric dimensions, and severity quantification indicators of the defect.

[0054] The environmental correlation parameter set ε represents the environmental factors affecting the tunnel structure state, including temperature and humidity inside the tunnel, surrounding soil stress, and train traffic frequency. By establishing a correlation model with the structural attitude parameter set θ, the influence weight of environmental factors on structural deformation is quantified.

[0055] By adopting the above technical solutions, the master-slave collaborative inspection unit is equipped with multiple types of sensors to achieve comprehensive acquisition of multi-source data such as LiDAR point clouds and high-definition images. Combined with time synchronization, the spatiotemporal consistency of the data is ensured. Furthermore, by employing technologies such as improved bilateral filtering, ESRGAN model, Transformer point cloud repair, and lightweight YOLOv8 model, data denoising, super-resolution reconstruction, sparse completion, and pre-identification of suspected disease areas are completed, providing a high-quality data foundation for subsequent analysis.

[0056] An adaptive, separable, four-parameter model is constructed. The structural morphology parameter set represents the inherent geometric and material properties of the tunnel, the structural attitude parameter set quantifies the spatial attitude and deformation, the local defect parameter set establishes a specific characterization sub-model for typical defects, and the environmental correlation parameter set quantifies the impact of environmental factors on deformation. Based on the design BIM model, drawings, and historical health data, the initial values ​​of the model are optimized by combining principal component analysis to adapt to construction errors and individual structural differences.

[0057] Hybrid alternating parameter optimization: A hybrid alternating optimization algorithm with coarse-fine tuning layer-by-layer iteration is adopted. In the coarse tuning stage, the structural morphology parameters are fixed, and the structural pose and environmental correlation parameters are optimized first to minimize the geometric deviation of the point cloud. In the fine tuning stage, based on the coarse tuning results, the structural morphology and local disease parameters are optimized to minimize the image feature matching deviation. By adapting the optimization algorithm to different parameter characteristics, the parameters are quickly converged and solved with high accuracy, ensuring low error matching between the model and the measured data.

[0058] Quantitative health assessment: Three core indicators—structural deformation, disease quantification, and environmental impact—are extracted from the optimal parameter set. The weights of the indicators are allocated using the analytic hierarchy process (AHP), and a comprehensive score is calculated based on clear scoring rules. A five-level health status assessment system is constructed, and the proportion of core indicators exceeding the standard limits and the rate of disease development are linked to trigger graded early warnings, thereby achieving accurate quantitative assessment and timely control of the tunnel structure's health status.

[0059] Example 5, the hybrid alternating optimization algorithm described in step 3 uses coarse-tuning and fine-tuning hierarchical iteration as its core framework, specifically including:

[0060] In the coarse adjustment stage, priority is given to optimizing dynamic parameters that have a significant impact on the structural state representation. The set of structural morphology parameters β is fixed, and the set of structural attitude parameters θ and the set of environmental correlation parameters ε are optimized in turn. The goal is to minimize the geometric deviation between the model and the measured 3D point cloud data. In the fine adjustment stage, based on the coarse-adjusted θ and ε, the set of structural morphology parameters β and the set of local disease parameters δ are optimized in turn. The goal is to minimize the feature matching deviation between the model and the diseased areas in the measured appearance image.

[0061] Adaptive optimization algorithms are selected based on the characteristics of different parameter sets. Gradient descent algorithms are used to achieve fast convergence for the structural attitude parameter set θ and the environmental correlation parameter set ε, while high-precision numerical optimization algorithms are used to improve the accuracy of parameter solution for the structural morphology parameter set β and the local disease parameter set δ.

[0062] A dual convergence threshold is set: the first threshold is the overall deviation between the model and the measured data, and the second threshold is the change in the overall deviation during multiple consecutive iterations. When both thresholds meet the preset requirements, the iteration stops and the optimal parameter set is output. .

[0063] By adopting the above technical solution, in the coarse adjustment stage, the structural morphology parameter set β is fixed, and the structural attitude parameter set θ and the environmental correlation parameter set ε are optimized first, with the goal of minimizing the geometric deviation between the model and the measured 3D point cloud data. Gradient descent algorithms are used to achieve rapid parameter convergence. In the fine adjustment stage, based on θ and ε obtained from the coarse adjustment, the structural morphology parameter set β and the local disease parameter set δ are further optimized, with the goal of minimizing the feature matching deviation between the model and the diseased areas in the measured appearance image. High-precision numerical optimization algorithms are selected to improve the solution accuracy. At the same time, the overall deviation between the model and the measured data and the change in deviation over multiple iterations are set as dual convergence thresholds. When both meet the preset requirements, the iteration stops, and the optimal parameter set is output. .

[0064] In Example 6, step 4, the weights of the three core indicators are determined using the analytic hierarchy process (AHP), with structural deformation indicators accounting for 40%, disease quantification indicators accounting for 45%, and environmental impact indicators accounting for 15%.

[0065] The overall score is calculated by multiplying the weights of each category of indicators by the sum of the average scores of each sub-indicator under the corresponding category of indicators, and then classifying the health status into five levels based on the overall score:

[0066] Scores of 85-100 correspond to Excellent, 70-84 to Good, 60-69 to Pass, 40-59 to Warning, and less than 40 to Danger. The triggering conditions for graded warnings are related to the proportion of core indicators exceeding the standard limits and the rate of disease development.

[0067] By adopting the above technical solution, the weight ratios of structural deformation indicators, disease quantification indicators, and environmental impact indicators are determined using the analytic hierarchy process (AHP) to be 40%, 45%, and 15%, respectively. The comprehensive score is calculated by multiplying the weights of each indicator by the mean scores of the sub-indicators under the corresponding category and then summing the results. Based on the comprehensive score, the health status is divided into five levels: 85-100 points corresponds to the excellent level, 70-84 points to the good level, 60-69 points to the qualified level, 40-59 points to the warning level, and less than 40 points to the dangerous level. The triggering conditions for the graded warning are linked to the proportion of core indicators exceeding the standard limits and the disease development rate, thereby achieving quantitative assessment of the health status and targeted early warning.

[0068] In Example 7, the LSTM time series prediction model in step 5 takes into account the optimal parameter set of the past 5 periods and outputs the structural attitude parameter θ(t) and the disease parameter δ(t) for the next 3 months, 6 months and 12 months.

[0069] By adopting the above technical solution, with the LSTM time series prediction model as the core: inputting the optimal parameter set of the past 5 periods, using the feature capture capability of the LSTM model for time series data, mining the change law of tunnel structure state parameters over time, and then outputting the structural attitude parameters θ(t) and the disease parameters δ(t) for the next 3 months, 6 months and 12 months, so as to realize the forward prediction of structural state changes.

[0070] Example 8: Intelligent 3D Modeling and Disease Diagnosis System for Subway Tunnel Structures. This system is used to realize intelligent 3D modeling and disease diagnosis methods for subway tunnel structures. It adopts a three-level architecture consisting of a perception layer, an edge computing layer, and a cloud application layer.

[0071] The perception layer is used for real-time acquisition and preliminary screening of multi-source heterogeneous data;

[0072] The edge computing layer is deployed at the inspection robot terminal and the nearest edge node of the tunnel for data preprocessing, suspected area identification and preliminary parameter optimization;

[0073] The cloud application layer includes a model calculation module, a diagnosis and assessment module, a predictive analysis module, a visualization management module, and an early warning push module. The cloud application layer is used for optimal parameter solving, health status assessment, time series prediction, visualization display, hierarchical early warning push, data archiving, and model iteration.

[0074] Example 9: The master-slave collaborative inspection robot is equipped with the master-slave collaborative inspection unit described in claim 2. It supports adaptive movement and data acquisition along the tunnel track or sidewall. It achieves data timestamp alignment of multiple devices through the PTP time synchronization protocol and completes spatial registration by using a feature point matching algorithm between lidar point cloud and camera image.

[0075] Example 10: The cloud application layer's visualization management module supports the automatic generation of high-precision 3D mesh models based on the optimal parameter set, providing functions such as highlighting diseased areas, scaling, rotating, and sectioning models; the early warning push module supports the push of graded early warning information through multiple channels such as SMS, APP, and system pop-ups; the system supports user permission classification, operation log recording, model parameter configuration, and standard update functions.

[0076] By adopting the above technical solution, the system employs a three-tier architecture: a perception layer, an edge computing layer, and a cloud application layer. The perception layer is responsible for real-time acquisition and preliminary screening of multi-source heterogeneous data. The edge computing layer is deployed at the inspection robot terminal and nearby edge nodes of the tunnel to complete data preprocessing, suspected area identification, and preliminary parameter optimization. The cloud application layer, through model calculation, diagnostic evaluation, predictive analysis, visualization management, and early warning push modules, achieves optimal parameter solving, health status assessment, time series prediction, visualization display, hierarchical early warning push, data archiving, and model iteration. Specifically, the master-slave collaborative inspection robot is equipped with a master-slave collaborative inspection unit, supporting adaptive movement and acquisition along the tunnel track or sidewall. It achieves multi-device data timestamp alignment through the PTP time series synchronization protocol and spatial registration using a feature point matching algorithm between LiDAR point clouds and camera images. The visualization management module of the cloud application layer supports the generation of high-precision 3D mesh models based on the optimal parameter set and provides disease annotation and model operation functions. The early warning push module supports multi-channel early warning information push. The system also features user permission hierarchy, operation log recording, model parameter configuration, and standard update functions, achieving intelligent collaboration and efficient management throughout the entire process.

[0077] The following specific embodiments illustrate the implementation principle of the present invention:

[0078] Taking a section of Metro Line 3 in a certain city as an example, this tunnel was constructed using the shield tunneling method. The outer diameter of the tunnel segments is 6m, the inner diameter is 5.4m, the total length is 1.2km, and the operating period is 5 years. This solution is required to achieve intelligent structural modeling and defect diagnosis. The specific implementation process is as follows:

[0079] Step 1: Multi-source data acquisition and preprocessing:

[0080] Data acquisition is performed using a master-slave collaborative inspection unit. The master unit is equipped with a 16-line LiDAR, a 4K high-definition industrial camera, a 640×512 resolution infrared thermal imager, and a 2MHz ultrasonic cross-section scanner. The slave units are equipped with fiber optic stress-strain sensors (measurement range 0-2000με) and temperature and humidity sensors (accuracy ±0.5℃ / ±2%RH). Data is collected adaptively along the tunnel track, with one set of data collected every 5 meters in the tunnel section. The PTP time synchronization protocol is used to align the timestamps of multiple devices, and the SIFT feature point matching algorithm between the LiDAR point cloud and the camera image is used to complete spatial registration.

[0081] The LSTM time series prediction model uses historical optimal parameter sets to predict structural state changes over the next 12 months, generating targeted maintenance recommendations, supporting predictive maintenance decisions, and reducing operational risks and maintenance costs caused by sudden defects.

[0082] Data preprocessing stage: An improved bilateral filtering algorithm is used to remove point cloud noise (filter window 5×5, standard deviation 1.2); the resolution of high-definition images is increased from 4K to 8K using the ESRGAN model to enhance the details of the defects; a Transformer-based point cloud repair algorithm is used to complete sparse areas (repair rate ≥92%); a lightweight YOLOv8 model (40% reduction in parameters) is used to pre-identify suspected defect areas, outputting 23 ROI coordinates, covering types such as cracks and spalling.

[0083] The three-tier architecture system achieves efficient integration of real-time multi-source data acquisition, rapid edge preprocessing, and deep cloud computing analysis through the collaborative division of labor among the perception layer, edge computing layer, and cloud application layer, thereby improving the system's data processing response speed. Visual management and multi-channel early warning push functions enhance the intuitiveness of diagnostic results display and the efficiency of early warning information delivery, thereby improving the intelligence and convenience of tunnel structure management.

[0084] An adaptive, separable four-parameter model M(β,θ,δ,ε) is constructed: the structural morphology parameter set β includes 12 indicators such as the tunnel cross-section design radius (inner 5400mm / outer 6000mm), the angle of the ring joint (12 locations, spaced 30° apart), the coordinates of reserved holes (8 equipment holes, coordinates extracted from the design BIM model), and the elastic modulus of the tunnel segment (35GPa). The initial values ​​are optimized using PCA with the design drawings and the data from the previous three health periods to accommodate construction errors (such as radius deviation ≤3mm); the structural attitude parameter set... θ is represented by a 6-DOF spatial transformation matrix, initially set to [0,0,0,0°,0°,0°]. The local disease parameter set δ constructs sub-models for cracks (location, length, width, etc.) and spalling (area, depth, etc.), initially set to a disease-free state. The environmental correlation parameter set ε includes the average temperature and humidity inside the tunnel (25℃, 65%RH), surrounding soil stress (0.8MPa), and train traffic frequency (120 trains / day), initialized based on the statistical average of historical data over the past year, and establishes a linear correlation model with θ.

[0085] Step 2: Construction and initialization of the four-parameter model:

[0086] Step 3: Mixed alternating parameter optimization:

[0087] A hybrid alternating optimization algorithm is used for iterative solution: In the coarse adjustment stage, β is fixed, and θ and ε are optimized first. Gradient descent is used to optimize θ (learning rate 0.001, 10 iterations), with the goal of point cloud chamfer distance error ≤ 0.5mm; then ε is optimized (learning rate 0.005), establishing a correlation weight with θ. In the fine adjustment stage, based on the coarse-adjusted θ and ε, the L-BFGS algorithm is used to optimize β (15 iterations), and coordinate descent is used to optimize δ (8 iterations), with the goal of image IoU ≥ 0.85 for the diseased area. A dual convergence threshold is set: overall deviation score ≥ 0.88 (point cloud deviation score × 0.6 + image deviation score × 0.4), deviation change ≤ 0.01 for 3 consecutive iterations, and the optimal parameter set is output after iteration stops. ,in The maximum settlement of the axis is 3.2mm. 18 actual defects were marked (5 cracks, 3 peelings, etc.).

[0088] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent 3D modeling and defect diagnosis of subway tunnel structures, characterized by: Includes the following steps: Step 1: The master-slave collaborative inspection unit collects lidar point cloud, high-definition images, infrared thermal images, ultrasonic data and environmental sensor data. After time synchronization and processing, suspected disease areas are pre-identified. Step 2: Construct an adaptive separable four-parameter model M(β,θ,δ,ε), where β is the set of structural morphology parameters, θ is the set of structural attitude parameters, δ is the set of local defects parameters, and ε is the set of environmental correlation parameters. The model is initialized based on the design BIM model and historical data. Step 3: Employ a hybrid alternating optimization algorithm, performing two rounds of coarse and fine tuning iterations to sequentially optimize θ, δ, β, and ε until the overall error between the model and the measured data meets the dual convergence thresholds, outputting the optimal parameter set. ; Step 4: Extract three types of indicators from the optimal parameter set: structural deformation, disease quantification, and environmental impact. Use the analytic hierarchy process (AHP) to establish a five-level health status assessment system and trigger graded early warnings. Step 5: Based on the LSTM time series prediction model, input the historical optimal parameter set, predict the structural state changes in the next 12 months, and generate predictive maintenance suggestions.

2. The intelligent three-dimensional modeling and defect diagnosis method for subway tunnel structures according to claim 1, characterized in that: In step 1, the master-slave collaborative inspection unit includes a master unit and a slave unit. The master unit is equipped with a lidar, a high-definition industrial camera, an infrared thermal imager, and an ultrasonic cross-section scanner. The slave unit is equipped with a fiber optic grating sensor and a temperature and humidity sensor. The acquisition frequency is dynamically adjusted according to the tunnel type.

3. The intelligent three-dimensional modeling and defect diagnosis method for subway tunnel structures according to claim 2, characterized in that: The preprocessing in step 1 includes: removing point cloud noise using an improved bilateral filtering algorithm, achieving image super-resolution reconstruction using the ESRGAN model, completing sparse completion using a Transformer-based point cloud inpainting algorithm, and pre-identifying suspected disease areas and outputting ROI coordinates using a lightweight YOLOv8 model.

4. The intelligent three-dimensional modeling and defect diagnosis method for subway tunnel structures according to claim 3, characterized in that: In step 2, the structural morphology parameter set β characterizes the inherent geometric and material properties of the tunnel structure, including the design reference dimensions of the tunnel cross section, the characteristics of the ring joints, the location of the reserved functional holes, and the mechanical parameters of the segment material. Based on the tunnel design drawings, BIM model, and historical health status data, the initial values ​​are optimized through principal component analysis to adapt to construction errors and individual structural differences. The structural attitude parameter set θ represents the spatial attitude and deformation state of the tunnel as a whole and in a local area, including the three-dimensional translation of the axis, the convergence value of the cross section and the spatial rotation angle of the ring section. A multi-degree-of-freedom spatial transformation matrix is ​​used to achieve quantitative representation. For typical tunnel defects, the local defect parameter set δ constructs a dedicated parameter sub-model. Each sub-model includes the location, geometric dimensions, and severity quantification indicators of the defect. The environmental correlation parameter set ε represents the environmental factors affecting the tunnel structure state, including temperature and humidity inside the tunnel, surrounding soil stress, and train traffic frequency. By establishing a correlation model with the structural attitude parameter set θ, the influence weight of environmental factors on structural deformation is quantified.

5. The intelligent three-dimensional modeling and defect diagnosis method for subway tunnel structures according to claim 4, characterized in that: The hybrid alternating optimization algorithm described in step 3 uses a coarse-tuning and fine-tuning hierarchical iterative framework as its core, specifically including: In the coarse adjustment stage, priority is given to optimizing dynamic parameters that have a significant impact on the structural state representation. The set of structural morphology parameters β is fixed, and the set of structural attitude parameters θ and the set of environmental correlation parameters ε are optimized in turn. The goal is to minimize the geometric deviation between the model and the measured 3D point cloud data. In the fine adjustment stage, based on the coarse-adjusted θ and ε, the set of structural morphology parameters β and the set of local disease parameters δ are optimized in turn. The goal is to minimize the feature matching deviation between the model and the diseased areas in the measured appearance image. Adaptive optimization algorithms are selected based on the characteristics of different parameter sets. Gradient descent algorithms are used to achieve fast convergence for the structural attitude parameter set θ and the environmental correlation parameter set ε, while high-precision numerical optimization algorithms are used to improve the accuracy of parameter solution for the structural morphology parameter set β and the local disease parameter set δ. A dual convergence threshold is set: the first threshold is the overall deviation between the model and the measured data, and the second threshold is the change in the overall deviation during multiple consecutive iterations. When both thresholds meet the preset requirements, the iteration stops and the optimal parameter set is output. .

6. The intelligent three-dimensional modeling and defect diagnosis method for subway tunnel structures according to claim 5, characterized in that: In step 4, the weights of the three core indicators are determined using the analytic hierarchy process (AHP), with structural deformation indicators accounting for 40%, disease quantification indicators accounting for 45%, and environmental impact indicators accounting for 15%. The overall score is calculated by multiplying the weights of each category of indicators by the sum of the average scores of each sub-indicator under the corresponding category of indicators, and then classifying the health status into five levels based on the overall score: Scores of 85-100 correspond to Excellent, 70-84 to Good, 60-69 to Pass, 40-59 to Warning, and less than 40 to Danger. The triggering conditions for graded warnings are related to the proportion of core indicators exceeding the standard limits and the rate of disease development.

7. The intelligent three-dimensional modeling and defect diagnosis method for subway tunnel structures according to claim 6, characterized in that: In step 5, the LSTM time series prediction model takes into account the optimal parameter set of the past 5 periods and outputs the structural attitude parameter θ(t) and the disease parameter δ(t) for the next 3 months, 6 months and 12 months.

8. A smart 3D modeling and defect diagnosis system for subway tunnel structures, characterized by: The method for intelligent three-dimensional modeling and defect diagnosis of subway tunnel structures as described in claim 7 adopts a three-level architecture of perception layer, edge computing layer, and cloud application layer, including: perception layer, edge computing layer, and cloud application layer; The perception layer is used for real-time acquisition and preliminary screening of multi-source heterogeneous data; The edge computing layer is deployed at the inspection robot terminal and the nearest edge node of the tunnel for data preprocessing, suspected area identification and preliminary parameter optimization; The cloud application layer includes a model calculation module, a diagnosis and assessment module, a predictive analysis module, a visualization management module, and an early warning push module. The cloud application layer is used for optimal parameter solving, health status assessment, time series prediction, visualization display, hierarchical early warning push, data archiving, and model iteration.

9. The intelligent three-dimensional modeling and defect diagnosis system for subway tunnel structures according to claim 8, characterized in that, The master-slave collaborative inspection robot is equipped with the master-slave collaborative inspection unit as described in claim 2. It supports adaptive movement and data acquisition along the tunnel track or sidewall, achieves data timestamp alignment of multiple devices through the PTP time synchronization protocol, and completes spatial registration by using a feature point matching algorithm between lidar point cloud and camera image.

10. The intelligent three-dimensional modeling and defect diagnosis system for subway tunnel structures according to claim 9, characterized in that, The cloud application layer's visualization management module supports the automatic generation of high-precision 3D mesh models based on the optimal parameter set, and provides functions such as highlighting diseased areas, scaling, rotating, and sectioning models. The early warning push module supports pushing tiered early warning information through multiple channels such as SMS, APP, and system pop-ups; the system supports user permission tiers, operation log recording, model parameter configuration, and standard update functions.

Images