Intelligent health maintenance and prediction method and system for bridge inspection

Abstract

The application provides a kind of bridge inspection intelligent health maintenance prediction method and system, related to bridge inspection technical field.According to the bridge picture uploaded by unmanned aerial vehicle inspection, the bridge point cloud is extracted, rendering is carried out, the crack center line skeleton is extracted, and the crack size is determined;Based on the crack, the health surface Gaussian body around the disease is fitted, the local reference plane is determined, and the directed distance from the Gaussian body of the disease area to the local reference plane is determined;According to the directed distance and the grid area, the volume of the disease area is determined;Based on multidimensional performance data, the bridge health index is determined;The bridge health index and time sequence characteristics are input into the disease evolution prediction model to obtain the predicted health index, and the degradation model and the remaining life estimation model are used to determine the probability distribution of the remaining life of the bridge.Thereby, the problem of lack of dynamic prediction ability in traditional inspection is solved, and the transformation from passive detection to active warning is realized.

Description

Technical Field

[0001] This invention relates to the field of bridge inspection technology, and more specifically, to an intelligent health maintenance prediction method and system for bridge inspection. Background Technology

[0002] As bridges age, structural defects are becoming increasingly prominent. Manual inspections, relying on bridge inspection vehicles and specialized personnel, are not only time-consuming (7-15 days for a single large bridge) and costly annually, but also pose significant safety risks due to working at heights. Traditional inspection methods face multiple challenges, including low efficiency, high costs, and prominent safety hazards.

[0003] While existing drone-based intelligent inspection technology has alleviated the above problems to some extent, the existing inspection system is mostly a periodic and passive condition assessment, that is, "inspect when it is due and repair when it is found", lacking dynamic prediction and proactive early warning mechanisms for the development trend of diseases. Summary of the Invention

[0004] The problem that this invention aims to solve is that the existing inspection system lacks a dynamic prediction and proactive early warning mechanism for the development trend of diseases.

[0005] To address the aforementioned problems, in a first aspect, the present invention provides an intelligent health maintenance prediction method for bridge inspection, comprising:

[0006] Based on the bridge images uploaded by the drone inspection, image feature points are extracted to obtain the bridge point cloud;

[0007] Render the bridge point cloud and extract the crack centerline skeleton;

[0008] Based on the extracted crack centerline skeleton, the crack size is determined, including crack length, crack width, and crack depth.

[0009] Based on the identified cracks, fit a local reference plane to the Gaussian volume of the healthy surface around the disease and determine the directed distance from the Gaussian volume of the diseased area to the local reference plane.

[0010] The affected area is gridded, and the volume of the affected area is determined based on the directed distance and grid area.

[0011] Based on defect performance data, structural performance data, and durability performance data, the fuzzy hierarchical analysis method is used to comprehensively evaluate the health status of bridges and determine the bridge health index. Defect performance data includes crack size and the volume of defect area.

[0012] The current bridge health index, disease performance data, structural performance data, durability performance data, and time series characteristics are input into the disease evolution prediction model to obtain the predicted health index corresponding to the time series characteristics.

[0013] Based on the predicted health index to generate degradation data, the degradation model and the remaining life estimation model are used to determine the probability distribution of the remaining life of the bridge, so as to facilitate the health maintenance of the bridge.

[0014] Optionally, the degradation model is a recurrent neural network;

[0015] The process of generating degradation data based on predicted health indices, and using degradation models and remaining life estimation models to determine the probability distribution of the bridge's remaining life includes:

[0016] The predicted health index and expected life expectancy are input into the trained recurrent neural network to obtain the deterioration health index.

[0017] Based on the deterioration health index and the deterioration threshold, the probability distribution of expected remaining lifespan below the deterioration threshold is determined in reverse.

[0018] Optionally, the intelligent health maintenance prediction method for bridge inspection further includes:

[0019] Construct a safety objective function based on the probability distribution of remaining lifetime;

[0020] Based on the total maintenance cost objective function, the safety objective function, and the traffic disturbance objective function, and under preset constraints, a multi-objective optimization function is solved to obtain multiple objective values ​​for the maintenance scheme, so that decision-makers can select the maintenance scheme based on the objective values.

[0021] Optionally, the crack length is

[0022]

[0023] in, It is a sequence of skeleton pixel coordinates;

[0024] The width of the crack is

[0025]

[0026] Where s is the skeleton arc length parameter, and GSD is the ground sampling distance;

[0027] Assuming the crack is a V-shaped groove, the crack depth is

[0028]

[0029] in, width, Due to the difference in perspective.

[0030] Optionally, the directed distance is

[0031]

[0032] Where n is the normal to the local reference plane, Let a point be on the local reference plane. Gaussian body The projection point on the local reference plane;

[0033] The volume of the diseased area is

[0034]

[0035] in, Let k be the average depth of the grid. This represents the grid area.

[0036] Optionally, the bridge health index is:

[0037]

[0038] in, As the indicator weight, Individual scoring;

[0039] Crack severity rating:

[0040]

[0041] in, The width of the crack.

[0042] Optionally, the disease evolution prediction model is a stacked long short-term memory network, which includes an input layer, a first long short-term memory layer, a discard layer, a second long short-term memory layer, a fully connected layer, and an output layer.

[0043] When training a stacked long short-term memory network, the dropout layer is retained, and a preset number of forward propagations are performed to determine the uncertainty estimation results of Monte Carlo dropout.

[0044] The expected value is

[0045]

[0046] Where N is the preset number of times, The predicted health index for the next h months is obtained from the nth training iteration;

[0047] The uncertainty of the forecast is

[0048]

[0049] The 95% confidence interval was obtained:

[0050] .

[0051] Optionally, the degenerative health index is

[0052]

[0053] in, For the degeneracy function of a recurrent neural network, For the parameters of the recurrent neural network, For observing noise, t represents time t;

[0054] Solve for the expected remaining lifetime.

[0055]

[0056] in, Indicates the current moment. express After a moment, Indicates the degradation threshold;

[0057] Uncertainty was quantified using Monte Carlo simulation, generating the expected remaining lifespan probability distribution as follows:

[0058]

[0059] in, This indicates that the expected remaining lifespan is less than or equal to The probability; N is the preset number of attempts; It is an indicator function, indicating If the condition within the function is true, the function value is 1; otherwise, it is 0. This represents the remaining lifetime value obtained from the nth Monte Carlo simulation.

[0060] Optionally, the security objective function is:

[0061]

[0062] in, For bridge i at time The probability of failure; , indicates whether bridge i needs maintenance, 0 means no maintenance, 1 means maintenance;

[0063] The objective function for total maintenance cost is:

[0064]

[0065] in, This includes labor costs, material costs, and losses due to traffic control measures.

[0066] The traffic disturbance objective function is:

[0067]

[0068] in, To maintain the impact on traffic flow;

[0069] The preset constraints include:

[0070] Budget constraints: Where B is the budget threshold;

[0071] Human resource constraints: ,in, Maintain project thresholds;

[0072] Failure risk constraints: ,in, This represents the failure probability threshold.

[0073] Secondly, the present invention also provides an intelligent health maintenance prediction system for bridge inspection, comprising:

[0074] The point cloud processing module is used to extract image feature points from bridge images uploaded by drones during inspections to obtain bridge point clouds.

[0075] The crack skeleton extraction module is used to render the bridge point cloud and extract the crack centerline skeleton.

[0076] The crack size determination module is used to determine the crack size based on the extracted crack centerline skeleton, wherein the crack size includes crack length, crack width and crack depth;

[0077] The distance determination module is used to fit a local reference plane to the Gaussian body of the healthy surface around the disease based on the identified cracks and determine the directed distance from the Gaussian body of the diseased area to the local reference plane.

[0078] The volume determination module is used to mesh the diseased area and determine the volume of the diseased area based on the directed distance and the mesh area.

[0079] The health index determination module is used to comprehensively evaluate the health status of bridges based on defect performance data, structural performance data, and durability performance data, and to determine the bridge health index. Defect performance data includes crack size and the volume of defect area.

[0080] The health index prediction module is used to input the current bridge health index, disease performance data, structural performance data, durability performance data and time series characteristics into the disease evolution prediction model to obtain the predicted health index corresponding to the time series characteristics;

[0081] The remaining life analysis module is used to generate degradation data based on predicted health indices, and to determine the probability distribution of the bridge's remaining life using degradation models and remaining life estimation models, so as to facilitate health maintenance of the bridge.

[0082] This invention provides an intelligent health maintenance prediction method and system for bridge inspection. Compared with existing technologies, it has the following advantages:

[0083] Bridge point clouds are obtained by extracting image feature points from bridge images uploaded by drones during inspections. This leverages the rapid coverage capability of drones to avoid the risks of high-altitude operations and generates high-precision 3D data as the basis for subsequent analysis. The bridge point cloud is then rendered to extract the crack centerline skeleton. Through visualization processing, the complex crack structure is simplified into a quantifiable skeleton, providing a clear path for accurate measurement. Crack dimensions, including length, width, and depth, are determined based on the extracted crack centerline skeleton, ensuring comprehensive capture of crack geometric features. Based on the identified cracks, a Gaussian fit is applied to the healthy surface surrounding the defect to a local reference plane, and a directed distance is determined. The defect area is then meshed, and the directed distance is calculated... The volume is determined by the area of ​​the grid, and continuous defects are discretized into computable units. Combining depth and area enables a three-dimensional measurement of defect severity. A bridge health index is determined based on defect performance data, structural performance data, and durability performance data, making health assessments more aligned with engineering realities. The current health index and related data are input into a defect evolution prediction model to obtain a predicted health index. Temporal characteristics are used to capture defect degradation patterns, giving the prediction results dynamic evolution capabilities. Degradation data is generated based on the predicted health index, and the probability distribution of remaining life is determined using a degradation model and a remaining life model. The uncertainty of maintenance timing is quantified through probabilistic output, providing a scientific basis for decision-making. By constructing a predictive maintenance process and using UAV inspection data to drive defect quantification and evolution prediction, the problem of traditional inspections lacking dynamic prediction capabilities is solved, realizing a shift from passive inspection to proactive early warning. Attached Figure Description

[0084] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0085] Figure 1 A flowchart illustrating an intelligent health maintenance prediction method for bridge inspection provided in an embodiment of the present invention;

[0086] Figure 2 This is a schematic diagram of the structure of an intelligent health maintenance prediction system for bridge inspection provided in an embodiment of the present invention. Detailed Implementation

[0087] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application are described clearly and completely. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0088] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0089] like Figure 1 As shown in the embodiment of this application, an intelligent health maintenance prediction method for bridge inspection includes:

[0090] S1: Based on the bridge images uploaded by the drone inspection, extract image feature points to obtain the bridge point cloud.

[0091] S2: Render the bridge point cloud and extract the crack centerline skeleton.

[0092] S3: Determine the crack size based on the extracted crack centerline skeleton, wherein the crack size includes crack length, crack width and crack depth.

[0093] S4: Based on the identified cracks, fit a local reference plane to the Gaussian body of the healthy surface around the disease and determine the directed distance from the Gaussian body of the diseased area to the local reference plane.

[0094] S5: Grid the diseased area and determine the volume of the diseased area based on the directed distance and grid area.

[0095] S6: Based on the data on defects, structural performance, and durability, the fuzzy hierarchical analysis method is used to comprehensively evaluate the health status of the bridge and determine the bridge health index. The data on defects include crack size and the volume of the defect area.

[0096] S7: Input the current bridge health index, disease performance data, structural performance data, durability performance data, and time series characteristics into the disease evolution prediction model to obtain the predicted health index corresponding to the time series characteristics.

[0097] S8: Based on the predicted health index, degradation data is generated. Using the degradation model and the remaining life estimation model, the probability distribution of the remaining life of the bridge is determined to facilitate health maintenance of the bridge.

[0098] In this optional embodiment, the method constructs a predictive maintenance process, utilizing UAV inspection data to drive the quantification and evolution prediction of defects, thereby solving the problem of traditional inspections lacking dynamic prediction capabilities and realizing a transformation from passive detection to proactive early warning. Specifically, image feature points are extracted from bridge images uploaded by UAV inspections to obtain bridge point clouds. This utilizes the rapid coverage capability of UAVs to avoid the risks of high-altitude operations and generates high-precision 3D data as the basis for subsequent analysis. The bridge point cloud is rendered to extract the crack centerline skeleton. Through visualization processing, the complex crack structure is simplified into a quantifiable skeleton, providing a clear path for accurate measurement. The crack size, including length, width, and depth, is determined based on the extracted crack centerline skeleton. This ensures the comprehensive capture of crack geometric features. The width is calculated based on the arc length parameter of the skeleton, and the depth is derived by combining the viewing angle difference, making the quantification results more consistent with the actual defect morphology. Based on the identified cracks, a Gaussian body is fitted to the healthy surface around the defect to form a local reference plane and a directed distance is determined. A local reference system is established by referencing the healthy area, so that the directed distance can objectively reflect the degree of deviation of the defect from the intact surface, avoiding the need for a complete measurement of the defect's position. Errors introduced by the local benchmark are addressed through a multi-layered approach. The diseased area is gridded, and its volume is determined based on directed distance and grid area. Continuous diseases are discretized into computable units, and depth and area are combined to achieve a three-dimensional measurement of disease severity. A health index is determined using fuzzy hierarchical analysis based on disease performance data (including crack size and volume), structural performance data, and durability performance data. This method handles the uncertainty of multi-source data through fuzzy processing and assigns reasonable weights to key parameters such as disease volume, making health assessments more consistent with engineering realities. The current health index and related data are input into a disease evolution prediction model to obtain a predicted health index. Temporal characteristics are used to capture disease degradation patterns, giving the prediction results dynamic evolution capabilities. Degradation data is generated based on the predicted health index, and the probability distribution of remaining life is determined using a degradation model and a remaining life model. The uncertainty of maintenance timing is quantified through probabilistic output, providing a scientific basis for decision-making. The entire process achieves full-cycle management of diseases from identification and quantification to prediction through a closed-loop data chain, fundamentally changing the passive maintenance model.

[0099] The following is a detailed description of each step.

[0100] S1: Based on the bridge images uploaded by the drone inspection, extract image feature points to obtain the bridge point cloud.

[0101] Specifically, the drone can employ a tightly coupled and integrated navigation architecture of V-SLAM / LiDAR-SLAM, RTK-GNSS, and IMU to achieve centimeter-level precise positioning in all scenarios and all weather conditions. For image acquisition, the drone can be equipped with a high-definition camera, an IMU (Inertial Measurement Unit, consisting of a 6-axis gyroscope and accelerometer), and other sensor data acquisition devices. The SIFT (Scale-Invariant Feature Transform) algorithm is used to extract image feature points. SIFT features are invariant to scale, rotation, and illumination changes, making them suitable for complex scenarios like bridges. The COLMAP algorithm can also be used to generate point clouds: selecting the image pair with the most matches, estimating the relative pose using the five-point method, and triangulating to generate an initial sparse point cloud; iteratively selecting the new image with the most matches to the reconstructed point cloud, estimating the camera pose using the PnP (Perspective-n-Point) algorithm; triangulating the matching points between the newly registered image and existing images, adding new 3D points, repeating this process multiple times, and outputting a sparse point cloud.

[0102] S2: Render the bridge point cloud and extract the crack centerline skeleton.

[0103] Specifically, each point is initialized as a 3D Gaussian distribution to obtain a reconstructed Gaussian model; for the reconstructed Gaussian model, a high-resolution orthophoto (GSD=1mm / pixel) is rendered along the normal direction of the bridge surface to eliminate perspective distortion; the DeepLabV3+ model is used to perform pixel-level segmentation on the orthophoto to extract the crack region mask; the crack mask is morphologically refined to extract the crack centerline skeleton.

[0104] S3: Determine the crack size based on the extracted crack centerline skeleton, wherein the crack size includes crack length, crack width and crack depth.

[0105] Specifically, the crack length is calculated cumulatively along the skeleton, and the crack length is...

[0106]

[0107] in, It is a sequence of skeleton pixel coordinates.

[0108] The crack boundary spacing was measured along the skeleton normal direction, and sub-pixel edge detection was used. The crack width was...

[0109]

[0110] Where s is the skeleton arc length parameter, and GSD is the ground sampling distance. The average width, maximum width, and width standard deviation can also be calculated further.

[0111] The crack depth is reconstructed from images taken from different angles using multi-view geometric constraints. Assuming the crack is a V-groove, the crack depth is...

[0112]

[0113] in, width, Due to the difference in viewing angle, the crack depth is estimated by fusing multi-view triangulation and Gaussian depth information, with an accuracy of ±0.5 mm.

[0114] S4: Based on the identified cracks, fit a local reference plane to the Gaussian body of the healthy surface around the disease and determine the directed distance from the Gaussian body of the diseased area to the local reference plane.

[0115] Specifically, for the Gaussian surface surrounding the diseased area, a local reference plane or surface is fitted, such as z=ax+by+c or z=ax. 2 +by 2 +cxy+dx+ey+f. RANSAC robust fitting is used to eliminate the influence of out-of-point points in the diseased area.

[0116] For the Gaussian body in the diseased area, calculate its directed distance to the reference plane, where the directed distance is...

[0117]

[0118] Where n is the normal to the local reference plane, Let a point be on the local reference plane. Gaussian body The projection point on the local reference plane.

[0119] S5: Grid the diseased area and determine the volume of the diseased area based on the directed distance and grid area.

[0120] Specifically, the affected area is meshed (e.g., a 1mm × 1mm mesh), and the volume is calculated cumulatively. The volume of the affected area is...

[0121]

[0122] in, Let k be the average depth of the grid. This represents the mesh area. Additionally, for complex shapes, Delaunay triangulation combined with tetrahedral volume summation is used.

[0123] Accuracy verification: Compared with the measured data of the laser scanner, the relative error of volume measurement is <8%, which meets the requirements of engineering evaluation.

[0124] S6: Based on the data on defects, structural performance, and durability, the fuzzy hierarchical analysis method is used to comprehensively evaluate the health status of the bridge and determine the bridge health index. The data on defects include crack size and the volume of the defect area.

[0125] Specifically, the Fuzzy Analytic Hierarchy Process (AHP) comprehensively assesses the health status of bridges. The assessment index system includes data on defects, structural performance, and durability. Defect performance data includes crack severity, spalling / exposed reinforcement degree, corrosion area ratio, and other defects; structural performance data includes load-bearing capacity, stiffness degradation, and changes in dynamic characteristics; durability performance data includes material aging degree, environmental corrosion effects, and fatigue cumulative damage.

[0126] The bridge health index is:

[0127]

[0128] in, As the indicator weight, For individual scoring, scores can be calculated using corresponding functions or based on experience. Indicator weights can be determined using expert questionnaires and AHP (Analytic Hierarchy Process). For example, the weight of the index corresponding to crack severity in the defect performance data can be set to 0.2, the weight of the index corresponding to spalling / exposed reinforcement degree can be set to 0.12, the weight of the index corresponding to load-bearing capacity in the structural performance data can be set to 0.15, and the weight of the index corresponding to material aging degree in the durability performance data can be set to 0.1. Other indicators can be assigned similar weights, but the overall weight is 1.

[0129] Taking cracks as an example, the severity score for cracks is:

[0130]

[0131] in, The crack width is represented by this piecewise function, which reflects the engineering grading standard for crack width (0.1mm for minor, 0.5mm for moderate, and 2.0mm for severe).

[0132] For example, the weighting is as follows: defect performance data 0.45, structural performance data 0.35, and durability performance data 0.20. Defect performance data: Crack severity: average width 3.2mm, score 48 (moderate deviation); Spalling / Exposed reinforcement: 43 instances of spalling in the main cable coating, score 62 (slight deviation); Rust area ratio: total rust area of ​​the steel box girder surface 5.96 m² / total surface area 12,450 m² = 0.048%, score 75 (good); Other defects: expansion joints, supports, bolts, etc., comprehensive score 68 (good). Structural performance data: Load-bearing capacity: Based on SHM strain data, the load-bearing capacity index is 0.92 (relative to design value), with a score of 82 (excellent); Stiffness degradation: Measured mid-span deflection / theoretical = 1.08, indicating slight stiffness degradation, with a score of 76 (good); Dynamic characteristics: First-order frequency is 0.186 Hz (design value 0.192 Hz), a frequency decrease of 3.1%, with a score of 78 (good). Durability performance data: Material aging: Concrete carbonation depth is 8.5 mm, steel corrosion rate is 0.048%, with a score of 72 (good); Environmental corrosion: Chloride ion corrosion from the Yangtze River, with 43 instances of peeling off the protective coating, with a score of 66 (caution); Fatigue cumulative damage: Calculated based on the Miner criterion, the cumulative damage degree is 0.23, with a score of 70 (good).

[0133] Bridge health index HI = 0.45 × 58 + 0.35 × 79 + 0.20 × 69 = 67.8

[0134] S7: Input the current bridge health index, disease performance data, structural performance data, durability performance data, and time series characteristics into the disease evolution prediction model to obtain the predicted health index corresponding to the time series characteristics.

[0135] Specifically, the disease evolution prediction model is a stacked long short-term memory (LSTM) network, which includes an input layer, a first LSTM layer, a dropout layer, a second LSTM layer, a fully connected layer, and an output layer. Multiple stacked LSTM networks can be trained to predict health indices for the next 6 / 12 / 24 months, and the predicted health index can be represented as...

[0136]

[0137] Monte Carlo dropout uncertainty estimation: When training a stacked long short-term memory network, retain the dropout layer and perform a preset number of forward propagations (N=100): ,

[0138] The expected value is

[0139]

[0140] Where N is the preset number of times, This is the predicted health index for the next h months obtained from the nth training iteration.

[0141] The uncertainty of the forecast is

[0142]

[0143] The 95% confidence interval was obtained:

[0144] .

[0145] The length of the confidence interval can characterize the fluctuation of the model; a longer interval indicates poorer model stability and accuracy.

[0146] S8: Based on the predicted health index, degradation data is generated. Using a degradation model and a remaining life estimation model, the probability distribution of the bridge's remaining life is determined to facilitate health maintenance of the bridge. The degradation model is a recurrent neural network. This step includes:

[0147] S81: Input the predicted health index and expected remaining lifespan into the trained recurrent neural network to obtain the deterioration health index.

[0148] Specifically, the degenerative health index is

[0149]

[0150] in, For the degeneracy function of a recurrent neural network, For the parameters of the recurrent neural network, For the observation of noise, t represents time t.

[0151] S82: Based on the deterioration health index and the deterioration threshold, determine the probability distribution of expected remaining lifespan that is less than the deterioration threshold.

[0152] Specifically, a degradation threshold of 30 is set, and when the degradation health index is below 30, it is considered that mandatory repair is required.

[0153] Solve for the expected remaining lifetime.

[0154]

[0155] in, Indicates the current moment. express After a moment, This represents the degradation threshold.

[0156] Uncertainty was quantified using Monte Carlo simulation, generating the expected remaining lifespan probability distribution as follows:

[0157]

[0158] in, This indicates that the expected remaining lifespan is less than or equal to The probability; N is the preset number of attempts; It is an indicator function, indicating If the condition within the function is true, the function value is 1; otherwise, it is 0. This represents the remaining lifetime value obtained from the nth Monte Carlo simulation.

[0159] By utilizing the inherent recurrent structure of recurrent neural networks to process sequential data, the nonlinear dependence of the health index on changes over time can be adaptively learned. This overcomes the limitation of traditional static models in failing to capture the dynamic characteristics of bridge degradation processes, making the model more closely aligned with the gradual development of defects in actual service. Secondly, the predicted health index and expected remaining lifespan are input into the trained recurrent neural network to obtain the degradation health index. This process, based on a joint input mechanism of historical prediction data and future expected targets, allows the degradation function to dynamically adjust the generation of the degradation trajectory, avoiding the one-sidedness of relying solely on current state predictions. This ensures that the generated degradation health index reflects both historical evolution trends and future degradation directions. Finally, the probability distribution of expected remaining lifespan below the degradation threshold is determined by inversely using the degradation health index and degradation threshold. Through a reverse inference mechanism, the degradation health index is correlated with a preset safety threshold, achieving an inverse probability mapping from degradation state to remaining lifespan. This not only quantifies the uncertainty range of the prediction results but also provides a risk assessment basis at the probability distribution level for maintenance decisions, thereby supporting more refined remaining lifespan management.

[0160] In an optional embodiment of this application, the intelligent health maintenance prediction method for bridge inspection further includes:

[0161] S9: Construct a safety objective function based on the probability distribution of remaining lifetime.

[0162] Specifically, the objective function for total maintenance cost (i.e., minimizing total maintenance cost) is:

[0163]

[0164] in, Costs include labor costs, material costs, and losses due to traffic control, and vary over time, with higher costs during peak seasons.

[0165] The security objective function (i.e., maximizing security, which is equivalent to minimizing the risk of failure) is:

[0166]

[0167] in, For bridge i at time The probability of failure; , indicates whether bridge i is to be maintained, 0 means no maintenance, 1 means maintenance.

[0168] The traffic interference objective function (i.e., minimizing traffic interference) is:

[0169]

[0170] in, To maintain the impact on traffic flow, historical traffic flow data is used.

[0171] The preset constraints include:

[0172] Budget constraints: Where B is the budget threshold.

[0173] Human resource constraints: ,in, Maintenance project threshold, maximum number of times during the same period Several projects are running in parallel.

[0174] Failure risk constraints: ,in, This represents the failure probability threshold.

[0175] S10: Based on the total maintenance cost objective function, safety objective function, and traffic disturbance objective function, and under preset constraints, solve the multi-objective optimization function to obtain multiple objective values ​​for the maintenance scheme, so that decision-makers can select the maintenance scheme based on the objective values.

[0176] Specifically, the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) is used to solve the multi-objective optimization problem.

[0177] Pareto Domination: Solution Dominate , recorded as ,like , ; .

[0178] Pareto optimal front: the set of all non-dominated solutions. NSGA-II maintains a diverse set of Pareto optimal solutions through fast non-dominated sorting and crowding distance selection, allowing decision-makers to choose according to their actual preferences.

[0179] The algorithm flow of Non-Dominated Sorting Genetic Algorithm II (NSGA-II):

[0180] 1. Initialize the population P0 (size N=100) and randomly generate a maintenance scheme.

[0181] 2. For generation t=1, 2, …, Tmax. Specifically includes:

[0182] Offspring Qt (size N) are generated through selection, crossover, and mutation; merging... (Scale 2N); Fast non-dominated sorting, calculate the dominance level F1, F2, … for each solution; select from low to high dominance level until the population is full; calculate the crowding distance for the last level, select solutions with high crowding to maintain diversity; form a new population Pt+1;

[0183] 3. Output the final Pareto front. Decision-makers can select the Pareto front; example output:

[0184] Option A: Low cost (¥2 million) but slightly higher risk (5% failure probability);

[0185] Option B: Moderate cost (¥3 million), low risk (1% failure probability), minimal traffic disruption;

[0186] Option C: High cost (¥4.5 million) but extremely low risk (0.2% failure probability) and minimal traffic disruption.

[0187] Choose flexibly based on budget and risk tolerance.

[0188] In this embodiment, a multi-objective optimization framework is constructed to address the problem of multi-dimensional objective synergistic optimization in maintenance decision-making, shifting the selection of maintenance schemes from experience-driven to data-driven. Specifically, a safety objective function is constructed based on the probability distribution of remaining lifespan. The dynamic characteristics of the probability distribution are used to accurately quantify the bridge failure risk, transforming the uncertainty of remaining lifespan into a calculable safety indicator. This avoids the rigidity of traditional static safety assessments and ensures that safety objectives can be adjusted in real time according to the bridge's degradation status. Optimization is performed based on the total maintenance cost objective function, the safety objective function, and the traffic interference objective function, integrating the three core dimensions of economy, safety, and social impact. Multi-objective synergy avoids the local optimum problem caused by single-objective optimization, enabling maintenance schemes to comprehensively reflect actual engineering needs. The multi-objective optimization function is solved under preset constraints, introducing hard constraints such as budget, human resources, and failure risk constraints to ensure that the optimization process strictly follows the engineering feasibility boundary, preventing resource overruns or safety risk exceeding thresholds. Multiple objective values ​​for the maintenance scheme are obtained for decision-makers to choose from, and a quantified set of objective values ​​is output, providing decision-makers with transparent and comparable optimization results. This supports data-driven trade-off analysis and significantly improves the objectivity and scientific nature of maintenance scheme selection.

[0189] like Figure 2 As shown in the embodiment of this application, an intelligent health maintenance prediction system for bridge inspection includes:

[0190] The point cloud processing module 10 is used to extract image feature points from bridge images uploaded by UAV inspections to obtain bridge point clouds.

[0191] The crack skeleton extraction module 20 is used to render the bridge point cloud and extract the crack centerline skeleton.

[0192] The crack size determination module 30 is used to determine the crack size based on the extracted crack centerline skeleton, wherein the crack size includes crack length, crack width and crack depth.

[0193] The distance determination module 40 is used to fit a local reference plane to the Gaussian body of the healthy surface around the disease based on the identified cracks and determine the directed distance from the Gaussian body of the diseased area to the local reference plane.

[0194] The volume determination module 50 is used to mesh the diseased area and determine the volume of the diseased area based on the directed distance and the mesh area.

[0195] The health index determination module 60 is used to comprehensively evaluate the health status of the bridge based on the defect performance data, structural performance data, and durability performance data, and to determine the bridge health index. The defect performance data includes crack size and the volume of the defect area.

[0196] The health index prediction module 70 is used to input the current bridge health index, disease performance data, structural performance data, durability performance data and time series characteristics into the disease evolution prediction model to obtain the predicted health index corresponding to the time series characteristics.

[0197] The remaining life analysis module 80 is used to generate degradation data based on the predicted health index, and to determine the probability distribution of the remaining life of the bridge using the degradation model and the remaining life estimation model, so as to facilitate health maintenance of the bridge.

[0198] In this embodiment, the system constructs a complete technology chain from data acquisition to predictive decision-making, realizing dynamic monitoring and proactive early warning of bridge health status, fundamentally changing the passive response mode of traditional inspections. The point cloud processing module 10 extracts image feature points from bridge images uploaded by UAV inspections to generate a bridge point cloud. This utilizes the efficient data acquisition capabilities of UAVs, avoiding the risks of manual high-altitude operations, and simultaneously transforms two-dimensional images into precise three-dimensional point cloud data, providing a high-fidelity geometric foundation for subsequent analysis, solving the inefficiency problem caused by traditional inspections relying on manual measurement. The crack skeleton extraction module 20 renders the bridge point cloud and extracts the crack centerline skeleton. This visualizes the crack features, structuring them and transforming the originally blurry image cracks into quantifiable linear skeletons, providing a clear path for size calculation and avoiding subjective errors from manual visual inspection. The crack size determination module 30 determines the crack size, including length, width, and depth, based on the extracted crack centerline skeleton. The length is calculated based on the skeleton pixel coordinate sequence, and the depth is derived by combining the ground sampling distance and viewing angle difference. This ensures the objectivity and accuracy of the size measurement. Viewing angle difference compensation is specifically introduced for depth calculation, overcoming the depth estimation bias caused by a single image viewpoint and providing a reliable quantitative basis for disease assessment. The distance determination module 40 fits a local reference plane to the Gaussian body of the healthy surface surrounding the disease based on the identified crack and determines the directed distance. This constructs a reference plane through a local healthy region, projects the Gaussian body of the diseased region onto this plane to calculate the directed distance, effectively isolating global deformation interference and accurately quantifying the local depth deviation of the disease, laying the geometric foundation for volume calculation. The volume determination module 50 meshes the diseased region and determines the volume of the diseased region based on the directed distance and grid area. This discretizes the continuous disease through grid division and combines the average depth and area of ​​each grid for precise integration, achieving a three-dimensional quantification of disease severity, reflecting structural damage more comprehensively than a single size indicator. The health index determination module 60 comprehensively assesses the health status based on disease performance data (including crack size and volume), structural performance data, and durability performance data using fuzzy hierarchical analysis. This method addresses the fuzziness and weight uncertainty between indicators through multi-source data fusion, incorporating quantitative results such as disease volume into the scoring system to generate an objective bridge health index, avoiding the one-sidedness of relying solely on surface diseases in traditional assessments. The health index prediction module 70 inputs the current health index, performance data, and time-series characteristics into the disease evolution prediction model to obtain the predicted health index. This method utilizes time-series characteristics to capture historical trends and dynamically extrapolates the evolution of the health status through models such as stacked long short-term memory networks, achieving a leap from static assessment to trend prediction and providing a basis for early warning.The remaining life analysis module 80 generates degradation data based on predicted health indices and uses degradation models and remaining life estimation models to determine the probability distribution of remaining life. This model uses recurrent neural networks to model the health degradation process and combines probability distributions to quantify the uncertainty of remaining life, enabling maintenance decisions to shift from experience-based judgment to data-driven risk management and supporting the proactive development of maintenance plans.

[0199] Experimental analysis was conducted, and a comparison table of the quantitative accuracy of disease was obtained, as shown in Table 1.

[0200] Table 1 Comparison of Disease Quantification Accuracy

[0201]

[0202] The method for quantifying defects based on Gaussian body parameters improves accuracy by 63%-75% compared to traditional mesh model-based measurement methods. Key innovations include: utilizing the continuous geometric representation of Gaussian bodies and multi-view depth information to estimate crack depth through triangulation, overcoming the limitation of traditional methods that can only measure two-dimensional parameters; and a complete technical chain of datum plane fitting, depth map generation, and volume integration, enabling automated calculation of spalling / corrosion volume, reducing the relative error from 15-28% to 5-8%, meeting the needs of refined engineering assessments.

[0203] The accuracy of disease evolution prediction (based on data from 20 bridges tracked for 36 consecutive months) is shown in Table 2.

[0204] Table 2 Prediction Accuracy Analysis Table

[0205]

[0206] By modeling the temporal evolution of the health index using Stacked LSTM, the mean absolute error (MAE) for 6-month short-term prediction is 3.2 (HI range 0-100), with a relative error of 3.2%; the MAE for 12-month medium-term prediction is 5.8, with a relative error of 5.8%; and the MAE for 24-month long-term prediction is 9.3, with a relative error of 9.3%, all meeting engineering decision-making requirements. The root mean square error (RMSE) for crack width prediction is <0.1mm within 6 months, providing early warning of whether cracks will exceed control thresholds (e.g., 0.5mm). The relative length of failure (RUL) prediction is ±1.5 months within 12 months, providing sufficient lead time for maintenance planning. Compared to traditional passive periodic inspections (inspecting on schedule and repairing upon discovery), this application achieves "predictive maintenance," which can detect potential risks 6-12 months in advance, avoiding sudden accidents; optimize maintenance timing, intervening in the early stages of disease development to reduce maintenance costs; and rationally allocate resources, avoiding reactive, piecemeal responses.

[0207] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting the intelligent health maintenance of bridges during inspection, characterized in that, include: Based on the bridge images uploaded by the drone inspection, image feature points are extracted to obtain the bridge point cloud; Render the bridge point cloud and extract the crack centerline skeleton; Based on the extracted crack centerline skeleton, the crack size is determined, including crack length, crack width, and crack depth. Based on the identified cracks, fit a local reference plane to the Gaussian volume of the healthy surface around the disease and determine the directed distance from the Gaussian volume of the diseased area to the local reference plane. The affected area is gridded, and the volume of the affected area is determined based on the directed distance and grid area. Based on defect performance data, structural performance data, and durability performance data, the fuzzy hierarchical analysis method is used to comprehensively evaluate the health status of bridges and determine the bridge health index. Defect performance data includes crack size and the volume of defect area. The current bridge health index, disease performance data, structural performance data, durability performance data, and time series characteristics are input into the disease evolution prediction model to obtain the predicted health index corresponding to the time series characteristics. Based on the predicted health index to generate degradation data, the degradation model and the remaining life estimation model are used to determine the probability distribution of the remaining life of the bridge, so as to facilitate the health maintenance of the bridge. The length of the crack is in, It is a sequence of skeleton pixel coordinates; The width of the crack is Where s is the skeleton arc length parameter, and GSD is the ground sampling distance; Assuming the crack is a V-shaped groove, the crack depth is in, width, Due to the difference in perspective; The directed distance is Where n is the normal to the local reference plane, Let a point be on the local reference plane. Gaussian body The projection point on the local reference plane; The volume of the diseased area is in, Let k be the average depth of the grid. This represents the grid area.

2. The intelligent health maintenance prediction method for bridge inspection as described in claim 1, characterized in that, The degradation model is a recurrent neural network; The process of generating degradation data based on predicted health indices, and using degradation models and remaining life estimation models to determine the probability distribution of the bridge's remaining life includes: The predicted health index and expected life expectancy are input into the trained recurrent neural network to obtain the deterioration health index. Based on the deterioration health index and the deterioration threshold, the probability distribution of expected remaining lifespan below the deterioration threshold is determined in reverse.

3. The intelligent health maintenance prediction method for bridge inspection as described in claim 2, characterized in that, Also includes: Construct a safety objective function based on the probability distribution of remaining lifetime; Based on the total maintenance cost objective function, the safety objective function, and the traffic disturbance objective function, and under preset constraints, a multi-objective optimization function is solved to obtain multiple objective values ​​for the maintenance scheme, so that decision-makers can select the maintenance scheme based on the objective values.

4. The intelligent health maintenance prediction method for bridge inspection as described in claim 1, characterized in that, The bridge health index is: in, As the indicator weight, Individual scoring; Crack severity rating: in, The width of the crack.

5. The intelligent health maintenance prediction method for bridge inspection as described in claim 1, characterized in that, The disease evolution prediction model is a stacked long short-term memory network, which includes an input layer, a first long short-term memory layer, a discard layer, a second long short-term memory layer, a fully connected layer, and an output layer. When training a stacked long short-term memory network, the dropout layer is retained, and a preset number of forward propagations are performed to determine the uncertainty estimation results of Monte Carlo dropout. The expected value is Where N is the preset number of times, The predicted health index for the next h months is obtained from the nth training iteration; The uncertainty of the forecast is The 95% confidence interval was obtained: 。 6. The intelligent health maintenance prediction method for bridge inspection as described in claim 3, characterized in that, The degenerative health index is: in, For the degeneracy function of a recurrent neural network, For the parameters of the recurrent neural network, For observing noise, t represents time t; Solve for the expected remaining lifetime. in, Indicates the current moment. express After a moment, Indicates the degradation threshold; Uncertainty was quantified using Monte Carlo simulation, generating the expected remaining lifespan probability distribution as follows: in, This indicates that the expected remaining lifespan is less than or equal to The probability; N is the preset number of attempts; It is an indicator function, indicating If the condition within the function is true, the function value is 1; otherwise, it is 0. This represents the remaining lifetime value obtained from the nth Monte Carlo simulation.

7. The intelligent health maintenance prediction method for bridge inspection as described in claim 6, characterized in that, The objective function for total maintenance cost is: in, This includes labor costs, material costs, and losses due to traffic control measures. The security objective function is: in, For bridge i at time The probability of failure; , indicates whether bridge i needs maintenance, 0 means no maintenance, 1 means maintenance; The traffic disturbance objective function is: in, To maintain the impact on traffic flow; The preset constraints include: Budget constraints: Where B is the budget threshold; Human resource constraints: ,in, Maintain project thresholds; Failure risk constraints: ,in, This represents the failure probability threshold.

8. An intelligent health maintenance prediction system for bridge inspection, characterized in that, include: The point cloud processing module is used to extract image feature points from bridge images uploaded by drones during inspections to obtain bridge point clouds. The crack skeleton extraction module is used to render the bridge point cloud and extract the crack centerline skeleton. The crack size determination module is used to determine the crack size based on the extracted crack centerline skeleton, wherein the crack size includes crack length, crack width and crack depth; The distance determination module is used to fit a local reference plane to the Gaussian body of the healthy surface around the disease based on the identified cracks and determine the directed distance from the Gaussian body of the diseased area to the local reference plane. The volume determination module is used to mesh the diseased area and determine the volume of the diseased area based on the directed distance and the mesh area. The health index determination module is used to comprehensively evaluate the health status of bridges based on defect performance data, structural performance data, and durability performance data, and to determine the bridge health index. Defect performance data includes crack size and the volume of defect area. The health index prediction module is used to input the current bridge health index, disease performance data, structural performance data, durability performance data and time series characteristics into the disease evolution prediction model to obtain the predicted health index corresponding to the time series characteristics; The remaining life analysis module is used to generate degradation data based on predicted health indices, and to determine the probability distribution of the remaining life of the bridge using degradation models and remaining life estimation models, so as to facilitate health maintenance of the bridge. The length of the crack is in, It is a sequence of skeleton pixel coordinates; The width of the crack is Where s is the skeleton arc length parameter, and GSD is the ground sampling distance; Assuming the crack is a V-shaped groove, the crack depth is in, width, Due to the difference in perspective; The directed distance is Where n is the normal to the local reference plane, Let a point be on the local reference plane. Gaussian body The projection point on the local reference plane; The volume of the diseased area is in, Let k be the average depth of the grid. This represents the grid area.

Images