A method and system for real-time location and adaptive repair triggering of leaks in metal roofs

Abstract

This application discloses a method and system for real-time location and adaptive repair triggering of leaks in metal roofs. The method includes: S1: real-time acquisition of status data through a multimodal sensing network deployed on the metal roof; S2: identification of potential leakage areas; S3: precise location of leakage points; S4: dynamic assessment and classification of leakage risk; S5: generation of adaptive repair decisions based on a probabilistic principal component analysis model according to the risk level, and determination of repair strategy parameters; S6: control of a repair execution device to accurately deliver repair materials according to the repair strategy parameters, and real-time monitoring of the repair effect. This application constructs a dynamic comprehensive risk value calculation model by integrating real-time leakage rate, impact range estimation, and environmental factor correction, realizing a leap from qualitative judgment to quantitative classification of leakage risk, providing a scientific basis for differentiated and precise repair decisions, and avoiding waste or inadequacy of repair resources.

Description

Technical Field

[0001] This application belongs to the field of building waterproofing and intelligent maintenance technology, specifically relating to a method and system for real-time location and adaptive repair triggering of leaks in metal roofs. Background Technology

[0002] Metal roofing is widely used in large industrial plants, transportation hubs, stadiums, and other buildings due to its advantages such as lightweight, high strength, and convenient construction. However, affected by factors such as thermal expansion and contraction of materials, structural deformation, fatigue of connection nodes, environmental corrosion, and extreme weather, metal roofing is prone to leakage at joints, fasteners, and gutters. Leaking water not only corrodes the insulation layer and reduces building energy efficiency, but may also damage indoor equipment, cause electrical short circuits, and even induce corrosion of the steel structure, seriously threatening the safety and service life of the building.

[0003] Currently, the detection of leaks in metal roofs mainly relies on methods such as regular manual inspections, infrared thermal imaging surveys, or localized potential detection. Manual inspections are highly subjective, inefficient, and struggle to detect hidden or micro-leaks. While infrared thermal imaging can detect leaks through temperature anomalies, it is easily affected by ambient temperature, sunlight, wind speed, and the reflectivity of roof materials, resulting in a high false alarm rate and difficulty in accurately locating the leak source. Single potential sensing networks can detect moisture intrusion through resistivity changes, but are susceptible to interference from environmental temperature and humidity, surface contamination, and other factors, easily generating false signals and failing to assess the severity of leaks. Furthermore, existing detection methods mostly remain at the "discovery" stage, lacking a closed-loop system linked to repair decisions. Repair work still heavily relies on human experience; the timing of repairs, material selection, and scope of work lack quantitative basis, leading to untimely or excessive repairs and hindering precise, efficient, and adaptive leak control.

[0004] Therefore, there is an urgent need to develop an integrated method and system that can achieve real-time and accurate leakage location and intelligent triggering and optimization of repair operations, so as to improve the automation, intelligence and long-term reliability of metal roof operation and maintenance. Summary of the Invention

[0005] This application provides a real-time location and adaptive repair triggering method and system for metal roof leakage, aiming to solve the problems of high false alarm rate, difficulty in accurately locating the source of leakage, susceptibility to interference from environmental temperature and humidity, surface contamination, etc., easy generation of false signals, and inability to assess the severity of leakage in the existing technology.

[0006] A first aspect includes a method for real-time location and adaptive repair triggering of leaks in metal roofs, the method comprising:

[0007] S1: Real-time status data is collected through a multimodal sensing network deployed on the metal roof. The multimodal sensing network includes a distributed potential sensing network, a visual monitoring unit, and an environmental parameter acquisition unit.

[0008] S2: Based on the data collected by the potential sensing network, resistivity time-series characteristics are compared to identify potential leakage areas;

[0009] S3: For the potential leakage area, perform multimodal data fusion analysis by fusing potential data, infrared thermal image data and visible light images to accurately locate the leakage point;

[0010] S4: Based on the location of the leak point, combined with environmental parameters and historical leak data, conduct dynamic assessment and classification of leak risk;

[0011] S5: Based on the risk level, generate an adaptive remediation decision using a probabilistic principal component analysis model and determine the remediation strategy parameters;

[0012] S6: Based on the repair strategy parameters, control the repair execution device to accurately deliver the repair material and monitor the repair effect in real time;

[0013] S7: Based on the verification results of the repair effect and historical data, perform self-learning and optimization of system parameters, and update the risk assessment model and repair decision model.

[0014] Optionally, in step S1, the distributed potential sensing network includes a microelectrode node array arranged in a row-column pattern, with a node spacing of 0.3-1.2m. The natural potential difference between adjacent nodes is measured using a four-wire constant current excitation method, and the data is uploaded to a cloud platform via wireless communication.

[0015] Optionally, S2 includes: measuring the initial resistivity reference value of each node under dry conditions; measuring the current resistivity value when environmental parameters trigger an early warning threshold; calculating the resistivity change rate, and determining a potential leakage area when the change rate exceeds a preset threshold.

[0016] Optionally, S3 includes: identifying temperature anomaly regions based on infrared thermal imaging data and generating a thermal anomaly confidence map;

[0017] Based on visible light images, leakage traces are identified using an improved YOLOv3 model, and a visual confidence map is generated.

[0018] The potential gradient is calculated based on the potential data and interpolated to generate a potential anomaly intensity map;

[0019] By integrating the thermal anomaly confidence map, visual confidence map, and potential anomaly intensity map, a three-dimensional probability distribution surface of the leakage point is constructed, and the precise coordinates are determined through optimization search.

[0020] Optionally, S4 includes: estimating the leakage rate and the leakage impact range;

[0021] The comprehensive risk value is calculated by combining environmental factor correction coefficients for temperature, humidity, rainfall intensity, and tilt angle.

[0022] Based on the comprehensive risk value, the risk is divided into three levels: low risk, medium risk, and high risk.

[0023] Optionally, in step S5, the probabilistic principal component analysis model is trained based on historical repair data. The inputs are leakage characteristics and environmental parameters, and the outputs include repair strategy parameters such as repair material type, repair radius, material usage, and triggering timing. Decision enhancement is performed based on risk level.

[0024] Optionally, S6 includes:

[0025] For low-risk leaks, apply microcapsule-type self-healing materials;

[0026] For medium-risk leaks, activate the pre-embedded liquid core fiber network repair system;

[0027] For high-risk leaks, a composite repair process is implemented.

[0028] Optionally, in step S6, the verification of the repair effect includes: monitoring the rate of change of resistivity in the repair area, and determining that the initial repair is completed when the continuous sampling value is lower than the threshold and the trend is stable;

[0029] Continuous monitoring is conducted for 24 hours after the repair. If the rate of change in resistivity meets the preset standard and there is no rebound after trend testing, the repair is considered successful.

[0030] Optionally, S7 includes: updating the probabilistic principal component analysis model based on newly added repair cases;

[0031] Threshold for determining the rate of change of resistivity under dynamic calibration;

[0032] Optimize visual inspection algorithm parameters;

[0033] Regularly generate roof health status assessment reports.

[0034] Secondly, a real-time location and adaptive repair triggering system for metal roof leakage is provided to implement the aforementioned real-time location and adaptive repair triggering method for metal roof leakage. The system includes: a multimodal sensing module for real-time acquisition of metal roof status data, including a distributed potential sensing network, a visual monitoring unit, and an environmental parameter acquisition unit.

[0035] The data fusion and positioning module receives data from the multimodal sensing module and performs resistivity time-series feature comparison and precise location of leakage points.

[0036] The risk assessment and decision-making module receives the location results, performs dynamic assessment and classification of leakage risk, and generates repair decisions based on a probabilistic principal component analysis model.

[0037] The repair control and execution module receives the repair decision, controls the repair device to deliver materials, and monitors the repair effect.

[0038] The self-learning and optimization module updates the system model and parameters based on the repair results and historical data, and generates a health status report.

[0039] Compared with the prior art, this application has at least the following beneficial effects:

[0040] This application constructs a distributed multi-source sensing network, integrates potential, infrared and visible light multimodal data, and combines resistivity time-series feature comparison with an improved visual recognition algorithm. This not only significantly improves the detection rate of micro-leakage, but also provides early warning in the early stages of leakage and enhances the positioning accuracy. It overcomes the shortcomings of single detection methods, such as susceptibility to interference and ambiguous positioning, and achieves high-precision, real-time leakage positioning and early warning.

[0041] This application constructs a dynamic comprehensive risk value calculation model by integrating real-time leakage rate, impact range estimation, and environmental factor correction. This model enables the leap from qualitative judgment to quantitative classification of leakage risk, providing a scientific basis for differentiated and precise remediation decisions and avoiding waste or inadequacy of remediation resources.

[0042] Based on historical repair big data, this application uses a probabilistic principal component analysis (PPCA) model to build a repair strategy generation engine. This engine can automatically recommend the optimal repair materials, dosage, scope and timing according to the specific characteristics and environmental conditions of the current leakage, and output confidence intervals. This shifts repair decisions from relying on experience to data-driven optimization, improving the targeting and success rate of repairs, and providing data-driven adaptive intelligent repair decision-making. Attached Figure Description

[0043] Figure 1 A flowchart illustrating a real-time location and adaptive repair triggering method for metal roof leakage, provided in one embodiment of this application;

[0044] Figure 2 This is a schematic diagram of the module connection of a real-time location and adaptive repair triggering system for metal roof leakage provided in one embodiment of this application. Detailed Implementation

[0045] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figure 1-2 The present application will be further described in detail with reference to the embodiments.

[0046] This application provides a method for real-time location and adaptive repair triggering of leaks in metal roofs, including the following steps:

[0047] S1: Real-time roof condition sensing and data acquisition, constructing a distributed multi-source sensing network on the metal roof surface, specifically including: arranging microelectrode sensing node arrays in a row-column pattern along the roof panel ribs and overlap areas. The node substrate is made of 316L stainless steel with a nano-hydrophobic coating, and the exposed area of ​​the electrode contacts is controlled within... The node spacing is dynamically set according to the roof structure characteristics (0.8-1.2m in normal areas, and 0.3-0.5m in areas such as gutters, eaves, and board joints). Each node uses a four-wire constant current excitation method to measure the natural potential difference between adjacent nodes, with an excitation current of 50μA±5%. The sampling frequency is initially set to 1Hz, and automatically increased to 10Hz when environmental parameters trigger the warning threshold. Simultaneously, visual monitoring units are installed at intervals of 3-5m on the roof support purlins. This unit integrates an uncooled infrared focal plane array (resolution 384×288, NETD≤50mK, temperature range -20℃~150℃) and a global shutter CMOS visible light module (2 million pixels, 1 / 2.8” target surface, minimum illuminance 0.01Lux). The optical axes of the two are calibrated to achieve pixel-level registration, and the field of view covers 120°×90°. The environmental parameter acquisition unit includes a PT1000 platinum resistance temperature sensor. The system includes a Class A accuracy sensor, a capacitive humidity sensor (with automatic temperature compensation), a tipping bucket rain gauge (double tipping bucket structure, measurement error ≤ ±2%), a MEMS dual-axis tilt sensor (range ±30°, nonlinearity 0.1%FS), and an ultrasonic anemometer (starting wind speed ≤ 0.15m / s). All sensor data undergoes time-stamp synchronization (IEEE1588v2 protocol), moving average filtering, and outlier removal via an edge computing gateway before being uploaded to the cloud platform via a LoRaWAN (open area) and 4GCat.1 (dense urban area) dual-mode communication link. The power supply system uses a combination of flexible copper indium gallium selenide (CIGS) thin-film photovoltaic cells (5W per cell) and lithium iron phosphate batteries (12V / 10Ah). The node housing meets IP68 protection level and GB / T10125 salt spray test requirements. The installation adopts a magnetic-mechanical dual-fixation structure to avoid secondary damage to the metal roof panel.

[0048] S2: Rapid general survey of leakage areas based on resistivity time-series characteristic comparison. Building upon real-time roof condition sensing and data acquisition, further in-depth analysis and processing of the data acquired by the potential sensing network are performed. First, under dry and stable environmental conditions (preferably after 72 hours of continuous sunny weather), the initial resistivity reference value of each microelectrode node is measured. This value is then stored in a cloud database as the basis for subsequent comparisons. Specifically, a current of 50μA±5% is applied using a four-wire constant current excitation method, and the voltage drop is read using a high-precision ADC chip to calculate the resistance value R between each node. Then, the resistance is calculated according to the formula... Converted to resistivity, where S represents the electrode contact area and L is the distance between adjacent electrodes; for nodes at different locations, differences in geometric parameters lead to... Fluctuations are corrected using the Z-score standardization method to ensure data consistency;

[0049] When the environmental monitoring unit indicates a potential risk of water seepage (such as before or after rainfall or a significant increase in air humidity), the dynamic resistivity change model is activated to update and calculate the current resistivity value in real time. In this process, the influence of environmental factors such as temperature and humidity on resistivity is considered, and a multiple regression correction model is established to eliminate spurious signal interference caused by non-leakage factors. Subsequently, according to The formula calculates the resistivity change rate distribution map for each monitoring point, providing a visual representation of the overall roof condition. A threshold value of 0.15 is set, meaning that when the resistivity change rate at any node... If the threshold is exceeded, the area is considered to have potential leakage risks. Furthermore, to improve detection sensitivity and accuracy, historical data trend analysis and local anomaly pattern recognition algorithms (such as the DBSCAN clustering algorithm) are used to conduct secondary confirmation of suspected leakage sections, ultimately forming preliminary leakage area survey results.

[0050] Furthermore, a segmentation strategy is implemented for large metal roofs (single area ≥ 2000㎡), specifically including the following technical measures: First, based on the roof's as-built BIM model or 3D laser scanning point cloud data (accuracy ≤ 2mm), the center lines of the slab joints, gutter axes, eaves outlines, equipment foundation boundaries, and structural deformation joints are extracted as initial physical boundaries. Combined with the roof slope raster map (resolution 0.5m × 0.5m) generated by the tilt sensor array, areas with a slope gradient change rate > 3° / m are automatically marked as the basis for sub-block division. The area of ​​a single logical monitoring block is limited to 80-150㎡, with quadrilaterals (length-to-width ratio ≤ 2.5:1) being preferred. For irregularly shaped areas, a constrained Delaunay triangulation algorithm is used for adaptive mesh division, and a region merging algorithm ensures that the minimum block area is ≥ 50㎡. Each logical block independently establishes a resistivity reference value library: During a continuous 72-hour dry and stable period (ambient humidity < 60% RH, no precipitation), data from all valid electrode nodes within the block are collected. After outliers are removed using the 3σ criterion, the block reference value is calculated using the weighted median method. (Weight coefficients are assigned according to the Gaussian kernel function:) , The distance from the node to the geometric center of the block. (Take 1 / 3 of the shorter side length of the block). Leakage is determined using intra-block spatial clustering rules: when three or more consecutive adjacent nodes appear within a block... Or a single point within the central sensitive area (a circular region centered at the geometric center with a radius equal to one-third of the shorter side). When a leakage warning is triggered for that block, a 1.0m wide overlapping monitoring strip is set at the boundary of adjacent blocks. The node data within the strip is evaluated by dual-attribution fusion using a Bayesian probability model (posterior probability weights are dynamically allocated according to the distance from the node to the center of the two blocks). The final judgment result is output through an inter-block voting mechanism (if two blocks issue warnings simultaneously, it is confirmed; if a single block issues a warning, it is verified by the visual unit).

[0051] S3: Precisely locates leak points based on multimodal data fusion. For the leak hazard logic block output by S2, a multimodal collaborative precise location process is initiated. First, radiometric calibration and non-uniformity correction are performed on the raw thermal images acquired by the infrared thermal imager: two-point correction is performed using a blackbody source (temperature settings 25.0℃±0.1℃ and 45.0℃±0.1℃), combined with lens distortion parameters (radial distortion coefficient). , Geometric correction is performed to generate a temperature matrix with a spatial resolution of 1.5 mm / pixel. Based on the thermophysical properties of the roofing material (thermal conductivity of aluminum-magnesium-manganese alloy is 160 W / (m·K), specific heat capacity is 880 J / (kg·K)), the local temperature gradient field is calculated. When the temperature of a pixel is continuously lower than the mean of its 5×5 neighborhood by more than 2.5℃ and the temperature drop rate is >0.8℃ / s, it is marked as an effective temperature anomaly point, and a normalized thermal anomaly confidence map is generated. ( For temperature difference, (This represents the maximum temperature difference within the block).

[0052] Secondly, an improved YOLOv3 leak feature detection method was implemented for the synchronously acquired visible light images: the backbone network adopted a customized Darknet-24 structure (the number of residual blocks in stages 3, 4, and 5 of the original Darknet-53 was reduced from 8 / 8 / 4 to 4 / 4 / 2, while retaining the complete structure of stages 1 and 2 to ensure the extraction of shallow texture features), and a 1 / 16 scale detection head (receptive field corresponding to a physical size of 8mm×8mm) was added to the feature pyramid network to specifically enhance the recognition of micro-leakage traces with a width of <3mm; the loss function introduced an illumination robustness constraint term. ( , Let KL divergence be the KL divergence. The model uses a local binary histogram to effectively suppress interference from metallic reflections. After fine-tuning the model using 5000 labeled images (including three types of leakage: water stains, rust crystallization, and seepage from board seams), it outputs the bounding box of leakage traces and category confidence scores. After non-maximum suppression (IoU threshold 0.4), the results are mapped to a unified coordinate system to generate a visual confidence map. .

[0053] Subsequently, the potential sensing data is fused: the natural potential gradient of the microelectrode nodes within the hazardous area is extracted. Surface-constrained inverse distance weighted interpolation (power parameter) is used. interpolation radius Weighting function By combining the local curvature parameters provided by the roof BIM model to correct the interpolation weights, a potential anomaly intensity map is generated. For node areas with a historical leakage frequency of ≥1 time / year, a dynamic enhancement coefficient is introduced. ( (This refers to the number of leakage records in the past 12 months).

[0054] Finally, a three-dimensional probability distribution surface of the leakage points is constructed: The weighting coefficient Dynamically adjust based on real-time environmental parameters (during rainfall:) When it is sunny and dry: (The weights are always 1.0). Optimization search using the golden ratio on curved surfaces: Starting from the geometric center of the potential hazard block, perform one-dimensional searches along the x and y directions respectively (the initial interval is the block boundary, with a shrinkage ratio of 0.618). The iteration terminates when the interval length is <0.3mm or after 3 consecutive iterations. Subpixel-level surfaces were constructed within the convergence point neighborhood (3×3mm) using Catmull-Rom bicubic spline interpolation, and the results were obtained by solving... and Determine the extreme coordinates After coordinate system transformation (based on the homography matrix H obtained by Zhang Zhengyou's calibration method, with a reprojection error of <0.25 pixels), the world coordinates of the leakage point are output. Verified by a third-party testing agency on a standard 10m×10m test roof (with Φ1.0mm artificial leakage holes), the positioning error can be controlled within 2mm under standard testing conditions.

[0055] S4: Dynamic assessment and classification of leakage risk, based on the precise coordinates of leakage points output by S3. By combining multimodal confidence data with real-time environmental parameters and a historical roof leakage database, a dynamic leakage risk assessment model is constructed. The leakage rate index R is estimated using a potential-visual dual-modal fusion method: in the potential mode, the resistivity change rate of microelectrode nodes within a 0.5m radius of the leakage point is extracted. According to the laboratory calibration curve (Unit: mL / min; Calibration conditions: 25℃ deionized water, aluminum-magnesium-manganese alloy plate thickness 0.9mm, tilt angle 5°±0.5°, calibration sample size n=120 groups, R) 2 =0.963) calculate the potential to estimate the leakage amount; in the visual modality, perform U-Net semantic segmentation on the leakage trace area in the visible light image (training set contains 3800 labeled images, Dice coefficient 0.89), combine the pixel area of ​​the low temperature region ΔT≥2.5℃ in the infrared thermal image, and use the area-flow mapping function Computer vision estimation of leakage (A is the projected area of ​​water stains, in mm) 2 v represents the roof water flow velocity, in mm / s; v is based on Manning's formula. The calculation, where h is the water film thickness, is performed using an infrared temperature difference inversion model. Sure, (Local tilt angle); the dual-modal weights are dynamically allocated based on ambient lighting conditions (when illuminance > 5000 lux). , Illuminance <1000 lux or during rainfall , ), final leakage rate ( The default continuous effective detection interval is 300 seconds.

[0056] The leakage influence range S is calculated based on the resistivity spatial diffusion model: a two-dimensional anisotropic diffusion equation is constructed with the leakage point as the source. The diffusion coefficient , (Based on the thermal conductivity of the roofing material of 160 W / (m·K) and surface roughness Ra=3.2 μm); the ADI (Alternating Direction Implicit) finite difference method was used for solution (time step 10 s, spatial step 10 mm), and the boundary condition was set as a third-type convective boundary (heat transfer coefficient h=15 W / (m·K)). 2 ·K)); When the simulation duration equals the current leakage duration t, extract The area of ​​the closed region enclosed by the contour lines is taken as S (unit: m). 2 ).

[0057] Specifically, the parameter calibration of the diffusion equation adopts a three-pronged calibration process of "experimental measurement - numerical inversion - cross-validation", and the specific implementation steps are as follows:

[0058] Step 1: Standard Leakage Diffusion Test. An aluminum-magnesium-manganese alloy test plate (1.0m × 1.0m, 0.9mm thick, fluorocarbon coated surface, roughness Ra = 3.2 ± 0.3μm measured by a profilometer) was selected, consistent with the actual engineering requirements. A 0.8mm leakage hole was pre-drilled in the center of the plate. The test environment was a constant temperature and humidity chamber (GB / T2423.1-2008): temperature 25.0 ± 0.5℃, relative humidity 50 ± 3%, with no wind disturbance. Deionized water with a conductivity < 5μS / cm was injected at a constant flow rate of 1.0mL / min using a micro-injection pump (Harvard Apparatus PhD2000, calibration certificate CAL-2023-WP-087). 25 calibrated microelectrode nodes (calibrated using a Fluke 752A standard resistance box) were arranged on the plate surface in a 50mm × 50mm grid, and the resistivity change rate was collected simultaneously. Time series data (sampling frequency 1Hz, duration 30min), each experiment was repeated 10 times and the average was taken to generate a spatial distribution benchmark dataset of 6 key time nodes t=60s / 120s / 300s / 600s / 90s / 1800s;

[0059] Step 2: COMSOL Multiphysics Coupled Model Construction. A two-dimensional axisymmetric model (with geometric dimensions consistent with the experimental sample) is built on the COMSOL 6.0 platform:

[0060] The “Dilute Substance Transfer” interface is enabled to simulate moisture diffusion, and the “Current” interface is coupled to calculate the resistivity response.

[0061] Establish the relationship between moisture concentration C and resistivity: Where k = 1.85 ± 0.08 (calibrated by a 5cm × 5cm sample titration experiment, R...) 2 =0.987);

[0062] Initial conditions: C=0; Boundary conditions: Leakage holes are set as concentration inflow boundaries (concentration value is converted from flow rate), and the surrounding area is set as zero flux boundaries.

[0063] Parameter to be calibrated: diffusion coefficient , (Initial guess value) , ), constraint range ;

[0064] Step 3: Parameter inversion optimization, defining the objective function:

[0065]

[0066] Calling the built-in Levenberg-Marquardt algorithm in the COMSOL optimization module (convergence tolerance) (Maximum 50 iterations) Iterative optimization based on experimental data. , The root mean square error (RMSE) between the simulated and measured values ​​was kept below 5% until the critical time points (t=300s / 600s / 900s).

[0067] Step 4: Mesh independence and cross-validation, refine the cell size sequentially (5mm → 1mm), and calculate the center point at t=600s. The value is considered convergent when the relative change rate of adjacent grid results is <1% (actual measurement: the change rate is 0.78% for 2mm grid, which meets the requirement).

[0068] Cross-validation: Five independent experimental data sets (flow rates of 0.5 mL / min and 2.0 mL / min) were used for validation. The coefficient of determination R between the predicted and measured values ​​was calculated. 2 >0.92;

[0069] Uncertainty analysis: Monte Carlo simulation (1000 samples) yielded... , The 95% confidence intervals are respectively , .

[0070] Step 5: Parameter solidification and applicability declaration, final calibration of diffusion coefficient: , This parameter is applicable to rolled aluminum-magnesium-manganese alloy plates with a surface roughness Ra of 3.0~3.5μm. (Originating from the x-directional diffusion enhanced by the rib groove, confirmed by SEM surface morphology observation), ambient temperature 20~30℃, leakage fluid conductivity <100μS / cm;

[0071] The environmental factor correction coefficient α is calculated using a piecewise normalization function: (T is the ambient temperature, °C); g(H) = clamp[(H-30) / 60, 0, 1] (H is the relative humidity, %RH); h(P) is mapped by rainfall level (no rain: 0, light rain 0.1~2.5mm / h: 0.3, moderate rain 2.6~8.0mm / h: 0.6, heavy rain 8.1~16.0mm / h: 0.9, torrential rain >16.0mm / h: 1.0); (θ is the local inclination angle of the leakage point, °); Comprehensive correction factor .

[0072] Overall Risk Value ,in (R is in mL / min, normalization interval [0.1, 10]). (S unit m) 2The normalized interval is [0.01, 1.0], ensuring... Based on the V value, a three-tiered risk assessment is implemented: V < 0.3 indicates low risk (slow leakage development, structural safety is assured, and the system generates a repair work order within 72 hours); 0.3 ≤ V < 0.7 indicates medium risk (continuous leakage development, with the risk of localized corrosion or insulation layer wetting, and the system triggers a repair warning within 24 hours and pushes it to the maintenance terminal); V ≥ 0.7 indicates high risk (leakage rate > 5 mL / min or affected area > 0.5 m). 2 If the structural safety or equipment below is threatened, the system will immediately sound an audible and visual alarm and trigger the emergency response process. The risk assessment results are simultaneously written into the roof's electronic health record, serving as the core basis for repair decisions and maintenance plans.

[0073] S5: Self-repair decision optimization based on probabilistic principal component analysis. For the leakage risk level determined in S4, an adaptive repair decision engine driven by probabilistic principal component analysis (PPCA) is constructed. First, a historical repair database is integrated (287 valid cases covering metal roof repair records of 5 large transportation hubs over the past 3 years). Each sample contains an 18-dimensional feature vector: leakage characteristics (leakage point coordinates x / y, rate R, impact range S, risk value V), environmental parameters (temperature T, humidity H, rainfall intensity P, tilt angle θ), and repair strategy (material type coding). Repair radius Material usage Trigger delay ) and repair effect (after 24 hours) Success rate (binary). Data preprocessing: Continuous variables are standardized using Z-score (mean μ, standard deviation). Categorical variables (calculated and fixed based on the training set) One-hot encoding was used; missing values ​​were imputed using the chained multiple interpolation method (MICE, 50 iterations, random forest as the imputation model). Outliers were removed based on the Local Outlier Factor (LOF, k=20, threshold 1.5), resulting in 263 sets of high-quality training samples.

[0074] Then, a PPCA probabilistic generative model is constructed: the 18-dimensional observation vector Modeling as Among them, latent variables (q=3, selected through 5-fold cross-validation: RMSE=0.183 when q=2, RMSE=0.097 when q=3, RMSE=0.095 when q=4, but the model complexity increases significantly). Let μ be the weight matrix and μ be the mean vector. ( The noise is Gaussian. The EM algorithm is used for iterative optimization (initialization: W represents the first 3 principal components of PCA, ...). Convergence condition: log-likelihood change (or 200 iterations), final convergence parameters: The reconstructed RMSE was 0.096. The model was validated using the leave-one-out method: the prediction of the 4-dimensional repair strategy sub-vector significantly improves the accuracy of repair decisions.

[0075] For the current leakage point, its leakage characteristics and environmental parameters (9 dimensions) are standardized according to preprocessing rules and used as part of the observation. The posterior distribution of the repair strategy subspace is calculated based on the PPCA conditional probability formula:

[0076]

[0077] in , The posterior mean is taken and inversely standardized to obtain the basic repair parameters: prediction radius. Dosage ,Delay Material type is encoded based on the maximum posterior probability (threshold: Valid; otherwise marked as "low confidence").

[0078] Enhanced decision-making based on risk level integration: For high risk (V≥0.7), mandatory... (Composite type) Set to 0 (trigger immediately) Multiply by a safety factor of 1.3; for medium risk (0.3 ≤ V < 0.7), retain the PPCA forecast value, but... The upper limit is set to 60 minutes; for low risk (V<0.3), Perform according to the predicted value (≤72h). The repair area boundary is a circular area centered on the leak point, with a radius... To avoid cross-block interference. Material usage correction: , This represents the number of similar repair failures for this block in the past 12 months (0 if there are no failures).

[0079] The final repair instruction with a 95% confidence interval is generated: the confidence interval for continuous parameters is calculated as follows. ( (This is the posterior covariance matrix); the classification parameters output the posterior probability distribution. The instruction is sent after digital signature. Send the data to the repair control module and mark it as either "high confidence" (confidence interval width of all parameters < 15% of the predicted value) or "requires review". Based on the preset data volume threshold or time period, trigger the PPCA model incremental update (retain 70% of historical data + 30% of new data for mixed retraining) to ensure the continuous evolution of the decision model.

[0080] S6: Precise delivery of self-healing materials and verification of repair effects. Based on the repair instructions with confidence intervals generated by S5, a six-degree-of-freedom collaborative robotic arm (model: UR10e, repeatability ±0.05mm) performs precise repair operations. The robotic arm's end effector integrates a three-in-one actuator: the center is a piezoelectric micro-droplet injection valve (nozzle orifice diameter...). 0.3mm, response frequency 1kHz), four material conveying channels are arranged symmetrically around the circumference (inner diameter) 1.0mm, with independent solenoid valve and pressure sensor), with a miniature industrial camera (resolution 1280×1024, working distance 50±5mm) and laser ranging module (accuracy ±0.1mm) embedded at the bottom, providing real-time feedback on the spray position and roof curvature;

[0081] Differentiated repair strategies are implemented for different risk levels, specifically including: (1) Low-risk leakage points (V<0.3): triggering the release of microcapsule-type self-healing materials. The microcapsules are made of urea-formaldehyde resin wall material (thickness 1.5±0.3μm) encapsulating sodium polyacrylate / polyvinyl alcohol composite water-absorbing resin core material (mass ratio 7:3), with a particle size distribution of 50-120μm (D50=85μm), prepared by in-situ polymerization (formaldehyde / urea molar ratio 1.2:1, pH=4.0-4.5, reaction at 60℃ for 4h); the material is prepared into a suspension with a solid content of 15% using ethanol as the dispersion medium, and injected into the storage tank after pre-filtration through a 0.22μm filter membrane; the injection parameters are: pressure 0.15MPa, pulse width 20ms, and single-point injection volume according to the formula. Calculate (r is the repair radius, (The theoretical film thickness is used). The spray trajectory adopts an Archimedean spiral (pitch 0.5 mm) to ensure coverage of the area around the leakage point within 1.2 times the predicted radius. Under ideal experimental conditions, it starts to expand within 30 seconds after contact with water. The measured expansion ratio in deionized water at 25℃ is 9.3±0.7 times (GB / T19631-2005), and the compressive strength of the expanded body is ≥0.8MPa.

[0082] (2) Medium-risk leakage points (0.3≤V<0.7): Activate the pre-embedded liquid core fiber network repair system. The network is made of hollow glass fiber (E-glass, inner diameter 200±20μm, wall thickness 30±5μm, tensile strength ≥1800MPa) woven into a 2D mesh structure according to the direction of the roof panel ribs (weaving density 500 fibers / m). 2The fiber cavity is pre-filled with a two-component epoxy repair agent (component A: epoxy resin E-51 / reactive diluent AGE = 95:5; component B: polyamide curing agent T31 / nano SiO2 = 90:10, nanoparticles modified with silane coupling agent KH-550); after the repair command is triggered, the probe at the end of the robotic arm precisely punctures the fiber adjacent to the leakage point (puncture depth 0.8mm±0.1mm, force control threshold 2.5N), and the repair agent flows autonomously along the crack to fill it under capillary action; simultaneously, the infrared heating module (wavelength 3.0-5.0μm, power 50W / cm²) is activated. 2 The repaired area is locally heated to 45±3℃ to accelerate curing (surface drying time at 25℃ ≤15min, 7d tensile shear strength ≥12MPa, GB / T7124-2008).

[0083] (3) High-risk leakage points (V≥0.7): Implement composite repair process. First stage: Spray fast-curing polyurethane sealant (isocyanate index NCO%=4.5±0.3%, thixotropic fumed silica addition 3wt%), spray pressure 0.35MPa, forming a conical sealant with a diameter 1.5 times the predicted radius (cone angle 60°), gel time ≤25s (ASTMD2471); Second stage: After an interval of 90±10s, spray a fluorocarbon modified acrylate long-lasting coating (solid content 62±3%, fluorine content 18wt%, adhesion grade 0, GB / T9286-2021), spray air pressure 0.25MPa, film thickness 180±20μm (calibrated by magnetic thickness gauge); between the two stages, the integrity of the sealant formation (edge ​​continuity >95%) is confirmed by the robotic arm vision module, and automatic re-spraying is performed if abnormality occurs;

[0084] Closed-loop monitoring is implemented during the repair process: a potential sensing network collects resistivity data of the repair area at a frequency of 10Hz, and when 5 consecutive samples meet the requirements... When the sensitivity is increased by 0.15 compared to the S2 threshold and the trend is stable (absolute slope < 0.001 / s), the initial repair is considered complete, and the robotic arm is withdrawn; the repair effect verification process is started simultaneously: monitoring is performed every 10 minutes within 0-2 hours after repair, and every hour within 2-24 hours, generating a resistivity time-series decay curve; if any monitoring point If the duration exceeds 8 minutes, the system will automatically trigger a secondary repair (material usage increased by 20%, repair radius expanded by 1.3 times) and generate an anomaly analysis report, which will be pushed to the maintenance terminal. The criteria for successful repair are: the end of 24 hours. And there is no rebound trend (according to the Mann-Kendall trend test, Z value < -1.645).

[0085] Furthermore, the resistivity change trend verification in the repair effect validation adopts the Mann-Kendall (MK) nonparametric one-sided trend test method. The specific implementation process is as follows: taking the repair completion time (t=0) as the starting point, extract the resistivity change trend data collected at a preset frequency within 24 hours after the repair. Effective monitoring sequence (n≥30, according to the S6 monitoring strategy: 1 time every 10 min from 0-2h, 1 time every hour from 2-24h; terminate this verification and mark the anomaly when the data missing rate is >10%). Sequence preprocessing includes: (1) removing outliers using Grubbs test (α=0.01); (2) sorting the remaining data in ascending order of time; (3) calculating the number of result groups g and the frequency of each group. (When the percentage of identical values ​​is greater than 5%, result correction is enabled). Calculate the MK statistic S:

[0086]

[0087] Calculate the corrected variance:

[0088]

[0089] When n > 10, S approximately follows a normal distribution. Calculate the standardized Z-value:

[0090]

[0091] Set one-sided test hypothesis: null hypothesis (There is an upward trend or no significant downward trend), alternative hypothesis (A significant downward trend exists); taking a significance level of α = 0.05, the critical value... Judgment rule: If Z < -1.645, reject. The resistivity change rate was confirmed to show a significant decreasing trend, meeting the "no rebound trend" requirement; if Z ≥ -1.645 (including no significant trend or an increasing trend), a rebound risk was identified, and the repair effect verification failed. This result is consistent with the result at the end of 24 hours. The threshold determination together constitutes a dual standard for successful repair (both must be met simultaneously).

[0092] S7: System self-learning and parameter optimization, based on the final judgment result of the S6 repair effect verification (success / failure / required review), the system automatically triggers the closed-loop self-learning process. First, construct a structured repair archive: each repair event generates a unique archive number (e.g., format: ROOF-YYYYMMDD-SEQ, where SEQ is the serial number of the day), and the archive content is stored in five structured categories: (1) Input parameter set (leakage point desensitization coordinates, risk value V, environmental parameters T / H / P / θ, PPCA prediction parameters and 95% confidence interval); (2) Execution parameter set (actual material type code, usage, trigger delay, key point coordinates of robotic arm trajectory); (3) Process monitoring set ( Time series data (10Hz sampling), infrared temperature sequence, visual verification image hash value); (4) Effect verification set (24h monitoring) (5) Manual review set (on-site confirmation photos, text notes, and digital signatures of reviewers uploaded by the operation and maintenance terminal). All data are encrypted by the SM4 national cryptographic algorithm and written into the time series database (InfluxDB 2.6 Enterprise), and an index associated with the block ID of the roof BIM model is established. The data retention strategy complies with the "Construction Engineering Document Archiving Specification" GB / T 50328-2014 (original monitoring data is retained for 24 months, and structured archives are permanently archived).

[0093] PPCA model incremental optimization is performed: When the cumulative number of newly added effective repair cases reaches 20 or the interval since the last update is ≥30 days (whichever comes first), the system automatically starts model retraining. A "sliding time window + time-weighted" strategy is adopted: historical cases from the most recent 18 months are retained (weight coefficient 1.0), and newly added cases are overlaid (weight coefficient 1.2) to construct a mixed training set; the PPCA modeling process described in S5 is invoked (latent variable dimension q=3, EM algorithm iterated to convergence), and the weight matrix W and noise variance are re-estimated. The updated model underwent leave-one-out cross-validation. If the key indicators (material type weighted F1-score, continuous parameter prediction R-value) are... 2 If the performance improvement is ≥1% or equal to the current production model, the production model will be replaced after digital signature (SM2 algorithm), and a version log (including version number, training set size, validation metrics, and hash verification value) will be generated. If the metrics decrease by ≥2%, a manual review alert will be triggered, and the model will be rolled back to the previous stable version. The entire model update process is executed in an isolated computing sandbox, and the update package is distributed to each edge node after being verified by the national cryptographic SM3 hash.

[0094] Resistivity change rate threshold dynamic calibration: The system performs adaptive threshold correction daily at 02:00 (off-peak business period). The calibration is based on three factors: (1) material aging coefficient. (Cumulative UV radiation is calculated by integrating historical data from the NASAPOWER meteorological database, in MJ / m³) 2 ; (1) Number of repairs in the block in the past 36 months); (2) Seasonal correction factor γ (based on the solar term table of GB / T 19167-2003: Spring Equinox-Summer Solstice γ=1.05, Summer Solstice-Autumn Solstice γ=0.95, Autumn Equinox-Winter Solstice γ=1.00, Winter Solstice-Spring Solstice γ=1.10); (3) Environmental baseline drift: Calculate the number of leak-free blocks in the current 7 days. mean with standard deviation ,when or At that time, baseline correction is initiated. Final threshold calculation: The calibration result is limited to the range of [0.10, 0.25]; the calibration result is written to the system configuration table and synchronized to all edge computing gateways via the MQTT protocol (response latency <500ms).

[0095] The improved YOLOv3 detection algorithm continues to evolve: the system automatically performs model fine-tuning every Sunday at 03:00. Data sources include: (1) images of leak areas that have been desensitized during the repair process (new samples); (2) new leak patterns marked by maintenance personnel (such as cracking of fluorocarbon coating and electrochemical corrosion around rivet holes); (3) synthetic samples generated by StyleGAN2-ADA (accounting for ≤15%, reviewed and entered into the database by experts). The annotation is performed by two independent annotations + cross-validation (IoU≥0.85), which complies with the "Building Leakage Image Annotation Specification V1.2" (Q / ROOF-TECH 001-2023). Fine-tuning employs transfer learning: using the current production model weights as initial values, the model is trained for 10 epochs on the new dataset with a learning rate of 1e-4 and a batch size of 16, using cosine annealing scheduling. After fine-tuning, the model is tested on the retained validation set (containing 20% ​​of the new data). If mailto:mAP@0.5 improves by ≥0.5% or the recall rate of novel leakage categories improves by ≥3%, the production model is replaced and the version number is updated (format: YOLOv3-ROOF-Vx.yz). The fine-tuning process is executed on an NVIDIA A10 GPU server, and training logs are archived on the MLflow tracking server.

[0096] Finally, a roof health status assessment report is generated: The system automatically generates the previous month's assessment report (PDF / A-3 format, conforming to the long-term preservation requirements of electronic files in GB / T 33482-2016) at 08:00 on the 1st of each month. The report includes: (1) global statistics (total number of leakage events, risk distribution, repair success rate, average response time); (2) block health index ( β represents the number of leaks per month, and β is the aging coefficient. The success rate of repair is calculated by the index-weighted moving average, λ=0.3); (3) Spatiotemporal trend analysis (heat map of leakage hotspot evolution, monthly change curve of key parameters); (4) Maintenance decision support (list of blocks with health index <60, priority ranking of preventive maintenance, early warning of material inventory). The report is pushed to the operation and maintenance platform after being encrypted with digital watermark, and blocks with H<70 are marked as "key focus", and the monitoring frequency is automatically increased to 5Hz in the next census. All self-learning operation logs are synchronously stored in Chang'an Chain, and the hash value is connected to the judicial blockchain platform.

[0097] In one embodiment, a real-time location and adaptive repair triggering system for metal roof leaks is provided, the system comprising:

[0098] The multimodal sensing module includes a distributed potential sensing network, a visual monitoring unit, and an environmental parameter acquisition unit. The distributed potential sensing network consists of a microelectrode array arranged at key locations on the roof, with an electrode spacing of 0.5-1.0m, encapsulated in corrosion-resistant materials, used for real-time monitoring of changes in the roof's natural potential. The visual monitoring unit includes an infrared thermal imager and a visible light camera, installed on the roof support structure, with a coverage angle greater than 120°, used to capture temperature anomalies and visual leakage characteristics. The environmental parameter acquisition unit includes a temperature sensor, a humidity sensor, a rain gauge, a tilt sensor, and an anemometer, used to provide environmental data support. Each unit is connected to the central processing module via a wireless communication module, with a data sampling frequency of 1-10Hz.

[0099] The data fusion and localization module receives data collected by the multimodal sensing module and performs resistivity time-series feature comparison and precise leakage point location. This module includes a signal preprocessing unit, a resistivity calculation unit, an improved Yolov3 detection unit, and a golden section localization unit. The signal preprocessing unit filters and denoises the original signal. The resistivity calculation unit performs resistivity change rate calculation for S2. The improved Yolov3 detection unit identifies leakage features from the visual image; its backbone network adopts a lightweight design, containing only 1×, 2×, 2×, 2×, 2× residual block structures to reduce computational load. The golden section localization unit combines potential gradient data and visual recognition results to determine the precise location of the leakage point using the golden section optimization search method, outputting the location coordinates and confidence interval.

[0100] The risk assessment and decision-making module receives the leak location results and performs dynamic leak risk assessment and repair decision optimization. This module includes a risk assessment unit, a PPCA optimization unit, and a decision generation unit. The risk assessment unit calculates the comprehensive risk value V based on S4 and classifies it into levels. The PPCA optimization unit constructs a probabilistic principal component analysis model based on historical repair data to generate the optimal repair strategy for the current leak. The decision generation unit integrates the risk level and the PPCA optimization results to generate repair parameters with confidence intervals, including the repair area boundary, material usage, and trigger timing, achieving adaptive repair decision-making.

[0101] The repair control and execution module receives instructions from the decision module and controls the precise delivery of repair materials. This module includes a repair material storage unit, a precise delivery control unit, and a repair effect monitoring unit. The repair material storage unit contains three types of repair materials: microcapsule-type self-healing material, liquid-core fiber network repair agent, and composite repair material, each corresponding to different risk levels. The precise delivery control unit uses a multi-axis positioning robotic arm with a positioning accuracy of ±1mm, controlling material delivery based on the leak point coordinates and boundaries. The repair effect monitoring unit assesses the repair effect by real-time monitoring of resistivity changes in the repair area, providing a completion signal when a preset standard is met and triggering a secondary repair if the standard is not met.

[0102] The self-learning and optimization module is responsible for updating system parameters and optimizing performance. This module includes a data storage unit, a model update unit, and a parameter adjustment unit. The data storage unit records data on each leakage event, repair process, and effect verification. The model update unit periodically updates the PPCA model using probabilistic principal component analysis to optimize repair decisions. The parameter adjustment unit dynamically adjusts the resistivity change rate threshold, YOLOv3 detection algorithm parameters, and repair material ratios to improve system adaptability and accuracy. This module also periodically generates roof health status assessment reports to provide data support for preventative maintenance.

[0103] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for real-time positioning and adaptive repair triggering of metal roof leakage, characterized in that, The method includes: S1: Real-time status data is collected through a multimodal sensing network deployed on the metal roof. The multimodal sensing network includes a distributed potential sensing network, a visual monitoring unit, and an environmental parameter acquisition unit. S2: Based on the data collected by the potential sensing network, resistivity time-series characteristics are compared to identify potential leakage areas; S3: For the potential leakage area, perform multimodal data fusion analysis by fusing potential data, infrared thermal image data and visible light images to accurately locate the leakage point; S4: Based on the location of the leak point, combined with environmental parameters and historical leak data, conduct dynamic assessment and classification of leak risk; S5: Based on the risk level, generate an adaptive remediation decision using a probabilistic principal component analysis model and determine the remediation strategy parameters; S6: Based on the repair strategy parameters, control the repair execution device to accurately deliver the repair material and monitor the repair effect in real time; S7: Based on the verification results of the repair effect and historical data, perform self-learning and optimization of system parameters, and update the risk assessment model and repair decision model.

2. The method of claim 1, wherein the method further comprises: In step S1, the distributed potential sensing network includes a microelectrode node array arranged in a row-column pattern with a node spacing of 0.3-1.2m. It uses a four-wire constant current excitation method to measure the natural potential difference between adjacent nodes and uploads the data to a cloud platform via wireless communication.

3. The method of claim 1, wherein the method further comprises: S2 includes: measuring the initial resistivity reference value of each node under dry conditions; measuring the current resistivity value when environmental parameters trigger an early warning threshold; calculating the resistivity change rate, and determining a potential leakage area when the change rate exceeds a preset threshold.

4. The method of claim 1, wherein, S3 includes: identifying temperature anomaly regions based on infrared thermal imaging data and generating a thermal anomaly confidence map; Based on visible light images, leakage traces are identified using an improved YOLOv3 model, and a visual confidence map is generated. The potential gradient is calculated based on the potential data and interpolated to generate a potential anomaly intensity map; By integrating the thermal anomaly confidence map, visual confidence map, and potential anomaly intensity map, a three-dimensional probability distribution surface of the leakage point is constructed, and the precise coordinates are determined through optimization search.

5. The method for real-time location and adaptive repair triggering of leaks in metal roofs according to claim 1, characterized in that, S4 includes: estimating the leakage rate and the extent of leakage impact; The comprehensive risk value is calculated by combining environmental factor correction coefficients for temperature, humidity, rainfall intensity, and tilt angle. Based on the comprehensive risk value, the risk is divided into three levels: low risk, medium risk, and high risk.

6. The method for real-time location and adaptive repair triggering of leaks in metal roofs according to claim 1, characterized in that, In S5, the probabilistic principal component analysis model is trained based on historical repair data. The input is leakage characteristics and environmental parameters, and the output includes repair strategy parameters such as repair material type, repair radius, material usage, and triggering timing. Decision enhancement is performed based on risk level.

7. The method for real-time location and adaptive repair triggering of leaks in metal roofs according to claim 1, characterized in that, S6 includes: For low-risk leaks, apply microcapsule-type self-healing materials; For medium-risk leaks, activate the pre-embedded liquid core fiber network repair system; For high-risk leaks, a composite repair process is implemented.

8. The method for real-time location and adaptive repair triggering of leaks in metal roofs according to claim 1, characterized in that, In step S6, the verification of the repair effect includes: monitoring the rate of change of resistivity in the repair area, and determining that the preliminary repair is completed when the continuous sampling value is lower than the threshold and the trend is stable; Continuous monitoring is conducted for 24 hours after the repair. If the rate of change in resistivity meets the preset standard and there is no rebound after trend testing, the repair is considered successful.

9. The method for real-time location and adaptive repair triggering of leaks in metal roofs according to claim 1, characterized in that, S7 includes: updating the probabilistic principal component analysis model based on newly added repair cases; Threshold for determining the rate of change of resistivity under dynamic calibration; Optimize visual inspection algorithm parameters; Regularly generate roof health status assessment reports.

10. A real-time location and adaptive repair triggering system for leaks in metal roofs, characterized in that, The system is used to implement the real-time location and adaptive repair triggering method for metal roof leakage as described in any one of claims 1-9. The system includes: a multimodal sensing module for real-time acquisition of metal roof status data, including a distributed potential sensing network, a visual monitoring unit, and an environmental parameter acquisition unit. The data fusion and positioning module receives data from the multimodal sensing module and performs resistivity time-series feature comparison and precise location of leakage points. The risk assessment and decision-making module receives the location results, performs dynamic assessment and classification of leakage risk, and generates repair decisions based on a probabilistic principal component analysis model. The repair control and execution module receives the repair decision, controls the repair device to deliver materials, and monitors the repair effect. The self-learning and optimization module updates the system model and parameters based on the repair results and historical data, and generates a health status report.

Images