Health monitoring method for curtain wall aluminum supporting structure based on three-dimensional surveying of unmanned aerial vehicle

Abstract

The application discloses a curtain wall aluminum supporting structure health monitoring method based on unmanned aerial vehicle three-dimensional surveying and mapping, relates to the technical field of intelligent monitoring, and comprises the following steps: acquiring building facade point cloud data through an unmanned aerial vehicle laser radar, performing three-dimensional coordinate parameterization processing on the point cloud by adopting a curved surface parameterization expression technology, setting constraint conditions according to the geometric characteristics of the building facade, and establishing an initial parameterized curved surface model containing boundary condition restrictions; performing geometric deviation calculation on the initial parameterized curved surface model and a design reference surface, obtaining the spatial deviation amount of each point by adopting a residual error calculation mode, assigning different weights to the deviation points in different regions by a weight coefficient distribution technology, and constructing a target function expression based on weighted residual error; the curtain wall aluminum supporting structure health monitoring method based on unmanned aerial vehicle three-dimensional surveying and mapping significantly improves the intelligent level and accuracy of curtain wall installation acceptance and long-term health monitoring.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology, specifically to a method for health monitoring of aluminum support structures for curtain walls based on UAV 3D mapping. Background Technology

[0002] As a crucial external envelope structure for modern high-rise buildings, the installation quality of the aluminum support structure of the curtain wall directly affects the structural safety and service life of the building. Traditional curtain wall acceptance mainly relies on manual measurement and visual inspection, which has significant shortcomings when dealing with complex three-dimensional spatial structures. Manual measurement often only obtains data from local key points, making it difficult to fully reflect the true state of the overall structure, while visual inspection is further limited by the experience and field of vision of the inspectors.

[0003] The core technical challenge in current curtain wall acceptance lies in the contradictory relationship between quality control of surveying data and benchmark establishment. The completeness, accuracy, and registration errors of point cloud data during surveying directly affect the accuracy of acceptance conclusions, but existing technologies lack a systematic quality control process to ensure these key parameters meet acceptance standards. More importantly, even with high-quality surveying data, transforming this data into reliable health record benchmarks to provide effective comparative references for subsequent long-term monitoring remains a pressing technical challenge. This technical dilemma is particularly pronounced in practical engineering projects. For example, during the acceptance of a curtain wall project, minor deviations were found in localized areas. However, due to the lack of complete 3D data support and standardized quality control processes, it was impossible to accurately determine whether these deviations were within acceptable limits, nor could reliable initial benchmark data be established for subsequent structural health monitoring. Summary of the Invention

[0004] The purpose of this invention is to provide a health monitoring method for aluminum support structures of curtain walls based on UAV three-dimensional mapping, thereby solving the problems existing in the prior art.

[0005] To achieve the above objectives, this invention provides the following technical solution: a health monitoring method for aluminum support structures of curtain walls based on UAV 3D mapping, comprising: acquiring point cloud data of the building facade using UAV lidar; performing 3D coordinate parameterization processing on the point cloud using surface parameterization technology; setting constraint conditions according to the geometric features of the building facade to establish an initial parameterized surface model including boundary condition constraints; calculating the geometric deviation between the initial parameterized surface model and the design reference plane; obtaining the spatial deviation of each point using residual calculation; assigning differentiated weights to deviation points in different regions using weight coefficient allocation technology; constructing an expression for a target function based on weighted residuals; if the gradient value of the target function exceeds the preset fitting accuracy control range, initiating the function gradient calculation program; determining the optimization termination condition using convergence threshold setting technology; controlling the parameter update amplitude each time through an iteration step size adjustment mechanism to establish an adaptive iterative optimization solution system; performing the first round of parameter update calculation according to the adaptive iterative optimization solution system; tracking the changing trend of the target function value in real time using error change monitoring technology; judging the current iteration efficiency through a convergence speed evaluation algorithm; and recording historical iteration data.

[0006] Preferably, the step of acquiring point cloud data of the building facade using UAV LiDAR, performing three-dimensional coordinate parameterization processing on the point cloud using surface parameterization technology, setting constraints based on the geometric features of the building facade, and establishing an initial parameterized surface model including boundary condition restrictions includes: scanning the building facade using UAV LiDAR, acquiring point cloud data from the scan results to obtain a point cloud dataset; performing three-dimensional coordinate parameterization processing on the point cloud dataset using surface parameterization technology, determining parameterized coordinate values ​​by mapping point cloud points to a parameter domain; setting constraints on the parameterized coordinate values ​​based on the geometric features of the building facade, obtaining a constraint parameter group by integrating feature line segments; establishing an initial parameterized surface model including boundary condition restrictions using the constraint parameter group, and judging the model boundary integrity; if the model boundary integrity is lower than a preset threshold, performing boundary interpolation adjustment on the initial parameterized surface model to obtain a boundary-complete surface model.

[0007] Preferably, the step of calculating the geometric deviation between the initial parametric surface model and the design datum, obtaining the spatial deviation of each point using residual calculation, and assigning differentiated weights to deviation points in different regions using weighted coefficient allocation technology to construct the objective function expression based on weighted residuals includes: obtaining the spatial deviation of each point by calculating the Euclidean distance through the difference in point coordinates using the residual calculation method based on the initial parametric surface model and the design datum, thus obtaining a set of deviations; assigning weighted coefficients to different regions using the set of deviations, setting proportional coefficients according to the curvature characteristics of the regions, and determining a set of differentiated weight values; integrating the weighted residual terms from the set of differentiated weight values, multiplying them by the corresponding deviations, and summing them to construct the objective function expression based on weighted residuals; introducing boundary constraints for the objective function expression, setting parameter boundary ranges, and obtaining an optimized set of function parameters; and correcting the deviation of the building facade surface using the set of function parameters, adjusting the point coordinate values, and obtaining a corrected surface representation.

[0008] Preferably, if the gradient value of the objective function exceeds the preset fitting accuracy control range, the function gradient calculation program is initiated. A convergence threshold setting technique is used to determine the optimization termination condition. An iteration step size adjustment mechanism controls the parameter update amplitude for each iteration. The adaptive iterative optimization solution system is established by: obtaining the current gradient value from the gradient threshold check; comparing the accuracy control range; activating function gradient calculation if the threshold is exceeded; calculating the gradient vector by differentiating the objective function and combining it with the current parameter value to obtain the updated gradient direction; using the convergence threshold setting technique, comparing the gradient vector magnitude with the preset threshold to determine the optimization termination condition and obtain the iteration step size adjustment value; controlling the parameter update amplitude using the iteration step size adjustment value; establishing an adaptive iterative optimization solution system by multiplying by the gradient direction and applying a decay factor to obtain the optimized parameter set; and performing deviation correction on the building facade surface from the optimized parameter set by adjusting the surface parameter values ​​and recalculating the point coordinates to obtain the corrected surface representation.

[0009] Preferably, the first round of parameter update calculation is performed according to the adaptive iterative optimization solution system. Error change monitoring technology is used to track the changing trend of the objective function value in real time. The current iteration efficiency is judged by the convergence speed evaluation algorithm. Historical iteration data is recorded, including the first round of parameter update calculation results obtained from the adaptive iterative optimization solution system. Error change monitoring technology is used to calculate the difference sequence by continuously sampling the objective function value sequence. The difference sequence is then fitted with a trend to obtain the changing trend. Based on the changing trend, the current iteration step size is compared with a preset threshold using the gradient descent algorithm. The gradient descent algorithm takes the changing trend and step size as inputs and outputs an efficiency index to determine the iteration efficiency.

[0010] Preferably, the step of performing the first round of parameter update calculation according to the adaptive iterative optimization solution system, using error change monitoring technology to track the changing trend of the objective function value in real time, judging the current iteration efficiency through a convergence speed evaluation algorithm, and recording iteration history data also includes extracting historical data records from the iteration efficiency, using a data storage mechanism to write the data into a preset buffer through a serialization method to obtain a record set; for the record set, performing subsequent analysis and processing, calculating the deviation vector by comparing the historical data with the current parameters to obtain the optimization adjustment value; applying the optimization adjustment value to the correction of the building facade surface parameters, using deviation correction technology to adjust the coordinates of the surface points by weighted average to obtain the correction representation.

[0011] Preferably, the method further includes performing numerical stability checks based on iterative historical data, using local optimum judgment techniques to identify whether the system has fallen into a local extreme state, evaluating whether the current optimization state meets the convergence requirements through iteration termination conditions, and outputting the optimal parameter combination if the convergence threshold is met. Specifically, this includes obtaining a numerical sequence from the iterative historical data, using a difference calculation method to obtain a difference sequence by subtracting adjacent values, averaging the absolute values ​​of the difference sequences to obtain a stability index, judging whether the preset threshold is met through the stability index, and determining the numerical stability state; for the numerical stability state, using a gradient analysis method to identify local extreme features by calculating the slope of the numerical sequence, extracting a deviation vector from the local extreme features, and summing the deviation vectors to obtain the extreme state identification result.

[0012] Preferably, the step of performing numerical stability checks based on iterative historical data, using local optimum judgment techniques to identify whether the system has fallen into a local extreme state, evaluating whether the current optimization state meets the convergence requirements through iteration termination conditions, and outputting the optimal parameter combination if the convergence threshold is met, further includes obtaining iteration termination condition parameters based on the extreme state identification results, and determining whether the system has fallen into a local extreme state by comparing the current optimization state with the iteration termination condition parameters; if the determination shows that the system has not fallen into a local extreme state, the optimal parameter combination is output from the optimization state evaluation.

[0013] Preferably, the process also includes reconstructing the optimized facade surface model based on the optimal parameter combination, converting the optimized parameters into corrected three-dimensional coordinates using point cloud coordinate reprojection technology, calculating the accuracy improvement before and after optimization through geometric deviation comparison analysis, and generating a building facade geometric accuracy optimization report. Specifically, this includes obtaining the coordinate transformation matrix from the optimal parameter combination, using the point cloud reprojection method to convert the coordinates of each point in the original point cloud data into corrected three-dimensional coordinates through matrix multiplication to obtain the optimized coordinate set; constructing the facade surface model for the optimized coordinate set, dividing the coordinate points into triangular meshes using mesh subdivision technology and interpolating and fusing them to generate a continuous surface, and determining the surface geometric structure; obtaining the corresponding point pairs between the surface geometric structure and the original facade model, comparing the deviations between point pairs using the Euclidean distance calculation method, and obtaining a geometric deviation sequence.

[0014] Preferably, the step of reconstructing the optimized facade surface model based on the optimal parameter combination, using point cloud coordinate reprojection technology to convert the optimized parameters into corrected three-dimensional coordinates, calculating the accuracy improvement before and after optimization through geometric deviation comparison analysis, and generating a building facade geometric accuracy optimization report also includes extracting the average deviation value and standard deviation from the geometric deviation sequence, obtaining the average deviation value by summing the sequence elements and dividing by the number of elements, obtaining the standard deviation by summing the deviation squared differences, dividing by the number of elements, and then taking the square root, using subtraction to evaluate the accuracy difference before and after optimization, and determining the accuracy improvement; integrating the deviation sequence and surface structure data based on the accuracy improvement to generate a building facade geometric accuracy optimization report, including a deviation distribution map and improvement indicators.

[0015] As can be seen from the above technical solution, the present invention has the following beneficial effects:

[0016] This method for health monitoring of aluminum-supported curtain wall structures based on UAV 3D mapping automates the entire process from mapping and analysis to health assessment by acquiring point cloud data of the curtain wall facade using UAV LiDAR and combining it with parametric surface modeling and adaptive optimization algorithms. This method comprehensively collects 3D data of complex facade spaces, significantly improving data integrity and accuracy. Through geometric deviation calculation and multi-region differentiated weight assignment, it accurately identifies local structural deformations. The introduction of function gradient convergence control and error adaptive iteration mechanism effectively reduces mapping errors and improves the stability of optimization solutions. Furthermore, by combining historical data comparison and convergence analysis, it can dynamically track changes in structural status and establish reliable health record benchmarks. Compared with traditional manual measurement and visual inspection methods, this invention significantly improves the intelligence and accuracy of curtain wall installation acceptance and long-term health monitoring. Attached Figure Description

[0017] Figure 1 This is a flowchart of the health monitoring method for aluminum support structures of curtain walls based on UAV three-dimensional mapping according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, this invention provides a technical solution: a health monitoring method for aluminum support structures of curtain walls based on UAV 3D mapping, including acquiring point cloud data of the building facade using UAV lidar, performing 3D coordinate parameterization processing on the point cloud using surface parameterization technology, setting constraint conditions according to the geometric features of the building facade, and establishing an initial parameterized surface model including boundary condition constraints; calculating the geometric deviation between the initial parameterized surface model and the design reference plane, obtaining the spatial deviation of each point using residual calculation, assigning differentiated weights to deviation points in different regions using weight coefficient allocation technology, and constructing an expression for a target function based on weighted residuals; if the gradient value of the target function exceeds the preset fitting accuracy control range, initiating the function gradient calculation program, determining the optimization termination condition using convergence threshold setting technology, and adjusting the iteration step size. The mechanism controls the magnitude of each parameter update and establishes an adaptive iterative optimization solution system. The first round of parameter update calculations is performed based on this system, using error change monitoring technology to track the changing trend of the objective function value in real time. A convergence speed evaluation algorithm is used to determine the efficiency of the current iteration, and historical iteration data is recorded for subsequent analysis. Numerical stability checks are performed based on the historical iteration data, using local optimum judgment technology to identify whether the system has fallen into a local extreme state. The current optimization state is evaluated to determine if it meets the convergence requirements using iteration termination conditions. If the convergence threshold is met, the optimal parameter combination is output. The optimized facade surface model is reconstructed based on the optimal parameter combination. Point cloud coordinate reprojection technology is used to convert the optimized parameters into corrected 3D coordinates. Geometric deviation comparison analysis is used to calculate the accuracy improvement before and after optimization, generating a building facade geometric accuracy optimization report.

[0020] This method utilizes an unmanned aerial vehicle (UAV) lidar system to acquire point cloud data of the curtain wall facade. It then transforms this point cloud information into a mathematically processable 3D coordinate parameter model using surface parametric representation technology. Constraints are set based on the geometric features of the building facade, forming an initial surface model with boundary restrictions. Subsequently, the initial surface is compared with the design reference plane, and spatial deviations are extracted using a residual algorithm. Differential weights are applied to different regions to construct a weighted objective function. During the optimization phase, when the gradient of the objective function exceeds the control range, an adaptive gradient optimization program is initiated, achieving automatic model convergence by dynamically adjusting the iteration step size and convergence threshold. This system combines error change monitoring and convergence speed evaluation to achieve real-time feedback and stability control of the iteration process, avoiding getting trapped in local optima. The final optimal parameter combination is used to reconstruct the optimized facade surface model, and the corrected coordinates are visualized using point cloud reprojection technology, resulting in a report on the improvement of curtain wall geometric accuracy.

[0021] This method achieves intelligent integration of UAV 3D mapping and curtain wall structural health monitoring. Through parametric modeling and adaptive iterative optimization, it effectively improves the accuracy and robustness of curtain wall facade geometry measurements. Its weighted residual calculation method based on weight allocation significantly reduces outlier interference and improves overall fitting quality. Adaptive step size control and convergence threshold judgment mechanisms ensure the stability and efficiency of the iterative process, avoiding overfitting or local optima traps. The final generated geometry accuracy optimization report provides highly reliable data support for curtain wall structural deformation analysis, health status assessment, and maintenance decisions, thereby improving the efficiency and safety of building structure monitoring.

[0022] Point cloud data of building facades is acquired using UAV LiDAR. 3D coordinate parameterization is performed on the point cloud using surface parametric representation technology. Constraints are set based on the geometric features of the building facade to establish an initial parametric surface model with boundary conditions. This process involves scanning the building facade with UAV LiDAR to obtain point cloud data, resulting in a point cloud dataset. 3D coordinate parameterization is performed on the point cloud dataset using surface parametric representation technology, mapping the point cloud points to a parameter domain to determine parametric coordinate values. Constraints are set on the parametric coordinate values ​​based on the geometric features of the building facade, and a set of constraint parameters is obtained by integrating feature line segments. An initial parametric surface model with boundary conditions is established using the constraint parameter set, and the model's boundary integrity is assessed. If the model's boundary integrity is below a preset threshold, boundary interpolation is performed on the initial parametric surface model to obtain a boundary-complete surface model.

[0023] In one possible implementation, the UAV's lidar performs a reciprocating scan covering the entire building facade. The system records the 3D coordinates and reflection intensity of each laser echo and generates a point cloud dataset. Each point in the dataset contains 3D coordinates and a time sequence number to ensure the orderliness and traceability of the points in subsequent processing. The surface parameterization takes the entire facade as the object. First, the horizontal range of the parameter domain is determined to be 0 to 1, and the vertical range is also determined to be 0 to 1. Then, using the circumscribed rectangle of the building facade as a reference, the point cloud is orthogonally projected along the facade normal to obtain the horizontal and vertical coordinates of each point. The normalized relative position ratio is used as the horizontal and vertical parameter values ​​for that point, respectively, thus completing the determination of the 3D coordinates to the parameter domain coordinates. To avoid parameter overlap, the system performs density equalization on multiple points within the same projection grid, and sparsifies them according to the rule that the minimum point spacing is not less than the average spacing of adjacent grids, retaining only representative points and recording the original 3D coordinate index of the retained points. Constraints are set according to the geometric features of the building facade, specifically identifying the facade outline, corner lines, seam lines, and opening edges in the point cloud. The identification rule is that the normal change of adjacent points exceeds a fixed threshold, and adjacent... A point is identified as a feature line segment if it simultaneously meets any two of the following three criteria: a sudden change in the first-order difference of its elevation or horizontal coordinates, or a significant peak in point density along a certain line. All confirmed feature line segments are integrated into a constraint parameter set after breakpoint piecing and orientation consistency checks. Each feature line segment consists of a sequence of feature points arranged in the parameter domain, along with its start and end parameter values ​​in the parameter domain and the corresponding point index in 3D space. An initial parametric surface model with boundary condition constraints is then established using the constraint parameter set. Specifically, a regular mesh within the parameter domain is used as the initial control framework to define the outer contour. Feature points are fixed at corresponding positions around the parameter domain, and corner and seam feature points are fixed at corresponding positions inside the parameter domain. At the same time, each fixed feature point is required to maintain a one-to-one correspondence between the three-dimensional space and the parameter domain. The system fills non-feature grid points in the order from the outside to the inside. The three-dimensional coordinates of non-feature grid points are interpolated by weighting the ratio of the spatial position of their adjacent known points to the relative position of the parameter domain. The weights are determined according to the proportion of the parameter distance from the adjacent known points to the point. The parameter distance is represented by the sum of the absolute values ​​of the horizontal parameter difference and the vertical parameter difference. The sum of the weights is fixed to 1 to ensure the uniqueness of the interpolation result.Immediately after the initial model is generated, the model boundary integrity is assessed. Integrity is defined as the ratio of the length of the established continuous boundary to the total expected boundary length given by the constraint parameter set. The length of the continuous boundary is obtained by accumulating segment by segment along the feature point sequence of the outer contour. If there is a discontinuity, the length of the discontinuity is included in the missing length. The total expected boundary length is the sum of the target lengths of the four sides of the outer contour and the feature line segments around the opening. The model boundary integrity threshold is fixed at 97%. The threshold is determined by using the allowable defect ratio of the curtain wall aluminum support structure not exceeding 3% as the control standard, and setting the upper limit of the acceptable boundary missing ratio at 3%. Therefore, the threshold is set to 97%, and this threshold remains unchanged in all projects to ensure consistency in judgment. When the model boundary integrity is lower than 97%, the system initiates boundary interpolation adjustment. The adjustment steps are as follows: each boundary discontinuity is processed sequentially. First, the known feature points at both ends of the discontinuity are located, the discontinuity length is calculated, and the number of boundary control points to be inserted is determined. The discontinuity length is divided by the average point spacing of the feature lines and rounded to the nearest integer. Then, equally spaced parameter positions are generated along the line connecting the two endpoints in the parameter domain. At each inserted parameter position, a linear interpolation is performed using the ratio of the direction and distance of the two endpoints in 3D space to determine the 3D coordinates. At the same time, the parameter distance between the inserted point and the adjacent point is not less than half of the original average parameter distance to prevent over-density. After all inserted points are added to the constraint parameter group, the continuity of the outer contour and the edge of the opening is updated again until all discontinuities are filled or the remaining missing length of a single discontinuity is less than 20% of the average point spacing of the feature lines. At this point, the boundary is considered to meet the continuity requirement. After all discontinuities are processed, the integrity of the model boundary is recalculated. When the integrity is not less than 97%, the boundary complete surface model is output. Otherwise, the same interpolation loop is continued. The upper limit of the loop is fixed at 5 times to prevent invalid repetition. When the upper limit is reached but the threshold is still not met, the missing positions are marked in the form of a list and a prompt is made that supplementary data needs to be collected. The horizontal and vertical ranges of the parameter domain are fixed at 0 to 1 to unify the scale and ensure consistency in subsequent comparisons. The horizontal and vertical parameter values ​​are the normalized positions of the points within the parameter domain, calculated directly from the horizontal and vertical relative distances of the points within the circumscribed rectangle of the elevation. The minimum point spacing is the lower limit of the spatial resolution of the point cloud, determined by the median distance between adjacent points in the point cloud and fixed as a representative value for the dataset. The normal variation threshold is a fixed threshold for identifying feature line segments, obtained by comparing the median and upper quartile values ​​of the normal variation extracted from the sample elevation and taking a constant value between them to ensure consistent judgment across different datasets. The first-order difference mutation threshold is fixed at twice the difference between the upper quartile and median values ​​of the coordinate differences between adjacent points, used to eliminate random noise interference. The peak point density threshold is fixed at 1.5 times the average point density of the region, used to identify high-density linear structures such as seams and polygonal lines. The average point spacing of feature lines is obtained by dividing the total length of the feature line segments by the number of feature points, used to determine the number and spacing of interpolation points.The upper limit of the interpolation loop is fixed at 5 times to limit computation time and avoid over-filling of boundaries in noisy areas.

[0024] The model boundary integrity threshold is set at 97%, based on the fact that boundary data loss caused by occlusion, glass reflection, and incident angle limitations during drone flight around buildings is typically no more than 3%. Setting the acceptable upper limit of loss at 3% ensures boundary continuity while avoiding over-patch of a small number of unavoidable gaps. The peak point density threshold is set at 1.5 times the average point density of the region, used to identify linear features such as seam lines and polygonal lines. The rationale is that linear structures form stable sampling clusters along the line direction, which usually significantly improves performance but is limited by scanning resolution; setting it to 1.5 times the average point density ensures better performance. It is stable enough to distinguish between structural clustering and random fluctuations without introducing a large number of false alarms; the first-order difference abrupt change threshold for adjacent point coordinates is twice the difference between the upper quartile and the median value. The reason is that this difference can characterize the tail dispersion of the data, and using twice as the abrupt change limit can preserve the true geometric transformation while shielding most of the occasional jumps caused by noise; the normal change threshold is selected as a fixed value between the median and the upper quartile value of the normal change. The reason is that the normal change in the large area of ​​the facade is concentrated near the median value, while the normal change at corners and broken lines is stably higher than the upper quartile value. The threshold is set between the two values ​​to stably capture the edge without excessively encroaching on flat areas. The lower limit of the minimum parameter distance for interpolation of non-feature grid points is half of the original average parameter distance. The reason is that while ensuring topological order and interpolation stability, moderate densification is allowed to suppress hole propagation and numerical oscillation. The upper limit of the remaining discontinuity allowed after boundary interpolation is 20% of the average point spacing of the feature lines. The reason is that when the remaining gap is much smaller than the typical sampling spacing, it will not have a substantial impact on subsequent deviation calculation and surface continuity evaluation. The benefit of continuing to fill in the gap is lower than the risk of introducing overfitting. The upper limit of the interpolation cycle is 5 times. The reason is that boundary defects have been significantly repaired in the first few rounds. The benefit of more than 5 times tends to decrease and will significantly increase the computation time. Setting an upper limit can prevent repeated interpolation in noise-dominated areas. The minimum point spacing for point cloud sparsification is taken as the median distance between adjacent points as a representative value. The reason is that this statistic is not sensitive to outliers and can stably reflect the actual resolution of the sensor. The horizontal and vertical ranges of the parameter domain are fixed at 0 to 1. The reason is that after unifying the scale, the parameter distances of different facades and different zones can be directly compared and the consistency of weight allocation can be maintained.

[0025] Geometric deviations are calculated based on the initial parametric surface model and the design datum. Spatial deviations at each point are obtained using residual calculation. Differential weights are assigned to deviation points in different regions using a weighted coefficient allocation technique. A weighted objective function expression is constructed based on these residuals. The process includes: calculating the Euclidean distance between the point coordinates using the residuals to obtain the spatial deviation at each point, resulting in a deviation set; allocating weighted coefficients for different regions based on the deviation set, setting proportional coefficients according to regional curvature characteristics to determine differentiated weight value groups; integrating weighted residual terms from these differentiated weight value groups, multiplying them by the corresponding deviations, and summing the results to construct the weighted objective function expression; introducing boundary constraints to define parameter boundary ranges and obtaining an optimized function parameter set; and using this function parameter set to correct deviations on the building facade surface, adjusting the point coordinate values ​​to obtain a corrected surface representation.

[0026] In one possible implementation, the initial parametric surface model, design reference plane, and point cloud point set are first input. For each point, its unique reference position on the design reference plane is determined by using the position corresponding to the shortest spatial distance from the point to the design reference plane as the reference position, and the value of this shortest distance is recorded as the spatial deviation of the point. The spatial deviations of all points are then arranged into a deviation set according to the point index order. Subsequently, the parameter domain is divided into regions. The division rule is to divide the parameter domain into a fixed number of equal-width regions in both the horizontal and vertical directions, and then combine them in pairs into rectangular partitions of equal size. The fixed number is... The number of partitions is determined by the point cloud size to ensure that each partition contains no fewer than 100 points. If the point cloud size is insufficient to meet this requirement, the number of partitions is automatically reduced until each partition meets the requirement of no fewer than 100 points. This lower limit is used to ensure the stability of subsequent curvature statistics. Within each partition, the curvature characteristics are calculated. The calculation steps are as follows: for each point, select several nearest neighbor points with the closest distance to its parameters. The number of neighbor points is fixed at 8. If a partition has fewer than 8 points, all points within that partition are used. The directional change and degree of bending between the point and each neighbor point are calculated sequentially. The sum of these two values ​​yields the local curvature measure of that point. Then, the median of the local curvature measurements for all points within the partition is taken as the representative curvature value for that partition. All representative curvature values ​​for all partitions are sorted from smallest to largest and divided into five equal quantile segments. These quantile segments correspond to proportionality coefficients from 1 to 5, which are fixed values ​​and remain constant throughout the same project. To avoid over-amplification or over-weakening, upper and lower limits are set before weight allocation: a lower limit of 1 and an upper limit of 5. When the proportionality coefficient of a partition is lower than 1, it is counted as 1; when it is higher than 5, it is counted as 5. The base weight is used to offset differences in sampling density. The base weight within each partition is calculated by multiplying the inverse of the number of points in that partition. The total number of partitions is multiplied by a fixed scaling factor to obtain the base weights of all partitions, which are summed to 1 across the entire domain. This fixed scaling factor is determined during the initial calculation and remains unchanged throughout the calculation process. The final weight of each point is equal to the product of the scaling factor of its partition and the base weight of that partition, resulting in a set of differentiated weight values. The set of differentiated weight values ​​is matched one by one with the set of deviations. For each point, the weighted residual term is calculated as the product of the final weight of that point and the spatial deviation of that point. The weighted residual terms of all points are summed to obtain a single target value, which is the numerical representation of the objective function based on the weighted residuals.After the target value calculation is completed, boundary constraints are introduced. These constraints include two types of rules: one is the parameter boundary range rule, which stipulates that both the lateral and longitudinal parameters of the parameter domain must not exceed the range of 0 to 1, and that boundary points must remain within the allowable zone of their original boundary positions. This allowable zone is defined as no more than 50% of the average distance from the original boundary to the nearest inner point along the boundary normal, both inside and outside. This proportion ensures that the boundary shape remains unchanged while allowing for minor corrections. The other type is the feature line sequence rule, which stipulates that the nesting relationship between the outer contour line and the opening edge line must not be changed, and requires that the minimum distance between adjacent feature lines be no less than 5% of the average distance between the obtained feature line points. 0%, this minimum spacing is used to prevent overlap or reversal after correction; after the boundary constraints are fixed, the optimized function parameter set is acquired. The acquisition process involves selecting a set of parameter values ​​to describe the current surface correction and adjusting these values ​​successively. Each adjustment aims to reduce the target value. If the target value does not decrease after an adjustment, the adjustment is rolled back and a smaller adjustment amount is used. The initial adjustment amount is the coordinate change corresponding to one-hundredth of the target value. If the target value is not reduced after three consecutive adjustments, the adjustment amount is reduced to 50% of the original value until the target value decreases; when the target value decreases after five consecutive adjustments, the adjustment is considered complete. When the previous target value is 1% and all parameters at all points have not touched the boundary constraints, the adjustment is stopped, and the set of parameter values ​​at this time is used as the optimized function parameter set. If a point touches the boundary constraints during the adjustment process, the closest position of that point within the boundary allowable zone is fixed, and only the parameters of other points are adjusted. After obtaining the function parameter set, deviation correction is performed. The correction step is to update the coordinates of each point along the direction of the line connecting it to the reference position according to the corresponding correction amount in the function parameter set. During the update, points located on the boundary and feature lines are processed first, followed by internal points. The correction amount of internal points is determined when its local curvature measure is higher than the median of its partition. The system takes the full correction value from the function parameter set. When the local curvature measure is lower than the median of the corresponding partition, 80% of the full correction value is used. This percentage is fixed to prevent overcorrection in flat areas. After all points are updated, the spatial deviation is recalculated for each partition, and the target value is quickly verified to be lower than before correction. If it is not lower, the system reverts to the pre-correction state and reduces the most recent adjustment of the function parameter set to 50% of the original value before re-performing the correction. This reverting and recalculation is repeated a maximum of three times to prevent invalid loops. The final output corrected surface is the set of point coordinates obtained after applying the function parameter set and their positions in the parameter domain. The spatial deviation measures the actual geometric difference between each point and the design reference surface, determined by the unique shortest distance from the point to the surface. The number of partitions ensures statistical stability of curvature, and a lower limit of at least 100 points is used to suppress small sample fluctuations. The number of neighboring points is fixed at 8 for stable calculation of local curvature measures. Five levels of scaling factors are used to form a weighted gradient from low to high curvature across the entire domain.The lower limit of weight is 1, and the upper limit of weight is 5 to limit extreme weights and ensure that all regions are included; the normalization rule of the basic weight is used to maintain the consistency of the total contribution across different partitions; the parameter boundary range is 0 to 1 to maintain a uniform scale of the parameter domain; the boundary allowable band is 50% of the average distance to allow necessary fine-tuning while maintaining the boundary shape; the minimum spacing is 50% of the average spacing between feature lines to avoid overlap of feature lines after correction; the initial adjustment value is the coordinate change corresponding to one-hundredth of the target value to set a stable starting step; the adjustment amount is reduced to 50% of the original value to refine the search when the descent is blocked; the stopping condition is that the descent amplitude is less than one-hundredth of the previous target value for 5 consecutive descents to determine convergence; internal points at low curvature take 80% of the full correction value to prevent overcorrection; rollback and recalculation are limited to a maximum of 3 times to limit the calculation time and ensure the result is deterministic.

[0027] Setting the minimum number of points in a single partition to 100 ensures sufficient sample size for curvature statistics to suppress small sample fluctuations and stabilize representative values. The fixed number of neighboring points (8) covers the main directions of local geometric changes while avoiding excessive far-field noise from too many neighboring points. The curvature representative value is graded into 5 quantiles and assigned a scaling factor from 1 to 5 to form a defined weighted gradient from flat to dramatically changing regions, ensuring higher attention is given to high-curvature areas while preventing excessive amplification of flat areas. Setting the lower weight limit to 1 and the upper weight limit to 5 limits instability caused by extreme weights and ensures all regions are included in the target value. The base weights are normalized based on the number of points in each partition and summed to 1 across the entire region to eliminate contribution imbalances caused by differences in sampling density and ensure the overall impact of different partitions is on the same scale. The parameter boundary range is fixed from 0 to 1 to unify the parameter scale for direct comparison of parameter distances between different facades and partitions. The boundary allows for the inclusion of average edge values. The boundary distance of 50% to the nearest inner point is to provide necessary fine-tuning space while maintaining the outer contour and opening shape, thus avoiding boundary jaggedness or reversal; the minimum spacing of feature lines is taken as 50% of the average spacing of feature lines to prevent overlap or order reversal after correction while preserving detailed shapes; the initial adjustment amount is taken as the coordinate change corresponding to 1 / 100 of the target value to reach the effective descent zone with a stable starting step without oscillation; convergence is determined when the target value decreases by less than 1 / 100 of the previous target value for 5 consecutive times to ensure that the benefit of subsequent adjustments is negligible and thus terminate the calculation; when the descent is blocked, the adjustment amount is reduced to 50% of the original value for refining the search and avoiding crossing the optimal valley; only 80% of the full correction value is applied to internal low curvature points to reduce the risk of overcorrection in flat areas and concentrate correction resources at the points of shape change; the upper limit of three backtracking and recalculation is to limit time costs and avoid repeated oscillations in noise-dominated local areas.

[0028] If the gradient value of the objective function exceeds the preset fitting accuracy control range, the function gradient calculation program is initiated. A convergence threshold setting technique is used to determine the optimization termination condition. An iteration step size adjustment mechanism controls the magnitude of each parameter update, establishing an adaptive iterative optimization solution system. This includes obtaining the current gradient value from the gradient threshold check, comparing it against the accuracy control range, and activating function gradient calculation if it exceeds the threshold. The gradient vector is calculated by differentiating the objective function and combining it with the current parameter values ​​to obtain the updated gradient direction. For the updated gradient direction, a convergence threshold setting technique is used. By comparing the gradient vector magnitude with the preset threshold, the optimization termination condition is determined, and the iteration step size adjustment value is obtained. The iteration step size adjustment value controls the parameter update magnitude. By multiplying by the gradient direction and applying a decay factor, an adaptive iterative optimization solution system is established, resulting in an optimized parameter set. From the optimized parameter set, deviation correction is performed on the building facade surface. By adjusting the surface parameter values ​​and recalculating the point coordinates, the corrected surface representation is obtained.

[0029] In one possible implementation, the current gradient value is first read from the gradient threshold check. The gradient value is a length measure of the overall change intensity of the objective function with respect to all parameters. The upper limit of the precision control range is set as the activation threshold. The activation threshold is determined at the beginning of this optimization by using one-twentieth of the initial gradient value obtained in the first round of trial calculations, that is, reducing the initial gradient value of the first round to five percent as the upper limit standard. When the current gradient value is greater than the activation threshold, the function gradient calculation process is immediately started. The calculation process involves performing positive and negative trials on each parameter in the parameter set with extremely small and fixed amplitudes. The perturbation is probed, with the amplitude of the perturbation set to one-thousandth of the allowable range of variation for the parameter. The change in the objective function after each perturbation is calculated, and this change is divided by the difference in perturbation amplitude to obtain the rate of change of the parameter's direction. The rate of change of all parameters' directions is then combined into a gradient vector in the original parameter order. The length of this vector is calculated, and its reverse direction is recorded as the update direction. Next, the convergence threshold is set and determined. The convergence threshold is one-fifth of the activation threshold, which is one-hundredth of the initial gradient value. The length of the gradient vector is compared with this convergence threshold; if the length is not greater than the threshold, the termination condition is met, and the result is output. The current parameter set is taken as the optimal value for the current stage. When the length exceeds the threshold, step size generation and adaptive adjustment are initiated. The initial step size is taken as the equivalent change in the parameter space mapped from one percent of the average deviation. If the average deviation is not provided, one percent of the average value of the allowable range of each parameter is used as a substitute. Subsequently, the step size is adjusted based on two indicators: the relative decrease of the objective function from the previous round to the current round and whether the decrease is continuous. When the relative decrease is not less than two percent and there are at least two consecutive decrease rounds, the step size is increased to twice that of the previous round, but does not exceed the upper limit step size. The upper limit step size is fixed at the initial value. The step size is quadrupled. When the relative decrease is less than 2% or a non-decline occurs, the step size is reduced to half of the previous round, but not less than the lower limit step size, which is fixed at one-quarter of the initial value. To suppress oscillations while maintaining the downward trend, the step size of this round is multiplied by the decay factor to obtain the actual update magnitude. The decay factor is initially set to 0.8. When a non-decline occurs or the decrease is less than 1%, the decay factor is multiplied by 0.8 and tightened once until it is not less than 0.2. When the decline continues for 3 consecutive rounds and the decrease in each round is not less than 2%, the decay factor is increased to 1.25 times that of the previous round, but not exceeding 1.0; After determining the actual update magnitude, apply a fixed increment to each parameter in the update direction. The new parameter value is equal to the old value after moving the actual update magnitude along the update direction. After completing one parameter update, immediately recalculate the objective function value and repeatedly calculate the new gradient vector and its length. At the same time, record the current objective function value, current gradient length, current step size, current decay factor, and whether it has decreased in the iteration history. Then, decide whether to continue iterating based on three conditions: whether the objective function has decreased strictly compared to the previous round, whether the gradient length has decreased by at least one percent compared to the previous round, and whether the step size is between the upper and lower limits. If any condition is not met, backtrack to the parameters of the previous round, reduce the step size to half of the previous round, multiply the decay factor by 0.8, and try to update again. Each backtracking counts once, and the number of backtrackings does not exceed 3. If more than 3 backtrackings are performed, retain the parameter set corresponding to the current minimum objective function and enter the next stage to avoid invalid loops. When the gradient length decreases to no more than the convergence threshold or the relative decrease of the objective function in 5 consecutive rounds is less than the threshold, the iteration continues. When the objective function reaches 1 / 100 of the previous round, the iteration stops and the parameter set at that time is determined as the optimized parameter set. After obtaining the optimized parameter set, deviation correction is performed. The correction process is to determine and adjust the surface parameter values ​​according to the correction amount corresponding to the optimized parameter set. The adjustment order is to process the parameter points located at feature positions such as the outer contour and the opening first, and then process the internal parameter points. The coordinate update amount of the internal parameter points is weighted according to their spatial proximity relationship with the surrounding updated feature points. The proximity relationship is determined by a fixed number of nearest neighbor points, with a fixed number of 8. When there are less than 8 available neighbor points, all available neighbor points are taken. The weight is directly used for coordinate weighting after normalization according to the reciprocal of the proximity distance. After obtaining the new point coordinates, they are merged into the corrected surface representation. The corrected objective function value is calculated to verify whether it has decreased compared to before correction. If it has not decreased, it reverts to the state before correction and reduces the step size to half of the current value and multiplies it by the decay factor by 0.8, and repeats the correction verification once. It is repeated a maximum of 2 times. Finally, the corrected surface representation with the objective function minimized and meeting the convergence threshold condition is output. The activation threshold is set to one-twentieth of the initial gradient value to trigger gradient calculation; the convergence threshold is set to one-hundredth of the initial gradient value to determine termination; the initial step size is set to one-hundredth of the average deviation or the average of the allowable range of variation to set the stable starting amplitude; the upper limit step size is four times the initial value and the lower limit step size is one-quarter of the initial value to limit the update scale; the decay factor adaptively varies between 0.2 and 1.0 to eliminate oscillations and accelerate convergence in the stable phase; the maximum number of backoffs is three and the maximum number of correction and verification repetitions is two to constrain time costs.

[0030] The gradient activation threshold is set to 5% of the initial gradient because the gradient is usually large in the early stages of error. Only when the current gradient exceeds this proportion does it indicate that the region is still clearly descentable. At this point, initiating precise gradient calculation can obtain a definite update direction and avoid unnecessary calculations caused by repeated triggering in the low gradient stage. The convergence threshold is set to 1% of the initial gradient to lock the termination criterion within a sensitive but stable range. Below this level, the improvement of the target value by any parameter update is less than the combined impact of measurement noise and point cloud dispersion. Continuing iteration has no engineering significance for improving the final surface geometric accuracy. The initial step size is set to 1% of the average deviation or the average of the allowable variation range to provide a controllable starting amplitude, so that the first parameter update can cross the numerical plateau without crossing the optimal neighborhood. The upper limit step size is set to 4 times the initial value and the lower limit step size is set to one-quarter of the initial value to amplify the exploration efficiency when there is a continuous and significant descent and to forcibly tighten the step size when the descent stalls or rebounds, thereby avoiding oscillations and exceeding the limits. The step size is increased when the decrease reaches 2% for two consecutive rounds. This is based on the empirical rule that the weighted residual often exhibits a monotonically slow change in the early and middle stages. Accelerating the process after confirming a continuous and effective decrease can significantly shorten the number of iterations. When the relative decrease is less than 2% or a non-decreasing condition occurs, the step size is reduced to ensure that the update falls back into the descent range and to reduce disturbances to boundary-limited points. The decay factor is initially set to 0.8 and adaptively adjusted between 0.2 and 1.0 to provide a stable buffer under unknown surface shapes. When the decrease is insufficient or there is a rebound, the decay factor is reduced to suppress oscillations. When the decrease is continuous and stable, it is increased to no more than 1.0 to proceed with the planned step size. The iteration is stopped when only a slight decrease of less than 1% occurs for five consecutive rounds because the objective function is in a flat region at this point, and the benefit of continuing the search is lower than the cost of computation and data reprojection. Parameter updates are always performed within the predetermined boundary range to ensure that the geometric relationship between the outer contour and boundary points such as the opening is not destroyed and to ensure that the correction results can be directly used for subsequent health monitoring.

[0031] The first round of parameter update calculations is performed based on the adaptive iterative optimization solution system. Error change monitoring technology is used to track the changing trend of the objective function value in real time. The current iteration efficiency is judged by the convergence speed evaluation algorithm. Historical iteration data is recorded, including the first round of parameter update calculation results obtained from the adaptive iterative optimization solution system. Error change monitoring technology is used to calculate the difference sequence by continuously sampling the objective function value sequence. The difference sequence is then fitted with a trend to obtain the change trend. Based on the change trend, the current iteration step size is compared with a preset threshold using the gradient descent algorithm. The gradient descent algorithm takes the change trend and step size as inputs and outputs an efficiency index to determine the iteration efficiency. Historical data records are extracted from the iteration efficiency. The data is serialized and written to a preset buffer using a data storage mechanism to obtain a record set. Subsequent analysis and processing are performed on the record set. The deviation vector is calculated by comparing the historical data with the current parameters to obtain the optimization adjustment value. The optimization adjustment value is applied to the correction of the building facade surface parameters. Deviation correction technology is used to adjust the coordinates of the surface points by weighted averaging to obtain the correction representation.

[0032] In one possible implementation, the first round of parameter update calculation results and their corresponding objective function values ​​are first read from the adaptive iterative optimization solution system. Then, a sequence of objective function values ​​arranged in chronological order is generated with a fixed sampling period, where sampling occurs once after each parameter update. A difference sequence is obtained by subtracting adjacent sampled values; the difference is the sum of the previous and subsequent objective function values. To obtain the trend, the system performs trend fitting on the difference sequence within a sliding window of length 10. This involves calculating the arithmetic mean of the differences within the window, calculating the sign of the linear slope in chronological order, and determining whether the two slopes are the same sign. If they are the same sign and negative, it indicates a stable downward trend; if they are the same sign and positive, it indicates a stable upward trend; and if they are different signs, it indicates an oscillating trend. Simultaneously, the results are recorded. The absolute magnitude of the average value is used as the descent strength index. After obtaining the trend, the algorithm proceeds to iterative efficiency evaluation. The input to the gradient descent algorithm is the trend and the current iteration step size, which is the actual step size retained from the previous stage. The algorithm calculates the efficiency index according to the following process: When the trend is a stable descent, the product of the descent strength index and the step size is used as the expected descent amount, and the ratio is taken with the actual descent amount in this round to obtain the efficiency index. When the trend is oscillating or steadily rising, the efficiency index is directly set to a fixed value of 0.5, which is less than 1, to trigger tightening. The efficiency index threshold is set to two levels: the first threshold is 1, and the second threshold is 0.8. When the efficiency index is greater than or equal to 1, the current step size is considered to match and the iteration efficiency is output as high. When the efficiency index is between 0.8 and 1, the iteration efficiency is output as medium. When the efficiency index is less than 0...At 8:00 AM, the output iteration efficiency was low. The system then compiled historical data records, including the current change trend, descent strength index, current iteration step size, efficiency index, objective function value, parameter values, and timestamp. The record field order was fixed as timestamp, objective function value, iteration step size, efficiency index, change trend, descent strength index, and parameter values. A serialization method was used to convert the scalars and vectors in this order into a continuous byte stream and write it to a preset buffer. The buffer capacity was fixed, storing the most recent 100 records. When full, the buffer was overwritten using a first-in, first-out (FIFO) approach, forming a record set. Subsequent analysis and processing were performed on the record set. First, the best historical record was determined, which was the record with the lowest objective function value and a step size no less than one-quarter of the initial step size. The parameter value recorded is taken as the historical best parameter. The current parameter is then subtracted from the historical best parameter one by one to obtain the components of the deviation vector. This is then robustly weighted by time, with the weighting rule being that the closer to the current value, the higher the weight. The weighting coefficients of the most recent 10 records are determined by an integer sequence from 10 to 1 and normalized before being applied to the contribution summation of the corresponding components. Based on this, the system calculates the optimization adjustment value. The determination principle is as follows: when the iteration efficiency is high and the trend is a stable decrease, the optimization adjustment value is taken as one-tenth of the magnitude in the opposite direction of the deviation vector to refine it; when the iteration efficiency is medium or the trend is oscillating, the optimization adjustment value is taken as one-fifth of the magnitude in the opposite direction of the deviation vector, and subsequent step size candidate values ​​are marked as needing to be scaled down; when the iteration efficiency is... When the trend is low or steadily rising, the optimized adjustment value is taken as half the amplitude in the same direction as the deviation vector to quickly return to the historical best neighborhood. After obtaining the optimized adjustment value, it is applied to the correction of the building facade surface parameters. The correction order is: priority for outer contour parameter points and opening parameter points, followed by internal parameter points. After the parameters are updated, the coordinates of the surface points are adjusted by weighted average according to the deviation correction technique. The weighting consists of two parts: the sum of regional weight and stability weight is 1. The regional weight is fixed at 0.6 to maintain the priority of high curvature and key structural areas, and the stability weight is fixed at 0.4 to express the degree of reversal of the parameter fluctuation of the point in the record set. The stability is determined by statistically analyzing the standard deviation of the parameters of the point in the most recent 20 records. The stability coefficient is obtained by reciprocal normalization, with higher stability resulting in greater weight. For each point, its eight nearest updated neighboring points in the parameter space are collected first. The reciprocal of the spatial distance to these neighbors is calculated and normalized to obtain interpolation coefficients. The regional weight and stability weight are then multiplied proportionally by the interpolation coefficients to obtain the final weight. The coordinates of the neighboring points are then weighted using this final weight to obtain the new coordinates of the point, thus forming the corrected representation. After completing one round of global weighted averaging, the new objective function value is immediately calculated and compared with the previous objective function value. If the new value is not higher than the old value and the efficiency index is greater than or equal to 0.8, the current round of correction is confirmed as effective and written into the record set as new historical data. If the new value is higher than the old value or the efficiency index is less than 0, the correction is considered effective.At 8:00, the algorithm reverts to the pre-correction state and reduces the current iteration step size to half of the previous round, then recalculates the optimized adjustment value and weighted average. The maximum number of reverts is 2; if this is exceeded, the state corresponding to the minimum objective function before the revert is retained for the next round of sampling. In this implementation, the sampling period is defined as one sampling after each round of updates to ensure synchronization with parameter changes. The sliding window length is 10 to stably identify trends in the short term. An efficiency threshold of 1 is used to determine if the step size is sufficiently matched, and an efficiency threshold of 0.8 is used to determine if a smaller step size is needed. A buffer capacity of 100 entries is used to balance traceability and memory usage. The lower limit of the historical best step size is one-quarter of the initial step size to exclude pseudo-optimal values ​​caused by accidental small step sizes. The number of nearest neighbors is 8 for stable interpolation calculations. The region weight ratio of 0.6 and the stability weight ratio of 0.4 are used to achieve a clear ratio between shape fidelity and numerical stability. The maximum number of reverts is 2 to limit time costs and prevent oscillation cycles.

[0033] The sampling period is set to one sample after each parameter update because the effective change in the objective function only has statistical significance after a complete update. Overly dense sampling will lead to noise dominating the difference and reduce the reliability of the judgment. The sliding window length is set to 10 to ensure sufficient sample size to stabilize the trend while covering short-term fluctuations. Less than 10 will amplify occasional jumps, while more than 10 will delay the response to trend reversals. During oscillations or upward movements, the efficiency index is directly assigned a value of 0.5 to force a tightening strategy with a definite low-efficiency signal, avoiding further amplification of the step size under unfavorable trends. The efficiency index grading threshold is set as follows: The thresholds are set to 1 and 0.8, where 1 is used to determine if the step size and descent rate perfectly match and allow for maintenance or even moderate amplification, while 0.8 is used to determine if the descent efficiency is insufficient and the step size needs to be reduced. This clear two-level threshold division avoids repeated jittering at boundary states. The buffer capacity is fixed at 100 records to ensure that at least 10 windows of history are covered to support robust backtracking while controlling storage usage and write overhead. Too small a capacity will lead to history loss, while too large a capacity will increase retrieval and serialization time. When selecting the best history, the step size is limited to no less than one-quarter of the initial step size to eliminate false optima caused by accidental local descents obtained through extremely small step sizes. The design prioritizes representativeness and sustainability of the reference solution. Eight nearest neighbors are used for coordinate weighting, providing the minimum sufficient neighborhood for anisotropic distribution under common densities in 3D facade point clouds. Fewer than eight neighbors weakens interpolation stability, while more than eight introduces far-field interference and dilutes local shape. The region weight and stability weight are fixed at 0.6 and 0.4 respectively, based on the engineering requirement of prioritizing high curvature and critical structural areas in curtain wall geometry control, while also considering the suppression of numerical oscillations by point temporal stability. The sum of these two weights is 1 to ensure dimensional consistency and facilitate auditing. The upper limit of backoff is set to two times to allow for adjustments after corrections occur. When efficiency is insufficient, limited self-correction opportunities are provided and time costs are strictly limited. If more than two self-corrections are made, it indicates that the current search direction or step size is inappropriate and the next round of strategy evaluation should be initiated. When the iteration efficiency is high and steadily decreasing, the optimization adjustment value adopts one-tenth of the magnitude in the opposite direction of the deviation vector. This is to refine the search without overshooting when approaching the optimal neighborhood. When the iteration efficiency is medium or oscillating, one-fifth of the magnitude is adopted and the step size is slightly reduced. This is to quickly suppress oscillations while maintaining a net advance towards the optimal. When the iteration efficiency is low or steadily increasing, half of the magnitude is adopted to retreat along the same direction as the deviation vector. This is to ensure a significant departure from the unfavorable region and return to the historical optimal neighborhood.

[0034] Numerical stability checks are performed based on iterative historical data. Local optimum detection techniques are used to identify whether the system has fallen into a local extreme state. The current optimization state is evaluated to determine if it meets convergence requirements using iteration termination conditions. If it meets a set convergence threshold, the optimal parameter combination is output. This involves obtaining a numerical sequence from the iterative historical data, using a difference calculation method to obtain a difference sequence by subtracting adjacent values, averaging the absolute values ​​of the difference sequences to obtain a stability index, and using the stability index to determine if a preset threshold is met, thus determining the numerical stability state. For the numerical stability state, gradient analysis is used to identify local extreme features by calculating the slope of the numerical sequence, extracting deviation vectors from these features, and summing these deviation vectors to obtain the extreme state identification result. Based on the extreme state identification result, iteration termination condition parameters are obtained. By comparing the current optimization state with the iteration termination condition parameters, it is determined whether the system has fallen into a local extreme state. If the determination shows that the system has not fallen into a local extreme state, the optimal parameter combination is output from the optimization state evaluation.

[0035] In one possible implementation, the objective function values ​​are first read from the iterative history data in chronological order to form a numerical sequence. Then, a difference sequence is obtained by subtracting the previous term from the next term for each adjacent term. The absolute values ​​of the differences are then averaged over the most recent 10 terms to obtain a stability index. The stability threshold is determined using a two-step method: first, the median of the absolute values ​​of the differences over the previous 10 terms is used as the baseline noise amplitude; second, this baseline is multiplied by 1.2. A stability threshold is obtained to balance noise fluctuations and misjudgment control. When the stability index is less than or equal to the stability threshold, the current state is considered stable; when the stability index is greater than the stability threshold, the state is considered unstable and observation continues. After stability determination, gradient analysis is performed using a sliding window of fixed length 10. Within each window, the window slope is obtained by subtracting the first item from the last item of the sequence and then dividing by the number of intervals within the window. The window slope is compared with zero to determine whether the overall trend is decreasing, flat, or increasing. Simultaneously, the sign change of the difference sequence is checked point by point within the window to identify local extrema. Specifically, a change in sign from negative to positive is recorded as a local minimum, and a change in sign from positive to negative is recorded as a local maximum. To avoid... To avoid noise-triggered misjudgments, a lower limit for extreme value determination is set at one-tenth of the stability threshold. An extreme value is confirmed only when the absolute value of the difference between the two sides before and after the extreme value is not less than this lower limit. At each confirmed local extreme value, a deviation vector is constructed, its direction determined by the orientation of the undesirable upward region after moving away from the extreme value. The deviation vector for the minimum value is oriented towards the direction of continued descent, and the deviation vector for the maximum value is oriented towards the direction of continued ascent to move away from the peak. The length of the deviation vector is the average difference of the absolute values ​​of the two differences to the left and right of the extreme value. If there are fewer than two usable terms, the average difference of the usable terms is used. All deviation vectors are summed in chronological order to obtain the total length and total direction of the extreme value state identification result. Subsequently, based on the extreme value state identification... The results and trend information generate an iteration termination condition parameter set, which includes four thresholds and criteria, each with its own determination method: The first is the stability threshold, derived from the aforementioned two-step method and unchanged; the second is the slope threshold, calculated by dividing the stability threshold by the number of intervals within the window, used to define the "approximately level" numerical band, where the absolute value of the window slope is less than or equal to this threshold; the third is the extreme value density threshold, setting the upper limit of the number of local extreme values ​​allowed within the last 10 terms to 2, with more than 2 considered extreme value density; the fourth is the extreme value intensity threshold, using the stability threshold as the lower limit of the length of the deviation vector summation, where a summation length greater than or equal to this threshold indicates a significant impact of extreme values; the system then compares the current optimization state with the four thresholds... The values ​​are compared one by one. If all four of the following conditions are met simultaneously, it is determined that the system has not fallen into a local extremum state and continues the process: the stability index is less than or equal to the stability threshold, the absolute value of the window slope is less than or equal to the slope threshold, the number of local extrema in the most recent 10 terms does not exceed 2, and the length of the sum of the deviation vectors is less than the extremum intensity threshold. If any condition is not met, it is marked as having a risk of local extrema and the system returns to the previous stage to execute convergence-related tightening and correction strategies. When it is determined that the system has not fallen into a local extremum state, the system outputs the optimal parameter combination from the optimization state evaluation. Specifically, it selects the parameter group with the smallest objective function value and the stability index not exceeding the stability threshold from the parameter groups corresponding to the most recent 10 terms as the optimal parameter combination. The sliding window length is 10 to stably identify trends in the short term, and the stability threshold is derived from the median of the absolute values ​​of the differences of the previous 10 terms multiplied by 1.2. To balance robustness and sensitivity, the lower limit of the extreme value judgment amplitude is one-tenth of the stability threshold to suppress small noise triggering. The slope threshold is taken from the stability threshold and amortized over the window interval to define the smooth region. The extreme value density threshold is 2 to limit frequent round trips within a short window. The extreme value intensity threshold is equal to the stability threshold to unify the dimensions and ensure consistency for cross-project reuse.

[0036] The stability threshold is calculated by multiplying the median of the absolute values ​​of the differences of the top 10 historical values ​​by 1.2. This is because the median robustly represents the baseline noise without being affected by extreme values, and multiplying by 1.2 provides a clear safety margin to distinguish random fluctuations from true instability. The sliding window length is set to 10, which, with an iteration-per-sampling rhythm, covers short-term fluctuations while stabilizing trend determination without lag. A length less than 10 would amplify occasional jumps, while a length greater than 10 would delay the response to reversals. The lower limit of the extreme value determination amplitude is set to one-tenth of the stability threshold to eliminate small fluctuations far below the noise band, ensuring that the turning points marked as extreme values ​​have engineering significance. The slope threshold is the result of amortizing the stability threshold within the window, used to represent "approximate flatness". The gradual change region is clearly separated from the significant rising or falling region to avoid misjudging the trend due to the accumulation of small residuals; the extreme value density threshold is limited to 2 within the most recent 10 terms, which is based on the reasonable upper limit of the turning frequency during the convergence process. More than 2 indicates frequent back-and-forth within the short window, with a high risk of getting trapped in local extrema, requiring tightening and backtracking; the extreme value intensity threshold is taken from the stability threshold, using a unified dimension to measure the overall impact of the sum of the deviation vectors. When the intensity is below this threshold, the extreme value influence is considered not to constitute substantial interference, thus avoiding overreaction; the output of the optimal parameter combination is limited to the most recent 10 terms and must simultaneously satisfy the stability not being worse than the stability threshold. This is to ensure that the solution is both in the latest convergence stage and has passed the stability screening, reducing the risk of mistaking accidental low values ​​as the global optimum.

[0037] The optimized facade surface model is reconstructed based on the optimal parameter combination. Point cloud coordinate reprojection technology is used to convert the optimized parameters into corrected 3D coordinates. Geometric deviation comparison analysis is used to calculate the accuracy improvement before and after optimization, generating a building facade geometric accuracy optimization report. This includes obtaining the coordinate transformation matrix from the optimal parameter combination, using point cloud reprojection to convert the coordinates of each point in the original point cloud data into corrected 3D coordinates through matrix multiplication, obtaining the optimized coordinate set. For the optimized coordinate set, a facade surface model is constructed. Mesh subdivision technology is used to divide the coordinate points into triangular meshes, which are then interpolated and fused to generate a continuous curve. The process involves: determining the surface geometry; obtaining corresponding point pairs between the surface geometry and the original facade model; comparing the deviations between point pairs using the Euclidean distance calculation method to obtain a geometric deviation sequence; extracting the average deviation value and standard deviation from the geometric deviation sequence; obtaining the average deviation value by summing the sequence elements and dividing by the number of elements; obtaining the standard deviation by summing the squared differences of deviations, dividing by the number of elements, and then taking the square root; evaluating the difference in accuracy before and after optimization using subtraction operations to determine the accuracy improvement; and integrating the deviation sequence and surface structure data based on the accuracy improvement to generate a building facade geometric accuracy optimization report, including a deviation distribution map and improvement indicators.

[0038] In one possible implementation, the scaling factor, rotation factor around three orthogonal directions, and translation factor in three directions are first read sequentially from the optimal parameter combination. These factors are all fixed values ​​stored with six decimal places of precision and form a coordinate transformation matrix in a fixed order of scaling, rotation, and translation. The original point cloud data is then reprojected point by point. For each point, the coordinates in the three directions are first scaled proportionally according to the scaling factor, then rigidly rotated in the three directions according to the rotation factor, and finally, the coordinates in the three directions are added according to the translation factor to obtain the corrected 3D coordinates of that point. All corrected coordinates are combined into an optimized coordinate set according to the original sampling order. Subsequently, a facade surface model is constructed. The parameters for the mesh subdivision technique are set as two categories: target side length and subdivision merging threshold. The target side length is taken from the optimized coordinate set. The median distance from each point to its nearest neighbor is used. The subdivision threshold for long and thin sides is 1.5 times the target side length, and the merging threshold for short sides is 0.5 times the target side length. These three values ​​are constants and remain unchanged during this reconstruction process. The mesh is constructed incrementally. Specifically, on the elevation projection plane, the row and column order of the points is used as the index. First, three adjacent points form an initial triangle. Then, new points are added one by one in the order of the points and connected to the two points of the existing mesh closest to them in the projection plane to form a new triangle. If any side length is greater than 1.5 times the target side length, a new vertex is inserted at the midpoint of that side, and the original triangle is divided into two triangles. If any side length is less than 0.5 times the target side length, the two endpoints of that side are merged into the midpoint, and the midpoint replaces the original two endpoints to update the adjacent triangles, until all side lengths are between 0.5 and 1.Between 5 times the target side length; to obtain a continuous surface, interpolation and fusion are performed on surface points at arbitrary positions within each triangle. The interpolation weights are determined according to the ratio of the area of ​​the opposite vertex triangle to the total area. The three weights are non-negative and sum to 1. The corrected 3D coordinates of the three vertices of the triangle are weighted using these weights to obtain a continuous set of points on the surface, thereby determining the surface geometry; when establishing the corresponding point pairs between the surface geometry and the original facade model, a nearest neighbor search is performed on each surface point in the point set of the original facade model. The nearest neighbor is defined as... The original point with the smallest 3D coordinate difference length is used. If multiple original points have the same distance, the one with the smallest index is selected as the corresponding point. The 3D coordinate difference length of each pair of corresponding points is recorded one by one, forming a geometric deviation sequence according to the order of the surface points. Two statistics are calculated for the geometric deviation sequence: the mean deviation is the sum of all deviation values ​​divided by the number of deviation values, and the standard deviation is the square root of the sum of the squares of the differences between each deviation value and the mean deviation value, divided by the number of deviation values. Both statistics are retained to three decimal places. This is used to evaluate the performance before optimization. For post-optimization accuracy differences, the system reads the average deviation and standard deviation from the pre-optimization deviation statistics. The accuracy improvement is calculated using a predetermined subtraction: the average deviation improvement is the difference between the pre-optimization average deviation and the post-optimization average deviation; the standard deviation improvement is the difference between the pre-optimization standard deviation and the post-optimization standard deviation. Both positive values ​​indicate improvement, zero indicates no change, and negative values ​​indicate degradation. When generating the building facade geometric accuracy optimization report, the curved surface points are projected onto the facade reference plane and divided into units using equidistant grids. The arithmetic mean of the geometric deviations within each grid unit is taken as the representative deviation for that unit. A deviation distribution map is plotted based on the representative deviations. The upper and lower limits of the color band are determined by the range from the post-optimization average deviation minus twice the post-optimization standard deviation to the post-optimization average deviation plus twice the post-optimization standard deviation. Values ​​exceeding the upper and lower limits are truncated at the boundaries to avoid extreme values ​​affecting the color band. The report text includes predetermined parameters such as the post-optimization average deviation, post-optimization standard deviation, average deviation improvement, standard deviation improvement, target side length, subdivision and merging thresholds, and coordinate transformation order, along with a deviation distribution map and a description of the corresponding point pairs.

[0039] The coordinate transformation values ​​are retained to six decimal places because the typical ranging noise and time synchronization error of UAV LiDAR are on the order of millimeters to centimeters. Six decimal places, when measured in meters, can reduce the rounding error to the order of micrometers, without amplifying noise and facilitating reproduction across rounds. The target side length for mesh subdivision is taken as the median of the nearest neighbor distance rather than the average. This is to suppress scale shifts caused by outliers. After using this median as the global scale benchmark, a 1.5x threshold for thin and long sides and a 0.5x threshold for merging excessively short sides are used to form a symmetrical tolerance band for the target side length. The pairing of 1.5x and 0.5x limits the side length to between half and one and a half times the target value, preventing excessively long and thin triangles that cause interpolation instability and avoiding excessively dense splitting that causes a surge in computation, thus achieving a verifiable balance between mesh quality and efficiency. The establishment of corresponding point pairs uses the nearest neighbor in three-dimensional space, and when the distances are equal, the one with the smallest index is selected. This is to provide... The system ensures consistent deviation sequences in repeated calculations by using single-valued, definitive matching results and eliminating randomness. The spatial boundary of invalid points is capped by the bounding box of the original facade model, preventing deviations beyond this box. This avoids outliers outside the engineering scope affecting the statistical distribution after reprojection. This boundary is directly determined from the original model, eliminating the need for additional out-of-limit ratios and ensuring simple, auditable rules. The upper and lower limits of the color bands in the deviation distribution map are the average deviation value plus or minus two standard deviations. This is based on the engineering fact that curtain wall measurement errors are approximately symmetrically distributed after aggregation. Two standard deviations cover most normal deviations and truncate extreme outliers, preventing a few anomalies from affecting the color bands or compressing the main color contrast, facilitating intuitive interpretation. Numerical statistics in the report are retained to three decimal places because this precision matches the tolerance level of on-site component installation and testing, allowing for stable comparison of differences before and after optimization without introducing false precision.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for health monitoring of aluminum support structures for curtain walls based on UAV 3D mapping, characterized in that, include: Point cloud data of building facades are acquired by UAV LiDAR, and three-dimensional coordinate parameterization is performed on the point cloud using surface parameterization technology. Constraints are set according to the geometric features of the building facade to establish an initial parameterized surface model with boundary condition restrictions. Geometric deviations are calculated based on the initial parametric surface model and the design reference surface. Spatial deviations at each point are obtained using residual calculation. Differentiated weights are assigned to deviation points in different regions using weight coefficient allocation techniques. An objective function based on weighted residuals is constructed. If the gradient value of the objective function exceeds the preset fitting accuracy control range, the function gradient calculation program is started, the convergence threshold setting technique is used to determine the optimization termination condition, the iteration step size adjustment mechanism is used to control the parameter update magnitude each time, and an adaptive iterative optimization solution system is established. The first round of parameter update calculation is performed based on the adaptive iterative optimization solution system. Error change monitoring technology is used to track the changing trend of the objective function value in real time. The current iteration efficiency is judged by the convergence speed evaluation algorithm, and historical iteration data is recorded.

2. The method for health monitoring of aluminum support structures for curtain walls based on UAV three-dimensional mapping according to claim 1, characterized in that: The process of acquiring building facade point cloud data via UAV LiDAR, performing three-dimensional coordinate parameterization on the point cloud using surface parametric representation technology, setting constraints based on the geometric features of the building facade, and establishing an initial parametric surface model including boundary condition restrictions includes: By scanning the building facade with a drone's LiDAR, point cloud data is obtained from the scanning results, resulting in a point cloud dataset. A surface parametric representation technique is used to perform three-dimensional coordinate parametric processing on the point cloud dataset. The parametric coordinate values ​​are determined by mapping the point cloud points to the parameter domain. Based on the geometric features of the building facade, constraint conditions are set for the parametric coordinate values, and the constraint parameter set is obtained by integrating the feature line segments. By using a set of constraint parameters, an initial parametric surface model containing boundary condition constraints is established, and the integrity of the model boundary is determined. If the boundary integrity of the model is lower than the preset threshold, boundary interpolation adjustment is performed on the initial parametric surface model to obtain a boundary-complete surface model.

3. The method for health monitoring of aluminum support structures for curtain walls based on UAV three-dimensional mapping according to claim 1, characterized in that: The geometric deviation calculation based on the initial parametric surface model and the design reference surface is performed. The spatial deviation at each point is obtained using residual calculation. Differential weights are assigned to deviation points in different regions using a weighted coefficient allocation technique. The objective function based on the weighted residual is constructed by including the following expression: Based on the initial parametric surface model and the design datum, the Euclidean distance is obtained by calculating the difference in point coordinates using the residual calculation method, and the spatial deviation of each point is obtained to get the set of deviations. By using the set of deviations, weight coefficients are allocated for different regions, and proportional coefficients are set according to the curvature characteristics of the regions to determine the differentiated weight value groups. From the differentiated weight value group, integrate the weighted residual terms, multiply them by the corresponding deviations, and sum them to construct the objective function expression based on the weighted residuals; For the objective function expression, boundary constraints are introduced, parameter boundary ranges are set, and the optimized function parameter set is obtained; By using a set of function parameters, deviations in the building facade surface are corrected, and the coordinate values ​​of the points are adjusted to obtain a corrected surface representation.

4. The method for health monitoring of aluminum support structures for curtain walls based on UAV three-dimensional mapping according to claim 1, characterized in that: If the gradient value of the objective function exceeds the preset fitting accuracy control range, the function gradient calculation program is started, the convergence threshold setting technique is used to determine the optimization termination condition, and the parameter update magnitude is controlled by the iteration step size adjustment mechanism. The adaptive iterative optimization solution system includes: The current gradient value is obtained from the gradient threshold check. A precision control range comparison is used. If it is exceeded, the gradient of the activation function is calculated. The gradient vector is calculated by taking the derivative of the objective function and combining it with the current parameter value to obtain the updated gradient direction. For the updated gradient direction, a convergence threshold setting technique is adopted. By comparing the gradient vector magnitude with the preset threshold, the optimization termination condition is determined and the iteration step size adjustment value is obtained. By adjusting the step size during iteration, the parameter update magnitude is controlled. By multiplying by the gradient direction and applying a decay factor, an adaptive iterative optimization solution system is established to obtain the optimized parameter set. From the optimized parameter set, deviation correction is performed on the building facade surface. By adjusting the surface parameter values ​​and recalculating the point coordinates, the corrected surface representation is obtained.

5. The method for health monitoring of aluminum support structures for curtain walls based on UAV three-dimensional mapping according to claim 1, characterized in that: The first round of parameter update calculation is performed according to the adaptive iterative optimization solution system. Error change monitoring technology is used to track the changing trend of the objective function value in real time. The current iteration efficiency is judged by the convergence speed evaluation algorithm. Historical iteration data is recorded, including: The first round of parameter update calculation results are obtained from the adaptive iterative optimization solution system. Error change monitoring technology is used to calculate the difference sequence by continuously sampling the objective function value sequence. The difference sequence is then fitted with a trend to obtain the change trend. To determine the iteration efficiency, the gradient descent algorithm is used to compare the current iteration step size with a preset threshold. The gradient descent algorithm takes the change trend and step size as inputs and outputs an efficiency index.

6. The method for health monitoring of aluminum support structures for curtain walls based on UAV three-dimensional mapping according to claim 5, characterized in that: The process of performing the first round of parameter update calculations based on the adaptive iterative optimization solution system, using error change monitoring technology to track the changing trend of the objective function value in real time, judging the current iteration efficiency through a convergence speed evaluation algorithm, and recording iterative history data also includes: Historical data records are extracted from the iteration efficiency, and the data is written to a preset buffer using a data storage mechanism and a serialization method to obtain the record set; For the record set, subsequent analysis and processing are performed. By comparing historical data with current parameters, a deviation vector is calculated to obtain the optimized adjustment value. The optimized adjustment values ​​are applied to the correction of building facade surface parameters. Deviation correction technology is used to adjust the coordinates of surface points by weighted average to obtain the correction representation.

7. The method for health monitoring of aluminum support structures for curtain walls based on UAV three-dimensional mapping according to claim 1, characterized in that, It also includes performing numerical stability checks based on iterative history data, using local optimum detection techniques to identify whether the system has fallen into a local extreme state, evaluating whether the current optimization state meets the convergence requirements through iteration termination conditions, and outputting the optimal parameter combination if the convergence threshold is met. Specifically, this includes: Numerical sequences are obtained from iterative historical data. A difference sequence is obtained by subtracting adjacent values ​​using a difference calculation method. The stability index is obtained by averaging the absolute values ​​of the difference sequences. The stability index is used to determine whether a preset threshold is met, thus determining the numerical stability state. For numerical stability, gradient analysis is used to identify local extrema by calculating the slope of the numerical sequence, extracting deviation vectors from the local extrema, and summing the deviation vectors to obtain the extrema state identification result.

8. The method for health monitoring of aluminum support structures for curtain walls based on UAV three-dimensional mapping according to claim 7, characterized in that, The process of performing numerical stability checks based on iterative historical data, using local optimum judgment techniques to identify whether the system has fallen into a local extreme state, evaluating whether the current optimization state meets the convergence requirements through iteration termination conditions, and outputting the optimal parameter combination if the convergence threshold is met, also includes: Based on the extreme state identification results, obtain the iteration termination condition parameters, and determine whether the current optimization state has fallen into a local extreme state by comparing the current optimization state with the iteration termination condition parameters. If the assessment shows that the system has not fallen into a local extremum, then the optimal parameter combination is output from the optimization state evaluation.

9. The method for health monitoring of aluminum support structures for curtain walls based on UAV three-dimensional mapping according to claim 1, characterized in that, This also includes reconstructing the optimized facade surface model based on the optimal parameter combination, using point cloud coordinate reprojection technology to convert the optimized parameters into corrected three-dimensional coordinates, calculating the accuracy improvement before and after optimization through geometric deviation comparison analysis, and generating a building facade geometric accuracy optimization report, specifically including: The coordinate transformation matrix is ​​obtained from the optimal parameter combination. The point cloud reprojection method is used to convert the coordinates of each point in the original point cloud data into corrected three-dimensional coordinates through matrix multiplication, thus obtaining the optimized coordinate set. For the optimized coordinate set, an elevation surface model is constructed. The coordinate points are divided into triangular meshes by mesh subdivision technology and interpolated and fused to generate a continuous surface, thus determining the surface geometry. Obtain the corresponding point pairs between the curved surface geometry and the original facade model, and use the Euclidean distance calculation method to compare the deviations between the point pairs to obtain the geometric deviation sequence.

10. The method for health monitoring of aluminum support structures for curtain walls based on UAV three-dimensional mapping according to claim 9, characterized in that, The process of reconstructing the optimized facade surface model based on the optimal parameter combination, converting the optimized parameters into corrected three-dimensional coordinates using point cloud coordinate reprojection technology, calculating the accuracy improvement before and after optimization through geometric deviation comparison analysis, and generating a building facade geometric accuracy optimization report also includes: The mean deviation and standard deviation are extracted from the geometric deviation sequence. The mean deviation is obtained by summing the sequence elements and dividing by the number of elements. The standard deviation is obtained by summing the squared differences of deviations, dividing by the number of elements, and then taking the square root. The difference in accuracy before and after optimization is evaluated by subtraction, and the amount of accuracy improvement is determined. Based on the accuracy improvement, the deviation sequence and surface structure data are integrated to generate a building facade geometric accuracy optimization report, which includes a deviation distribution map and improvement indicators.

Images