Bridge construction support deformation monitoring and early warning method based on unmanned aerial vehicle laser radar

Abstract

This invention relates to the field of bridge deformation monitoring technology, and provides a method for monitoring and early warning of bridge construction support deformation based on UAV lidar. The method includes: identifying the support monitoring object and solidifying the reference period point cloud; setting the flight altitude, flight path spacing, speed, and overlap rate; controlling the UAV equipped with lidar to fly and collect the current period point cloud; performing time alignment, extrinsic parameter calculation, noise reduction, and downsampling on the two period point clouds to obtain a usable point cloud for the support; selecting a reference point set for the pier or hardened ground, registering the rigid body to unify coordinates, and recording the residuals; extracting the centerline and nodes from the cylindrical and linear fitting of the control component point set, and calculating the deformation amount by the normal distance difference based on the cross section; calculating the LoD95 from local noise and registration residuals and gating it; aggregating and generating settlement, lateral displacement, and tilt sequence, and establishing a three-level threshold based on allowable limits or stable period statistics; sequentially verifying and linearly extrapolating the remaining time; generating re-measurement flight paths and handling suggestions according to anomaly weights; and outputting alarm commands.

Description

Technical Field

[0001] This invention relates to the field of bridge deformation monitoring technology, specifically a method, equipment, and medium for monitoring and early warning of bridge construction support deformation based on UAV lidar. Background Technology

[0002] Deformation monitoring of bridge construction supports involves installing sensors and combining them with IoT, laser, or traditional measurement technologies to acquire real-time data on the settlement, displacement, and tilt of the supports. This allows for analysis of deformation patterns, early warning of exceeding limits, and ensures construction safety and bridge quality. Common methods include laser monitoring, sensor monitoring, and the application of IoT data platforms.

[0003] In the monitoring scenario of cast-in-place continuous beam construction across terrain, existing bridge casting stage monitoring of construction supports relies mainly on a small number of measuring points or manual leveling re-measurement, which makes it difficult to continuously acquire the full-area three-dimensional geometric state under conditions of large-scale supports, dense components, and environmental constraints. When early micro-deformations such as uneven settlement, lateral displacement, and local buckling occur, existing solutions cannot distinguish between actual deformation and measurement errors, resulting in difficulty in setting early warning thresholds and insufficient alarm reliability. Therefore, there is an urgent need for a support deformation monitoring and early warning method based on UAV lidar to achieve full-area non-contact point cloud acquisition, stable registration, difference discrimination, and interpretable early warning output. Summary of the Invention

[0004] In view of the existing problems mentioned above, a method, equipment and medium for monitoring and early warning of deformation of bridge construction supports based on UAV lidar are proposed.

[0005] To at least partially solve the above-mentioned technical problems, in a first aspect, the present invention provides a method for monitoring and early warning of bridge construction support deformation based on UAV lidar, comprising:

[0006] Obtain point cloud data for the baseline period and the current period, preprocess and filter to obtain usable point cloud data. The usable point cloud data includes a stable reference point set. Using the stable reference point set as a constraint, unify the point cloud data for the baseline period and the current period to the same coordinate system, and calculate the registration residual.

[0007] Extract the centerline of the control component, determine the local normal of each control component using the centerline, calculate the distance difference along the local normal between the reference period and the current period point cloud data, and obtain the normal deformation.

[0008] A deformation threshold is constructed by combining local noise and registration residual. The normal deformation is gating and judged with the deformation threshold as the boundary. The normal deformation that passes the judgment is aggregated into an index sequence according to the component and the section.

[0009] Based on the preset graded early warning threshold table, a sequential trend test is performed on the indicator sequence to calculate the remaining time to reach the early warning threshold.

[0010] As a preferred embodiment, the preprocessing and screening to obtain usable point clouds includes:

[0011] The three-dimensional point cloud data of the baseline period and the current period are acquired respectively, and the two periods of point cloud data are preprocessed. The preprocessing process includes timestamp sorting and initial coordinate calculation of the point cloud, statistical outlier removal and radius neighborhood filtering to remove isolated noise points, threshold screening of point cloud intensity and echo information to remove low confidence echoes, and voxel downsampling of the point cloud to obtain uniform point density.

[0012] Ground and non-support target segmentation is performed on the two phase point clouds. Ground point sets are extracted and removed based on height layering and plane fitting. The remaining point sets are clustered according to connectivity and geometric shape. The point sets that meet the geometric characteristics of the support components are retained as usable point clouds for the support structure. Stable reference point sets are selected from the multi-phase point clouds and confirmed by spatial range selection and feature consistency verification.

[0013] In a preferred embodiment, the calculation of the registration residual includes:

[0014] Using the stable reference point set as a constraint, a rigid body registration algorithm is used to solve for the rotation matrix and translation vector from the current point cloud to the reference point cloud. The obtained rigid body transformation is then applied to the available point cloud in the current period to unify it to the reference coordinate system. After registration, the registration residual is calculated based on the distance between corresponding point pairs before and after registration using the stable reference point set. The registration residual is then associated with and stored in relation to the current registration result.

[0015] In a preferred embodiment, the distance difference between the calculation baseline period and the current period point cloud data along the local normal includes:

[0016] Based on the available point cloud of the baseline period and the available point cloud of the current period with unified coordinates, the corresponding component point set is extracted from the point cloud according to the component identifier; for each component point set, geometric fitting is first performed to extract the component centerline and node features. When the component is a pole or diagonal brace, a cylindrical model is used to fit the component axis and record the cylinder radius and axis direction vector. When the component is a horizontal connection or the edge line segment of a beam, a straight line model is used to fit the centerline and record the line direction vector.

[0017] During the fitting process, outliers are filtered out using a preset inlier distance threshold, and the intersections of components, the neighborhood of connecting plates, or the points where the fitting residual changes abruptly are selected as candidate node regions. The node feature point set is determined through spatial clustering and geometric constraints. After obtaining the component centerlines for the baseline period and the current period, sampling sections are generated along the component length direction with a preset step size, using the baseline period centerline as a reference. A local coordinate system is constructed at each sampling section based on the centerline direction, and the section plane perpendicular to the centerline is taken as the component local normal reference surface. The corresponding local point set is extracted in the neighborhood of each sampling section in the baseline period point cloud and the current period point cloud, respectively, and the distance difference along the section normal is calculated for the two local point sets. The distance difference is used as the normal deformation at the section position.

[0018] As a preferred embodiment, the local normal includes:

[0019] During the fitting process, outliers are screened out using a preset inlier distance threshold, and candidate regions for nodes are selected. The set of node feature points is determined through spatial clustering and geometric constraints. After obtaining the center lines of the components in the baseline period and the current period, sampling sections are generated along the length of the components with a preset step size, using the center line of the baseline period as a reference. A local coordinate system is constructed at each sampling section based on the center line direction, and the section plane perpendicular to the center line is taken as the local normal reference surface of the component.

[0020] As a preferred embodiment, the gating determination of the normal deformation amount based on the deformation threshold includes:

[0021] Within the local point set, the residual from the point to the fitted model is calculated, and the local noise parameters are statistically obtained. The registration residual of the current point cloud is read as the registration error term. The deformation threshold at the corresponding position is calculated by combining the local noise parameters and the registration error term, as shown in the formula:

[0022] ,

[0023] Where k represents the component index. This represents the 95% confidence level deformation threshold for the k-th component, and 1.96 is the coefficient of a normally distributed system at the 95% confidence level. This represents the standard deviation of local noise in the k-th component during the reference period. This represents the number of valid points participating in the statistics within the k-th component during the baseline period. This represents the standard deviation of local noise in the k-th component during the current period. This represents the number of valid points included in the statistics within the k-th component in the current period. This represents the registration residual term obtained after rigid body registration of the two point clouds under the constraint of a stable reference point set;

[0024] The normal deformation at each cross-section is compared with its deformation threshold. When the normal deformation exceeds the deformation threshold, the position is marked as gated pass, and the gated judgment result is recorded together with the component identifier, cross-section position index and corresponding threshold.

[0025] As a preferred implementation, the process of generating the preset graded early warning threshold table includes:

[0026] Read the allowable deformation limits for settlement, lateral displacement, and tilt angle given in the construction design documents, and write the allowable deformation limits into the highest-level threshold item of the threshold table; when the special plan does not give the allowable deformation limit for a certain indicator, select multi-period indicator samples from the stable period; calculate the mean and standard deviation of the indicator respectively, and generate three-level thresholds of prompt, early warning, and alarm according to a preset multiple relationship, forming a hierarchical early warning threshold table containing indicator type, component identification, threshold level, and threshold value, and store the threshold table in association with the component ledger and period identification.

[0027] The process of performing a sequential trend test on the indicator sequence and calculating the remaining time to reach the warning threshold includes:

[0028] Read the settlement, lateral displacement, and tilt angle index sequences of each component and their corresponding graded early warning threshold tables. For each type of index of each component, construct a sliding analysis window with the period identifier as the time index. Perform sequential trend test on the index sequence within the sliding analysis window, calculate the incremental index sequence within the window and generate cumulative statistics. When the cumulative statistics exceed the preset trigger threshold, record that the component and the index have entered the trend state. After entering the trend state, perform linear fitting on the index sequence within the window to obtain the slope of change, and calculate the remaining time to reach the early warning threshold using the current index value, the slope of change, and the corresponding early warning threshold. Record the remaining time as a lead time and associate it with the component identifier, index type, and threshold level.

[0029] In a second aspect, this application provides an electronic device, including: a processor; and a memory storing program instructions that, when executed by the processor, cause the electronic device to implement one or more embodiments of the first aspect described above.

[0030] In a third aspect, this application provides a storage medium having computer-readable instructions stored thereon, which, when executed by one or more processors, implement one or more embodiments of the first aspect described above.

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] 1. By using rigid body registration constrained by a stable reference point set, coordinates across multiple periods are unified, and the registration residual is output as an error term to participate in subsequent threshold calculations, ensuring that deformation in each period is calculated under the same coordinate datum.

[0033] 2. Perform sequential trend testing and linear fitting on the indicator sequence, calculate the remaining time to reach the warning threshold as the lead time output, and generate the next adaptive retest route based on the weight of the abnormal area to realize the closed-loop process of monitoring, judgment and retesting. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0035] Figure 1 This is a schematic diagram of a method for monitoring and early warning of bridge construction support deformation based on UAV lidar provided in an embodiment of the present invention;

[0036] Figure 2 This is a comparison diagram of the effects of a bridge construction support deformation monitoring and early warning method based on UAV lidar provided in an embodiment of the present invention and the prior art, wherein gray bars represent the prior art and black bars represent the present invention. Detailed Implementation

[0037] To make the technical means, creative features, objectives, and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention.

[0038] To achieve the above objectives, please refer to Figures 1 to 2 This invention provides a method for monitoring and early warning of deformation of bridge construction supports based on UAV lidar, the method comprising the following steps:

[0039] Step S1: Identify the monitoring target and establish a baseline, set the flight path parameters and collect point cloud data;

[0040] Step S2: Preprocess the point cloud and filter available points; perform rigid body registration to unify multi-phase coordinates.

[0041] Step S3: Fit and extract the component center, calculate the normal deformation, and perform gating judgment based on the deformation threshold;

[0042] Step S4: Generate an index sequence based on components and cross-sections, and establish a graded early warning threshold table;

[0043] Step S5: Calculate the remaining time before the warning threshold is reached, generate an adaptive retest route based on the weight of the abnormal area, and provide handling suggestions.

[0044] This method is implemented in the order of S1–S5, and its overall process is as follows:

[0045] Obtain point cloud data for the baseline period and the current period, preprocess and filter to obtain usable point cloud data. The usable point cloud data includes a stable reference point set. Using the stable reference point set as a constraint, unify the point cloud data for the baseline period and the current period to the same coordinate system, and calculate the registration residual.

[0046] Extract the centerline of the control component, determine the local normal of each control component using the centerline, calculate the distance difference along the local normal between the reference period and the current period point cloud data, and obtain the normal deformation.

[0047] A deformation threshold is constructed by combining local noise and registration residual. The normal deformation is gating and judged with the deformation threshold as the boundary. The normal deformation that passes the judgment is aggregated into an index sequence according to the component and the section.

[0048] Based on the preset graded early warning threshold table, a sequential trend test is performed on the indicator sequence to calculate the remaining time to reach the early warning threshold.

[0049] S1. The specific steps for identifying the monitoring target and establishing a baseline, and setting up route parameters for point cloud acquisition are as follows:

[0050] Based on the construction drawings and the special plan for the support, the set of monitoring objects for bridge construction support is determined. The set of monitoring objects includes uprights, transverse connecting components, scissor braces, and key connection nodes, such as the nodes between uprights and horizontal bars, the nodes between uprights and diagonal braces, the nodes where uprights are extended, the nodes where wall ties are attached, and the nodes where top supports are supported by distribution beams, etc. A unique component identifier is assigned to each monitoring object.

[0051] Specifically, the method of assigning a unique component identifier to each monitored object adopts a deterministic generation rule that combines field concatenation set sequence numbers and verification. First, a component ledger table is established, and the support components are divided and classified according to the construction coordinate system. The partition fields include working surface identifier, mileage identifier, and elevation interval identifier, and the classification field includes component type identifier. The component type identifier is obtained by mapping from a preset dictionary, which includes uprights, horizontal bars, scissor braces, wall ties, top supports, distribution beam support points, and node plates. Under the constraints of the same partition field and the same type identifier, a sequence number is generated according to the sorting rule of the component's geometric center point in the construction coordinate system. The sorting rule is to first sort by X coordinate in ascending order, then by Y coordinate in ascending order, and then by Z coordinate in ascending order.

[0052] During the period when the support structure is erected and in a baseline state, a baseline data acquisition task is set. The baseline state refers to the period during which at least two repeated scans are performed and registered within the same acquisition timeframe. The normal difference components of the two scans are calculated for the corresponding cross-sections of each component within the monitored object set, and all normal difference components do not exceed the corresponding deformation threshold. Simultaneously, the registration residual for this period does not exceed a preset upper limit. This period is then defined as the baseline state, and the baseline point cloud is solidified. External parameter calibration parameters, coordinate system definitions, and data recording formats for the UAV and LiDAR are configured. The UAV is controlled to fly within the monitoring area along a preset baseline flight path and simultaneously acquire LiDAR point clouds, forming baseline three-dimensional point cloud data. The preset baseline flight path is defined by dividing the outer boundary of the support monitoring area into strips. The system generates round-trip routes, calculates flight altitude, route spacing, and flight speed based on the target point spacing, and sets a fixed overlap rate for adjacent flight strips and bilateral observation angles to form a set of repeatable route parameters. Specifically, the target point spacing is used as the sampling density requirement. The scanning frequency, field of view, and effective ranging range of the lidar are used to calculate the ground strip width and lateral sampling interval at different flight altitudes. The flight altitude is then selected based on the lateral sampling interval not being greater than the target point spacing. Subsequently, the relationship between the number of points sampled per second in the forward direction and the UAV's forward speed is used to ensure that the longitudinal sampling interval is not greater than the target point spacing to determine the flight speed. Finally, the spacing between adjacent routes is calculated and route parameters are generated based on the strip width corresponding to the selected flight altitude and the preset overlap rate.

[0053] In the current data acquisition task, flight acquisition parameters are set according to the coverage and point density requirements of the baseline acquisition. These parameters include flight altitude, flight path spacing, flight speed, scanning frequency, and field of view. A current flight path consistent with the baseline is generated. The UAV equipped with lidar is controlled to perform automatic flight according to the current flight path. During the flight, lidar ranging data and corresponding pose information are recorded with a unified timestamp. The ranging data and pose information are associated and encapsulated, and the three-dimensional point cloud data of the bridge construction support in the current period is output and associated with the component identifier for storage.

[0054] S2. The specific steps for preprocessing the point cloud and selecting usable points, and for rigid body registration to unify multi-stage coordinates are as follows:

[0055] The three-dimensional point cloud data of the baseline period and the current period are acquired respectively, and the two periods of point cloud data are preprocessed. The preprocessing process includes timestamp sorting and initial coordinate calculation of the point cloud, statistical outlier removal and radius neighborhood filtering to remove isolated noise points, threshold screening of point cloud intensity and echo information to remove low confidence echoes, and voxel downsampling of the point cloud to obtain uniform point density.

[0056] Specifically, the preprocessing process includes the following: both the baseline and current raw point clouds contain the three-dimensional coordinates, intensity, echo number, and timestamp of each point, along with the UAV's pose time sequence. During preprocessing, the point timestamps and pose timestamps are first unified to the same time base and sorted in ascending order. For each point, the corresponding pose is obtained by interpolation in the pose sequence based on its timestamp. Then, the external parameter calibration parameters of the lidar and the airframe, including the arm and mounting angle, are read. The points are sequentially transformed from the radar coordinate system to the airframe coordinate system and then to the construction coordinate system to obtain a point cloud with unified coordinate representation. Points exceeding the effective ranging range are removed during this stage. The effective ranging range is set as a distance threshold interval based on the nominal minimum and maximum ranging range of the lidar used. The radial distance between each point and the radar transmission center is calculated; points whose radial distance falls within this threshold interval are considered valid points. Finally, statistical outlier removal is performed on the point cloud after coordinate calculation. Specifically, for each point, several nearest neighbors are counted. The average distance of the points is used to construct a threshold using the global mean and standard deviation, for example, 20 points. Isolated points with an average distance exceeding the threshold are removed. Then, radius neighborhood filtering is performed. Specifically, the number of neighborhood points for each point is searched with a preset radius, and points with insufficient neighborhood points are removed. The radius is first estimated based on the local point density according to the distance between target points in the point cloud, and then set to 3 times the distance between target points in the point cloud. After noise removal, threshold filtering is performed on the point cloud intensity and echo information. Specifically, upper and lower limits of intensity are set, and low-confidence echo points are removed by combining the quantile threshold of intensity distribution. In this embodiment, the 5th percentile and 95th percentile are selected as the upper and lower limits of the threshold. At the same time, the first echo and the last echo point are retained according to the echo number to reduce floating scattering points. Finally, voxel downsampling is performed on the filtered point cloud. The space is divided into a cubic grid with a side length of 3cm. Only one representative point is retained in each grid. The coordinates of the representative point are taken as the mean of the points in the grid, and the intensity is taken as the median, so as to obtain a preprocessed point cloud with relatively uniform point density.

[0057] Ground and non-support target segmentation is performed on the two-phase point clouds. The segmentation includes extracting and removing ground point sets based on height layering and plane fitting, and clustering the remaining point sets according to connectivity and geometric shape, retaining the point sets that meet the geometric characteristics of the support components as the usable point clouds of the support structure. Subsequently, a stable reference point set is selected from the multi-phase point clouds. The stable reference point set consists of pre-determined stable structure point cloud fragments, such as fixed ground hardening areas, bridge piers, etc., and is confirmed by spatial range selection and feature consistency verification. Using the stable reference point set as a constraint, a rigid body registration algorithm is used to solve the rotation matrix and translation vector from the current phase point cloud to the reference phase point cloud, and the obtained rigid body transformation is applied to the current phase usable point cloud to unify it to the reference phase coordinate system. After registration, the registration residual is calculated based on the distance statistics of corresponding point pairs before and after registration of the stable reference point set, and the registration residual is associated with and stored with the current phase registration result.

[0058] Specifically, the rigid body registration algorithm is used to solve for the rotation matrix and translation vector from the current point cloud to the reference point cloud. A two-stage rigid body registration is employed to unify the coordinates of the current and reference point clouds. First, reference sub-point clouds are selected in both the reference and current point clouds based on pre-labeled stable structures, resulting in a reference reference point set and a current reference point set, respectively. Unified downsampling is then performed, for example, using a voxel side length of 0.05m and a normal estimation neighborhood radius of 0.20m. Subsequently, coarse registration is performed, specifically by extracting multiple feature units with stable geometric structures from the reference and current reference point sets, including planar segments and line-angle features. Random consistency sampling is used to generate candidate pairs between feature units. Candidate pairs satisfy a distance threshold (e.g., 0.50m) and a normal angle threshold (e.g., 15°) to form an inlier set. The initial rotation and initial translation are then calculated from the inlier set to form the initial... Initial rigid body transformation; after obtaining the initial rigid body transformation, fine registration is performed. Fine registration uses an iterative nearest neighbor algorithm to establish the nearest neighbor correspondence between the current reference point set and the baseline reference point set. The nearest neighbor search uses a KD-tree acceleration structure, sets a maximum corresponding distance threshold, and removes corresponding point pairs exceeding the threshold, for example, 0.30m. At the same time, corresponding point pairs are sorted in ascending order of corresponding distance and the bottom 20% are discarded as outlier removal rules. In each iteration, the incremental rigid body transformation is solved based on the retained corresponding point pairs. The solution method uses the least squares method from point to plane to solve rotation and translation. The incremental rigid body transformation is left-multiplied to update the current transformation and applied to the current reference point set. The iteration termination condition is set to reach the maximum number of iterations, for example, 50 times. After the iteration ends, the rotation matrix and translation vector from the current point cloud to the baseline point cloud are output and encapsulated as a homogeneous transformation matrix and stored in association with the period number.

[0059] S3. The specific steps for fitting and extracting the component center, calculating the normal deformation, and performing gating judgment based on the deformation threshold are as follows:

[0060] Based on the available point cloud of the baseline period and the available point cloud of the current period with unified coordinates, the corresponding component point set is extracted from the point cloud according to the component identifier; for each component point set, geometric fitting is first performed to extract the component centerline and node features. When the component is a vertical pole or diagonal brace, a cylindrical model is used to fit the component axis and record the cylinder radius and axis direction vector. When the component is a horizontal connecting rod or the edge line segment of the distribution beam, a straight line model is used to fit the centerline and record the line direction vector.

[0061] Specifically, for each component point set, geometric fitting is first performed to extract the component centerline and node features. First, the component point set is extracted from the unified coordinate point cloud based on the spatial bounding box corresponding to the component identifier. Voxel downsampling is then performed on the extracted point set, with the voxel side length set to 0.02m. The component bounding box is formed by reading the endpoint coordinates of the component in the construction BIM and expanding the component radius. Subsequently, the covariance matrix of the point set is calculated and its eigenvalues ​​are obtained. Specifically, the mean of the three-dimensional coordinates of the component point set is removed, and then pairwise products are multiplied along each dimension and normalized to obtain a 3×3 covariance matrix. Eigenvalue decomposition is performed on this matrix, and the matrix is ​​sorted in descending order of eigenvalues. The principal directions and orthogonal directions of the point set are obtained. Linear components are screened based on the ratio of eigenvalues. Point sets satisfying the condition that the ratio of the second eigenvalue to the first eigenvalue is less than 0.15 and the ratio of the third eigenvalue to the first eigenvalue is less than 0.05 are considered linear point sets suitable for axis fitting. For this linear point set, the principal component directions are used as the initial axis, and robust fitting is performed using random consistency sampling. The number of iterations is set to 1000, with 50 points randomly selected each time to generate candidate models. The orthogonal distance from a point to a candidate model is no greater than 0.05m as the criterion for interior points. Candidate models with no less than 500 interior points enter the refinement stage. When the component... When the component type is a vertical pole or diagonal brace, the refinement stage uses least squares optimization with a cylindrical model. The goal is to minimize the sum of squared residuals from the point to the cylinder's side surface, updating the cylinder's axis direction and the coordinates of a point on the axis to obtain the component's centerline. When the component type is a horizontal connecting rod or a distribution beam edge segment, the refinement stage uses linear total least squares fitting, outputting the straight line direction vector and the coordinates of a point on the line as the centerline. After the centerline is determined, the component's point set is projected onto the centerline direction. The positions corresponding to the minimum and maximum values ​​of the projected scalar are taken as the initial endpoints at both ends of the component, and an end neighborhood is extracted at each endpoint with a radius of 0.20m. Point set; simultaneously, calculate the shortest distance for the centerlines of all components. When the shortest distance between two centerlines is no greater than 0.10m and the projection position of the corresponding shortest distance point falls within the neighborhood of their respective endpoints, the midpoint of the shortest distance points of the two centerlines is determined as the connection node feature point, and the component identifiers involved in the connection are written into the association list of the node; for pole extension nodes, the point density is statistically analyzed along the centerline direction using a sliding window of 0.05m. If there is a continuous window interval with a length of 0.08m whose point density is less than 20% of the average point density of adjacent windows, the center position of the low-density interval is determined as the extension node feature point;

[0062] During the fitting process, outliers are filtered out using a preset inlier distance threshold, such as 0.05m. Intersections of components, the neighborhood of connecting plates, or abrupt changes in the fitting residual are selected as candidate node regions. The node feature point set is determined through spatial clustering and geometric constraints. After obtaining the component centerlines for the baseline and current periods, sampling sections are generated along the component length direction with a preset step size, such as 0.1m, using the baseline centerline as a reference. A local coordinate system is constructed at each sampling section based on the centerline direction, and the section plane perpendicular to the centerline is taken as the component's local normal reference plane. Corresponding local point sets are extracted from the neighborhood of each sampling section in both the baseline and current point clouds. The distance difference along the section normal is calculated for the two local point sets, and this distance difference is used as the normal deformation at that section location. Simultaneously, the residual from the point to the fitting model is calculated within the local point set, and local noise parameters are statistically obtained. The registration residual of the current point cloud is read as the registration error term. The deformation threshold at the corresponding location is calculated by combining the local noise parameters and the registration error term, using the following formula:

[0063] ,

[0064] Where k represents the component index. This represents the 95% confidence level deformation threshold for the k-th component, and 1.96 is the coefficient of a normally distributed system at the 95% confidence level. This represents the standard deviation of local noise in the k-th component during the reference period. This represents the number of valid points participating in the statistics within the k-th component during the baseline period. This represents the standard deviation of local noise in the k-th component during the current period. This represents the number of valid points included in the statistics within the k-th component in the current period. This represents the registration residual term obtained after rigid body registration of the two point clouds under the constraint of a stable reference point set;

[0065] The normal deformation at each cross-section is compared with its deformation threshold. When the normal deformation exceeds the deformation threshold, the position is marked as gated pass, and the gated judgment result is recorded together with the component identifier, cross-section position index and corresponding threshold.

[0066] S4. The specific steps for generating an index sequence based on component and section aggregation and establishing a graded early warning threshold table are as follows:

[0067] The normal deformation amount that passes the gate control judgment is obtained and collected according to the component identifier and cross-section position index; a cross-section sequence is established for the normal deformation amount of each cross-section of the same component, and the cross-section sequence includes cross-section coordinates, normal deformation amount and corresponding time identifier;

[0068] Specifically, the process of establishing a cross-sectional sequence for the normal deformation of each section of the same component involves: first, reading the two endpoints and direction vector of the centerline of each component in the reference period, and defining the centerline as the length axis from endpoint to endpoint; second, generating a set of cross-sectional indices on the length axis starting from the starting point with a preset step size of 0.1m, assigning a unique cross-sectional identifier to each cross-sectional index, and recording the parameter position of the cross-section on the centerline and the corresponding three-dimensional cross-sectional center coordinates; then, extracting the point set within the neighborhood of each cross-section from the point cloud of the reference period and the current period, respectively, using the center of each cross-section as the center and a preset neighborhood radius as the range, and calculating the normal deformation, deformation threshold, and gating judgment result corresponding to the cross-section based on the neighborhood point set; sorting the results of all cross-sections within the same component by cross-sectional index from smallest to largest, and using the component identifier combined with the cross-sectional identifier as the key, writing the cross-sectional center coordinates, cross-sectional index, period identifier, normal deformation, and gating mark into the cross-sectional record table; and then splicing multiple cross-sectional records of the same component in the cross-sectional record table according to the period identifier to form the cross-sectional sequence of the component.

[0069] Based on the local coordinate system of the component centerline, the normal deformation is decomposed into vertical and horizontal components. The vertical component is used as the component settlement, and the horizontal component is statistically analyzed according to the X and Y directions of the construction coordinate system as the component lateral displacement. For each component, the settlement and lateral displacement of multiple sections in the same period are aggregated using the median to form the settlement index and lateral displacement index of the component in that period. The indices of each period are spliced ​​together in chronological order to generate the settlement index sequence and the lateral displacement index sequence. In the process of generating the tilt index, the component centerline direction vectors of the reference period and each period are extracted respectively. The angle between the direction vector of each period and the direction vector of the reference period is calculated as the tilt index of the component, and the tilt index sequence is formed in chronological order.

[0070] During the establishment of the graded early warning threshold table, the allowable deformation limits for settlement, lateral displacement, and tilt angle given in the construction design documents are read, and the allowable deformation limits are written into the highest-level threshold item of the threshold table. When the special plan does not give the allowable deformation limit for a certain indicator, a multi-period indicator sample of the stable period is selected. The stable period is five consecutive monitoring periods after the preloading and support adjustment are completed. When the settlement, lateral displacement, and tilt angle indicators of each component do not reach the prompt threshold within the five periods, the five periods of data are determined as the stable period sample set. The mean and standard deviation of the indicator are calculated respectively, and three levels of thresholds of prompt, early warning, and alarm are generated according to the preset multiple relationship to form a graded early warning threshold table containing indicator type, component identification, threshold level, and threshold value. The threshold table is then stored in association with the component ledger and period identification.

[0071] Specifically, the method of generating three levels of thresholds for prompts, warnings, and alarms according to a preset multiple relationship, and establishing a statistical baseline by selecting a sample of the indicator for a continuous period of m periods during the stable period (e.g., m=6), first calculating the mean value during the stable period. with standard deviation Subsequently, three levels of thresholds—prompt, warning, and alarm—are generated according to a preset multiple relationship, with the prompt threshold set to... +3 The warning threshold is taken as +5 Alarm threshold is set +7 The aforementioned thresholds are then written into the tiered early warning threshold table; for indicators where the value decreases with increasing risk, such as carrying capacity indicators, a symmetrical approach is adopted. -3 , -5 , -7 Write it into the threshold table.

[0072] S5. The specific steps for calculating the remaining time to reach the warning threshold, generating adaptive retest routes based on the weight of abnormal areas, and providing handling suggestions are as follows:

[0073] The settlement, lateral displacement, and tilt angle index sequences of each component and their corresponding graded early warning threshold tables are read. For each type of index of each component, a sliding analysis window is constructed using the period identifier as the time index. The window length of the sliding analysis window is taken as the most recent 5 periods and is updated incrementally according to the period as new data arrives. Within the sliding analysis window, a sequential trend test is performed on the index sequence. The trend test includes calculating the incremental sequence of the index within the window and generating a cumulative statistic. When the cumulative statistic exceeds a preset trigger threshold, the component and the index are recorded as entering a trend state. For example, 5 times the standard deviation of the stable period is used as the preset trigger threshold. After entering the trend state, linear fitting is performed on the index sequence within the window to obtain the slope of change. The remaining time to reach the early warning threshold is calculated using the current index value, the slope of change, and the corresponding early warning threshold. The remaining time is recorded as a lead time and associated with the component identifier, index type, and threshold level.

[0074] Specifically, the linear fitting of the index sequence within the window is performed to obtain the slope of change. Within the sliding analysis window, the time index of each period and the corresponding index value are taken to form a point pair sequence. The least squares linear fitting is used to fit the point pair sequence into a trend line, and the slope of the trend line is used as the slope of change. The time index can be taken in equal increments according to the period order. During the fitting process, all window samples participate in the calculation and output a unique slope parameter.

[0075] The abnormal cross-sections passing through the gate are clustered into an abnormal region set according to their spatial location, and a weight is calculated for each abnormal region. The weight is determined by the number of gate passing points in the abnormal region, the importance level of the involved components, the threshold trigger level, and the point cloud coverage confidence. The next adaptive retest route is generated based on the weight of the abnormal region. According to the component type and indicator type corresponding to the abnormal region, the disposal suggestion entries are retrieved from the pre-set disposal rule library. The disposal suggestion entries are bound with the abnormal region identifier, component identifier, and threshold level and output. The disposal rule library is formed by the structured organization of disposal clauses that can be directly executed in the construction organization design, support special plan, and project safety management system. Before the start of construction, the project technical person in charge encodes the clauses into the rule library according to the mapping relationship of component type, indicator type, threshold level, and suggested action items. At the same time, each rule is configured with a version number, effective scope, and applicable working condition identifier. When the threshold level reaches the preset alarm level, an alarm command containing component identifier, indicator type, threshold level, current indicator value, lead time, abnormal region location, and disposal suggestion is generated and written into the alarm queue for downstream system call.

[0076] Specifically, the adaptive retest route is generated in tiers according to the weight of abnormal areas, with W representing the weight of abnormal areas. When W ≥ 0.8, the route spacing is set to 0.5 times the baseline route spacing and repeated twice; when 0.5 ≤ W < 0.8, the route spacing is set to 0.7 times the baseline route spacing and repeated once; when W < 0.5, the baseline route spacing and number of coverages are used; the overlap rate of the flight strip and the flight altitude of each tier are kept consistent with the baseline route.

[0077] like Figure 2 A comparison of the effects of a bridge construction support deformation monitoring and early warning method based on UAV lidar is presented. To facilitate comparability under different dimensions, the performance of existing technologies in four indicators is uniformly set to 100. The early warning lead time index characterizes the degree to which the present invention outputs usable alarms when signs of overall support instability appear. The false alarm suppression index characterizes the ability of the present invention to reduce false alarms through deformation threshold gating and consistency constraints when point cloud occlusion, sparsity, and registration errors exist. The deformation positioning accuracy index characterizes the spatial positioning and quantification accuracy of key deformations such as component settlement, lateral displacement, and tilt angle. The coverage integrity index characterizes the effective point cloud coverage ratio and usable point extraction integrity in multi-component areas of high supports. The comparison results show that the present invention, under the combined process of unified coordinate registration, component skeleton extraction, normal deformation and detectable threshold gating, and trend verification and adaptive retesting closed loop, outperforms existing technologies in all four dimensions: early warning, false alarm control, positioning accuracy, and coverage integrity.

[0078] Example 2:

[0079] Based on the above embodiment 1, a method for monitoring and early warning of bridge construction support deformation based on UAV lidar in the monitoring scenario of cross-terrain cast-in-place continuous beam construction is as follows:

[0080] Step 1: Based on the specific plan, the support poles, transverse connections, scissor braces, top supports, distribution beam support points, and pole extension nodes along the pier side are identified as the monitoring object set, and component ledgers are established on both working surfaces. Component identification is generated using the rules of working surface, mileage segment, elevation segment, component type, and serial number. The serial number is obtained by arranging the center point of the component within the working surface from small to large along the beam axis and is written into the ledger and associated with the component bounding box range. After preloading is completed and unloaded to the working state, two UAV lidar repeated flights are organized at the same time period. The registration is completed by constructing a stable reference point set on the outer surface of the pier body and the abutment. When the normal difference on each component section of the two scans does not exceed the corresponding deformation threshold and the registration residual does not exceed the upper limit, it is solidified as the baseline point cloud. Subsequently, the baseline flight path is set according to the principle of densification in the valley bottom obstruction area and double-sided observation in the high support area next to the pier. In subsequent periods, the point cloud of the current period is collected along the same flight path and archived and stored according to the component identification.

[0081] Step 2: Read the timestamps, echo numbers, intensity fields, and UAV pose sequences from the two point cloud datasets. Sort the point cloud in ascending order by time and interpolate the pose of each point based on adjacent pose frames. Transform the points from the radar coordinate system to the construction coordinate system using extrinsic parameters, and remove out-of-range points according to the equipment's nominal ranging interval. In the point cloud after coordinate calculation, remove outliers using nearest neighbor statistics and filter isolated points using a fixed radius neighborhood. Then, retain reliable echoes based on intensity quantile thresholds and echo numbers, and uniformly downsample the point cloud using a voxel grid. On the preprocessed point cloud, layer according to elevation and planar simulation... The valley floor points are removed, and the remaining points are clustered according to connectivity. Only the point clusters that meet the geometric characteristics of the slender steel pipe support are retained as the usable point cloud of the support. The stable reference point set is selected from the point cloud fragments of the pier body, the outer surface of the abutment, and the hardened platform on the valley side away from construction disturbance. It is determined by spatial selection and feature consistency verification. With this reference point set as a constraint, the initial pose is obtained by coarse feature registration. Then, the rotation and translation from the current period to the reference period are solved by iterative nearest point fine registration. The registration residual is calculated by the root mean square distance between the corresponding point pairs of the reference point set. The rigid body transformation and residual are stored in the database period by period.

[0082] Step 3: Based on the two available point clouds unified to the reference period coordinate system, extract component point sets for the pier-side uprights, valley-side scissor braces, and top support points according to the bounding boxes in the component ledger, and perform voxel downsampling on each point set; extract the axis using cylindrical fitting for the upright and diagonal brace point sets, and extract the centerline using linear fitting for the transverse connection point sets. During fitting, use 0.05m as the interior point distance threshold and select the model with the largest number of interior points as the initial model in iterative sampling. Then, perform least squares refinement on the initial model to obtain the centerline parameters; based on the shortest distance between the centerline endpoint neighborhood and the adjacent component centerlines, approximately... The system extracts node feature points at the extension positions of the uprights, the connection points between the uprights and horizontal bars, and the supports of the top support and distribution beams, and writes them into the node table. Then, with the center line of the reference period as a reference, sampling sections are generated every 0.10m along the length direction. Local point sets of the two periods are extracted in the neighborhood of each section, and the distance difference along the normal direction of the section is calculated as the normal deformation. At the same time, the local noise of the two periods is obtained by statistically analyzing the residuals from the local points to the fitted model, and the deformation threshold LoD95 is calculated by superimposing the registration residuals of the current period. Sections with normal deformation exceeding LoD95 are marked as gated and associated with the component identifier, section index, and threshold record and stored in the database.

[0083] Step 4: Read the gated cross-section records and aggregate them into a cross-section sequence according to component identifier, cross-section index, and period identifier. The cross-section sequence includes the cross-section center coordinates, normal deformation, and corresponding LoD95 threshold. Using the centerline of the component in the reference period as the length axis, generate cross-section identifiers every 0.10m from the starting point and write them into the cross-section record table. For the same component, arrange them in ascending order by cross-section index and splice them by period. Then, in the component's local coordinate system, decompose the normal deformation into vertical and horizontal components. The vertical component is the settlement of the component, and the horizontal component is decomposed according to the X and Y directions of the construction coordinate system. The lateral displacement was statistically analyzed, and the median aggregation of results from multiple sections within the same period was used to obtain the component-level settlement index and lateral displacement index. At the same time, the angle between the centerline direction vector of the component in the current period and the baseline period was calculated as the component tilt angle index. The three types of indexes were arranged into an index sequence according to the period order. When the threshold table was established, the allowable deformation limit given by the special plan was written first. If no limit was given, the sample of five consecutive periods after preloading and unloading, in which none of the indicators reached the warning threshold, was selected as the stable period. The mean and standard deviation were calculated to generate three levels of thresholds: warning, early warning and alarm. The component-level graded early warning threshold table was formed and entered into the database.

[0084] Step 5: Read the settlement, lateral displacement, and tilt angle index sequences and their threshold tables for the pier-side uprights, valley-side braces, and top support points. For each index, construct a sliding analysis window based on the most recent 5 periods and update it with each new period. Within the window, generate a one-sided cumulative statistic using CUSUM recursion. The trigger threshold is set to 5 times the standard deviation of the stable period. When the cumulative statistic exceeds the trigger threshold, mark the corresponding component index as a trend state. After entering the trend state, perform least-squares linear fitting on the index points within the window, numbered at equal intervals according to periods, to obtain the slope of change. Calculate the remaining time to reach the warning threshold using the current index value and the warning threshold, and record it as a lead time. The gated cross-sections are clustered using DBSCAN to generate abnormal regions, with a cluster radius of 0.30m and a minimum sample size of 10. The weight W of the abnormal regions is calculated based on the number of gated cross-sections, threshold level, and component importance. The next retest route is generated based on W. When W≥0.8, the route spacing is halved and repeated twice. When 0.5≤W<0.8, the route spacing is set to 0.7 times and repeated once. The baseline route is used for the remaining regions. At the same time, the handling suggestion entries that match the component type and indicator type are retrieved from the handling rule base, and alarm commands are generated and queued by binding the abnormal region location and alarm level.

[0085] This embodiment provides an electronic device, which includes a processor and a memory. The memory is communicatively connected to the processor and stores program instructions that can be executed by the processor. The processor is used to execute the program instructions to implement all the steps of the bridge construction support deformation monitoring and early warning method described in the above method embodiment.

[0086] The processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., with data processing capabilities; the general-purpose processor can be a microprocessor or any conventional processor.

[0087] The memory may include various media capable of storing program instructions, such as read-only memory, random access memory, magnetic disk or optical disk; the memory and the processor may be integrated on the same chip, or they may be set separately and connected by bus communication.

[0088] This embodiment provides a medium storing computer-readable instructions, which, when executed by one or more processors, implement all the steps of the bridge construction support deformation monitoring and early warning method described in the above method embodiment.

[0089] The medium may be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk or solid-state drive, etc.; the computer-readable storage medium may also be various combinations of the above-mentioned storage media.

[0090] The embodiments of the present invention described above are subject to modification and change of method by those skilled in the art without departing from the embodiments and broader aspects of the present invention. The appended claims are intended to include all such modifications and changes of method that do not depart from the present invention.

Claims

1. A method for monitoring and early warning of deformation of bridge construction supports based on UAV lidar, characterized in that, include: Obtain point cloud data for the baseline period and the current period, preprocess and filter to obtain usable point cloud data. The usable point cloud data includes a stable reference point set. Using the stable reference point set as a constraint, unify the point cloud data for the baseline period and the current period to the same coordinate system, and calculate the registration residual. Extract the centerline of the control component, determine the local normal of each control component using the centerline, calculate the distance difference along the local normal between the reference period and the current period point cloud data, and obtain the normal deformation. A deformation threshold is constructed by combining local noise and registration residual. The normal deformation is gating and judged with the deformation threshold as the boundary. The normal deformation that passes the judgment is aggregated into an index sequence according to the component and the section. Based on the preset graded early warning threshold table, a sequential trend test is performed on the indicator sequence to calculate the remaining time to reach the early warning threshold.

2. The method for monitoring and early warning of deformation of bridge construction supports according to claim 1, characterized in that, The preprocessing and filtering process yields usable point clouds including: The three-dimensional point cloud data of the baseline period and the current period are acquired respectively, and the two periods of point cloud data are preprocessed. The preprocessing process includes timestamp sorting and initial coordinate calculation of the point cloud, statistical outlier removal and radius neighborhood filtering to remove isolated noise points, threshold screening of point cloud intensity and echo information to remove low confidence echoes, and voxel downsampling of the point cloud to obtain uniform point density. Ground and non-support target segmentation is performed on the two phase point clouds. Ground point sets are extracted and removed based on height layering and plane fitting. The remaining point sets are clustered according to connectivity and geometric shape. The point sets that meet the geometric characteristics of the support components are retained as usable point clouds for the support structure. Stable reference point sets are selected from the multi-phase point clouds and confirmed by spatial range selection and feature consistency verification.

3. The method for monitoring and early warning of deformation of bridge construction supports according to claim 1, characterized in that, The calculation of registration residuals includes: Using the stable reference point set as a constraint, a rigid body registration algorithm is used to solve for the rotation matrix and translation vector from the current point cloud to the reference point cloud. The obtained rigid body transformation is then applied to the available point cloud in the current period to unify it to the reference coordinate system. After registration, the registration residual is calculated based on the distance between corresponding point pairs before and after registration using the stable reference point set. The registration residual is then associated with and stored in relation to the current registration result.

4. The method for monitoring and early warning of deformation of bridge construction supports according to claim 1, characterized in that, The distance difference along the local normal between the baseline period and the current period point cloud data includes: Based on the available point cloud of the baseline period and the available point cloud of the current period with unified coordinates, the corresponding component point set is extracted from the point cloud according to the component identifier; for each component point set, geometric fitting is first performed to extract the component centerline and node features. When the component is a pole or diagonal brace, a cylindrical model is used to fit the component axis and record the cylinder radius and axis direction vector. When the component is a horizontal connection or the edge line segment of a beam, a straight line model is used to fit the centerline and record the line direction vector. During the fitting process, outliers are filtered out using a preset inlier distance threshold, and the intersections of components, the neighborhood of connecting plates, or the points where the fitting residual changes abruptly are selected as candidate node regions. The node feature point set is determined through spatial clustering and geometric constraints. After obtaining the component centerlines for the baseline period and the current period, sampling sections are generated along the component length direction with a preset step size, using the baseline period centerline as a reference. A local coordinate system is constructed at each sampling section based on the centerline direction, and the section plane perpendicular to the centerline is taken as the component local normal reference surface. The corresponding local point set is extracted in the neighborhood of each sampling section in the baseline period point cloud and the current period point cloud, respectively, and the distance difference along the section normal is calculated for the two local point sets. The distance difference is used as the normal deformation at the section position.

5. The method for monitoring and early warning of deformation of bridge construction supports according to claim 4, characterized in that, The local normal includes: During the fitting process, outliers are screened out using a preset inlier distance threshold, and candidate regions for nodes are selected. The set of node feature points is determined through spatial clustering and geometric constraints. After obtaining the center lines of the components in the baseline period and the current period, sampling sections are generated along the length of the components with a preset step size, using the center line of the baseline period as a reference. A local coordinate system is constructed at each sampling section based on the center line direction, and the section plane perpendicular to the center line is taken as the local normal reference surface of the component.

6. The method for monitoring and early warning of deformation of bridge construction supports according to claim 1, characterized in that, The gating determination of the normal deformation amount with the deformation threshold as the boundary includes: Within the local point set, the residual from the point to the fitted model is calculated, and the local noise parameters are statistically obtained. The registration residual of the current point cloud is read as the registration error term. The deformation threshold at the corresponding position is calculated by combining the local noise parameters and the registration error term, as shown in the formula: ， Where k represents the component index. This represents the 95% confidence level deformation threshold for the k-th component, and 1.96 is the coefficient of a normally distributed system at the 95% confidence level. This represents the standard deviation of local noise in the k-th component during the reference period. This represents the number of valid points participating in the statistics within the k-th component during the baseline period. This represents the standard deviation of local noise in the k-th component during the current period. This represents the number of valid points included in the statistics within the k-th component in the current period. This represents the registration residual term obtained after rigid body registration of the two point clouds under the constraint of a stable reference point set; The normal deformation at each cross-section is compared with its deformation threshold. When the normal deformation exceeds the deformation threshold, the position is marked as gated pass, and the gated judgment result is recorded together with the component identifier, cross-section position index and corresponding threshold.

7. The method for monitoring and early warning of deformation of bridge construction supports according to claim 1, characterized in that, The process of generating the preset graded early warning threshold table includes: The normal deformation amount that passes the gate control judgment is obtained and collected according to the component identifier and cross-section position index; a cross-section sequence is established for the normal deformation amount of each cross-section of the same component, and the cross-section sequence includes cross-section coordinates, normal deformation amount and corresponding time identifier; Read the allowable deformation limits for settlement, lateral displacement, and tilt angle given in the construction design documents, and write the allowable deformation limits into the highest-level threshold item of the threshold table; when the special plan does not give the allowable deformation limit for a certain indicator, select multi-period indicator samples from the stable period; calculate the mean and standard deviation of the indicator respectively, and generate three-level thresholds of prompt, early warning, and alarm according to a preset multiple relationship, forming a hierarchical early warning threshold table containing indicator type, component identification, threshold level, and threshold value, and store the threshold table in association with the component ledger and period identification.

8. The method for monitoring and early warning of deformation of bridge construction supports according to claim 1, characterized in that, The process of performing a sequential trend test on the indicator sequence and calculating the remaining time to reach the warning threshold includes: Read the settlement, lateral displacement, and tilt angle index sequences of each component and their corresponding graded early warning threshold tables. For each type of index of each component, construct a sliding analysis window with the period identifier as the time index. Perform sequential trend test on the index sequence within the sliding analysis window, calculate the incremental index sequence within the window and generate cumulative statistics. When the cumulative statistics exceed the preset trigger threshold, record that the component and the index have entered the trend state. After entering the trend state, perform linear fitting on the index sequence within the window to obtain the slope of change, and calculate the remaining time to reach the early warning threshold using the current index value, the slope of change, and the corresponding early warning threshold. Record the remaining time as a lead time and associate it with the component identifier, index type, and threshold level.

9. An electronic device, characterized in that, include: processor; as well as A memory storing program instructions that, when executed by the processor, cause the electronic device to perform the method according to any one of claims 1-8.

10. A storage medium, characterized in that, The medium is a computer-readable storage medium storing computer-readable instructions thereon, which, when executed by one or more processors, implement the method as described in any one of claims 1-8.

Images