A steel box girder truss arch bridge point cloud denoising method based on multi-source feature fusion

Abstract

The present application relates to the technical field of point cloud denoising, and discloses a steel box girder truss arch bridge point cloud denoising method based on multi-source feature fusion, comprising the following steps: step S1, splicing to generate a normalized feature vector; step S2, constructing a candidate point set; step S3, calculating a tangential dissipation feature; step S4, generating noise classification probability and structure semantic probability; step S5, generating intermediate refined coordinates; and step S6, generating a final denoised point cloud. The present application realizes differential correction of point cloud positions by means of multi-source feature splicing and tangential dissipation feature construction, combined with a double-branch network sharing a coding backbone, synchronous output of noise classification probability and structure semantic probability of the point cloud, weighted fusion of three types of coordinates, namely normal backfolding, heterogeneous compression and curvature smoothing, and division of hole rim and weld point sets according to the structure semantic probability, so that the output point cloud data meets the engineering application requirements of bridge geometry verification and reverse modeling.

Description

Technical Field

[0001] This invention relates to the field of point cloud denoising technology, and more specifically, to a point cloud denoising method for steel box girder truss arch bridges based on multi-source feature fusion. Background Technology

[0002] Steel box girder truss arch bridges are characterized by a variety of component types, complex node relationships, a high proportion of thin-walled plates, and mixed connection methods. During long-term service, these bridges often exhibit concentrated weld seams, dense bolt hole distribution, aging surface coatings, and localized rough corrosion. With the increasing demand for digital bridge inspection and reverse modeling, multi-station terrestrial 3D laser scanning combined with near-range supplementary scanning to acquire point clouds of node areas has become an important method for bridge construction and pre-assembly analysis. Since the point cloud data is subsequently used directly for hole location verification, weld contour judgment, and component spatial relationship reconstruction, the reliability of the subsequent analysis results is directly affected by whether the point cloud denoising results retain the true structural details.

[0003] However, point cloud noise in the steel bridge node area is not merely a few isolated flying points. The outer surfaces of steel box girders, node plates, and arch ribs are mostly reflective metallic surfaces, which, under conditions of tilted viewing angles, multi-station acquisition, and local obstruction, easily form an attached layer of pseudo-points near the actual structural surface. Simultaneously, spatter particles and burrs may adhere around welds, and localized corrosion areas can cause subtle normal undulations. These anomalies do not completely detach from the actual structural surface but are often mixed with the inner walls of bore edges, weld toe transition zones, and plate edges. While methods based solely on distance, curvature, or uniform smoothing radius can eliminate some noise, they can also flatten the actual bore edges and welds, thereby compromising key dimensional features. Furthermore, on-site acquisition of long-span bridges typically requires multiple stations to overcome obstruction, resulting in significant density differences between near and far stations, and a multi-scale sampling coexistence between locally supplemented areas and the overall scanned area. This causes the same structural boundary to exhibit different point spacings and observation directions in different areas, further increasing the difficulty of noise identification. Existing technologies lack a holistic approach that can simultaneously utilize geometric features, observation angle features, strength correction features, and semantic recognition results, and continue to apply geometric constraints to protect the hole edges and welds after denoising. Therefore, it is necessary to propose a point cloud denoising method for steel box girder truss arch bridges based on multi-source feature fusion to address the challenge of simultaneously suppressing noise and preserving the structure in complex nodal areas. Summary of the Invention

[0004] This invention provides a point cloud denoising method for steel box girder truss arch bridges based on multi-source feature fusion, which solves the technical problems mentioned in the background.

[0005] This invention provides a point cloud denoising method for steel box girder truss arch bridges based on multi-source feature fusion, comprising the following steps: Step S1: Extract the three-dimensional coordinates, normal vector, geometric curvature, incident angle, correction intensity and neighborhood density of each initial spatial point in the multi-source point cloud set, and concatenate them to generate a normalized feature vector. Step S2: Calculate the average distance from each initial spatial point to its nearest neighbor. Obtain a filtering threshold based on the median and absolute deviation of the distribution of all average distances. Select initial spatial points whose average distance is less than the filtering threshold to generate a candidate point set. Step S3: Calculate the one-sided outward thickness and the near-side supplementary thickness of each retained point along the normal vector in the candidate point set. Based on the difference between the two, combined with the correction intensity and geometric curvature, calculate and generate the tangential dissipation feature. Step S4: The normalized feature vector and the tangential dissipative feature are concatenated to form a composite feature, which is then input into the dual-branch network to calculate and generate the noise classification probability and structural semantic probability for the retained points. Step S5: For the retained points, the normal shrinkage coordinates, anisotropic compression coordinates, and curvature smoothing coordinates are calculated independently. The three coordinates are then weighted and fused using the noise classification probability component values ​​to generate intermediate refined coordinates. Step S6: Extract the hole edge point set and weld point set based on the structural semantic probability division. Project the intermediate fine-tuning coordinates in the hole edge point set onto the cylindrical fitting surface. Perform structural projection on the intermediate fine-tuning coordinates in the weld point set based on the axial direction and the lateral distance. Summarize all coordinates to generate the final denoised point cloud.

[0006] The beneficial effects of this invention are as follows: This invention forms a composite feature system adapted to the point cloud characteristics of steel box girder truss arch bridges through multi-source feature splicing and tangential dissipative feature construction. Combined with a dual-branch network sharing a coding backbone, it simultaneously outputs the noise classification probability and structural semantic probability of the point cloud. Based on the probability components of different noise types, it performs weighted fusion of three types of coordinates: normal contraction, anisotropic compression, and curvature smoothing, to achieve differentiated correction of point cloud positions. Simultaneously, it divides the hole edge and weld seam point sets according to the structural semantic probability, and implements targeted protection for these two types of key structures through geometric projection constraints. This method can adapt to non-uniform density point clouds formed by multi-station acquisition. When dealing with anomalies such as mixed-distribution attached pseudo-points, weld attachments, and corrosion undulations, it can reduce the disturbance to the geometric features of key bridge structures. The output point cloud data can be directly adapted to the engineering application requirements of subsequent bridge geometric verification and reverse modeling. Attached Figure Description

[0007] Figure 1 This is a flowchart illustrating the calculation process of a point cloud denoising method for steel box girder truss arch bridges based on multi-source feature fusion, according to the present invention. Detailed Implementation

[0008] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0009] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" indicate that the element or object preceding the term encompasses the elements or objects listed following the term and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0010] like Figure 1 As shown, a point cloud denoising method for steel box girder truss arch bridge based on multi-source feature fusion includes the following steps: Step S1: Extract the three-dimensional coordinates, normal vector, geometric curvature, incident angle, correction intensity and neighborhood density of each initial spatial point in the multi-source point cloud set, and concatenate them to generate a normalized feature vector. Step S2: Calculate the average distance from each initial spatial point to its nearest neighbor. Obtain a filtering threshold based on the median and absolute deviation of the distribution of all average distances. Select initial spatial points whose average distance is less than the filtering threshold to generate a candidate point set. Step S3: Calculate the one-sided outward thickness and the near-side supplementary thickness of each retained point along the normal vector in the candidate point set. Based on the difference between the two, combined with the correction intensity and geometric curvature, calculate and generate the tangential dissipation feature. Step S4: The normalized feature vector and the tangential dissipative feature are concatenated to form a composite feature, which is then input into the dual-branch network to calculate and generate the noise classification probability and structural semantic probability for the retained points. Step S5: For the retained points, the normal shrinkage coordinates, anisotropic compression coordinates, and curvature smoothing coordinates are calculated independently. The three coordinates are then weighted and fused using the noise classification probability component values ​​to generate intermediate refined coordinates. Step S6: Extract the hole edge point set and weld point set based on the structural semantic probability division. Project the intermediate fine-tuning coordinates in the hole edge point set onto the cylindrical fitting surface. Perform structural projection on the intermediate fine-tuning coordinates in the weld point set based on the axial direction and the lateral distance. Summarize all coordinates to generate the final denoised point cloud.

[0011] In one embodiment of the present invention, the calculation formula for step S1 includes: The formula for calculating the local covariance matrix based on the 3D coordinates of the initial spatial point and the 3D coordinates of its nearest neighbors is as follows: ; in It is a local covariance matrix. The number of nearest neighbors. For the nearest neighbor set, The three-dimensional coordinates of the nearest neighbor points The average center coordinates of the nearest neighbor set. The three-dimensional coordinates of the initial spatial point. This is the matrix transpose symbol; The formula for calculating the normal vector is as follows, based on the eigenvector corresponding to the smallest eigenvalue of the local covariance matrix and the scan view vector: ; in It is the normal vector. For sign determination function, The eigenvector corresponding to the smallest eigenvalue of the local covariance matrix. For the scanning view vector; The formula for calculating geometric curvature based on all eigenvalues ​​of the local covariance matrix is ​​as follows: in For geometric curvature, It is the smallest eigenvalue. The second eigenvalue, It is the largest eigenvalue; The formula for calculating the incident angle based on the normal vector and the scanning view vector is as follows: in The angle of incidence, It is an inverse cosine function. The sign for calculating absolute value or vector magnitude. It is a minimal non-zero positive constant; The formula for calculating the corrected intensity is as follows, based on the original echo intensity, scanning distance, and incident angle: ; in To correct the intensity, The original echo intensity, For scanning distance, For reference distance, It is a cosine function; The formula for calculating the neighborhood density is as follows, based on the three-dimensional coordinates of the initial spatial point and the three-dimensional coordinates of its nearest neighbors: ; in The neighborhood density; The formula for calculating the normalized eigenvector by concatenating the three-dimensional coordinates, normal vector, geometric curvature, incident angle, correction intensity, and neighborhood density is as follows: ; in For normalized eigenvectors, This is a symbol for sequence splicing and combination.

[0012] It should be noted that the multi-source point cloud set is the sum of all point cloud data acquired and uniformly registered by terrestrial 3D laser scanning equipment, short-range handheld scanning equipment, or other compatible point cloud acquisition equipment in the same bridge inspection task. This can be achieved through joint acquisition via multi-station terrestrial 3D laser scanning and local short-range supplementary scanning, followed by inter-station registration to form a unified point cloud set. The initial spatial point's 3D coordinates are the spatial position values ​​of a point to be processed in the point cloud within a unified coordinate system. These coordinates serve as the reference position for local feature calculation and subsequent denoising updates. The spatial coordinates of each point can be directly acquired by a terrestrial 3D laser scanner or a handheld laser scanner and written into the unified coordinate system after registration. The nearest neighbor point's 3D coordinates are the spatial position values ​​of a group of neighboring points participating in local statistical calculations around the current target point. These coordinates are used to construct the covariance matrix, distance statistics, and local geometric description. The coordinates of the corresponding point set can be obtained by performing a nearest neighbor search on the acquired point cloud; the original coordinates themselves are acquired by the scanning equipment. The spatial coordinate difference matrix is ​​an array of differences formed by subtracting the target point's coordinates or the neighborhood center coordinates from the nearest neighbor point coordinates. It describes the discrete structure of the local point group in three coordinate directions. The local covariance matrix is ​​a third-order statistical matrix formed by multiplying and averaging the spatial coordinate difference matrices. It characterizes the diffusion strength of the neighborhood point group in three orthogonal directions. The number of nearest neighbors is the number of neighborhood points participating in the local statistical calculation of the current target point. It controls the stability of normal estimation, curvature calculation, and density estimation, and is preferably set between 20 and 60, meaning this range can balance statistical stability and detail preservation among the steel bridge deck surface, hole edge, and weld seam neighborhoods. The average center coordinate of the nearest neighbor set is the arithmetic mean of the coordinates of all nearest neighbors in the current neighborhood, used to eliminate the influence of coordinate translation on the covariance statistical results. The minimum eigenvalue is the smallest of the three eigenvalues ​​of the local covariance matrix, used to characterize the dispersion of the neighborhood point group along the weakest diffusion direction along the normal. The corresponding eigenvector is a direction vector corresponding one-to-one with the minimum eigenvalue, reflecting the spatial orientation of the minimum dispersion direction of the local point group.

[0013] It should be noted that the scanning origin is the emission center position of the laser scanning equipment acquiring the current point cloud data at this exposure or station position. It is used to construct the scanning direction vector and incident angle, and can be obtained from the station attitude parameters output by the scanning equipment, the results of external control point calculations, or the registered equipment pose record. The scanning direction vector is a spatial direction vector formed by the scanning origin pointing to the target point, used to describe the observation direction of the laser beam reaching the structural surface. The normal vector is a unit direction quantity perpendicular to the local surface, used to distinguish between surface normal displacement and tangential displacement, and to support subsequent retraction, smoothing, and projection calculations. Geometric curvature is a local bending index constructed based on the proportion of eigenvalues ​​of the local covariance matrix, used to represent the strength of geometric undulations of the point cloud surface in the current neighborhood. The incident angle is the numerical value of the angle between the scanning direction vector and the normal vector, used to characterize the degree of tilt of the surface irradiated by the laser beam. The original echo intensity is the uncorrected energy response value received and output by the equipment after the laser beam interacts with the bridge surface, used to characterize the surface reflectivity, and can be directly read from the intensity channel, reflectivity channel, or original point cloud attribute file of the scanning equipment. The actual scanning distance is the actual distance measured between the scanning origin and the target point. It is used to correct for the influence of distance attenuation on the intensity value. It can be directly read from the ranging results of the scanning equipment or calculated from the coordinates of the scanning origin and the target point. The reference distance is the standard distance benchmark used for intensity normalization. It is used to convert intensity values ​​from different stations or under different measurement ranges to a comparable scale. A preferred value is 10 to 20 meters; this range is commonly used as a stable distance range for equipment intensity calibration and engineering comparison in near-mid distance ground scanning of bridges. The corrected intensity is the intensity characterization value obtained after jointly correcting the original echo intensity for distance and incident angle. It is used to improve the comparability between different stations and different observation angles. Neighborhood density is a local sampling density index composed of the reciprocal of the average distance from the target point to all nearest neighbors. It is used to indicate whether the point cloud in the current area is dense. Minimal non-zero positive constants are minimal stable terms introduced during division, inverse cosine, or projective normalization to avoid zero denominators. They are preferably set between 0.000001 and 0.00000001, meaning they must be sufficient to prevent numerical overflow without causing significant perturbations to the true geometric magnitude. Normalized eigenvectors are unified input feature representations formed by concatenating three-dimensional coordinates, normals, curvature, incident angles, correction intensities, and neighborhood densities in a predetermined order.

[0014] It should be noted that the unified coordinate system for the multi-source point cloud set is established as follows: First, at least three stable control points or target markers are set up at the bridge site, and ground station scanning and close-range supplementary scanning are completed respectively; then, a global bridge coordinate system is established by combining target registration with local iterative nearest point fine registration; then, the longitudinal direction of the main bridge is used as the principal axis of the coordinate system, the transverse direction of the bridge deck is used as the secondary axis, and the vertical direction is used as the elevation axis to unify the data sources; finally, point cloud files and station pose files under the unified coordinate system are output for each data source; for example, a Leica total station can be used to establish a control network, a target marker can be used to complete coarse registration, and then the iterative nearest point algorithm can be used to refine the overlap error between the plate surface and the node surface. The selection method and the setting method of the nearest neighbor set are as follows: a spatial index structure is established for each target point under a unified coordinate system, such as a voxel grid index or a 1000-dimensional tree index; then, the nearest 20 to 60 neighbor points are retrieved in a fixed neighbor number mode. If the point cloud of the target area is extremely uneven, a hybrid retrieval using a fixed radius and a fixed neighbor number can be used. For densely populated areas such as hole edges and weld seams, 30 to 40 neighbor points are preferred, and for sparsely spaced plate surfaces, 40 to 60 neighbor points are preferred. If the number of neighbor points is less than the lower limit, the search radius is expanded to 2 to 3 times the average point spacing and the search is repeated.

[0015] It should be noted that the scanning origin and scanning view vector are obtained as follows: for ground scanning data, the coordinates of the station center or the exterior orientation elements exported by the device are directly read; for handheld scanning data, the sampling center positions of each frame after trajectory calculation are read. Then, the station center or sampling center is connected with the target point coordinates to obtain the scanning view vector pointing from the scanning origin to the target point. If the device file only provides the distance measurement and horizontal and vertical angles, the direction vector from the scanning origin to the target point can also be obtained through polar coordinate calculation. For example, if a bridge deck steel box girder node consists of 2 ground stations and 1 handheld supplementary scan, then each point should be bound to the source station number and source pose of the point to ensure that the source of the view vector is clear. The calibration and outlier control methods for the corrected intensity are as follows: First, a test scan is performed using a standard reflector or a uniform steel plate with known on-site reflection characteristics to record the original echo intensity at different distances and angles; then, the intensity is normalized using a correction model based on the reference distance and incident angle to establish an equipment-level correction curve; for outlier control, saturation points, no-echo points, and intensity missing points are first eliminated, and then quantile truncation is performed on the corrected intensity, prioritizing the truncation of values ​​below the 1st percentile and above the 99th percentile to the boundary value; for example, local extreme bright spots caused by paint peeling should be prevented from dominating subsequent optical compensation by quantile truncation.

[0016] Specifically, point cloud denoising for steel bridges does not require a strict principal curvature in the differential geometry of a continuous surface, but rather a stable, fast, and locally responsive bending index that adapts to changes in the neighborhood morphology. Therefore, this invention uses the ratio of the minimum eigenvalue to the sum of all eigenvalues ​​as the geometric curvature. For example, this ratio is small in flat plate areas, increases in weld protrusion areas, and further increases in bolt hole edges and edge transition areas. Thus, this definition is suitable for direct embedding into point cloud feature pipelines. Furthermore, laser echo intensity is affected by both propagation distance and illumination angle. If the original echo intensity is used directly, the same real surface at different stations and tilt angles will exhibit numerical deviations that cannot be directly compared. Therefore, this invention incorporates the square of the distance ratio and the cosine of the incident angle into the intensity correction. For example, when the same painted steel plate is collected from near and far stations, the uncorrected intensity values ​​will be significantly different. After joint correction, the intensity values ​​can more closely approximate the reflectivity of the real surface, making it more suitable for identifying reflective false shells.

[0017] In one embodiment of the present invention, the calculation formula for step S2 includes: The formula for calculating the average distance based on the three-dimensional coordinates of the initial spatial point and the three-dimensional coordinates of its nearest neighbors is as follows: ; in Average distance; The formula for calculating the absolute deviation is as follows, based on the mean distance and the median of the distribution: in For absolute deviation, To find the median operator, The median of the distribution; The formula for calculating the filtering threshold is as follows, based on the median and absolute deviation of the distribution: in The filtering threshold is... For fixed-scale magnification; The formula for calculating the candidate point set based on the average distance and the filtering threshold is as follows: in For the candidate point set, It is a collection of point clouds from multiple sources.

[0018] It should be noted that the average distance is the arithmetic mean of the straight-line distances from the target point to all its nearest neighbors, used as a statistic of dispersion in the coarse filtering stage. The median of the distribution is the middle value of all average distances in the overall sample, used to construct a statistical benchmark with strong resistance to outliers. The deviation is the difference between a single average distance and the median of the distribution, used to describe the discrete offset of that point relative to the center of the population. The absolute deviation is the median statistical result of the absolute values ​​of all deviations, used to characterize the robust fluctuation scale of the average distance distribution. The fixed scale magnification factor is a scale correction coefficient that converts the absolute deviation into a scale that can be directly used in threshold construction, with a preferred value of 1.4826, which is a commonly used correction factor in robust statistics to map the absolute deviation to an approximate standard deviation scale. The filtering threshold is the upper limit of the coarse filtering judgment formed by the superposition of the distribution median and the magnified absolute deviation, used to remove overly discrete points. The candidate point set is a subset of the point cloud retained after coarse filtering, representing the structural candidate data that enters the subsequent feature calculation and refinement stages.

[0019] It should be noted that the calculation domain for the average distance is determined as follows: all initial spatial points in a unified coordinate system are used as the statistical population. For each point, the average Euclidean distance to its nearest neighbor is calculated only within the structural neighborhood of the same data batch. Areas with obvious missing measurements, background objects, and cross-component occlusion are not directly included in the same neighborhood. If a point is located in the bridge boundary area, the average distance is calculated using neighboring points within the same structural plane. For example, points on the inner wall of bolt holes should not be mixed with distant points on the outer plate surface of the holes when calculating the average distance, otherwise the coarse filtering threshold will be raised. The filtering threshold and candidate point set are performed as follows: the average distance of each initial spatial point is compared with the filtering threshold point by point. Points with an average distance greater than the threshold are marked as coarse filtering rejection points, and points with an average distance less than or equal to the threshold are marked as candidate retention points. Then, the retention points are re-summarized according to the original point number order, retaining all attributes required for their coordinates, intensity, source station location, and normal calculation, to generate a candidate point set. For example, in a local scan block of a steel node, background scattered points and isolated flying points will be rejected first, while attached reflective pseudo-layer points will still be retained in the subsequent fine-tuning stage.

[0020] Specifically, the goal of the coarse filtering stage of this invention is to remove only significant environmental outliers, rather than eliminating attached reflective noise in the early stages. Therefore, a candidate point set is formed by globally comparing the average distance of all initial spatial points. For example, although some reflective pseudo-shell points are in abnormal positions, they are still closely attached to the real surface, and their average distance will not increase as drastically as background flying points. Therefore, they will be retained for subsequent fine-tuning stages, which will not be elaborated here.

[0021] In one embodiment of the present invention, the calculation formula for step S3 includes: The formula for calculating the unilateral outward thickness is as follows, based on the 3D coordinates of the nearest neighbor set, the 3D coordinates of the retained points, and the normal vector: in The thickness is unilateral outward. This is the scale layer number. For the nearest neighbor set, Numbering within the nearest neighbor set. This is the spatial decay weighting coefficient. To extract non-negative and positive extreme values, The three-dimensional coordinates within the nearest neighbor set. To preserve the three-dimensional coordinates of the point, It is the normal vector. This is the symbol for spatial dot product; The formula for calculating the proximal augmentation thickness is as follows, based on the 3D coordinates of the nearest neighbor set, the 3D coordinates of the retained points, and the normal vector: in To supplement the thickness on the proximal side; The formula for calculating the tangential dissipation term is as follows, based on the internal normal vector of the nearest neighbor set, the normal vector, and the geometric curvature: in For tangential dissipation terms, It is the identity matrix. The normal vector inside the nearest neighbor set. For a specific constant parameter, Geometric curvature; The calculation formula for generating the tangential dissipation characteristics is as follows, based on the unilateral outward thickness, the near-side supplementary thickness, the incident angle, the corrected intensity, and the tangential dissipation term: in It is a tangential dissipation characteristic. Set the total number of scale levels. For scale weight parameters, The angle of incidence, To correct the intensity, To minimize the constant that prevents it from returning to zero.

[0022] It should be noted that the retained point is the target point currently selected within the candidate point set for calculating tangential dissipation features and subsequent coordinate refinement. The scale neighborhood is a set of local spatial ranges around the retained point, divided according to a preset radius level or a preset nearest neighbor level, preferably set to 3 to 5 levels. Too few multi-scale levels weaken cross-scale stability, while too many increase computational burden and introduce redundant smoothing. The nearest neighbor set is the set of points within a certain scale neighborhood that participate in the calculation of unilateral outflow thickness, near-side supplementary thickness, and tangential dissipation term. The 3D coordinates of the retained point are the spatial positions of the current target point in the candidate point set within a unified coordinate system, used as reference coordinates for scale feature calculation. The spatial coordinate deviation vector is the displacement vector formed by subtracting the retained point coordinates from the nearest neighbor coordinates, used to decompose the normal and tangential components. The spatial attenuation weight coefficient is a weight set based on the spatial distance difference between the nearest neighbor and the retained point, used to ensure that points closer to the nearest point contribute more to local features. The unilateral outflow thickness is the unilateral normal thickening amount formed by weighted averaging of the forward projection components of the nearest neighbor points in the opposite direction of the normal. The proximal augmentation thickness is the proximal normal augmentation amount formed by weighted averaging of the positive projection components of the nearest point in the positive normal direction.

[0023] It should be noted that the normal vector within the nearest neighbor set is the local surface normal estimated by each nearest neighbor in the previous steps, used to describe the consistency of the normal distribution within the neighborhood. The included angle projection deviation difference is a measure of the deviation between the nearest neighbor normal and the preserved point normal after projection onto the tangential plane, used to describe the tangential structural change. The specific constant parameter is an adjustment coefficient for the degree of influence of amplified geometric curvature in the tangential dissipation term, preferably set to 0.1 to 0.5, meaning that this range allows the curvature contribution of the actual hole edge and weld to be preserved, while avoiding the curvature term excessively dominating the dissipation ratio. The tangential dissipation term is a structural change index composed of the average deviation of the normal projection and the curvature compensation term, used to suppress the misleading influence of the actual hole edge and weld on noise judgment. The optical compensation term is a radiation compensation amount composed of the incident angle change and the correction intensity, used to reflect the enhancement effect of reflected noise under observed geometric changes. The thickness subtraction difference calculation value is the normal asymmetric thickening amount obtained by subtracting the near-side supplementary thickness from the unilateral outward thickness. The scale-specific independent dissipation ratio is a single-scale discriminant ratio constructed by using normal asymmetric thickening and optical compensation as the numerator and tangential dissipation as the denominator. Multiple scale weight parameters are the set of weights used when weighting and summing the scale-specific independent dissipation ratios. Preferred values ​​are 0.2, 0.3, and 0.5 at the three-scale level, meaning the near-scale level preserves details, the mid-scale level stabilizes the transition, and the far-scale level suppresses random noise. Therefore, a weight configuration that gradually increases from near to far is preferable. The tangential dissipation feature is a comprehensive discriminant feature formed by weighted summation of multiple scale-specific independent dissipation ratios, used to characterize the relative relationship between the tendency of a unilateral pseudo-shell and the tangential changes of the real structure. The total scale level setting is the total number of scale levels involved in the tangential dissipation feature calculation, preferably set to 3, as three scale levels are usually sufficient to cover local to mid-scale structural changes in the hole edge, weld, and plate transition zone.

[0024] It should be noted that the scale neighborhood is divided as follows: multiple spherical neighborhoods or multiple nearest neighbor neighborhoods are constructed with the retained point as the center, with a three-layer scale preferred; the radius of the first layer is 2 to 3 times the average point spacing, used to capture hole edge and weld toe details; the radius of the second layer is twice that of the first layer, used to express the weld transition zone; the radius of the third layer is twice that of the second layer, used to perceive the pseudo-shell adhesion trend on the plate surface; if a fixed nearest neighbor number mode is used, the number of neighbor points in the three layers can be set to 24, 48, and 96 respectively. The spatial attenuation weight coefficient is constructed as follows: based on the Euclidean distance from the nearest neighbor point to the retained point, an exponential weight or Gaussian weight that decreases monotonically with distance is constructed, preferably using the form of dividing the square of the distance by the square of the scale radius and taking the negative exponent; for extremely close points, a minimum weight lower limit can be added to avoid infinite amplification of a single coincident point; for example, in the first layer scale, the weight of a point 1 mm away from the retained point should be significantly higher than that of a point 3 mm away, thus ensuring that the unilateral outward thickness mainly reflects the near-field adhesion structure.

[0025] It should be noted that the calculation method for the angular projection deviation difference is as follows: first, the difference between the normal of the nearest point and the normal of the retained point is calculated, then the difference is projected onto the tangent plane with the normal of the retained point as the normal, and finally the magnitude or equivalent angle deviation of the projection vector is calculated as the tangential deviation metric. If the normal is not normalized, it should be normalized first. For example, if the nearest normal in a certain neighborhood is only flipped in the opposite direction of the normal but the change in the tangent plane is very small, the projection deviation difference should be kept low to avoid mistakenly treating a single pseudo-layer as a real tangential structure. The method for setting specific constant parameters and multiple different scale weight parameters is as follows: First, perform a mesh search on the labeled reflective pseudo-layer, real hole edge, real weld and ordinary plate surface samples; the specific constant parameter is preferentially tested with a step size of 0.05 between 0.1 and 0.5; multiple different scale weight parameters are preferentially used in combination from near to far, provided that the sum is 1, such as 0.2, 0.3, 0.5; the final parameters are determined by comparing the comprehensive performance of different combinations in pseudo-shell recognition rate and hole edge retention rate. The optical compensation term and the specific scale independent dissipation ratio are realized as follows: First, the angle compensation factor is calculated using the incident angle, and then multiplied by the correction intensity to form the optical compensation term; then, the normal asymmetry is obtained by subtracting the near-side supplementary thickness from the single-sided outward thickness, and added to the optical compensation term as the numerator; then, the tangential dissipation term and the minimum stability term are added as the denominator to complete the single-scale ratio calculation; for example, for the edge region of a strong reflector, if the angle compensation factor is large and the correction intensity is high, the optical compensation term in this region will be significantly improved, thereby amplifying the discrimination score of the reflective pseudo-shell.

[0026] Specifically, a significant characteristic of reflective pseudo-shells is that the point cloud mainly thickens along one side of the normal, while real hole edges and real welds often exhibit more symmetrical or tangential geometric changes. Therefore, this invention decomposes the normal displacement into two parts: unilateral outward thickness and proximal supplementary thickness. For example, if an adhesion point layer appears on the opposite side of a reflective plate, the unilateral outward thickness increases significantly, while the proximal supplementary thickness does not increase accordingly. If it is a real weld bulge, the statistical relationship on both sides and the tangential structure term will change together. Although real hole edges and real welds also show normal differences, their geometric changes are often accompanied by significant tangential structural extension and continuous curvature changes. Therefore, this invention incorporates the normal projection deviation difference and the geometric curvature compensation term into the tangential dissipation term. For example, when the weld extends continuously along the axial direction, the neighborhood normal will show organized changes, and the curvature will maintain a structural distribution. At this time, the tangential dissipation term increases, thereby suppressing the false deletion of real structures in noise discrimination. Furthermore, because the steel bridge node area contains both millimeter-level details and centimeter-level structural transitions, a single-scale feature cannot simultaneously capture hole edge details, weld extension, and plate surface reflective adhesion. Therefore, this invention weighted and fused the independent dissipation ratios of multiple scales. For example, the near-scale can capture welding spatter and hole edge burrs, the mid-scale can express the weld transition zone, and the far-scale can reflect the overall adhesion trend of the plate surface pseudo-shell. The tangential dissipation feature obtained after fusion is more robust.

[0027] In one embodiment of the present invention, the calculation formula for step S4 includes: The formula for calculating the composite feature generated by concatenating and combining the normalized eigenvector and the tangential dissipative feature is as follows: ;in It is a composite feature. For normalized eigenvectors, It is a tangential dissipation characteristic. Symbols for combining and permuting sequences; The formula for calculating the multidimensional shared hidden layer features is as follows, based on the composite feature input encoding backbone extraction component: ;in To share hidden layer features across multiple dimensions, Extract component mapping operators for the encoding backbone; Based on the multidimensional shared hidden layer features, the formula for calculating the noise classification probability for the retained points is as follows: ;in For noise classification probability, It is a nonlinear normalized exponential squeezing function. The noise mapping weight matrix is... This is a noise mapping bias variable; Based on the multidimensional shared hidden layer features, the formula for calculating the structural semantic probability of the retained points is as follows: ;in For structural semantic probability, This is the semantic mapping weight matrix. This is a semantic mapping bias variable.

[0028] It should be noted that composite features are joint input features formed by concatenating normalized feature vectors and tangential dissipative features, possessing geometric, radial, and specific discriminative information. The dual-branch network is a neural network structure that shares the same encoding backbone and is divided into a noise discrimination branch and a structural semantic branch at the back end. The encoding backbone extraction component is the front-end network module responsible for extracting multi-scale local structural representations from the composite features. The multi-scale neighborhood spatial perception aggregation technique is a feature modeling method that simultaneously extracts and fuses local context across different neighborhood ranges to improve adaptability to non-uniformly sampled point clouds. Multi-dimensional shared hidden layer features are high-dimensional latent representations output by the encoding backbone extraction component, used to simultaneously support noise classification and structural semantic recognition. The noise mapping weight matrix is ​​a trainable matrix parameter that maps shared hidden layer features to the noise category space; its values ​​are learned through supervised training. The noise mapping bias variable is a trainable offset parameter used in conjunction with the noise mapping weight matrix to correct the baseline response for different noise categories. The noise classification probability is the output of the dual-branch network, indicating whether the current point belongs to a real structural surface, a reflective pseudo-shell, welding adhesion noise, or rust and roughness noise. The semantic mapping weight matrix is ​​a trainable matrix parameter that maps shared hidden layer features to the structural semantic category space. The semantic mapping bias variable is a trainable offset parameter used to correct prior responses for different structural semantic categories. The structural semantic probability is the network's output probability that the current point belongs to a semantic entity such as a normal plate surface, bolt hole edge, real weld, or structural edge. The nonlinear normalized exponential squeezing function is a normalization function that converts the linear mapping result into a category probability distribution with a sum of 1.

[0029] It should be noted that the dimension normalization and input organization of composite features are as follows: First, zero-mean unit variance normalization or minimum-maximum normalization is performed on the coordinates, curvature, incident angle, correction intensity, neighborhood density, and tangential dissipation features respectively; then, they are concatenated into an input vector of uniform length in a fixed order; if the dimensions of different source devices are significantly different, standardization should be performed once by grouping by device before overall normalization is performed; for example, 3D coordinates, 3D normal, 1D curvature, 1D incident angle, 1D correction intensity, 1D neighborhood density, and 1D tangential dissipation feature can be combined to form a 10-dimensional composite feature. The dual-branch network and encoding backbone extraction component are constructed as follows: the front end adopts a point cloud encoding backbone based on hierarchical neighborhood sampling, and prioritizes the use of a 3-level neighborhood abstraction structure; each level includes neighborhood sampling, local point feature upscaling, and cross-scale aggregation; the back end is divided into a noise classification branch and a structural semantic branch, each branch contains 1 to 2 fully connected mapping layers; for example, a smaller neighborhood can be used in the first layer to preserve hole edge details, and weld seam and plate surface context can be gradually introduced in the second and third layers, and then 4 types of noise probabilities and 4 types of semantic probabilities are output respectively.

[0030] It should be noted that the training sample annotation method for noise classification probability and structural semantic probability is as follows: first, representative point cloud blocks are collected from the node areas of multiple steel box girder truss arch bridges; then, at least two annotators with bridge inspection experience perform double-person cross-annotation; the noise categories include at least real structural surfaces, reflective pseudo-shells, welding adhesion noise, and rust roughness noise; the structural semantic categories include at least ordinary plate surfaces, bolt hole edges, real welds, and structural edges; divergent samples are reviewed by a third party; for example, if a point is geometrically located in the transition zone on both sides of a weld but has spatter particles adhering to its surface, the noise label should be labeled as welding adhesion noise, while the semantic label can still be labeled as the neighborhood of the real weld. The training constraints and loss design of the dual-branch network are as follows: cross-entropy loss is applied to the noise branch and the semantic branch respectively, and class weights are assigned to minority class samples; the total loss is the weighted sum of the losses of the two branches, with priority given to assigning weights similar to those of the noise branch and the semantic branch; during training, batch random sampling, rotation perturbation, slight intensity perturbation, and density perturbation are used to enhance sample robustness; for example, when the number of reflective pseudo-shell samples is less than that of ordinary plate samples, the loss weight of this class should be increased to avoid the network only learning the majority class. The probability output and class usage in the inference stage are as follows: the network simultaneously outputs noise classification probability and structural semantic probability for each retained point; the noise probability is directly used for the four-class coordinate weighted fusion in the refinement stage without hard threshold truncation; the structural semantic probability takes the class corresponding to the highest probability to generate local semantic entity classification labels, while saving the complete probability vector for uncertainty analysis; for example, when the probability of reflective pseudo-shell at a certain point is 0.55 and the probability of welding attachment noise is 0.35, the coordinate fusion should be carried out by both probabilities in the refinement stage, rather than processing it separately according to the largest class.

[0031] Specifically, both noise discrimination and structural semantic recognition rely on the same local geometric basis, but their output targets are different. Establishing two completely independent networks would increase training costs and weaken feature sharing. Therefore, this invention adopts a dual-branch network with a shared coding backbone. For example, the same set of local features can be used to determine whether a point belongs to a reflective pseudo-shell, or to determine whether the point belongs to a hole edge or a weld, thus enabling denoising and protection to be completed collaboratively in the same representation space. In addition, because the sampling density of bridge point clouds is uneven and the local structure has obvious multi-scale characteristics, it is difficult to retain both near-range details and mid-range context if only single-layer neighborhood aggregation is used. Therefore, this invention adopts a multi-scale neighborhood spatial perception aggregation technology. For example, the inner wall of a hole edge needs to rely on a smaller neighborhood to represent the cylindrical boundary, while the transition of node plates and welds needs a larger neighborhood to reflect the overall direction. Multi-scale aggregation can cover both of these needs simultaneously.

[0032] In one embodiment of the present invention, the calculation formula for step S5 includes: The formula for calculating the local spatial support center point based on the structural surface weight values ​​and the three-dimensional coordinates of the nearest points is as follows: in As the center point of local spatial support, These are the structural surface weight values. These are the spatial weighting coefficients. The three-dimensional coordinates of the nearest neighbor points For the set of nearest neighbors; Based on the reflective shell component values, tangential dissipation characteristics, local spatial support center point, and the original three-dimensional coordinates and normal vector of the retained point, the formula for calculating the normal shrinkage coordinates is as follows: in To control the retraction ratio, These are the values ​​for the reflective shell component. It is a tangential dissipation characteristic. For normal-reduced coordinates, To preserve the original three-dimensional coordinates of the point, It is the normal vector; The formula for calculating the normal vector of the cross-section pair, based on the normal vector and the weld axial direction, is as follows: in For the cross-section pair normal, For three-dimensional spatial difference operators, The axial direction of the weld; Based on the spatial difference between the preserved point and its nearest neighbor, and by coordinating and combining the three orthogonal scale attenuation parameters across the three orthogonal scale components, the formula for calculating the local three-dimensional spatial anisotropic smooth attenuation term is as follows: in For local three-dimensional spatial anisotropic smooth attenuation term, It is an exponential function with the natural constant as the base. To preserve the spatial difference between a point and its nearest neighbor, The axial dimension is the attenuation parameter. The attenuation parameter along the subnormal scale, The attenuation parameter is the scale-down parameter along the normal direction; The weighted compression offset adjustment is performed based on the local three-dimensional anisotropic smooth attenuation term. The formula for calculating the anisotropic compression coordinates is as follows: in For heterogeneous compressed coordinates, This is the subnormal compressibility parameter; Based on the spatial straight-line distance and the magnitude of the multi-scale geometric curvature scattering deviation, the formula for calculating the composite smoothing control factor is as follows: in It is a composite smoothing control factor. To smoothly attenuate the influence radius of spatial location, Let the geometric curvature of the nearest neighbor point be . To preserve the geometric curvature of the point, This is a constant representing the range of curvature difference control. Based on the composite smoothing control factor, the corrosion roughness attenuation value, the geometric curvature, and the normal vector, the formula for calculating the curvature smoothing coordinates is as follows: in This is the scaling displacement suppression coefficient. This represents the value of corrosion and roughness attenuation. For curvature smoothing coordinates; The formula for calculating the intermediate refined coordinates is as follows: The three coordinates are weighted and fused using the noise classification probability component values. ; in For intermediate fine-tuning coordinates, This represents the probability value of spatial spatter welding adhesion noise.

[0033] It should be noted that the structural surface weight value is the probability component corresponding to the real structural surface in the noise classification probability, used to characterize the confidence level that a point is more likely to belong to the real plate surface or the real structural support surface. The spatial weighting coefficient is the weight assigned to the nearest neighbor points according to distance or neighborhood structure when calculating the local spatial support center point. The local spatial support center point is the local real structural center position obtained by weighting the structural surface weight and the spatial weight, used as the support reference for the reflection pseudo-layer retraction. The original 3D coordinates corresponding to the retained point are the original spatial positions of the candidate points retained before entering the fine-tuning calculation, used to construct the initial benchmark for regression, compression and smoothing. The difference deviation vector is the spatial difference vector between the original coordinates of the retained point and the local spatial support center point, used to calculate the normal projection regression amount. The projection scalar length value is the projection length of the difference deviation vector in the normal direction, used to quantify the degree of normal deviation of the current point relative to the local real structural surface. The reflection shell component value is the probability component corresponding to the reflection pseudo-shell category in the noise classification probability, used to control the normal retraction intensity. The retraction control ratio is a retraction adjustment coefficient composed of the reflective shell component value and the tangential dissipation characteristic, used to determine the strength of the normal retraction. The spatial single-step retraction amount is the normal single-step retraction displacement formed by multiplying the projected scalar length value by the retraction control ratio. The normal retraction coordinates are candidate corrected coordinates obtained by retracting the retained point along the opposite direction of the normal by one spatial single-step retraction amount.

[0034] It should be noted that the weld axial direction is the principal axis extracted from the main extension trend of the actual weld nearest neighbor set, used to constrain the anisotropic compression of points near the weld. The cross-sectional subnormal direction is the third orthogonal direction obtained by spatial cross product of the normal vector and the weld axial direction, used to represent the lateral direction of the weld cross section. The spatial difference between the retained point and the nearest neighbor is the displacement of the neighboring point relative to the retained point, used to decompose local differences in the weld axial, subnormal, and normal directions. The axial dimension attenuation parameter is a dimension parameter that controls the allowable propagation distance of the weld direction in anisotropic smoothing, preferably 3 to 6 times the average distance between neighboring points, meaning that the weld direction usually needs to maintain continuity, so the axial dimension should be larger than the normal dimension. The subnormal dimension attenuation parameter is a dimension parameter that controls the lateral compression range of the cross section, preferably 1.5 to 3 times the average distance between neighboring points, meaning that the subnormal needs to retain the transverse profile of the weld, but also suppress lateral bulging caused by spatter adhesion. The normal-scale attenuation parameter controls the depth of normal compression and smoothing. It is preferably set to 1 to 2 times the average distance between neighboring points, meaning the normal treatment should be relatively conservative to avoid excessively flattening the actual weld height or hole edge thickness. The local three-dimensional anisotropic smoothing attenuation term is a direction-sensitive weight constructed based on the component differences between the retained point and its nearest neighbor in the axial, subnormal, and normal directions. The subnormal compression coefficient parameter is an adjustment coefficient controlling the lateral compression amplitude of the anisotropic compression coordinates in the cross-section. It is preferably set to 0.2 to 0.6, meaning this range can suppress weld spatter and burrs without damaging the transverse morphology of the weld. The anisotropic compression coordinates are candidate coordinates obtained by directional compression of the retained point in the normal and subnormal directions using the anisotropic smoothing attenuation term. The composite smoothing control factor is a smoothing control quantity composed of spatial distance weights and curvature difference weights, used to jointly constrain rust roughness and smoothing amplitude. The rust roughness attenuation value is the probability component in the noise classification probability corresponding to the surface rust roughness noise category, used to control the curvature smoothing intensity.

[0035] It should be noted that the spatial location smoothing attenuation influence radius is a radius parameter in the composite smoothing control factor used to describe the range of distance influence. A preferred value is 2 to 5 times the average distance between neighboring points. This range covers the locally connected range of rust roughness undulations without easily crossing the boundaries of the actual structure. The curvature difference span control range constant is a limiting constant in the composite smoothing control factor that controls whether points with different curvatures are allowed to smooth each other. A preferred value is 0.01 to 0.1. This means that the curvature difference between the steel plate plane and the weld or hole edge is usually small but separable, making this range suitable for distinguishing roughness undulations from the actual geometric transition. The scaling displacement suppression coefficient is a smoothing suppression amount composed of the rust roughness attenuation value and the geometric curvature, used to limit the normal displacement in high curvature regions. The curvature smoothing coordinates are candidate coordinates obtained by slightly smoothing the retained points along the normal direction according to the composite smoothing control factor. The probability value belonging to spatial spatter welding attachment noise is the probability component of the corresponding welding attachment noise category in the noise classification probability, used to control the proportion of anisotropic compressed coordinates in the fusion. The intermediate refinement coordinates are the refinement result obtained by weighting and fusing the original coordinates, normal shrinkage coordinates, anisotropic compression coordinates, and curvature smoothing coordinates according to different noise probabilities.

[0036] It should be noted that the weighted construction method for the local spatial support center point is as follows: within the neighborhood of the retained point, the nearest neighbor points with a higher probability of being true structural surfaces are first selected based on the structural surface weight values. Then, a Gaussian distance weight is constructed based on the spatial distance, multiplied by the structural surface weight values, and the coordinates of the nearest neighbor points are normalized and weighted summed. If there are insufficient effective structural surface points, the neighborhood is expanded by one layer and recalculated. For example, when there are a large number of reflective pseudo-layer points around a certain point, only points with a high probability of being true structural surfaces will significantly affect the support center position. The extraction method for the weld axial direction is as follows: first, the nearest neighbor points with a probability of being true welds higher than a set threshold are selected from the structural semantic probability, such as above 0.6. Then, principal component analysis is performed on these points, and the eigenvector corresponding to the largest eigenvalue is taken as the weld axial direction. If the number of points is insufficient, samples are supplemented along the neighborhood of the previous scale. For example, the largest principal component direction of a linear weld will stably extend along the weld length direction and can be directly used for anisotropic compression. The parameter setting method for the local three-dimensional spatial anisotropic smoothing attenuation term is as follows: First, statistically analyze the point distribution scale in the axial length direction, cross section direction, and normal direction of the real weld in the training set; then set the attenuation parameters along the axial direction, the attenuation parameters along the sub-normal direction, and the attenuation parameters along the normal direction respectively; prioritize ensuring that the axial scale is the largest, the sub-normal scale is the second largest, and the normal scale is the smallest; for example, it can be initialized at 5 times, 2 times, and 1.5 times the average point spacing, and then fine-tuned according to the weld conformation preservation effect of the validation set.

[0037] It should be noted that the calculation method of the composite smoothing control factor and the scaling displacement suppression coefficient is as follows: First, construct the first attenuation term using spatial distance, then construct the second attenuation term using the difference between the geometric curvature of the nearest point and the geometric curvature of the retained point, and multiply the two to obtain the composite smoothing control factor; then, divide the corrosion roughness attenuation value by 1 and the sum of the geometric curvature of the retained point to obtain the scaling displacement suppression coefficient; for example, even if the probability of corrosion roughness is not low at high curvature hole edge points, over-smoothing will be avoided due to the reduction of the scaling displacement suppression coefficient. The fusion constraint method of the four types of candidate coordinates is as follows: keep the sum of the four types of noise probability components equal to 1, and use this probability as the fusion weight of the original coordinates, normal shrinkage coordinates, anisotropic compression coordinates, and curvature smoothing coordinates; before fusion, it is necessary to ensure that the four types of candidate coordinates are all in a unified coordinate system, and the normal direction of each candidate coordinate is defined consistently; for example, when the probability of the true structural surface is the highest, the intermediate fine-tuning coordinates should be closer to the original coordinates; when the probability of the reflective pseudo-shell increases, the intermediate fine-tuning coordinates should be shifted towards the normal shrinkage coordinates. In addition, when the retained point is located at the hole edge fracture, occluded edge, or extremely sparse region, resulting in the number of effective neighbors being less than the preset lower limit, the neighborhood is first expanded by one layer and the search is repeated; if it is still insufficient, the complex refinement steps are frozen, and only the original coordinates and the light correction of normal shrinkage are retained; if the normal estimation is unstable, the normal result of the previous stable scale is directly inherited; for example, the occluded boundary point of the anchorage area of ​​the suspender often has a missing neighborhood on one side, and it is not advisable to forcibly perform anisotropic compression in this case.

[0038] Specifically, not all points require the same intensity of normal retraction. Simply retracting at a fixed step size can easily flatten the actual weld or hole edge. Therefore, this invention utilizes the reflective shell component value and tangential dissipation characteristics to jointly form the retraction control ratio. For example, for points with a high probability of reflective false shells and weak tangential dissipation, the retraction control ratio will be larger, and the normal retraction will be more obvious. For points with a high probability of actual welds, the retraction control ratio will be smaller to avoid excessive retraction. In addition, since welding adhesion noise usually adheres to the vicinity of the actual weld, it should not be indiscriminately smoothed along the weld direction, but should be mainly compressed in the normal and transverse directions. Therefore, this invention first extracts the axial direction of the weld, and then constructs the cross-sectional sub-normal from the normal and axial directions. For example, when a fillet weld extends along the edge of the component, the weld length direction should be maintained as much as possible, while the bulges of spatter particles in the transverse and normal directions can be directionally compressed.

[0039] Specifically, because the compressibility of noise near the weld differs from that of the actual structure in different directions, the axial direction needs to be preserved more, the subnormal direction needs to be moderately compressed, and the normal direction needs the strongest constraint. Therefore, this invention constructs a local three-dimensional spatial anisotropic smoothing attenuation term, assigning different attenuation scales to the three orthogonal directions. For example, in areas where the weld direction is continuous and transverse burrs are obvious, axial differences are treated more leniently, while transverse and normal differences are attenuated more aggressively. In addition, because corrosion roughness noise manifests as local normal undulations, but the actual weld toe and hole edge may also have high curvature, ordinary smoothing would damage the real details without additional constraints. Therefore, this invention constructs a composite smoothing control factor using spatial distance and curvature difference, and then constructs a scaling displacement suppression coefficient using the corrosion roughness attenuation value and the current curvature. For example, in plate areas with high corrosion probability but not extreme curvature, moderate smoothing can be performed; in high curvature real hole edge areas, even if there is roughness noise nearby, it will be suppressed by excessive smoothing.

[0040] In one embodiment of the present invention, the calculation formula for step S6 includes: Based on the intermediate finishing coordinates within the hole edge point set, a parameter optimization fitting operation is performed to extract the coordinates of the cylinder center reference point, the unit vector of the cylinder's spatial axial direction, and the calculation formulas for the spatial cylinder radius parameters, as follows: ; in The coordinates of the cylinder's center reference point. Let be the unit vector along the axis of the cylindrical space. For the radius dimension parameters of the spatial cylinder, To approximate the analytical minimum extreme value error solution operator, For the collection of points around the hole, A unit-order full-rank square matrix. For intermediate fine-tuning coordinates, The transpose operator; Based on the intermediate finishing coordinates within the weld point set, the calculation formula for extracting the axial direction and the axial center reference positioning point is as follows: ; in In the axial direction, To approximate the analytical maximum extremum distribution solution operator, To attempt to search for the detection direction vector in three dimensions, For weld point set, This serves as the axial center reference positioning point. Based on the three-dimensional coordinates of the retained point, the axial center reference positioning point, and the axial direction, the formula for calculating the lateral distance is as follows: ; in The horizontal distance. To preserve the three-dimensional coordinates of the point; The calculation formula for generating the final denoised point cloud by projecting the intermediate fine-tuning coordinates within the hole edge point set onto the cylindrical fitting surface and performing structural projection on the intermediate fine-tuning coordinates within the weld seam point set, and summing all coordinates, is as follows: in This is the radial distance deviation difference vector. The normal displacement vector of the cross section. To retain the 3D coordinate points after the hard constraint replacement process is completed, For the smallest non-zero constant compensation variable, For the final denoised point cloud, This calculates the total number of extreme values ​​of the three-dimensional discrete data points in the system's output reconstructed summary.

[0041] It should be noted that the local semantic entity classification label is a discrete semantic label generated from the category corresponding to the highest probability in the structural semantic probability, used to distinguish hole edges, welds, and other structural entities. The hole edge point set is a point set formed by aggregating points that are determined by the local semantic entity classification label to be bolt hole edges of assembly nodes. The weld point set is a point set formed by aggregating points that are determined by the local semantic entity classification label to be actual welds. The cylinder center reference point coordinates are the coordinates of the reference point located on the cylinder axis in the hole edge cylinder fitting result, used to define the position of the hole axis. The cylinder spatial axial unit vector is the axial direction unit vector obtained from the hole edge cylinder fitting, used to describe the penetration direction of the bolt hole. The spatial cylinder radius dimension parameter is the hole radius value in the hole edge cylinder fitting result, used to constrain the radial distance after the hole edge point is projected. The cylinder fitting surface is the target projection surface defined by the cylinder center reference point coordinates, the cylinder spatial axial unit vector, and the spatial cylinder radius dimension parameter. The radial distance deviation difference vector is the correction direction vector corresponding to the radial distance error from the hole edge point to the target cylinder surface. The axial direction is the most significant main direction extending from the weld point set in the overall spatial distribution, used to describe the weld orientation. The axial center reference point is the central reference point for the spatial distribution of the weld point set, used to establish the central benchmark for weld projection and lateral distance calculation. The lateral distance is the orthogonal span length of the retained point relative to the weld axial center reference point after removing the axial component, used to maintain the lateral profile dimensions of the weld. The cross-sectional normal displacement vector is the outward displacement vector of the weld point in the cross-sectional plane after removing the axial component, used to perform weld structure projection. The three-dimensional coordinate points retained after the reconstruction of rigid constraint replacement are the final single-point coordinates of the hole edge or weld point after completing the rigid geometric projection. The final denoised point cloud is the final output point cloud result formed by re-integrating the hole edge projection points, weld projection points, and other unchanged points. The total number of extreme values ​​of the sum of the three-dimensional discrete data points calculated and reconstructed by the system output is the total number of points contained in the final denoised point cloud, used to represent the data scale of the final result.

[0042] Specifically, bolt holes in steel plates typically exhibit an approximate cylindrical hole wall structure. Even after denoising, without geometric constraints, the hole edges may still exhibit local undulations, ellipticization, or scattered edges. Therefore, this invention fits a cylindrical surface to the hole edge point set and performs rigid projection. For example, some hole edge points may experience slight bulging due to reflections or spatter; after projection onto the fitted cylindrical surface, they will return to a uniform radius position, thus restoring the geometric consistency of the hole. Since the goal of weld protection is not to flatten the weld but to preserve its lateral profile relative to the axial center, this invention uses the original lateral distance as a constraint, projecting the weld points to the target structural position that maintains this lateral distance. For example, when there is attached noise on the weld surface, projection will suppress disordered burrs but will not eliminate the lateral bulges that the weld body should have. Furthermore, the processing methods for hole edge points, weld points, and ordinary structural surface points differ, and the final output must maintain the complete correspondence and unified coordinate system of all points. Therefore, after rigid projection, this invention uniformly reassembles the hole edge projection points, weld projection points, and the remaining unchanged intermediate fine-tuning coordinates to generate the final denoised point cloud.

[0043] It should be noted that the clustering and organization methods for the hole edge point set and weld seam point set are as follows: First, candidate points for the hole edge and weld seam are screened out based on local semantic entity classification labels; then, continuous point sets are organized using adjacency distance-based clustering methods; the hole edge point set is preferentially clustered according to circumferential connectivity, and the weld seam point set is preferentially clustered according to axial continuity; isolated clusters with too few points are removed; for example, when there are multiple holes in the same node area, each hole edge should be split into an independent point set based on connectivity. The iterative optimization method for the cylinder fitting surface is as follows: First, the initial cylinder axis is estimated based on the principal direction of the hole edge point set, and then the radius is initialized based on the average distance from the hole edge point to the axis; then, the coordinates of the cylinder center reference point, the axial vector, and the radius parameters are optimized using least squares iterative optimization; after each iteration, outliers with a distance residual exceeding 5% to 10% of the current radius are removed until the parameter change is less than 0.0001 or the number of iterations reaches 50; for example, when the bolt hole edge is affected by local burrs, outlier removal can improve the cylinder fitting stability. The method for obtaining the axial direction and axial center reference positioning point is as follows: perform principal component analysis on the weld point set, and take the eigenvector corresponding to the largest eigenvalue as the axial direction; then take the arithmetic mean of all intermediate fine-tuning coordinates of the weld point set to obtain the axial center reference positioning point; if there is obvious bending in the weld, it should be segmented according to the connected regions first, and then the local axial direction and local center point should be obtained separately; for example, when long welds and short welds are mixed together, a global direction should not be used to directly cover all points.

[0044] It should be noted that the execution methods for hole edge projection and weld structure projection are as follows: For hole edge projection, the intermediate fine-tuning coordinates are first decomposed into cylindrical axial and radial components, then the radial component is normalized to the fitted radius while keeping the axial component unchanged. For weld structure projection, the axial component is first removed, the displacement direction of the transverse plane is calculated, and then the weld point is projected to the corresponding length boundary position according to the original transverse distance of the retained points. For example, if the hole edge point bulges outward, it will be pressed back to the target radius, and the surface burrs of the weld point will be projected to the target position that maintains the transverse contour of the weld. The final output organization of the denoised point cloud is as follows: the three-dimensional coordinate points retained after the hard constraint replacement processing after hole edge projection, the corresponding coordinate points after weld projection, and the remaining unchanged intermediate fine-tuning coordinates are recombined and output in the original point index order; the output file must contain at least the final three-dimensional coordinates, source point number, noise category probability, and structural semantic label.

[0045] Specifically, the actual deployment and use of this invention typically begins with the design of the on-site data acquisition scheme. Implementers first determine the areas requiring priority processing based on the bridge structure diagram and existing inspection data, prioritizing areas such as the junction of the arch rib and steel box girder, the connection area between the wind brace and the node plate, the anchorage area of ​​the hangers, areas with dense bolt holes, and areas with concentrated welds. Subsequently, a unified control coordinate system is established at the bridge site, and stable control points and several scanning stations are deployed. For large-scale skeletal areas, ground-based 3D laser scanning equipment is used for multi-station data acquisition to ensure complete coverage of the main bridge outline, node plate positions, and overall component relationships. For detailed areas such as weld surfaces, hole edges, and plate edges, close-range supplementary scanning is used to supplement high-density point clouds. During the acquisition phase, point coordinates, original echo intensity, scanning distance, equipment station information, and acquisition time information should be simultaneously retained to facilitate subsequent view vector calculation, intensity correction, and multi-source registration.

[0046] Specifically, after the data enters the processing system, it first performs coordinate unification, abnormal file checks, and point attribute integrity verification. After verification, a normalized feature vector is constructed for each point, including 3D coordinates, normal vector, geometric curvature, incident angle, correction intensity, and neighborhood density. Then, a robust statistical method is used to perform the first layer of coarse filtering, removing only obviously outlier background scattered points and flying points. Next, the system calculates the unilateral outflow thickness, near-side augmentation thickness, tangential dissipation term, and optical compensation term in the neighborhood at multiple scales, generating a tangential dissipation feature. This feature is concatenated with the normalized feature vector and fed into a dual-branch network, where the system simultaneously outputs the noise classification probability and structural semantic probability for each point. Based on the noise probability, the system generates normal contraction coordinates, anisotropic compression coordinates, and curvature smoothing coordinates, which are then fused with the original coordinates to form intermediate refined coordinates. Based on structural semantic probability, the system divides the points into a set of hole edge points, a set of weld points, and the remaining points. It performs cylindrical constraint projection on the set of hole edge points and axial and lateral distance preservation projection on the set of weld points, finally forming the final denoised point cloud.

[0047] Specifically, the final output is point cloud data with unified coordinates, noise suppression, and preservation of key structural geometry. The output includes not only the final coordinates of each point but also noise category probabilities, structural semantic labels, and source point numbers. For example, in a steel box girder node plate area, the original point cloud might simultaneously contain surface flare caused by reflections, spatter attachments around welds, and localized rough undulations due to rust. After processing by this invention, the pseudo-shell points on the plate surface will recede towards the real structural surface, the irregular attachments around the welds will be directionally compressed, and the sharp undulations caused by rust roughness will be gently reduced. Meanwhile, the bolt hole edges will maintain a contour close to the cylindrical hole wall, and the real welds will continue to extend along the weld direction with a stable lateral contour. For subsequent reverse modeling software, the output point cloud can be directly used for plate fitting, hole position verification, weld size evaluation, and node assembly relationship analysis.

[0048] It should be noted that in engineering applications, the deployment of this invention can be either offline or semi-real-time. Offline processing is suitable for comprehensive review before major repairs, establishment of component digital archives, and pre-assembly analysis before construction. Semi-real-time processing is suitable for rapid screening within nighttime road closure windows; after on-site data collection, the denoising results for key node areas can be output first, allowing technicians to determine whether additional scanning is needed. If used in conjunction with a bridge operation and maintenance platform, the system can also compare the final denoised point cloud with historical detection models to identify hole edge wear, weld shape changes, and local corrosion expansion trends. Since the output results maintain point-level semantics and coordinate correspondence, subsequent model reconstruction, dimension extraction, and defect recording can be directly reused without repeated segmentation, which will not be elaborated upon here.

[0049] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0050] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A method for denoising point clouds of steel box girder truss arch bridges based on multi-source feature fusion, characterized in that, Includes the following steps: Step S1: Extract the three-dimensional coordinates, normal vector, geometric curvature, incident angle, correction intensity and neighborhood density of each initial spatial point in the multi-source point cloud set, and concatenate them to generate a normalized feature vector. Step S2: Calculate the average distance from each initial spatial point to its nearest neighbor. Obtain a filtering threshold based on the median and absolute deviation of the distribution of all average distances. Select initial spatial points whose average distance is less than the filtering threshold to generate a candidate point set. Step S3: Calculate the one-sided outward thickness and the near-side supplementary thickness of each retained point along the normal vector in the candidate point set. Based on the difference between the two, combined with the correction intensity and geometric curvature, calculate and generate the tangential dissipation feature. Step S4: The normalized feature vector and the tangential dissipative feature are concatenated to form a composite feature, which is then input into the dual-branch network to calculate and generate the noise classification probability and structural semantic probability for the retained points. Step S5: For the retained points, the normal shrinkage coordinates, anisotropic compression coordinates, and curvature smoothing coordinates are calculated independently. The three coordinates are then weighted and fused using the noise classification probability component values ​​to generate intermediate refined coordinates. Step S6: Extract the hole edge point set and weld point set based on the structural semantic probability division. Project the intermediate fine-tuning coordinates in the hole edge point set onto the cylindrical fitting surface. Perform structural projection on the intermediate fine-tuning coordinates in the weld point set based on the axial direction and the lateral distance. Summarize all coordinates to generate the final denoised point cloud.

2. The point cloud denoising method for steel box girder truss arch bridges based on multi-source feature fusion according to claim 1, characterized in that, Extract the 3D coordinates of the initial spatial points and the 3D coordinates of the nearest neighbor points within a multi-source point cloud set; The spatial coordinate difference matrix is ​​calculated based on the three-dimensional coordinates of the initial spatial point and the three-dimensional coordinates of the nearest neighbor points. The spatial coordinate difference matrix is ​​multiplied by the transpose of the spatial coordinate difference matrix and the average calculation result is obtained to generate the local covariance matrix. Calculate the minimum eigenvalue and corresponding eigenvector of the local covariance matrix, and extract the scanning gaze vector formed by the line connecting the scanning origin to the initial spatial point; The direction of the eigenvector is reversed by using the dot product of the eigenvector corresponding to the smallest eigenvalue and the scanning view vector to generate the normal vector; Calculate the minimum eigenvalue divided by the sum of all eigenvalues ​​in the local covariance matrix to generate the geometric curvature; The absolute value of the dot product of the normal vector and the scanning view vector is divided by the spatial magnitude of the scanning view vector to obtain the scaling result. The inverse cosine of the scaling result is then calculated to generate the incident angle. Extract the original echo intensity and the actual scanning distance, calculate the square of the ratio of the actual scanning distance to the reference distance set by the system, multiply the squared result by the original echo intensity and divide by the cosine of the incident angle to generate the corrected intensity. Calculate the arithmetic mean of the linear physical distances between the initial spatial point and all its nearest neighbors, extract the reciprocal of the arithmetic mean, and generate the neighborhood density. The three-dimensional coordinates, normal vector, geometric curvature, incident angle, correction intensity, and neighborhood density of the initial spatial points are spliced ​​together in sequence to generate a normalized feature vector.

3. The point cloud denoising method for steel box girder truss arch bridges based on multi-source feature fusion according to claim 1, characterized in that, Obtain the physical distances of the straight lines from each initial spatial point to all nearest neighbor points, calculate the average sum of the physical distances of the straight lines, and generate the average distance. Extract the average distances corresponding to all initial spatial points within the multi-source point cloud set, arrange the numerical values ​​sequentially, extract the value at the center of the sequence, and generate the median of the distribution. Calculate the deviation value by subtracting the median of the distribution from the average distance for each distance, and extract the absolute non-negative value of the deviation value. All absolutely non-negative values ​​are arranged sequentially, and the absolutely non-negative value at the center of the sequence is extracted to generate the absolute deviation. The absolute deviation is multiplied by a fixed scale magnification factor to obtain the magnified calculation result. The magnified calculation result is then added to the median of the distribution to generate the filtering threshold. Compare the average distance corresponding to each initial spatial point with the value of the filtering threshold; Initial spatial points with an average distance greater than the filtering threshold are filtered out, while initial spatial points with an average distance less than or equal to the filtering threshold are retained, merged, and recombined to generate a candidate point set.

4. The point cloud denoising method for steel box girder truss arch bridges based on multi-source feature fusion according to claim 1, characterized in that, Extract the nearest neighbor set contained in the scale neighborhood of the retained point within the candidate point set, and construct the spatial coordinate deviation vector by subtracting the three-dimensional coordinates of the retained point from the three-dimensional coordinates of the nearest neighbor set. The positive component value obtained by calculating the dot product of the spatial coordinate deviation vector and the normal vector in opposite directions is used to calculate the weighted average of the positive component value using the spatial attenuation weighting coefficient, thereby generating the unilateral outward thickness. The positive component value is obtained by calculating the positive direction dot product of the spatial coordinate deviation vector and the normal vector. The positive component value is then weighted and averaged using the spatial attenuation weighting coefficient to generate the near-side supplementary thickness.

5. A point cloud denoising method for steel box girder truss arch bridges based on multi-source feature fusion according to claim 4, characterized in that, Extract the angular projection deviation difference between the normal vector inside the nearest neighbor set and the normal vector of the retained point. Use the spatial attenuation weight coefficient to perform a weighted average operation on the angular projection deviation difference and superimpose the geometric curvature adjusted and amplified by a specific constant parameter to generate a tangential dissipation term. The cotangent difference is obtained by subtracting the cosine of the incident angle from the numerical constant, and then multiplying the cotangent difference by the correction intensity to generate the optical compensation term. The difference in thickness is obtained by subtracting the proximal supplementary thickness from the unilateral outward thickness. The configuration is constructed by adding the optical compensation term to the thickness subtraction difference calculation value as the numerator, and the configuration is constructed by adding the minimum anti-zero constant to the tangential dissipation term as the denominator. The numerator is divided by the denominator to generate the scale-independent dissipation ratio. By using multiple pre-defined weight parameters of different scales, a weighted summation and merging operation is performed on the independent dissipation ratios of specific scales calculated for different distance ranges to generate tangential dissipation features.

6. The point cloud denoising method for steel box girder truss arch bridges based on multi-source feature fusion according to claim 1, characterized in that, Extract the normalized feature vector and tangential dissipative feature corresponding to the retained point, and perform a spatial combination operation on the normalized feature vector and tangential dissipative feature according to the data feature dimension sequence to generate composite features; The composite features are transmitted into the encoding backbone extraction component nested inside the dual-branch network, and multi-dimensional shared hidden layer features are extracted and generated using the multi-scale neighborhood spatial perception aggregation technology. The multidimensional shared hidden layer features are fed into a linear mapping processing path specifically designed for noise identification in the dual-branch network; The mapping processing values ​​are obtained by multiplying the configured noise mapping weight matrix with multidimensional shared hidden layer features and superimposing the corresponding noise mapping bias variables. A nonlinear normalized exponential squeezing function is used to perform classification space probability transformation mapping calculation on the mapped numerical values ​​to generate noise classification probabilities for the retained points; Multidimensional shared hidden layer features are synchronously and in parallel fed into a linear mapping processing path specifically designed for structural semantic discrimination in a dual-branch network; The semantic mapping calculation value is obtained by multiplying the configured semantic mapping weight matrix with multidimensional shared hidden layer features and superimposing the corresponding semantic mapping bias variables. By using a nonlinear normalized exponential squeezing function to perform classification space probability transformation mapping calculations on the semantic mapping numerical values, structural semantic probabilities for the retained points are generated.

7. The point cloud denoising method for steel box girder truss arch bridges based on multi-source feature fusion according to claim 1, characterized in that, Extract the portion of the noise classification probability that belongs to the physical real structure surface as the structure surface weight value, and use the structure surface weight value to perform a spatial weighted summation and average operation on the three-dimensional coordinates of the nearest point to obtain the local spatial support center point. Calculate the projection scalar length of the difference vector between the original three-dimensional coordinates of the retained point and the local spatial support center point in the direction of the normal vector; The portion of the noise classification probability that belongs to the reflective pseudo-shell formed by reflection is extracted as the reflective shell component value. The reflective shell component value is multiplied by the tangential dissipation feature and then divided by the combined value of the number one and the tangential dissipation feature to generate the shrinkage control ratio. Multiply the length of the projected scalar by the shrinkage control ratio to generate a single-step spatial backoff amount. Drive the original 3D coordinates of the retained point to move along the opposite direction of the normal vector. Calculate and generate normal shrinkage coordinates independently. Extract the principal axis vector with the most significant structural spatial extension characteristics within the distribution domain of the nearest point set representing the actual weld type, and generate the weld axial direction; Calculate the three-dimensional spatial product vector of the normal vector and the weld axial direction to generate the cross-sectional sub-normal.

8. The point cloud denoising method for steel box girder truss arch bridges based on multi-source feature fusion according to claim 7, characterized in that, By coordinating and combining the independent projection components of the spatial difference between the retained point and the nearest neighbor point in the weld axial direction, cross-sectional sub-normal direction, and normal vector, and combining three orthogonal scale attenuation parameters, a local three-dimensional spatial anisotropic smooth attenuation term is constructed. In the direction of the normal vector and the cross-sectional subnormal vector, the spatial difference between the retained point and the nearest neighbor point is adjusted by weighted inward compression translation according to the local three-dimensional space anisotropic smooth attenuation term, and anisotropic compression coordinates are generated independently. A composite smoothing control factor is constructed by extracting the spatial straight-line distance between the retained point and the associated points in the surrounding area, as well as the amplitude of the multi-scale geometric curvature distribution deviation. The portion of noise classification probability that belongs to surface rust and roughness noise is extracted as the rust and roughness attenuation value, and the scaling displacement suppression coefficient is calculated by combining the geometric curvature of points in space. Using a composite smoothing control factor, the spatial difference between the retained point and the nearest neighbor point is projected onto the normal vector to perform an overall weighted summation calculation. The overall weighted summation calculation result is multiplied by the scaling displacement suppression coefficient and superimposed onto the original three-dimensional coordinates corresponding to the retained point to perform slight curvature undulation reduction smoothing. Curvature smooth coordinates are generated independently. Using the structural surface weight values, reflective shell component values, probability values ​​of spatial spatter welding adhesion noise, and rust roughness attenuation values, the three-dimensional coordinates of the multiplication retention points, normal shrinkage coordinates, anisotropic compression coordinates, and curvature smoothing coordinates are independently matched, the results of all independent multiplications are directly summarized, added, and merged to calculate and generate intermediate refinement coordinates.

9. A point cloud denoising method for steel box girder truss arch bridges based on multi-source feature fusion according to claim 1, characterized in that, Extract the position of the maximum probability value of the record within the structural semantic probability to represent the local semantic entity classification label to which it belongs; The local semantic entity classification label is used to identify the point group set of the assembly node bolt hole edge, which is then extracted, clustered and merged to generate the hole edge point set. The local semantic entity classification label is used to identify the point group set that refers to the actual weld of the structural component. The point group set is extracted, clustered and merged to generate the weld point set. Perform a geometric projection parameter analysis optimization iterative fitting operation on the intermediate fine-tuning coordinates contained in the hole edge point set, calculate and extract the cylinder center reference point coordinates, cylinder spatial axial unit vector and spatial cylinder radius size parameters obtained by fitting, and collaboratively construct the cylinder fitting surface; Calculate the radial distance deviation vector from the intermediate finishing coordinates within the hole edge point set to the coordinates of the cylinder center reference point in a unit vector plane perpendicular to the cylinder's spatial axial direction; The intermediate fine-tuning coordinates within the drive hole edge point set are truncated and translated along the radial distance deviation difference vector direction, and the spatial scaling is moved to a fixed position in space where the truncation distance is exactly equal to the radius dimension parameter of the spatial cylinder. The intermediate fine-tuning coordinates within the hole edge point set are then projected onto the cylinder fitting surface.

10. A point cloud denoising method for steel box girder truss arch bridges based on multi-source feature fusion according to claim 9, characterized in that, For the intermediate finishing coordinates contained within the weld point set, calculate and extract the spatial distribution extreme value extension feature direction to generate the axial direction, and calculate the spatial distribution mean focus of the intermediate finishing coordinate cluster to generate the axial center reference positioning point. The orthogonal span length of the vertical cross-section after stripping away and eliminating axial direction component interference is obtained by extracting the three-dimensional coordinates of the retained point to the axial center reference positioning point, and generating the lateral distance. Calculate the normal displacement vector of the cross section after peeling off the axial direction component interference, connecting the intermediate fine-tuning coordinates within the weld point set to the axial center reference positioning point. The intermediate finishing coordinates within the drive weld point set are stretched, scaled, and moved outward along the direction of the cross-sectional normal displacement vector to a fixed position in space where the length boundary is equal to the lateral distance. Based on the axial direction and the lateral distance, a precise structural projection is completed. The coordinates of the hole edge point set after rigid projection constraint processing, the coordinates of the weld point set after structural dimension projection reset, and the remaining scattered and independent intermediate fine-tuning coordinates that remain unchanged are directly recombined and seamlessly integrated to generate the final denoised point cloud by summing all the coordinates.

Images