A method, device and medium for quantitative evaluation of a jacket damage based on registration and fusion of data of different scales

Abstract

The application discloses a kind of based on the registration and fusion of heteroscale data, method, device and medium of quantitative evaluation of jacket damage, comprising: acquisition jacket monocular vision image sequence and laser radar data, respectively three-dimensional reconstruction obtains scaleless vision dense point cloud and has absolute scale laser point cloud;Three-dimensional feature is extracted and is registered by scale and pose decoupling strategy heteroscale, and high-precision datum point cloud model before construction is constructed;According to the same procedure, obtain the point cloud to be measured in construction, and the weight of large residual point pair in local deformation area is obtained by dynamic inhibition, and difference point cloud is obtained;Surface distance is calculated using a specific algorithm, and initial abnormal point cloud is extracted according to threshold value;Independent damage instance is separated by adaptive density clustering combined with spatial position and deformation depth constraint, and the surface area and maximum physical recess depth of each instance are quantified.The application realizes the complementary advantages of cross-scale data, solves the registration distortion problem of traditional method, and provides high-precision damage positioning and quantitative evaluation means.

Description

Technical Field

[0001] This invention relates to the field of quantitative assessment technology for catheter stent damage, and in particular to a method, device and medium for quantitative assessment of catheter stent damage based on heteroscale data registration and fusion. Background Technology

[0002] As a core foundation structure of offshore platforms, jacket structures are constantly exposed to the complex marine environment, including wind, waves, and currents, making them highly susceptible to collision damage. This directly threatens the safety of marine engineering construction and the long-term stability of the structure. Traditional damage detection relies on divers' exploration or visual inspection by underwater robots. These methods are heavily constrained by the marine environment, posing high operational risks and making it difficult to accurately quantify the extent of damage. With the increasing size and depth of marine engineering equipment, the complexity of jacket structures continues to rise, creating an increasingly urgent need for timely, high-precision positioning and quantitative assessment of damage detection. Therefore, a highly efficient damage assessment technology adapted to the complex marine environment is urgently needed to provide reliable technical support for construction safety management and structural maintenance.

[0003] Existing technologies suffer from two major drawbacks: First, they lack adaptability to multi-source data fusion. Traditional vision and lidar fusion methods rely on rigid hardware connections and offline calibration, lacking the ability to flexibly register data at different scales. They struggle to balance the dense texture advantages of visual data with the absolute scale characteristics of lidar, and are prone to feature matching failures in weakly textured marine environments. Second, they have weak resistance to deformation interference. When the jacket structure experiences local collision deformation, conventional registration algorithms are susceptible to interference from local anomalies, leading to global model misalignment. This prevents accurate spatial alignment between the undamaged main structure and the structure under test, resulting in damage identification errors. Furthermore, they are difficult to effectively separate multiple concurrent damage instances, failing to meet the high-precision quantitative assessment requirements of engineering projects. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of the existing technology, the present invention provides a method, device and medium for quantitative assessment of duct stent damage based on heteroscale data registration and fusion.

[0005] The technical solution adopted in this invention is a quantitative assessment method for jacket damage based on heteroscale data registration and fusion, comprising the following steps: S1, acquiring monocular visual image sequences and lidar data of offshore wind turbine jackets respectively; S2, performing three-dimensional reconstruction on the monocular visual image sequences and lidar data respectively to form a scale-free visual dense point cloud and a lidar point cloud with absolute scale; S3, performing three-dimensional feature extraction on the visual dense point cloud and the lidar point cloud, performing heteroscale registration through an optimization strategy of decoupling scale factor and pose variable, and fusing to construct a high-precision benchmark point cloud model before construction that has both dense texture and absolute scale; S4, during construction, according to S2... -S3 The reconstruction and fusion process acquires the current state of the test point cloud data. A robust registration strategy against deformation interference is adopted to align the test point cloud data with the reference point cloud model in space. In the iteration, the point pair residual is calculated and the registration weight of large residual point pairs in local deformation areas is dynamically suppressed to obtain spatially aligned difference point cloud data. S5 Based on the difference point cloud data, the surface distance is calculated using the M3C2 algorithm, and the initial abnormal point cloud is extracted according to the distance threshold. S6 For the initial abnormal point cloud, an adaptive density clustering algorithm based on the joint constraints of spatial location and deformation depth is used to separate multiple independent damage instance point clouds, and the surface area and maximum physical indentation depth of each damage instance are quantified respectively.

[0006] Furthermore, in S2, when performing 3D reconstruction of monocular vision image sequences, a water surface specular mask is constructed. The formula is: ,in, This represents the absolute brightness value of a pixel in a grayscale image. The average brightness of the local spatial neighborhood background. To achieve high light sensitivity, a high-frequency wind-induced adaptive filter is introduced into the original acceleration sequence during lidar data reconstruction to smooth the acceleration. The calculation formula is: ,in These are the current raw acceleration observations. The smooth acceleration from the previous moment. It is a dynamic smoothing factor, and The high-frequency variance of acceleration within the sliding time window. is the wind vibration damping constant.

[0007] Furthermore, in S3, heteroscale registration extracts topological nodes using Gaussian curvature. The calculation formula is: ,in, For point The two principal curvatures of the locally quadratic fitted surface are used; a spatial side-length topology graph is constructed for consistency verification, with scale-invariant topology error. The formula is: ,in, For the three-dimensional coordinates of the visual point cloud topology nodes, The three-dimensional coordinates of the topological nodes corresponding to the laser point cloud. It represents the Euclidean distance in space.

[0008] Furthermore, the deformation-resistant robust registration strategy in S4 introduces a deformation-aware adaptive truncation kernel function to allocate registration weights. The formula is:

[0009]

[0010] in, For the residual of the point pair in the k-th iteration, The environmental noise threshold. To determine the cutoff threshold for confirmed deformation, The dynamic scaling factor is used; the weighted error objective function is: ; S5 also includes dynamic calculation of the detection limit threshold. The formula is: ,in These are the local surface roughness of the baseline point cloud and the difference point cloud, respectively. To correspond to the number of valid point clouds in the neighborhood, This refers to global spatial alignment error; S6 also includes a joint clustering distance measure. The formula is: ,in It is a three-dimensional Euclidean distance. The normal distance of M3C2 This is the deformation depth penalty coefficient; S6 also includes adaptive clustering neighborhood radius. The calculation formula is: ,in It is the spatial connectivity expansion factor. The number of nearest neighbors. The coordinates of the nearest neighbor points, This is the depth gradient tolerance coefficient. For point The corresponding detection limit threshold; the surface area of ​​the damaged instance is obtained by accumulating the area of ​​the triangular mesh through three-dimensional triangular mesh reconstruction, and the maximum physical indentation depth is determined by extracting the maximum absolute value by traversing the M3C2 distance.

[0011] Further, S3 includes the following sub-steps: S31, acquiring the visual dense point cloud and laser point cloud generated in S2, respectively serving as the source point cloud and target point cloud, and clarifying the differences in data source attributes and scale characteristics between the two types of point clouds; S32, calculating the local principal curvature of the two types of point clouds, selecting saddle surface feature regions based on Gaussian curvature to construct a set of topological nodes, and extracting three-dimensional common feature descriptors in the local spatial neighborhood of the nodes through the SpinNet network; S33, establishing an initial correspondence based on the feature descriptors, constructing a spatial side length topology graph, and eliminating mismatched point pairs through scale-invariant topology error verification, and inputting the pure correspondence into the TEASER++ algorithm to solve the global scale factor, rotation matrix, and translation vector; S34, based on coarse registration, using the SymmetricICP algorithm to introduce bidirectional surface normal constraints to perform fine registration, eliminating local density asymmetric residuals, and forming a high-precision benchmark point cloud model before construction.

[0012] Further, S4 includes the following sub-steps: S41, during construction, visual image sequences and lidar data are collected simultaneously, strictly following the independent reconstruction process of S2 and the heteroscale fusion process of S3, to generate current state test point cloud data consistent with the scale of the reference point cloud model; S42, extracting regions in the test point cloud and the reference point cloud model that meet the saddle surface feature conditions, and constructing K-type, T-type, and Y-type topological node sets as undamaged stiffness extreme value regions; S43, using the SpinNet network to extract three-dimensional feature descriptors within the topological node sets and establishing initial correspondences, inputting the TEASER++ algorithm to eliminate outliers and mismatches through truncated least squares estimation, and solving the global rotation matrix and translation vector at the same scale; S44, using the point-to-surface distance as the error objective function, introducing a deformation-aware adaptive truncated kernel function to identify local deformation regions, and iteratively minimizing the weighted error objective function to complete the accurate spatial alignment of the main structure and obtain the difference point cloud data.

[0013] Further, S5 includes the following sub-steps: S51, acquiring the spatial alignment difference point cloud data output by S4, defining a cylindrical search neighborhood with each point on the surface of the reference point cloud model as the core, along the surface normal vector direction, and clarifying the spatial range and geometric parameters of the neighborhood; S52, using the M3C2 algorithm to calculate the mean projection distance of the difference point cloud in the cylindrical search neighborhood, obtaining the M3C2 distance corresponding to each point; S53, combining the local surface roughness, point cloud density, and global spatial alignment error of the reference point cloud and the difference point cloud, dynamically calculating the 95% confidence level detection limit threshold corresponding to each point; S54, filtering out the point clouds with an absolute value of M3C2 distance greater than the corresponding detection limit threshold to form an initial abnormal point cloud, clarifying the spatial distribution range of the abnormal point cloud.

[0014] A device for quantitative assessment of jacket structure damage based on heterogeneous data registration and fusion is disclosed. This device is applied to a method for quantitative assessment of jacket structure damage based on heterogeneous data registration and fusion. It includes: a multi-source heterogeneous data acquisition and transmission module for independently acquiring monocular visual image sequences and LiDAR data of offshore wind turbine jacket structures, performing lossless transmission and temporary storage of the two types of raw data through a data transmission interface; a dual-branch independent 3D reconstruction module connected to the multi-source heterogeneous data acquisition and transmission module, performing feature matching, camera pose calculation, and depth estimation operations on the received monocular visual image sequences to generate visually dense point clouds, and performing high-frequency vibration filtering, tightly coupled pose calculation, and global map update operations on the LiDAR data to generate LiDAR point clouds; and a heterogeneous point cloud registration and fusion module connected to the dual-branch independent 3D reconstruction module, extracting the 3D topological features of the two types of point clouds and decoupling them through scale factors and pose variables. The strategy executes heteroscale registration and fuses to generate a high-precision benchmark point cloud model before construction. The deformation-resistant robust registration and alignment module connects to the heteroscale point cloud registration and fusion module and the dual-branch independent 3D reconstruction module, receiving the point cloud to be measured during construction and the benchmark point cloud model. It then uses a dynamic weight adjustment strategy to accurately align the main structure and output the difference point cloud. The anomaly point cloud extraction and damage separation module connects to the deformation-resistant robust registration and alignment module. It uses the M3C2 algorithm to calculate surface distances and combines them with dynamic thresholds to extract initial anomaly point clouds. It then uses an adaptive density clustering algorithm with joint constraints of spatial location and deformation depth to separate independent damage instances. The damage parameter quantitative calculation module connects to the anomaly point cloud extraction and damage separation module. It performs triangular mesh reconstruction on the point clouds of each independent damage instance to calculate the surface area, traverses the M3C2 distances to extract the maximum physical indentation depth, and outputs a complete quantitative damage assessment result.

[0015] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a method for quantitative assessment of duct stent damage based on heteroscale data registration and fusion.

[0016] A computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for quantitative assessment of duct stent damage based on heteroscale data registration and fusion.

[0017] Beneficial Effects: This invention proposes a method, device, and medium for quantitative assessment of duct stent damage based on heteroscale data registration and fusion. At the method level, multi-source data is collected stepwise and 3D reconstruction is performed separately. A heteroscale registration strategy, decoupling scale factors and pose variables, overcomes the limitations of rigid hardware connections and offline calibration, achieving complementary advantages between dense texture visual data and the absolute scale of LiDAR, thus solving the problem of feature matching failure in weak texture environments. A robust registration strategy resistant to deformation interference is adopted, dynamically suppressing the weights of large residual points in local deformation regions during iteration, avoiding global model misalignment, and achieving accurate alignment between the undamaged main structure and the structure under test. Anomaly point clouds are extracted through dynamic thresholding, and an adaptive density clustering algorithm, constrained by spatial location and deformation depth, accurately separates multiple concurrent damage instances, enabling quantitative calculation of damage surface area and maximum indentation depth. At the device level, six functional modules are sequentially connected and work collaboratively, forming a complete technical chain from data acquisition, independent reconstruction, heteroscale fusion to registration and alignment, damage separation, and quantitative calculation, significantly improving the accuracy and efficiency of damage detection. The overall technical solution avoids the high risk and low accuracy problems of traditional detection methods, while making up for the shortcomings of existing technologies in multi-source data adaptability and anti-deformation interference capability, providing reliable technical support for damage assessment of offshore wind turbine jackets. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a flowchart of method step S3 of the present invention; Figure 3 This is a flowchart of method step S4 of the present invention; Figure 4 This is a flowchart of step S5 of the method of the present invention; Figure 5 This is a diagram showing the modular composition of the device of the present invention; Figure 6 This is a schematic diagram of the computer device of the present invention. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] like Figure 1As shown, a quantitative assessment method for jacket damage based on heteroscale data registration and fusion includes the following steps: S1, acquiring monocular visual image sequences and lidar data of offshore wind turbine jackets respectively; S2, performing 3D reconstruction on the monocular visual image sequences and lidar data respectively to form a scale-free visual dense point cloud and a lidar point cloud with absolute scale; S3, extracting 3D features from the visual dense point cloud and the lidar point cloud, performing heteroscale registration through an optimization strategy that decouples scale factors and pose variables, and fusing to construct a high-precision benchmark point cloud model with both dense texture and absolute scale before construction; S4, during construction, according to the steps in S2-S3... The process involves establishing and fusing the current state of the point cloud data to be tested. A robust registration strategy resistant to deformation interference is used to spatially align the point cloud data to be tested with the reference point cloud model. In the iteration, the point pair residuals are calculated and the registration weights of large residual point pairs in local deformation areas are dynamically suppressed to obtain spatially aligned difference point cloud data. In step S5, based on the difference point cloud data, the surface distance is calculated using the M3C2 algorithm, and the initial abnormal point cloud is extracted according to the distance threshold. In step S6, for the initial abnormal point cloud, an adaptive density clustering algorithm based on the joint constraints of spatial location and deformation depth is used to separate multiple independent damage instance point clouds, and the surface area and maximum physical indentation depth of each damage instance are quantified respectively.

[0021] Step S1 completes the accurate acquisition of multi-source heterogeneous data on the offshore wind turbine jacket, providing high-quality raw data support for subsequent 3D reconstruction and damage assessment. During implementation, a drone equipped with a high-resolution monocular camera and a LiDAR sensor was used as the data acquisition platform. This eliminates the need for rigid hardware connections or offline calibration of the two sensors, allowing for independent, flexible, and efficient data acquisition. The monocular camera must have a pixel resolution of at least 20 million pixels. During shooting, it takes 360-degree, multi-angle shots around the jacket along a preset flight path, with the overlap between adjacent images controlled at 60% to 70%, ensuring coverage of all surface areas of the jacket and forming a complete monocular visual image sequence. The images are captured using a lossless compression format to preserve original texture information. The lidar sensor used is a pulsed lidar with a ranging accuracy better than two centimeters. The scanning frequency is set to 100,000 to 300,000 points per second, and the scanning angle range is 360 degrees horizontally and -15 degrees to 90 degrees vertically. During the UAV's flight, it continuously acquires three-dimensional spatial information of the guide frame surface, simultaneously recording laser echo intensity data and raw data from the inertial measurement unit. Extreme wind and wave weather must be avoided during the acquisition process to ensure the stability of the sensor's operating environment. The acquired lidar data is stored in a standard point cloud format for easy subsequent data processing. This step, through optimizing the acquisition parameters and flight path design, ensures that the acquired monocular visual image sequence has clear texture and uniform lidar data density, laying the foundation for the accuracy of subsequent independent 3D reconstruction.

[0022] Step S2 involves independently reconstructing 3D from the acquired monocular vision image sequence and LiDAR data, generating two types of point cloud data with complementary characteristics. For the monocular vision image sequence, a water surface specular mask is first constructed to eliminate interference from strong light reflections and splashing water. The specular sensitivity coefficient is set to a range of 1.5 to 2.5. A two-dimensional Gaussian smoothing filter is used to calculate the average background brightness in the local spatial neighborhood. Pixels exceeding the set multiple of the average background brightness are marked as specular regions and masked, retaining only the jacket structure area for feature matching. Subsequently, the LoFTR algorithm is used for detectorless dense feature matching, combined with the LO-RANSAC algorithm to remove mismatched points, resulting in clean feature matching data. Based on this data, the incremental SFM algorithm is used to calculate the camera pose, and then the MVSformer++ network is introduced for depth estimation and dynamic geometric consistency verification. Finally, a scale-free visual dense point cloud is generated, with a point density of 50 to 100 points per square centimeter, which can accurately reproduce the surface texture details of the jacket structure. For LiDAR data, a high-frequency wind vibration adaptive filter is first introduced to smooth the raw acceleration sequence input by the inertial measurement unit. The sliding time window is set to 50 to 100 milliseconds, and the smoothing factor is dynamically adjusted according to the real-time wind vibration intensity. The value of the wind vibration damping constant is determined based on the jacket stiffness and the background noise level of the IMU sensor. Then, the preprocessed inertial data and the raw LiDAR data are input into the FAST-LIO2 algorithm. Tightly coupled pose calculation and mapping are performed through the error state Kalman filter front end. An incremental kd-tree is used to maintain the global map and generate a LiDAR point cloud with absolute scale. The absolute scale error of the point cloud is controlled within 1 cm, providing an accurate scale reference for subsequent heteroscale fusion.

[0023] Step S3 achieves high-precision registration and fusion of point clouds at different scales, constructing a high-precision benchmark point cloud model for pre-construction that combines dense texture and absolute scale. First, 3D features are extracted from both the visually dense point cloud and the laser-generated point cloud. The local principal curvature of each point in both types of point clouds is calculated. Point cloud regions satisfying the saddle surface feature conditions are selected using Gaussian curvature. These regions correspond to K-type, T-type, or Y-type pipe nodes of the jacket, constructing a topological node set. This set effectively avoids registration interference caused by repetitive textures on the cylindrical surface of the jacket. Subsequently, within the local spatial neighborhood of the topological node set, rotation-invariant 3D common feature descriptors are extracted using the SpinNet network. Initial correspondences are established based on the Euclidean distance of the feature descriptors, constructing a spatial side-length topological graph. Consistency is verified by calculating scale-invariant topological errors, setting the topological consistency tolerance threshold to 0.05 to 0.1. Mismatched point pairs caused by axial slippage due to repetitive textures on the cylindrical surface are removed, retaining pure correspondences. The pure correspondence is input into the TEASER++ algorithm, and the global scale factor, rotation matrix, and translation vector are calculated using truncated least squares estimation to complete the coarse registration at different scales. The root mean square error of the coarse registration is controlled within 5 cm. Based on the coarse registration, the SymmetricICP algorithm is used to introduce bidirectional surface normal constraints for fine registration. The number of iterations is set to 50 to 100. Each iteration updates the pose parameters and calculates the projection error from the bidirectional points to the tangent plane until the error converges. Finally, the local density asymmetry residuals are eliminated, generating a high-precision reference point cloud model before construction. This model integrates the dense texture of the visual point cloud and the absolute scale of the laser point cloud, with an overall registration error of less than 2 cm, providing an accurate reference for subsequent damage detection.

[0024] Step S4 involves acquiring the test point cloud and performing robust registration against deformation during construction to achieve precise spatial alignment with the reference point cloud model. First, at key nodes in the jacket construction (such as lowering, hoisting, and base placement), following the independent reconstruction process of step S2 and the heterogeneous scale fusion process of step S3, monocular visual image sequences and LiDAR data are simultaneously collected on-site to generate the test point cloud data in its current state. This ensures that the test point cloud and the reference point cloud are consistent in data format and scale characteristics, making them comparable. Subsequently, pipe node regions satisfying the saddle surface characteristic conditions are extracted from the test point cloud and the reference point cloud model. K-type, T-type, and Y-type topological node sets are constructed as undamaged stiffness extrema regions. These regions have high stiffness and are not easily deformed, providing stable feature support for registration. The SpinNet network is used to extract 3D feature descriptors within the topological node set and establish initial correspondences. These initial correspondences are then input into the TEASER++ algorithm. Outlier mismatches caused by collisions are eliminated through truncated least squares estimation. The global rotation matrix and translation vector at the same scale are calculated to complete the initial pose alignment, with the initial alignment error controlled within 3 to 5 cm. In the fine registration stage, the distance from a point to a surface is used as the error objective function. A deformation-aware adaptive truncation kernel function is introduced, setting the environmental noise threshold to 0.5 to 1 cm and the confirmed deformation truncation threshold to 3 to 5 cm. The dynamic scale factor is calculated using the estimated absolute median difference of all residuals in the current iteration. In each iteration, registration weights are dynamically allocated based on the point-pair residuals to suppress the influence of large residual point pairs in local deformation regions. During the iteration process, joint iteration termination conditions are set, including the weighted error change rate of two consecutive iterations being less than 0.001, the magnitude of the incremental translation vector being less than 0.1 cm and the incremental rotation angle being less than 0.1 degrees, or the number of iterations reaching the maximum safe number of iterations of 200. By iteratively minimizing the weighted error objective function, the main structure of the stripped local deformation region is accurately spatially aligned, and spatially aligned difference point cloud data is obtained. This data can accurately reflect the geometric difference between the point cloud to be measured and the reference point cloud, providing a direct basis for damage identification.

[0025] Step S5 calculates surface distances based on the difference point cloud data and extracts initial anomalous point clouds using dynamic thresholds to accurately locate potential damage areas. First, spatially aligned difference point cloud data output from step S4 is acquired. Using a high-precision pre-construction benchmark point cloud model as a reference, the local surface normal vector is calculated for each core point on the benchmark point cloud surface. The radius of the cylindrical search neighborhood is set according to the density of the benchmark point cloud, typically 1 to 3 cm, and the neighborhood height is set to 2 to 5 cm. A cylindrical search neighborhood extending along the normal vector direction is constructed. This neighborhood accurately defines the local contrast range and reduces interference from irrelevant areas. Subsequently, the M3C2 algorithm is used to calculate the mean projection distance of the difference point cloud within the cylindrical search neighborhood. This algorithm calculates the local geometric centroids of the benchmark and difference point cloud subsets within the neighborhood, and uses the projection length of the vector connecting the two centroids onto the cylindrical axis (i.e., the benchmark point normal vector) as the M3C2 distance, effectively reflecting the surface normal differences between the two point clouds. By combining the local surface roughness of the baseline point cloud and the difference point cloud within the cylindrical search neighborhood, the effective number of point clouds, and the global spatial alignment error output by the deformation-resistant robust registration module, the detection limit threshold at a 95% confidence level for each point is dynamically calculated. This threshold can adapt to the point cloud quality and registration accuracy of different regions, avoiding missed or false detections caused by a fixed threshold. Finally, the point cloud set whose absolute value of the M3C2 normal distance is greater than the detection limit threshold at the corresponding position is extracted as the initial anomalous point cloud. By setting a minimum threshold for the number of anomalous points (usually 50 to 100 points), false anomalous regions composed of a small number of noise points are eliminated, ensuring that the initial anomalous point cloud can accurately cover potential damage areas, laying the foundation for subsequent damage instance separation.

[0026] Step S6 achieves accurate separation and quantitative calculation of multiple independent damage instances, outputting key damage parameters. First, for the initial anomalous point cloud extracted in Step S5, an adaptive density clustering algorithm based on joint constraints of spatial location and deformation depth is employed to construct a joint clustering distance measure. This measure considers the difference between the Euclidean distance and the M3C2 normal distance between two points in 3D space, setting the deformation depth penalty coefficient to 0.3 to 0.7 to balance the constraint weights of spatial connectivity and deformation depth continuity, ensuring that the clustering results accurately reflect the physical boundaries of the damage. Based on the local spatial resolution and depth noise tolerance of the initial anomalous point cloud, an adaptive clustering neighborhood radius is dynamically constructed for each core point. This radius is jointly determined by the average spatial distance from the core point to its k nearest neighbors (k ranges from 10 to 20), the spatial connectivity expansion factor (ranges from 1.2 to 1.5), the depth gradient tolerance coefficient (ranges from 0.5 to 0.8), and the corresponding detection limit threshold, enabling adaptation to changes in point cloud density and damage depth in different regions. The initial anomalous point cloud is traversed, and based on the connectivity condition that the joint clustering distance metric is less than the adaptive clustering neighborhood radius of the corresponding core point, the initial anomalous point cloud is segmented into multiple independent damage instance point clouds. Each instance corresponds to an independent collision damage, effectively avoiding confusion between multiple concurrent damages. For each independent damage instance sub-point cloud, a Poisson surface reconstruction algorithm is used for 3D triangular mesh reconstruction. The maximum side length of the triangular mesh is set to 0.5 to 1 cm. The local surface area of ​​the damage instance is calculated by accumulating the areas of all triangular mesh faces. Simultaneously, the M3C2 distance of all points within the sub-point cloud is traversed, and the distance with the largest absolute value is extracted as the maximum physical indentation depth of the damage instance. The above calculation steps are repeated to traverse all damage instance sub-point clouds in turn, outputting the precise location, local surface area, and maximum physical indentation depth of each independent damage instance. The local surface areas of all instances are then globally accumulated to obtain the total surface area of ​​the jacket collision damage, providing accurate quantitative data support for construction safety assessment and structural maintenance.

[0027] Preferably, in S2, when performing 3D reconstruction of the monocular vision image sequence, a water surface specular mask is constructed. The formula is: ,in, This represents the absolute brightness value of a pixel in a grayscale image. The average brightness of the local spatial neighborhood background. To achieve high light sensitivity, a high-frequency wind-induced adaptive filter is introduced into the original acceleration sequence during lidar data reconstruction to smooth the acceleration. The calculation formula is: ,in These are the current raw acceleration observations. The smooth acceleration from the previous moment. It is a dynamic smoothing factor, and The high-frequency variance of acceleration within the sliding time window. is the wind vibration damping constant.

[0028] Specifically, the core requirement of independent 3D reconstruction of two types of data in S2 is to construct a dedicated formula to overcome interference from the marine environment. The water surface specular mask formula is based on the characteristics of strong light reflection and water splash at sea. By comparing the difference between the absolute brightness of pixels and the average brightness of the local background, specular interference areas are screened and shielded. The average brightness of the local background is calculated using a two-dimensional Gaussian smoothing filter to ensure the stability of the neighborhood brightness. The specular sensitivity coefficient is set to a range of 1.5 to 2.5, a value determined through multiple experiments based on the significant difference in reflectivity between the steel pipe surface and the sea surface water splash, accurately distinguishing specular noise from the jacket structure's solid area. The high-frequency wind vibration adaptive filtering formula combines the high-frequency vibration characteristics of hardware caused by marine wind loads. It uses an exponential smoothing model to dynamically adjust the smoothing factor. The high-frequency variance of acceleration within the sliding time window directly represents the real-time wind vibration intensity. The wind vibration damping constant is adapted to the jacket stiffness and the background noise level of the IMU sensor, ensuring that the filter can adapt to different wind vibration intensities. During implementation, a water surface highlight mask is first superimposed onto the image sequence to constrain the feature matching space. A visually dense point cloud is generated through feature matching and geometric verification. For the lidar data, the original IMU acceleration sequence is first preprocessed by high-frequency wind vibration adaptive filtering, and then input into relevant algorithms to generate a laser point cloud. The two types of formulas specifically address the highlight interference in visual reconstruction and the vibration interference in laser reconstruction, respectively, to ensure the accuracy of independent reconstruction.

[0029] Preferably, in S3, heteroscale registration extracts topological nodes using Gaussian curvature. The calculation formula is: ,in, For point The two principal curvatures of the locally quadratic fitted surface are used; a spatial side length topology graph is constructed for consistency verification, while the scale remains unchanged and the topology error is not considered. The formula is: ,in, For the three-dimensional coordinates of the visual point cloud topology nodes, The three-dimensional coordinates of the topological nodes corresponding to the laser point cloud. It represents the Euclidean distance in space.

[0030] Specifically, the core requirement of S3 heteroscale registration is to achieve accurate registration. The Gaussian curvature formula is based on the intrinsic relationship between principal curvature and Gaussian curvature in differential geometry. By calculating the product of the two principal curvatures of the local quadratic fitted surface of the point cloud, saddle-shaped feature regions are selected. These regions correspond to the pipe nodes of the guide frame, possessing stable topological properties and avoiding interference from repetitive textures on cylindrical surfaces. The scale-invariant topology error formula is based on the principle of invariance of spatial side length ratios. By comparing the difference in side length ratios of the triangles formed by the topological nodes of the visual point cloud and the laser point cloud, the consistency of feature matching is verified. The topology consistency tolerance threshold is set to 0.05 to 0.1, a value verified through extensive experiments, which effectively eliminates mismatched point pairs due to axial slippage. In implementation, the Gaussian curvature of the two types of point clouds is first calculated to extract the topology node set. Then, the SpinNet network is used to extract feature descriptors to establish an initial correspondence. The scale-invariant topology error formula is used to verify and retain the pure correspondence. This is then input into the TEASER++ algorithm to solve for the scale factor, rotation matrix, and translation vector, completing the coarse registration. The two types of formulas respectively solve the problems of the specificity of feature extraction and the effectiveness of matching verification in heteroscale registration, laying the foundation for subsequent fine registration and ultimately achieving high-precision heteroscale fusion of visual point clouds and laser point clouds.

[0031] Preferably, in S4, the deformation-resistant robust registration strategy introduces a deformation-aware adaptive truncation kernel function to allocate registration weights. The formula is:

[0032] in, For the residual of the point pair in the k-th iteration, The environmental noise threshold. To determine the deformation cutoff threshold, The dynamic scaling factor is used; the weighted error objective function is: ; S5 also includes dynamic calculation of the detection limit threshold. The formula is: ,in These are the local surface roughness of the baseline point cloud and the difference point cloud, respectively. To correspond to the number of valid point clouds in the neighborhood, This refers to global spatial alignment error; S6 also includes a joint clustering distance measure. The formula is: ,in It is a three-dimensional Euclidean distance. The normal distance of M3C2 This is the deformation depth penalty coefficient; S6 also includes adaptive clustering neighborhood radius. The calculation formula is: ,in It is the spatial connectivity expansion factor. The number of nearest neighbors. The coordinates of the nearest neighbor points, This is the depth gradient tolerance coefficient. For point The corresponding detection limit threshold; the surface area of ​​the damaged instance is obtained by accumulating the area of ​​the triangular mesh through three-dimensional triangular mesh reconstruction, and the maximum physical indentation depth is determined by extracting the maximum absolute value by traversing the M3C2 distance.

[0033] Specifically, the S4 deformation-resistant robust registration requirement aims to achieve precise alignment of the main structure. The deformation-aware adaptive truncation kernel function, based on the residual distribution characteristics in iterative registration, divides the residuals into noise, transition, and deformation intervals. Different weights are assigned to point pairs in different intervals. The environmental noise threshold is set to 0.5 to 1 cm, and the confirmed deformation truncation threshold is set to 3 to 5 cm. The dynamic scaling factor is estimated through the absolute median difference of the residuals in the current iteration, ensuring that the weight allocation adapts to the differences between noise and deformation. The weighted error objective function is based on the least squares principle, using weighted optimization to suppress the influence of large residual point pairs in local deformation regions, allowing iterative optimization to focus on the undamaged main structure. In implementation, the topological node sets of the point cloud to be measured and the reference point cloud are first extracted to establish an initial correspondence. After coarse registration, the point-to-surface distance is used as the error objective function. The deformation-aware adaptive truncation kernel function is introduced to dynamically allocate weights, iteratively minimizing the weighted error objective function until the joint iteration termination condition is met. The two types of formulas work together to solve the problem of global registration distortion caused by local deformation, and achieve precise spatial alignment of the main structure.

[0034] The requirements of initial anomalous point cloud extraction (S5) and damage instance separation (S6) are met to achieve accurate screening and separation. The detection limit threshold formula is based on the calculation principle of confidence intervals in statistics. It combines the local surface roughness of the baseline and difference point clouds, the number of effective point clouds, and the global spatial alignment error to calculate the detection limit at a 95% confidence level. This threshold can adapt to the point cloud quality and registration accuracy of different regions, avoiding missed or false detections caused by a fixed threshold. The joint clustering distance measure formula combines the constraints of spatial connectivity and deformation depth continuity, weighted and fused the difference between 3D Euclidean distance and M3C2 normal distance. The deformation depth penalty coefficient is set to 0.3 to 0.7 to balance the weights of the two types of constraints. In implementation, the detection limit threshold formula is used in S5 to calculate a dynamic threshold, extracting point clouds with absolute M3C2 distance values ​​exceeding the threshold as initial anomalous point clouds. In S6, the joint clustering distance measure formula is used to measure the correlation between two points in the initial anomalous point cloud, providing a distance criterion for subsequent adaptive density clustering. The two types of formulas respectively address the issues of accuracy in abnormal point cloud extraction and rationality in damage instance separation, providing a reliable data foundation for quantitative damage assessment.

[0035] The S6 adaptive density clustering and damage parameter quantification requirements are addressed. The adaptive clustering neighborhood radius formula is based on local point cloud density and depth noise tolerance, combined with the average distance from the core point to its nearest neighbors, spatial connectivity expansion factor (1.2 to 1.5), depth gradient tolerance coefficient (0.5 to 0.8), and detection limit threshold. The neighborhood radius is dynamically adjusted, with the number of nearest neighbors, k, set to 10 to 20 to ensure the neighborhood radius adapts to the point cloud characteristics of different regions. The damage parameter quantification method is based on 3D geometric reconstruction and distance statistics. Surface area is obtained by accumulating the area of ​​facets through triangular mesh reconstruction. The maximum side length of the triangular mesh is set to 0.5 to 1 cm, and the maximum physical indentation depth is determined by extracting the maximum absolute value through traversing the M3C2 distance. In implementation, the adaptive clustering neighborhood radius formula is first used to assign a dynamic neighborhood radius to the core points of the initial abnormal point cloud. Independent damage instances are separated based on the joint clustering distance measure and neighborhood radius. Then, triangular mesh reconstruction is performed on the point cloud of each damage instance, and the surface area and maximum physical indentation depth are calculated. The formula and quantification method work together to solve the adaptive problem of separating multiple damage instances and the accuracy problem of damage parameter quantification, and finally output the quantitative indicators of each damage instance.

[0036] Preferred, such as Figure 2 As shown, step S3 includes the following sub-steps: S31, acquiring the visual dense point cloud and laser point cloud generated in S2, respectively serving as the source point cloud and target point cloud, and clarifying the differences in data source attributes and scale characteristics between the two types of point clouds; S32, calculating the local principal curvature of the two types of point clouds, selecting saddle surface feature regions based on Gaussian curvature to construct a set of topological nodes, and extracting three-dimensional common feature descriptors in the local spatial neighborhood of the nodes through the SpinNet network; S33, establishing an initial correspondence based on the feature descriptors, constructing a spatial side length topology graph, and eliminating mismatched point pairs through scale-invariant topology error verification, and inputting the pure correspondence into the TEASER++ algorithm to solve the global scale factor, rotation matrix, and translation vector; S34, based on coarse registration, using the SymmetricICP algorithm to introduce bidirectional surface normal constraints to perform fine registration, eliminating local density asymmetric residuals, and forming a high-precision benchmark point cloud model before construction.

[0037] Specifically, the S3 heteroscale registration and baseline model construction includes four sub-steps. S31, as the initial step, clearly identifies the two types of point clouds received from the S2 output, defining the attribute differences between the source and target point clouds to provide a clear data foundation for subsequent registration. This step requires no additional parameter settings; its core is to complete data reception and classification. In S32, a quadratic surface fitting method is used to calculate the local principal curvature. The threshold for Gaussian curvature screening of saddle surface feature regions is set to a negative number. Extensive experiments are conducted to determine a reasonable range for accurate identification of tube nodes. The constructed topological node set avoids interference from repetitive cylindrical textures. The local neighborhood radius for SpinNet network feature descriptor extraction is set to 3 to 5 centimeters to ensure that the extracted features possess rotational invariance and cross-modal commonality. After establishing the initial correspondence in S33, the consistency verification threshold for the spatial side length topology map is set to 0.05 to 0.1, effectively eliminating mismatched point pairs due to axial slippage. The maximum error tolerance of the truncation least squares in the TEASER++ algorithm is set to 2 to 3 cm to ensure the robustness of the scale factor, rotation matrix, and translation vector calculations. In S34, the SymmetricICP algorithm for fine registration iterations is set to 50 to 100 times, with a convergence threshold of 0.01 cm for each iteration. By minimizing the symmetric projection error through bidirectional surface normal constraints, the asymmetric residuals between the visual point cloud and the laser point cloud are eliminated. The final high-precision benchmark point cloud model constructed before construction has a registration error of less than 2 cm, possessing both the dense texture of visual data and the absolute scale of LiDAR data, providing a reliable benchmark for subsequent damage assessment.

[0038] Preferred, such as Figure 3 As shown, S4 includes the following sub-steps: S41, during construction, visual image sequences and lidar data are collected simultaneously, strictly following the independent reconstruction process of S2 and the heteroscale fusion process of S3, to generate current state test point cloud data consistent with the scale of the reference point cloud model; S42, extracting regions in the test point cloud and the reference point cloud model that meet the saddle surface feature conditions, and constructing K-type, T-type, and Y-type topological node sets as undamaged stiffness extreme value regions; S43, using the SpinNet network to extract three-dimensional feature descriptors in the topological node sets and establish initial correspondences, inputting the TEASER++ algorithm to eliminate outliers and mismatches through truncated least squares estimation, and solving the global rotation matrix and translation vector at the same scale; S44, using the point-to-surface distance as the error objective function, introducing a deformation-aware adaptive truncated kernel function to identify local deformation regions, and iteratively minimizing the weighted error objective function to complete the accurate spatial alignment of the main structure and obtain the difference point cloud data.

[0039] Specifically, the S4 deformation-resistant robust registration includes four sub-steps, ensuring precise alignment of the main structure by clearly defining parameters and implementation procedures. In S41, when collecting field data, the monocular camera's overlap rate is maintained at 60% to 70%, and the LiDAR scanning frequency is set to 100,000 to 300,000 points per second. The independent reconstruction process of S2 and the heteroscale fusion process of S3 are strictly reused to ensure that the measured point cloud and the reference point cloud are completely consistent in data format and scale characteristics, providing a basis for direct comparison. The core of this step is to ensure the uniformity of the data generation mechanism and avoid registration deviations caused by process differences. In S42, when extracting the saddle surface feature region, the Gaussian curvature screening method is used. The constructed K-type, T-type, and Y-type topological node sets serve as the undamaged stiffness extreme value region. The selection of this region is based on the structural mechanical properties of the jacket structure; the stiffness of the pipe nodes is much higher than that of the middle area of ​​the pipe segment, making them less prone to deformation and providing stable feature support for registration. The number of topological nodes is limited during the screening process to ensure a sufficient matching base. In S43, the parameters of the SpinNet network for extracting feature descriptors are consistent with those in S32, ensuring the consistency and comparability of features. The truncation least squares estimation threshold of the TEASER++ algorithm is set to 3 to 5 cm, effectively eliminating outlier mismatched point pairs caused by collisions. The calculated global rotation matrix and translation vector can achieve preliminary alignment between the point cloud to be measured and the reference point cloud, with the initial alignment error controlled within 3 to 5 cm. S44 uses the point-to-surface distance as the objective error function, sets the environmental noise threshold to 0.5 to 1 cm, and the confirmed deformation cutoff threshold to 3 to 5 cm. The dynamic scaling factor is calculated in real time using the absolute median difference of the current iteration residual. The iteration termination conditions for the weighted error objective function include an error change rate of less than 0.001, an incremental translation vector magnitude of less than 0.1 cm and an incremental rotation angle of less than 0.1 degrees, or an iteration count of 200. Through this dynamic weight adjustment strategy, interference from local deformation regions is effectively isolated, allowing registration to focus on the undamaged main structure. The resulting difference point cloud data accurately reflects the geometric differences between the tested structure and the reference structure, providing a direct basis for damage identification.

[0040] Preferred, such as Figure 4 As shown, step S5 includes the following sub-steps: S51, acquiring the spatial alignment difference point cloud data output by S4, defining a cylindrical search neighborhood along the surface normal direction with each point on the reference point cloud model surface as the core, and clarifying the spatial range and geometric parameters of the neighborhood; S52, using the M3C2 algorithm to calculate the mean projection distance of the difference point cloud in the cylindrical search neighborhood, obtaining the M3C2 distance corresponding to each point; S53, combining the local surface roughness, point cloud density, and global spatial alignment error of the reference point cloud and the difference point cloud, dynamically calculating the 95% confidence level detection limit threshold corresponding to each point; S54, filtering out the point clouds with an absolute value of M3C2 distance greater than the corresponding detection limit threshold to form an initial abnormal point cloud, clarifying the spatial distribution range of the abnormal point cloud.

[0041] Specifically, the initial anomaly point cloud extraction in S5 includes four sub-steps, with parameter optimization and process standardization ensuring the accuracy of the extraction. S51: After acquiring the difference point cloud data, a cylindrical search neighborhood is constructed with each point on the surface of the baseline point cloud model as the core. The neighborhood radius is set to 1-3 cm based on the baseline point cloud density, and the neighborhood height is set to 2-5 cm. This parameter range has been experimentally verified to balance the accuracy of local comparison with computational efficiency, ensuring sufficient coverage of local areas while avoiding interference from irrelevant point clouds. The direction of the neighborhood strictly follows the normal vector of the baseline point cloud surface, ensuring the accuracy of distance calculation. S52: When calculating the projected mean distance using the M3C2 algorithm, the local geometric centroids of the baseline and difference point cloud subsets are solved separately. The projection of the line connecting the two centroids in the normal vector direction is used as the M3C2 distance. This calculation method effectively reflects the relative displacement of the two point clouds in the normal direction, avoiding interference from in-plane translation. During the calculation process, the number of valid points within the neighborhood is screened, with a minimum of 30-50 valid points set to ensure the stability of the centroid calculation. When dynamically calculating the detection limit threshold in S53, the local surface roughness, effective point cloud quantity, and global spatial alignment error of the reference point cloud and the difference point cloud are comprehensively considered. The local surface roughness is calculated by the variance of the distance from the neighboring points to the fitting plane, and the global spatial alignment error is taken from the output of the deformation-resistant robust registration module. The 95% confidence level setting meets the statistical significance test standard and can effectively distinguish between real damage and system noise. In S54, when screening the initial abnormal point cloud, point clouds with an absolute M3C2 distance greater than the corresponding detection limit threshold are included. At the same time, a minimum threshold for the number of abnormal points is set to 50 to 100, and a small number of false abnormal areas composed of noise points are removed. This step ensures that the initial abnormal point cloud can accurately cover the potential damage area through double screening, which avoids missing real damage and reduces the subsequent processing pressure caused by false detection, laying a high-quality data foundation for the damage instance separation in S6.

[0042] like Figure 5As shown, a device for quantitative assessment of jacket damage based on heteroscale data registration and fusion is presented. This device is applied to a method for quantitative assessment of jacket damage based on heteroscale data registration and fusion, and includes: a multi-source heterogeneous data acquisition and transmission module for independently acquiring monocular visual image sequences and lidar data of offshore wind turbine jackets, performing lossless transmission and temporary storage of the two types of raw data through a data transmission interface; a dual-branch independent 3D reconstruction module connected to the multi-source heterogeneous data acquisition and transmission module, performing feature matching, camera pose calculation, and depth estimation operations on the received monocular visual image sequences to generate visually dense point clouds, and performing high-frequency vibration filtering, tightly coupled pose calculation, and global map update operations on the lidar data to generate lidar point clouds; and a heteroscale point cloud registration and fusion module connected to the dual-branch independent 3D reconstruction module, extracting the 3D topological features of the two types of point clouds, and using scale factors and pose variables... The decoupling strategy performs heteroscale registration and fuses the data to generate a high-precision benchmark point cloud model before construction. The deformation-resistant robust registration and alignment module connects to both the heteroscale point cloud registration and fusion module and the dual-branch independent 3D reconstruction module. It receives the point cloud to be measured during construction and the benchmark point cloud model, performs precise alignment of the main structure using a dynamic weight adjustment strategy, and outputs the difference point cloud. The anomaly point cloud extraction and damage separation module connects to the deformation-resistant robust registration and alignment module. It uses the M3C2 algorithm to calculate surface distances and combines dynamic thresholds to extract initial anomaly point clouds. It separates independent damage instances using an adaptive density clustering algorithm constrained by both spatial location and deformation depth. The damage parameter quantitative calculation module connects to the anomaly point cloud extraction and damage separation module. It performs triangular mesh reconstruction on the point clouds of each independent damage instance to calculate the surface area, traverses the M3C2 distances to extract the maximum physical indentation depth, and outputs a complete quantitative damage assessment result.

[0043] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a method for quantitative assessment of duct stent damage based on heteroscale data registration and fusion.

[0044] A computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for quantitative assessment of duct stent damage based on heteroscale data registration and fusion.

[0045] like Figure 6The computer device described includes a processor, memory, communication interface, and display components that are interconnected via a system bus. It should be noted that only a computer device with four components is shown in the figure; however, it should be understood that implementing all shown components is not required, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described here is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions. Its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, and deep learning computing platforms including GPUs.

[0046] Computer devices can include desktop computers, laptops, high-performance workstations, handheld computers, and cloud servers. Computer devices can interact with users through keyboards, mice, remote controls, touchpads, or voice-activated devices.

[0047] In some embodiments, the processor may be a central processing unit (CPU), a graphics processing unit (GPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is typically used to control the overall operation of a computer device. In this embodiment, the processor is used to run program code stored in memory or process data, such as running the program code for the aforementioned method for quantitative assessment of duct stent damage based on heteroscale data registration and fusion, to implement a method for quantitative assessment of duct stent damage based on heteroscale data registration and fusion.

[0048] The memory includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory can be an internal storage unit of a computer device, such as the hard disk or RAM of the computer device. In other embodiments, the memory can also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. Of course, the memory can also include both internal storage units and external storage devices of the computer device. In this embodiment, the memory is typically used to store the operating system and various application software installed on the computer device, such as the program code of a method for quantitative assessment of duct stent damage based on heteroscale data registration and fusion. In addition, the memory is also used to temporarily store various types of data that have been output or will be output.

[0049] The communication interface may include a wireless network interface or a wired network interface. This communication interface is typically used to establish a communication connection between computer equipment and other electronic devices (such as UAV-borne monocular cameras and lidar sensors) to enable real-time or offline transmission of heterogeneous data.

[0050] The display component can be a liquid crystal display, an organic light-emitting diode (OLED) display, or the like. In this embodiment, the display component is used to intuitively display the 3D reconstructed point cloud model, the registered difference heat map, and a quantitative assessment report including the damaged surface area and the maximum indentation depth, providing a visual interface for human-computer interaction.

[0051] A method, device, and medium for quantitative assessment of duct stent damage based on heteroscale data registration and fusion are presented. This method employs a step-by-step approach to acquire and independently reconstruct multi-source data, combined with a heteroscale registration strategy that decouples scale and pose. It eliminates the need for rigid hardware connections and offline calibration, flexibly leveraging the complementary advantages of dense texture visual data and the absolute scale of LiDAR, successfully addressing the problem of feature matching failure in weakly textured environments. An innovative deformation-resistant robust registration mechanism is introduced, dynamically adjusting the weights of point pairs in locally deformed regions during iteration, effectively mitigating the risk of global model misalignment and achieving precise spatial alignment between the undamaged main structure and the structure under test. Through dynamic threshold filtering and joint constraint clustering algorithms, multiple concurrent damage instances are accurately separated, enabling refined quantification of damage parameters and overcoming the limitations of traditional methods in quantitative assessment.

[0052] At the device level, a complete technical chain from data acquisition to result output is constructed through the coordinated operation of six functional modules. Each module is closely connected and has a specialized function, significantly improving the efficiency and reliability of damage detection. This technical solution completely changes the traditional high-risk detection mode that relies on manual probing. At the same time, it addresses the core shortcomings of existing technologies, such as poor adaptability to multi-source data and weak resistance to deformation interference. Through key technological innovations such as topological feature extraction, dynamic weight adjustment, and adaptive clustering, it achieves high-precision positioning and quantitative assessment of jacket damage. This provides more efficient and reliable technical support for the safety management and control of marine engineering construction, and significantly improves the scientific nature and pertinence of marine platform structural maintenance.

[0053] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0054] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for quantitative assessment of duct stent damage based on heteroscale data registration and fusion, characterized in that, Includes the following steps: S1, which collects monocular visual image sequences and lidar data of offshore wind turbine jacket foundations respectively; S2, performs 3D reconstruction on monocular visual image sequences and LiDAR data respectively, forming scale-free visual dense point clouds and LiDAR point clouds with absolute scale; S3 extracts three-dimensional features from visually dense point clouds and laser point clouds, performs heteroscale registration through an optimization strategy that decouples scale factors and pose variables, and fuses and constructs a high-precision benchmark point cloud model for construction that combines dense texture and absolute scale. S4. During the construction process, the current state of the point cloud data to be measured is obtained according to the reconstruction and fusion process of S2-S3. A robust registration strategy against deformation interference is adopted to make the point cloud data to be measured spatially aligned with the reference point cloud model. In the iteration, the point pair residual is calculated and the registration weight of large residual point pairs in the local deformation area is dynamically suppressed to obtain spatially aligned difference point cloud data. S5, based on the difference point cloud data, uses the M3C2 algorithm to calculate the surface distance, and extracts the initial abnormal point cloud according to the distance threshold; S6. For the initial abnormal point cloud, an adaptive density clustering algorithm based on the joint constraints of spatial location and deformation depth is used to separate multiple independent damage instance point clouds, and the surface area and maximum physical indentation depth of each damage instance are quantified respectively.

2. The method for quantitative assessment of duct stent damage based on heteroscale data registration and fusion according to claim 1, characterized in that, When performing 3D reconstruction on a monocular vision image sequence in S2, a water surface specular mask is constructed. The formula is: ,in, This represents the absolute brightness value of a pixel in a grayscale image. The average brightness of the local spatial neighborhood background. To achieve high light sensitivity, a high-frequency wind-induced adaptive filter is introduced into the original acceleration sequence during lidar data reconstruction to smooth the acceleration. The calculation formula is: ,in These are the current raw acceleration observations. The smooth acceleration from the previous moment. It is a dynamic smoothing factor, and The high-frequency variance of acceleration within the sliding time window. is the wind vibration damping constant.

3. The method for quantitative assessment of duct stent damage based on heteroscale data registration and fusion according to claim 1, characterized in that, In S3, heteroscale registration extracts topological nodes using Gaussian curvature. The calculation formula is: ,in, For point The two principal curvatures of the locally quadratic fitted surface are used; a spatial side length topology graph is constructed for consistency verification, while the scale remains unchanged and the topology error is not considered. The formula is: ,in, For the three-dimensional coordinates of the visual point cloud topology nodes, The three-dimensional coordinates of the topological nodes corresponding to the laser point cloud. It represents the Euclidean distance in space.

4. The method for quantitative assessment of duct stent damage based on heteroscale data registration and fusion according to claim 1, characterized in that, In S4, a deformation-resistant robust registration strategy is introduced, which employs a deformation-aware adaptive truncation kernel function to allocate registration weights. The formula is: in, For the residual of the point pair in the k-th iteration, The environmental noise threshold. To determine the deformation cutoff threshold, The dynamic scaling factor is used; the weighted error objective function is: ; S5 also includes dynamic calculation of the detection limit threshold. The formula is: ,in These are the local surface roughness of the baseline point cloud and the difference point cloud, respectively. To correspond to the number of valid point clouds in the neighborhood, This refers to global spatial alignment error; S6 also includes a joint clustering distance measure. The formula is: ,in It is a three-dimensional Euclidean distance. The normal distance of M3C2 This is the deformation depth penalty coefficient; S6 also includes adaptive clustering neighborhood radius. The calculation formula is: ,in It is the spatial connectivity expansion factor. The number of nearest neighbors. The coordinates of the nearest neighbor points, This is the depth gradient tolerance coefficient. For point The corresponding detection limit threshold; the surface area of ​​the damaged instance is obtained by accumulating the area of ​​the triangular mesh through three-dimensional triangular mesh reconstruction, and the maximum physical indentation depth is determined by extracting the maximum absolute value by traversing the M3C2 distance.

5. The method for quantitative assessment of duct stent damage based on heteroscale data registration and fusion according to claim 1, characterized in that, S3 includes the following steps: S31, acquire the visual dense point cloud and laser point cloud generated in S2, and use them as the source point cloud and target point cloud respectively, to clarify the differences in data source attributes and scale characteristics between the two types of point clouds; S32, calculate the local principal curvature of two types of point clouds, select saddle surface feature regions based on Gaussian curvature to construct a set of topological nodes, and extract three-dimensional common feature descriptors in the local spatial neighborhood of the nodes through the SpinNet network; S33: Based on the feature descriptor, an initial correspondence is established, a spatial side length topology graph is constructed, and mismatched point pairs are eliminated through scale-invariant topology error verification. The pure correspondence is then input into the TEASER++ algorithm to solve the global scale factor, rotation matrix, and translation vector. S34, based on coarse registration, uses the SymmetricICP algorithm to introduce bidirectional surface normal constraints to perform fine registration, eliminate local asymmetric residuals, and form a high-precision benchmark point cloud model before construction.

6. The method for quantitative assessment of duct stent damage based on heteroscale data registration and fusion according to claim 1, characterized in that, S4 includes the following sub-steps: S41, during the construction process, visual image sequences and lidar data are collected simultaneously, strictly following the independent reconstruction process of S2 and the heteroscale fusion process of S3, to generate current state point cloud data to be measured that is consistent with the scale of the benchmark point cloud model. S42, extract the regions that satisfy the saddle surface feature conditions in the point cloud to be measured and the reference point cloud model, and construct K-type, T-type and Y-type topological node sets as the undamaged stiffness extreme value regions; S43 uses the SpinNet network to extract three-dimensional feature descriptors within the set of topological nodes and establishes an initial correspondence. The TEASER++ algorithm is then used to eliminate outliers and false matches through truncated least squares estimation and to solve for the global rotation matrix and translation vector at the same scale. S44 uses the point-to-surface distance as the error objective function, introduces a deformation-aware adaptive truncation kernel function to identify local deformation regions, and achieves accurate spatial alignment of the main structure and obtains difference point cloud data by iteratively minimizing the weighted error objective function.

7. The method for quantitative assessment of duct stent damage based on heteroscale data registration and fusion according to claim 1, characterized in that, S5 includes the following steps: S51: Obtain the spatial alignment difference point cloud data output by S4. Using each point on the surface of the reference point cloud model as the core, define a cylindrical search neighborhood along the surface normal vector direction to clarify the spatial range and geometric parameters of the neighborhood. S52, the M3C2 algorithm is used to calculate the mean distance of the projection of the difference point cloud in the cylindrical search neighborhood, and the M3C2 distance of each point is obtained. S53, combining the local surface roughness, point cloud density and global spatial alignment error of the reference point cloud and the difference point cloud, dynamically calculates the 95% confidence level detection limit threshold for each point; S54 filters out point clouds whose absolute distance from M3C2 is greater than the corresponding detection limit threshold to form an initial abnormal point cloud and clarify the spatial distribution range of the abnormal point cloud.

8. A device for quantitative assessment of duct stent damage based on heteroscale data registration and fusion, characterized in that, This device is applied to the quantitative assessment method for duct stent damage based on heteroscale data registration and fusion as described in claim 1, comprising: The multi-source heterogeneous data acquisition and transmission module is used to independently acquire monocular visual image sequences and lidar data of offshore wind turbine jackets, and to perform lossless transmission and temporary storage of the two types of raw data through the data transmission interface. The dual-branch independent 3D reconstruction module connects to the multi-source heterogeneous data acquisition and transmission module. It performs feature matching, camera pose calculation and depth estimation operations on the received monocular visual image sequence to generate visual dense point clouds, and performs high-frequency vibration filtering, tightly coupled pose calculation and global map update operations on the lidar data to generate lidar point clouds. The heteroscale point cloud registration and fusion module connects to the dual-branch independent 3D reconstruction module, extracts the 3D topological features of the two types of point clouds, performs heteroscale registration through the decoupling strategy of scale factor and pose variable, and fuses to generate a high-precision benchmark point cloud model before construction. The deformation-resistant robust registration and alignment module is connected to the heteroscale point cloud registration and fusion module and the dual-branch independent 3D reconstruction module, respectively. It receives the point cloud to be measured and the reference point cloud model during construction, and performs accurate alignment of the main structure through a dynamic weight adjustment strategy and outputs the difference point cloud. The abnormal point cloud extraction and damage separation module is connected to the deformation-resistant robust registration and alignment module. It uses the M3C2 algorithm to calculate the surface distance and combines dynamic thresholds to extract the initial abnormal point cloud. It separates independent damage instances through an adaptive density clustering algorithm constrained by spatial location and deformation depth. The damage parameter quantitative calculation module connects the abnormal point cloud extraction and damage separation modules. It performs triangular mesh reconstruction on the point cloud of each independent damage instance to calculate the surface area, traverses the M3C2 distance to extract the maximum physical indentation depth, and outputs a complete quantitative damage assessment result.

9. A computer device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a method for quantitative assessment of duct stent damage based on heteroscale data registration and fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements a method for quantitative assessment of duct stent damage based on heteroscale data registration and fusion as described in any one of claims 1 to 7.

Images