Tower body and peripheral attachment recognition method fusing visualized data

Abstract

The application discloses a tower body and peripheral attachment recognition method fusing visual data, and relates to the technical field of infrastructure inspection, and comprises the following steps: identifying and extracting at least one surface area where a suspected attachment is located from three-dimensional point cloud data of a tower, as an attached area; constructing an attachment domain for data fusion analysis in a three-dimensional space with the attached area as the center; acquiring multi-source visual data of the tower, and mapping the multi-source visual data to the attachment domain for feature fusion to form a unified multi-modal evidence set; based on the multi-modal evidence set, constructing an attachment graph representing the correlation between the attachment and the tower components, and analyzing the attachment graph to output an analysis result. The application realizes joint cognition and analysis of a "tower-attachment" spatial unity, can accurately determine whether a foreign object is "actually attached" and its specific position, and provides direct technical support for grading and disposal of high-risk attachments.

Description

Technical Field

[0001] This invention relates to the field of infrastructure inspection technology, and more specifically, to a method for identifying the pole body and surrounding attachments by integrating visualized data. Background Technology

[0002] Transmission towers, as crucial supporting structures for power grid lines, directly impact the safety and stability of the lines through their operational status. Composed of multiple components including the tower body, crossarms, insulator strings, and hardware, towers are susceptible to structural damage or safety hazards due to long-term outdoor operation, including wind, rain, bird damage, and external attachments. For example, bird nests, hanging objects, tree branches, and plastic film attached to crossarms, insulators, or tower legs can cause discharges, short circuits, or tripping. With the development of smart grids and digital twin technologies, visualization and identification methods integrating laser point clouds, visible light images, infrared thermal images, and multispectral data have become a core focus. These methods enable automated identification of towers and surrounding attachments, providing technical support for accurate line condition assessment, proactive risk warning, and full lifecycle management. This forms a crucial foundation for intelligent inspection and safe operation of power lines.

[0003] However, existing technologies still face severe challenges in practice. Their bottleneck is not simply insufficient recognition accuracy, but rather a lack of joint understanding and analytical capabilities regarding the unified spatial entity of "tower-attached objects." Specifically: First, the problem of "pseudo-fusion" is prominent. Multi-source data (such as point clouds and images) are often processed in parallel rather than deeply fused, failing to generate a unified feature expression that supports accurate spatial relationship determination. This results in the algorithm being able to identify "what is there," but struggling to determine whether foreign objects are "truly attached" and "where they are attached." Second, the limitations of "decoupled recognition" are obvious. Treating tower body recognition and attachment detection as two independent tasks ignores their inherent spatial constraints, causing the system to be unable to understand critical risk scenarios such as "vines entwining insulators" or "hanging objects bridging conductors." The output results cannot directly support the accurate classification and maintenance of high-risk attachments, significantly reducing practicality. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for identifying the pole body and surrounding attachments by integrating visualized data. First, an attachment domain is established based on three-dimensional point cloud data. A multimodal evidence set is formed through multi-source data mapping and feature fusion. Then, a dependency map representing the relationship between the attachments and pole components is constructed. The dependency map is then analyzed, and the analysis results are output to solve the problems of existing methods being unable to accurately determine whether foreign objects are "truly attached" and their spatial location, as well as the difficulty in understanding complex risk scenarios.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for identifying the main body of a power pole and its surrounding attachments by integrating visualized data includes the following steps: identifying and extracting at least one surface area containing a suspected attachment from the three-dimensional point cloud data of the power pole, as the attachment area; constructing a dependency domain in three-dimensional space centered on the attachment area for data fusion analysis; acquiring multi-source visualized data of the power pole and mapping the multi-source visualized data to the dependency domain for feature fusion to form a unified multimodal evidence set; constructing a dependency map representing the relationship between the attachment and the power pole components based on the multimodal evidence set, parsing the dependency map, and outputting the analysis results.

[0006] In a preferred embodiment, the method for acquiring the three-dimensional point cloud data includes using lidar, structured light, or multi-view stereo photography to acquire three-dimensional point cloud data of the tower and its surrounding environment.

[0007] In a preferred embodiment, the step of constructing a dependency domain for data fusion analysis in three-dimensional space with the attachment region as the center specifically involves: performing semantic segmentation on the overall three-dimensional point cloud data of the tower to identify and extract point cloud clusters of key components; obtaining the main structural line representing the main attachment direction of the component through three-dimensional geometric fitting based on each point cloud cluster; determining the core point of the attachment region based on the spatial position of the main structural line and its connection relationship with the main body of the tower, and generating the corresponding dependency domain according to the component type and structural characteristics with the core point as the spatial center.

[0008] In a preferred embodiment, the key components include an insulator string, an equalizing ring, conductor suspension hardware, and a crossarm.

[0009] In a preferred embodiment, the step of mapping multi-source visualization data to a dependency domain for feature fusion to form a unified multimodal evidence set specifically involves: extracting the image region corresponding to the dependency domain and performing spatial projection matching between the image region and the dependency domain, so that each dependency domain point obtains its corresponding texture and color attributes; for the projected dependency domain points, calculating and constructing a local cross-modal feature group based on the point's normal vector, curvature, reflection intensity, and image texture gradient; performing attachment sensitivity discrimination on the cross-modal feature group to screen primary attachment evidence; and performing spatial consistency aggregation on all primary attachment evidence within the dependency domain, forming a multimodal evidence set by evaluating the directional consistency, connectivity, and cross-modal overlap between the evidence.

[0010] In a preferred embodiment, the screening of primary attachment evidence includes: constructing a local perturbation field for the attachment region based on the endogenous coupling relationship of local cross-modal feature groups, and tracking the response trajectory of each feature under the perturbation field; determining a subset of features that are highly sensitive to attachment behavior by jointly measuring the change amplitude and direction consistency of the response trajectory; and outputting the features in the subset that maintain a monotonic evolution trend before and after the perturbation as primary attachment evidence.

[0011] In a preferred embodiment, the step of constructing a dependency graph representing the relationship between the attachment and the tower components based on a multimodal evidence set, wherein the nodes of the graph represent dependency contacts and the edges represent attachment paths, specifically involves: constructing a dependency contact potential field within the dependency domain based on the multimodal evidence set, and using the peak points of the potential field as candidate dependency contacts; performing local causal inference on the evidence combinations of the candidate dependency contacts to determine the dependency graph nodes; constructing a virtual adhesion force field based on the evidence flow direction between dependency graph nodes, and simulating and generating candidate attachment paths through path line integration; mapping the graph topology formed by the dependency graph nodes and candidate attachment paths to an energy function, and obtaining the final dependency graph by minimizing the overall energy.

[0012] In a preferred embodiment, the construction of the virtual adhesion field based on the evidence flow between nodes in the attached diagram specifically involves: constructing a field source characterizing the node association strength based on the coupling consistency of the multimodal features between nodes; using the coupling consistency as an attraction source and the spatial dispersion between nodes as a repulsion source to synthesize a dynamic non-uniform virtual adhesion field; in the virtual adhesion field, the force line direction is determined by the dominant change direction of the coupling consistency, and the force line strength is a function of the coupling consistency strength and the spatial dispersion.

[0013] In a preferred embodiment, the step of parsing the attached drawing and outputting the analysis results specifically involves: matching the attached drawing with a predefined attachment mechanism template to identify the attachment category; parsing the edges in the final attached drawing after energy minimization optimization as attachment paths; querying the spatial distribution of evidence strength in the multimodal evidence set along the attachment paths to determine the attachment depth; and mapping the nodes in the final attached drawing to the semantic 3D model of the tower component to determine the affected component.

[0014] A system for identifying the pole body and surrounding attachments by integrating visualized data includes: an identification construction unit, used to identify and extract at least one surface area where a suspected attachment is located from the three-dimensional point cloud data of the pole, as the attachment area; and to construct a dependency domain in three-dimensional space centered on the attachment area for data fusion analysis; a feature fusion unit, used to acquire multi-source visualized data of the pole, and to map the multi-source visualized data to the dependency domain for feature fusion to form a unified multimodal evidence set; and an analysis output unit, used to construct a dependency map representing the relationship between the attachment and pole components based on the multimodal evidence set, and to parse the dependency map and output the analysis results.

[0015] The technical effects and advantages of the method for identifying the tower body and surrounding attachments by integrating visualized data in this invention are as follows: 1. This invention generates a unified multimodal evidence set by constructing an "attachment domain" and fusing features from multi-source visualization data under this spatial benchmark, thus solving the "pseudo-fusion" problem of existing technologies. This approach deeply correlates heterogeneous data such as point clouds and images in a unified three-dimensional space, enabling the system not only to identify what the attachment is, but also to accurately determine its three-dimensional spatial attachment relationship with tower components, achieving a cognitive leap from "what is there" to "whether it is truly attached" and "where it is attached".

[0016] 2. This invention effectively overcomes the limitations of "decoupled identification" by constructing a "graph-based" model based on a multimodal evidence set and performing global consistency analysis. This method places tower components and attachments within a unified graph structure for joint reasoning, enabling a complete understanding of the spatial topological relationships in complex risk scenarios such as "vines entwining insulators." This results in a structured output including attachment type, path, depth, and affected components, providing a direct and reliable decision-making basis for the accurate classification and efficient operation and maintenance of high-risk attachments, greatly enhancing the practical value of the method. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the method for identifying the tower body and surrounding attachments that integrates visualized data according to the present invention. Figure 2 This is a schematic diagram of the system structure of the method for identifying the tower body and surrounding attachments that integrates visualized data according to the present invention. Figure 3 This is a schematic diagram illustrating the process of constructing a virtual adhesion field in the method for identifying the tower body and surrounding attachments that integrates visualized data according to the present invention. Detailed Implementation

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

[0019] Example 1, Figure 1 The present invention provides a method for identifying the tower body and surrounding attachments by integrating visualization data, comprising the following steps: S1. From the three-dimensional point cloud data of the tower, identify and extract at least one surface area where a suspected attachment is located as the attachment area; with the attachment area as the center, construct an attachment domain in three-dimensional space for data fusion analysis. In this embodiment, the method for acquiring the three-dimensional point cloud data includes using lidar, structured light, or multi-view stereo photography to acquire three-dimensional point cloud data of the tower and its surrounding environment.

[0020] In this embodiment, a DJI Matrice 350 RTK drone platform was used, equipped with a Zenmuse L1 LiDAR module for data acquisition. The flight altitude was set at 50-70 meters above the top of the tower, and the flight speed did not exceed 3 m / s to ensure a point cloud density of no less than 200 points / square meter. By designing a flight path around the tower, unobstructed 3D point cloud data of 360° around the tower and within a 50-meter radius was acquired. The point cloud data includes 3D coordinates (X, Y, Z) and laser reflection intensity information.

[0021] It should be noted that the acquisition of 3D point cloud data is not limited to the devices and methods mentioned above. For example, a ground-based Faro Focus S 350 ground laser scanner can be used for high-precision static scanning. Alternatively, photogrammetry-based techniques can be employed, such as using a DJI P1 full-frame aerial survey camera to capture high-definition images from multiple angles, and then using software like ContextCapture for aerial triangulation and dense matching to generate high-precision 3D point cloud data.

[0022] The process involves identifying and extracting at least one surface region containing a suspected attachment from the 3D point cloud data of the tower, designating it as the attachment region; and constructing an attachment domain in 3D space centered on the attachment region for data fusion analysis, specifically: Specifically, the semantic segmentation of the overall three-dimensional point cloud data of the tower, and the identification and extraction of point cloud clusters of key components, involves: The key components include insulator strings, equalizing rings, conductor suspension hardware, and crossarms.

[0023] After acquiring the raw 3D point cloud data, a statistical outlier removal algorithm was first used to denoise the data. Subsequently, a deep learning-based semantic segmentation model was used to perform component-level identification on the overall 3D point cloud data of the tower, employing the RandLA-Net network as the underlying architecture. During model training, a self-built 3D point cloud dataset of power components was used for transfer learning. This dataset contains over 100,000 manually labeled tower component points, with labels including, but not limited to, insulator strings, equipotential rings, conductor suspension hardware, and crossarms. The denoised tower point cloud data was input into the trained RandLA-Net model, and the network automatically assigned a semantic label to each point cloud data point. After segmentation, connected component analysis was used to extract sets of spatially continuous points belonging to the same component into independent point cloud clusters, resulting in insulator string point cloud clusters, equipotential ring point cloud clusters, conductor suspension hardware point cloud clusters, and crossarm point cloud clusters. Specifically, the method of obtaining the principal structural lines representing the main attachment directions of the component based on each point cloud cluster through three-dimensional geometric fitting is as follows: Taking a crossarm point cloud cluster as an example, principal component analysis (PCA) is used to analyze the coordinates of the crossarm point cloud cluster. The direction of the first principal component (i.e., the direction of the eigenvector with the largest eigenvalue) is the principal direction of the crossarm. The line connecting the two extreme points of the crossarm point cloud cluster in this direction is defined as the principal structural line of the crossarm. For ring structures such as equalizing rings, the RANSAC algorithm is used to fit a circle or ellipse to obtain the center and normal vector of the ring surface. The straight line perpendicular to the fitted circle surface is taken as its principal structural line, and its direction reflects the central axis of the equalizing ring. Insulator strings are approximately vertically linearly distributed, and their principal structural lines can be obtained by RANSAC straight line fitting or PCA principal direction analysis. For conductor suspension hardware, the local principal direction (usually arc-shaped or short rod-shaped) of its point cloud is obtained through PCA as the principal structural line.

[0024] Specifically, the determination of the core point of the attachment area based on the spatial location of the main structural line and its connection relationship with the tower body, and the generation of corresponding attachment domains based on component type and structural characteristics using the core point as the spatial center, are as follows: For the crossarm, the core point is located at the connection point between its main structural line and the main structure of the tower. A flat, rectangular attachment domain is generated with the core point as the center and the main structural line of the crossarm as the central axis. The length of the domain is set along the crossarm direction, and the thickness is determined according to the crossarm's cross-sectional dimensions. For the insulator string, there are two core points, located at its upper attachment point (where it connects to the crossarm) and lower attachment point (where it connects to the conductor). A slender cylindrical attachment domain is constructed with the line connecting the two core points as the central axis to cover the upper and lower spaces of the insulator string. The scope is as follows: For the equalizing ring, the core point is the center of its fitted circle. A spherical attachment domain is generated with the core point as the center. The radius of the sphere is adaptively determined by the diameter of the fitted circle and the density of the point cloud. For the conductor suspension fitting, the core point is located at its theoretical contact point with the conductor model. A conical domain is generated with the core point as the vertex and the cone axis along the direction of its main structure line (i.e., the opposite direction of gravity, pointing to the sky). For the crossarm, a flat cuboid domain is generated with the main structure line as the central axis. For the equalizing ring, a spherical domain is generated.

[0025] S2, acquire multi-source visualization data of the tower, and map the multi-source visualization data to the dependency domain for feature fusion to form a unified multimodal evidence set; In this embodiment, the step of acquiring multi-source visualization data of the tower, mapping the multi-source visualization data to the dependency domain for feature fusion, and forming a unified multimodal evidence set specifically involves: A visible light image region acquired synchronously in space and time with the dependent domain is extracted. Each pixel of the image is precisely projected onto the point cloud of the dependent domain using forward intersection and collinearity equations, specifically through the following coordinate transformations:

[0026] in, These are image pixel coordinates. These are the three-dimensional coordinates of the point cloud within the dependent domain. It is the camera's intrinsic parameter matrix, including focal length and principal point offset. R is the extrinsic parameter matrix of the camera, R is the rotation matrix, and t is the translation vector. It is a non-zero scale factor, also known as the projection depth.

[0027] This projection enables each attached point to acquire its corresponding texture and color attributes, thus achieving attribute binding between three-dimensional geometry and two-dimensional texture.

[0028] For each dependent domain point after projection Calculate its multimodal features and construct a local cross-modal feature set describing the characteristics of the point and its neighborhood. This feature set is composed of the point's normal vector, curvature, reflection intensity, and image texture gradient feature vector, where the normal vector... By calculating this point (usually taken) The covariance matrix of the nearest neighbor point cloud Then perform eigenvalue decomposition, and the eigenvector corresponding to the smallest eigenvalue is the normal vector. .

[0029]

[0030] in, For point of Nearest neighbor set The first in The three-dimensional coordinates of the points Set of nearest neighbors The mean of the three-dimensional coordinates.

[0031] Based on the above covariance matrix eigenvalues Calculate curvature The greater the curvature, the more pronounced the surface change at that point (sharp, convex):

[0032] Reflection intensity The reflection intensity attribute and image texture gradient are directly derived from the laser point cloud. To capture the color or texture differences between the attached object and the tower surface, the pixel coordinates are obtained by projecting the attachment points onto the image plane. Use the Sobel operator to calculate the grayscale or color channel gradient:

[0033] in, For pixel grayscale or color intensity, and These represent the gradients along the horizontal and vertical directions, respectively, and the image texture gradient. The larger the value, the more obvious the local texture changes, and the possible presence of attached substances.

[0034] Through the above steps, for each point This yields a local cross-modal feature set. .

[0035] Specifically, the attachment sensitivity of the cross-modal feature group is judged to screen primary attachment evidence, which includes: The intrinsic coupling relationship of the local cross-modal feature group refers to the inherent correlation between points in the neighborhood in terms of space, direction, and physical attributes, and this is a comprehensive coupling weight. To quantify this, the weight is the product of the following three types of weights, specifically:

[0036]

[0037]

[0038]

[0039] in, Spatial proximity coupling weights, derived from points and The Euclidean distance between them determines, It is a scale parameter that controls the rate of spatial coupling decay. The geometric coupling weights are determined by... and normal vector and The dot product determines, For attribute similarity coupling weights, by point and eigenvectors and The Euclidean distance determines, The normalization coefficient is... , .

[0040] Based on the calculated integrated endogenous coupling weights, for the point Calculate a local perturbation vector This vector is the weighted sum of the deviations between the feature vectors of all points in the neighborhood and the feature vector of the center point; the weights are the comprehensive coupling weights. The set of local perturbation vectors of all points within the entire dependent domain constitutes the local perturbation field. After constructing the local perturbation field, the eigenvalues ​​in the local cross-modal feature group are driven to evolve within the local perturbation field. Specifically, feature changes are simulated iteratively, and the sequence of changes of each feature component with the number of iterations is recorded. This sequence is the dynamic response trajectory of the feature, and the dynamic response trajectories of all features are jointly measured. Those features whose change amplitude is greater than a preset first threshold are selected. The direction of change is consistent with that of at least one other feature (by calculating the Pearson correlation coefficient and determining that it is greater than a preset second threshold). High-sensitivity features constitute a subset of highly sensitive features. From this subset, further selection is made of feature points that maintain a monotonic and stable trend throughout the entire perturbation evolution process (e.g., the Spearman rank correlation coefficient of their response trajectories). Greater than the preset third threshold These feature points serve as reliable primary attachment evidence output.

[0041] All primary attachment evidence within the attachment domain is spatially aggregated for consistency. A multimodal evidence set is formed by evaluating directional consistency, connectivity, and cross-modal overlap among the evidence. Directional consistency refers to analyzing the local surface orientation of each primary attachment evidence point. Specifically, for a given evidence point, the system simultaneously checks whether the surface orientation of other evidence points in its surrounding area maintains the same or similar trend as that point's orientation. If, within the same local space, the surface orientation of most evidence points exhibits similar variation patterns, it indicates that these evidence points are more likely to belong to the same attached surface structure; otherwise, they may simply be single noise points or isolated interference points.

[0042] Regarding spatial connectivity, the set of neighboring points within the attachment domain for each evidence point is first determined, and it is then determined whether these points can be connected to each other through continuous spatial paths. If an evidence point can be connected to a larger group of evidence points through neighboring points, it indicates that the group of evidence points may originate from the same attached entity; while isolated points that lack continuous spatial connections with other evidence points are considered more likely to be false evidence caused by noise, stray reflections, or missegmentation of point clouds.

[0043] Regarding cross-modal overlap, responses sensitive to attachments in different modalities (such as point cloud shape features, reflection intensity features, and image texture features) are compared. For example, for the same primary evidence point, if its geometric features (such as curvature changes), laser reflection intensity features, and image texture gradient features all simultaneously show anomalies or abrupt changes, it indicates that multiple modalities jointly point to the presence of attachments at that point, and its credibility is significantly higher than that of a point that only shows anomalies in a single modality. Conversely, if there is a lack of consistency in the trends of change across different modalities, the point is more likely to be a misjudgment.

[0044] The system jointly evaluates all primary attachment evidence based on the comprehensive consistency performance of three dimensions: directional consistency, connectivity, and cross-modal overlap. Points that exhibit coordinated changes in direction, maintain coherent distribution in space, and jointly demonstrate prominent features in multimodal characteristics will eventually be aggregated into the same attachment region to form a multimodal evidence set.

[0045] S3, based on a multimodal evidence set, constructs a diagram representing the relationship between the attachments and tower components, analyzes the diagram, and outputs the analysis results.

[0046] In this embodiment, the construction of a graph representing the relationship between the attachment and the tower components based on a multimodal evidence set, wherein the nodes of the graph represent attachment points and the edges represent attachment paths, specifically: Based on multimodal evidence set Constructing a potential field of dependency contacts within the dependency domain. The value of the potential field at any point of attachment is formed by the sum of the contributions of all primary attachment evidence points within a certain neighborhood of that point, and the density is estimated using a Gaussian kernel function. After calculating the potential field, a local extremum search is performed on the potential field to extract all potential field peak points as candidate attachment points.

[0047] For each candidate attachment point, multimodal evidence combinations within its neighborhood are extracted for local causal inference to determine whether the evidence combination is caused by genuine "attachment," rather than the tower structure itself or noise. Specifically, a PC algorithm based on conditional independence testing is used, defining three key variables for causal analysis and processing these variables. It is a binary variable representing whether there are attachments at that point. This indicates the presence of attached substances. Indicates the absence of attachments; outcome variable That is, the evidence combination at this point is a multidimensional vector composed of multiple features (such as anomalous geometry, reflectivity, and texture); confounding variables These are a set of context variables that can simultaneously influence both the processing variables and the outcome variables. Examples include component type (insulator strings are inherently more prone to generating complex geometric signals than smooth tower bodies) and lighting conditions (shadows at certain angles can create textures resembling foreign objects in images). Testing is conducted on given confounding variables... Time processing variables With outcome variable Whether conditional independence (significance level) If the test result rejects the conditional independence hypothesis, then a causal effect from "attachment" to the evidence set is considered to exist, thus identifying the candidate attachment contact as a formal attachment node. .

[0048] The construction of the virtual adhesion field based on the evidence flow direction between the attached nodes is specifically as follows: Nodes are calculated using weighted cosine similarity. and The similarity of the corresponding evidence combinations is the node. and Coupling consistency between Based on the coupling consistency of multimodal characteristics between nodes, a field source characterizing the node association strength is constructed. Using the coupling consistency as an attractive force source and the spatial dispersion between nodes as a repulsive force source, a dynamic, non-uniform virtual adhesion field is synthesized. The spatial dispersion... That is, the three-dimensional Euclidean distance between nodes. For any point in space, the "adhesion force" it experiences is determined by the combined attraction and repulsion from all nodes. right The direction of the generated force is from point to The unit vector, its intensity Coupling consistency and spatial dispersion Functions:

[0049] in, It is the repulsive force intensity coefficient. The direction of the force line is determined by the direction of the resultant force.

[0050] In the virtual adhesion force field, starting from each node in the attached diagram, pathline integration is performed using the fourth-order Runge-Kutta method (RK4) to simulate the force field streamlines. These streamlines constitute the candidate adhesion paths connecting different nodes, forming a candidate edge set. .

[0051] The current node set and candidate edge sets The resulting graph topology mapping is an energy function that comprehensively considers data fit, path smoothness, and topological simplicity. It minimizes the overall energy through optimization algorithms such as graph cutting or Monte Carlo, eliminates redundant and unreliable edges, and finally obtains the final graph.

[0052] The analysis of the attached diagram and the output of the analysis results are as follows: The analysis results include adhesion type, adhesion path, adhesion depth, and affected components.

[0053] The final attached graph is matched against predefined attachment mechanism templates. These templates define typical graph patterns (e.g., star-shaped, chain-like, network-like) for different types of attachments (e.g., bird nests, vines, hanging objects). By calculating the similarity of the graph structures (as shown by the kernel function), the best-matching attachment category is identified. The edges retained in the final attached graph after energy minimization optimization are directly parsed into vector line segments in three-dimensional space, i.e., attachment paths. A series of points are sampled along these attachment paths in three-dimensional space. For each sampled point, the evidence strength (e.g., potential field value) at the corresponding location in the multimodal evidence set is queried. (Or feature similarity). The attachment depth is determined by statistical analysis of these strength values ​​(such as taking the maximum value or integrating). The spatial relationship between the three-dimensional coordinates of each node in the attached figure and the semantic three-dimensional model of the tower component is determined. By calculating the nearest distance from the node to the surface of each component model, the node is mapped to the nearest component, thereby determining the affected component. For example, if a node is closest to the "insulator string" model, then the insulator string is determined to be the affected component.

[0054] Example 2, Figure 2The present invention provides a system for identifying the tower body and surrounding attachments by integrating visualized data, comprising: The identification building unit is used to identify and extract at least one surface area containing a suspected attachment from the three-dimensional point cloud data of the tower, as the attachment area; and to construct an attachment domain in three-dimensional space centered on the attachment area for data fusion analysis. The feature fusion unit is used to acquire multi-source visualization data of the tower and map the multi-source visualization data to the dependency domain for feature fusion to form a unified multimodal evidence set; The analysis output unit is used to construct a diagram representing the relationship between the attachments and tower components based on a multimodal evidence set, and to parse the diagram to output the analysis results.

[0055] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on 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.

[0056] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0057] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0058] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0059] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0060] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying the tower body and surrounding attachments by integrating visualized data, characterized in that, Includes the following steps: From the three-dimensional point cloud data of the tower, identify and extract at least one surface area where a suspected attachment is located as the attachment area; Centered on the attached region, a dependency domain is constructed in three-dimensional space for data fusion analysis; Multi-source visualization data of the tower is acquired and mapped to the attachment domain for feature fusion to form a unified multimodal evidence set. Specifically, the following steps are taken: the image region corresponding to the attachment domain is extracted, and the image region is spatially projected and matched with the attachment domain so that each attachment domain point obtains its corresponding texture and color attributes; for the projected attachment domain points, local cross-modal feature groups are calculated and constructed based on the point's normal vector, curvature, reflection intensity, and image texture gradient; the cross-modal feature groups are subjected to attachment sensitivity discrimination to screen primary attachment evidence; all primary attachment evidence in the attachment domain is spatially consistent and aggregated, and a multimodal evidence set is formed by evaluating the directional consistency, connectivity, and cross-modal overlap between the evidences. The screening of primary attachment evidence includes: constructing a local perturbation field for the attachment region based on the endogenous coupling relationship of local cross-modal feature groups, and tracking the response trajectory of each feature under the perturbation field; determining a subset of features that are highly sensitive to attachment behavior by jointly measuring the change amplitude and direction consistency of the response trajectory; and outputting the features of this subset that maintain a monotonic evolution trend before and after the perturbation as primary attachment evidence. Based on a multimodal evidence set, a diagram representing the relationship between the attachments and tower components is constructed, and the diagram is analyzed to output the analysis results.

2. The method for identifying the tower body and surrounding attachments by integrating visualized data according to claim 1, characterized in that, The method for acquiring the three-dimensional point cloud data includes using lidar, structured light, or multi-view stereo photography to acquire three-dimensional point cloud data of the tower and its surrounding environment.

3. The method for identifying the tower body and surrounding attachments by integrating visualized data according to claim 2, characterized in that, The construction of a dependency domain for data fusion analysis in three-dimensional space, centered on the attachment region, specifically involves: Semantic segmentation is performed on the overall 3D point cloud data of the tower to identify and extract point cloud clusters of key components; Based on the point cloud clusters, the main structural lines characterizing the main attachment directions of the components are obtained through three-dimensional geometric fitting. Based on the spatial location of the main structural line and its connection relationship with the main tower body, the core point of the attachment area is determined, and the corresponding attachment domain is generated according to the component type and structural characteristics with the core point as the spatial center.

4. The method for identifying the tower body and surrounding attachments by integrating visualized data according to claim 3, characterized in that, The key components include insulator strings, equalizing rings, conductor suspension hardware, and crossarms.

5. The method for identifying the tower body and surrounding attachments by integrating visualized data according to claim 4, characterized in that, The graph, based on a multimodal evidence set, constructs a graph representing the relationship between the attachments and tower components. The nodes in the graph represent attachment points, and the edges represent attachment paths. Specifically: Based on a multimodal evidence set, a potential field for dependency contacts is constructed within the dependency domain, and the peak point of the potential field is used as a candidate dependency contact. Local causal inference is performed on the evidence combination of candidate attachment points to determine the attachment node; A virtual adhesion force field is constructed based on the evidence flow direction between nodes in the attached diagram, and candidate adhesion paths are generated by path line integration. The graph topology formed by the nodes in the attached graph and the candidate attachment paths is mapped to an energy function, and the final attached graph is obtained by minimizing the overall energy.

6. The method for identifying the tower body and surrounding attachments by integrating visualized data according to claim 5, characterized in that, The construction of the virtual adhesion field based on the evidence flow direction between the attached nodes is specifically as follows: Based on the coupling consistency of multimodal features between nodes, a field source characterizing the strength of node association is constructed; Using the aforementioned coupling consistency as the attraction source and the spatial dispersion between nodes as the repulsion source, a dynamic non-uniform virtual adhesion field is synthesized. In the virtual adhesion force field, the direction of the force lines is determined by the dominant change direction of the coupling consistency, and the intensity of the force lines is a function of the intensity of the coupling consistency and the spatial dispersion.

7. The method for identifying the tower body and surrounding attachments by integrating visualized data according to claim 6, characterized in that, The analysis of the attached diagram and the output of the analysis results are as follows: The attachment type will be identified by matching the attached diagram with a predefined attachment mechanism template. The edges in the final attached diagram, optimized by minimizing energy, are resolved into attachment paths; Along the attachment path, query the spatial distribution of evidence strength in the multimodal evidence set to determine the attachment depth; The nodes in the final attached diagram are mapped to the semantic 3D model of the tower components to determine the affected components.

8. A system for identifying the pole body and surrounding attachments using the method for integrating visual data as described in any one of claims 1-7, characterized in that, include: The identification building unit is used to identify and extract at least one surface area containing a suspected attachment from the three-dimensional point cloud data of the tower, as the attachment area; Centered on the attached region, a dependency domain is constructed in three-dimensional space for data fusion analysis; The feature fusion unit is used to acquire multi-source visualization data of the tower and map the multi-source visualization data to the dependency domain for feature fusion to form a unified multimodal evidence set; The analysis output unit is used to construct a diagram representing the relationship between the attachments and tower components based on a multimodal evidence set, and to parse the diagram to output the analysis results.

Images