Method and system for attribute assignment of tower fittings based on point cloud model

Abstract

The application discloses a kind of based on point cloud model's tower hardware's attribute empowerment method and system, including following steps: the three-dimensional point cloud data of tower scene is collected, and pre-processing is executed;Extract target point cloud area, and execute semantic segmentation and instance extraction, generate hardware point cloud set and corresponding geometric attribute vector;Extract the feature of each hardware point cloud, construct multi-modal feature tensor, execute attribute prediction;If hardware point cloud exists anomaly, carry out point cloud diffusion completion, reconstruct missing area, generate the hardware point cloud after completion and infer semantic attribute label;Infer future operation risk, generate behavior attribute label;Structured attribute data is constructed, and structured attribute data is mapped to the corresponding structure node of digital tower.The application fuses three-dimensional point cloud recognition and behavior prediction technology, realizes tower hardware attribute intelligent empowerment, with the advantages of high recognition precision, strong risk perception, complete empowerment process.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of digital modeling of power equipment, and in particular to a method and system for attribute empowerment of tower fittings based on a point cloud model. BACKGROUND

[0002] In a power transmission system, tower fittings, as key components connecting conductors and tower structures, directly affect the safety and stability of the transmission line. However, in the traditional transmission tower operation and maintenance system, the acquisition of the properties and state assessment of the fittings mainly rely on manual inspection, two-dimensional image recognition or low-precision structure modeling. Such methods are limited by factors such as light, angle of view and human experience, making it difficult to achieve high-precision modeling and all-around attribute recognition of the fittings, especially in complex tower structures or high-altitude environments, where the recognition accuracy and work safety are significantly insufficient.

[0003] In the prior art, although some studies attempt to use three-dimensional point cloud data to model the tower structure, they mainly focus on overall shape reconstruction and tower body identification, lacking the ability to extract, model and analyze the properties of the fittings components in detail. Moreover, the fitting structure in the point cloud data is easily affected by occlusion, sparse sampling and boundary ambiguity, resulting in low material recognition accuracy and difficulty in determining the aging state. More importantly, the current technology has not established a multi-level information empowerment mechanism that integrates geometric properties, semantic properties and behavior properties, lacking the ability to predict changes in the operating state of the fittings, and unable to achieve early warning and intelligent management of potential failure risks. SUMMARY

[0004] One object of the present application is to provide a method and system for attribute empowerment of tower fittings based on a point cloud model. The present application integrates three-dimensional point cloud recognition and behavior prediction technology to achieve intelligent empowerment of tower fitting attributes, with the advantages of high recognition accuracy, strong risk perception and complete empowerment process.

[0005] According to an embodiment of the present application, a method for attribute empowerment of tower fittings based on a point cloud model comprises the following steps:

[0006] Collect three-dimensional point cloud data of the tower scene and perform preprocessing on the collected three-dimensional point cloud data to generate standardized point cloud data;

[0007] Divide the standardized point cloud data into spatial regions, extract the target point cloud region, and perform semantic segmentation and instance extraction on the target point cloud region to generate a set of fitting point clouds and corresponding geometric attribute vectors;

[0008] For each fitting point cloud, extract the reflection intensity feature, point cloud density feature and edge blur feature, construct a multi-modal feature vector, input it into a material recognition model to perform attribute prediction, and generate a semantic attribute label;

[0009] Determine whether the hardware point cloud has structural missing, scanning occlusion, or morphological abnormalities. If so, perform point cloud diffusion completion, reconstruct the missing area, generate the completed hardware point cloud, and infer semantic attribute labels;

[0010] Use the behavior prediction network to infer future operational risks and generate behavior attribute labels;

[0011] The geometric attribute vectors, semantic attribute labels, completed semantic attribute labels and behavioral attribute labels of the hardware point cloud are integrated to construct structured attribute data, and the structured attribute data is mapped to the corresponding structural nodes of the digital tower.

[0012] Optionally, the preprocessing includes noise removal, spatial downsampling and coordinate registration.

[0013] Optionally, the generation of the hardware point cloud set and the corresponding geometric attribute vectors specifically includes:

[0014] Perform spatial region partitioning on the standardized point cloud data, dividing the entire point cloud space into three-dimensional voxel grids with fixed side lengths, and counting the number of points in each voxel grid. If the number of points is greater than a preset density threshold, the corresponding voxel grid is marked as the target point cloud area;

[0015] In the target point cloud area, a local neighborhood point set is constructed for each point, and local geometric features are extracted based on the neighborhood information;

[0016] The local geometric features of each point are concatenated to form a feature vector set, a feature map of the target point cloud area is constructed, and the feature map is input into the point cloud recognition network to generate a semantic label set;

[0017] All points marked as hardware are filtered from the semantic label set to form a hardware semantic point set. A density-connectivity-based clustering method is used to perform instance extraction on the hardware semantic point set to generate multiple hardware point clouds.

[0018] An axis-aligned three-dimensional bounding box is constructed for each hardware point cloud, and the spatial geometric attribute parameters of the three-dimensional bounding box are extracted and combined to generate a geometric attribute vector.

[0019] Optionally, performing the instance extraction operation on the hardware semantic point set specifically includes:

[0020] Set the minimum neighborhood point count threshold and the distance threshold between points. For each semantic point in the hardware semantic point set, search for a neighborhood semantic point set within the distance threshold, and count the number of semantic points in the neighborhood. If the number of semantic points in the neighborhood of a semantic point is greater than or equal to the minimum neighborhood point count threshold, then use the semantic point as a cluster seed semantic point to construct the initial cluster subset.

[0021] From the initial cluster sub-set, continuously search for semantic points that meet the inter-point distance threshold in its neighborhood, and add the semantic points that meet the conditions to the current cluster sub-set, while repeating the neighborhood search process until the cluster sub-set no longer expands;

[0022] Select a new cluster seed semantic point from the remaining semantic points that do not participate in clustering, repeat the above process, and finally obtain multiple structurally coherent and mutually independent cluster sub-sets, and take the cluster sub-set as the fitting point cloud.

[0023] Optionally, the generation of the semantic attribute label specifically includes:

[0024] For each fitting point cloud, respectively extract the reflection intensity feature, the point cloud density feature and the edge blur feature;

[0025] The reflection intensity feature, the point cloud density feature and the edge blur feature are normalized, spliced into a feature vector in a fixed order, and a corresponding multi-modal feature vector is constructed for each fitting point cloud;

[0026] The multi-modal feature vector is input into the material recognition model to perform deep feature extraction and classification inference, and the corresponding semantic attribute label is output.

[0027] Optionally, the generating the completed fitting point cloud and inferring the semantic attribute label specifically includes:

[0028] Performing structural integrity detection on each fitting point cloud to determine whether there is structural loss, occluded area or abnormal morphology, if there is any abnormality, the fitting point cloud is determined as an abnormal state point cloud;

[0029] For the abnormal state point cloud, the diffusion direction and diffusion area are determined using the normal vector distribution and local curvature of the edge point, and the completion point is gradually generated along the main structure direction to form a candidate completion point cloud set;

[0030] The candidate completion point cloud set and the adjacent fitting point cloud are analyzed for structural consistency, and the abnormal points that do not meet the spatial connection logic are filtered out, and the effective points with good topological connectivity and smooth spatial curvature transition are retained to generate the first stage completion point cloud;

[0031] The first stage completion point cloud and the original fitting point cloud are fused to form a preliminary completion fitting point cloud, and a semantic attribute label generation operation is performed to output a preliminary completion point cloud semantic attribute label set;

[0032] The preliminary completion fitting point cloud and the semantic attribute label set are input into the secondary generation mechanism of the completion point cloud semantic attribute label, and a semantic re-labeling process based on context consistency optimization is performed;

[0033] For the boundary region in the preliminary completed accessory point cloud, a neighborhood point set with a fixed radius is constructed for each accessory semantic point. All semantic points in the neighborhood point set are traversed, and the number distribution of each semantic attribute label is counted. If there are two or more different types of semantic attribute labels in the neighborhood, and the semantic attribute labels of adjacent semantic points frequently switch, the semantic point is marked as a semantic boundary candidate point;

[0034] Among all the semantic boundary candidate points, a label mutation point is further identified.

[0035] The semantic attribute label set in the neighborhood of the label mutation point is extracted, the label distribution density is counted according to the frequency of the label, the semantic attribute label with the highest frequency is selected as the corrected label, and the corrected label is assigned to the current mutation point to replace the original semantic attribute label.

[0036] Finally, a complete point cloud semantic attribute label set with contextual coherence and semantic consistency is generated.

[0037] Optionally, the semantic re-labeling process specifically includes:

[0038] For each preliminary completed point in the preliminary completed accessory point cloud, a neighborhood point set with a fixed radius centered on the preliminary completed point is constructed, the semantic attribute label types and corresponding numbers of all points in the neighborhood point set are counted, and a neighborhood label distribution vector is generated.

[0039] The number of occurrences of the semantic attribute label of the current completed point in the neighborhood label distribution vector is divided by the total number of neighborhood points to obtain the confidence score of the semantic attribute label.

[0040] If the confidence score is less than a pre-set confidence threshold, it is determined that the current semantic label does not have neighborhood consistency, and is marked as a candidate re-labeling point.

[0041] Further, a neighborhood accessory semantic point set is constructed within the three-dimensional spatial neighborhood range of each candidate re-labeling point. For each point in the neighborhood accessory semantic point set, the current semantic attribute label and the corresponding confidence score are extracted to form a data set containing label categories and confidence scores. According to the confidence score corresponding to each semantic attribute label as the voting weight, all semantic attribute labels in the three-dimensional spatial neighborhood are weighted and counted, and the weighted total value of each type of semantic attribute label is calculated. Among all the semantic attribute labels participating in the voting, the label category with the largest weighted total value is selected as the final semantic attribute label of the candidate re-labeling point.

[0042] Optionally, the weighted total value is equal to the sum of the confidence scores of all accessory points with semantic attribute labels in the neighborhood.

[0043] Optionally, the behavior attribute label specifically includes:

[0044] collecting a history time sequence posture of each fitting point cloud, the history time sequence posture including a direction vector of a fitting space orientation, a centroid position change sequence, and a connection angle change amount;

[0045] extracting a spatial connection state feature between the fitting point cloud and the tower structure;

[0046] splicing the preprocessed history time sequence posture sequence and the spatial connection state feature to construct a joint tensor;

[0047] inputting the joint tensor into a behavior prediction network based on an Informer structure, performing multi-scale attention encoding and key feature aggregation within a time window, inferring a running stability change trend of a target fitting in a future time interval, and generating a behavior attribute label.

[0048] According to an attribute enabling system for a tower fitting based on a point cloud model, the system comprises:

[0049] a data acquisition module configured to acquire three-dimensional point cloud data of a tower scene and perform preprocessing to generate standardized point cloud data;

[0050] a region extraction module configured to divide a spatial region of the standardized point cloud data and extract a target point cloud region;

[0051] a semantic recognition module configured to perform semantic segmentation and instance extraction on the target point cloud region to generate a fitting point cloud set and a semantic label;

[0052] a geometric feature extraction module configured to generate a geometric attribute vector;

[0053] a semantic attribute recognition module configured to recognize a fitting material type and an aging grade based on a multi-modal feature vector to generate a semantic attribute label;

[0054] a structure integrity detection module configured to determine whether the fitting point cloud has a structural anomaly and recognize an abnormal state point cloud;

[0055] a point cloud completion module configured to perform point cloud diffusion and structure-guided completion operations on the abnormal state point cloud to generate a completed fitting point cloud;

[0056] a semantic label optimization module configured to perform semantic label consistency optimization on the completed fitting point cloud to improve label accuracy and continuity;

[0057] a structure stability verification module configured to analyze the stability of a connection structure of the fitting to determine whether the connection angle and stress are reasonable;

[0058] a behavior prediction module configured to infer a future running risk based on a history posture and a connection state to generate a behavior attribute label;

[0059] An attribute mapping module is configured to integrate various attribute labels and map them to corresponding structure nodes of the digital tower model, thereby completing the attribute assignment of the fittings.

[0060] The present application has the following advantages:

[0061] Firstly, the present application provides a method for attribute assignment of tower fittings based on a point cloud model, which fully utilizes the spatial structure information of three-dimensional point cloud data to realize the full-process automatic processing of tower fittings from geometric modeling, semantic understanding to behavior prediction, thereby improving the fitting recognition accuracy, material classification accuracy and operation risk assessment capability. By introducing the spatial region division, point cloud semantic segmentation and instance extraction mechanism, the fitting instances can be effectively separated and the key geometric attributes can be extracted in the dense and complex tower structure, thereby laying a solid foundation for subsequent attribute recognition and behavior modeling. Meanwhile, the material recognition model is constructed by fusing multi-modal features, thereby improving the distinguishing ability of different metal components and aging states and realizing the precise assignment of material types and aging levels.

[0062] Secondly, in terms of processing abnormal fitting structures, a structure-guided point cloud diffusion completion mechanism is introduced to dynamically generate the missing areas in combination with the main direction vector of the fitting, the normal consistency and the curvature information, thereby effectively reconstructing the structure defect point cloud caused by occlusion or fracture. In addition, a secondary generation mechanism of the semantic attribute label of the completed point cloud and a context consistency optimization strategy are introduced, thereby improving the spatial continuity and semantic stability of the semantic label in the completed area and solving the problems of boundary label confusion and label jumping in the traditional method.

[0063] In addition, the present application designs a behavior prediction mechanism for the operation situation, constructs an Informer network inference model by comprehensively considering the historical time sequence posture and spatial connection state features, accurately captures the future failure trend, connection slip and poor contact risk, and the introduction of the behavior attribute label enables the digital tower to not only have the static attribute description capability, but also have the prediction and response capability of the dynamic operation state, thereby enhancing the active perception and decision support capability of the digital twin system in the equipment-level operation and maintenance management. In summary, the present application effectively breaks through the complete chain from point cloud data to equipment attributes and from static recognition to dynamic evaluation, thereby promoting the intelligentization, automation and refinement development of the tower fitting management. BRIEF DESCRIPTION OF DRAWINGS

[0064] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application together with the embodiments thereof, and explain the present application, but do not constitute a limitation of the present application. In the drawings:

[0065] Figure 1 A flowchart of a method for attribute assignment of tower fittings based on a point cloud model according to the present application is shown. DETAILED DESCRIPTION

[0066] The application will be described in further detail below with reference to the drawings. These drawings are simplified schematic diagrams and only show the basic structure of the application in a schematic manner, and thus only show the components relevant to the application.

[0067] Reference Figure 1 A method for assigning attributes to tower fittings based on a point cloud model, comprising the following steps:

[0068] Collect three-dimensional point cloud data of a tower scene, and perform preprocessing on the collected three-dimensional point cloud data to generate standardized point cloud data;

[0069] Divide the standardized point cloud data into spatial regions, extract a target point cloud region, and perform semantic segmentation and instance extraction on the target point cloud region to generate a fitting point cloud set and a corresponding geometric attribute vector;

[0070] For each fitting point cloud, extract reflection intensity features, point cloud density features, and edge blur features, construct a multi-modal feature vector, input it into a material recognition model to perform attribute prediction, and generate a semantic attribute label;

[0071] Determine whether the fitting point cloud has structural defects, scanning occlusion, or abnormal morphology, and if so, perform point cloud diffusion completion, reconstruct the missing area, generate a completed fitting point cloud, and infer the semantic attribute label;

[0072] Use a behavior prediction network to infer future operation risks and generate a behavior attribute label;

[0073] Integrate the geometric attribute vector, semantic attribute label, completed semantic attribute label, and behavior attribute label of the fitting point cloud, construct structured attribute data, and map the structured attribute data to the corresponding structure node of the digital tower to complete the attribute assignment of the tower fittings.

[0074] In this embodiment, the preprocessing includes noise removal, spatial downsampling, and coordinate registration.

[0075] In this embodiment, the generation of the fitting point cloud set and the corresponding geometric attribute vector specifically includes:

[0076] Perform spatial region division processing on the standardized point cloud data, divide the overall point cloud space into a three-dimensional voxel grid with a fixed edge length, count the number of points in each voxel grid, and if the number of points is greater than a preset density threshold, mark the corresponding voxel grid as a target point cloud region;

[0077] In the target point cloud region, construct a local neighborhood point set for each point, extract local geometric features based on neighborhood information, including:

[0078] The normal vector consistency feature represents the average value of the included angle between the normal vectors of the current point and the points in the neighborhood, and reflects the consistency of the normal direction.

[0079] The curvature feature represents the ratio between the minimum eigenvalue of the local covariance matrix and the sum of all eigenvalues, and characterizes the bending degree of the local surface.

[0080] The edge jump degree feature represents the average amplitude of the gradient change of the spatial distance between the neighborhood points, and is used to identify the sensitivity of the boundary region.

[0081] The generation of the curvature feature comprises:

[0082] For each point in the target point cloud region, a neighborhood point set with a fixed radius as the neighborhood range is constructed.

[0083] For each point, the average value of all three-dimensional coordinates in the neighborhood is calculated to obtain the spatial barycenter of the neighborhood points.

[0084] The difference between the coordinates of each neighborhood point and the barycenter coordinates is combined according to the three coordinate axis directions to construct a set of local spatial offset vectors.

[0085] Based on the set of local spatial offset vectors, the degree of cooperative change between the directions is calculated to form a local covariance matrix, which is used to describe the discrete directionality and distribution trend of the neighborhood points in the three-dimensional space.

[0086] The local covariance matrix is subjected to eigenvalue decomposition, the eigenvalues are arranged in descending order, the minimum eigenvalue is extracted, and the ratio between the minimum eigenvalue and the sum of all eigenvalues is calculated to obtain the curvature feature of the current point, which describes the bending degree of the local point cloud surface.

[0087] The local geometric features of each point are spliced to form a feature vector set, a feature map of the target point cloud region is constructed, and the feature map is input into a point cloud recognition network to generate a semantic label set, the point cloud recognition network uses a PointNet++ network to perform a semantic segmentation operation on the target point cloud region, and outputs the semantic label set, the semantic label set includes fittings, towers, conductors and backgrounds.

[0088] All points labeled as fittings are filtered from the semantic label set to form a fitting semantic point set, a density connectivity-based clustering method is used to perform instance extraction on the fitting semantic point set to generate multiple fitting point clouds.

[0089] For each fitting point cloud, an axis-aligned three-dimensional bounding box is constructed, and the spatial geometric attribute parameters of the three-dimensional bounding box are extracted to form a geometric attribute vector, the spatial geometric attribute parameters include a bounding box size vector, a main direction vector, a point cloud density indicator and a barycenter coordinate.

[0090] The construction of the three-dimensional bounding box comprises:

[0091] For each fitting point cloud, the minimum coordinate value and the maximum coordinate value of all points in each coordinate axis direction in the three-dimensional space are counted to determine the spatial position of the bounding boundary respectively;

[0092] According to the distance between the minimum value and the maximum value of each coordinate axis direction, the boundary span of the fitting point cloud in each direction is calculated;

[0093] Based on the boundary span and the boundary point position of the three directions, an axis-aligned three-dimensional bounding box consistent with the three-dimensional coordinate axis direction is constructed, so that the fitting point cloud is completely wrapped inside the three-dimensional bounding box.

[0094] In the embodiment, the instance extraction operation performed on the fitting semantic point set specifically comprises:

[0095] A minimum neighborhood point threshold and a point distance threshold are set, for each semantic point in the fitting semantic point set, a neighborhood semantic point set within the point distance threshold range is searched, and the number of semantic points in the neighborhood is counted, if the number of semantic points in the neighborhood of a certain semantic point is greater than or equal to the minimum neighborhood point threshold, the semantic point is taken as a clustering seed semantic point, and an initial clustering sub-set is constructed;

[0096] Starting from the initial clustering sub-set, the neighborhood of the semantic points satisfying the point distance threshold is continuously searched, and the semantic points satisfying the condition are added to the current clustering sub-set, and the neighborhood search process is repeated, until the clustering sub-set is no longer expanded;

[0097] A new clustering seed semantic point is selected from the remaining semantic points not participating in clustering, and the above process is repeated, and finally a plurality of structurally coherent and mutually independent clustering sub-sets are obtained, and the clustering sub-set is taken as the fitting point cloud.

[0098] In the embodiment, the generation of the semantic attribute label specifically comprises:

[0099] For each fitting point cloud, the reflection intensity feature, the point cloud density feature and the edge blur feature are extracted respectively, the reflection intensity feature represents the signal intensity returned by the laser on the fitting surface, the point cloud density feature reflects the distribution density of points in a unit space, and the edge blur feature measures the blur degree of the fitting boundary through the change degree of the neighborhood normal vector;

[0100] The reflection intensity feature, the point cloud density feature and the edge blur feature are normalized, spliced into a feature vector in a fixed order, and a corresponding multi-modal feature vector is constructed for each fitting point cloud;

[0101] The multi-modal feature vector is input into a material recognition model to perform deep feature extraction and classification inference, and a corresponding semantic attribute label is output, the material recognition model adopts a multi-layer perceptron network, and the semantic attribute label includes a material type and an aging level, the material type includes cast iron, hot-dip galvanized steel, aluminum alloy, copper alloy, stainless steel and synthetic composite material, and is used to represent the structural material composition of the fitting, and the aging level includes no aging, mild aging, moderate aging, severe aging and failure, and is used to represent the physical degradation degree of the fitting in long-term operation.

[0102] In the embodiment, the generating the completed fitting point cloud and inferring the semantic attribute label specifically includes:

[0103] The structural integrity of each fitting point cloud is detected to determine whether there is structural loss, occluded area or abnormal morphology, and if any of the abnormalities exist, the fitting point cloud is determined as an abnormal state point cloud, and the structural integrity detection is based on the spatial density continuity, the main direction vector stability and the three-dimensional bounding box closure of the point cloud;

[0104] For the abnormal state point cloud, the diffusion direction and diffusion area are determined by using the normal vector distribution and local curvature of the edge point, and the candidate completion point cloud set is formed by guiding the completion point to generate gradually along the structural main direction;

[0105] The generation of the candidate completion point cloud set includes:

[0106] An edge point set is extracted from the current abnormal state fitting point cloud, the edge point refers to a boundary point whose number of points in the neighborhood range of the fitting point cloud is less than a preset density threshold, or the local curvature value exceeds a preset curvature anomaly threshold, or the angle between the normal vector and the average neighborhood normal vector is greater than a set angle threshold;

[0107] The local normal vector of each edge point is calculated, the local normal vectors of all edge points are combined into a statistical set, and the principal component analysis method is used to analyze and identify the main direction component in the statistical set as the structural main direction of the edge point set;

[0108] Around each edge point, a fixed diffusion step is defined based on the structural main direction, and a multi-round point expansion operation is performed along the structural main direction, in each round of expansion, the position of a new completion point is calculated according to the average normal direction and the curvature gradient of the current edge point, and a candidate completion point cloud set is formed;

[0109] The position of the new completion point calculated according to the average normal direction and the curvature gradient of the current edge point includes:

[0110] For the current edge point, the normal vector of each point in the neighborhood is calculated, the average of all normal vectors is taken as the average normal direction of the current edge point, the diffusion step is set, and the position of the new completed point is defined as a shift in the weighted synthesis direction of the average normal direction and the curvature gradient direction, specifically: taking the current edge point as the center, shifting the distance Δ along the weighted synthesis direction to generate the position coordinates of the new completed point, the generation direction of the new completed point satisfies the structure continuity and curvature smoothness constraint, and the Δ is a preset diffusion step;

[0111] The candidate completed point cloud set is analyzed for structural consistency with the adjacent fitting point cloud, abnormal points that do not satisfy the spatial connection logic are screened out, effective points with good topological connectivity and smooth spatial curvature transition are retained, and a first-stage completed point cloud is generated;

[0112] The first-stage completed point cloud is fused with the original fitting point cloud to form a preliminary completed fitting point cloud, and a semantic attribute label generation operation is performed once to output a preliminary completed point cloud semantic attribute label set, and the semantic label set covers all fitting point clouds;

[0113] The preliminary completed fitting point cloud and the semantic attribute label set are input into a completed point cloud semantic attribute label secondary generation mechanism to perform a semantic re-labeling process based on context consistency optimization;

[0114] For the boundary region in the preliminary completed fitting point cloud, a neighborhood point set with a fixed radius is constructed for each fitting semantic point, all semantic points in the neighborhood point set are traversed, the number distribution of each semantic attribute label is counted, and if there are two or more different types of semantic attribute labels in the neighborhood and the semantic attribute labels of adjacent semantic points frequently switch, the semantic point is marked as a semantic boundary candidate point;

[0115] Among all the semantic boundary candidate points, a label mutation point is further identified, specifically: a semantic neighborhood set with a fixed radius is constructed around each semantic boundary candidate semantic point, the semantic attribute label sequence of the semantic points in the semantic neighborhood set is extracted, and the semantic attribute labels are arranged in ascending order according to the spatial Euclidean distance between the semantic points in the semantic neighborhood set and the semantic boundary candidate semantic point, and a label change sequence is constructed. The switching frequency of adjacent semantic attribute labels in the label change sequence is calculated to obtain a label switching frequency value; if the label switching frequency value exceeds a set switching frequency threshold, and the switching direction of the semantic attribute label presents an unordered distribution mode, the current semantic boundary candidate semantic point is determined as a label mutation point;

[0116] The determination that the switching direction of the semantic attribute label presents an unordered distribution mode includes:

[0117] For the semantic point pair with different adjacent semantic attribute labels in the label change sequence, a spatial displacement vector between the semantic point pair is extracted to represent the spatial direction corresponding to the label switching;

[0118] The spatial displacement vectors of all semantic point pairs are counted to construct a direction distribution vector set, and a direction concentration index is calculated, which is the average direction module of the unitized direction vector;

[0119] If the direction concentration index is lower than a preset direction disorder threshold, it is determined that there is a label switching with disordered distribution mode in the current label change sequence, indicating that the corresponding semantic point belongs to a semantic attribute label space distribution chaotic region;

[0120] The semantic attribute label set in the neighborhood of the label mutation point is extracted, the label distribution density is counted according to the frequency of the label, the semantic attribute label with the highest frequency is selected as the corrected label, and the corrected label is assigned to the current mutation point to replace the original semantic attribute label, thereby completing the label smoothing adjustment of the semantic boundary region and improving the spatial consistency and semantic coherence of the completed point cloud semantic attribute label set.

[0121] Finally, a completed point cloud semantic attribute label set with context coherence and semantic consistency is generated.

[0122] In the embodiment, the semantic re-labeling process specifically includes:

[0123] For each preliminary completed point in the preliminary completed fitting point cloud, a neighborhood point set with the preliminary completed point as the center and a fixed radius is constructed, the semantic attribute label types and corresponding quantities of all points in the neighborhood point set are counted, and a neighborhood label distribution vector is generated.

[0124] The number of occurrences of the semantic attribute label of the current completed point in the neighborhood label distribution vector is divided by the total number of neighborhood points to obtain a confidence score of the semantic attribute label.

[0125] If the confidence score is less than a preset confidence threshold, it is determined that the current semantic label does not have neighborhood consistency and is marked as a candidate re-labeling point.

[0126] Further, a neighborhood fitting semantic point set is constructed within the three-dimensional spatial neighborhood range of each candidate re-labeling point, for each point in the neighborhood fitting semantic point set, the current semantic attribute label and the corresponding confidence score are extracted to form a data set containing label categories and confidence pairs, and according to the confidence score corresponding to each semantic attribute label as the voting weight, all semantic attribute labels in the three-dimensional spatial neighborhood are weighted and counted, and the weighted total value of each type of semantic attribute label is calculated. In all the semantic attribute labels participating in the voting, the label category with the largest weighted total value is selected as the final semantic attribute label of the candidate re-labeling point.

[0127] In this embodiment, the weighted vote total value is equal to the sum of the confidence scores of all semantic attribute tags of the fitting points in the neighborhood, which realizes the updating of the semantic attribute tags of the candidate re-labeled points. Through the semantic attribute tag re-estimation mechanism based on the neighborhood context consistency, the spatial coherence and classification consistency of the semantic attribute tags in the completed point cloud region are effectively improved, and the overall label quality and recognition stability are enhanced.

[0128] In this embodiment, the behavior attribute tag specifically includes:

[0129] The historical time sequence posture of each fitting point cloud is collected, and the historical time sequence posture includes a direction vector of a fitting space orientation, a centroid position change sequence, and a connection angle change amount;

[0130] The spatial connection state features between the fitting point cloud and the tower structure are extracted, and the spatial connection state features include normal direction consistency, boundary fitting degree, and stable time sequence index. The normal direction consistency is quantified by the standard deviation amount of the connection region normal vector angle, the boundary fitting degree is represented by the nearest neighbor distance between the connection boundary point clouds, and the stable time sequence index is measured according to the change standard deviation of the connection angle in the historical time sequence. The three jointly reflect the spatial matching and dynamic stability of the fitting connection structure;

[0131] The preprocessed historical time sequence posture sequence and the spatial connection state features are spliced to construct a joint tensor. The preprocessing includes performing missing value interpolation, outlier removal, time alignment, and normalization;

[0132] The joint tensor is input into a behavior prediction network based on the Informer structure to perform multi-scale attention encoding and key feature aggregation within a time window, infer the running stability change trend of the target fitting in the future time interval, and generate a behavior attribute tag. The behavior attribute tag includes failure probability, connection slip probability, and poor contact risk probability.

[0133] An attribute enabling system for a tower fitting based on a point cloud model, comprising:

[0134] A data acquisition module for acquiring three-dimensional point cloud data of a tower scene and preprocessing to generate standardized point cloud data;

[0135] A region extraction module for dividing the spatial region of the standardized point cloud data and extracting a target point cloud region;

[0136] A semantic recognition module for performing semantic segmentation and instance extraction on the target point cloud region to generate a fitting point cloud set and a semantic tag;

[0137] A geometric feature extraction module for generating a geometric attribute vector;

[0138] a semantic attribute recognition module configured to recognize the fitting material type and aging grade based on the multi-modal feature vector and generate a semantic attribute label;

[0139] a structural integrity detection module configured to determine whether the fitting point cloud has structural abnormalities and recognize abnormal state point clouds;

[0140] a point cloud completion module configured to perform point cloud diffusion and structure-guided completion operations on the abnormal state point clouds and generate completed fitting point clouds;

[0141] a semantic label optimization module configured to optimize the semantic label consistency of the completed fitting point clouds and improve the label accuracy and continuity;

[0142] a structural stability verification module configured to analyze the stability of the connection structure of the fitting and determine whether the connection angle and stress are reasonable;

[0143] a behavior prediction module configured to infer future operation risks based on historical postures and connection states and generate behavior attribute labels;

[0144] an attribute mapping module configured to integrate various attribute labels and map them to corresponding structure nodes of a digital tower model to complete the attribute empowerment of the fitting.

[0145] Embodiment 1:

[0146] To verify the feasibility of the present application in implementation, the present application is applied to the digital management of tower assets in the 220kV power transmission line inspection operation, in which scenario, the inspection unit needs to collect data, analyze the structure and evaluate the material of hundreds of tower fittings to ensure the stability and safety of the long-term service state of the fittings. However, due to the complex terrain and poor traffic in the mountainous area, only airborne laser radar or unmanned aerial vehicle carrying three-dimensional scanning equipment can be used for remote sensing data acquisition, and the quality of the collected point cloud data is uneven, with problems such as local occlusion, structural loss and uneven density, which makes it difficult for traditional point cloud processing methods to effectively complete fitting recognition and state inference.

[0147] The system of the present application is deployed in the operation terminal background data center, and for the uploaded original point cloud data, first, standardization processing is completed, including noise elimination, spatial downsampling and coordinate registration, to ensure that the point cloud data has a unified structure and processing basis. The processed point cloud is divided into multiple local spatial regions, and through density analysis and spatial connection feature extraction, the target region containing the fitting is recognized. Further, the system uses a point cloud recognition network to perform semantic segmentation and cluster instance extraction on the target region, effectively distinguishing various fitting point clouds from the background tower structure and conductors, while constructing an axis-aligned bounding box to extract geometric attribute vectors and improve the accuracy of structure recognition.

[0148] In the attribute prediction link, the present application extracts the reflection intensity, density feature and edge blur degree for each fitting point cloud, and inputs them into the material identification module based on the PointNet++ network, outputs the structural material and aging grade semantic label, and automatically judges the abnormal area of the shape for the fitting point cloud with structural loss, combines the main direction analysis result, performs the structure guided diffusion completion operation, generates the completed point cloud layer by layer, and performs the secondary generation and re-estimation of the semantic label of the context consistency, to ensure the semantic coherence and accurate label of the completed area.

[0149] Subsequently, the system constructs a time sequence posture sequence based on the historical scanning data of the fitting, extracts spatial connection state features including connection angle change, boundary fitting degree and normal consistency, fuses them into a behavior feature vector, inputs it into the behavior prediction network, infers the failure probability, slip risk and poor contact trend in the future operation process, and finally aligns and maps all attribute data and tower structure digital model to form a standard format semantic data file that can be called by the operation and maintenance system.

[0150] In order to verify the performance of the present application, the traditional method is compared, and the comparison results are shown in Table 1.

[0151] Table 1 Comparison of performance of the present application and traditional method

[0152] Performance index The present invention Conventional method Accuracy rate of fitting identification (%) 93.8％ 85.3％ Completeness rate of structure complement (%) 82.4％ 68.7％ Accuracy rate of semantic label (%) 89.7％ 72.1％ Detection rate of abnormal fitting (%) 87.2％ 65.5％ Single-tower processing duration (min) 10.5 18.2 Number of data compatible devices 6 2

[0153] The present application achieves 93.8% in fitting recognition accuracy, which is significantly improved compared with 85.3% of the traditional method. The main reason is that the present application uses a semantic segmentation method based on PointNet++ network, and combines geometric attributes, reflection intensity and multi-modal features for joint reasoning, which enhances the discrimination ability of different types of fittings and reduces the misidentification and missed identification phenomenon.

[0154] The structure completion integrity rate of the present application is 82.4%, which is greatly improved compared with 68.7% of the traditional method. The advantage comes from the introduction of the guided point diffusion completion mechanism and the secondary generation process of the semantic label of the completed point cloud, which can reconstruct the complete fitting point cloud according to the normal vector direction and the structure main shaft rule when there is occlusion or scanning loss, so as to ensure the spatial continuity and morphological integrity.

[0155] The semantic label accuracy is improved from 72.1% of the traditional method to 89.7%. This improvement benefits from the context consistency optimization strategy used in the present application. Especially in the semantic re-labeling process of the completed area, through the neighborhood voting mechanism and label mutation point repair, the semantic label distribution disorder phenomenon is eliminated, and the semantic consistency and coherence are guaranteed.

[0156] The detection rate of the abnormal gold fitting of the present application is 87.2%, which is higher than the 65.5% of the traditional method. The reason is that the present application not only relies on point cloud density or shape anomaly, but also introduces multi-dimensional features such as historical time sequence posture and spatial connection state, which improves the recognition ability of the dynamic abnormal state of the loose, sliding and poor contact of the gold fitting structure.

[0157] The processing time is shortened from 18.2 minutes of the traditional method to 10.5 minutes. The efficiency improvement is mainly due to the modular design of the structured point cloud processing flow in the present application, combined with the efficient behavior prediction network and label optimization mechanism, which significantly compresses the calculation and analysis time while improving the accuracy.

[0158] The present application method supports the compatible processing of 6 types of data acquisition equipment, while the traditional method only supports 2 types, indicating that it has stronger adaptability in multi-source data access and system integration. This is due to the use of a standardized process in the point cloud preprocessing and coordinate registration link in the present application, which is suitable for multi-source input of laser radar, structured light and unmanned aerial vehicle platform, and has strong expansibility.

[0159] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can make equivalent replacements or changes within the technical scope disclosed by the present application according to the technical solution and inventive concept of the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for enabling the properties of tower hardware based on a point cloud model, characterized in that: The steps include: Collect 3D point cloud data of the tower scene, and perform preprocessing on the collected 3D point cloud data to generate standardized point cloud data; Perform spatial region division on the standardized point cloud data, extract the target point cloud area, perform semantic segmentation and instance extraction on the target point cloud area, and generate a hardware point cloud set and the corresponding geometric attribute vector; For each hardware point cloud, the reflection intensity feature, point cloud density feature, and edge fuzzy feature are extracted to construct a multimodal feature vector. This vector is then input into the material recognition model to perform attribute prediction and generate semantic attribute labels. Determine whether the hardware point cloud has structural missing, scanning occlusion, or morphological abnormalities. If so, perform point cloud diffusion completion, reconstruct the missing area, generate the completed hardware point cloud, and infer semantic attribute labels; Use the behavior prediction network to infer future operational risks and generate behavior attribute labels; Integrate the geometric attribute vectors, semantic attribute labels, completed semantic attribute labels, and behavioral attribute labels of the hardware point cloud to construct structured attribute data, and map the structured attribute data to the corresponding structural nodes of the digital tower; The generating behavior attribute label specifically includes: Collect the historical time series posture of each hardware point cloud, which includes the direction vector of the hardware spatial orientation, the center of mass position change sequence and the connection angle change; Extract the spatial connection status features between the hardware point cloud and the tower structure; The pre-processed historical temporal posture sequence and spatial connection state features are spliced ​​to construct a joint tensor; The joint tensor is input into the behavior prediction network based on the Informer structure, multi-scale attention encoding and key feature aggregation within the time window are performed, the operating stability change trend of the target hardware in the future time interval is inferred, and the behavior attribute label is generated.

2. The method for enabling attributes of tower hardware based on a point cloud model according to claim 1, characterized in that: The preprocessing includes noise removal, spatial downsampling and coordinate registration.

3. The method for enabling attributes of tower hardware based on a point cloud model according to claim 1, characterized in that: The generation of the hardware point cloud set and the corresponding geometric attribute vector specifically includes: Perform spatial region partitioning on the standardized point cloud data, dividing the entire point cloud space into three-dimensional voxel grids with fixed side lengths, and counting the number of points in each voxel grid. If the number of points is greater than a preset density threshold, the corresponding voxel grid is marked as the target point cloud area; In the target point cloud area, a local neighborhood point set is constructed for each point, and local geometric features are extracted based on the neighborhood information; The local geometric features of each point are concatenated to form a feature vector set, a feature map of the target point cloud area is constructed, and the feature map is input into the point cloud recognition network to generate a semantic label set; All points marked as hardware are filtered from the semantic label set to form a hardware semantic point set. A density-connectivity-based clustering method is used to perform instance extraction on the hardware semantic point set to generate multiple hardware point clouds. An axis-aligned three-dimensional bounding box is constructed for each hardware point cloud, and the spatial geometric attribute parameters of the three-dimensional bounding box are extracted and combined to generate a geometric attribute vector.

4. The method for enabling attributes of tower hardware based on a point cloud model according to claim 3, characterized in that: The example extraction operation performed on the hardware semantic point set specifically includes: Set the minimum neighborhood point count threshold and the distance threshold between points. For each semantic point in the hardware semantic point set, search for a neighborhood semantic point set within the distance threshold, and count the number of semantic points in the neighborhood. If the number of semantic points in the neighborhood of a semantic point is greater than or equal to the minimum neighborhood point count threshold, then use the semantic point as a cluster seed semantic point to construct the initial cluster subset. Starting from the initial cluster subset, continuously search for semantic points in its neighborhood that meet the distance threshold between points, and add the semantic points that meet the conditions to the current cluster subset. Repeat the neighborhood search process until the cluster subset no longer expands; New cluster seed semantic points are selected from the remaining semantic points that are not involved in the clustering, and the above process is repeated to finally obtain multiple structurally coherent and mutually independent cluster subsets, and the cluster subsets are used as the hardware point cloud.

5. The method for enabling attributes of tower hardware based on a point cloud model according to claim 1, characterized in that: The generation of the semantic attribute label specifically includes: For each hardware point cloud, the reflection intensity feature, point cloud density feature and edge fuzzy feature are extracted respectively; Normalize the reflection intensity features, point cloud density features, and edge fuzziness features, concatenate them into feature vectors in a fixed order, and construct a corresponding multimodal feature vector for each hardware point cloud. The multimodal feature vector is input into the material recognition model to perform deep feature extraction and classification reasoning, and the corresponding semantic attribute label is output.

6. The method for enabling attributes of tower hardware based on a point cloud model according to claim 1, characterized in that: Generating the completed hardware point cloud and inferring semantic attribute labels specifically includes: Perform structural integrity testing on each hardware point cloud to determine whether there are structural defects, blocked areas, or abnormal shapes. If any of these abnormalities exist, the hardware point cloud is considered to be in an abnormal state. For abnormal point clouds, the normal vector distribution and local curvature of edge points are used to determine the diffusion direction and diffusion area, guiding the gradual generation of completion points along the main direction of the structure to form a set of candidate completion point clouds; Perform structural consistency analysis on the candidate completion point cloud set and the adjacent hardware point cloud, filter out abnormal points that do not meet the spatial connection logic, retain valid points with good topological connectivity and smooth transition of spatial curvature, and generate the first stage completion point cloud; The first-stage completed point cloud is merged with the original hardware point cloud to form a preliminary completed hardware point cloud, and a semantic attribute label generation operation is performed to output a preliminary completed point cloud semantic attribute label set; The preliminary completed hardware point cloud and semantic attribute label set are input into the completed point cloud semantic attribute label secondary generation mechanism to execute the semantic re-labeling process based on context consistency optimization; To preliminarily complete the boundary area in the hardware point cloud, a fixed-radius neighborhood point set is constructed for each hardware semantic point. All semantic points in the neighborhood point set are traversed, and the number distribution of each semantic attribute label is counted. If there are two or more semantic attribute labels of different categories in the neighborhood, and the semantic attribute labels of adjacent semantic points frequently switch, the semantic point is marked as a candidate semantic boundary point. Among all semantic boundary candidate points, label mutation points are further identified; Extract the set of semantic attribute labels in the neighborhood of the label mutation point, calculate the label distribution density according to the frequency of label appearance, select the semantic attribute label with the highest frequency as the correction label, and assign the correction label to the current mutation point, replacing the original semantic attribute label; Finally, a set of completed point cloud semantic attribute labels with contextual coherence and semantic consistency is generated.

7. The method for enabling attributes of tower hardware based on a point cloud model according to claim 6, characterized in that: The semantic re-labeling process specifically includes: For each preliminary completed point in the preliminary completed hardware point cloud, a neighborhood point set with a fixed radius centered on the preliminary completed point is constructed. The types and corresponding quantities of semantic attribute labels of all points in the neighborhood point set are counted to generate a neighborhood label distribution vector. Divide the number of occurrences of the semantic attribute label of the current completion point in the neighborhood label distribution vector by the total number of neighborhood points to obtain the confidence score of the semantic attribute label; If the confidence score is less than the preset confidence threshold, the current semantic label is determined to be inconsistent with the neighborhood and marked as a candidate re-labeling point; Furthermore, a neighborhood hardware semantic point set is constructed within the three-dimensional spatial neighborhood of each candidate re-labeling point. For each point in the neighborhood hardware semantic point set, the current semantic attribute label and the corresponding confidence score are extracted to form a data set containing label category and confidence pairs. Based on the confidence score corresponding to each semantic attribute label as the voting weight, weighted statistics are performed on all semantic attribute labels in the three-dimensional spatial neighborhood, and the weighted total vote value of each type of semantic attribute label is calculated. Among all the semantic attribute labels participating in the vote, the label category with the largest weighted total vote value is selected as the final semantic attribute label of the candidate re-labeling point.

8. The method for enabling attributes of tower hardware based on a point cloud model according to claim 7, characterized in that: The total weighted vote value is equal to the sum of the confidence scores of all hardware points with semantic attribute labels in the neighborhood.

9. A tower hardware attribute enabling system based on a point cloud model, which executes the tower hardware attribute enabling method based on a point cloud model according to any one of claims 1 to 8, characterized in that: include: The data acquisition module is used to collect and pre-process the three-dimensional point cloud data of the tower scene to generate standardized point cloud data; The region extraction module is used to divide the spatial region of the standardized point cloud data and extract the target point cloud region; The semantic recognition module is used to perform semantic segmentation and instance extraction on the target point cloud area to generate a hardware point cloud set and semantic labels; Geometric feature extraction module, used to generate geometric attribute vectors; Semantic attribute recognition module, used to identify the material type and aging level of hardware based on multimodal feature vectors and generate semantic attribute labels; The structural integrity detection module is used to determine whether there are structural anomalies in the hardware point cloud and identify abnormal point clouds; The point cloud completion module is used to perform point cloud diffusion and structure-guided completion operations on abnormal point clouds to generate completed hardware point clouds; The semantic label optimization module is used to optimize the semantic label consistency of the completed hardware point cloud to improve label accuracy and coherence; The structural stability verification module is used to analyze the stability of the connection structure of the hardware to determine whether its connection angle and force are reasonable; The behavior prediction module is used to infer future operational risks based on historical posture and connection status and generate behavior attribute labels; The attribute mapping module is used to integrate various attribute tags and map them to the corresponding structural nodes of the digital tower model to complete the attribute empowerment of the hardware.

Images