Method for extracting and classifying disease features of point cloud data of key components of catenary

Abstract

The application discloses a disease feature extraction and classification method for point cloud data of key components of a catenary, comprising the following steps: obtaining original point cloud of a vehicle-mounted laser radar and filtering and height segmentation to obtain a point cloud data set of catenary components; calculating linear and planar features of the point cloud to construct a geometric feature vector and performing initial classification through a random forest to obtain point cloud subsets of each component; fitting catenary wire and catenary cable catenary trajectories, constructing a dropper connection based on a projection relationship and clustering into independent dropper units; extracting a positioner plane area and constructing an exclusion mask to obtain precisely classified catenary wire point cloud; extracting cross-region features and performing dropper disease classification; and calculating wear depth based on a catenary wire cross section and a standard contour and performing an alarm. The application solves the problems of linear feature confusion in a cross region and misjudgment of adjacent region planar features, makes the boundaries of point cloud of various catenary components clearer, and makes the classification results more accurate, thereby providing a reliable data basis for subsequent fine disease identification and state evaluation.

Description

Technical Field

[0001] This invention relates to the field of overhead contact line inspection technology, and in particular to a method for extracting and classifying defect features from point cloud data of key components of overhead contact lines. Background Technology

[0002] As a core component of the electrified railway traction power supply system, the structural condition of its key components (including contact wire, catenary, droppers, and positioners) directly affects the safety and stability of train operation. With the widespread application of LiDAR technology in railway inspection, the automated extraction and condition detection of contact wire components based on high-precision three-dimensional point cloud data has become a current research hotspot.

[0003] In existing technologies, the processing of contact network point cloud data typically employs classification methods based on geometric features. For example, Chinese patent CN112529044A discloses a method for extracting and classifying railway contact networks based on onboard LiDAR. This method utilizes a multi-scale adaptive feature classification algorithm, leveraging the linear, planar, and discrete features of the point cloud under principal component analysis (PCA), to achieve coarse classification and extraction of the catenary, contact line, and return line. Chinese patent CN118447098A further proposes a contact line segmentation method based on three-dimensional point cloud data. This method calculates the linearity of local neighborhood point sets using PCA and combines it with the RANSAC straight line detection algorithm to achieve fine segmentation of the contact line. The core idea of ​​these methods is to utilize the geometrical differences of various components of the contact network in three-dimensional space (e.g., the contact line and catenary are linearly distributed, while the locators are planarly distributed) and classify components through feature threshold determination.

[0004] However, in practical engineering applications, the above-mentioned classification methods based on geometric features still have the following key technical problems: First, there is the problem of linear feature confusion. In the catenary structure, the dropper, as a key component connecting the catenary and the contact wire, has a slender linear structure that is highly similar in geometry to the contact wire and catenary. Especially near the intersection of the dropper and the contact wire, the local point cloud contains information from both linear structures, causing the point cloud in this area to exhibit significant linear features in principal component analysis. When existing methods rely solely on local PCA feature values ​​for judgment, it is difficult to effectively distinguish whether the point cloud near the intersection belongs to the contact wire, catenary, or dropper, thus causing component classification confusion. This confusion not only affects the independent extraction accuracy of each component but also interferes with the accuracy of subsequent defect identification based on component relationships (such as dropper slack and local wear of the contact wire).

[0005] Second, there is the problem of misjudging planar features. The locator is a key component in the overhead contact system used to fix the spatial position of the contact wire, and its main structure is planar. Existing methods (such as CN202410486234) extract and remove the locator by calculating planar features to avoid interference with contact wire extraction. However, in actual scanning data, the edge region of the locator often spatially adjoins or even mixes with the contact wire or dropper point cloud, causing the point cloud in this region to simultaneously possess some planar features and some linear features. In this case, a single planar feature threshold is insufficient to achieve complete and accurate removal of the locator: setting the threshold too strictly will result in locator residue, which may be misidentified as the contact wire during subsequent RANSAC line detection; setting the threshold too wide will mistakenly remove adjacent contact wire or dropper point clouds as planar points, causing the loss of data for key components.

[0006] In summary, existing contact wire point cloud classification methods based on geometric features suffer from linear feature confusion in the intersection area of ​​the droppers and the contact line, and planar feature misclassification in the area where the locator and the contact line are adjacent. These problems lead to insufficient boundary accuracy in component extraction, and the classification results show mixed point clouds of different components, making it difficult to support subsequent refined defect feature extraction and classification of individual components. Therefore, there is an urgent need to propose a method that can solve the aforementioned feature confusion problem in interactive areas and achieve high-precision classification and defect feature extraction of key contact wire components. Summary of the Invention

[0007] To address the aforementioned problems, this invention provides a method for extracting and classifying defect features from point cloud data of key components of overhead contact lines.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This application discloses a method for extracting and classifying defect features from point cloud data of key components of overhead contact lines, including the following steps: S1. Obtain the original point cloud data of the key components of the catenary collected by the vehicle-mounted lidar, and perform filtering, noise reduction and elevation threshold segmentation on the original point cloud data to obtain a point cloud dataset of catenary components including the contact wire, catenary wire, dropper and positioner. S2. Based on the point cloud dataset of the contact wire components, calculate the linear feature value and planar feature value of each point cloud under a preset neighborhood radius, and construct the geometric feature vector of each point cloud according to the numerical relationship between the linear feature value and the planar feature value. Input the geometric feature vector into a random forest classifier for initial classification to obtain the initial classification of contact wire point cloud subset, initial classification of catenary point cloud subset, initial classification of drop wire point cloud subset and initial classification of locator point cloud subset; S3. For the initial classification of the contact line point cloud subset and the initial classification of the catenary point cloud subset, extract the contact line catenary trajectory and the catenary catenary trajectory respectively, and project each point cloud in the initial classification of the dropper point cloud subset onto the contact line catenary trajectory and the catenary catenary trajectory respectively. Construct a dropper connection relationship index table according to the projection distance, and cluster the point clouds in the initial classification of the dropper point cloud subset located on the same dropper structure into independent dropper point cloud units according to the dropper connection relationship index table. S4. For the initial classification locator point cloud subset, extract the locator plane main body region and calculate the outer boundary contour of the locator plane main body region. Construct a locator exclusion mask with the outer boundary contour as a constraint. Perform spatial overlay analysis on the locator exclusion mask and the initial classification contact line point cloud subset, and remove the point clouds in the initial classification contact line point cloud subset that fall within the locator exclusion mask to obtain the refined classification contact line point cloud subset. S5. Using the refined classification of the contact line point cloud subset as a spatial index, traverse the independent dropper point cloud unit, extract the intersection area point cloud of each independent dropper point cloud unit and the refined classification of the contact line point cloud subset, calculate the spatial distribution width of the intersection area point cloud along the line direction and the offset in the vertical direction, and construct the dropper structural integrity feature vector and the dropper spatial attitude feature vector. S6. Input the structural integrity feature vector of the dropper and the spatial posture feature vector of the dropper into the dropper disease classification model to obtain the classification results of the normal state, the slack state and the broken state of the dropper. S7. Based on the refined contact line point cloud subset, extract the point cloud profile of the contact line cross section along the normal direction of the contact line catenary trajectory, calculate the wear depth sequence between the point cloud profile and the standard design profile, and output contact line wear defect alarm information when the maximum wear depth in the wear depth sequence exceeds the preset wear threshold.

[0009] Furthermore, step S1 includes: S11. Collect raw scanning point cloud data along the railway line using the vehicle-mounted lidar, and associate and store the spatial three-dimensional coordinates of each point cloud in the raw scanning point cloud data with the laser reflection intensity information to construct the raw point cloud data of the key components of the catenary. S12. The statistical filtering algorithm is used to remove outliers from the original point cloud data of the key components of the overhead contact system. The neighborhood search radius of each point cloud is set as the first neighborhood threshold. The number of neighboring point clouds within the neighborhood search radius is counted. Point clouds with a number of neighboring point clouds less than the first number threshold are identified as outlier noise points and removed to obtain filtered point cloud data. S13. Perform elevation value statistical analysis on the filtered point cloud data to obtain an elevation histogram of the filtered point cloud data in the vertical direction. Determine the ground elevation threshold based on the elevation distribution interval corresponding to the ground point cloud in the elevation histogram. Use a pass-through filtering algorithm to remove point clouds in the filtered point cloud data whose elevation values ​​are less than the ground elevation threshold to obtain the point cloud dataset of the contact wire components including the contact wire, catenary wire, dropper, and positioner.

[0010] Furthermore, step S2 includes: S21. Obtain each point cloud in the contact network component point cloud dataset, set N neighborhood radii arranged in ascending order to form a multi-scale radius set, and construct a local neighborhood point set for each point cloud using each neighborhood radius in the multi-scale radius set to obtain N local neighborhood point sets corresponding to each point cloud. S22. For each local neighborhood point set, the covariance matrix is ​​calculated using principal component analysis. The covariance matrix is ​​then decomposed into eigenvalues ​​to obtain a first eigenvalue, a second eigenvalue, and a third eigenvalue. The first eigenvalue is greater than the second eigenvalue, and the second eigenvalue is greater than the third eigenvalue. Based on the first eigenvalue, the second eigenvalue, and the third eigenvalue, the linear eigenvalue and planar eigenvalue of the local neighborhood point set are calculated to obtain N linear eigenvalues ​​and N planar eigenvalues ​​corresponding to each point cloud at N scales. S23. Concatenate the N linear feature values ​​and N planar feature values ​​corresponding to each point cloud to construct a 2N-dimensional geometric feature vector for each point cloud. Input the 2N-dimensional geometric feature vectors of all point clouds in the contact wire component point cloud dataset into a pre-trained random forest classifier. The random forest classifier outputs the component category label for each point cloud. The component category label includes the contact wire category label, the catenary cable category label, the dropper category label, and the locator category label. S24. Based on the component category labels, divide the contact wire component point cloud dataset into the initial classification contact line point cloud subset, the initial classification catenary point cloud subset, the initial classification dropper point cloud subset, and the initial classification locator point cloud subset.

[0011] Furthermore, step S3 includes: S31. For the initial classification of the contact line point cloud subset, a catenary model fitting algorithm based on spatial continuity constraints is adopted. An upper limit for the number of iterations and a distance tolerance threshold are set. The contact line catenary trajectory is extracted by least squares iterative fitting. The contact line catenary trajectory describes the three-dimensional orientation of the contact line in space in the form of catenary equation parameters. S32. For the initial classification of the catenary point cloud subset, the catenary model fitting algorithm based on spatial continuity constraints is adopted. The same upper limit of the number of iterations and the distance tolerance threshold as in step S31 are set. The catenary trajectory of the catenary is extracted by least squares iterative fitting. The catenary trajectory of the catenary describes the three-dimensional orientation of the catenary in space in the form of catenary equation parameters. S33. Project each point cloud in the initial classification dropper point cloud subset onto the contact line catenary trajectory and the catenary trajectory of the catenary cable, respectively. Calculate the first projection distance from each point cloud to the contact line catenary trajectory and the second projection distance to the catenary trajectory of the catenary cable. When both the first projection distance and the second projection distance are less than the connection distance threshold, record the point cloud in the dropper connection relationship index table and associate and store the first projection distance and the second projection distance corresponding to the point cloud. S34. Traverse the dropper connection relationship index table. Based on the projection distance relationship recorded in the dropper connection relationship index table, cluster the point clouds in the initial classification dropper point cloud subset that correspond to the same catenary trajectory interval of the same catenary and the same contact line catenary trajectory interval into the same independent dropper point cloud unit, and obtain multiple independent dropper point cloud units. Each independent dropper point cloud unit corresponds to the point cloud representation of a complete dropper structure.

[0012] Furthermore, step S4 includes: S41. For the initial classification locator point cloud subset, a region growth algorithm based on normal vector consistency clustering is adopted. First, the normal vector of each point cloud in the initial classification locator point cloud subset is calculated. The point cloud with the smallest curvature in the initial classification locator point cloud subset is selected as the initial seed point. The normal vector angle threshold and the growth distance threshold are set. Neighboring point clouds that satisfy the normal vector angle being less than the normal vector angle threshold and the spatial Euclidean distance being less than the growth distance threshold are included in the same clustering region. The growth is iterated until no new point clouds are added, and the main body region of the locator plane is extracted. S42. Perform two-dimensional convex hull calculation on the main body region of the locator plane to obtain the minimum convex polygon boundary of the main body region of the locator plane on the projection plane. Extend the minimum convex polygon boundary outward by a preset extension distance along the normal direction of the main body region of the locator plane to obtain the outer boundary contour of the main body region of the locator plane. S43. Using the outer boundary contour of the main body area of ​​the locator plane as the spatial constraint boundary, construct a locator exclusion mask. The locator exclusion mask is a binary spatial mask matrix. In the locator exclusion mask, the pixels located inside the outer boundary contour are assigned the first mask value, and the pixels located outside the outer boundary contour are assigned the second mask value. S44. Perform spatial overlay analysis on the locator exclusion mask and the initial classification contact line point cloud subset, traverse each point cloud in the initial classification contact line point cloud subset, project each point cloud in the initial classification contact line point cloud subset onto the projection plane where the locator exclusion mask is located, obtain the mask value corresponding to the projection point, and when the mask value is the first mask value, remove the point cloud from the initial classification contact line point cloud subset to obtain the refined classification contact line point cloud subset.

[0013] Furthermore, step S5 includes: S51. Using the refined contact line point cloud subset as a spatial index, traverse each independent dropper point cloud unit. For the currently traversed independent dropper point cloud unit, calculate the spatial centroid coordinates of all point clouds in the independent dropper point cloud unit. Using the spatial centroid coordinates as a reference point, set the forward search radius and the backward search radius along the line direction. Extract point clouds located within the range of the forward search radius and the backward search radius from the refined contact line point cloud subset to form the contact line local point cloud subset corresponding to the independent dropper point cloud unit. Perform spatial intersection between the contact line local point cloud subset and the independent dropper point cloud unit, and extract the point clouds that exist simultaneously in the contact line local point cloud subset and the independent dropper point cloud unit as the intersection area point cloud. S52. The principal component analysis method is used to calculate the principal direction features of the point cloud in the intersection area. Based on the principal direction features, the principal direction vector along the line direction is determined. All points in the point cloud in the intersection area are projected onto a plane perpendicular to the principal direction vector. The two-dimensional convex hull area of ​​the projected point cloud on the plane is calculated. The two-dimensional convex hull area is divided by the span value of the point cloud in the intersection area on the principal direction vector to obtain the spatial distribution width of the point cloud in the intersection area along the line direction. S53. Calculate the spatial Euclidean distance from each point cloud in the intersection region point cloud to the subset of the fine classification contact line point cloud. Statistically analyze the spatial Euclidean distances of all point clouds and take the difference between the maximum and minimum spatial Euclidean distances as the offset of the intersection region point cloud in the vertical direction. S54. After normalizing the spatial distribution width of the intersection area point cloud along the line direction and the offset of the intersection area point cloud in the vertical direction, the points are spliced ​​together to construct the dropper structural integrity feature vector and dropper spatial attitude feature vector corresponding to each independent dropper point cloud unit.

[0014] Furthermore, step S6 includes: S61. Obtain multiple historical independent dropper point cloud units with labeled defect types from historical catenary inspection data. For each historical independent dropper point cloud unit, extract the historical dropper structural integrity feature vector and the historical dropper spatial attitude feature vector according to the method in step S5. After concatenating the historical dropper structural integrity feature vector and the historical dropper spatial attitude feature vector, associate and store them with the corresponding defect type label to construct a dropper defect training sample set. S62. Construct a classification model for string disease. The classification model for string disease uses a multi-class support vector machine as the classifier kernel. The kernel function of the multi-class support vector machine is set as a radial basis function kernel function. The kernel parameters and penalty parameters of the radial basis function kernel function are set. The string disease training sample set is input into the string disease classification model for training. The kernel parameters and penalty parameters are optimized by a grid search algorithm until the classification accuracy of the string disease classification model on the validation set reaches a preset accuracy threshold, thus obtaining a pre-trained string disease classification model. S63. The feature vector of the integrity of the dropper structure and the feature vector of the spatial posture of the dropper are concatenated and then input into the pre-trained dropper defect classification model. The pre-trained dropper defect classification model outputs one of the following as the classification result: dropper normal state category label, dropper slack state category label, and dropper broken state category label.

[0015] Furthermore, step S7 includes: S71. Based on the refined contact line point cloud subset, multiple sampling cross-section positions are set at equal intervals along the contact line catenary trajectory with a preset sampling step size. For each sampling cross-section position, with the sampling cross-section position as the center, the normal direction perpendicular to the tangent direction of the contact line catenary trajectory is set as the cross-section extraction direction. A cross-section thickness threshold is set. All point clouds in the refined contact line point cloud subset that are spatially less than the cross-section thickness threshold and located within the range on both sides of the normal direction are extracted, thus forming the contact line cross-section point cloud contour corresponding to the sampling cross-section position. S72. Obtain the standard contact line design profile, which is a geometric profile template of the standard contact line cross-section in a wear-free state. Iterate the nearest point registration between the point cloud profile of each contact line cross-section and the standard contact line design profile. Calculate the rigid body transformation parameters that minimize the root mean square error between the point cloud profile of the contact line cross-section and the standard contact line design profile. Align the point cloud profile of the contact line cross-section and the standard contact line design profile to the same coordinate system according to the rigid body transformation parameters. S73. In the aligned coordinate system, calculate the directed distance from each point cloud in the point cloud profile of the contact line section to the standard contact line design profile along the normal direction of the standard contact line design profile. Take the maximum value of the directed distance as the wear depth value at the sampling section position. Traverse all sampling section positions to obtain the wear depth sequence. S74. Compare the maximum wear depth in the wear depth sequence with a preset wear threshold. When the maximum wear depth is greater than the preset wear threshold, output contact wire wear fault alarm information. The contact wire wear fault alarm information includes the sampling section position coordinates corresponding to the maximum wear depth and the maximum wear depth value.

[0016] This invention addresses the problem of feature confusion in existing contact network point cloud classification methods based on geometric features in intersecting and adjacent regions. By introducing a joint modeling mechanism of multi-scale feature representation and structural constraints, it achieves high-precision separation and defect feature extraction of key components of the contact network, with the following beneficial effects: By constructing multi-scale principal component analysis features and combining them with a random forest classifier, the point cloud was transformed from single-scale geometric discrimination to multi-scale comprehensive discrimination. This effectively alleviated the classification confusion problem caused by the similarity of linear features between the droppers and the contact wire in the intersection area, and improved the discriminative power and stability of the initial classification results. By introducing a catenary model based on spatial continuity constraints, global structural fitting was performed on the contact wire and the catenary. Inverse constraints were applied using the spatial connection relationship between the droppers and both, enabling the reconstruction of the dropper point cloud from a disordered distribution to structural units. This significantly improved the accuracy and structural integrity of dropper identification in the intersection area. Furthermore, a region growing method based on normal vector consistency was employed to improve... By extracting the main planar region of the locator and constructing a locator exclusion mask with outer boundary constraints, the interference of the locator on the contact line is transformed into a spatial region culling problem, effectively avoiding misjudgment and missed judgment caused by the instability of planar feature thresholds and improving the boundary purity of the contact line point cloud. By constructing the structural integrity and spatial attitude features of the dropper based on the spatial distribution characteristics of the intersection area of ​​the contact line and the dropper, the transformation from geometric point cloud to physical state features is realized, providing a high-discrimination input for defects such as dropper slack and breakage. Through the alignment analysis based on the contact line cross-sectional profile and the standard design profile, the quantitative calculation and spatial positioning of the contact line wear depth are realized, improving the accuracy and interpretability of defect detection.

[0017] In summary, this invention effectively solves the problems of confusion of linear features in cross regions and misjudgment of planar features in adjacent regions, making the point cloud boundaries of various catenary components clearer and the classification results more accurate, thus providing a reliable data foundation for subsequent refined defect identification and condition assessment. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0019] In the attached diagram: Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0020] The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0021] like Figure 1 As shown, this invention discloses a method for constructing a predictive model for postoperative cognitive decline in elderly patients, including: S1. Obtain the original point cloud data of the key components of the catenary collected by the vehicle-mounted lidar, and perform filtering, noise reduction and elevation threshold segmentation on the original point cloud data to obtain a point cloud dataset of catenary components including the contact wire, catenary wire, dropper and positioner. S2. Based on the point cloud dataset of the contact wire components, calculate the linear feature value and planar feature value of each point cloud under a preset neighborhood radius, and construct the geometric feature vector of each point cloud according to the numerical relationship between the linear feature value and the planar feature value. Input the geometric feature vector into a random forest classifier for initial classification to obtain the initial classification of contact wire point cloud subset, initial classification of catenary point cloud subset, initial classification of drop wire point cloud subset and initial classification of locator point cloud subset; S3. For the initial classification of the contact line point cloud subset and the initial classification of the catenary point cloud subset, extract the contact line catenary trajectory and the catenary catenary trajectory respectively, and project each point cloud in the initial classification of the dropper point cloud subset onto the contact line catenary trajectory and the catenary catenary trajectory respectively. Construct a dropper connection relationship index table according to the projection distance, and cluster the point clouds in the initial classification of the dropper point cloud subset located on the same dropper structure into independent dropper point cloud units according to the dropper connection relationship index table. S4. For the initial classification locator point cloud subset, extract the locator plane main body region and calculate the outer boundary contour of the locator plane main body region. Construct a locator exclusion mask with the outer boundary contour as a constraint. Perform spatial overlay analysis on the locator exclusion mask and the initial classification contact line point cloud subset, and remove the point clouds in the initial classification contact line point cloud subset that fall within the locator exclusion mask to obtain the refined classification contact line point cloud subset. S5. Using the refined classification of the contact line point cloud subset as a spatial index, traverse the independent dropper point cloud unit, extract the intersection area point cloud of each independent dropper point cloud unit and the refined classification of the contact line point cloud subset, calculate the spatial distribution width of the intersection area point cloud along the line direction and the offset in the vertical direction, and construct the dropper structural integrity feature vector and the dropper spatial attitude feature vector. S6. Input the structural integrity feature vector of the dropper and the spatial posture feature vector of the dropper into the dropper disease classification model to obtain the classification results of the normal state, the slack state and the broken state of the dropper. S7. Based on the refined contact line point cloud subset, extract the point cloud profile of the contact line cross section along the normal direction of the contact line catenary trajectory, calculate the wear depth sequence between the point cloud profile and the standard design profile, and output contact line wear defect alarm information when the maximum wear depth in the wear depth sequence exceeds the preset wear threshold.

[0022] Specifically, step 1 includes: The core of step S1 lies in extracting a point cloud dataset of contact wire components with engineering significance from the original vehicle-mounted LiDAR scanning data, providing high-quality input data for subsequent geometric feature analysis and component classification. This step achieves point cloud data purification and target region focusing through three levels: original data construction, outlier noise removal, and elevation constraint segmentation.

[0023] In S11, a vehicle-mounted LiDAR continuously scans along the railway line to acquire a spatially discrete set of three-dimensional point clouds. Each point cloud contains not only spatial three-dimensional coordinate information, but also... It also records the corresponding laser reflection intensity information simultaneously. Three-dimensional coordinates are used to characterize the geometric position of the catenary structure in space, while laser reflection intensity information reflects differences in material properties and surface reflection characteristics, which helps to distinguish different types of components in subsequent processing. By storing spatial coordinates and reflection intensities in a one-to-one correspondence, a unified data structure is constructed to form the original point cloud data of key components of the catenary, providing a complete multi-attribute observation basis for subsequent processing.

[0024] In step S12, outlier removal based on statistical characteristics is performed on the original point cloud data. For each point cloud, a fixed neighborhood search radius is set as the first neighborhood threshold, and the number of neighboring point clouds within this neighborhood is counted. Since the point cloud distribution on a real structural surface has spatial continuity, the number of points within its neighborhood is usually maintained within a certain range. However, noise points generated by measurement errors, reflection anomalies, or environmental interference often exhibit isolated distributions, with significantly fewer points within their neighborhoods. Based on this, by setting the first threshold, point clouds with fewer neighboring points than this threshold are identified as outlier noise points and removed. This process essentially uses local spatial density as a criterion to automatically identify and delete sparse outliers, thereby effectively improving the structural integrity and statistical stability of the point cloud data.

[0025] In step S13, after outlier noise removal, elevation information is further used to perform layered filtering of the point cloud data. Statistical analysis is performed on the vertical elevation distribution of the filtered point cloud data to construct an elevation histogram. Since ground point clouds typically form obvious high-density clusters in the overall data, their corresponding elevation intervals appear as significant peak areas in the histogram. By identifying these peak intervals, the elevation range of the ground point clouds can be determined, and a ground elevation threshold can be set accordingly. Subsequently, a pass-through filtering strategy is used to evaluate all point clouds point by point, removing those with elevation values ​​below the ground elevation threshold, retaining only the point cloud data located within the height range of the contact wire structure. This process effectively removes the influence of the ground and low-lying interfering objects from the full-scene scan data, ensuring that the dataset primarily contains point clouds of high-altitude structures such as the contact wire, catenary, droppers, and positioners.

[0026] After the above three sub-steps, the original point cloud data is transformed from a mixed dataset containing a large amount of noise and irrelevant structures into a point cloud dataset with continuous spatial distribution, significantly reduced noise, and focused on key components of the overhead contact system. This dataset has a clearer spatial structure and more stable statistical properties, providing a reliable data foundation for subsequent multi-scale principal component analysis and fine classification, while also creating the preconditions for solving the problem of feature confusion in cross-regions.

[0027] Specifically, step 2 includes: Step S2 builds upon the point cloud dataset of catenary components obtained in Step S1. Its goal is to transform the local spatial structural differences of the point cloud into distinguishable numerical representations through multi-scale geometric feature modeling, and to achieve preliminary classification of different catenary components using a supervised learning model. The key to this step is overcoming the insufficient expressive power of geometric features at a single scale, and enhancing the discriminative ability for complex structural regions through multi-scale information fusion.

[0028] In step S21, each point cloud in the contact wire component point cloud dataset output from step S1 is processed point-by-point. For each point cloud, a multi-scale neighborhood structure is constructed centered on that point. Specifically, N neighborhood radii are set in ascending order to form a multi-scale radius set. Different scales correspond to different spatial perception ranges; small-scale neighborhoods can characterize the local fine structural features of the point cloud, while large-scale neighborhoods reflect a more stable overall geometric trend. For any given point cloud, under the constraint of each neighborhood radius, all point clouds within its neighborhood range are obtained through Euclidean distance search, and N local neighborhood point sets are constructed respectively. This processing ensures that the same point cloud has an independent geometric description basis at different spatial scales, which helps to alleviate the problem of easy confusion of single-scale features in the intersection area of ​​the dropper and the contact wire.

[0029] In S22, principal component analysis is performed on each local neighborhood point set to extract the principal directional features of its spatial distribution. For any local neighborhood point set, the covariance matrix of its three-dimensional coordinates is first calculated: in, This indicates the number of point clouds in the neighborhood. Indicates the first The three-dimensional coordinates of the points This represents the centroid coordinates of the neighborhood point cloud. The covariance matrix reflects the dispersion of the point cloud along the three principal directions. Eigenvalue decomposition of this covariance matrix yields the first, second, and third eigenvalues, ordered by magnitude, denoted as: The first eigenvalue corresponds to the maximum dispersion of the point cloud along the main direction, while the third eigenvalue reflects the minimum dispersion perpendicular to the main direction.

[0030] Based on the above eigenvalues, a geometric discriminant index is constructed. Linear eigenvalues ​​are used to characterize whether a point cloud exhibits a thin, linear distribution; their definition is: When the point cloud is significantly distributed along a single direction Much larger This value approaches 1. Planar eigenvalues ​​are used to characterize whether a point cloud exhibits a planar distribution; they are defined as follows: This value is large when the point cloud has large dispersion in the two principal directions and is concentrated in the third direction. By calculating the linear eigenvalues ​​and planar eigenvalues ​​at each scale, the structural representation of the same point cloud at different spatial scales can be obtained, thus forming differentiated feature responses in the intersection regions.

[0031] In S23, the linear eigenvalue sequences obtained at N scales for each point cloud are concatenated with the planar eigenvalue sequences to construct a unified geometric eigenvector. For any point cloud, its geometric eigenvector is represented as... This 2N-dimensional vector comprehensively reflects the linear and planar structural features of the point cloud in multi-scale space. Compared to single-scale features, this multi-scale feature vector can demonstrate scale response differences in the intersection region of the dropper and the contact line. For example, the dropper is more likely to exhibit isolated linear features at small scales, while the contact line has stronger continuity at large scales, thereby enhancing classification and discrimination capabilities.

[0032] The geometric feature vectors of all point clouds are input into a pre-trained random forest classifier for classification. The random forest classifier consists of multiple decision trees, constructed by randomly sampling different feature subsets and sample subsets, effectively handling high-dimensional feature spaces and reducing the risk of overfitting. During classification, each decision tree independently outputs a class prediction result, and the final component class label for each point cloud is determined through a majority voting mechanism. The output class labels include contact wire, catenary, dropper, and locator class labels. Because random forests can learn nonlinear combinations of features across multiple scales, they still possess good discriminative ability when faced with structures that have similar geometric features but different spatial contexts.

[0033] In step S24, the point cloud dataset of the overhead contact system components is divided based on the component category labels output by the random forest classifier. Point clouds identified as contact lines are assigned to the initial contact line subset; point clouds identified as catenary wires are assigned to the initial catenary wire subset; point clouds identified as droppers are assigned to the initial dropper subset; and point clouds identified as locators are assigned to the initial locator subset. This division achieves the initial separation of key components of the overhead contact system, providing a classification foundation for subsequent fine-grained reconstruction based on structural constraints and extraction of defect features.

[0034] Through the processing in step S2, the point cloud data is transformed from its original form, which only contains spatial coordinate information, into a structured expression containing multi-scale geometric semantics. This completes the mapping from unlabeled data to data with clear component category labels, effectively alleviating the classification confusion problem caused by local geometric similarity. At the same time, it provides reliable initial classification results for introducing structural relationship constraints in subsequent steps.

[0035] Specifically, step 3 includes: Step S3 builds upon the point cloud purification and spatial filtering completed in step S1 and the initial classification of multi-scale geometric features completed in step S2. Its core objective is to introduce global geometric constraints on the catenary structure, restore the continuous trajectory of the contact wire and the catenary cable through the catenary model, and use the prior knowledge of this structure to perform structured reorganization of the drop wire point cloud, thereby solving the problem of cross-confusion that is difficult to distinguish by relying solely on local geometric features.

[0036] In step S31, a catenary trajectory is fitted to the initially classified contact line point cloud subset obtained in step S2. In actual engineering, the contact line, under the influence of gravity, exhibits a typical catenary shape, which can be described in a two-dimensional plane by the following function. Where parameters Controlling the curvature of the curve Indicates horizontal offset. This represents vertical translation. In three-dimensional space, by introducing a line direction parameter, the model can be extended to a spatial curve that continuously varies along the line direction. For the initial classification of the contact line point cloud subset, a local coordinate system is first established based on the line direction, so that the point cloud approximately satisfies the two-dimensional catenary distribution characteristics in this coordinate system. Then, the least squares criterion is used, with the spatial distance from the point cloud to the catenary model as the error term, to construct the objective function and perform iterative optimization. During the iteration process, a spatial continuity constraint is introduced, requiring the fitting parameters to change smoothly within adjacent sampling intervals, thereby suppressing the interference of local misclassified points on the overall fitting. When the number of iterations reaches a preset upper limit or the residual change is less than the distance tolerance threshold, the iteration stops, and the contact line catenary trajectory is output. This trajectory describes the three-dimensional orientation of the contact line in space in the form of a set of continuous parametric functions, realizing the mapping from discrete point clouds to a continuous structural model.

[0037] In step S32, the initial classification subset of the catenary point cloud is processed using the same method as in S31. The catenary also satisfies the catenary mechanical properties, but its spatial position is usually higher than the contact line, and its sag parameters differ. Using the same least-squares iterative fitting framework, the catenary trajectory of the catenary is extracted under spatial continuity constraints. Since the initial classification results in step S2 have already distinguished the catenary from the contact line to some extent, the global constraints of the catenary model can further eliminate the influence of local misclassified points, making the fitting results more stable. Thus, two spatially corresponding and continuously smooth structural main lines are obtained: the catenary trajectory of the contact line and the catenary trajectory of the catenary.

[0038] In S33, the spatial connection between the dropper and the two catenaries is introduced. For each point cloud in the initially classified dropper point cloud subset, the shortest distance projection is made onto the contact line catenary trajectory and the catenary trajectory of the catenary. For any point cloud... The distance from it to the catenary trajectory can be expressed as: in This represents the set of corresponding catenary trajectories. By numerically solving for the minimum distance, the first projected distance from the point cloud to the contact line and the second projected distance to the catenary can be obtained. In the actual structure, the dropper connects the catenary and the contact line, and its point cloud should simultaneously approximate both catenary trajectories. Therefore, when both the first and second projected distances are less than a preset connection distance threshold, the point cloud can be determined to belong to a valid dropper connection area. Point clouds that meet this condition are recorded in the dropper connection relationship index table, and their corresponding two types of projected distances are stored together. This index table essentially establishes a dual spatial association between the point cloud and the two main structures, transforming discrete points into connection units with structural semantics.

[0039] In S34, the dropper connection relationship index table is structurally parsed to achieve clustering and reconstruction of the dropper point cloud. All record point clouds in the index table are traversed, and point clouds with similar projection intervals are grouped together based on their projection positions on the catenary trajectory of the contact line and the catenary trajectory of the catenary. Specifically, if multiple point clouds have projection positions on the catenary trajectory within the same local interval, and their projection positions on the contact line trajectory also fall within the corresponding interval, then these point clouds are considered to constitute the same dropper structure. In this way, the initially classified dropper point cloud subset is divided into multiple independent dropper point cloud units. Each independent dropper point cloud unit spatially connects a section of catenary cable and a section of contact line, forming a complete structural representation.

[0040] Through the processing in step S3, the initial classification results obtained in step S2 based solely on local geometric features are further reconstructed by incorporating global structural constraints. The contact line and catenary are transformed from discrete point clouds into continuous catenary trajectories, and the dropper point cloud is transformed from an unordered set into structural units with clear connections. This shift from being driven by local features to being driven by structural relationships effectively avoids the problem of linear feature confusion in the intersection area of ​​the dropper and the contact line, while providing an accurate structural basis for subsequent dropper condition analysis and contact line wear detection.

[0041] Specifically, step 4 includes: Step S4, building upon the point cloud purification, multi-scale initial classification, and structural trajectory recovery completed in steps S1 to S3, introduces an explicit planar structure extraction and spatial exclusion mechanism to address the misjudgment of planar features in the spatially adjacent region between the locator and the contact line. This step transforms the locator's interference with the contact line point cloud into a controllable spatial mask constraint by constructing the locator's main planar region and its spatial extension boundary, thereby achieving further refined screening of the contact line point cloud.

[0042] In step S41, a local geometric consistency analysis is performed on the subset of point clouds obtained in step S2 for the initial classification locator. For each point cloud, based on its neighborhood point set, the local covariance matrix is ​​calculated through principal component analysis, and the eigenvector corresponding to the minimum eigenvalue is extracted as the normal vector of that point. The normal vector reflects the orientation information of the local surface where the point cloud is located, and the curvature can be characterized by the ratio of the minimum eigenvalue to the sum of eigenvalues. The smaller the curvature, the closer the point is to the planar region. Based on this, the point with the smallest curvature is selected from the subset of point clouds for the initial classification locator as the initial seed point for region growing. Around this seed point, a normal vector angle threshold and a spatial distance threshold are set to filter the point clouds in its neighborhood. When the angle between the normal vectors of the neighboring point clouds and the point clouds in the current region satisfies: Furthermore, when the spatial Euclidean distance between two points is less than the growth distance threshold, the neighboring point is included in the current clustering region. As the region expands, the same judgment condition is repeated for newly added points until no more new point clouds satisfy the condition, completing a full region growth process. The point cloud region extracted by this method has high consistency in the normal vector direction and spatial continuity, thus forming the planar main region of the locator and effectively eliminating non-planar point clouds interspersed in the edge mixed regions.

[0043] In S42, boundary modeling is performed on the extracted locator planar main region. First, the point cloud within this main region is projected onto its local principal plane coordinate system, transforming the 3D problem into a 2D planar problem. Within the projection plane, convex hull calculation is performed on all point clouds to obtain the smallest convex polygon boundary enclosing the point set. This convex polygon completely covers the outer contour of the locator planar main region. Considering the sparse sampling and incomplete boundary conditions during actual point cloud acquisition, this boundary is expanded outward along the plane normal direction. The expansion distance is empirically set based on the locator's structural dimensions, ensuring the boundary covers any potentially cluttered areas at the locator's edges. The expanded boundary becomes the outer boundary contour of the locator planar main region, providing geometric constraints for subsequent spatial exclusion.

[0044] In step S43, the aforementioned outer boundary contour is transformed into a mask structure that can be used for spatial determination. Using the outer boundary contour as a constraint, a two-dimensional discrete mesh is constructed on its corresponding projection plane, forming a binary spatial mask matrix. For regions located inside the outer boundary contour, a uniform first mask value is assigned to identify the locator's influence area; for regions located outside the outer boundary contour, a second mask value is assigned to identify non-locator regions. This mask spatially creates a clear segmentation of the locator region, transforming previously difficult-to-distinguish mixed regions using local geometric features into directly determinable spatial regions.

[0045] In S44, the locator exclusion mask is applied to the initially classified contact line point cloud subset to achieve fine-grained filtering of the contact line point cloud. For each point cloud in the initially classified contact line point cloud subset, its spatial coordinates are projected onto the projection plane corresponding to the locator exclusion mask to obtain its position index in the mask matrix, and the corresponding mask value is read. When the mask value corresponding to the point cloud is the first mask value, it means that the point is located within the coverage area of ​​the locator's outer boundary contour. This point cloud is very likely to originate from the edge of the locator or be mixed with the locator, and should be removed from the contact line point cloud subset; when the mask value is the second mask value, the point cloud is retained. Through point-by-point traversal and filtering, the finely classified contact line point cloud subset after removing locator interference is obtained.

[0046] After step S4, the contact line point cloud subset is further purified from the initial classification results, significantly reducing the interference of the locator edge point cloud on the contact line geometry. Compared to methods that rely solely on planar feature thresholds, this step explicitly constructs the locator planar region and its spatial extension boundary, transforming the complex local feature discrimination problem into a spatial region determination problem, thereby avoiding misjudgments and missed judgments caused by inappropriate threshold selection. Simultaneously, this processing provides a cleaner and more structurally consistent point cloud foundation for subsequent wear detection based on contact line cross-section analysis.

[0047] Specifically, step 5 includes: Step S5 builds upon the point cloud purification, initial classification, structural trajectory recovery, and locator interference removal completed in steps S1 to S4. Its core objective is to meticulously characterize the local interaction areas between each independent dropper and the contact line, starting from the structural connection relationships, and transforming the original spatial geometric relationships into quantifiable defect discrimination features. This step achieves a joint representation of the dropper structure's integrity and spatial attitude by constructing point clouds of intersecting areas and extracting their spatial distribution characteristics.

[0048] In S51, the refined contact line point cloud subset obtained in step S4 is used as a spatial reference index to process each independent dropper point cloud unit obtained in step S3. For the current independent dropper point cloud unit, the spatial barycenter coordinates of all its point clouds are first calculated: in This represents the number of points in the dropper point cloud unit. The spatial centroid can be considered as the approximate center of the area where the dropper connects to the contact line. Using this centroid as a reference, and combining the line direction information determined in step S3, forward and backward search radii are set in this direction respectively. Point clouds within this range are extracted from the refined contact line point cloud subset, forming a local point cloud subset of the contact line corresponding to the current dropper. This local subset spatially covers the area where the dropper connects to the contact line. Subsequently, this local point cloud subset of the contact line is spatially intersected with the current independent dropper point cloud unit, and the set of point clouds that belong to both in spatial location is selected and defined as the intersection area point cloud. This intersection area point cloud reflects the actual contact or adjacency state between the dropper and the contact line and is the core data carrier for subsequent feature calculations.

[0049] In S52, the spatial distribution morphology of the point cloud in the intersection area is analyzed. First, principal component analysis is used to calculate the covariance matrix of the point cloud set, and the eigenvector corresponding to the largest eigenvalue is extracted as the principal direction vector. Since the intersection area has extension along the contact line direction, this principal direction vector can be considered as an approximate expression along the line direction. Using this principal direction vector as a reference, all intersection area point clouds are projected onto a plane perpendicular to this direction, thus transforming the three-dimensional distribution into a two-dimensional distribution. Within this projection plane, convex hull calculation is performed on the projected point clouds to obtain their envelope area (A). Simultaneously, the projection span (L) of the intersection area point cloud on the principal direction vector is calculated, which is the difference between the maximum and minimum projection values ​​of all points in this direction. This is achieved by constructing a ratio relationship: The spatial distribution width of the point cloud in the intersection area along the line direction is obtained. This width reflects the lateral expansion of the area where the dropper connects to the contact wire. When the dropper structure is intact and the connection is stable, this width remains within a small range; when slack or structural abnormalities occur, this area may show a diffusion trend, resulting in an increase in width.

[0050] In S53, the vertical dispersion of the point cloud in the intersection region is measured. For each point cloud in the intersection region, the minimum spatial Euclidean distance to the refined contact line point cloud subset is calculated. This distance represents the degree to which the point cloud deviates from the ideal position of the contact line. The distance values ​​for all point clouds are statistically analyzed, and the maximum and minimum values ​​are extracted, with their difference calculated. This difference characterizes the thickness variation or dispersion of the point cloud in the intersection area in the direction perpendicular to the contact line. When the suspension wire is under uniform stress and the connection is normal, this value is usually small; when the suspension wire is slack or locally deformed, the distribution of the point cloud in the vertical direction will expand significantly, resulting in an increase in this offset.

[0051] In S54, the extracted spatial distribution width and vertical offset are uniformly expressed. To eliminate the influence of different dimensions and numerical ranges, the two types of features are normalized to map them to a unified scale space. Subsequently, the normalized spatial distribution width and vertical offset are concatenated to construct the feature vector corresponding to each independent dropper point cloud unit. This feature vector describes the dropper state from two complementary dimensions: on the one hand, it reflects the structural integrity through the spatial distribution width, and on the other hand, it reflects the spatial attitude stability through the vertical offset, thus forming the dropper structural integrity feature vector and the dropper spatial attitude feature vector, respectively.

[0052] Through the processing in step S5, the dropper state information, which is difficult to directly determine in the original point cloud, is transformed into quantitative features with clear physical meaning. This feature extraction method based on structural interaction regions explicitly introduces the geometric relationship between the dropper and the contact line into the analysis process, effectively avoiding the uncertainty caused by simply relying on the dropper's own point cloud. At the same time, the feature vector output by this step provides a highly discriminative input for the subsequent disease classification model, enabling a clear distinction between states such as dropper slack and breakage in the feature space.

[0053] Specifically, step 6 includes: Step S6, based on the point cloud preprocessing, structural decomposition, and feature construction completed in steps S1 to S5, further maps the geometric features of the dropper to specific disease types, realizing the transformation from "structural representation" to "state discrimination." The core of this step lies in constructing a classification model with good generalization ability, so that the feature differences exhibited by the dropper under different operating states can form a stable and separable distribution in the feature space.

[0054] In step S61, the foundation of sample data for model training is first constructed. Multiple historical independent dropper point cloud units with completed defect annotations are obtained from historical catenary inspection data. Each sample corresponds to a specific state category, including normal, relaxed, and broken states. For each historical independent dropper point cloud unit, the same processing flow as in step S5 is used to extract its dropper structural integrity feature vector and dropper spatial attitude feature vector. Since the features defined in step S5 have clear physical meanings, these two types of features respectively reflect the lateral distribution characteristics and vertical stability of the dropper connection area. The two types of features are concatenated to form a unified feature representation vector: in Indicates the width of the normalized spatial distribution. This represents the normalized vertical offset. Subsequently, this feature vector is stored in a one-to-one correspondence with the corresponding disease type label to construct a training sample set for string disease. This method allows different disease states to form a labeled distribution structure in the feature space, providing a foundation for supervised learning.

[0055] In S62, a classification model for dropper string defects is constructed. The model uses a multi-class support vector machine as the core classifier. Its basic idea is to construct multiple optimal separating hyperplanes in the feature space to maximize the margin between samples of different classes. Considering that dropper string condition features often exhibit non-linear separability in actual distributions, a radial basis function kernel is introduced to perform a non-linear mapping on the original feature space. Its expression is as follows: in These are kernel parameters used to control the range of influence between samples. Penalty parameter. Used to balance the relationship between maximizing the classification margin and the classification error, when When the size is large, the model tends to reduce misclassification. When the spacing is smaller, more emphasis is placed on the smoothness of the overall interval.

[0056] The constructed training sample set of stringer disease is input into the support vector machine model, and the optimal classification boundary is determined through the training process. To obtain the optimal model parameters, a grid search strategy is used to traverse the kernel parameters and penalty parameters in a preset parameter space, and the classification performance of each parameter set is evaluated on the validation set. By continuously comparing the classification accuracy, the parameter combination that makes the validation set performance reach a preset accuracy threshold is selected as the optimal parameter configuration. After parameter optimization, a pre-trained stringer disease classification model is obtained. This model forms a stable classification boundary in the feature space, making different disease states clearly separated in space.

[0057] In step S63, the feature vector of dropper structural integrity extracted from the current detection data in step S5 is concatenated with the feature vector of dropper spatial attitude to form the feature input vector to be classified. This feature vector is then input into a pre-trained dropper defect classification model, which determines the defect based on its positional relationship in the feature space and outputs the corresponding category label. The output result is one of the following: dropper in normal state, dropper in slack state, or dropper in broken state. This classification result directly reflects the current operating status of the dropper and provides a basis for subsequent maintenance decisions.

[0058] Through the processing in step S6, the suspension wire is further transformed from a simple geometric structural representation into a disease status label with engineering significance. This process realizes a complete closed loop from data-driven feature extraction to model-driven state recognition. Since the input features are derived from the high-discriminatory features extracted based on structural interaction relationships in step S5, and the classification model uses a support vector machine with strong nonlinear mapping capabilities, it can effectively distinguish different disease types with subtle differences in geometric morphology, thus improving the overall recognition accuracy and stability.

[0059] Specifically, step 7 includes: Step S7, based on the point cloud purification, structure extraction, component fine classification, and dropper defect identification completed in steps S1 to S6, performs a quantitative analysis of the wear state of the contact wire body. This step uses the finely classified contact wire point cloud subset and the recovered contact wire catenary trajectory as geometric benchmarks. Through cross-section extraction, standard contour alignment, and directed distance calculation, it achieves precise quantification of wear depth and forms a wear distribution sequence along the line direction.

[0060] In step S71, based on the catenary trajectory of the contact line obtained in step S3, it is discretized at equal intervals in its curve parameter space according to a preset sampling step size to obtain multiple sampling section positions. Each sampling section position corresponds to a local cross-sectional analysis region on the contact line. For any sampling section position, with that position as the center point, the tangent direction vector of the catenary at that point is calculated, and the normal direction perpendicular to the tangent direction is used as the section extraction direction. This normal direction defines the plane where the local section is located, ensuring that the section is always orthogonal to the contact line axis, thereby ensuring that sections at different positions have consistent geometric meaning. Based on this, a section thickness threshold is set, and all point clouds that meet the following conditions are selected from the refined contact line point cloud subset: their spatial distance to the sampling section position is less than the section thickness threshold, and they are located within a local range on both sides of the normal direction. All point clouds that meet the conditions together constitute the contact line cross-sectional point cloud contour corresponding to the sampling section position. In this way, the three-dimensional continuous point cloud can be decomposed into a series of local two-dimensional sections along the line direction, providing standardized input for wear analysis.

[0061] In S72, a standard contact line design profile is introduced as a comparison benchmark. This standard profile is derived from the design cross-sectional geometry of the contact line under wear-free conditions, typically represented as a high-precision curve or discrete point set. To eliminate deviations in spatial position, orientation, and scale of the actual collected point cloud, rigid body registration is required between the point cloud profile of each contact line cross-section and the standard design profile. An iterative nearest-point method is employed, repeatedly establishing point-to-point correspondences and minimizing the root mean square error between the two sets of points to solve for the optimal rigid body transformation parameters. This error can be expressed as: in For points in the cross-sectional point cloud, For the corresponding points in the standard profile, This represents the rigid body transformation function. Through iterative optimization, the error is converged to a minimum, thereby aligning the contact line cross-section point cloud profile with the standard design profile to the same coordinate system. This alignment process ensures that subsequent distance calculations have a unified reference datum.

[0062] In S73, after coordinate alignment, the degree of wear is quantitatively calculated. Using the standard contact line design profile as a reference surface, for each point cloud in the cross-sectional point cloud profile, its directed distance to the standard profile is calculated along the local normal direction. This distance not only reflects the degree of deviation of the point cloud from the standard profile but also distinguishes between concave and convex directions by its sign. For contact line wear, the focus is on the distance change in the inward concavity direction. The directed distances of all point clouds are statistically analyzed, and the maximum value is taken as the wear depth value at that sampling section location. The above process is repeated along all sampling section locations to obtain a set of wear depth sequences arranged in line order. This sequence fully describes the spatial wear distribution of the contact line, enabling precise location of local abnormal areas.

[0063] In S74, a threshold determination analysis is performed on the wear depth sequence. The maximum wear depth in the sequence is extracted and compared with a preset wear threshold. This threshold is set based on the contact wire material properties and safe operation standards. When the maximum wear depth exceeds this threshold, it is determined that the contact wire has an over-limit wear risk, triggering a wear damage alarm. The output alarm information includes the sampling section location coordinates corresponding to the maximum wear depth and the specific wear depth value, used to indicate the precise spatial location and severity of the damage.

[0064] Through the processing in step S7, the contact line is further transformed from a geometric structure representation into continuous wear state distribution information. This method utilizes the catenary trajectory to establish a unified spatial reference, eliminates the uncertainties brought about by the complex three-dimensional structure through cross-sectional analysis, and achieves quantitative comparison by combining standard contours, thus avoiding the insufficient accuracy problem of traditional methods based on overall morphology or experience-based judgment. Simultaneously, the construction of the wear depth sequence enables contact line defects not only to be detected but also to be precisely located and quantified, providing direct data support for contact network maintenance.

[0065] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for extracting and classifying defect features from point cloud data of key components of overhead contact lines, characterized in that, Includes the following steps: S1. Obtain the original point cloud data of the key components of the catenary collected by the vehicle-mounted lidar, and perform filtering, noise reduction and elevation threshold segmentation on the original point cloud data to obtain a point cloud dataset of catenary components including the contact wire, catenary wire, dropper and positioner. S2. Based on the point cloud dataset of the contact wire components, calculate the linear feature value and planar feature value of each point cloud under a preset neighborhood radius, and construct the geometric feature vector of each point cloud according to the numerical relationship between the linear feature value and the planar feature value. Input the geometric feature vector into a random forest classifier for initial classification to obtain the initial classification of contact wire point cloud subset, initial classification of catenary point cloud subset, initial classification of drop wire point cloud subset and initial classification of locator point cloud subset; S3. For the initial classification of the contact line point cloud subset and the initial classification of the catenary point cloud subset, extract the contact line catenary trajectory and the catenary catenary trajectory respectively, and project each point cloud in the initial classification of the dropper point cloud subset onto the contact line catenary trajectory and the catenary catenary trajectory respectively. Construct a dropper connection relationship index table according to the projection distance, and cluster the point clouds in the initial classification of the dropper point cloud subset located on the same dropper structure into independent dropper point cloud units according to the dropper connection relationship index table. S4. For the initial classification locator point cloud subset, extract the locator plane main body region and calculate the outer boundary contour of the locator plane main body region. Construct a locator exclusion mask with the outer boundary contour as a constraint. Perform spatial overlay analysis on the locator exclusion mask and the initial classification contact line point cloud subset, and remove the point clouds in the initial classification contact line point cloud subset that fall within the locator exclusion mask to obtain the refined classification contact line point cloud subset. S5. Using the refined classification of the contact line point cloud subset as a spatial index, traverse the independent dropper point cloud unit, extract the intersection area point cloud of each independent dropper point cloud unit and the refined classification of the contact line point cloud subset, calculate the spatial distribution width of the intersection area point cloud along the line direction and the offset in the vertical direction, and construct the dropper structural integrity feature vector and the dropper spatial attitude feature vector. S6. Input the structural integrity feature vector of the dropper and the spatial posture feature vector of the dropper into the dropper disease classification model to obtain the classification results of the normal state, the slack state and the broken state of the dropper. S7. Based on the refined contact line point cloud subset, extract the point cloud profile of the contact line cross section along the normal direction of the contact line catenary trajectory, calculate the wear depth sequence between the point cloud profile and the standard design profile, and output contact line wear defect alarm information when the maximum wear depth in the wear depth sequence exceeds the preset wear threshold.

2. The method for extracting and classifying defect features from point cloud data of key components of overhead contact lines according to claim 1, characterized in that, Step S1 includes: S11. Collect raw scanning point cloud data along the railway line using the vehicle-mounted lidar, and associate and store the spatial three-dimensional coordinates of each point cloud in the raw scanning point cloud data with the laser reflection intensity information to construct the raw point cloud data of the key components of the catenary. S12. The statistical filtering algorithm is used to remove outliers from the original point cloud data of the key components of the overhead contact system. The neighborhood search radius of each point cloud is set as the first neighborhood threshold. The number of neighboring point clouds within the neighborhood search radius is counted. Point clouds with a number of neighboring point clouds less than the first number threshold are identified as outlier noise points and removed to obtain filtered point cloud data. S13. Perform elevation value statistical analysis on the filtered point cloud data to obtain an elevation histogram of the filtered point cloud data in the vertical direction. Determine the ground elevation threshold based on the elevation distribution interval corresponding to the ground point cloud in the elevation histogram. Use a pass-through filtering algorithm to remove point clouds in the filtered point cloud data whose elevation values ​​are less than the ground elevation threshold to obtain the point cloud dataset of the contact wire components including the contact wire, catenary wire, dropper, and positioner.

3. The method for extracting and classifying defect features from point cloud data of key components of overhead contact lines according to claim 1, characterized in that, Step S2 includes: S21. Obtain each point cloud in the contact network component point cloud dataset, set N neighborhood radii arranged in ascending order to form a multi-scale radius set, and construct a local neighborhood point set for each point cloud using each neighborhood radius in the multi-scale radius set to obtain N local neighborhood point sets corresponding to each point cloud. S22. For each local neighborhood point set, the covariance matrix is ​​calculated using principal component analysis. The covariance matrix is ​​then decomposed into eigenvalues ​​to obtain a first eigenvalue, a second eigenvalue, and a third eigenvalue. The first eigenvalue is greater than the second eigenvalue, and the second eigenvalue is greater than the third eigenvalue. Based on the first eigenvalue, the second eigenvalue, and the third eigenvalue, the linear eigenvalue and planar eigenvalue of the local neighborhood point set are calculated to obtain N linear eigenvalues ​​and N planar eigenvalues ​​corresponding to each point cloud at N scales. S23. Concatenate the N linear feature values ​​and N planar feature values ​​corresponding to each point cloud to construct a 2N-dimensional geometric feature vector for each point cloud. Input the 2N-dimensional geometric feature vectors of all point clouds in the contact wire component point cloud dataset into a pre-trained random forest classifier. The random forest classifier outputs the component category label for each point cloud. The component category label includes the contact wire category label, the catenary cable category label, the dropper category label, and the locator category label. S24. Based on the component category labels, divide the contact wire component point cloud dataset into the initial classification contact line point cloud subset, the initial classification catenary point cloud subset, the initial classification dropper point cloud subset, and the initial classification locator point cloud subset.

4. The method for extracting and classifying defect features from point cloud data of key components of overhead contact lines according to claim 1, characterized in that, Step S3 includes: S31. For the initial classification of the contact line point cloud subset, a catenary model fitting algorithm based on spatial continuity constraints is adopted. An upper limit for the number of iterations and a distance tolerance threshold are set. The contact line catenary trajectory is extracted by least squares iterative fitting. The contact line catenary trajectory describes the three-dimensional orientation of the contact line in space in the form of catenary equation parameters. S32. For the initial classification of the catenary point cloud subset, the catenary model fitting algorithm based on spatial continuity constraints is adopted. The same upper limit of the number of iterations and the distance tolerance threshold as in step S31 are set. The catenary trajectory of the catenary is extracted by least squares iterative fitting. The catenary trajectory of the catenary describes the three-dimensional orientation of the catenary in space in the form of catenary equation parameters. S33. Project each point cloud in the initial classification dropper point cloud subset onto the contact line catenary trajectory and the catenary trajectory of the catenary cable, respectively. Calculate the first projection distance from each point cloud to the contact line catenary trajectory and the second projection distance to the catenary trajectory of the catenary cable. When both the first projection distance and the second projection distance are less than the connection distance threshold, record the point cloud in the dropper connection relationship index table and associate and store the first projection distance and the second projection distance corresponding to the point cloud. S34. Traverse the dropper connection relationship index table. Based on the projection distance relationship recorded in the dropper connection relationship index table, cluster the point clouds in the initial classification dropper point cloud subset that correspond to the same catenary trajectory interval of the same catenary and the same contact line catenary trajectory interval into the same independent dropper point cloud unit, and obtain multiple independent dropper point cloud units. Each independent dropper point cloud unit corresponds to the point cloud representation of a complete dropper structure.

5. The method for extracting and classifying defect features from point cloud data of key components of overhead contact lines according to claim 1, characterized in that, Step S4 includes: S41. For the initial classification locator point cloud subset, a region growth algorithm based on normal vector consistency clustering is adopted. First, the normal vector of each point cloud in the initial classification locator point cloud subset is calculated. The point cloud with the smallest curvature in the initial classification locator point cloud subset is selected as the initial seed point. The normal vector angle threshold and the growth distance threshold are set. Neighboring point clouds that satisfy the normal vector angle being less than the normal vector angle threshold and the spatial Euclidean distance being less than the growth distance threshold are included in the same clustering region. The growth is iterated until no new point clouds are added, and the main body region of the locator plane is extracted. S42. Perform two-dimensional convex hull calculation on the main body region of the locator plane to obtain the minimum convex polygon boundary of the main body region of the locator plane on the projection plane. Extend the minimum convex polygon boundary outward by a preset extension distance along the normal direction of the main body region of the locator plane to obtain the outer boundary contour of the main body region of the locator plane. S43. Using the outer boundary contour of the main body area of ​​the locator plane as the spatial constraint boundary, construct a locator exclusion mask. The locator exclusion mask is a binary spatial mask matrix. In the locator exclusion mask, the pixels located inside the outer boundary contour are assigned the first mask value, and the pixels located outside the outer boundary contour are assigned the second mask value. S44. Perform spatial overlay analysis on the locator exclusion mask and the initial classification contact line point cloud subset, traverse each point cloud in the initial classification contact line point cloud subset, project each point cloud in the initial classification contact line point cloud subset onto the projection plane where the locator exclusion mask is located, obtain the mask value corresponding to the projection point, and when the mask value is the first mask value, remove the point cloud from the initial classification contact line point cloud subset to obtain the refined classification contact line point cloud subset.

6. The method for extracting and classifying defect features from point cloud data of key components of overhead contact lines according to claim 1, characterized in that, Step S5 includes: S51. Using the refined contact line point cloud subset as a spatial index, traverse each independent dropper point cloud unit. For the currently traversed independent dropper point cloud unit, calculate the spatial centroid coordinates of all point clouds in the independent dropper point cloud unit. Using the spatial centroid coordinates as a reference point, set the forward search radius and the backward search radius along the line direction. Extract point clouds located within the range of the forward search radius and the backward search radius from the refined contact line point cloud subset to form the contact line local point cloud subset corresponding to the independent dropper point cloud unit. Perform spatial intersection between the contact line local point cloud subset and the independent dropper point cloud unit, and extract the point clouds that exist simultaneously in the contact line local point cloud subset and the independent dropper point cloud unit as the intersection area point cloud. S52. The principal component analysis method is used to calculate the principal direction features of the point cloud in the intersection area. Based on the principal direction features, the principal direction vector along the line direction is determined. All points in the point cloud in the intersection area are projected onto a plane perpendicular to the principal direction vector. The two-dimensional convex hull area of ​​the projected point cloud on the plane is calculated. The two-dimensional convex hull area is divided by the span value of the point cloud in the intersection area on the principal direction vector to obtain the spatial distribution width of the point cloud in the intersection area along the line direction. S53. Calculate the spatial Euclidean distance from each point cloud in the intersection region point cloud to the subset of the fine classification contact line point cloud. Statistically analyze the spatial Euclidean distances of all point clouds and take the difference between the maximum and minimum spatial Euclidean distances as the offset of the intersection region point cloud in the vertical direction. S54. After normalizing the spatial distribution width of the intersection area point cloud along the line direction and the offset of the intersection area point cloud in the vertical direction, the points are spliced ​​together to construct the dropper structural integrity feature vector and dropper spatial attitude feature vector corresponding to each independent dropper point cloud unit.

7. The method for extracting and classifying defect features from point cloud data of key components of overhead contact lines according to claim 1, characterized in that, Step S6 includes: S61. Obtain multiple historical independent dropper point cloud units with labeled defect types from historical catenary inspection data. For each historical independent dropper point cloud unit, extract the historical dropper structural integrity feature vector and the historical dropper spatial attitude feature vector according to the method in step S5. After concatenating the historical dropper structural integrity feature vector and the historical dropper spatial attitude feature vector, associate and store them with the corresponding defect type label to construct a dropper defect training sample set. S62. Construct a classification model for string disease. The classification model for string disease uses a multi-class support vector machine as the classifier kernel. The kernel function of the multi-class support vector machine is set as a radial basis function kernel function. The kernel parameters and penalty parameters of the radial basis function kernel function are set. The string disease training sample set is input into the string disease classification model for training. The kernel parameters and penalty parameters are optimized by a grid search algorithm until the classification accuracy of the string disease classification model on the validation set reaches a preset accuracy threshold, thus obtaining a pre-trained string disease classification model. S63. The feature vector of the integrity of the dropper structure and the feature vector of the spatial posture of the dropper are concatenated and then input into the pre-trained dropper defect classification model. The pre-trained dropper defect classification model outputs one of the following as the classification result: dropper normal state category label, dropper slack state category label, and dropper broken state category label.

8. The method for extracting and classifying defect features from point cloud data of key components of overhead contact lines according to claim 1, characterized in that, Step S7 includes: S71. Based on the refined contact line point cloud subset, multiple sampling cross-section positions are set at equal intervals along the contact line catenary trajectory with a preset sampling step size. For each sampling cross-section position, with the sampling cross-section position as the center, the normal direction perpendicular to the tangent direction of the contact line catenary trajectory is set as the cross-section extraction direction. A cross-section thickness threshold is set. All point clouds in the refined contact line point cloud subset that are spatially less than the cross-section thickness threshold and located within the range on both sides of the normal direction are extracted, thus forming the contact line cross-section point cloud contour corresponding to the sampling cross-section position. S72. Obtain the standard contact line design profile, which is a geometric profile template of the standard contact line cross-section in a wear-free state. Iterate the nearest point registration between the point cloud profile of each contact line cross-section and the standard contact line design profile. Calculate the rigid body transformation parameters that minimize the root mean square error between the point cloud profile of the contact line cross-section and the standard contact line design profile. Align the point cloud profile of the contact line cross-section and the standard contact line design profile to the same coordinate system according to the rigid body transformation parameters. S73. In the aligned coordinate system, calculate the directed distance from each point cloud in the point cloud profile of the contact line section to the standard contact line design profile along the normal direction of the standard contact line design profile. Take the maximum value of the directed distance as the wear depth value at the sampling section position. Traverse all sampling section positions to obtain the wear depth sequence. S74. Compare the maximum wear depth in the wear depth sequence with a preset wear threshold. When the maximum wear depth is greater than the preset wear threshold, output contact wire wear fault alarm information. The contact wire wear fault alarm information includes the sampling section position coordinates corresponding to the maximum wear depth and the maximum wear depth value.

Images