An on-line identification method for insulation layer crack of overhead insulated cable

Abstract

The application discloses an online identification method for cracks in an insulation layer of an overhead insulated cable, and relates to the technical field of crack identification. The method comprises the following steps: collecting initial insulation layer profile data, and establishing a reference microtopography model; collecting real-time insulation layer profile data, and aligning the real-time insulation layer profile data with the reference model in space through an iterative nearest point algorithm to obtain aligned insulation layer profile data; extracting a crack feature vector of the aligned insulation layer profile data, and comparing the crack feature vector with a threshold value to determine a candidate crack region; screening the candidate region through a periodic stability determination mechanism to obtain a real crack; and constructing a prediction model based on a long short-term memory network, inputting real crack time series data into the prediction model, and outputting a crack trend prediction result. The application divides the cracks into risk levels based on the crack trend prediction result, and realizes online identification of the cracks in the insulation layer.

Description

Technical Field

[0001] This invention relates to the field of crack identification technology, specifically to an online method for identifying cracks in the insulation layer of overhead insulated cables. Background Technology

[0002] In the operation and maintenance management of overhead insulated cables, achieving early and accurate identification and prediction of the evolution trend of insulation layer cracks is a key technical aspect of ensuring the safe and stable operation of the power grid and avoiding major accidents caused by insulation failure. Especially for lines deployed in complex natural environments, developing an online, automated, and reliable crack monitoring method has significant engineering value.

[0003] Currently, a typical online method for identifying insulation layer cracks in engineering practice is based on two-dimensional image recognition. This method uses a high-definition camera mounted on an inspection robot to acquire two-dimensional digital images of the insulation layer surface along the cable route. Subsequently, digital image processing algorithms such as edge detection and texture analysis, or deep learning models trained on a large number of samples, are used to identify and segment texture abnormalities in the image, thereby locating potential crack positions. However, this two-dimensional image recognition method, due to its inherent reliance on two-dimensional projection information, cannot directly obtain the three-dimensional geometric features of the crack in the depth direction. This makes it difficult to establish a baseline contour and real-time contour model that can be used for accurate comparison. Ultimately, this results in an inherent deficiency in distinguishing real cracks from visual artifacts caused by lighting shadows, stains, etc., seriously affecting the reliability of the identification results. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an online method for identifying insulation layer cracks in overhead insulated cables, solving the problem of the inability to effectively integrate the inherent, stable, and unchanging macroscopic spatial anchor points on the cable surface as registration constraints.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an online identification method for insulation layer cracks in overhead insulated cables, comprising the following steps: Step S1: Obtain the initial insulation layer contour data of the overhead insulated cable and establish a reference micro-morphology model of the insulation layer; Step S2: Obtain real-time insulation layer contour data. Based on the reference micro-topography model, use the iterative nearest point algorithm to spatially align the real-time insulation layer contour data to obtain aligned insulation layer contour data. Step S3: Extract features from the aligned insulation layer contour data to obtain crack feature vectors, and compare the crack feature vectors with a preset threshold to determine candidate crack regions; Step S4: Construct a periodic stability determination mechanism, and use the periodic stability determination mechanism to determine the cracks in the candidate crack region and filter out the real cracks. Step S5: Based on a long short-term memory network, construct a crack trend prediction model, obtain the characteristic time-series data of the real crack, input the characteristic time-series data of the real crack into the crack trend prediction model, output the crack trend prediction result, and realize online identification of insulation layer cracks based on the crack trend prediction result.

[0006] Preferably, obtaining the initial insulation layer profile data of the overhead insulated cable includes: Select several cross-sectional sections along the length of the target overhead insulated cable. Where k is the index of the cross-sectional segment order, and K is the total number of cross-sectional segments; in each segment The inner intervals are arranged at fixed circumferential angles. and axial spacing Plan out dense scanning trajectories ,in For trajectory indexing Used to plan the scanning trajectory around the cable surface. Used to plan the scanning trajectory along the length of the cable.

[0007] Then, using the micro-surface profile acquisition components deployed on the inspection device, the laser line scanner strictly follows the planned trajectory. The outer surface of the insulation layer is scanned. The initial contour data of the insulation layer obtained in this process is a discrete three-dimensional point cloud, which can be represented as:

[0008] in, Representative at the Section, No. The collection of raw point clouds collected along the trajectory. These are the coordinates of a single 3D point in a point cloud; It is a single data point in the set, representing the coordinates of a specific location on the surface of the insulating layer measured in three-dimensional space; It is a point coordinates, These represent the three coordinate components of the point in the three-dimensional coordinate system; the superscript T indicates transpose. This symbol represents a point. It belongs to the three-dimensional real number space; It is an index variable used to iterate through each point in the set; This indicates that in this specific scan trajectory The total number of points collected.

[0009] Preferably, establishing a baseline micromorphological model of the insulating layer includes: For two adjacent point clouds and An optimal rigid body transformation is found by using the iterative nearest point algorithm. ,in It is a 3x3 rotation matrix. It is a 3x1 translation vector. The goal of this transformation is to minimize the Euclidean distance between corresponding points. Its mathematical expression is:

[0010] in, and These are the optimal rotation matrix and optimal translation vector, respectively, obtained through iterative optimization algorithms. It is the source cloud The first in A three-dimensional point, It is a target point cloud Midpoint The corresponding nearest neighbor will be determined by the algorithm. Find the nearest point in another point cloud as... ; For point pairs participating in the calculation The total quantity; The optimal rotation matrix is ​​obtained by iteratively solving this optimization problem. and optimal translation vector This transforms the source point cloud into the coordinate system of the target point cloud.

[0011] in, Points in the source point cloud After optimal transformation and The new coordinates obtained later; Finally, by sequentially executing the above registration process, all local point clouds are... Unified transformation to global coordinate system Below, the registered point cloud is obtained. All By integrating the data, a baseline micromorphological model of the cable under healthy conditions was established. .

[0012] Preferably, obtaining real-time insulation layer contour data includes: During real-time inspection, the laser line scanner strictly follows the planned trajectory, utilizing the micro-surface profile acquisition components deployed on the inspection device. The outer surface of the insulation layer is scanned to obtain real-time contour data of the cable insulation layer. .

[0013] Preferably, the iterative nearest-point algorithm is used to spatially align the real-time insulation layer contour data, resulting in aligned insulation layer contour data including: An improved Iterative Closest Point (ICP) algorithm is used for fine registration. The objective function of this algorithm adds a spatial anchor point constraint term to the standard point-pair distance term to improve the stability and accuracy of the registration.

[0014] in, It is the fine transformation matrix to be solved; It refers to points in a real-time point cloud. yes The corresponding nearest neighbor; This indicates the number of points in the real-time point cloud that participated in the registration. This is the total number of anchor point pairs; This is a standard ICP data item used to minimize the distance of the overall point cloud. It is a weighting factor used to adjust each point pair The importance of this, based on the points of reference The matching quality is determined; It is an anchor point constraint term. and These are matched spatial anchor point pairs. This ensures that these significant feature point pairs are always pulled towards alignment during the optimization process, enhancing the robustness of the algorithm. and These are hyperparameters used to balance the weights of data items and anchor constraint items, and are adjusted experimentally. It is a spatial anchor index; Let be the transformation function, representing the substitution of the source points into the optimal transformation matrix; by iteratively optimizing this objective function, the optimal high-precision transformation matrix is ​​finally obtained. ; Finally, the optimal transformation is applied to correct the real-time insulation layer profile data:

[0015] in, It is the same as the baseline micro-morphology model Precisely aligned insulation layer contour data. It is real-time insulation layer contour data.

[0016] Preferably, feature extraction of the aligned insulating layer contour data to obtain the crack feature vector includes: for Each sampling point in and in Corresponding points in Calculate the following geometric features, including height residual, local curvature variation, normal vector angle deviation, and local plane fitting residual: For the height residual, calculate the absolute height deviation between the current point and the reference point:

[0017] in, Indicates the first The height residual of each point; Indicates real-time micromorphological data The first in A three-dimensional point; Indicating in the baseline micromorphology model Zhongyu The exact corresponding number A three-dimensional point; This is a function that retrieves the normal direction coordinates of a point; the Z-axis of the preset scan is roughly aligned with the normal direction of the cable surface, thus extracting the point coordinates. In Quantity; The formula for calculating curvature change is:

[0018] Local curvature change at a point for The smallest eigenvalue, for The smallest eigenvalue; For the deviation of the normal vector angle, first reuse the neighborhood point set calculated by curvature. and covariance matrix After decomposition, the smallest eigenvalue The corresponding eigenvector is the direction of the surface normal vector. If the eigenvector points to the outside / inside of the cable and is opposite to the reference normal vector, the correction direction is determined by the dot product to ensure that it is consistent with the orientation of the reference normal vector; the normal vector is divided by its magnitude to obtain the unit normal vector.

[0019] in, The unit normal vector in real-time micro-topography data. The unit normal vector in the baseline micro-topography model. for of Corresponding feature vectors; for of For the corresponding eigenvectors, ||…|| are norm symbols, which are used here to calculate the magnitude of the vector; Then, calculate the angle between the surface normal vectors at the two points:

[0020] in, This represents the angle deviation between the normal vectors at the i-th point. It is an inverse cosine function, and the corresponding angle is calculated based on the cosine value. Texture breakage or misalignment will cause a significant change in the direction of the normal vector. For the local plane fitting residuals, the points are calculated using the least squares method. and local neighborhood point set and Find the best-fitting plane, and then calculate the difference in average distances from the point cloud to its fitted plane:

[0021] in, This represents the change in the local plane fitting residual in the neighborhood of the i-th point; Point The set of neighborhood points; Point The set of neighborhood points; Represented by neighborhood point set The best-fit plane; Represented by neighborhood point set The best-fit plane; is the distance from the point to the plane; k is the number of points in the neighborhood; Since each feature targets a specific geometric distortion that a crack may cause, and Z-score normalization is used to transform the multi-dimensional crack features into dimensionless values ​​in terms of standard deviation, a fair benchmark is established for anomaly detection based on a unified statistical threshold. Together, these features constitute the crack feature vector for identifying candidate cracks. in, The average value of the high residuals in the healthy sample serves as the benchmark. The standard deviation of high residuals in healthy samples; The average value of the curvature change in healthy samples serves as a benchmark. The standard deviation of the curvature change in healthy samples is used as a measure. Typical average of the angle between the normal vectors of healthy samples; The natural fluctuation range of the angle between the normal vectors of healthy samples; The standard deviation scale is used to measure the local plane fitting residuals of healthy samples. is the standard deviation scale of the local plane fitting residuals of healthy samples, which are sample points in a healthy state randomly extracted from the baseline model.

[0022] Preferably, comparing the crack feature vector with a preset threshold to determine candidate crack regions includes: Based on crack feature vector To identify outliers, a threshold is set for each feature dimension. If any feature value of a point exceeds its corresponding threshold, then that point is marked as a preliminary outlier.

[0023] in, This is a preliminary outlier detection function. =1 indicates that the i-th point is marked as an initial outlier. =0 indicates that the i-th point is marked as a normal point; It is the high residual threshold; It is the curvature change threshold; It is the normal vector deviation threshold; It is the roughness variation threshold, if the local plane fitting residual of the point If the value exceeds this threshold, it is considered abnormal. All four thresholds are preset constants, derived by analyzing historical data. To avoid misclassifying isolated noise points or minute attachments as cracks, a spatial continuity constraint is introduced. Only regions with a continuous distribution of initial anomalies are identified as candidate crack regions. .

[0024] Preferably, a periodic stability determination mechanism is constructed, which is used to determine the cracks in the candidate crack region and screen out the real cracks, including: For each candidate region Create a time series archive Record its current and historical significance The state of existence in each inspection cycle, and a binary variable is defined. Indicates the region In the cycle Was it detected?

[0025] This is an existence indicator function, a binary variable used to label candidate regions. In a specific Whether it was detected during each inspection cycle; It means the first One candidate crack region; This indicates the time cycle number of the inspection. Index of the candidate crack region order; Next, calculate the region in Total number of cycles within a given period:

[0026] in, Indicates candidate region exist The total number of cycles within a given period; This indicates the total number of inspection cycles used for judgment. Greater than or equal to a preset persistence threshold ,Right now Then the candidate crack region is determined. The cracks are identified as real cracks and added to the set of real crack regions. Otherwise, the region is determined to be an occasional disturbance and is removed from the candidate set.

[0027] Preferably, the crack trend prediction model based on long short-term memory networks includes: The crack trend prediction model is a long short-term memory network based on an attention mechanism, for each real crack... From its time series archives Get the past The characteristic time-series data of each continuous inspection cycle are used as the model input; An attention mechanism is introduced at the top layer of the standard LSTM encoder, enabling the model to dynamically focus on historical key points most relevant to future predictions. This mechanism is calculated as follows: First, calculate the attention score for each time step of the LSTM encoder output. Hidden state Calculate its correlation score :

[0028] in, It is a learnable parameter matrix whose function is to perform a linear transformation on the concatenated long vector, mapping it to a new vector space; It is a learnable parameter row vector that acts on the result after tanh activation, mapping it from a vector to a scalar; It is the final state after the encoder has processed the entire input sequence; it encodes summary information of the entire sequence. ( This is a nonlinear activation function that performs a nonlinear mapping on the result of a linear transformation; T is the transpose sign. Then, a weighted context vector is generated, and the scores are normalized into attention weights using the Softmax function. The hidden states are then weighted and summed to obtain the context vector:

[0029] Where c is the context vector. and Both represent the index of the time step, with values ​​ranging from 1 to the total sequence length L; It is an exponential function; The final predicted output is obtained by using the context vector. The hidden state of the last time step splicing, outputting the future through a fully connected layer. Predicted value of the cycle :

[0030] in, It is a learnable parameter matrix that performs a linear transformation on the concatenated long feature vector, mapping it to the dimension of the prediction target; This is the output bias vector; To predict the step size.

[0031] Preferably, based on the crack trend prediction results, online identification of insulation layer cracks includes: For each real crack The feature sequence is input into the trained model to obtain its predicted output. ; Subsequently, a comprehensive risk classification is performed, combining the instantaneous geometric characteristics of the crack with its dynamic propagation trend to assess the risk. The formula for calculating the risk index RI is as follows:

[0032] in, This indicates the rate at which the prediction is normalized and expanded. Step S4 represents the actual crack. The normalized multidimensional geometric feature vector of the current state; Let be the norm of this vector, used to quantify the current overall severity of the crack. These are weighting coefficients used to balance the impact of expansion speed and current severity. They are preset parameters determined after training and optimization based on historical crack data. Based on the calculated risk index They are classified into different risk levels as follows: High risk Medium risk Low risk ,in, and It is the threshold for the risk level; Ultimately, a structured early warning message is generated for each crack. This message not only includes the location and risk level of the crack area, but more importantly, it includes trend judgment based on the prediction model, thereby realizing online identification of insulation layer cracks.

[0033] This invention provides an online method for identifying insulation layer cracks in overhead insulated cables, involving machine learning and deep learning technologies, which has the following beneficial effects: (1) The online identification method for insulation layer cracks in overhead insulated cables uses an iterative nearest point algorithm to spatially align the real-time insulation layer contour data with the benchmark micro-morphology model, effectively solving the problem of spatial position deviation caused by cable posture changes, wind-induced swaying, etc., and making the contour data of different inspection cycles have accurate spatial comparability, providing a unified spatial reference basis for subsequent crack feature extraction and identification, and greatly improving the spatial accuracy of crack identification. The traditional method has a crack feature matching accuracy of only 65% ​​due to spatial deviation, while the matching accuracy of the method of this invention is improved to 92%, and the spatial alignment accuracy is improved by 27 percentage points.

[0034] (2) The online identification method for insulation layer cracks in overhead insulated cables, by introducing multi-objective function optimization with spatial anchor point constraints, further enhances the stability and robustness of spatial alignment based on the iterative nearest point algorithm, effectively avoiding alignment failure caused by local non-rigid deformation of the cable surface, so that spatial alignment is not only accurate in macroscopic posture, but also consistent in microscopic local features, further improving the reliability of crack identification. The feature alignment error rate of the traditional iterative nearest point algorithm is 8%, while the error rate of the method of this invention is reduced to 3%, and the alignment robustness is improved by 5 percentage points.

[0035] (3) The online identification method for insulation layer cracks in overhead insulated cables uses an LSTM prediction model with attention mechanism to predict crack trends. It can adaptively focus on key change nodes in the crack feature time series data and accurately capture the time evolution law of crack propagation. Compared with the traditional time series prediction model ARIMA, the prediction accuracy and foresight of crack propagation trend are significantly improved, providing more valuable predictive information for operation and maintenance decisions.

[0036] (4) The online identification method for insulation layer cracks in overhead insulated cables effectively solves the problem of excessive sensitivity of the traditional loss function MSE to outliers in crack feature time series data, measurement errors in a certain inspection, and sudden crack expansion when the training process of the crack trend prediction model is optimized by adopting Huber loss function. This makes the model have stronger anti-interference ability while ensuring prediction accuracy and improves the stability of crack trend prediction. Attached Figure Description

[0037] Figure 1 This is a flowchart of an online method for identifying insulation layer cracks in overhead insulated cables, as proposed in this invention.

[0038] Figure 2 The present invention provides a hierarchical diagram of aligned insulation layer contour data for an online identification method for insulation layer cracks in overhead insulated cables.

[0039] Figure 3 This is a hierarchical diagram showing the crack trend prediction results obtained in the online identification method for insulation layer cracks in overhead insulated cables proposed in this invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Please see Figure 1-3 This invention provides a technical solution: an online method for identifying insulation layer cracks in overhead insulated cables. Specifically, the following online method for identifying insulation layer cracks in overhead insulated cables is provided; please refer to [link / reference]. Figure 1 The method includes the following steps: Step S1: Obtain the initial insulation layer contour data of the overhead insulated cable and establish a reference micro-morphological model of the insulation layer.

[0042] First, this step aims to establish a high-precision, long-term reusable benchmark micro-morphology model as the basis for comparison in subsequent data identification and processing. Therefore, the establishment of the benchmark micro-morphology model must be based on data collected when the target overhead insulated cable is in a healthy state, preferably in the initial stage after cable installation.

[0043] Specifically, several cross-sectional sections are selected along the length of the target overhead insulated cable. Where k is the index of the cross-sectional segment order, and K is the total number of cross-sectional segments; in each segment Inside, at fixed circumferential angular intervals (Unit: degrees) and axial spacing (Unit: mm) Plan out dense scanning trajectories ,in For trajectory indexing Used to plan the scanning trajectory around the cable surface. Used to plan the scanning trajectory along the length of the cable.

[0044] Then, using a micro-surface profile acquisition component and a laser line scanner deployed on the inspection device, the system strictly follows the planned trajectory. The outer surface of the insulation layer is scanned. The initial insulation layer contour data obtained in this process is a discrete three-dimensional point cloud, which can be represented as:

[0045] in, Representative at the Section, No. The collection of raw point clouds collected along the trajectory. These are the coordinates of a single 3D point in a point cloud; It is a single data point in the set, representing the coordinates of a specific location on the surface of the insulating layer measured in three-dimensional space; It is a point coordinates, These represent the three coordinate components of the point in a three-dimensional coordinate system; the superscript T indicates transpose, which here means it will be written as the coordinates of a row vector. Represented as a column vector, this is the standard mathematical form for handling point coordinates; This symbol represents a point. It belongs to the three-dimensional real number space; It is an index variable used to iterate through each point in the set; This indicates that in this specific scan trajectory The total number of point clouds collected.

[0046] Because the device pose may differ during each scan, several independent point clouds are acquired. They exist in their own independent local coordinate systems. To unify them into a single global coordinate system... To construct a complete and consistent baseline micro-topography model, point cloud registration is required.

[0047] Specifically, for two adjacent point clouds and (Or with a registered reference point cloud), an optimal rigid body transformation is found using the iterative nearest-point algorithm. ,in It is a 3x3 rotation matrix. It is a 3x1 translation vector. The goal of this transformation is to minimize the Euclidean distance between corresponding points, and its mathematical expression is:

[0048] in, and These are the optimal rotation matrix and optimal translation vector, respectively, obtained through iterative optimization algorithms. It is the source cloud The first in A three-dimensional point, It is a target point cloud Midpoint The corresponding nearest neighbor will be determined by the algorithm. Find the nearest point in another point cloud as... ; The optimal rotation matrix is ​​obtained by iteratively solving this optimization problem. and optimal translation vector This transforms the source point cloud into the coordinate system of the target point cloud.

[0049] in, Points in the source point cloud After optimal transformation and The new coordinates obtained later; Finally, by sequentially executing the above registration process, all local point clouds are... Unified transformation to global coordinate system Below, the registered point cloud is obtained. All By integrating these elements, a baseline micromorphological model of the cable under healthy conditions can be established. This model is a high-precision three-dimensional digital surface within a unified spatial reference framework, providing an accurate geometric comparison benchmark for subsequent online inspections.

[0050] Step S2: Obtain real-time insulation layer contour data. Based on the reference micro-topography model, use the iterative nearest point algorithm to spatially align the real-time insulation layer contour data to obtain aligned insulation layer contour data.

[0051] During real-time inspection, the laser line scanner strictly follows the planned trajectory, utilizing the micro-surface profile acquisition components deployed on the inspection device. The outer surface of the insulation layer is scanned to obtain real-time contour data of the cable insulation layer. However, the orientation of overhead insulated cables may shift due to wind-induced swaying, temperature expansion and contraction, or tower deformation, requiring real-time insulation profile data. The baseline micromorphology model established in step S1 There are significant differences in spatial pose between them, making it impossible to make a direct and detailed comparison between the two.

[0052] Therefore, the core task of this step is to achieve... and To unify the high-precision coordinate system, this step does not simply apply the standard iterative nearest point algorithm, but introduces a new approach from... Spatial anchor points extracted Key improvements were made to the algorithm to address the problems of the iterative nearest point algorithm being sensitive to the initial position and easily getting trapped in local optima.

[0053] In the baseline micromorphology model After the system is established, a set of stable feature points that remain unchanged throughout the cable's lifecycle needs to be defined for subsequent registration, i.e., the spatial anchor point set. The stable and unchanging relative reference is the inherent macroscopic structure formed during the cable manufacturing stage (such as mold seams and embossed sheath lettering). Under the premise of normal cable operation and maintenance and no physical damage (impact, tearing, strong abrasion), the relative geometric positions (circumferential angle, longitudinal spacing) of these feature points on the insulation surface remain permanently unchanged and can serve as the constraint reference for long-term registration, spatial anchor point set. The calculation process is as follows: For the model any point on the surface This requires calculating the local curvature within its neighborhood. An eigenvalue method based on covariance analysis is used for estimation. First, the point is determined. of The nearest neighbor points constitute its neighborhood point set. Then calculate the covariance matrix of the neighborhood. :

[0054] in, It is a neighborhood point set The point in the middle, It is the centroid of the neighboring points. is the number of nearest neighbors, and j is used as the index variable for the summation loop in this formula.

[0055] Next, the covariance matrix Eigenvalue decomposition yields three eigenvalues. (Sorted from smallest to largest, i.e.) ),point The change in surface curvature (an efficient approximation of curvature). It can be derived from the smallest eigenvalue The calculation yielded:

[0056] in, The change in surface curvature represents the point. The degree of curvature of the surface. The larger the value, the greater the deviation of the neighborhood of that point from the tangent plane, and it may be located in a feature region with large curvature such as convexity, depression, or edge. The minimum eigenvalue is used. Characteristic points The principle for determining the degree of surface curvature is based on the eigenvalues ​​of the covariance matrix. Reflects the dispersion of the distribution of neighborhood points. Minimal (close to 0); neighborhood points on curved surfaces (such as anchor points, cracks) deviate from the tangent plane. Significantly increased. Therefore, It can effectively distinguish between flat and curved areas, and reasonably characterize the degree of curvature.

[0057] Then, based on the calculated surface changes at each point... To filter candidate feature points, for a point , and With all its nearest neighbors The changes in the surface were compared.

[0058] If point of It is significantly greater than the corresponding values ​​of all its neighboring points, thus satisfying the condition:

[0059] Or it is significantly smaller than the corresponding values ​​of all its neighboring points, thus satisfying the condition:

[0060] in, It is a preset tolerance threshold used to define "significance"; Neighboring points The curvature metric value; points that satisfy one of the above conditions. These points are marked as candidate feature points. These points correspond to locations with significant geometric features, such as mold seams, edges created by printing on the sheath, and concave / convex vertices.

[0061] Calculate the surface variation of all candidate feature points that meet the above conditions. (Based on minimum eigenvalue) ),according to Sort the data from largest to smallest, select the top 20% of highly significant points, and sample the cable cross-sections to ensure spatial uniformity. The circumferential direction (0~360°) is divided into Equal parts (e.g.) (divided into 1 part at 45° intervals) and longitudinally (length direction) into Equal parts (e.g.) (1 copy per 200 mm), at least one highly significant point is selected within each "circumferential-longitudinal" grid; for the initially selected points, their curvature change is checked in 3 consecutive scans (change <5%), unstable points (such as temporary high curvature points caused by surface deposits) are removed, and the final selection is made. Points (each cross-sectional section) Select 5 to 10 points to form a spatial anchor point set. superscript The term "base" indicates that it belongs to the baseline micromorphological model.

[0062] To provide a good initial value for subsequent fine registration, it is necessary to use real-time insulation layer contour data. Quickly identify spatial anchor point sets The corresponding point.

[0063] Using the same eigenvalue method based on covariance analysis as described above, for A rapid scan is performed to identify a set of feature points. Then, a baseline anchor point set is established using FPFH (Fast Point Feature Histogram) combined with the RANSAC (Random Sample Consensus) algorithm. With real-time anchor set Precise matching to form a real-time spatial anchor point set The specific process is as follows: right and Each anchor point in and Based on neighborhood radius (Adjustable based on point cloud density) Determine the normal vectors of points in its neighborhood, and statistically analyze the distribution of the angle between the normal vectors of points in the neighborhood and the distance from the point to the anchor point, generating an 11-dimensional FPFH descriptor, denoted as the baseline descriptor. and real-time descriptors .

[0064] calculate Each anchor point in Euclidean distance of all anchor points in the descriptor:

[0065] in, ( () is the similarity function, which is used to calculate the similarity between two points by comparing the differences in their FPFH descriptors. Select the one with the smallest distance As the initial matching pair.

[0066] Then, robust matching is performed using the RANSAC algorithm. Three initial matching pairs are randomly selected, and statistical transformations are performed to ensure that the results satisfy the following conditions. The number of interior points; iterate 1000 times, retain the transformation matrix with the highest interior point rate, and use its corresponding matching pair as the final result.

[0067] Next, by solving a rigid body transformation, we can... and Alignment. This transformation. By a rotation matrix and a translation vector The composition is obtained by minimizing the following objective function:

[0068] in, It is the j-th anchor point in the real-time point cloud. It is the j-th anchor point in the reference point cloud. It is the total number of anchor point pairs. and The optimal rotation matrix and translation vector constitute the initial coarse registration transformation. Real-time insulation layer contour data Initial alignment to the reference coordinate system lays the foundation for the next step of fine registration and effectively avoids getting trapped in local optima.

[0069] Then, starting from the initial transformation described above, an improved Iterative Closest Point (ICP) algorithm is used for fine registration. The objective function of this algorithm adds a spatial anchor point constraint term to the standard point-pair distance term to improve the stability and accuracy of the registration.

[0070] in, It is the fine transformation matrix to be solved; It refers to points in a real-time point cloud. yes The corresponding nearest neighbor; This indicates the number of points in the real-time point cloud that participated in the registration. This is the total number of anchor point pairs; This is a standard ICP data item used to minimize the distance of the overall point cloud. It is a weighting factor used to adjust each point pair The importance of this, based on the points of reference The matching quality is determined; It is an anchor point constraint term. and These are matched spatial anchor point pairs. This ensures that these significant feature point pairs are always pulled towards alignment during the optimization process, enhancing the robustness of the algorithm. and These are hyperparameters used to balance the weights of data items and anchor constraint items, and are adjusted experimentally. It is a spatial anchor index; Let be the transformation function, representing the substitution of the source points into the optimal transformation matrix; by iteratively optimizing this objective function, the optimal high-precision transformation matrix is ​​finally obtained. .

[0071] It should be noted that, and The stability and invariance refer to the invariance of local geometric features (curvature, normal vector), rather than the invariance of the global spatial position. The wind-induced swaying and temperature expansion and contraction of overhead cables can lead to... Position offset and transformation in the global coordinate system Its function is to correct the offset, so that and Global coordinate alignment, therefore It is usually not 0, and only approaches 0 when the cable has no attitude deviation.

[0072] Finally, the optimal transformation is applied to correct the real-time insulation layer profile data:

[0073] in, It is the same as the baseline micro-morphology model Precisely aligned insulation layer contour data. It is real-time insulation layer contour data.

[0074] This step addresses the issue of pose discrepancies between real-time scan data and the baseline model caused by cable swaying due to wind and thermal expansion and contraction. Its innovation lies in improving the traditional iterative nearest-point algorithm. Starting from predefined, stable, and unchanging "spatial anchor points" (such as seams or logo protrusions) in the baseline model, it quickly matches corresponding anchor points in the real-time data, calculating an initial coarse registration transformation. This provides a high-quality starting point for subsequent fine registration, avoiding getting trapped in local optima. Then, in the objective function of fine registration, not only is the overall point cloud distance minimized, but an anchor point constraint term is also added, forcing these significant feature points to align, thus significantly improving the accuracy and robustness of the registration. This ensures spatial comparability of data collected at different time points, accurately "resetting" data "misaligned" due to environmental disturbances, and providing the geometric basis for the point-by-point fine comparison in step S3.

[0075] Step S3: Extract features from the aligned insulation layer contour data to obtain crack feature vectors, and compare the crack feature vectors with a preset threshold to determine candidate crack regions.

[0076] After successfully aligning the real-time micro-topography data with the baseline micro-topography model in step S2, the goal of this step is to identify potentially crack-related local anomaly regions from the massive dataset through meticulous comparison. The core of this step lies in extracting geometric features that characterize material damage and focusing on candidate crack regions based on these features.

[0077] Specifically, instead of simply comparing height differences, features are extracted from the aligned data from several geometric dimensions to enhance sensitivity to early micro-cracks; Since the spatial alignment in step S2 only addresses the global pose deviation between the real-time point cloud and the baseline model, but cannot guarantee a direct correspondence between point-by-point indices, the following steps are required to determine the corresponding points. : Set search radius (Ensure the deviation range of the corresponding theoretical points is covered); then in the benchmark model In China, search and The Euclidean distance is the smallest and The point, as the corresponding point If there is no reference point within the search radius (e.g., local data of the reference model is missing), mark the real-time point as a point to be verified and temporarily exclude it from feature extraction to avoid misjudgment without a reference.

[0078] for Each sampling point in and in Corresponding points in Calculate the following geometric features, including height residual, local curvature variation, normal vector angle deviation, and local plane fitting residual: For the height residual, calculate the absolute height deviation between the current point and the reference point:

[0079] in, Indicates the first The height residual of each point; Indicates real-time micromorphological data The first in A three-dimensional point. Indicating in the baseline micromorphology model Zhongyu The exact corresponding number A three-dimensional point; This is a function that retrieves the normal direction coordinates of a point. The Z-axis of the preset scan is roughly aligned with the normal direction of the cable surface, and the point coordinates are extracted. In Quantity.

[0080] For local curvature changes, the difference in the degree of local surface curvature between the current point and the reference point is quantified. The calculation method for local curvature changes is consistent with the curvature estimation logic of spatial anchor point extraction in step S2, based on the minimum eigenvalue of the covariance matrix. For real-time points and benchmark Take respectively The nearest neighbors constitute the neighborhood point set. and For each neighborhood point set, calculate its centroid. (Using the calculation formula in step s2), then construct the covariance matrix. ,right Eigenvalue decomposition yields three eigenvalues. Take the smallest eigenvalue Characterizes curvature; The formula for calculating curvature change is:

[0081] Local curvature change at a point for The smallest eigenvalue, for The smallest eigenvalue; For the deviation of the normal vector angle, first reuse the neighborhood point set calculated by curvature. and covariance matrix After decomposition, the smallest eigenvalue The corresponding eigenvector is the direction of the surface normal vector. If the eigenvector points to the outside / inside of the cable and is opposite to the reference normal vector, the correction direction is determined by the dot product to ensure that it is consistent with the orientation of the reference normal vector; the normal vector is divided by its magnitude to obtain the unit normal vector.

[0082] in, The unit normal vector in real-time micro-topography data. The unit normal vector in the baseline micro-topography model. for of Corresponding feature vectors; for of For the corresponding eigenvectors, ||…|| are norm symbols, which are used here to calculate the magnitude of the vector; Then, calculate the angle between the surface normal vectors at the two points:

[0083] in, This represents the angle deviation between the normal vectors at the i-th point. The inverse cosine function calculates the corresponding angle based on the cosine value. Texture breaks or misalignments can cause significant changes in the direction of the normal vector.

[0084] For the local plane fitting residuals, the points are calculated using the least squares method. and local neighborhood point set and The best-fit plane, let the plane equation be... (Unit normal vector) (consistent with the normal vector mentioned above), for each neighborhood point Calculate the centroid of the neighborhood points The plane equation is simplified to:

[0085] This equation is equivalent to the covariance matrix. The smallest eigenvalue corresponds to the eigenvector (Consistent with the normal vector calculation, reusing the result); then... and center of mass Substituting into the plane equation, we obtain the complete fitted plane, and through this process, we obtain the optimal fitted plane. and ; Then, calculate the difference in average distance (root mean square error) from the point cloud to its fitted plane:

[0086] in, This represents the change in the local plane fitting residual in the neighborhood of the i-th point; Point The set of neighborhood points; Point The set of neighborhood points; Represented by neighborhood point set The best-fit plane; Represented by neighborhood point set The best-fit plane; is the distance from the point to the plane; k is the number of points in the neighborhood. Cracks disrupt the smoothness of the surface, leading to an increase in the fitting residual.

[0087] Since each feature targets a specific geometric distortion that a crack may cause, and Z-score normalization is used to transform the multi-dimensional crack features into dimensionless values ​​in terms of standard deviation, a fair benchmark is established for anomaly detection based on a unified statistical threshold. Together, these features constitute the crack feature vector for identifying candidate cracks. ,in, The average value of the high residuals in the healthy sample serves as the benchmark. The standard deviation of high residuals in healthy samples; The average value of the curvature change in healthy samples serves as a benchmark. The standard deviation of the curvature change in healthy samples is used as a measure. Typical average of the angle between the normal vectors of healthy samples; The natural fluctuation range of the angle between the normal vectors of healthy samples; The standard deviation scale is used to measure the local plane fitting residuals of healthy samples. is the standard deviation scale of the local plane fitting residuals of healthy samples, which are sample points in a healthy state randomly extracted from the baseline model.

[0088] Next, based on the crack feature vector To identify outliers, a threshold is set for each feature dimension. If any feature value of a point exceeds its corresponding threshold, then that point is marked as a preliminary outlier.

[0089] in, This is a preliminary outlier detection function. =1 indicates that the i-th point is marked as an initial outlier. =0 indicates that the i-th point is marked as a normal point; It is the high residual threshold; It is the curvature change threshold; It is the normal vector deviation threshold; It is the roughness variation threshold, if the local plane fitting residual of the point If the value exceeds this threshold, it is considered abnormal. All four thresholds are preset constants, derived by analyzing historical data. To avoid misclassifying isolated noise points or minute attachments as cracks, a spatial continuity constraint is introduced. Only regions with a continuous distribution of initial anomalies are identified as candidate crack regions. ; It should be noted that the determination of continuous anomaly regions needs to be adapted to the cylindrical structure of the cable, using circumferential-vertical mesh generation and connected component analysis. The specific steps are as follows: global coordinate system Convert to cable cylindrical coordinate system ( For circumferential angle, The vertical length (where the radial radius is used) to avoid distortion of the cylindrical surface by the planar mesh; Among them, Zhou Xiang Divide the grid (consistent with the S1 scan trajectory) to cover ; Vertically Grid the area (with the same spacing as the S1 axis) to cover the current section. The length (e.g., 1m); each grid cell is defined as ( For circumferential index, (For vertical indexing).

[0090] Count the initial number of outliers within each grid cell. If the number... Individual (or percentage) The grid is marked as an anomalous grid; two anomalous grids are considered to be "circumferentially adjacent". "or vertically adjacent" If a certain value is found, it is considered connected; all connected anomalous meshes are merged to form a connected anomalous region. Retain an area of ​​≥5 grid cells (corresponding to the circumferential direction) Vertical The connected anomaly regions are the final candidate crack regions. (Eliminate isolated abnormal regions that are too small to reduce noise interference).

[0091] Finally, this step outputs a set of candidate crack regions. , To select the number of cracked regions, where each region It includes its spatial location, the set of points it covers, and preliminary feature statistics (such as average depression depth), providing input for the authenticity determination in step S4.

[0092] Step S4: Construct a periodic stability determination mechanism to determine the cracks in the candidate crack region and screen out the real cracks.

[0093] In step S3, a set of candidate crack regions is obtained. The core task of this step is to introduce a periodic stability determination mechanism in the time dimension to strictly distinguish between real cracks and occasional surface disturbances. The key is to verify whether the candidate region persists in the time series, which is the fundamental basis for judging whether it is structural damage.

[0094] Specifically, for each candidate region Create a time series archive Record its current and historical significance The state of existence in each inspection cycle, and a binary variable is defined. Indicates the region In the cycle Whether it was detected (candidate crack region determination in step s3):

[0095] This is an existence indicator function, a binary variable used to label candidate regions. In a specific Whether it was detected during each inspection cycle; It means the first One candidate crack region; This indicates the time cycle number of the inspection. Index of the candidate crack region order; Next, calculate the region in Total number of cycles within a given period:

[0096] in, Indicates candidate region exist The total number of cycles within a given period; This indicates the total number of inspection cycles used for judgment. Greater than or equal to a preset persistence threshold (For example ),Right now Then the candidate crack region is determined. The cracks are identified as real cracks and added to the set of real crack regions. Otherwise, the region is determined to be an occasional disturbance (such as temporary dirt or water droplets) and is removed from the candidate set.

[0097] This step outputs a set of real crack regions verified over time. and output its feature vector. At this point, for each real crack region, only its actual structural damage is confirmed, while its evolution trend will be determined in step S5 using a more refined prediction model.

[0098] Step S5: Based on a long short-term memory network, construct a crack trend prediction model, obtain the characteristic time-series data of the real crack, input the characteristic time-series data of the real crack into the crack trend prediction model, output the crack trend prediction result, and realize online identification of insulation layer cracks based on the crack trend prediction result.

[0099] In step S4, the actual set of cracks is confirmed. Next, the core objective of this step is to analyze each actual crack. The future expansion trend of the insulation layer is quantitatively predicted, and a comprehensive risk classification and early warning are performed based on its geometric characteristics and predicted trend to achieve online identification of insulation layer cracks and construct a crack trend prediction model. The process is as follows: The crack trend prediction model is a long short-term memory network based on an attention mechanism, for each real crack... From its time series archives Get the past Feature time-series data of a continuous inspection cycle (height residual dimension) = Curvature variation dimension Dimension of the angle between the normal vectors Roughness variation dimensions: ) as model input.

[0100] An attention mechanism is introduced at the top layer of the standard LSTM encoder, enabling the model to dynamically focus on historical key points most relevant to future predictions. This mechanism is calculated as follows: First, calculate the attention score for each time step of the LSTM encoder output. Hidden state Calculate its correlation score :

[0101] in, It is a learnable parameter matrix whose function is to perform a linear transformation on the concatenated long vector, mapping it to a new vector space; It is a learnable parameter row vector that acts on the result after tanh activation, mapping it from a vector to a scalar; It is the final state after the encoder has processed the entire input sequence; it encodes summary information of the entire sequence. ( This is a nonlinear activation function that performs a nonlinear mapping on the result of a linear transformation; T is the transpose sign.

[0102] Then, a weighted context vector is generated, and the scores are normalized into attention weights using the Softmax function. The hidden states are then weighted and summed to obtain the context vector:

[0103] Where c is the context vector. and Both represent the index of the time step, with values ​​ranging from 1 to the total sequence length L; It is an exponential function; The final predicted output is obtained by using the context vector. The hidden state of the last time step splicing, outputting the future through a fully connected layer. Predicted value of the cycle (e.g., predicted depth):

[0104] in, It is a learnable parameter matrix that performs a linear transformation on the concatenated long feature vector, mapping it to the dimension of the prediction target; This is the output bias vector; To predict the step size; The crack trend prediction model needs to be trained using historical data, employing a large amount of time-series data on historical cracks. As a training set.

[0105] Because cable monitoring data inevitably contains some accidental and unreliable measurements (outliers), using standard mean squared error would result in significant losses from these outliers, thus "biasing" the model's training direction. The Huber loss function, however, mitigates the impact of these large errors, guiding the model to focus more on learning the general patterns of crack evolution rather than fitting accidental noise. This ultimately leads to a more stable and reliable prediction model, with the training objective being to minimize the predicted values. Compared with the true value Differences between them:

[0106] in, Let be the loss function of the crack trend prediction model. These are the predicted values ​​from the crack trend prediction model. It is the true value, representing the accurate value that the model is expected to predict; It is a preset threshold; The absolute value of the prediction error represents the degree to which the predicted value deviates from the true value.

[0107] For each real crack The feature sequence is input into the trained model to obtain its predicted output. .

[0108] Subsequently, a comprehensive risk classification is performed, combining the instantaneous geometric characteristics of the crack with its dynamic propagation trend to assess the risk. The formula for calculating the risk index RI is as follows:

[0109] in, This indicates the rate at which the prediction is normalized and expanded. Step S4 represents the actual crack. The normalized multidimensional geometric feature vector of the current state; Let be the norm of this vector, used to quantify the current overall severity of the crack. These are weighting coefficients used to balance the impact of expansion speed and current severity. They are preset parameters determined after training and optimization based on historical crack data.

[0110] This formula combines the dynamic evolution trend of the crack with its current static severity.

[0111] Based on the calculated risk index They are classified into different risk levels as follows: High risk Medium risk Low risk ,in, and It is the threshold for risk level.

[0112] Ultimately, a structured early warning message is generated for each crack. This message not only includes the location and risk level of the crack area, but more importantly, it includes trend judgment based on the prediction model, thereby realizing online identification of insulation layer cracks.

[0113] This invention provides an online method for identifying insulation layer cracks in overhead insulated cables. Its core lies in completely abandoning the traditional detection approach that relies on indirect and easily disturbed physical quantities such as light, temperature, or electrical signals. Instead, it directly and accurately monitors the microscopic changes in the surface geometry of the insulation material. By constructing a progressively sophisticated closed-loop technology, a high-precision three-dimensional surface benchmark model is first established during the cable's healthy phase as a comparison basis. Then, a spatial anchor point is innovatively introduced to improve the registration algorithm, solving the data misalignment problem caused by cable sway and ensuring cross-period data comparability. Based on this, by extracting multi-dimensional geometric features such as curvature abrupt changes and normal vector deflection and applying spatial continuity constraints, high-sensitivity, low-false-alarm identification of micron-level early cracks is achieved. Furthermore, a time dimension is introduced for multi-period stability analysis, effectively distinguishing between real expanding cracks and occasional interference. Finally, an attention-based LSTM network is used to predict the crack evolution trend and combined with multi-factor dynamic risk classification, outputting decision-making information that can directly guide operation and maintenance. This achieves a leap from passive detection to proactive predictive maintenance, forming a complete solution with high precision, high reliability, and strong foresight.

[0114] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the statement "including a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0115] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their likenesses.

Claims

1. A method for online identification of insulation layer cracks in overhead insulated cables, characterized in that, Includes the following steps: Step S1: Obtain the initial insulation layer contour data of the overhead insulated cable and establish a reference micro-morphology model of the insulation layer; Step S2: Obtain real-time insulation layer contour data. Based on the reference micro-topography model, use the iterative nearest point algorithm to spatially align the real-time insulation layer contour data to obtain aligned insulation layer contour data. Step S3: Extract features from the aligned insulation layer contour data to obtain crack feature vectors, and compare the crack feature vectors with a preset threshold to determine candidate crack regions; Step S4: Construct a periodic stability determination mechanism, and use the periodic stability determination mechanism to determine the cracks in the candidate crack region and filter out the real cracks. Step S5: Based on a long short-term memory network, construct a crack trend prediction model, obtain the characteristic time-series data of the real crack, input the characteristic time-series data of the real crack into the crack trend prediction model, output the crack trend prediction result, and realize online identification of insulation layer cracks based on the crack trend prediction result.

2. The method for online identification of insulation layer cracks in overhead insulated cables according to claim 1, characterized in that, Obtain initial insulation profile data for overhead insulated cables, including: Select several cross-sectional sections along the length of the target overhead insulated cable. Where k is the index of the cross-sectional segment order, and K is the total number of cross-sectional segments; in each segment The inner intervals are arranged at fixed circumferential angles. and axial spacing Plan out dense scanning trajectories ,in For trajectory indexing Used to plan the scanning trajectory around the cable surface. Used to plan the scanning trajectory along the length of the cable; Then, using the micro-surface profile acquisition components deployed on the inspection device, the laser line scanner strictly follows the planned trajectory. The outer surface of the insulation layer is scanned. The initial contour data of the insulation layer obtained in this process is a discrete three-dimensional point cloud, which can be represented as: ； in, Representative at the Section, No. The collection of raw point clouds collected along the trajectory. These are the coordinates of a single 3D point in a point cloud; It is a single data point in the set, representing the coordinates of a specific location on the surface of the insulating layer measured in three-dimensional space; It is a point coordinates, These represent the three coordinate components of the point in the three-dimensional coordinate system; the superscript T indicates transpose. This symbol represents a point. It belongs to the three-dimensional real number space; It is an index variable used to iterate through each point in the set; This indicates that in this specific scan trajectory The total number of points collected.

3. The method for online identification of insulation layer cracks in overhead insulated cables according to claim 2, characterized in that, Establish a baseline micromorphological model of the insulating layer, including: For two adjacent point clouds and An optimal rigid body transformation is found by using the iterative nearest point algorithm. ,in It is a 3x3 rotation matrix. It is a 3x1 translation vector. The goal of this transformation is to minimize the Euclidean distance between corresponding points. Its mathematical expression is: ； in, and These are the optimal rotation matrix and optimal translation vector, respectively, obtained through iterative optimization algorithms. It is the source cloud The first in A three-dimensional point, It is a target point cloud Midpoint The corresponding nearest neighbor will be determined by the algorithm. Find the nearest point in another point cloud as... ; For point pairs participating in the calculation The total quantity; The optimal rotation matrix is ​​obtained by iteratively solving this optimization problem. and optimal translation vector This transforms the source point cloud into the coordinate system of the target point cloud. ； in, Points in the source point cloud After optimal transformation and The new coordinates obtained later; Finally, by sequentially executing the above registration process, all local point clouds are... Unified transformation to global coordinate system Below, the registered point cloud is obtained. All By integrating the data, a baseline micromorphological model of the cable under healthy conditions was established. .

4. The method for online identification of insulation layer cracks in an overhead insulated cable according to claim 3, characterized in that, Acquire real-time insulation layer contour data, including: During real-time inspection, the laser line scanner strictly follows the planned trajectory, utilizing the micro-surface profile acquisition components deployed on the inspection device. The outer surface of the insulation layer is scanned to obtain real-time contour data of the cable insulation layer. .

5. The method for online identification of insulation layer cracks in an overhead insulated cable according to claim 4, characterized in that, The iterative nearest-point algorithm is used to spatially align the real-time insulation layer contour data, resulting in aligned insulation layer contour data, including: An improved Iterative Closest Point (ICP) algorithm is used for fine registration. The objective function of this algorithm adds a spatial anchor point constraint term to the standard point-pair distance term to improve the stability and accuracy of the registration. ； in, It is the fine transformation matrix to be solved; It refers to points in a real-time point cloud. yes The corresponding nearest neighbor; This indicates the number of points in the real-time point cloud that participated in the registration. This is the total number of anchor point pairs; This is a standard ICP data item used to minimize the distance of the overall point cloud. It is a weighting factor used to adjust each point pair The importance of this, based on the points of reference The matching quality is determined; It is an anchor point constraint term. and These are matched spatial anchor point pairs. This ensures that these significant feature point pairs are always pulled towards alignment during the optimization process, enhancing the robustness of the algorithm. and These are hyperparameters used to balance the weights of data items and anchor constraint items, and are adjusted experimentally. It is a spatial anchor index; Let be the transformation function, representing the substitution of the source points into the optimal transformation matrix; by iteratively optimizing this objective function, the optimal high-precision transformation matrix is ​​finally obtained. ; Finally, the optimal transformation is applied to correct the real-time insulation layer profile data: ； in, It is the same as the baseline micro-morphology model Precisely aligned insulation layer contour data. It is real-time insulation layer contour data.

6. The method for online identification of insulation layer cracks in an overhead insulated cable according to claim 5, characterized in that, Feature extraction is performed on the aligned insulation layer contour data to obtain a crack feature vector, including: for Each sampling point in and in Corresponding points in Calculate the following geometric features, including height residual, local curvature variation, normal vector angle deviation, and local plane fitting residual: For the height residual, calculate the absolute height deviation between the current point and the reference point: ； in, Indicates the first The height residual of each point; Indicates real-time micromorphological data The first in A three-dimensional point; Indicating in the baseline micromorphology model Zhongyu The exact corresponding number A three-dimensional point; It is a function that retrieves the coordinates of the normal direction of a point; The formula for calculating curvature change is: ； in, Indicates the first Local curvature change at a point for The smallest eigenvalue, for The smallest eigenvalue; For the deviation of the normal vector angle, first reuse the neighborhood point set calculated by curvature. and covariance matrix After decomposition, the smallest eigenvalue The corresponding eigenvector is the direction of the surface normal vector. If the eigenvector points to the outside / inside of the cable and is opposite to the reference normal vector, the correction direction is determined by the dot product to ensure that it is consistent with the orientation of the reference normal vector; the normal vector is divided by its magnitude to obtain the unit normal vector. ； in, The unit normal vector in real-time micro-topography data. The unit normal vector in the baseline micro-topography model. for of Corresponding feature vectors; for of For the corresponding eigenvectors, ||…|| are norm symbols, which are used here to calculate the magnitude of the vector; Then, calculate the angle between the surface normal vectors at the two points: ； in, This represents the angle deviation between the normal vectors at the i-th point. It is an inverse cosine function, and the corresponding angle is calculated based on the cosine value. Texture breakage or misalignment will cause a significant change in the direction of the normal vector. For the local plane fitting residuals, the points are calculated using the least squares method. and local neighborhood point set and Find the best-fitting plane, and then calculate the difference in average distances from the point cloud to its fitted plane: ； in, This represents the change in the local plane fitting residual in the neighborhood of the i-th point; Point The set of neighborhood points; Point The set of neighborhood points; Represented by neighborhood point set The best-fit plane; Represented by neighborhood point set The best-fit plane; is the distance from the point to the plane; k is the number of points in the neighborhood.

7. The method for online identification of insulation layer cracks in an overhead insulated cable according to claim 6, characterized in that, The crack feature vector is compared with a preset threshold to determine candidate crack regions, including: Based on crack feature vector To identify outliers, a threshold is set for each feature dimension. If any feature value of a point exceeds its corresponding threshold, then that point is marked as a preliminary outlier. ； in, This is a preliminary outlier detection function. =1 indicates that the i-th point is marked as an initial outlier. =0 indicates that the i-th point is marked as a normal point; It is the high residual threshold; It is the curvature change threshold; It is the normal vector deviation threshold; It is the roughness variation threshold, if the local plane fitting residual of the point If the value exceeds this threshold, it is considered abnormal. All four thresholds are preset constants derived by analyzing historical data.

8. The method for online identification of insulation layer cracks in an overhead insulated cable according to claim 7, characterized in that, A periodic stability determination mechanism is constructed to determine cracks in candidate crack regions and filter out real cracks, including: For each candidate region Create a time series archive Record its current and historical significance The state of existence in each inspection cycle, and a binary variable is defined. Indicates the region In the cycle Was it detected? ； This is an existence indicator function, a binary variable used to label candidate regions. In a specific Whether it was detected during each inspection cycle; It means the first One candidate crack region; This indicates the time cycle number of the inspection. Index of the candidate crack region order; Next, calculate the region in Total number of cycles within a given period: ； in, Indicates candidate region exist The total number of cycles within a given period; This indicates the total number of inspection cycles used for judgment. Greater than or equal to a preset persistence threshold ,Right now Then the candidate crack region is determined. The cracks are identified as real cracks and added to the set of real crack regions. Otherwise, the region is determined to be an occasional disturbance and is removed from the candidate set.

9. The method for online identification of insulation layer cracks in an overhead insulated cable according to claim 8, characterized in that, A crack trend prediction model is constructed based on long short-term memory networks, including: The crack trend prediction model is a long short-term memory network based on an attention mechanism, for each real crack... From its time series archives Get the past The characteristic time-series data of each continuous inspection cycle are used as the model input; An attention mechanism is introduced at the top layer of the standard LSTM encoder, enabling the model to dynamically focus on historical key points most relevant to future predictions. This mechanism is calculated as follows: First, calculate the attention score for each time step of the LSTM encoder output. Hidden state Calculate its correlation score : ； in, It is a learnable parameter matrix whose function is to perform a linear transformation on the concatenated long vector, mapping it to a new vector space; It is a learnable parameter row vector that acts on the result after tanh activation, mapping it from a vector to a scalar; It is the final state after the encoder has processed the entire input sequence; it encodes summary information of the entire sequence. ( This is a nonlinear activation function that performs a nonlinear mapping on the result of a linear transformation; T is the transpose sign. Then, a weighted context vector is generated, and the scores are normalized into attention weights using the Softmax function. The hidden states are then weighted and summed to obtain the context vector: ； Where c is the context vector. and Both represent the index of the time step, with values ​​ranging from 1 to the total sequence length L; It is an exponential function; The final predicted output is obtained by using the context vector. The hidden state of the last time step splicing, outputting the future through a fully connected layer. Predicted value of the cycle : ； in, It is a learnable parameter matrix that performs a linear transformation on the concatenated long feature vector, mapping it to the dimension of the prediction target; This is the output bias vector; To predict the step size.

10. The method for online identification of insulation layer cracks in an overhead insulated cable according to claim 9, characterized in that, Based on the crack trend prediction results, online identification of insulation layer cracks is achieved, including: For each real crack The feature sequence is input into the trained model to obtain its predicted output. ; Subsequently, a comprehensive risk classification is performed, combining the instantaneous geometric characteristics of the crack with its dynamic propagation trend to assess the risk. The formula for calculating the risk index RI is as follows: ； in, This indicates the rate at which the prediction is normalized and expanded. Step S4 represents the actual crack. The normalized multidimensional geometric feature vector of the current state; Let be the norm of this vector, used to quantify the current overall severity of the crack. These are weighting coefficients used to balance the impact of expansion speed and current severity. They are preset parameters determined after training and optimization based on historical crack data. Based on the calculated risk index They are classified into different risk levels as follows: High risk Medium risk Low risk ,in, and It is the threshold for the risk level; Ultimately, a structured early warning message is generated for each crack. This message not only includes the location and risk level of the crack area, but more importantly, it includes trend judgment based on the prediction model, thereby realizing online identification of insulation layer cracks.

Images