An overhead power line icing thickness detection method and system based on a structure perception framework

Abstract

The application discloses an overhead power line ice thickness detection method and system based on a structure perception framework, first acquires power transmission line point cloud data, and inputs the point cloud data into a structure perception semantic segmentation network after pretreatment; the network adopts a four-level encoder-decoder architecture, the encoder extracts multi-scale features through local spatial coding, cross-scale spatial attention fusion and local aggregation operation, the decoder realizes feature fusion in combination with a skip mechanism, and outputs a conductor semantic segmentation result; based on the segmentation result, conductor point sets are extracted and Euclidean clustering is performed to divide the conductor into strands; double-plane projection and polynomial fitting are performed on each strand of conductor to construct a three-dimensional center line; finally, point-to-center line distance is calculated, peripheral point sets are selected to calculate conductor envelope radius, and ice thickness is obtained in combination with the radius of a bare conductor. The application effectively solves the measurement deviation caused by point cloud occlusion, curved conductor and asymmetric ice, realizes millimeter-level precision non-contact ice thickness detection in a complex environment, and provides reliable technical support for power grid disaster prevention.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and in particular relates to a method and system for detecting the icing thickness of overhead power lines based on a structure-aware framework. Background Technology

[0002] Currently, ice thickness detection mainly relies on image processing or physical sensing methods. One common approach is image-based geometric contour extraction, which uses drones to capture images of power transmission lines and employs image enhancement, edge detection, and contour fitting techniques to estimate the outer radius of the conductor under icing conditions. This radius is then compared with the radius of a known bare conductor to infer the ice thickness. However, this method is highly dependent on the external environment; images are easily affected by factors such as occlusion, uneven lighting, and changes in viewing angle, leading to decreased accuracy. Furthermore, it cannot obtain three-dimensional structural information, limiting its practical application in complex environments. Another existing technology is a physical model-based sensing method, such as deploying tension sensors or inclinometers on the conductor. This involves combining meteorological data to calculate changes in conductor stress and deformation, and then establishing a conductor-ice physical response model for retrieving ice thickness. While this method offers high accuracy, it requires sophisticated sensor deployment and calibration, resulting in high equipment costs, complex maintenance, and difficulty in large-scale deployment. It also lacks non-contact detection capabilities.

[0003] In recent years, with the development of UAV-LiDAR technology, some studies have proposed using 3D point cloud modeling methods to achieve non-contact ice thickness estimation. This typically employs a two-stage strategy: first, semantic segmentation of the point cloud is performed to extract the guide wire region; then, centerline fitting and cross-sectional analysis are conducted to calculate the ice thickness based on radial expansion. However, in real-world complex corridor scenarios, guide wire point clouds are characterized by their elongated shape, sparseness, and significant occlusion. Existing segmentation methods struggle to accurately preserve the guide wire's geometric structure, often resulting in breaks and blurred boundaries. This leads to large deviations in centerline fitting, unstable extraction of outer shell points, and discontinuous and erroneous thickness estimation results. Therefore, existing technologies still have significant shortcomings in terms of point cloud structure preservation, boundary clarity, and estimation accuracy, necessitating a new ice thickness detection method with strong structural continuity, high boundary discrimination capability, and robustness. This invention addresses these issues by proposing an ice thickness estimation method that integrates structure-aware semantic segmentation and geometric modeling analysis, effectively solving the degradation of structure and estimation accuracy caused by guide wire sparsity and occlusion. Summary of the Invention

[0004] To address the above technical problems, this invention provides a method and system for detecting the icing thickness of overhead power lines based on a structure-sensing framework.

[0005] The technical solution adopted by this invention to solve its technical problem is:

[0006] A method for detecting icing thickness on overhead power lines based on a structure-sensing framework, the method comprising the following steps:

[0007] S100: The UAV's lidar collects point cloud data of overhead power lines, preprocesses the point cloud data, and uses a multilayer perceptron to perform feature upscaling on the preprocessed point cloud data.

[0008] S200: The upscaled point cloud data is input into the structure-aware semantic segmentation network, and the semantic features of the wire point cloud are extracted through the encoder-decoder architecture. The encoder contains a four-layer structure, and each layer performs local spatial encoding, cross-scale spatial attention fusion and local aggregation operations in sequence to generate multi-scale fusion features. The decoder contains a symmetrical four-layer structure, which upsamples through nearest neighbor interpolation and fuses with the encoder features. Finally, the wire semantic segmentation result is output through a shared fully connected classification head.

[0009] S300: Extract the conductor point set based on the semantic segmentation results, separate the multiple conductors through Euclidean clustering, and fit the three-dimensional centerline trajectory of each conductor.

[0010] S400: For each conductor strand, calculate the shortest Euclidean distance from each conductor point to the fitted centerline. Based on the statistical distribution of the distances from the conductor points to the centerline, extract the shell point set and calculate the conductor envelope radius. Combine this with the bare conductor radius to estimate the icing thickness.

[0011] Preferably, each encoder in S200 includes a local spatial coding module, a cross-level spatial coding module, and a local aggregation module;

[0012] After downsampling the features of the input encoder, the local spatial coding module aggregates the neighborhood features through grouping operations, and extracts the local geometric structure through max pooling and MLP.

[0013] The cross-level spatial coding module concatenates the output of the current layer's local spatial coding module with the features of the previous layer, constructs a query vector Q, a key vector K, and a value vector V through the MLP-ReLU module, and calculates attention weights through matrix multiplication and then weighted and fused them.

[0014] The local aggregation module integrates the output features of the local spatial coding module and the output features of the cross-level spatial coding module.

[0015] Preferably, the local spatial coding module includes a dual-branch processing. In the first branch, the input features are aggregated into neighborhood points through a grouping operation, and local features are extracted through max pooling and MLP. In the second branch, the input features are directly transformed by MLP, and the outputs of the two branches are concatenated and then generated by MLP to produce local enhanced features.

[0016] Preferably, the cross-level spatial coding module concatenates the output of the current layer's local spatial coding module with the features from the previous layer, specifically as follows:

[0017] ;

[0018] in, For features of the next higher level, For the current layer features, T(·) represents the projection function. Features after splicing;

[0019] Cross-scale spatial correlation features are generated through the attention weight calculation module, specifically:

[0020] ;

[0021] in, , , Both consist of two MLP layers, used to map input features to query, key, and value representations in the attention space. This represents the attention weight matrix, used to characterize the strength of the relationships between points, and is used for weighted aggregation of features. Indicates the dimension of the key vector;

[0022] The weighted fused features are output to the next encoding layer, specifically as follows:

[0023] ;

[0024] in, These are adjustable fusion weights used to balance the importance of local features and global attention-enhanced features. This indicates the output of the cross-level spatial encoding module.

[0025] Preferably, S300 includes:

[0026] S310: Extract the conductor category point set based on the semantic segmentation results output by S200, and perform isolated point removal and connectivity filtering on the conductor category point set to remove local noise and misclassified points;

[0027] S320: Based on the adjacency relationship constructed by the Euclidean distance between points, the region growing clustering method is used to divide the point set of the conductor category into several spatially connected subsets, each subset corresponding to different strands of conductor; in the clustering process, a distance threshold is set to ensure the spatial continuity of the same strand of conductor, and a minimum cluster size is set to filter fragmented clustering results;

[0028] S330: For each set of conductor points, a three-dimensional centerline trajectory is constructed using a local polynomial fitting method based on its spatial distribution characteristics.

[0029] Preferably, S330 includes:

[0030] S331: Project the guide point to , On a plane;

[0031] S332: Perform nth-order polynomial fitting on the coordinate points in each projection plane to construct the centerline trajectory model:

[0032] ;

[0033] in, These are coordinate components in three-dimensional space; and They are respectively plane and The polynomial coefficients of the plane-fitted curve represent the th... The second term and the first The weight of the next term; and The order of the polynomial used controls the complexity of the fitted curve. , This is the index of the number of terms in a polynomial expansion;

[0034] S333: Using the least squares method to fit the coefficients in the centerline fitting model and Solve to minimize the vertical distance residual of the traverse point cloud relative to the fitted curve in the corresponding projection plane; when... Taking a plane as an example, let the set of points of the projected traverse be... Then the fitting coefficient This was obtained by solving the following optimization problem:

[0035] ;

[0036] Similarly, in Construct a fitting model in the plane, and let the projection point be... Then the coefficient satisfy:

[0037] ;

[0038] Both of the above minimization problems are linear least squares fitting problems. By constructing a system of normal equations or a pseudo-inverse matrix and solving it, the final fitting coefficients are used to reconstruct the centerline trajectory equation of the conductor in three-dimensional space. Specifically:

[0039] ;

[0040] in, For the curve parameters on the center line, take a one-dimensional variable. .

[0041] Preferably, S400 includes:

[0042] S410: For each conductor strand Calculate each traverse point The shortest Euclidean distance to the fitted centerline:

[0043] ;

[0044] in, Indicates the first From the first guide point to the second Minimum radial distance from the centerline of a conductor strand; For the first The three-dimensional spatial coordinates of each guide point; Indicates the first The centerline trajectory function of a conductor. The curve parameters are those along the center line; Represents the Euclidean norm in two-dimensional or three-dimensional space;

[0045] S420: Based on the statistical feature extraction rules, select the outermost points with the lowest ranking among all radius values ​​to form the outer shell point set. ,Right now:

[0046] ;

[0047] Among them, percentile function Indicates all After sorting the values ​​from smallest to largest, points ranked between 85% and 100% are selected as peripheral candidate points, while internal points in dense areas or with noise interference are excluded.

[0048] S430: Shell point set Based on this, an estimated value for the envelope radius of the conductor is constructed. Specifically:

[0049] ;

[0050] S440: The bare conductor radius corresponding to the actual conductor model. Calculate the final estimated ice thickness. Specifically:

[0051] ;

[0052] in, To obtain the statistically significant radius of the conductor's outer shell envelope containing ice, The radius of the conductor body can be obtained from a database or through on-site measurement. The radial icing thickness is calculated in millimeters.

[0053] An overhead power line icing thickness detection system based on a structure-sensing framework includes:

[0054] The point cloud data acquisition and preprocessing module is used to collect point cloud data of overhead power lines by UAV lidar, preprocess the point cloud data, and perform feature upscaling on the preprocessed point cloud data through a multilayer perceptron.

[0055] The structure-aware semantic segmentation module is used to input the upscaled point cloud data into the structure-aware semantic segmentation network and extract the semantic features of the wire point cloud through an encoder-decoder architecture. The encoder consists of four layers, each of which performs local spatial encoding, cross-scale spatial attention fusion and local aggregation operations in sequence to generate multi-scale fusion features. The decoder symmetrically consists of four layers, which upsample through nearest neighbor interpolation and fuse with the encoder features. Finally, the wire semantic segmentation results are output through a shared fully connected classification head.

[0056] The 3D centerline trajectory fitting module is used to extract the conductor point set based on the semantic segmentation results, separate multiple conductors through Euclidean clustering, and fit the 3D centerline trajectory of each conductor.

[0057] The icing thickness estimation module is used to calculate the shortest Euclidean distance from each conductor point to the fitted centerline for each conductor strand. Based on the statistical distribution of the distances from the conductor points to the centerline, it extracts the shell point set and calculates the conductor envelope radius, and estimates the icing thickness by combining the bare conductor radius.

[0058] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of a method for detecting the icing thickness of overhead power lines based on a structure-aware framework.

[0059] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for detecting the icing thickness of overhead power lines based on a structure-aware framework.

[0060] The aforementioned method and system for detecting ice thickness on overhead power lines based on a structure-aware framework significantly improves the detection accuracy and robustness of ice thickness on overhead power lines in complex environments through the collaborative innovation of a structure-aware semantic segmentation network and geometric modeling. The structure-aware segmentation network effectively enhances the semantic continuity of the conductor point cloud in occluded and fractured scenarios through a cross-scale spatial attention fusion mechanism, providing a highly complete conductor point set for subsequent geometric modeling. The 3D centerline segmentation fitting method, combined with dual-plane projection and local multinomial regression, overcomes the distortion problem of curved conductor trajectories, reducing the centerline reconstruction error to the millimeter level. The outer shell point statistical ice thickness estimation strategy significantly suppresses measurement deviations caused by point cloud sparsity and asymmetric icing by extracting some peripheral points and calculating the squared average radius. The entire method achieves non-contact, fully automated detection, supporting millimeter-level precision ice thickness analysis for transmission lines, providing highly reliable technical support for power grid anti-icing and disaster prevention. Attached Figure Description

[0061] Figure 1 This is a flowchart of a method for detecting the icing thickness of overhead power lines based on a structure-sensing framework, according to an embodiment of the present invention.

[0062] Figure 2 This is a schematic diagram of the structure of a structure-aware semantic segmentation network in one embodiment of the present invention. Detailed Implementation

[0063] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0064] In one embodiment, such as Figure 1 and 2 As shown, a method for detecting the icing thickness of overhead power lines based on a structure-sensing framework includes the following steps:

[0065] S100: The UAV's lidar collects point cloud data of overhead power lines, preprocesses the point cloud data, and uses a multilayer perceptron to perform feature upscaling on the preprocessed point cloud data.

[0066] S200: The upscaled point cloud data is input into the structure-aware semantic segmentation network, and the semantic features of the wire point cloud are extracted through the encoder-decoder architecture. The encoder contains a four-layer structure, and each layer performs local spatial encoding, cross-scale spatial attention fusion and local aggregation operations in sequence to generate multi-scale fusion features. The decoder contains a symmetrical four-layer structure, which upsamples through nearest neighbor interpolation and fuses with the encoder features. Finally, the wire semantic segmentation result is output through a shared fully connected classification head.

[0067] S300: Extract the conductor point set based on the semantic segmentation results, separate the multiple conductors through Euclidean clustering, and fit the three-dimensional centerline trajectory of each conductor.

[0068] S400: For each conductor strand, calculate the shortest Euclidean distance from each conductor point to the fitted centerline. Based on the statistical distribution of the distances from the conductor points to the centerline, extract the shell point set and calculate the conductor envelope radius. Combine this with the bare conductor radius to estimate the icing thickness.

[0069] Specifically, this invention proposes a Structure-Aware Framework (SAF) for detecting ice thickness on overhead transmission lines. This framework integrates a two-stage processing flow of semantic segmentation and geometric modeling. It constructs an ice thickness estimation method with strong structure preservation and robustness, taking into account the sparsity, elongation and occlusion sensitivity of the conductor point cloud structure.

[0070] To address the issues of conductor structure breakage and boundary ambiguity in complex environments, a 3D point cloud semantic segmentation network based on an encoder-decoder structure was designed. This network integrates a key improvement module: Cross-level Spatial Encoding (CSE). This module introduces information interaction between multiple resolutions during feature extraction, enhancing the spatial continuity representation of the conductor structure at different scales and avoiding semantic interruptions caused by downsampling. Through this network, a complete conductor point set can be accurately extracted from the original point cloud, providing a reliable foundation for geometric modeling. Based on high-quality conductor point clouds, this invention introduces a radial thickness estimation method based on centerline fitting and shell point statistical analysis. By combining spatial structure continuity modeling with slice-level thickness analysis, this invention effectively suppresses estimation errors caused by local noise and point cloud sparsity, achieving millimeter-level ice thickness detection accuracy.

[0071] Furthermore, in S100, input point cloud data ,in For batch size, For points, To input the number of feature channels, a multilayer perceptron is used. Ascend to Point cloud data The input is fed into a segmentation network employing an encoder-decoder structure. Four layers of encoder and four layers of decoder sequentially extract and fuse multi-scale semantic features. Finally, the input is fed into a classification head to obtain the semantic prediction result for each point, and the output is... .

[0072] In one embodiment, each encoder in S200 includes a local spatial coding module, a cross-level spatial coding module, and a local aggregation module;

[0073] After downsampling the features of the input encoder, the local spatial coding module aggregates the neighborhood features through grouping operations, and extracts the local geometric structure through max pooling and MLP.

[0074] The cross-level spatial coding module concatenates the output of the current layer's local spatial coding module with the features of the previous layer, constructs a query vector Q, a key vector K, and a value vector V through the MLP-ReLU module, and calculates attention weights through matrix multiplication and then weighted and fused them.

[0075] The local aggregation module integrates the output features of the local spatial coding module and the output features of the cross-level spatial coding module.

[0076] Specifically, the point cloud features input to the encoder First, it undergoes a grid downsampling operation to be adjusted to Then it enters the local spatial coding module to obtain feature output. .

[0077] The feature result will be processed along two paths: one path inputs to the cross-level spatial coding module (CSE), and the other path concatenates the CSE module's output with the input of the next layer encoder. The input of the CSE module consists of two sources: one is the local encoded feature of the current stage, and the other is the local feature output from the previous level encoder. The two are concatenated to generate a cross-level structural representation.

[0078] The features output by the CSE module and the downsampling results are then fused again by the local aggregation module to generate the fused features. .

[0079] In one embodiment, the local spatial coding module includes a two-branch process. In the first branch, the input features are aggregated into neighborhood points through a grouping operation, and local features are extracted through max pooling and MLP. In the second branch, the input features are directly transformed by MLP, and the outputs of the two branches are concatenated and then generated by MLP to produce local enhanced features.

[0080] In one embodiment, the cross-level spatial coding module concatenates the output of the current layer's local spatial coding module with the features from the previous layer, specifically:

[0081] ;

[0082] in, For features of the next higher level, For the current layer features, T(·) represents the projection function. Features after splicing;

[0083] Cross-scale spatial correlation features are generated through the attention weight calculation module, specifically:

[0084] ;

[0085] in, , , Both consist of two MLP layers, used to map input features to query, key, and value representations in the attention space. This represents the attention weight matrix, used to characterize the strength of the relationships between points, and is used for weighted aggregation of features. Indicates the dimension of the key vector;

[0086] The weighted fused features are output to the next encoding layer, specifically as follows:

[0087] ;

[0088] in, These are adjustable fusion weights used to balance the importance of local features and global attention-enhanced features. This represents the output of the cross-level spatial coding module. Experimental comparisons under multiple typical power transmission scenarios show that... This setting achieves optimal structural continuity and segmentation accuracy, and has been used as the default hyperparameter during model training.

[0089] In one embodiment, S200 involves upsampling via nearest neighbor interpolation and fusing it with encoder features via a jump connection, specifically as follows:

[0090] The input to the decoder stage is dimensionality reduced by an MLP layer, and the features are restored to point cloud shape using nearest neighbor interpolation. The feature is then fused with the encoder output features obtained from the upper layer jump connection and used as the decoder output.

[0091] Specifically, for the decoder stage, the input features are Dimensionality reduction through MLP layer And nearest neighbor interpolation is used to restore the features to the shape of the point cloud. The encoder output features obtained by the upper-layer jumper are fused together and used as the decoder output.

[0092] In one embodiment, S300 includes:

[0093] S310: Extract the conductor category point set based on the semantic segmentation results output by S200, and perform isolated point removal and connectivity filtering on the conductor category point set to remove local noise and misclassified points;

[0094] S320: Based on the adjacency relationship constructed by the Euclidean distance between points, the region growing clustering method is used to divide the point set of the conductor category into several spatially connected subsets, each subset corresponding to different strands of conductor; in the clustering process, a distance threshold is set to ensure the spatial continuity of the same strand of conductor, and a minimum cluster size is set to filter fragmented clustering results;

[0095] S330: For each set of conductor points, a three-dimensional centerline trajectory is constructed using a local polynomial fitting method based on its spatial distribution characteristics.

[0096] In one embodiment, S330 includes:

[0097] S331: Project the guide point to , On a plane;

[0098] S332: Perform nth-order polynomial fitting on the coordinate points in each projection plane to construct the centerline trajectory model:

[0099] ;

[0100] in, These are coordinate components in three-dimensional space; and They are respectively plane and The polynomial coefficients of the plane-fitted curve represent the th... The second term and the first The weight of the next term; and The order of the polynomial used controls the complexity of the fitted curve. , This is the index of the number of terms in a polynomial expansion;

[0101] S333: Using the least squares method to fit the coefficients in the centerline fitting model and Solve to minimize the vertical distance residual of the traverse point cloud relative to the fitted curve in the corresponding projection plane; when... Taking a plane as an example, let the set of points of the projected traverse be... Then the fitting coefficient This was obtained by solving the following optimization problem:

[0102] ;

[0103] Similarly, in Construct a fitting model in the plane, and let the projection point be... Then the coefficient satisfy:

[0104] ;

[0105] Both of the above minimization problems are linear least squares fitting problems. By constructing a system of normal equations or a pseudo-inverse matrix and solving it, the final fitting coefficients are used to reconstruct the centerline trajectory equation of the conductor in three-dimensional space. Specifically:

[0106] ;

[0107] in, For the curve parameters on the center line, take a one-dimensional variable. .

[0108] Specifically, the core objective of this step is to generate a geometrically smooth and structurally continuous center reference line for modeling the actual structure of the conductor, thereby enabling relative geometric measurements of icing. Compared to the traditional cross-sectional intercept method, this method offers stronger trajectory integrity and robustness, making it particularly suitable for centerline reconstruction in curved conductors and obstructed areas.

[0109] In one embodiment, S400 includes:

[0110] S410: For each conductor strand Calculate each traverse point The shortest Euclidean distance to the fitted centerline:

[0111] ;

[0112] in, Indicates the first From the first guide point to the second Minimum radial distance from the centerline of a conductor strand; For the first The three-dimensional spatial coordinates of each guide point; Indicates the first The centerline trajectory function of a conductor. The curve parameters are those along the center line; Represents the Euclidean norm in two-dimensional or three-dimensional space;

[0113] Furthermore, all The value is used to construct the radial distribution of points to support subsequent shell point extraction and thickness statistical analysis.

[0114] S420: Based on the statistical feature extraction rules, select the outermost points with the lowest ranking among all radius values ​​to form the outer shell point set. ,Right now:

[0115] ;

[0116] Among them, percentile function Indicates all After sorting the values ​​from smallest to largest, points ranked between 85% and 100% are selected as peripheral candidate points, while internal points in dense areas or with noise interference are excluded.

[0117] S430: Shell point set Based on this, an estimated value for the envelope radius of the conductor is constructed. Specifically:

[0118] ;

[0119] S440: The bare conductor radius corresponding to the actual conductor model. Calculate the final estimated ice thickness. Specifically:

[0120] ;

[0121] in, To obtain the statistically significant radius of the conductor's outer shell envelope containing ice, The radius of the conductor body can be obtained from a database or through on-site measurement. The radial icing thickness is calculated in millimeters.

[0122] The above-mentioned method for detecting ice thickness on overhead power lines based on a structure-aware framework constructs a high-precision, non-contact ice thickness detection framework for practical engineering applications through a combined approach of "semantic structure enhancement + shell geometric modeling". This not only improves the structural integrity of conductor extraction but also enhances the robustness of geometric modeling, making it widely applicable to automatic inspection tasks of transmission lines under different voltage levels and terrain conditions.

[0123] In one embodiment, an overhead power line icing thickness detection system based on a structure-sensing framework is also provided, comprising:

[0124] The point cloud data acquisition and preprocessing module is used to collect point cloud data of overhead power lines by UAV lidar, preprocess the point cloud data, and perform feature upscaling on the preprocessed point cloud data through a multilayer perceptron.

[0125] The structure-aware semantic segmentation module is used to input the upscaled point cloud data into the structure-aware semantic segmentation network and extract the semantic features of the wire point cloud through an encoder-decoder architecture. The encoder consists of four layers, each of which performs local spatial encoding, cross-scale spatial attention fusion and local aggregation operations in sequence to generate multi-scale fusion features. The decoder symmetrically consists of four layers, which upsample through nearest neighbor interpolation and fuse with the encoder features. Finally, the wire semantic segmentation results are output through a shared fully connected classification head.

[0126] The 3D centerline trajectory fitting module is used to extract the conductor point set based on the semantic segmentation results, separate multiple conductors through Euclidean clustering, and fit the 3D centerline trajectory of each conductor.

[0127] The icing thickness estimation module is used to calculate the shortest Euclidean distance from each conductor point to the fitted centerline for each conductor strand. Based on the statistical distribution of the distances from the conductor points to the centerline, it extracts the shell point set and calculates the conductor envelope radius, and estimates the icing thickness by combining the bare conductor radius.

[0128] Specific limitations regarding the structure-sensing framework-based overhead power line icing thickness detection system can be found in the above-described limitations of the structure-sensing framework-based overhead power line icing thickness detection method, and will not be repeated here. Each module in the aforementioned structure-sensing framework-based overhead power line icing thickness detection system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0129] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of a method for detecting the icing thickness of overhead power lines based on a structure-aware framework.

[0130] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for detecting the icing thickness of overhead power lines based on a structure-aware framework.

[0131] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0132] The foregoing has provided a detailed description of the method and system for detecting ice thickness on overhead power lines based on a structure-sensing framework, as provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention, and the descriptions of the embodiments above are merely for the purpose of helping to understand the core ideas of this invention. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this invention.

Claims

1. A method for detecting the icing thickness of overhead power lines based on a structure-sensing framework, characterized in that, The method includes the following steps: S100: The UAV's lidar collects point cloud data of overhead power lines, preprocesses the point cloud data, and uses a multilayer perceptron to perform feature upscaling on the preprocessed point cloud data. S200: The upscaled point cloud data is input into the structure-aware semantic segmentation network, and the semantic features of the wire point cloud are extracted through the encoder-decoder architecture. The encoder contains a four-layer structure, and each layer performs local spatial encoding, cross-scale spatial attention fusion and local aggregation operations in sequence to generate multi-scale fusion features. The decoder contains a symmetrical four-layer structure, which upsamples through nearest neighbor interpolation and fuses with the encoder features. Finally, the wire semantic segmentation result is output through a shared fully connected classification head. S300: Extract the conductor point set based on the semantic segmentation results, separate the multiple conductors through Euclidean clustering, and perform biplane projection and polynomial fitting on each conductor to construct a three-dimensional centerline; S400: For each conductor strand, calculate the shortest Euclidean distance from each conductor point to the fitted centerline. Based on the statistical distribution of the distances from the conductor points to the centerline, extract the shell point set and calculate the conductor envelope radius. Combine this with the bare conductor radius to estimate the icing thickness.

2. The method according to claim 1, characterized in that, Each encoder in the S200 includes a local spatial coding module, a cross-level spatial coding module, and a local aggregation module; After downsampling the features of the input encoder, the local spatial coding module aggregates the neighborhood features through grouping operations, and extracts the local geometric structure through max pooling and MLP. The cross-level spatial coding module concatenates the output of the current layer's local spatial coding module with the features of the previous layer, constructs a query vector Q, a key vector K, and a value vector V through the MLP-ReLU module, and calculates attention weights through matrix multiplication and then weighted and fused them. The local aggregation module integrates the output features of the local spatial coding module and the output features of the cross-level spatial coding module.

3. The method according to claim 2, characterized in that, The local spatial coding module includes a two-branch process. In the first branch, the input features are grouped to aggregate neighborhood points, and local features are extracted through max pooling and MLP. In the second branch, the input features are directly transformed by MLP, and the outputs of the two branches are concatenated and then generated by MLP to produce local enhanced features.

4. The method according to claim 3, characterized in that, The cross-level spatial coding module concatenates the output of the current layer's local spatial coding module with the features from the previous layer, specifically: ； in, For features of the next higher level, For the current layer features, T(·) represents the projection function. Features after splicing; Cross-scale spatial correlation features are generated through the attention weight calculation module, specifically: ； in, , , Both consist of two MLP layers, used to map input features to query, key, and value representations in the attention space. This represents the attention weight matrix, used to characterize the strength of the relationships between points, and is used for weighted aggregation of features. Indicates the dimension of the key vector; The weighted fused features are output to the next encoding layer, specifically as follows: ； in, These are adjustable fusion weights used to balance the importance of local features and global attention-enhanced features. This indicates the output of the cross-level spatial encoding module.

5. The method according to claim 4, characterized in that, The S300 includes: S310: Extract the conductor category point set based on the semantic segmentation results output by S200, and perform isolated point removal and connectivity filtering on the conductor category point set to remove local noise and misclassified points; S320: Based on the adjacency relationship constructed by the Euclidean distance between points, the region growing clustering method is used to divide the point set of the conductor category into several spatially connected subsets, each subset corresponding to different strands of conductor; in the clustering process, a distance threshold is set to ensure the spatial continuity of the same strand of conductor, and a minimum cluster size is set to filter fragmented clustering results; S330: For each set of conductor points, a three-dimensional centerline trajectory is constructed using a local polynomial fitting method based on its spatial distribution characteristics.

6. The method according to claim 5, characterized in that, The S330 includes: S331: Project the guide point to , On a plane; S332: Perform nth-order polynomial fitting on the coordinate points in each projection plane to construct the centerline trajectory model: ； in, These are coordinate components in three-dimensional space; and They are respectively plane and The polynomial coefficients of the plane-fitted curve represent the th... The second term and the first The weight of the next term; and The order of the polynomial used controls the complexity of the fitted curve. , This is the index of the number of terms in a polynomial expansion; S333: Using the least squares method to fit the coefficients in the centerline fitting model and Solve to minimize the vertical distance residual of the traverse point cloud relative to the fitted curve in the corresponding projection plane; when... Taking a plane as an example, let the set of points of the projected traverse be... Then the fitting coefficient This was obtained by solving the following optimization problem: ； Similarly, in Construct a fitting model in the plane, and let the projection point be... Then the coefficient satisfy: ； Fit coefficients Sum of coefficients Both minimization problems are linear least squares fitting problems, which are solved by constructing a system of normal equations or a pseudo-inverse matrix. The resulting fitting coefficients are used to reconstruct the centerline trajectory function of the conductor in three-dimensional space. Specifically: ； in, For the curve parameters on the center line, take a one-dimensional variable. .

7. The method according to claim 6, characterized in that, The S400 includes: S410: For each conductor strand Calculate each traverse point The shortest Euclidean distance to the fitted centerline: ； in, Indicates the first From the first guide point to the second Minimum radial distance from the centerline of a conductor strand; For the first The three-dimensional spatial coordinates of each guide point; Indicates the first The centerline trajectory function of a conductor. The curve parameters are those along the center line; Represents the Euclidean norm in two-dimensional or three-dimensional space; S420: Based on the statistical feature extraction rules, select the outermost points with the lowest ranking among all radius values ​​to form the outer shell point set. ,Right now: ； Among them, percentile function Indicates all After sorting the values ​​from smallest to largest, points ranked between 85% and 100% are selected as peripheral candidate points, while internal points in dense areas or with noise interference are excluded. S430: Shell point set Based on this, an estimated value for the envelope radius of the conductor is constructed. Specifically: ； S440: The bare conductor radius corresponding to the actual conductor model. Calculate the final estimated ice thickness. Specifically: ； in, To obtain the statistically significant radius of the conductor's outer shell envelope containing ice, The radius of the conductor body can be obtained from a database or through on-site measurement. The radial icing thickness is calculated in millimeters.

8. A system for detecting the icing thickness of overhead power lines based on a structure-sensing framework, characterized in that, include: The point cloud data acquisition and preprocessing module is used to collect point cloud data of overhead power lines by UAV lidar, preprocess the point cloud data, and perform feature upscaling on the preprocessed point cloud data through a multilayer perceptron. The structure-aware semantic segmentation module is used to input the upscaled point cloud data into the structure-aware semantic segmentation network and extract the semantic features of the wire point cloud through an encoder-decoder architecture. The encoder consists of four layers, each of which performs local spatial encoding, cross-scale spatial attention fusion and local aggregation operations in sequence to generate multi-scale fusion features. The decoder symmetrically consists of four layers, which upsample through nearest neighbor interpolation and fuse with the encoder features. Finally, the wire semantic segmentation results are output through a shared fully connected classification head. The 3D centerline trajectory fitting module is used to extract the conductor point set based on the semantic segmentation results, separate multiple conductors through Euclidean clustering, and fit the 3D centerline trajectory of each conductor. The icing thickness estimation module is used to calculate the shortest Euclidean distance from each conductor point to the fitted centerline for each conductor strand. Based on the statistical distribution of the distances from the conductor points to the centerline, it extracts the shell point set and calculates the conductor envelope radius, and estimates the icing thickness by combining the bare conductor radius.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images