Angle steel tower defect identification method based on graph neural network

Abstract

The invention discloses an angle steel tower defect identification method based on a graph neural network. The method comprises the following steps: step 1, constructing an initial structure graph; 2, constructing visual structure joint features of dominant nodes to form a first node feature set; step 3, constructing label missing perceptual coding features of the hidden nodes to form a second node feature set; 4, performing feature estimation on the second node feature set by using a feature probability migration module to form a third node feature set; 5, correcting the third node feature set to obtain final hidden node features, and forming a fourth node feature set; 6, fusing the first node feature set and the fourth node feature set to generate a graph embedding vector set; step 7, performing defect type prediction on each node; and 8, outputting a defect identification result. According to the invention, the graph neural network and the feature probability migration module are combined to realize structural level defect intelligent identification of the unobservable area of the angle steel tower.

Description

Technical Field

[0001] This invention relates to the field of structural defect identification technology, and in particular to a method for identifying defects in angle steel towers based on graph neural networks. Background Technology

[0002] With the expansion of power transmission scale and the continuous extension of communication network architecture, tall steel structures such as angle steel towers are widely used in power transmission lines and communication base stations. In order to ensure their structural safety and stable operation, efficient and reliable defect identification has become a necessary structural operation and maintenance method. At present, the defect identification of angle steel towers mainly relies on manual inspection or image recognition methods, which generally have the following technical problems:

[0003] Existing methods are mostly based on two-dimensional images for recognition, lacking the ability to model the topological connections between components and failing to reflect the spatial layout of the overall structure and the interdependence of components. When some components are missing from the image due to occlusion, angle limitations, or blind spots, existing recognition methods cannot effectively determine the structural state of that area, often resulting in false positives or false negatives. Traditional deep learning methods usually classify each component region separately, ignoring the important influence of structural context information on the formation and propagation of defects, making it difficult to identify related defects or structural deformations. In scenarios where node information is incomplete or some areas are unobservable, existing models generally lack the ability to intelligently reason and complete missing areas.

[0004] Therefore, how to provide a method for identifying defects in angle steel towers based on graph neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a defect identification method for angle steel towers based on graph neural networks. This invention introduces structural conjugate modeling and feature probability transfer modules based on graph neural networks to achieve intelligent estimation and classification of defects in unobservable areas of angle steel towers. It has the advantages of reasonable structural modeling, accurate defect identification, and wide applicability.

[0006] A method for identifying defects in angle steel towers based on graph neural networks according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Collect image data of the angle steel tower structure and the connection relationship of its components, and construct an initial structural diagram including explicit nodes and implicit nodes;

[0008] Step 2: Extract the visual features of the corresponding image data for each explicit node, and combine them with geometric dimensions, spatial location, component type and connection method to construct the joint visual structural features of each explicit node, forming the first node feature set;

[0009] Step 3: For each latent node, based on the topological connection relationship, obtain the set of adjacent explicit nodes, construct the label missing awareness coding feature of each latent node, and form the second node feature set;

[0010] Step 4: Using the first node feature set as a conditional input, the feature probability transfer module is used to estimate the features of the second node feature set, and the estimated features of each hidden node are obtained to form the third node feature set.

[0011] Step 5: Match the subgraph templates with the same topology and component functions in the labeled historical samples with the initial structure graph, and modify the third node feature set according to the matching results to obtain the final hidden node features, forming the fourth node feature set;

[0012] Step 6: Merge the feature sets of the first node and the feature sets of the fourth node to construct a complete node feature input graph, and input it into the graph neural network model to generate a graph embedding vector set;

[0013] Step 7: Based on the graph embedding vector set, perform defect identification and classification operations to predict the defect type for each node;

[0014] Step 8: Output the defect identification results for all nodes in the angle steel tower structure.

[0015] Optionally, step one specifically includes:

[0016] Collect image data of the angle steel tower from multiple perspectives, combine sensor pose information to perform spatial registration of the images, and generate an image coverage area mapping corresponding to the three-dimensional model of the angle steel tower.

[0017] Based on the image coverage area mapping, it is determined whether each component has been completely captured. If the component has been completely captured, it is marked as a visible node; otherwise, it is marked as a hidden node.

[0018] Based on the 3D model of the angle steel tower, the physical connection relationship between the components is identified, and the connection edges between the components are constructed.

[0019] Based on the explicit nodes, implicit nodes, and connecting edges between components, an initial structural diagram is constructed.

[0020] Optionally, step two specifically includes:

[0021] For each explicit node, visual features are extracted based on a convolutional neural network for the corresponding image region. These visual features include surface texture distribution, edge contours, and color gradient information.

[0022] Alternatively, the geometric dimensions of the component corresponding to the explicit node, including the component's length, cross-sectional width, thickness, and geometric shape category;

[0023] Extract the spatial position information of the components in the overall structure of the angle steel tower. The spatial position information includes the three-dimensional coordinate position, installation height, spatial orientation, and relative positional relationship with adjacent components.

[0024] The component types are determined, including angle steel members, connecting plates, stiffening ribs, and bolts, and the connection methods with other components are identified, including welded connections, bolted connections, and plug-in connections.

[0025] The visual features, geometric dimension parameters, spatial location information, component type, and connection method are encoded to form a joint visual structure feature, which serves as the node feature vector of the explicit node. The node feature vectors of all explicit nodes are then aggregated to construct the first node feature set.

[0026] Optionally, step three specifically includes:

[0027] For each hidden node, based on the topological connection relationship in the initial structure graph, retrieve all explicit nodes directly connected to the hidden node and construct an adjacent explicit node set;

[0028] Based on the connecting edges between the hidden nodes and each adjacent explicit node, calculate the spatial direction, edge length, included angle, and center point coordinates of the connecting edges to form an adjacency geometric relationship vector;

[0029] For each latent node, a label missing awareness coding feature is constructed, wherein the label missing awareness coding feature includes:

[0030] The average visual structural feature vector of adjacent dominant nodes;

[0031] Statistical characteristics of adjacency geometric relationship vectors, including mean, variance, maximum and minimum values;

[0032] The estimated spatial coordinates of the hidden nodes are calculated by three-dimensional interpolation based on the positions of the adjacent explicit nodes;

[0033] The label missing status indicator is used to explicitly indicate that the current node is a node with missing image information;

[0034] The label-missing awareness encoding features of all latent nodes are summarized to form the second node feature set.

[0035] Optionally, the feature probability transfer module specifically includes a structural conjugate modeling unit, a reverse reasoning generation unit, and a multi-solution distribution reordering unit;

[0036] The structural conjugate modeling unit is used to construct a set of higher-order conjugate neighborhoods in the initial graph structure for each latent node. The higher-order conjugate neighborhoods contain a set of all explicit nodes with a path length of no more than k to the latent node, and obtain the structural conjugate tensor within the neighborhood.

[0037] The reverse reasoning generation unit is used to deduce the defect propagation path dependency between components based on the structural conjugate tensor and the first node feature set. It takes the explicit node features with defect abnormal response as the defect propagation starting node and uses the overall credibility of path propagation as the edge weight to construct a reverse defect causal graph with implicit nodes as target nodes. The reverse defect causal graph is a directed weighted graph.

[0038] Based on the reverse defect causal graph, a variational autoencoder structure is adopted. The input is the visual structure joint feature and edge weight combination vector of the defect propagation starting node, and the output is the set of multi-solution estimated feature distribution of the latent node. The set of multi-solution estimated feature distribution is a set of candidate feature vectors simulated and generated in the feature space.

[0039] The multi-solution distribution reordering unit is used to construct a ranking scoring function from the multi-solution estimated feature distribution set to score and reorder the candidate feature vectors of each latent node. The ranking scoring function is the average cosine similarity between the candidate feature vector and the feature vector of the corresponding defect propagation starting node, and the candidate feature vector with the highest score is selected as the estimated feature of the latent node, and the third node feature set is output.

[0040] Optionally, step five specifically includes:

[0041] Extract subgraph templates with the same topological structure and component functions from labeled historical samples, and construct a subgraph template library. The subgraph template includes node connection relationships, component types, component functions, and corresponding node feature representations.

[0042] The local region containing hidden nodes in the initial structure graph is taken as the target subgraph, and subgraph matching is performed with the subgraph templates in the subgraph template library to determine the structurally identical subgraph templates.

[0043] Using the template features of the corresponding nodes in the matched subgraph template as reference features, the estimated features of the latent nodes in the third node feature set are corrected. The correction methods include linear interpolation fusion and feature bias compensation to obtain the final latent node features, forming the fourth node feature set.

[0044] Optionally, step six specifically includes:

[0045] The explicit node features in the first node feature set and the final implicit node features in the fourth node feature set are fused according to the node index to form a complete node feature set.

[0046] Combining the node and edge information in the initial structural graph, a complete node feature input graph is constructed, which includes the node features and connection relationships corresponding to all components.

[0047] The complete node feature input graph is input into the graph neural network model, and the node and adjacency information are aggregated using graph convolution operations to generate a set of graph embedding vectors for all nodes in the graph.

[0048] Optionally, step seven specifically includes:

[0049] The graph embedding vector of each node in the graph embedding vector set is used as the input feature of the classification model to predict the defect type.

[0050] The classification module includes multiple fully connected neural network layers, which are used to perform feature transformation and classification output on the graph embedding vector of each node;

[0051] Output the defect type prediction result for each node, which includes corrosion, cracks, loosening, missing parts and structural deformation.

[0052] Optionally, the defect identification result includes the node identifier of each node, the spatial location of the node, and the corresponding defect type prediction result.

[0053] The beneficial effects of this invention are:

[0054] This invention addresses the problems of component occlusion, image gaps, and weak structural correlation modeling in angle steel tower defect identification by constructing a graph structure representation that integrates explicit and implicit nodes, combined with label missing awareness coding and structural conjugate modeling. It proposes a structural-level defect reasoning method based on graph neural networks. Label missing awareness vectors are constructed by combining the joint features of explicit node visual structures and topological relationships. A structure-guided inverse defect causal graph and variational generative network are used to complete features in unobservable regions. Subgraph template transfer is introduced for feature correction based on structural continuity and functional consistency, ultimately achieving feature fusion modeling and node-level defect classification of the entire structural graph. Compared to traditional image-based independent recognition methods, this invention enhances the ability to judge the state of hidden node defects through high-order conjugate neighborhood modeling and a multi-solution generation and reordering mechanism, and possesses the ability to structurally perceive and reason about complex defect paths. Through a structural template alignment mechanism, it further improves the feature estimation accuracy of missing regions and the overall recognition robustness, and finally achieves accurate output of the defect type, spatial location and recognition confidence of each node of the angle steel tower. It is suitable for application scenarios such as intelligent inspection of angle steel tower structures and automatic recognition of images captured by UAVs, and has the advantages of strong comprehensiveness, high structural adaptability and strong practical deployment capability. Attached Figure Description

[0055] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0056] Figure 1 This is an overall flowchart of a defect identification method for angle steel towers based on graph neural networks proposed in this invention;

[0057] Figure 2 This is a schematic diagram of the feature probability transfer module of the angle steel tower defect identification method based on graph neural network proposed in this invention. Detailed Implementation

[0058] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0059] refer to Figure 1 and Figure 2 A method for identifying defects in angle steel towers based on graph neural networks includes the following steps:

[0060] Step 1: Collect image data of the angle steel tower structure and the connection relationship of its components, and construct an initial structural diagram including explicit nodes and implicit nodes;

[0061] Step 2: Extract the visual features of the corresponding image data for each explicit node, and combine them with geometric dimensions, spatial location, component type and connection method to construct the joint visual structural features of each explicit node, forming the first node feature set;

[0062] Step 3: For each latent node, based on the topological connection relationship, obtain the set of adjacent explicit nodes, construct the label missing awareness coding feature of each latent node, and form the second node feature set;

[0063] Step 4: Using the first node feature set as a conditional input, the feature probability transfer module is used to estimate the features of the second node feature set, and the estimated features of each hidden node are obtained to form the third node feature set.

[0064] Step 5: Match the subgraph templates with the same topology and component functions in the labeled historical samples with the initial structure graph, and modify the third node feature set according to the matching results to obtain the final hidden node features, forming the fourth node feature set;

[0065] Step 6: Merge the feature sets of the first node and the feature sets of the fourth node to construct a complete node feature input graph, and input it into the graph neural network model to generate a graph embedding vector set;

[0066] Step 7: Based on the graph embedding vector set, perform defect identification and classification operations to predict the defect type for each node;

[0067] Step 8: Output the defect identification results for all nodes in the angle steel tower structure.

[0068] In this embodiment, step one specifically includes:

[0069] Collect image data of the angle steel tower from multiple perspectives, combine sensor pose information to perform spatial registration of the images, and generate an image coverage area mapping corresponding to the three-dimensional model of the angle steel tower.

[0070] Based on the image coverage area mapping, it is determined whether each component has been completely captured. If the component has been completely captured, it is marked as a visible node; otherwise, it is marked as a hidden node.

[0071] Based on the 3D model of the angle steel tower, the physical connection relationship between the components is identified, and the connection edges between the components are constructed.

[0072] Based on the explicit nodes, implicit nodes, and connecting edges between components, an initial structural diagram is constructed.

[0073] In this embodiment, step two specifically includes:

[0074] For each explicit node, visual features are extracted based on a convolutional neural network for the corresponding image region. These visual features include surface texture distribution, edge contours, and color gradient information.

[0075] Alternatively, the geometric dimensions of the component corresponding to the explicit node, including the component's length, cross-sectional width, thickness, and geometric shape category;

[0076] Extract the spatial position information of the components in the overall structure of the angle steel tower. The spatial position information includes the three-dimensional coordinate position, installation height, spatial orientation, and relative positional relationship with adjacent components.

[0077] The component types are determined, including angle steel members, connecting plates, stiffening ribs, and bolts, and the connection methods with other components are identified, including welded connections, bolted connections, and plug-in connections.

[0078] The visual features, geometric dimension parameters, spatial location information, component type, and connection method are encoded to form a joint visual structure feature, which serves as the node feature vector of the explicit node. The node feature vectors of all explicit nodes are then aggregated to construct the first node feature set.

[0079] In this embodiment, step three specifically includes:

[0080] For each hidden node, based on the topological connection relationship in the initial structure graph, retrieve all explicit nodes directly connected to the hidden node and construct an adjacent explicit node set;

[0081] Based on the connecting edges between the hidden nodes and each adjacent explicit node, calculate the spatial direction, edge length, included angle, and center point coordinates of the connecting edges to form an adjacency geometric relationship vector;

[0082] For each latent node, a label missing awareness coding feature is constructed, wherein the label missing awareness coding feature includes:

[0083] The average visual structural feature vector of adjacent dominant nodes;

[0084] Statistical characteristics of adjacency geometric relationship vectors, including mean, variance, maximum and minimum values;

[0085] The estimated spatial coordinates of the hidden nodes are calculated by three-dimensional interpolation based on the positions of the adjacent explicit nodes;

[0086] The label missing status indicator is used to explicitly indicate that the current node is a node with missing image information;

[0087] The label-missing awareness encoding features of all latent nodes are summarized to form the second node feature set.

[0088] In this embodiment, the feature probability transfer module specifically includes a structural conjugate modeling unit, a reverse reasoning generation unit, and a multi-solution distribution reordering unit;

[0089] The structural conjugate modeling unit is used to construct a set of higher-order conjugate neighborhoods in the initial graph structure for each latent node. The higher-order conjugate neighborhoods contain a set of all explicit nodes with a path length of no more than k to the latent node, and obtain the structural conjugate tensor within the neighborhood.

[0090] The reverse reasoning generation unit is used to deduce the defect propagation path dependency between components based on the structural conjugate tensor and the first node feature set. It takes the explicit node features with defect abnormal response as the defect propagation starting node and uses the overall credibility of path propagation as the edge weight to construct a reverse defect causal graph with implicit nodes as target nodes. The reverse defect causal graph is a directed weighted graph.

[0091] In this invention, to construct a reverse defect causal graph targeting hidden nodes, it is first necessary to identify nodes with obvious defect symptoms from the explicit nodes, which serve as the starting points for defect propagation. This process is accomplished by calculating the defect anomaly response score for each explicit node.

[0092] Specifically, the defect anomaly response score R for each explicit node is calculated. i :

[0093] R i =σ(W f ·F i +b f );

[0094] Among them, F i Indicates display node v i The joint feature vector of visual structure, W f b represents the weights of the fully connected layer. f σ represents the bias, and σ represents the sigmoid activation function, used to normalize to the interval [0,1].

[0095] When R i >θ r When, display node v i The defective abnormal response is considered to be denoted as the propagation starting point node, where θ is the defect propagation starting point node. r This indicates a preset response score threshold;

[0096] After determining the starting node of the propagation, it is necessary to calculate the probability that the node will transmit the defect information to the hidden node through the path in the initial structure graph, which will be used for the edge weight assignment in the reverse causal graph.

[0097] For a line starting from the origin node vs To the target hidden node v t The path sequence P, whose path propagation probability T(P) is defined as the overall reliability T(P) of transmitting defect information on that path:

[0098] T(P) = Π (u,v)∈P α(u,v);

[0099] Where α(u,v) represents the local propagation weight of edge (u,v) in path sequence P, and its value range is [0,1].

[0100] α(u,v)=λ1·Conn(u,v)+λ2·exp(-d uv )+λ3·cos(θ uv );

[0101] Where Conn(u,v) represents the weight score of the component connection type, with bolted connections receiving 1.0, welded connections receiving 0.8, and plug-in connections receiving 0.5. uv θ represents the three-dimensional spatial distance between node u and node v. uv Let λ1, λ2, and λ3 represent the angle between the side direction and the conduction direction, exp represent the exponential function, and λ1, λ2, and λ3 are normalized weighting coefficients that satisfy λ1 + λ2 + λ3 = 1.

[0102] Based on the reverse defect causal graph, a variational autoencoder structure is adopted. The input is the visual structure joint feature and edge weight combination vector of the defect propagation starting node, and the output is the set of multi-solution estimated feature distribution of the latent node. The set of multi-solution estimated feature distribution is a set of candidate feature vectors simulated and generated in the feature space.

[0103] In this invention, to achieve defect feature estimation of unobservable latent nodes under structure-aware conditions, a variational autoencoder structure based on a reverse defect causal graph is designed. The structure receives input including joint visual structural feature vectors of multiple defect propagation starting points and edge weight combination vectors of the paths to latent nodes; these are concatenated to form a joint input tensor. This joint input tensor is fed into an encoder network to generate parameters representing the latent node's potential feature distribution, including a mean vector and a log-variance vector. A reparameterized sampling technique is used to sample multiple latent vector samples from the Gaussian distribution, representing the possible feature states of latent nodes under different defect propagation paths. Each sample is mapped to a candidate feature vector by the decoder, forming a set of multiple-solution estimated feature distributions for latent nodes. This set of multiple-solution estimated feature distributions reflects the various component states that latent nodes may exhibit in the feature space under the current structural and defect propagation information, possessing uncertainty modeling capabilities. This facilitates subsequent selection of the optimal estimate through a sorting mechanism, thereby improving the overall accuracy and reliability of defect identification.

[0104] The multi-solution distribution reordering unit is used to construct a ranking scoring function from the multi-solution estimated feature distribution set to score and reorder the candidate feature vectors of each latent node. The ranking scoring function is the average cosine similarity between the candidate feature vector and the feature vector of the corresponding defect propagation starting node, and the candidate feature vector with the highest score is selected as the estimated feature of the latent node, and the third node feature set is output.

[0105] In this embodiment, step five specifically includes:

[0106] Extract subgraph templates with the same topological structure and component functions from labeled historical samples, and construct a subgraph template library. The subgraph template includes node connection relationships, component types, component functions, and corresponding node feature representations.

[0107] The local region containing hidden nodes in the initial structure graph is taken as the target subgraph, and subgraph matching is performed with the subgraph templates in the subgraph template library to determine the structurally identical subgraph templates.

[0108] Using the template features of the corresponding nodes in the matched subgraph template as reference features, the estimated features of the latent nodes in the third node feature set are corrected. The correction methods include linear interpolation fusion and feature bias compensation to obtain the final latent node features, forming the fourth node feature set.

[0109] In this embodiment, step six specifically includes:

[0110] The explicit node features in the first node feature set and the final implicit node features in the fourth node feature set are fused according to the node index to form a complete node feature set.

[0111] Combining the node and edge information in the initial structural graph, a complete node feature input graph is constructed, which includes the node features and connection relationships corresponding to all components.

[0112] The complete node feature input graph is input into the graph neural network model, and the node and adjacency information are aggregated using graph convolution operations to generate a set of graph embedding vectors for all nodes in the graph.

[0113] In this embodiment, step seven specifically includes:

[0114] The graph embedding vector of each node in the graph embedding vector set is used as the input feature of the classification model to predict the defect type.

[0115] The classification module includes multiple fully connected neural network layers, which are used to perform feature transformation and classification output on the graph embedding vector of each node;

[0116] In this invention, a classification module based on a fully connected neural network is constructed to achieve automatic identification and classification of defect types for each node. This module takes the graph embedding vector set output by the graph neural network as input. The embedding vector represents the comprehensive feature information of each node within the overall graph structure context, possessing strong semantic expressive power. For each node, its graph embedding vector is first input to the first fully connected neural network layer for dimension mapping and nonlinear feature transformation. Subsequently, through one or more subsequent fully connected layers, high-level features related to the defect type are gradually extracted. Nonlinear transformations are performed between layers using the LeakyReLU activation function, and Dropout is used to enhance the model's generalization ability. Finally, a fully connected output layer with an output dimension corresponding to the number of defect types is combined with the Softmax function to achieve probabilistic prediction of multiple defect types. The classification result is the defect type label corresponding to each node, used for subsequent structural defect labeling and maintenance decision support. This classification module features simple structure, high inference efficiency, and ease of deployment, making it suitable for batch node parallel prediction scenarios.

[0117] Output the defect type prediction result for each node, which includes corrosion, cracks, loosening, missing parts and structural deformation.

[0118] In this embodiment, the defect identification result includes the node identifier of each node, the spatial location of the node, and the corresponding defect type prediction result.

[0119] Example 1:

[0120] To verify the feasibility of this invention in practice, it was applied to the inspection of angle steel towers on a typical 110kV transmission line in a certain region. The terrain of this line is complex, with numerous angle steel towers, some of which are located in densely vegetated or obstructed areas. Traditional methods relying on remote observation or image recognition are insufficient to achieve high coverage and accuracy in defect detection of all components, especially for structural areas (hidden nodes) that are difficult to see or assess due to unclear images or severe obstruction, creating long-standing blind spots.

[0121] In this embodiment, an unmanned aerial vehicle (UAV) platform equipped with a multi-view vision system and IMU synchronization function is used to collect image data of the angle steel tower. Visual positioning and spatial registration algorithms are used to align the images with the spatial information of the components. Subsequently, in the method described in this invention, based on image quality assessment and coverage determination, components that are completely covered are automatically designated as explicit nodes, while the remaining areas are marked as implicit nodes, thus constructing a structural diagram. Each explicit node has its texture, edge, and color features extracted by the image analysis module, and these features are combined with its geometric dimensions, relative spatial position, component type, and connection method to encode a joint feature vector. This feature vector serves as the node input information in the model processing flow.

[0122] For latent nodes that cannot be directly observed, the feature probability transfer module of this invention identifies their higher-order neighborhood components through structural conjugate modeling units. It then propagates the features of adjacent explicit nodes to the target latent node via backpropagation graph convolution. Combining multiple-solution estimation and distribution reordering strategies, it simulates and selects features to ultimately estimate the possible feature vectors of the latent node. Simultaneously, it matches subgraph templates with consistent topology from the historical sample library to further correct the estimated features of the latent node, thereby enhancing the rationality and structural consistency of the estimation results.

[0123] After the above processing, the feature sets of the first and fourth nodes are fused to construct a complete node feature map, which is then input into a multi-layer graph convolutional network model. The graph neural network performs feature aggregation and embedding generation on node information and adjacency structure, extracting high-order context features of the entire graph. In the classification stage, a three-layer fully connected neural network is used to process the graph embedding vector of each node and output its defect type prediction result.

[0124] The method of this invention not only achieves effective feature modeling and recognition of hidden nodes that are difficult to handle by traditional methods, but also significantly improves the accuracy and completeness of overall defect recognition. To verify the recognition performance of this method, 30 angle steel towers were randomly selected for comparative experiments. The comparison items included traditional CNN image recognition methods, GAT methods, and the GNN inference fusion method proposed in this invention. The recognition accuracy statistics are shown in Table 1 below.

[0125] Table 1 Comparison of Defect Identification Methods for Angle Steel Towers

[0126]

[0127] As shown in Table 1 above, different methods exhibit significant differences in their accuracy in recognizing explicit nodes, implicit nodes, and overall performance. Traditional CNN image recognition methods achieve an accuracy of 89.2% in recognizing explicit nodes, but only 36.5% in recognizing implicit nodes. This indicates that while the method has some ability to recognize visible areas of the image, its performance in recognizing missing or invisible areas (i.e., implicit nodes) is severely inadequate, limiting its reliability in complex structural scenarios.

[0128] The GAT method achieved a 91.7% accuracy rate in identifying explicit nodes, which is a slight improvement over traditional CNNs. However, its accuracy rate in identifying implicit nodes was only 54.3%. Although this is an improvement, it still suffers from significant defects, such as the loss of propagation path information, making it difficult to accurately estimate the state of implicit components.

[0129] In comparison, the GNN inference fusion method proposed in this invention achieves an accuracy of 94.5% in identifying explicit nodes and a significant improvement to 83.2% in identifying implicit nodes, with an overall accuracy of 90.1%. This result demonstrates that this invention not only retains the high recognition capability of traditional image recognition for explicit regions but also effectively infers the component defect features of missing image regions through feature probability transfer and causal inference mechanisms, significantly enhancing the robustness and comprehensiveness of the system.

[0130] Meanwhile, to verify the difference in recognition effect of the present invention under different defect types, the defect categories of the recognition results were statistically analyzed, as shown in Table 2 below.

[0131] Table 2 Statistical Table of Defect Type Identification Results

[0132] Defect type Actual number of existence Number of recognition by conventional method Number of recognition by the present invention Accuracy improvement Rust 184 139 170 +16.8％ Crack 93 61 85 +25.8％ Loose component 112 74 101 +24.1％ Missing component 43 29 39 +23.2％ Structure deformation 67 44 58 +20.9％

[0133] Based on the data in Table 2 above, it can be seen that the method of the present invention significantly outperforms traditional methods in various defect identification tasks. In rust type identification, the actual number of rust types was 184, while the traditional method identified 139, and the present invention identified 170, resulting in a 16.8% improvement in accuracy. This improvement reflects the superior perceptual ability of the present method in extracting texture features of surface corrosion areas and jointly modeling visible and hidden nodes.

[0134] For crack identification, traditional methods can identify 61 cracks, while this invention increases the number to 85, improving accuracy by 25.8%. Cracks are often subtle in shape and highly directional, making it difficult for traditional methods to accurately extract their edge information. This invention, however, effectively enhances the ability to perceive crack boundaries through graph neural network structure propagation and feature transfer.

[0135] In the identification of loose components, the actual number was 112, while the traditional method identified 74, and the present invention identified 101, improving the accuracy by 24.1%. Component loosening usually requires a comprehensive judgment based on spatial relationships and connection methods. The present invention has advantages in adjacency geometric relationship modeling and causal propagation path construction, thus improving the identification accuracy.

[0136] This invention also demonstrates good performance in addressing the problem of missing components. Of the 43 actual missing components, traditional methods identified 29, while this method identified 39, improving accuracy by 23.2%. Identifying hidden component missing components places extremely high demands on the model's feature completion capabilities. This method effectively solves the identification difficulties caused by missing labels through subgraph matching and estimated feature correction mechanisms.

[0137] Regarding structural deformation, traditional methods only identified 44 deformations, while the method of this invention improves this to 58, representing a 20.9% improvement. Structural deformation involves the perception of overall spatial attitude and mechanical offset. This invention achieves higher-dimensional information fusion and classification accuracy through the comprehensive encoding of node spatial positions and relative angles.

[0138] This invention outperforms traditional identification methods in all five common defect types of angle steel towers, with accuracy improvements ranging from 16.8% to 25.8%. It demonstrates particularly strong robustness and generalization ability in complex defect patterns and hidden node scenarios. By combining structural conjugate modeling with multi-solution feature estimation, it effectively overcomes the limitations of traditional image-based methods in structural understanding and defect propagation modeling.

[0139] This embodiment proposes a new paradigm for defect identification in complex structural scenarios of angle steel towers by deeply integrating graph neural network technology with structural feature modeling methods. In the initial stage, the invention constructs a graph structure model of explicit and implicit nodes, comprehensively integrating multi-source heterogeneous features such as visual information, spatial location, and connection methods of components to achieve high-fidelity modeling of structural details. During feature estimation, structural conjugate modeling and reverse defect causal graphs are introduced. For implicit nodes in areas with missing information or blind spots in image capture, path-dependent multi-solution feature generation is performed, significantly improving the model's ability to represent blurred areas of the local structure. Furthermore, feature correction is achieved by combining historical template subgraph matching, enhancing the semantic consistency and functional rationality of the estimated features. By uniformly fusing explicit and implicit node features through graph convolutional networks, structural-level defect type identification is achieved, demonstrating good generalization performance in multi-category defect scenarios. The method of this invention possesses stronger feature completion capabilities, higher identification accuracy, and better robustness in challenging scenarios such as missing node information, complex structural relationships, and incomplete image acquisition, providing an efficient and reliable technical means for the health inspection of practical engineering structures.

[0140] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for defect identification of angle steel towers based on graph neural networks, characterized in that, Includes the following steps: Step 1: Collect image data of the angle steel tower structure and the connection relationship of its components, and construct an initial structural diagram including explicit nodes and implicit nodes; Step 2: Extract the visual features of the corresponding image data for each explicit node, and combine them with geometric dimensions, spatial location, component type and connection method to construct the joint visual structural features of each explicit node, forming the first node feature set; Step 3: For each latent node, based on the topological connection relationship, obtain the set of adjacent explicit nodes, construct the label missing awareness coding feature of each latent node, and form the second node feature set; Step 4: Using the first node feature set as a conditional input, the feature probability transfer module is used to estimate the features of the second node feature set, and the estimated features of each hidden node are obtained to form the third node feature set. Step 5: Match the subgraph templates with the same topology and component functions in the labeled historical samples with the initial structure graph, and modify the third node feature set according to the matching results to obtain the final hidden node features, forming the fourth node feature set; Step 6: Merge the feature sets of the first node and the feature sets of the fourth node to construct a complete node feature input graph, and input it into the graph neural network model to generate a graph embedding vector set; Step 7: Based on the graph embedding vector set, perform defect identification and classification operations to predict the defect type for each node; Step 8: Output the defect identification results for all nodes in the angle steel tower structure.

2. The method for defect identification of angle steel towers based on graph neural networks according to claim 1, characterized in that, Step one specifically involves: Collect image data of the angle steel tower from multiple perspectives, combine sensor pose information to perform spatial registration of the images, and generate an image coverage area mapping corresponding to the three-dimensional model of the angle steel tower. Based on the image coverage area mapping, it is determined whether each component has been completely captured. If the component has been completely captured, it is marked as a visible node; otherwise, it is marked as a hidden node. Based on the 3D model of the angle steel tower, the physical connection relationship between the components is identified, and the connection edges between the components are constructed. Based on the explicit nodes, implicit nodes, and connecting edges between components, an initial structural diagram is constructed.

3. The method for defect identification of angle steel towers based on graph neural networks according to claim 1, characterized in that, Step two specifically involves: For each explicit node, visual features are extracted based on a convolutional neural network for the corresponding image region. These visual features include surface texture distribution, edge contours, and color gradient information. Alternatively, the geometric dimensions of the component corresponding to the explicit node, including the component's length, cross-sectional width, thickness, and geometric shape category; Extract the spatial position information of the components in the overall structure of the angle steel tower. The spatial position information includes the three-dimensional coordinate position, installation height, spatial orientation, and relative positional relationship with adjacent components. The component types are determined, including angle steel members, connecting plates, stiffening ribs, and bolts, and the connection methods with other components are identified, including welded connections, bolted connections, and plug-in connections. The visual features, geometric dimension parameters, spatial location information, component type, and connection method are encoded to form a joint visual structure feature, which serves as the node feature vector of the explicit node. The node feature vectors of all explicit nodes are then aggregated to construct the first node feature set.

4. The method for defect identification of angle steel towers based on graph neural networks according to claim 1, characterized in that, Step three specifically involves: For each hidden node, based on the topological connection relationship in the initial structure graph, retrieve all explicit nodes directly connected to the hidden node and construct an adjacent explicit node set; Based on the connecting edges between the hidden nodes and each adjacent explicit node, calculate the spatial direction, edge length, included angle, and center point coordinates of the connecting edges to form an adjacency geometric relationship vector; For each latent node, a label missing awareness coding feature is constructed, wherein the label missing awareness coding feature includes: The average visual structural feature vector of adjacent dominant nodes; Statistical characteristics of adjacency geometric relationship vectors, including mean, variance, maximum and minimum values; The estimated spatial coordinates of the hidden nodes are calculated by three-dimensional interpolation based on the positions of the adjacent explicit nodes; The label missing status indicator is used to explicitly indicate that the current node is a node with missing image information; The label-missing awareness encoding features of all latent nodes are summarized to form the second node feature set.

5. The method for defect identification of angle steel towers based on graph neural networks according to claim 1, characterized in that, The feature probability transfer module specifically includes a structural conjugate modeling unit, a reverse reasoning generation unit, and a multi-solution distribution reordering unit; The structural conjugate modeling unit is used to construct a set of higher-order conjugate neighborhoods in the initial graph structure for each latent node. The higher-order conjugate neighborhoods contain a set of all explicit nodes with a path length of no more than k to the latent node, and obtain the structural conjugate tensor within the neighborhood. The reverse reasoning generation unit is used to deduce the defect propagation path dependency between components based on the structural conjugate tensor and the first node feature set. It takes the explicit node features with defect abnormal response as the defect propagation starting node and uses the overall credibility of path propagation as the edge weight to construct a reverse defect causal graph with implicit nodes as target nodes. The reverse defect causal graph is a directed weighted graph. Based on the reverse defect causal graph, a variational autoencoder structure is adopted. The input is the visual structure joint feature and edge weight combination vector of the defect propagation starting node, and the output is the set of multi-solution estimated feature distribution of the latent node. The set of multi-solution estimated feature distribution is a set of candidate feature vectors simulated and generated in the feature space. The multi-solution distribution reordering unit is used to construct a ranking scoring function from the multi-solution estimated feature distribution set to score and reorder the candidate feature vectors of each latent node. The ranking scoring function is the average cosine similarity between the candidate feature vector and the feature vector of the corresponding defect propagation starting node, and the candidate feature vector with the highest score is selected as the estimated feature of the latent node, and the third node feature set is output.

6. The method for defect identification of angle steel towers based on graph neural networks according to claim 1, characterized in that, Step five specifically involves: Extract subgraph templates with the same topological structure and component functions from labeled historical samples, and construct a subgraph template library. The subgraph template includes node connection relationships, component types, component functions, and corresponding node feature representations. The local region containing hidden nodes in the initial structure graph is taken as the target subgraph, and subgraph matching is performed with the subgraph templates in the subgraph template library to determine the structurally identical subgraph templates. Using the template features of the corresponding nodes in the matched subgraph template as reference features, the estimated features of the latent nodes in the third node feature set are corrected. The correction methods include linear interpolation fusion and feature bias compensation to obtain the final latent node features, forming the fourth node feature set.

7. The method for defect identification of angle steel towers based on graph neural networks according to claim 1, characterized in that, Step six specifically involves: The explicit node features in the first node feature set and the final implicit node features in the fourth node feature set are fused according to the node index to form a complete node feature set. Combining the node and edge information in the initial structural graph, a complete node feature input graph is constructed, which includes the node features and connection relationships corresponding to all components. The complete node feature input graph is input into the graph neural network model, and the node and adjacency information are aggregated using graph convolution operations to generate a set of graph embedding vectors for all nodes in the graph.

8. The method for defect identification of angle steel towers based on graph neural networks according to claim 1, characterized in that, Step seven specifically involves: The graph embedding vector of each node in the graph embedding vector set is used as the input feature of the classification model to predict the defect type. The classification module includes multiple fully connected neural network layers, which are used to perform feature transformation and classification output on the graph embedding vector of each node; Output the defect type prediction result for each node, which includes corrosion, cracks, loosening, missing parts and structural deformation.

9. A method for defect identification of angle steel towers based on graph neural networks according to claim 1, characterized in that, The defect identification results include the node identifier of each node, the spatial location of the node, and the corresponding defect type prediction results.

Images