An orbit surface defect detection method and system

Abstract

The present invention provides a method and system for detecting rail surface defects, relating to the technical field of rail defect detection. The method includes: performing validity analysis on an initial rail surface image to obtain rail surface data, including the rail surface image and the corresponding defect classification label; extracting image node features, constructing an original graph adjacency matrix and an original graph weight matrix, dividing cluster groups and reconstructing the internal structure of the cluster groups, and constructing a cluster group adjacency matrix and a cluster group weight matrix; inputting them into a hypergraph convolutional model, and performing topology through a persistent homology mechanism to obtain the original graph topological features and the cluster group graph topological features respectively, fusing them, and outputting the current defect detection data of the rail surface image; comparing the current defect detection data with the numerical values of the defect classification labels, calculating the prediction error, and iteratively optimizing to obtain the defect detection data. The present invention can solve the technical problems of low accuracy and efficiency in the traditional manual inspection and non-destructive detection of rail surface defects in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of track defect detection, and particularly relates to a method and system for detecting track surface defects. Background Art

[0002] During the long-term operation of railway tracks, various defects such as cracks, spalls, pits, and corrugations are likely to occur on the surface due to factors such as train loads, environmental erosion, and material aging. If these defects cannot be detected and repaired in time, they may lead to damage to the track structure, a decline in the riding comfort, and even cause railway safety accidents.

[0003] At present, the detection of railway track surface defects mainly relies on manual inspections and non-destructive testing technologies such as ultrasonic testing, eddy current testing, and magnetic particle testing. Manual inspections are limited by subjective judgments, with low efficiency and susceptibility to external environmental interference, making it difficult to ensure the comprehensiveness and consistency of detections. Although traditional non-destructive testing technologies can provide relatively high detection accuracy, they usually require special equipment, have a slow detection speed, and limited adaptability to different types of defects. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technology, the purpose of the present invention is to provide a method and system for detecting track surface defects, aiming to solve the technical problems of low accuracy and efficiency in the traditional manual inspection and non-destructive testing of railway track surface defects in the existing technology.

[0005] On the one hand, the present invention provides a method for detecting track surface defects, the method comprising:

[0006] Obtaining a plurality of initial railway track surface images, performing validity analysis on the initial railway track surface images to obtain railway track surface data, the railway track surface data including the track surface images and the corresponding defect classification labels;

[0007] Extracting the image node features of the track surface images, constructing an original graph adjacency matrix and an original graph weight matrix, partitioning clusters and reconstructing the internal structure of the clusters, and constructing a cluster adjacency matrix and a cluster weight matrix;

[0008] Inputting the original graph adjacency matrix and the original graph weight matrix, as well as the cluster adjacency matrix and the cluster weight matrix into a hypergraph convolutional model respectively, and performing topology through a persistent homology mechanism to obtain original graph topological features and cluster graph topological features respectively, fusing the original graph topological features and the cluster graph topological features, and outputting the current defect detection data of the track surface images;

[0009] Comparing the current defect detection data with the numerical values of the defect classification labels, calculating the prediction error, and performing iterative optimization based on the prediction error to obtain the defect detection data.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: Through the rail surface defect detection method provided by the present invention, by extracting the image node features of the rail surface image, constructing the original graph adjacency matrix and the original graph weight matrix to extract global information, dividing the cluster groups and reconstructing the internal structure of the cluster groups, constructing the cluster group adjacency matrix and the cluster group weight matrix to amplify and extract local information, inputting the original graph adjacency matrix and the original graph weight matrix, as well as the cluster group adjacency matrix and the cluster group weight matrix into the hypergraph convolutional model respectively, the hypergraph convolutional model can effectively express the high-order relationship between local information and global information, can greatly improve the expression ability of local information and global information, can effectively capture the complex multiple relationships between nodes, and improve the accuracy and efficiency of detection; and perform topology through the persistent homology mechanism to obtain the original graph topological features and the cluster group graph topological features respectively, fuse the original graph topological features and the cluster group graph topological features, output the current defect detection data of the rail surface image, effectively capture the topological features in local information and global information, deeply excavate the topological information at different scales in local information and global information, significantly improve the expression ability of local information and global information, and further improve the accuracy and efficiency of detection, thereby solving the technical problems of low accuracy and efficiency in traditional manual inspection and non-destructive detection of rail surface defects in the prior art.

[0011] According to one aspect of the above technical solution, the steps of obtaining a plurality of initial rail surface images, performing validity analysis on the initial rail surface images to obtain rail surface data, where the rail surface data includes rail surface images and the corresponding defect classification labels specifically include:

[0012] Obtain a plurality of initial rail surface images, perform integrity verification and quality verification on the initial rail surface graphics to obtain rail surface images;

[0013] Perform defect detection analysis on the rail surface images, and mark defect classification labels on the rail surface images to obtain rail surface data.

[0014] According to one aspect of the above technical solution, the steps of extracting the image node features of the rail surface image, constructing the original graph adjacency matrix and the original graph weight matrix, dividing the cluster groups and reconstructing the internal structure of the cluster groups, constructing the cluster group adjacency matrix and the cluster group weight matrix specifically include:

[0015] Extract the image node features of the rail surface image to construct the original graph adjoint matrix;

[0016] Based on the original graph adjoint matrix, construct hyperedges to obtain the original graph adjacency matrix and the original graph weight matrix;

[0017] According to the connection between each hyperedge, divide the image node features into several cluster groups to construct the cluster group adjoint matrix;

[0018] Based on the cluster adjoint matrix, representative node features are screened, hyperedges are reconstructed, and a cluster adjacency matrix and a cluster weight matrix are obtained.

[0019] According to one aspect of the above technical solution, the steps of screening representative node features based on the cluster adjoint matrix, reconstructing hyperedges, and obtaining a cluster adjacency matrix and a cluster weight matrix specifically include:

[0020] The cluster adjoint matrix includes a weight matrix and a node feature matrix. The node feature matrix is input into the embedded Si-Net network for deep feature mapping. After being processed by the global feature extraction layer, a global feature matrix after non-linear projection is obtained;

[0021] Based on the global feature matrix, an attention weight matrix is calculated. According to the attention weight matrix and the global feature matrix, a weighted feature matrix is calculated. The calculation formula is as follows:

[0022] ,

[0023] where, is the weighted feature matrix, is the global feature matrix, is the attention weight matrix, , are trainable parameters, is the activation function, is the rectified linear unit function;

[0024] The weighted feature matrix is screened, the maximum feature difference degree is calculated, representative node features are screened, hyperedges are reconstructed, and a cluster adjacency matrix is obtained;

[0025] The weighted feature matrix is weighted by a distance-aware kernel function to obtain a cluster weight matrix. The calculation formula is as follows:

[0026] ,

[0027] where, is the cluster weight matrix, is the node in the weighted feature matrix vector, is the node in the weighted feature matrix vector, is and the distance between, is the node in the weighted feature matrix in the The element of dimension is a node in the weighted feature matrix at the dimension is the weighted feature matrix of the total dimension and is a learnable matrix parameter.

[0028] According to one aspect of the above technical solution, the steps of inputting the original graph adjacency matrix, the original graph weight matrix, the cluster group adjacency matrix, and the cluster group weight matrix into the hypergraph convolution model respectively and performing topology through the persistent homology mechanism to obtain the original graph topological features and the cluster group graph topological features specifically include:

[0029] Input the original graph adjacency matrix, the original graph weight matrix, the cluster group adjacency matrix, and the cluster group weight matrix into the hypergraph convolution model respectively to obtain the original graph convolution matrix and the cluster group convolution matrix. The calculation formula is as follows:

[0030] ,

[0031] where is the original graph convolution matrix or the cluster group convolution matrix of the layer, is the original graph convolution matrix or the cluster group convolution matrix of the layer, is the hypergraph degree matrix, is the original graph adjacency matrix or the cluster group adjacency matrix, is the transposed matrix of is the original graph weight matrix or the cluster group weight matrix, is the activation function;

[0032] Input the original graph convolution feature matrix and the cluster group convolution matrix into the persistent homology mechanism for topology respectively to obtain the original graph node topological features and the cluster group graph topological features.

[0033] According to one aspect of the above technical solution, the steps of inputting the original graph convolution feature matrix and the cluster group convolution matrix into the persistent homology mechanism for topology respectively to obtain the original graph node topological features and the cluster group graph topological features, and fusing the original graph topological features and the cluster group graph topological features to output the current defect detection data of the orbital surface image specifically include:

[0034] Input the original graph convolution feature matrix and the cluster group convolution matrix into the persistent homology mechanism for topology respectively. The topology includes node-level topology and edge-level topology. The calculation formula is as follows:

[0035] ,

[0036] Among them, represents the nodes of the original graph convolution matrix or the cluster convolution matrix, is the multi-layer perceptron function of the node level for mapping, is for the node of the node-level topological feature matrix,

[0037] ,

[0038] Among them, is the hyperedge of the original graph convolution matrix or the cluster convolution matrix connected node representation, , are the two nodes connected by the hyperedge, is the multi-layer perceptron function of the edge level for mapping, is for the node of the node-level topological feature matrix, is the hyperedge of the edge-level topological feature matrix;

[0039] Input the data output by the node-level topology and the edge-level topology into the DeepSet network respectively for dimension compression and information aggregation, and obtain the original graph node topology features at the node level and the edge level, and the cluster graph topology features at the node level and the edge level. The calculation formula is as follows:

[0040] ,

[0041] Among them, is the original graph node topology feature or the cluster graph topology feature at the node level, is the original graph node topology feature or the cluster graph topology feature at the edge level;

[0042] Fuse the original graph node topology features at the node level and the edge level, and the cluster graph topology features at the node level and the edge level, and output the current defect detection data of the track surface image.

[0043] According to one aspect of the above technical solution, the steps of comparing the current defect detection data with the numerical value of the defect classification label, calculating the prediction error, and performing iterative optimization based on the prediction error to obtain the defect detection data specifically include:

[0044] Compare the current defect detection data with the numerical value of the defect classification label and calculate the prediction error;

[0045] Based on the prediction error, use The loss function calculates a loss value and determines whether the loss value meets a preset condition;

[0046] If so, update the iteration count, adjust the trainable parameters and the learnable matrix parameters according to the loss value, and return to the steps of extracting the image node features of the track surface image, constructing the original graph adjacency matrix and the original graph weight matrix;

[0047] If not, output the current defect detection data as the defect detection data.

[0048] According to one aspect of the above technical solution, based on the prediction error, using The steps of calculating the loss value by the loss function and determining whether the loss value meets the preset condition specifically include:

[0049] ,

[0050] wherein, is the iteration count, is the maximum iteration count, is the loss value of the iteration.

[0051] According to one aspect of the above technical solution, after the step of comparing the current defect detection data with the numerical value of the defect classification label and calculating the prediction error, it further includes:

[0052] Obtain the iteration state of the current defect detection data and determine whether the convergence condition is met;

[0053] If so, continue to execute the step of calculating the loss value by using the loss function based on the prediction error;

[0054] If not, return to the steps of extracting the image node features of the track surface image, constructing the original graph adjacency matrix and the original graph weight matrix.

[0055] Another aspect of the present invention lies in providing a track surface defect detection system for implementing the above track surface defect detection method, and the system includes:

[0056] An image acquisition module for acquiring a plurality of initial rail surface images, performing validity analysis on the initial rail surface images to obtain rail surface data, where the rail surface data includes a track surface image and the corresponding defect classification label;

[0057] A matrix construction module for extracting the image node features of the track surface image, constructing the original graph adjacency matrix and the original graph weight matrix, dividing the cluster groups and reconstructing the internal structure of the cluster groups, and constructing the cluster group adjacency matrix and the cluster group weight matrix;

[0058] A defect detection module, configured to input the original image adjacency matrix, the original image weight matrix, the cluster group adjacency matrix, and the cluster group weight matrix into a hypergraph convolutional model respectively, perform topology through a persistent homology mechanism to obtain the original image topology feature and the cluster group graph topology feature respectively, fuse the original image topology feature and the cluster group graph topology feature, and output the current defect detection data of the track surface image;

[0059] A detection optimization module, configured to compare the current defect detection data with the numerical value of the defect classification label, calculate a prediction error, and perform iterative optimization based on the prediction error to obtain defect detection data. Description of the Drawings

[0060] The above and / or additional aspects and advantages of the present invention will become obvious and easy to understand from the description of the embodiments in conjunction with the following drawings, where:

[0061] Figure 1 It is a schematic flowchart of the track surface defect detection method in the first embodiment of the present invention;

[0062] Figure 2 It is a structural block diagram of the track surface defect detection system in the second embodiment of the present invention;

[0063] Description of the symbols of the components in the drawings:

[0064] An image acquisition module 100, a matrix construction module 200, a defect detection module 300, a detection optimization module 400. Detailed Embodiments

[0065] To make the objectives, features, and advantages of the present invention more obvious and understandable, the following provides a detailed description of the specific embodiments of the present invention in conjunction with the drawings. Several embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

[0066] Embodiment 1

[0067] Please refer to Figure 1 , showing a track surface defect detection method provided by the first embodiment of the present invention, the method including steps S10 - step S13:

[0068] Step S10, obtain a plurality of initial rail surface images, perform validity analysis on the initial rail surface images to obtain rail surface data, where the rail surface data includes a track surface image and the corresponding defect classification label;

[0069] Specifically, in step S100, a number of initial rail surface images are obtained, and the integrity and quality of the initial rail surface images are verified to obtain rail surface images;

[0070] By way of example and not limitation, the integrity verification is to detect whether there are missing, damaged or abnormal conditions in the initial rail surface images, and the quality verification is to detect whether the initial rail surface images meet the set standards such as clarity, contrast, brightness, etc., so as to exclude low-quality data caused by factors such as acquisition equipment failures and environmental light interference. The initial rail surface images for which the integrity verification or quality verification fails are determined as invalid images, and these initial rail surface images are deleted.

[0071] In step S101, the rail surface images are subjected to defect detection and analysis, and defect classification labels are marked on the rail surface images to obtain rail surface data.

[0072] The defect classification labels include categories such as cracks, spalls, pits, corrugations, and no defects.

[0073] In step S11, the image node features of the rail surface images are extracted, an original graph adjacency matrix and an original graph weight matrix are constructed, clusters are divided, and the internal structure of the clusters is reconstructed to construct a cluster adjacency matrix and a cluster weight matrix;

[0074] Specifically, in step S110, the image node features of the rail surface images are extracted to construct an original graph adjoint matrix;

[0075] For example, the ResNet-50 model is used to extract the image node features of the rail surface images, and the convolutional layers of the ResNet-50 model are used to extract multi-scale features. The feature extraction process includes multiple residual blocks, and the input image is processed through 3×3 convolutional kernels and pooling operations to generate feature maps of different scales, and the original information is retained through skip connections. The features extracted by ResNet-50 have a dimension of 2048.

[0076] In step S111, based on the original graph adjoint matrix, hyperedges are constructed to obtain an original graph adjacency matrix and an original graph weight matrix;

[0077] It should be noted that for nodes of the same modality, if they are spatially adjacent, they are considered to be similar, and an original graph adjacency matrix is formed by constructing hyperedges to represent the adjacent relationship between nodes.

[0078] In addition, it should be noted that the cluster adjoint matrix, whose structure is a subset of the original graph adjoint matrix, and the weight of each hyperedge follows the original weight. Therefore, the calculation formula of the original graph weight matrix is the same as that of the cluster weight matrix.

[0079] Step S112: Divide the image node features into several clusters according to the connections between each hyperedge, and construct a cluster adjoint matrix.

[0080] Step S113: Based on the cluster adjoint matrix, screen out representative node features, reconstruct the hyperedges, and obtain a cluster adjacency matrix and a cluster weight matrix.

[0081] Specifically, in step S1130, the cluster adjoint matrix includes a weight matrix and a node feature matrix. Input the node feature matrix into the embedded Si-Net network for deep feature mapping. After being processed by the global feature extraction layer, obtain the globally feature matrix after non-linear projection.

[0082] By way of example rather than limitation, the representation of the cluster adjoint matrix includes: the original graph: the original graph adjoint matrix before clustering; the extended subgraph: successively includes the cluster adjoint matrix composed of 3 nodes in the original graph, the cluster adjoint matrix composed of 4 nodes, and so on until it is extended to the preset optimal node scale ( ) is the total number of nodes in the original graph, and the cluster adjoint matrix of the extended subgraph inherits the hyperedge connection relationship of the original graph.

[0083] Step S1131: Based on the globally feature matrix, calculate the attention weight matrix. According to the attention weight matrix and the globally feature matrix, calculate the weighted feature matrix. The calculation formula is as follows:

[0084] ,

[0085] Where is the weighted feature matrix, is the globally feature matrix, is the attention weight matrix, , are trainable parameters, is the activation function, is the rectified linear unit function;

[0086] Step S1132: Screen the weighted feature matrix, calculate the maximum feature difference degree, screen out representative node features, reconstruct the hyperedges, and obtain a cluster adjacency matrix;

[0087] Specifically, screen the weighted feature matrix, calculate the maximum feature difference degree, and screen out the node pair with the largest maximum feature difference degree as the representative node pair;

[0088] Among them, the representative node pair is expressed as , and the calculation formula is: , where , is a node pair in the weighted feature matrix.

[0089] Connect the representative node pairs, and reconstruct the hyperedges for the remaining nodes and their corresponding representative nodes respectively to obtain the cluster adjacency matrix;

[0090] Step S1133, perform distance-aware kernel function weighting on the weighted feature matrix to obtain the cluster weight matrix, and the calculation formula is as follows:

[0091] ,

[0092] where, is the cluster weight matrix, is the vector of node in the weighted feature matrix , is the vector of node in the weighted feature matrix , is and the distance between, is the th dimension element of node in the weighted feature matrix , is the th dimension element of node in the weighted feature matrix , is the total dimension of the weighted feature matrix , is the learnable matrix parameter.

[0093] Among them, the cluster weight matrix constructs the weighted coefficient of the adjacency edge through the difference degree between the vectors in the weighted feature matrix.

[0094] It should be noted that through the cluster partitioning technology, the image node features are locally partitioned, representative nodes are selected for each independent cluster, and a hyperedge relationship with adaptive weights is constructed. This method realizes the localization processing of feature information, effectively avoids the over-dependence of local features on global information, and thus significantly improves the expression effect of track defect features.

[0095] Step S12, input the original graph adjacency matrix, the original graph weight matrix, the cluster adjacency matrix, and the cluster weight matrix into the hypergraph convolution model respectively, and perform topology through the persistent homology mechanism to obtain the original graph topological feature and the cluster graph topological feature respectively. Then fuse the original graph topological feature and the cluster graph topological feature, and output the current defect detection data of the track surface image;

[0096] Specifically, in step S120, the original graph adjacency matrix, the original graph weight matrix, the cluster group adjacency matrix, and the cluster group weight matrix are respectively input into the hypergraph convolutional model to obtain the original graph convolutional matrix and the cluster group convolutional matrix. The calculation formula is as follows:

[0097] ,

[0098] where, is the original graph convolutional matrix or the cluster group convolutional matrix of the th layer, is the original graph convolutional matrix or the cluster group convolutional matrix of the th layer, is the hypergraph degree matrix, is the original graph adjacency matrix or the cluster group adjacency matrix, is 's transpose matrix, is the original graph weight matrix or the cluster group weight matrix, is the activation function;

[0099] It should be noted that the hypergraph convolutional model can achieve the feature aggregation and propagation of node information. Using the hypergraph convolutional model to embed and express the high-order relationship between the local information of the cluster group adjacency matrix and the global information of the original graph adjacency matrix can greatly improve the expression ability of local information and global information. As a high-order structure, the hypergraph convolutional model can effectively capture the complex multiple relationships between nodes. Compared with traditional models, the hypergraph convolutional model can more accurately represent the multiple relationships between nodes and edges, as well as the state where these relationships are intertwined and interact at a higher level. Therefore, using the hypergraph convolutional model for embedding can not only finely process local information, but also optimize the high-order dependence relationship between nodes in global information, thereby enhancing the expression ability and reasoning ability, and improving the accuracy and efficiency of track defect detection.

[0100] In step S121, the original graph convolutional feature matrix and the cluster group convolutional matrix are respectively input into the persistent homology mechanism for topology to obtain the original graph node topology feature and the cluster group graph topology feature respectively.

[0101] Specifically, in step S1210, the original graph convolutional feature matrix and the cluster group convolutional matrix are respectively input into the persistent homology mechanism for topology. The topology includes node-level topology and edge-level topology. The calculation formula is as follows:

[0102] ,

[0103] where, represents the node of the original graph convolutional matrix or the cluster group convolutional matrix, is the multi-layer perceptron function at the node level for mapping, is the node The node-level topological feature matrix of

[0104] ,

[0105] where, is the hyperedge of the original graph convolution matrix or the cluster group convolution matrix The nodes connected represent , are the two nodes connected by the hyperedge, is the multi-layer perceptron function at the edge level for mapping, is for node The node-level topological feature matrix of is the hyperedge The edge-level topological feature matrix of;

[0106] Step S1211, input the data output by the node-level topology and the edge-level topology into the DeepSet network respectively for dimension compression and information aggregation, to obtain the original graph node topology features and cluster group graph topology features at the node level and the edge level. The calculation formula is as follows:

[0107] ,

[0108] where, is the original graph node topology feature or cluster group graph topology feature at the node level, is the original graph node topology feature or cluster group graph topology feature at the edge level.

[0109] Step S122, fuse the original graph node topology features at the node level and the edge level and the cluster group graph topology features at the node level and the edge level, and output the current defect detection data of the track surface image.

[0110] It should be noted that introducing the Persistent Homology mechanism can effectively capture topological features, significantly improve the expression ability of local features and global features. This method can deeply mine topological information at different scales, enhance the sensitivity and recognition ability to subtle local defects. Through the multi-scale topological analysis of persistent homology, it can more accurately capture the long-term stable structural features, avoiding details and small changes that may be ignored by traditional clustering models. After the enhanced local feature topological features (cluster group convolution matrix) are fused with the global topological features (original graph topological features), the overall expressiveness and reasoning accuracy are further improved, optimizing its application effect in complex image structures, and further improving the accuracy and efficiency of track defect detection.

[0111] Furthermore, efficient data classification is achieved through the cooperation between local magnification (features extracted from the cluster adjacency matrix via the hypergraph convolutional model and persistent homology learning mechanism) and global embedding (features extracted from the original graph adjacency matrix via the hypergraph convolutional model and persistent homology learning mechanism), significantly enhancing the accuracy and robustness of data mining and information extraction. When dealing with high-dimensional or heterogeneous data, single global analysis or local feature extraction is difficult to effectively capture the internal structure and fine features of the data, often resulting in limited classification or recognition performance. The local magnification method proposed in the present invention can conduct a detailed analysis of local features with key differences in complex data, capturing subtle but crucial pattern differences. At the same time, the global embedding method captures the global structural information of the data from an overall perspective, ensuring that important overall features are not missed during the analysis process. The cooperation between the two methods enables in-depth understanding of the data at different scales from macro to micro.

[0112] For example, in the analysis of high-order network data, local magnification can reveal local information or the microscopic connection patterns between nodes, while global embedding reveals the overall topological structure features of the network; in the image classification task, local magnification can highlight important detail features, and global embedding provides global visual background information. This cooperation mechanism ensures more accurate and efficient classification analysis of high-order complex data, providing a more solid theoretical basis and application support for data-driven decision-making analysis.

[0113] Step S13: Compare the current defect detection data with the numerical value of the defect classification label, calculate the prediction error, and perform iterative optimization based on the prediction error to obtain the defect detection data.

[0114] Specifically, in step S130, compare the current defect detection data with the numerical value of the defect classification label and calculate the prediction error;

[0115] In step S131, based on the prediction error, use the loss function to calculate the loss value and determine whether the loss value meets the preset conditions;

[0116] ,

[0117] where is the number of iterations, is the maximum number of iterations, is the loss value of the

[0118] If so, update the number of iterations, adjust the trainable parameters and the learnable matrix parameters according to the loss value, and return to the step of extracting the image node features of the orbital surface image and constructing the original graph adjacency matrix and the original graph weight matrix;

[0119] Otherwise, output the current defect detection data as the defect detection data.

[0120] After the step of comparing the current defect detection data with the value of the defect classification label and calculating the prediction error, the method further includes:

[0121] Obtain the iteration state of the current defect detection data and determine whether the convergence condition is satisfied;

[0122] If so, continue to execute the step of calculating the loss value using the loss function based on the prediction error;

[0123] Otherwise, return to the step of extracting the image node features of the track surface image and constructing the original graph adjacency matrix and the original graph weight matrix.

[0124] Furthermore, traditional clustering-based models usually rely on local distance metrics or local structure-based assumptions to partition clusters, thus emphasizing local similarity. However, this model has limitations in dealing with complex data relationships and capturing high-order spatial dependencies.

[0125] In contrast, this embodiment utilizes the cluster + hypergraph convolutional model + persistent homology learning mechanism to optimize the spatial locality problem caused during modeling and enhance the information interaction ability between different clusters. It can not only retain the local clustering structure of the data but also adaptively adjust the hyperedge connection relationship through iterative optimization (adjusting the trainable parameters and the learnable matrix parameters according to the loss value) to more precisely capture the spatial hierarchical characteristics of the data, making it more suitable for application scenarios with strong spatial dependency relationships, such as complex system analysis, intelligent recommendation, and image recognition. Compared with traditional clustering-based models, the method of this embodiment can not only model more accurately, solve the spatial locality problem, but also effectively integrate global information to achieve richer high-order feature expressions, providing more refined and dynamic support for spatial data analysis and intelligent decision-making.

[0126] Compared with the prior art, by adopting the track surface defect detection method shown in this embodiment, the image node features of the track surface image are extracted, the original graph adjacency matrix and the original graph weight matrix are constructed to extract global information, the cluster is divided and the internal structure of the cluster is reconstructed, the cluster adjacency matrix and the cluster weight matrix are constructed to amplify and extract local information. The original graph adjacency matrix, the original graph weight matrix, the cluster adjacency matrix and the cluster weight matrix are respectively input into the hypergraph convolution model. The hypergraph convolution model can effectively express the high-order relationship between local information and global information, greatly improve the expression ability of local information and global information, effectively capture the complex multiple relationships between nodes, and improve the accuracy and efficiency of detection. Through the persistent homology mechanism for topology, the original graph topology feature and the cluster graph topology feature are respectively obtained, and the original graph topology feature and the cluster graph topology feature are fused to output the current defect detection data of the track surface image, effectively capture the topology features in local information and global information, deeply excavate the topology information at different scales in local information and global information, significantly improve the expression ability of local information and global information, and further improve the accuracy and efficiency of detection, thus solving the technical problems of low accuracy and efficiency in traditional manual inspection and non-destructive detection of rail surface defects in the prior art.

[0127] Embodiment 2

[0128] Please refer to Figure 2 , which shows a track surface defect detection system provided by the second embodiment of the present invention. The system includes:

[0129] An image acquisition module 100, configured to acquire a plurality of initial rail surface images, perform validity analysis on the initial rail surface images to obtain rail surface data, where the rail surface data includes a track surface image and a corresponding defect classification label;

[0130] A matrix construction module 200, configured to extract the image node features of the track surface image, construct an original graph adjacency matrix and an original graph weight matrix, divide the cluster and reconstruct the internal structure of the cluster, and construct a cluster adjacency matrix and a cluster weight matrix;

[0131] A defect detection module 300, configured to respectively input the original graph adjacency matrix, the original graph weight matrix, the cluster adjacency matrix and the cluster weight matrix into a hypergraph convolution model, perform topology through a persistent homology mechanism to respectively obtain an original graph topology feature and a cluster graph topology feature, fuse the original graph topology feature and the cluster graph topology feature, and output the current defect detection data of the track surface image;

[0132] The detection optimization module 400 is used to compare the current defect detection data with the numerical values of the defect classification labels, calculate the prediction error, and perform iterative optimization based on the prediction error to obtain the defect detection data.

[0133] Compared with the prior art, by using the rail surface defect detection system shown in this embodiment, the matrix construction module is used to extract global information and amplify and extract local information. The hypergraph convolution model of the defect detection module can effectively express the high-order relationship between local information and global information, greatly improving the expression ability of local information and global information, effectively capturing the complex multiple relationships between nodes, and improving the accuracy and efficiency of detection. And through the persistent homology mechanism of the defect detection module, the topological features in local information and global information are effectively captured, the topological information at different scales in local information and global information is deeply mined, the expression ability of local information and global information is significantly improved, and the accuracy and efficiency of detection are further improved, thus solving the technical problems of low accuracy and efficiency in the traditional manual inspection and non-destructive testing of rail surface defects in the prior art.

[0134] The technical features of the above various embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0135] Those skilled in the art can understand that the logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable storage medium for an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can obtain instructions

[0136] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0137] The above-described embodiments merely represent several implementation manners of the present invention. The description thereof is relatively specific and detailed, but it should not be construed as a limitation to the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the appended claims.

Claims

1. A method for detecting track surface defects, characterized in that: The method comprises: Acquire a number of initial rail surface images, perform validity analysis on the initial rail surface images, and obtain rail surface data, wherein the rail surface data includes the rail surface images and corresponding defect classification labels; Extracting image node features of the track surface image, constructing an original image adjacency matrix and an original image weight matrix, dividing clusters and reconstructing the internal structure of the clusters, and constructing a cluster adjacency matrix and a cluster weight matrix, including: Extract the image node features of the track surface image and construct the adjoint matrix of the original image. Based on the original image adjoint matrix, a hyperedge is constructed to obtain the original image adjacency matrix and the original image weight matrix. According to the connection between each hyperedge, the image node features are divided into several clusters, and the cluster adjoint matrix is ​​constructed. Based on the cluster adjoint matrix, representative node features are screened out, hyperedges are reconstructed, and a cluster adjacency matrix and a cluster weight matrix are obtained; The original image adjacency matrix and the original image weight matrix, as well as the cluster adjacency matrix and the cluster weight matrix are respectively input into the hypergraph convolution model, and topology is performed through the persistent homology mechanism to obtain the original image topology features and the cluster graph topology features respectively, and the original image topology features and the cluster graph topology features are fused to output the current defect detection data of the track surface image; The current defect detection data is compared with the value of the defect classification label, a prediction error is calculated, and iterative optimization is performed based on the prediction error to obtain defect detection data.

2. The rail surface defect detection method according to claim 1, characterized in that: The steps of acquiring a plurality of initial rail surface images, performing validity analysis on the initial rail surface images, and obtaining rail surface data, wherein the rail surface data includes the rail surface images and corresponding defect classification labels, specifically include: Acquire a number of initial rail surface images, perform integrity check and quality check on the initial rail surface images, and obtain rail surface images; The track surface image is subjected to defect detection and analysis, and a defect classification label is marked on the track surface image to obtain rail surface data.

3. The rail surface defect detection method according to claim 1, characterized in that: Based on the cluster adjoint matrix, representative node features are screened out, hyperedges are reconstructed, and the steps of obtaining a cluster adjacency matrix and a cluster weight matrix specifically include: The cluster adjoint matrix includes a weight matrix and a node feature matrix. The node feature matrix is ​​input into the embedded Si-Net network for deep feature mapping, and processed by the global feature extraction layer to obtain a global feature matrix after nonlinear projection. Based on the global feature matrix, the attention weight matrix is ​​calculated. According to the attention weight matrix and the global feature matrix, the weighted feature matrix is ​​calculated. The calculation formula is as follows: ， in, is the weighted feature matrix, is the global feature matrix, is the attention weight matrix, , is a trainable parameter, is the activation function, is a linear rectification function; The weighted feature matrix is ​​screened, the maximum feature difference is calculated, representative node features are screened, hyperedges are reconstructed, and a cluster adjacency matrix is ​​obtained; The weighted feature matrix is ​​weighted by the distance-aware kernel function to obtain a cluster weight matrix, and the calculation formula is as follows: ， in, is the cluster weight matrix, For Node In the weighted feature matrix The vector in For Node In the weighted feature matrix The vector in for and The distance between For Node In the weighted feature matrix Middle The element of dimension, For Node In the weighted feature matrix Middle The element of dimension, is the weighted feature matrix The total dimensions of is the learnable matrix parameter.

4. The rail surface defect detection method according to claim 3, characterized in that: The steps of inputting the original image adjacency matrix and the original image weight matrix, as well as the cluster adjacency matrix and the cluster weight matrix into the hypergraph convolution model respectively, and performing topology through the persistent homology mechanism to obtain the original image topology features and the cluster image topology features respectively, specifically include: The original image adjacency matrix and the original image weight matrix, as well as the cluster adjacency matrix and the cluster weight matrix are respectively input into the hypergraph convolution model to obtain the original image convolution matrix and the cluster convolution matrix. The calculation formula is as follows: ， in, For the The original image convolution matrix or cluster convolution matrix of the layer, For the The original image convolution matrix or cluster convolution matrix of the layer, is the hypergraph degree matrix, is the original graph adjacency matrix or cluster adjacency matrix, for The transposed matrix of is the original image weight matrix or cluster weight matrix, is the activation function; The original image convolution feature matrix and the cluster group convolution matrix are respectively input into the persistent homology mechanism for topology, and the original image node topology features and the cluster group image topology features are respectively obtained.

5. The rail surface defect detection method according to claim 4, characterized in that: The steps of inputting the original image convolution feature matrix and the cluster convolution matrix into the persistent homology mechanism for topology, respectively obtaining the original image node topology features and the cluster image topology features, respectively fusing the original image topology features and the cluster image topology features, and outputting the current defect detection data of the track surface image specifically include: The original image convolution feature matrix and the cluster convolution matrix are respectively input into the persistent homology mechanism for topology, and the topology includes node-level topology and edge-level topology. The calculation formula is as follows: ， in, A node representing the original image convolution matrix or the cluster convolution matrix, is the node-level multilayer perceptron function used for mapping, For Node The node-level topological feature matrix, ， in, is the hyperedge of the original image convolution matrix or the cluster convolution matrix The connected nodes represent , are two nodes connected by a hyperedge, is the multilayer perceptron function at the edge level used for mapping, For Node The node-level topological feature matrix, For super edge The edge-level topological feature matrix of ; The data output by the node-level topology and the edge-level topology are respectively input into the DeepSet network for dimension compression and information aggregation to obtain the original graph node topology features at the node level and edge level and the cluster graph topology features at the node level and edge level. The calculation formula is as follows: ， in, is the original graph node topology feature or cluster graph topology feature at the node level, It is the original graph node topology feature or cluster graph topology feature at the edge level; The original image node topology features at the node level and the edge level and the cluster image topology features at the node level and the edge level are fused to output the current defect detection data of the track surface image.

6. The rail surface defect detection method according to claim 5, characterized in that: The steps of comparing the current defect detection data with the value of the defect classification label, calculating the prediction error, and performing iterative optimization based on the prediction error to obtain the defect detection data specifically include: The original image topological features and the cluster image topological features are merged to output current defect detection data of the track surface image; Comparing the current defect detection data with the value of the defect classification label to calculate a prediction error; Based on the prediction error, The loss function calculates the loss value and determines whether the loss value meets the preset conditions; If yes, then update the number of iterations, adjust the trainable parameters and the learnable matrix parameters according to the loss value, return to the step of extracting the image node features of the track surface image, and constructing the original image adjacency matrix and the original image weight matrix; If not, the current defect detection data is output as defect detection data.

7. The rail surface defect detection method according to claim 6, characterized in that: Based on the prediction error, The loss function calculates the loss value and determines whether the loss value meets the preset conditions, specifically including: ， in, is the number of iterations, is the maximum number of iterations, For the The loss value for the iteration.

8. The rail surface defect detection method according to claim 6, characterized in that: After the step of comparing the current defect detection data with the value of the defect classification label and calculating the prediction error, the method further includes: Obtaining the iteration status of the current defect detection data to determine whether a convergence condition is met; If so, continue to execute based on the prediction error, using The steps of loss function to calculate loss value; If not, return to the step of extracting the image node features of the track surface image and constructing the original image adjacency matrix and the original image weight matrix.

9. A rail surface defect detection system, characterized in that: The system is used to implement the rail surface defect detection method according to any one of claims 1 to 8, and the system comprises: An image acquisition module is used to acquire a number of initial rail surface images, perform validity analysis on the initial rail surface images, and obtain rail surface data, wherein the rail surface data includes the rail surface images and corresponding defect classification labels; A matrix construction module is used to extract the image node features of the track surface image, construct an original image adjacency matrix and an original image weight matrix, divide the clusters and reconstruct the internal structure of the clusters, and construct a cluster adjacency matrix and a cluster weight matrix; A defect detection module is used to input the original image adjacency matrix and the original image weight matrix, as well as the cluster adjacency matrix and the cluster weight matrix into the hypergraph convolution model respectively, and perform topology through the persistent homology mechanism to obtain the original image topology features and the cluster graph topology features respectively, fuse the original image topology features and the cluster graph topology features, and output the current defect detection data of the track surface image; The detection optimization module is used to compare the current defect detection data with the value of the defect classification label, calculate the prediction error, perform iterative optimization based on the prediction error, and obtain the defect detection data.

Images