A road facility semantic topology lightweight graph neural network modeling method

Abstract

The application belongs to the technical field of electric digital data processing, and relates to a road facility semantic topology lightweight graph neural network modeling method. Road facility data is parsed into a topology coordinate sequence and a semantic feature sequence; a topology and semantic double-flow branch is constructed, a graph Laplacian matrix is constructed in the topology branch according to the coordinate sequence, a first-order Chebyshev polynomial without a nonlinear activation function is used for graph filtering to extract a topology representation vector; in the semantic branch, the coordinate information is stripped, one-dimensional convolution is performed on the semantic feature sequence to extract a semantic representation vector; the two are spliced to generate a fusion representation vector; a pre-trained large graph neural network with an isomorphic topology is used as a teacher model, a mean square error of a teacher vector and the fusion representation vector is calculated to construct a distillation loss function, and network weights are updated through back propagation to obtain a lightweight model. The application reduces the amount of nonlinear operation and floating point operation, blocks the back propagation of topology noise, and shortens the inference response time.

Description

Technical Field

[0001] This invention belongs to the field of electronic digital data processing technology, and relates to a lightweight graph neural network modeling method for semantic topology of road facilities. Background Technology

[0002] In the field of graph neural network modeling of road facilities, existing conventional solutions typically employ graph attention networks or graph convolutional networks to process road facility data. The implementation of such solutions involves: receiving the spatial coordinates and category attribute data of the road facilities; constructing a graph adjacency matrix based on a distance threshold between the spatial coordinates; and transforming the category attributes into an initial feature matrix. In a single hidden layer of the graph convolutional network, the topology represented by the graph adjacency matrix and the attribute features represented by the initial feature matrix are simultaneously introduced and multiplied. Immediately after the multiplication, a nonlinear activation function is applied for nonlinear mapping. By stacking multiple hidden layers containing the multiplication and nonlinear mapping, the feature representations of the nodes are iteratively updated, ultimately outputting a road facility feature vector that integrates topology and attributes.

[0003] The aforementioned existing technical solutions highly couple and aggregate spatial topological relationships and semantic attribute features in the same hidden layer of graph convolution. Since topological distributions and attribute distributions belong to different data feature spaces, existing techniques must stack multiple hidden layers and retain a large number of network parameters in order to fit both hybrid distributions simultaneously in a single network. This feature-coupled encoding method requires intensive nonlinear and floating-point operations during forward inference, resulting in redundant model parameters and excessively high inference latency, failing to meet the deployment requirements of edge computing devices for lightweight models and rapid response. Summary of the Invention

[0004] The purpose of this invention is to provide a lightweight graph neural network modeling method for semantic topology of road facilities, which can effectively solve the problems in the background art mentioned above.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A lightweight graph neural network modeling method for semantic topology of road facilities includes: receiving road facility data and parsing the road facility attributes into a sequence of topological location coordinates and a sequence of semantic attribute features;

[0007] A topological feature extraction branch and a semantic feature extraction branch are constructed. In the topological feature extraction branch, a graph Laplacian matrix is ​​constructed based solely on the topological position coordinate sequence, and a graph filtering operation is performed on the graph Laplacian matrix using a first-order Chebyshev polynomial without a nonlinear activation function to extract the topological structure representation vector.

[0008] In the semantic feature extraction branch, the topological location coordinate sequence is stripped, and only the semantic attribute feature sequence is subjected to a one-dimensional convolution operation to extract the semantic attribute representation vector;

[0009] The topological structure representation vector and the semantic attribute representation vector are concatenated to generate a fused representation vector.

[0010] During the model training phase, a pre-trained large graph neural network with isomorphic topological structure to the topological feature extraction branch and the semantic feature extraction branch is constructed as a teacher model. A distillation loss function is constructed by calculating the mean square error between the teacher vector output by the teacher model and the fused representation vector. The network weights of the topological feature extraction branch and the semantic feature extraction branch are updated by backpropagation using the distillation loss function to obtain a lightweight road facility semantic topology model.

[0011] Preferably, the step of receiving road facility data and parsing the road facility attributes into a topological location coordinate sequence and a semantic attribute feature sequence includes: obtaining the absolute latitude and longitude coordinates and facility category label of the road facility in physical space, mapping the absolute latitude and longitude coordinates to a preset discretized spatial grid, obtaining the row and column index values ​​of the discretized spatial grid, arranging the row and column index values ​​according to the connection order of the road facilities in the road network, and generating the topological location coordinate sequence;

[0012] The facility category label is converted into a one-hot encoded vector that matches the dimension of the preset category dictionary. Multiple one-hot encoded vectors corresponding to the same road facility are concatenated along the feature dimension to generate the semantic attribute feature sequence.

[0013] Preferably, the step of constructing the graph Laplacian matrix based solely on the topological location coordinate sequence includes: extracting the grid row and column index values ​​of each road facility node from the topological location coordinate sequence, calculating the Euclidean distance between any two road facility nodes, and for each road facility node, retaining the K nearest neighbor nodes with the smallest Euclidean distance, and constructing a K nearest neighbor undirected graph adjacency matrix.

[0014] An augmented adjacency matrix is ​​obtained by adding a self-loop matrix to the adjacency matrix of the K-nearest neighbor undirected graph. The degree matrix is ​​constructed by summing the diagonal elements of the augmented adjacency matrix. The symmetric normalized graph Laplacian matrix is ​​calculated based on the product of the inverse square root matrix of the degree matrix, the transpose of the inverse square root matrix of the degree matrix, and the augmented adjacency matrix.

[0015] Preferably, the step of performing graph filtering operation on the graph Laplacian matrix using a first-order Chebyshev polynomial without a nonlinear activation function includes: scaling the eigenvalues ​​in the graph Laplacian matrix to the interval -1 to 1 to obtain a scaled Laplacian matrix; replacing the first-order terms of the first-order Chebyshev polynomial with the difference between the scaled Laplacian matrix and the identity matrix; and replacing the zero-order terms of the first-order Chebyshev polynomial with the identity matrix.

[0016] The initial feature matrix of the road facility node is multiplied by the identity matrix corresponding to the zero-order term and the difference matrix corresponding to the first-order term, respectively. The results of the two multiplication operations are added together, and the sum is multiplied by the learnable weight parameter matrix. The topology representation vector is directly output by skipping the nonlinear activation function.

[0017] Preferably, performing one-dimensional convolution operation only on the semantic attribute feature sequence includes: inputting the semantic attribute feature sequence into a semantic extraction network composed of multiple one-dimensional dilated convolutional layers; in the one-dimensional dilated convolutional layers, performing one-dimensional convolution processing on the semantic attribute feature sequence in order of increasing dilation rate; wherein the kernel size of each one-dimensional dilated convolution operation remains fixed, and the receptive field of the semantic attribute feature sequence covered by a single convolutional kernel is expanded by increasing the dilation rate.

[0018] After each one-dimensional dilated convolutional operation, a batch normalization operation is performed. The output features after the batch normalization operation are used as the input of the next one-dimensional dilated convolutional layer, until the last one-dimensional dilated convolutional layer outputs the semantic attribute representation vector.

[0019] Preferably, the step of concatenating the topology representation vector and the semantic attribute representation vector includes: obtaining a first feature dimension corresponding to the topology representation vector and a second feature dimension corresponding to the semantic attribute representation vector; when the first feature dimension and the second feature dimension are inconsistent, inputting the topology representation vector into a first fully connected layer and inputting the semantic attribute representation vector into a second fully connected layer; and mapping the first feature dimension and the second feature dimension to a preset target fusion dimension through the first fully connected layer and the second fully connected layer.

[0020] The mapped topological structure representation vector and the mapped semantic attribute representation vector are concatenated end-to-end along the target fusion dimension to generate the fusion representation vector.

[0021] Preferably, retaining the top K neighbor nodes with the smallest Euclidean distance for each road facility node includes: for the current road facility node, calculating the set of Euclidean distances from the current road facility node to all other road facility nodes, and calculating the mean and standard deviation of the distances in the set of Euclidean distances;

[0022] The cutoff distance is obtained by subtracting the standard deviation of the distance by a preset multiple from the mean distance. The number of Euclidean distances in the Euclidean distance set that are less than the cutoff distance is counted. The number of Euclidean distances is assigned to parameter K in the first K neighbor nodes. When the number of Euclidean distances is zero, parameter K is set to a fixed constant.

[0023] Preferably, the step of skipping the nonlinear activation function and directly outputting the topological structure representation vector includes: after multiplying the summation result with the learnable weight parameter matrix, obtaining the intermediate topological vector obtained by multiplication, and performing an element-wise addition operation on the intermediate topological vector with the initial feature matrix before performing graph filtering operation to obtain the residual topological vector;

[0024] The residual topology vector is subjected to L2 norm normalization, and the residual topology vector after L2 norm normalization is used as the topology structure representation vector finally output by the topology feature extraction branch.

[0025] Preferably, the step of performing one-dimensional convolution processing on the semantic attribute feature sequence in ascending order of dilation rate includes: between two adjacent one-dimensional dilated convolutional layers, performing a pooling downsampling operation with a preset stride on the feature map output by the previous one-dimensional dilated convolutional layer, performing a bilinear interpolation upsampling operation on the downsampled feature map to restore it to the size before downsampling, performing an element-wise addition operation between the restored feature map and the feature map before downsampling, and using the result of the element-wise addition operation as the input feature map of the next one-dimensional dilated convolutional layer.

[0026] Preferably, the step of uniformly mapping the first feature dimension and the second feature dimension to a preset target fusion dimension through the first fully connected layer and the second fully connected layer includes: during the forward propagation process of the first fully connected layer and the second fully connected layer, extracting the first weight matrix of the first fully connected layer and the second weight matrix of the second fully connected layer, calculating the product matrix of the first weight matrix and the transpose of the first weight matrix, calculating the product matrix of the second weight matrix and the transpose of the second weight matrix, using the Frobenius norm difference between the product matrix and the identity matrix as an orthogonal penalty term, and superimposing the orthogonal penalty term into the distillation loss function composed of the mean square error.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] 1. This invention physically separates the topological position coordinate sequence and the semantic attribute feature sequence at the input end by constructing a topological feature extraction branch and a semantic feature extraction branch. In the topological feature extraction branch, only the topological position coordinate sequence is used to construct the graph Laplacian matrix, and a first-order Chebyshev polynomial without nonlinear activation functions is used for graph filtering to extract the topological structure representation vector. In the semantic feature extraction branch, the topological position coordinate sequence is stripped, and only the semantic attribute feature sequence is subjected to one-dimensional convolution operation to extract the semantic attribute representation vector. The two-stream decoupling architecture blocks the back propagation of topological noise to the semantic feature extraction process, reduces the number of nonlinear operations and floating-point operations in the model's forward inference process, reduces the number of model parameters, and shortens the inference response time. Combining a pre-trained large graph neural network with isomorphic topological structure as the teacher model, and using mean squared error to construct a distillation loss function to update the network weights through backpropagation, the model size is further compressed while maintaining the ability to represent complex semantic topological relationship features.

[0029] 2. When constructing the graph Laplacian matrix, absolute latitude and longitude coordinates are mapped to a discretized spatial grid to obtain row and column index values ​​to construct a coordinate sequence. Based on Euclidean distance, the first few neighboring nodes are retained to construct a nearest-neighbor undirected graph, and a symmetric normalized graph Laplacian matrix is ​​calculated. This is combined with scaling eigenvalues ​​to the interval -1 to 1 to replace Chebyshev polynomial terms, introducing residual connections and norm normalization, thus improving the numerical stability of the topological feature extraction process. In the semantic feature extraction branch, a one-dimensional dilated convolutional layer with increasing dilution rate is used to extract semantic features. A local feature fusion operation consisting of pooling downsampling and bilinear interpolation upsampling is introduced between adjacent layers, expanding the receptive field covered by a single convolutional kernel while preserving detailed features. Fully connected layers uniformly map features of different dimensions to the target fusion dimension. An orthogonal penalty term is introduced into the weight matrix calculation of the fully connected layer and superimposed on the distillation loss function, avoiding information redundancy during feature splicing and ensuring the feature decoupling independence of the two-stream branch in the knowledge distillation process. Attached Figure Description

[0030] Figure 1 This is the main flowchart of a lightweight graph neural network modeling method for semantic topology of road facilities according to the present invention;

[0031] Figure 2 This is a flowchart of the data parsing process of the present invention;

[0032] Figure 3 This is a flowchart illustrating the construction process of the graph Laplace matrix of this invention.

[0033] Figure 4 This is a flowchart of the topology branch graph filtering operation of the present invention;

[0034] Figure 5This is a flowchart of the semantic branch one-dimensional convolution operation of the present invention;

[0035] Figure 6 This is a flowchart of the feature splicing and model distillation training process of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0037] Please refer to Figure 1 This embodiment provides a lightweight graph neural network modeling method for semantic topology of road facilities. First, it receives road facility data, which originates from the full dataset of road ancillary facilities stored on an urban geographic information system platform. This data covers all road facilities corresponding to road network nodes within the urban built-up area, including but not limited to traffic lights, road signs, prohibitory signs, warning signs, streetlights, underground pipeline manholes, road monitoring equipment, bus stops, road barriers, emergency call equipment, and other road ancillary facilities with fixed physical locations. The received road facility data includes basic data fields such as unique identifier information, physical location information, facility attribute classification information, facility operation and maintenance status information, and facility specification parameters for each road facility. The received road facility data undergoes preprocessing operations, including missing value imputation, outlier removal, and duplicate data removal. Missing value imputation uses the attribute mean of facilities of the same category and road segment. Outlier removal uses an outlier detection method based on quartile intervals to remove abnormal data exceeding a reasonable value range. Duplicate data removal is performed based on dual verification of the facility's unique identifier information and physical location information to delete duplicate data.

[0038] The preprocessed road facility attributes are parsed into a topological location coordinate sequence and a semantic attribute feature sequence. Specifically, spatial location-related attribute fields are extracted from the basic data fields of each road facility, and these are arranged into a topological location coordinate sequence in a preset order. Attribute fields related to facility classification, specifications, and status are extracted from the basic data fields of each road facility, and these are arranged into a semantic attribute feature sequence matching the length of the topological location coordinate sequence. Each element in the topological location coordinate sequence corresponds one-to-one with each element in the semantic attribute feature sequence, both corresponding to the same road facility node. The topological location coordinate sequence contains only spatial location-related information of the road facility node, while the semantic attribute feature sequence contains only non-spatial attribute-related information of the road facility node. The two are physically separated at the data level, with no information overlap.

[0039] Next, a topology feature extraction branch and a semantic feature extraction branch are constructed. These two branches are independent parallel network structures, and there is no interaction of intermediate features between them during the forward propagation process. The feature vectors output are fused only after both branches have completed all feature extraction operations. The input to the topology feature extraction branch is only the topology location coordinate sequence obtained from the previous analysis, and the input to the semantic feature extraction branch is only the semantic attribute feature sequence obtained from the previous analysis. The input data of the two branches are completely isolated, preventing the cross-branch propagation of topology noise and semantic noise.

[0040] In the topology feature extraction branch, the graph Laplacian matrix is ​​constructed solely based on the topology location coordinate sequence. No data from the semantic attribute feature sequence is introduced during the construction process; the graph structure is built solely based on the spatial location information of the road facility nodes. Specifically, each element in the topology location coordinate sequence is treated as a node in the graph structure. The number of nodes is consistent with the length of the topology location coordinate sequence and also with the total number of road facilities. Based on the topology location coordinates corresponding to the nodes, the spatial distance between nodes is calculated. The connection relationships between nodes are determined based on the spatial distance, and an adjacency matrix of the undirected graph is constructed. The degree matrix is ​​calculated based on the adjacency matrix, and the graph Laplacian matrix is ​​constructed based on the adjacency matrix and the degree matrix. The graph Laplacian matrix has the same dimension as the number of nodes, being an N×N square matrix, where N is the total number of road facility nodes.

[0041] After constructing the graph Laplacian matrix, a graph filtering operation is performed on the matrix using a first-order Chebyshev polynomial without nonlinear activation functions to extract the topological structure representation vector. Specifically, the expansion of the first-order Chebyshev polynomial only includes zero-order and first-order terms, excluding higher-order polynomial terms, thus reducing the computational complexity of the graph filtering operation. Throughout the graph filtering process, no nonlinear activation functions are introduced, including but not limited to commonly used nonlinear activation functions such as ReLU, LeakyReLU, Sigmoid, Tanh, and GELU. The graph filtering operation is performed solely through linear matrix operations, thereby reducing the amount of nonlinear computation in the forward inference process and lowering computational latency. The input to the graph filtering operation is the initial feature matrix of the road facility nodes. This initial feature matrix is ​​generated based on the topological location coordinate sequence and only includes spatial location-related features of the nodes, without introducing semantic attribute-related features. The output of the graph filtering operation is the topological structure representation vector corresponding to each road facility node. The dimension of the topological structure representation vector is a preset fixed dimension, and the dimension of the topological structure representation vector of all nodes remains consistent.

[0042] In the semantic feature extraction branch, the topological location coordinate sequence is stripped, and only the semantic attribute feature sequence is subjected to one-dimensional convolution operations to extract semantic attribute representation vectors. Specifically, throughout the entire forward propagation of the semantic feature extraction branch, no data from the topological location coordinate sequence is used, nor is any graph structure-related data constructed based on the topological location coordinates introduced; feature extraction is performed solely based on the semantic attribute feature sequence. The semantic attribute feature sequence is a one-dimensional sequence of data, with a length consistent with the total number of road facility nodes. Each element in the sequence is the semantic attribute feature vector of the corresponding road facility node. One-dimensional convolution operations are used to extract features from the semantic attribute feature sequence at multiple scales. The convolution kernel of the one-dimensional convolution operation slides along the length dimension of the sequence to extract semantic attribute association features between different nodes. During the one-dimensional convolution operation, multiple one-dimensional convolutional layers can be stacked to extract higher-dimensional semantic attribute features layer by layer. The output of each one-dimensional convolutional layer serves as the input to the next one-dimensional convolutional layer, until the last one-dimensional convolutional layer outputs the semantic attribute representation vector corresponding to each road facility node. The semantic attribute represents the dimension of the vector as a preset fixed dimension, and the semantic attribute represents the vector dimension of all nodes are consistent.

[0043] After extracting the topology representation vector and semantic attribute representation vector, a concatenation operation is performed on these vectors to generate a fused representation vector. Specifically, for each road facility node, the corresponding topology representation vector and semantic attribute representation vector are obtained. These two vectors are then concatenated end-to-end along the feature dimension to generate the fused representation vector for that node. The fused representation vectors of all nodes are arranged in node order to form a fused representation matrix. The number of rows in the fused representation matrix is ​​the total number of road facility nodes, and the number of columns is the sum of the dimensions of the topology representation vector and the semantic attribute representation vector.

[0044] During the model training phase, a pre-trained large-scale graph neural network with an isomorphic topological structure to the topological feature extraction and semantic feature extraction branches is constructed as the teacher model. Specifically, the teacher model adopts a two-stream decoupled architecture completely consistent with the aforementioned lightweight model, also containing independent topological feature extraction and semantic feature extraction branches, with the input data isolation rules for the two branches remaining consistent with the lightweight model. The topological isomorphism between the teacher model and the lightweight model is specifically reflected in the following: the topological feature extraction branches of the teacher model and the lightweight model use the same graph filtering operation logic, and the semantic feature extraction branches of the teacher model and the lightweight model use the same one-dimensional convolution operation logic. The core difference between the two is that the topological feature extraction and semantic feature extraction branches of the teacher model both stack more network layers, setting a larger feature dimension, possessing a larger number of parameters and stronger feature representation capabilities. The teacher model is pre-trained on a large-scale cross-city road facility dataset. The pre-training process adopts a fully supervised training method, with the classification task of road facility nodes and the prediction task of road network topology connection as the pre-training tasks. After training, the teacher model converges to the optimal state and has the ability to represent the topological structure and semantic attributes of road facilities with high precision.

[0045] The training dataset is input into both the teacher model and the lightweight model. The training dataset contains multiple batches of road facility data, each batch parsed into a sequence of topological location coordinates and semantic attribute features. Based on the input training data, the teacher model outputs a teacher vector for each road facility node. This teacher vector is the high-dimensional feature vector that integrates topological and semantic features. The lightweight model, based on the same input data, outputs a fused representation vector for each road facility node. A distillation loss function is constructed by calculating the mean squared error between the teacher vectors output by the teacher model and the fused representation vectors output by the lightweight model. During the calculation of the distillation loss function, for each batch of training data, the mean squared error between the teacher vectors and fused representation vectors of all nodes within that batch is calculated and used as the distillation loss value for that batch.

[0046] Backpropagation is performed using the distillation loss function to update the network weights of the topological feature extraction branch and the semantic feature extraction branch. Specifically, a gradient descent algorithm is used to calculate the gradient of the distillation loss function relative to the learnable weight parameters of each network layer in the lightweight model. Based on the gradient value and a preset learning rate, the learnable weight parameters are iteratively updated. During backpropagation, the weight parameters of the teacher model remain fixed and do not participate in the update; only the teacher vector output by the teacher model is used as a soft label for distillation learning to guide the weight update of the lightweight model. After completing a preset number of iterations, training stops when the value of the distillation loss function converges to a preset threshold range, resulting in the trained lightweight road infrastructure semantic topology model.

[0047] This embodiment employs a two-stream decoupled network architecture to physically separate the topological and semantic attribute features of road facilities at the input end. These features are then extracted through independent branches, preventing noise cross-transmission between different feature spaces. This reduces the amount of nonlinear and floating-point computation during the model's forward inference, thereby lowering the model's parameter count and inference latency. Knowledge distillation is performed using a teacher model with a homogeneous topology, further compressing the model's size while maintaining its feature representation capabilities. This results in a lightweight model that is suitable for deployment on edge computing devices.

[0048] Table 1 Mapping Relationship Between Basic Road Infrastructure Data and Feature Sequences

[0049]

[0050] Table 1 illustrates the one-to-one correspondence between the original basic data of road facilities, which is parsed into a sequence of topological location coordinates and a sequence of semantic attribute features, thus clarifying the physical separation rules between topological and semantic information. Specifically, the elements of the topological location coordinate sequence only contain spatial location-related information of the road facilities, while the elements of the semantic attribute feature sequence only contain non-spatial attribute-related information of the road facilities. There is no information overlap between the two, providing independent input data for the two-stream decoupled feature extraction branches and ensuring the decoupling of the feature extraction process.

[0051] In a preferred embodiment, reference Figures 2 to 4The process of parsing road facility data is refined. The absolute latitude and longitude coordinates and facility category labels of the road facilities in physical space are obtained. These absolute latitude and longitude coordinates are then mapped to a pre-defined discretized spatial grid. The row and column index values ​​of the discretized spatial grid are obtained, and these index values ​​are arranged according to the connection order of the road facilities in the road network to generate a topological location coordinate sequence. Specifically, the pre-defined discretized spatial grid is a two-dimensional planar grid covering the built-up area of ​​the target city. The horizontal and vertical directions of the grid correspond to the longitude and latitude directions of the geographic coordinate system, respectively. The total width of the grid is W grid cells, and the total height is H grid cells, where W and H are positive integers and can be set according to the scope of the target area and the required modeling accuracy. The longitude range of the target city's built-up area is... Latitude range is ,in The minimum longitude value for the target area. The maximum longitude value of the target area. The minimum latitude value of the target area. This represents the maximum latitude value of the target area. For each road facility node, extract the corresponding absolute longitude value. with absolute latitude value The horizontal row and column index value x and the vertical row and column index value y of the node in the discretized spatial grid are calculated using the following mapping formula:

[0052]

[0053]

[0054] Where x takes values ​​in the range of 1 / 2. The range of values ​​for y is x and y together constitute the row and column index values ​​of the road facility node in the discretized spatial grid. The row and column index values ​​are two-dimensional integer vectors that only contain the discretized spatial location information and do not introduce any semantic attribute information.

[0055] After calculating the row and column indices of all road facility nodes, the row and column indices are arranged according to the connection order of the road facilities in the road network to generate a topological location coordinate sequence. Specifically, the connection order of road facilities in the road network is determined based on the topological connection relationship of the urban road network. A directed topological graph of the road network is constructed, using road intersections as nodes and road segments as edges. Following the depth-first traversal order of the directed topological graph, the row and column indices corresponding to the road facility nodes along the traversal path are extracted sequentially and arranged in traversal order to form the topological location coordinate sequence. The length of the topological location coordinate sequence is the same as the total number of road facility nodes N, and the i-th element in the sequence is the row and column index of the i-th traversed road facility node. The range of values ​​for i is 1. .

[0056] Simultaneously, facility category labels are converted into one-hot encoded vectors matching the dimensions of a pre-defined category dictionary. Multiple one-hot encoded vectors corresponding to the same road facility are concatenated along the feature dimension to generate a semantic attribute feature sequence. Specifically, the pre-defined category dictionary contains primary and secondary category labels for road facilities. Primary category labels include major categories such as traffic control facilities, lighting facilities, pipeline facilities, monitoring facilities, public transport service facilities, and road isolation facilities. Each primary category label corresponds to multiple secondary category labels; for example, the category of traffic control facilities corresponds to secondary category labels such as traffic lights, directional signs, prohibitory signs, and warning signs. The total dimension of the pre-defined category dictionary is C, where C is the sum of the number of primary and secondary category labels. Each category label corresponds to a unique index position in the category dictionary. For each road facility node, the primary and secondary category labels corresponding to that node are extracted, and both are converted into C-dimensional one-hot encoded vectors. In the one-hot encoded vector, only the element corresponding to the category label index position has a value of 1, while the elements at other positions have a value of 0. The first-level category one-hot encoded vector and the second-level category one-hot encoded vector corresponding to the same road facility node are concatenated along the feature dimension to generate a 2C-dimensional semantic attribute feature vector for that node. The semantic attribute feature vectors of all road facility nodes are arranged in the same node order as the topological location coordinate sequence to generate a semantic attribute feature sequence. The length of the semantic attribute feature sequence is the same as the length of the topological location coordinate sequence, both being N. Each element in the sequence corresponds to a semantic attribute feature vector of a road facility node, containing only the category attribute information of the facility and not introducing any spatial location-related information.

[0057] In the topology feature extraction branch, based on the aforementioned generated topology location coordinate sequence, a graph Laplacian matrix is ​​constructed solely based on this sequence. Specifically, the process involves extracting the grid row and column indices of each road facility node from the topology location coordinate sequence, calculating the Euclidean distance between any two road facility nodes, and for each road facility node, retaining the K nearest neighbors with the smallest Euclidean distance to construct a K-nearest neighbor undirected graph adjacency matrix. More specifically, for the i-th road facility node and the j-th road facility node, the row and column indices of the i-th node are extracted. The row and column index values ​​of the j-th node The Euclidean distance between two nodes is calculated using the following formula. :

[0058]

[0059] Where the values ​​of i and j are all within the range of , The value of is a non-negative real number, representing the spatial distance between two nodes in the discretized spatial grid.

[0060] For each road facility node, calculate the Euclidean distance from that node to all other N-1 road facility nodes, forming the Euclidean distance set corresponding to that node. For the current road facility node, calculate the mean distance of its corresponding Euclidean distance set. and distance standard deviation The cutoff distance is calculated using the following formula. :

[0061]

[0062] in, This is a preset multiple constant, taking the value of a positive real number, which can be adjusted according to the spatial density requirements of the modeling scene. It counts the Euclidean distances less than the cutoff distance in the set of distances. The number of Euclidean distances is calculated and assigned to the parameter K in the first K neighbor nodes. When the number of Euclidean distances is zero, the parameter K is set to a fixed constant. The fixed constant is a positive integer, usually 3 or 5, to ensure that each node has at least a fixed number of neighbor nodes and to avoid isolated nodes.

[0063] After determining the parameter K for each node, for each node, sort the remaining nodes in ascending order of Euclidean distance, retain the top K nodes as its neighbors, and construct the adjacency matrix A of the K-nearest neighbor undirected graph. The adjacency matrix A is... A square matrix of dimension i, where the elements in the i-th row and j-th column are... The rule for determining the value is: when the j-th node is one of the top K neighbor nodes of the i-th node, or when the i-th node is one of the top K neighbor nodes of the j-th node, The value of is 1; otherwise, The value of is 0. The adjacency matrix A is a symmetric matrix, satisfying . The connection relationships corresponding to an undirected graph, the diagonal elements of the adjacency matrix. The value is 0, and does not include self-loop connections.

[0064] The augmented adjacency matrix is ​​obtained by adding a self-loop matrix to the adjacency matrix A of the K-nearest neighbor undirected graph. The self-loop matrix is ​​the identity matrix I with the same dimension as the adjacency matrix A. The calculation formula is:

[0065]

[0066] Augmented adjacency matrix The diagonal elements are all 1s, and the remaining elements are consistent with the adjacency matrix A. Calculate the augmented adjacency matrix. The degree matrix D is constructed by summing the diagonal elements. A diagonal matrix of dimension i, where the diagonal elements in the i-th row and i-th column are... The value of is the augmented adjacency matrix. The sum of all elements in the i-th row is calculated using the following formula:

[0067]

[0068] The off-diagonal elements of the degree matrix D all have a value of 0.

[0069] The Laplacian matrix of the symmetric normalized graph is calculated by multiplying the inverse square root of the degree matrix, the transpose of the inverse square root of the degree matrix, and the augmented adjacency matrix. The calculation formula is:

[0070]

[0071] in, Let be the inverse square root matrix of the degree matrix D, and be a diagonal matrix whose diagonal elements in the i-th row and i-th column take values ​​of . I is An identity matrix of dimension 1. The Laplacian matrix of a symmetric normalized graph. It is a symmetric matrix with eigenvalues ​​ranging from [0,2], exhibiting good numerical stability and effectively reducing the risk of numerical overflow during graph filtering operations.

[0072] After constructing the symmetric normalized graphical Laplacian matrix, a graph filtering operation is performed on the matrix using a first-order Chebyshev polynomial without nonlinear activation functions. Specifically, the eigenvalues ​​in the graphical Laplacian matrix are scaled to the interval -1 to 1 to obtain the scaled Laplacian matrix. The difference between the scaled Laplacian matrix and the identity matrix is ​​used to replace the first-order terms of the first-order Chebyshev polynomial, and the identity matrix is ​​used to replace the zero-order terms of the first-order Chebyshev polynomial. In detail, the symmetric normalized graphical Laplacian matrix is ​​first extracted. Maximum eigenvalue The following formula is used to... Eigenvalue scaling yields the scaled Laplacian matrix. :

[0073]

[0074] Scaling Laplacian The eigenvalues ​​range from [-1, 1], which meets the input value requirements of Chebyshev polynomials.

[0075] The standard expansion of a first-order Chebyshev polynomial is:

[0076]

[0077] in, It is a zero-order Chebyshev polynomial. ; It is a first-order Chebyshev polynomial. ; and These are learnable polynomial coefficient parameters. In this embodiment, the zeroth-order term of the zeroth-order Chebyshev polynomial is replaced by the identity matrix I, and the scaling Laplace matrix is ​​used. Difference from the identity matrix I Replacing the first-order terms of the first-order Chebyshev polynomial, the simplified first-order Chebyshev polynomial graph filter kernel is:

[0078]

[0079] The initial feature matrix of the road facility node is multiplied by the identity matrix corresponding to the zero-order term and the difference matrix corresponding to the first-order term, respectively. The results of the two multiplications are added together, and the sum is multiplied by the learnable weight parameter matrix. This skips the nonlinear activation function and directly outputs the topological structure representation vector. Specifically, the initial feature matrix of the road facility node... for A matrix of dimensions, where N is the number of nodes, F is the initial feature dimension, and the initial feature matrix is... The initial feature vector for each node, obtained by transforming the topological position coordinate sequence, is the row and column index value of that node. The resulting two-dimensional vector, where F is 2, is used. The formula for graph filtering is:

[0080]

[0081] Where W is A learnable weight parameter matrix of dimension. The preset output dimension for topological structure representation vectors. for Intermediate topological vector of dimension and These are learnable polynomial coefficient parameters that are updated via backpropagation during model training.

[0082] After multiplying the sum by the learnable weight parameter matrix, the intermediate topological vector obtained from the multiplication is obtained. , the intermediate topological vector The initial feature matrix before performing graph filtering operations Perform element-wise addition to obtain the residual topological vector. The initial feature matrix is ​​used. Before performing the element-wise addition operation, the vector is mapped to the intermediate topology vector through a fully connected layer. The same feature dimension ensures dimension matching for element-wise addition operations. L2 norm normalization is performed on the residual topology vector, and the L2 norm-normalized residual topology vector is used as the final topology representation vector output by the topology feature extraction branch. The calculation formula is:

[0083]

[0084] in, for The residual mapping weight matrix of dimension, The L2 norm operator ensures that the L2 norm of the topological representation vector for each node is 1 after normalization, further improving the numerical stability of the feature vector. Throughout the topological feature extraction branch, no nonlinear activation functions are introduced; all operations are linear matrix operations, effectively reducing the model's computational complexity and inference latency.

[0085] This embodiment details the entire process of parsing road facility data, constructing the graph Laplacian matrix, and performing first-order Chebyshev polynomial graph filtering. It standardizes the topological location coordinates through discretized spatial grid mapping, constructs a K-nearest neighbor graph with dynamic K values ​​to adapt to the distribution of road facility nodes in different density regions, ensures the numerical stability and computational efficiency of topological feature extraction through symmetric normalized graph Laplacian matrix and inactive first-order Chebyshev filtering, and further enhances the representation ability and generalization performance of the topological structure representation vector through residual connectivity and L2 norm normalization.

[0086] Table 2. Parameters for constructing the K-nearest neighbor undirected graph and rules for generating the adjacency matrix.

[0087]

[0088] Table 2 illustrates the construction process and parameter calculation logic of the K-nearest neighbor undirected graph for road facility nodes, clarifying the method for determining the dynamic K value and the rules for generating the adjacency matrix. This table clearly presents the process by which different nodes adaptively adjust the number of neighbor nodes based on their spatial distribution characteristics, ensuring that the constructed graph structure can accurately represent the topological connections of road facilities in areas of different densities, thus providing a reliable graph structure foundation for subsequent graph filtering operations.

[0089] In another preferred embodiment, reference Figure 5The feature extraction process of the semantic feature extraction branch is refined. The semantic attribute feature sequence is input into a semantic extraction network composed of multiple one-dimensional dilated convolutional layers. In each one-dimensional dilated convolutional layer, the semantic attribute feature sequence is processed by one-dimensional convolution in order of increasing dilation rate. The kernel size of each one-dimensional dilated convolutional operation remains fixed. The receptive field of the semantic attribute feature sequence covered by a single convolutional kernel is expanded by increasing the dilation rate. Specifically, the semantic extraction network consists of M sequentially cascaded one-dimensional dilated convolutional layers, where M is a positive integer greater than or equal to 2 and can be set according to the accuracy requirements of semantic feature extraction. In this embodiment, M is set to 4. The semantic attribute feature sequence is... The input feature matrix has dimensions N, where N is the total number of road facility nodes, corresponding to the length dimension of the sequence. The dimension of the semantic attribute feature vector corresponding to each node corresponds to the feature dimension of the sequence. The length dimension of the input feature matrix is ​​independent of the feature dimension. The one-dimensional convolution operation only performs a sliding operation on the length dimension.

[0090] The core of one-dimensional dilated convolution is to introduce a fixed number of zero-valued fillers, or holes, between the elements of the kernel in a standard one-dimensional convolution. This expands the receptive field of the kernel without increasing the number of kernel parameters. For an input one-dimensional feature sequence x, the element at the nth position in the output feature sequence of the one-dimensional dilated convolution operation is calculated using the following formula:

[0091]

[0092] Where K is the size of the convolution kernel, that is, the number of weight elements contained in the convolution kernel. The size K of the convolution kernel in all one-dimensional dilated convolutional layers remains fixed. In this embodiment, the value of K is 3. r is the dilation rate, which represents the number of zero-value padding between the convolution kernel elements. The dilation rate r is a positive integer. When r=1, the one-dimensional dilated convolution degenerates into a standard one-dimensional convolution. is the k-th weight element of the convolution kernel, a learnable parameter updated via backpropagation during model training; n is the position index of the output feature sequence, with a value ranging from 1 to 2. .

[0093] In M cascaded one-dimensional dilated convolutional layers, the dilation rate of each layer is set in ascending order. In this embodiment, the dilation rates of the four one-dimensional dilated convolutional layers are set to 1, 2, 4, and 8 respectively, increasing exponentially. With the kernel size K fixed at 3, the receptive field size of the first one-dimensional dilated convolutional layer with a dilation rate of 1 is 3, the second one with a dilation rate of 2 is 5, the third one with a dilation rate of 4 is 9, and the fourth one with a dilation rate of 8 is 17. Through this ascending dilation rate setting, as the network depth increases, the length range of semantic attribute feature sequences that a single convolutional kernel can cover continuously expands. Without increasing the number of convolutional kernel parameters or computational complexity, this achieves effective extraction of long-distance semantic association features, while avoiding the loss of sequence detail features caused by pooling operations.

[0094] After each one-dimensional dilated convolutional operation, batch normalization is performed. The output features after batch normalization are used as the input for the next one-dimensional dilated convolutional layer, until the last one-dimensional dilated convolutional layer outputs a semantic attribute representation vector. Specifically, the batch normalization operation standardizes the feature values ​​along the feature dimension for each batch of input features, adjusting the mean of the feature values ​​to 0 and the variance to 1. This alleviates the internal covariate shift problem during network training and accelerates the network's convergence speed. The feature matrix Z output by a certain one-dimensional dilated convolutional operation is given by Z, where Z is... The tensor has dimensions B, where B is the size of the training batch, N is the sequence length, and D is the feature dimension. The formula for calculating batch normalization is:

[0095]

[0096] in, Let Z be the mean of the feature matrix Z in this batch along the feature dimension. Let Z be the variance of the feature matrix Z in the feature dimension of this batch. This is a local constant used to avoid cases where the denominator is zero; it is typically set to a value of [value missing]. , This is the standardized feature matrix after batch normalization. The batch normalization operation also includes two learnable parameters: a scaling parameter γ and a translation parameter β. A linear transformation is performed on the standardized feature matrix to obtain the final batch-normalized output feature matrix. The scaling parameter γ and the translation parameter β are updated through backpropagation during model training, ensuring that the network can autonomously adjust the distribution range of features.

[0097] Between two adjacent one-dimensional dilated convolutional layers, a pooling downsampling operation with a preset stride is performed on the feature map output by the previous one-dimensional dilated convolutional layer. The downsampled feature map is then subjected to bilinear interpolation upsampling to restore it to its original size. The restored feature map is then element-wise added to the original feature map, and the result of this element-wise addition is used as the input feature map for the next one-dimensional dilated convolutional layer. Specifically, the preset pooling downsampling stride is 2, max pooling is used, and pooling is performed along the length dimension of the sequence. The size of the pooling window is consistent with the stride, both being 2. This process is applied to the feature map output by the previous one-dimensional dilated convolutional layer. Its dimensions are Where N is the sequence length, D is the feature dimension, and the output feature map of the pooling downsampling operation. The dimension is The sequence length is reduced to half of its original length, while the feature dimension remains unchanged.

[0098] downsampled feature map Perform bilinear interpolation upsampling with a scaling factor of 2 to restore the sequence length of the feature map to N before downsampling, resulting in the upsampled feature map. Its dimension is the same as the feature map before downsampling. Completely identical, for Bilinear interpolation upsampling generates upsampled feature values ​​by performing a weighted average of adjacent feature values. This method can preserve the detailed information of the feature map to the greatest extent and avoid feature distortion during the upsampling process.

[0099] Feature map after restoration Feature map before downsampling Perform element-wise addition to obtain the fused feature map. The calculation formula is:

[0100]

[0101] This operation constitutes a local residual feature fusion structure. It extracts global semantic features from the feature map through pooling downsampling, restores the detailed information of the feature map through bilinear interpolation upsampling, and then fuses the global semantic features with the original detailed features through residual addition. This not only expands the receptive field of feature extraction but also effectively preserves the detailed features of the semantic attribute feature sequence, avoids the loss of detailed information caused by downsampling, and improves the extraction accuracy of semantic attribute features.

[0102] Feature map after local residual feature fusion operation The feature map, used as input to the next 1D dilated convolutional layer, is then used to perform the 1D dilated convolution operation. This local residual feature fusion structure is set between every two adjacent 1D dilated convolutional layers until the last 1D dilated convolutional layer completes the operation, outputting the final semantic attribute representation vector. , for A matrix of dimensions, where N is the total number of road facility nodes. The preset output dimension is used for the semantic attribute representation vector, and the dimension of the semantic attribute representation vector for each node remains consistent.

[0103] This embodiment refines the entire process of one-dimensional dilated convolution feature extraction in the semantic feature extraction branch. By using one-dimensional dilated convolutional layers with increasing dilation rates, the receptive field of feature extraction is expanded without increasing the number of parameters, enabling effective extraction of long-distance semantic association features. Batch normalization after each convolutional layer alleviates the internal covariate shift problem during network training and accelerates the network's convergence speed. The local residual feature fusion structure composed of pooling downsampling and bilinear interpolation upsampling between adjacent layers effectively preserves the detailed features of the sequence while extracting global semantic features, thus improving the representation accuracy of semantic attribute representation vectors.

[0104] Table 3. Parameter Configuration Table for One-Dimensional Dilated Convolutional Semantic Extraction Network Layer

[0105]

[0106] Table 3 illustrates the parameter configurations and data flow logic of each layer in the one-dimensional dilated convolutional semantic extraction network within the semantic feature extraction branch. It clarifies the method for increasing the dilation rate, the expansion pattern of the receptive field, and the execution location of the local residual feature fusion structure. This table clearly presents the hierarchical structure and parameter setting rules of the semantic feature extraction network, ensuring the feasibility and accuracy of the semantic attribute feature extraction process.

[0107] In yet another preferred embodiment, reference is made to Figure 6 The process of concatenating and fusing topological structure representation vectors and semantic attribute representation vectors, as well as the construction process of the distillation loss function during model training, are refined. The first feature dimension corresponding to the topological structure representation vector and the second feature dimension corresponding to the semantic attribute representation vector are obtained. When the first and second feature dimensions are inconsistent, the topological structure representation vector is input into the first fully connected layer, and the semantic attribute representation vector is input into the second fully connected layer. The first and second fully connected layers then map the first and second feature dimensions to a preset target fusion dimension. Specifically, the topological structure representation vector output by the topological feature extraction branch... for A matrix of dimensions, where N is the total number of road facility nodes. The first feature dimension of the topological structure representation vector; the semantic attribute representation vector output by the semantic feature extraction branch. for A matrix of dimension This is the second feature dimension of the semantic attribute representation vector. The preset target fusion dimension is... , which is a positive integer and can be set according to the accuracy requirements of feature fusion.

[0108] When the first feature dimension With the second feature dimension When inconsistencies exist, the two feature vectors are mapped to their dimensions using two independent fully connected layers. The first fully connected layer corresponds to the topological structure representation vector, and the second fully connected layer corresponds to the semantic attribute representation vector. There is no parameter sharing between the two fully connected layers, ensuring that the feature mapping processes of the two branches are independent and maintaining the decoupling of the two-stream branches. The forward propagation calculation formula for the first fully connected layer is:

[0109]

[0110] in, for The first weight matrix of dimension is the learnable parameters of the first fully connected layer; for The first bias vector of dimension is the learnable parameter of the first fully connected layer; for The topological structure representation vector after dimensional mapping has its feature dimensions unified as the target fusion dimension. .

[0111] The formula for calculating the forward propagation of the second fully connected layer is:

[0112]

[0113] in, for The second weight matrix of dimension is the learnable parameters of the second fully connected layer; for The second bias vector of dimension is the learnable parameter of the second fully connected layer; for The semantic attribute representation vector after dimensional mapping has its feature dimension unified as the target fusion dimension. .

[0114] When the first feature dimension With the second feature dimension Consistent, and all equal to the target fusion dimension At this time, there is no need to perform dimension mapping operations; the topological structure representation vector and semantic attribute representation vector can be directly used for subsequent concatenation operations.

[0115] The mapped topology representation vector and the mapped semantic attribute representation vector are concatenated end-to-end along the target fusion dimension to generate a fused representation vector. Specifically, for each road facility node, the mapped topology representation vector and the mapped semantic attribute representation vector corresponding to that node are concatenated end-to-end along the feature dimension, and the dimension of the concatenated feature vector is... The concatenated eigenvectors of all nodes are arranged in node order to form a fusion representation matrix. Its dimensions are The calculation formula is:

[0116]

[0117] in, This represents a concatenation operation on the feature dimension. The concatenation operation is only performed after all feature extraction and dimension mapping are completed in both branches. During the forward propagation process, there is no feature interaction between the two branches, which ensures the decoupling of the two-stream branches.

[0118] During the model training phase, a pre-trained large graph neural network with a topological structure isomorphic to the topological feature extraction and semantic feature extraction branches is constructed as the teacher model. The weight parameters of the teacher model remain fixed during training, and it only outputs teacher vectors as soft labels for distillation learning. Road facility data from the training batch is simultaneously input into both the teacher model and the lightweight model. The teacher model outputs teacher vectors corresponding to all nodes in that batch. , for A dimensional matrix, and a fused representation matrix of the lightweight model output. The dimensions are completely consistent. A distillation loss function is constructed by calculating the mean squared error between the teacher vector output by the teacher model and the fused representation vector. The calculation formula is:

[0119]

[0120] in, To fuse the eigenvalues ​​of the i-th node in the d-th dimension of the matrix, Let be the eigenvalue of the d-th dimension of the i-th node in the teacher vector matrix, and N be the total number of road facility nodes in the batch. The feature dimension of the fused representation vector is defined by the distillation loss function, which characterizes the difference between the fused representation vector output by the lightweight model and the teacher vector output by the teacher model. The core objective of the training process is to minimize this distillation loss function, enabling the lightweight model to learn the strong feature representation capabilities of the teacher model.

[0121] During the forward propagation between the first and second fully connected layers, the first weight matrix of the first fully connected layer and the second weight matrix of the second fully connected layer are extracted. The product matrix of the first weight matrix and its transpose is calculated, as is the product matrix of the second weight matrix and its transpose. The Frobenius norm difference between the product matrix and the identity matrix is ​​used as an orthogonality penalty term, which is then added to the distillation loss function based on the mean squared error. Specifically, the orthogonality penalty term aims to constrain the first and second weight matrices to be orthogonal, ensuring that the feature mapping spaces of the two fully connected layers are mutually orthogonal. This avoids information redundancy between topological and semantic features during dimensionality mapping, guaranteeing the feature decoupling and independence of the two branches. The calculation formula is:

[0122]

[0123] in, The first weight matrix The transpose of the matrix, The second weight matrix The transpose of the matrix, for An identity matrix of dimension 1 This is the Frobenius norm operator. The Frobenius norm is the square root of the sum of the squares of all elements in a matrix. When the weight matrix is ​​an orthogonal matrix, the product of the weight matrix and its transpose is the identity matrix. In this case, the orthogonality penalty term is 0. The larger the value of the orthogonality penalty term, the greater the degree to which the weight matrix deviates from the orthogonal matrix.

[0124] The orthogonal penalty term is added to the distillation loss function to construct the total loss function for model training. The calculation formula is:

[0125]

[0126] Where β is the weight coefficient of the orthogonal penalty term, which is a non-negative real number and can be adjusted according to the need for feature decoupling during training. When β is 0, the total loss function degenerates into a pure distillation loss function.

[0127] Backpropagation is performed using the total loss function to update all learnable weight parameters of the topological feature extraction branch, the semantic feature extraction branch, and the first and second fully connected layers. Specifically, gradient descent updates are performed using the Adam optimizer. The exponential decay rate of the first moment estimate of the Adam optimizer is set to 0.9, the exponential decay rate of the second moment estimate is set to 0.999, and the numerical stability constant is set to... In each training batch, a forward propagation operation is first performed to calculate the total loss function value for that batch; then a backpropagation operation is performed to calculate the gradient of the total loss function with respect to all learnable weight parameters; finally, based on the update rules of the Adam optimizer and combined with a preset learning rate, all learnable weight parameters are iteratively updated. All parameters of the teacher model remain fixed throughout the training process and do not participate in gradient calculation or parameter updates.

[0128] During training, a validation dataset is set up to verify the model's generalization performance. The validation dataset consists of road facility data that does not overlap with the training dataset. After a fixed number of training iterations, the validation dataset is input into the currently trained lightweight model, and the distillation loss function value on the validation set is calculated. When the distillation loss function value on the validation set no longer decreases for several consecutive iterations, an early stopping mechanism is triggered to stop training and prevent overfitting. After training is complete, all learnable weight parameters of the lightweight model are saved, resulting in the final lightweight road facility semantic topology model.

[0129] This embodiment details the entire process of feature concatenation and fusion, as well as the loss function construction process during model training. A fully connected layer enables unified mapping of features across different dimensions, ensuring dimensionality matching in the feature concatenation operation. By introducing an orthogonal penalty term into the weight matrix of the fully connected layer, the feature mapping spaces of the two branches are constrained to be orthogonal, avoiding information redundancy during feature fusion and ensuring the decoupling independence of the two-stream branches. By superimposing the orthogonal penalty term onto the distillation loss function to construct the total loss function, synergistic optimization of feature representation capability and feature decoupling is achieved. This allows the trained lightweight model to maintain a low parameter count and inference latency while possessing high-precision feature representation capabilities.

[0130] Table 4. Hyperparameter and loss function configuration during model training.

[0131]

[0132] Table 4 illustrates the configuration rules and physical meanings of each hyperparameter during the model training phase, clarifying the collaborative optimization configuration method of the distillation loss function and the orthogonal penalty term. This table clearly presents the hyperparameter setting and optimization logic during model training, ensuring the feasibility and convergence of the model training process, and enabling the trained model to achieve the expected lightweight effect and feature representation accuracy.

Claims

1. A lightweight graph neural network modeling method for semantic topology of road facilities, characterized in that, include: Receive road facility data and parse the road facility attributes into a sequence of topological location coordinates and a sequence of semantic attribute features; A topological feature extraction branch and a semantic feature extraction branch are constructed. In the topological feature extraction branch, a graph Laplacian matrix is ​​constructed based solely on the topological position coordinate sequence, and a graph filtering operation is performed on the graph Laplacian matrix using a first-order Chebyshev polynomial without a nonlinear activation function to extract the topological structure representation vector. In the semantic feature extraction branch, the topological location coordinate sequence is stripped, and only the semantic attribute feature sequence is subjected to a one-dimensional convolution operation to extract the semantic attribute representation vector; The topological structure representation vector and the semantic attribute representation vector are concatenated to generate a fused representation vector. During the model training phase, a pre-trained large graph neural network with isomorphic topological structure to the topological feature extraction branch and the semantic feature extraction branch is constructed as a teacher model. A distillation loss function is constructed by calculating the mean square error between the teacher vector output by the teacher model and the fused representation vector. The network weights of the topological feature extraction branch and the semantic feature extraction branch are updated by backpropagation using the distillation loss function to obtain a lightweight road facility semantic topology model.

2. The lightweight graph neural network modeling method for semantic topology of road facilities according to claim 1, characterized in that, The received road facility data is parsed into a sequence of topological location coordinates and a sequence of semantic attribute features, including: Obtain the absolute latitude and longitude coordinates and facility category labels of road facilities in physical space, map the absolute latitude and longitude coordinates to a preset discretized spatial grid, obtain the row and column index values ​​of the discretized spatial grid, arrange the row and column index values ​​according to the connection order of road facilities in the road network, and generate the topological location coordinate sequence. The facility category label is converted into a one-hot encoded vector that matches the dimension of the preset category dictionary. Multiple one-hot encoded vectors corresponding to the same road facility are concatenated along the feature dimension to generate the semantic attribute feature sequence.

3. The lightweight graph neural network modeling method for semantic topology of road facilities according to claim 1, characterized in that, The construction of the graph Laplacian matrix based solely on the topological position coordinate sequence includes: Extract the grid row and column index values ​​of each road facility node from the topological location coordinate sequence, calculate the Euclidean distance between any two road facility nodes, and for each road facility node, retain the K nearest neighbor nodes with the smallest Euclidean distance to construct the K nearest neighbor undirected graph adjacency matrix. An augmented adjacency matrix is ​​obtained by adding a self-loop matrix to the adjacency matrix of the K-nearest neighbor undirected graph. The degree matrix is ​​constructed by summing the diagonal elements of the augmented adjacency matrix. The symmetric normalized graph Laplacian matrix is ​​calculated based on the product of the inverse square root matrix of the degree matrix, the transpose of the inverse square root matrix of the degree matrix, and the augmented adjacency matrix.

4. The lightweight graph neural network modeling method for semantic topology of road facilities according to claim 1, characterized in that, The graphical filtering operation on the graphical Laplacian matrix using a first-order Chebyshev polynomial without a nonlinear activation function includes: The eigenvalues ​​in the graph Laplacian matrix are scaled to the interval between -1 and 1 to obtain the scaled Laplacian matrix. The difference between the scaled Laplacian matrix and the identity matrix is ​​used to replace the first-order terms of the first-order Chebyshev polynomial. The identity matrix is ​​used to replace the zero-order terms of the first-order Chebyshev polynomial. The initial feature matrix of the road facility node is multiplied by the identity matrix corresponding to the zero-order term and the difference matrix corresponding to the first-order term, respectively. The results of the two multiplication operations are added together, and the sum is multiplied by the learnable weight parameter matrix. The topology representation vector is directly output by skipping the nonlinear activation function.

5. The lightweight graph neural network modeling method for semantic topology of road facilities according to claim 1, characterized in that, The step of performing a one-dimensional convolution operation only on the semantic attribute feature sequence includes: The semantic attribute feature sequence is input into a semantic extraction network composed of multiple one-dimensional dilated convolutional layers. In the one-dimensional dilated convolutional layers, the semantic attribute feature sequence is processed by one-dimensional convolution in order of increasing dilation rate. The kernel size of each one-dimensional dilated convolutional operation is kept fixed. The receptive field of the semantic attribute feature sequence covered by a single convolutional kernel is expanded by increasing the dilation rate. After each one-dimensional dilated convolutional operation, a batch normalization operation is performed. The output features after the batch normalization operation are used as the input of the next one-dimensional dilated convolutional layer, until the last one-dimensional dilated convolutional layer outputs the semantic attribute representation vector.

6. The lightweight graph neural network modeling method for semantic topology of road facilities according to claim 1, characterized in that, The step of concatenating the topological structure representation vector with the semantic attribute representation vector includes: Obtain the first feature dimension corresponding to the topology representation vector and the second feature dimension corresponding to the semantic attribute representation vector. When the first feature dimension and the second feature dimension are inconsistent, input the topology representation vector into the first fully connected layer and input the semantic attribute representation vector into the second fully connected layer. Through the first fully connected layer and the second fully connected layer, map the first feature dimension and the second feature dimension to the preset target fusion dimension. The mapped topological structure representation vector and the mapped semantic attribute representation vector are concatenated end-to-end along the target fusion dimension to generate the fusion representation vector.

7. The lightweight graph neural network modeling method for semantic topology of road facilities according to claim 3, characterized in that, For each road facility node, the K neighbor nodes with the smallest Euclidean distance are retained, including: For the current road facility node, calculate the set of Euclidean distances from the current road facility node to all other road facility nodes, and calculate the mean and standard deviation of the distances in the set of Euclidean distances. The cutoff distance is obtained by subtracting the standard deviation of the distance by a preset multiple from the mean distance. The number of Euclidean distances in the Euclidean distance set that are less than the cutoff distance is counted. The number of Euclidean distances is assigned to parameter K in the first K neighbor nodes. When the number of Euclidean distances is zero, parameter K is set to a fixed constant.

8. The lightweight graph neural network modeling method for semantic topology of road facilities according to claim 4, characterized in that, The step of skipping the nonlinear activation function and directly outputting the topological structure representation vector includes: After multiplying the summation result with the learnable weight parameter matrix, the intermediate topological vector obtained by multiplication is obtained. The intermediate topological vector is then added element-wise to the initial feature matrix before the graph filtering operation to obtain the residual topological vector. The residual topology vector is subjected to L2 norm normalization, and the residual topology vector after L2 norm normalization is used as the topology structure representation vector finally output by the topology feature extraction branch.

9. A lightweight graph neural network modeling method for semantic topology of road facilities according to claim 5, characterized in that, The step of performing one-dimensional convolution processing on the semantic attribute feature sequence in ascending order of hole rate includes: Between two adjacent one-dimensional dilated convolutional layers, a pooling downsampling operation with a preset stride is performed on the feature map output by the previous one-dimensional dilated convolutional layer. A bilinear interpolation upsampling operation is then performed on the downsampled feature map to restore it to the size before downsampling. The restored feature map is then added element-wise to the feature map before downsampling. The result of the element-wise addition operation is used as the input feature map for the next one-dimensional dilated convolutional layer.

10. A lightweight graph neural network modeling method for semantic topology of road facilities according to claim 6, characterized in that, The step of mapping the first feature dimension and the second feature dimension to a preset target fusion dimension through the first fully connected layer and the second fully connected layer includes: During the forward propagation process between the first fully connected layer and the second fully connected layer, the first weight matrix of the first fully connected layer and the second weight matrix of the second fully connected layer are extracted. The product matrix of the first weight matrix and the transpose of the first weight matrix is ​​calculated, and the product matrix of the second weight matrix and the transpose of the second weight matrix is ​​calculated. The Frobenius norm difference between the product matrix and the identity matrix is ​​used as an orthogonal penalty term, and the orthogonal penalty term is superimposed on the distillation loss function composed of the mean square error.

Images