Key road section identification method based on road network dynamic feature sampling aggregation

Abstract

The invention provides a key road section identification method based on road network dynamic feature sampling and aggregation. The method comprises the following steps: selecting a traffic database X, and preprocessing the X by adopting a network workx library and a pandas library to obtain a traffic network line graph G containing traffic features; a key road recognition model M is constructed, then G is input into the M, # imgabs0 # imgabs1 # is obtained through a data preprocessing module, # imgabs2 # imgabs3 # is obtained through a GraphDFSAGE module, and a final key road list is obtained through a key list output module. Compared with the prior art, the method has higher accuracy in identifying the key road in an unknown area, and fewer parameters are used, so that a plurality of traffic networks with less data volume can obtain more accurate key road recommendation results, and better reference suggestions are provided for planning and utilization of traffic resources.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a key road section identification method based on road network dynamic feature sampling and aggregation. Background Art

[0002] During the traffic control process, road conditions often change rapidly due to emergencies. Therefore, it is more important to find the key sections in the road network more quickly. How to reduce the complexity of extracting key sections, respond to the requirements of obtaining key sections more quickly, and locate the key sections of the road network more quickly and accurately in the actual traffic control of Chongqing West Railway Station is an urgent problem to be solved.

[0003] The Path2vec model, a key segment identification model based on dynamic traffic flow direction walk sampling, has a core concept that improves on the graph embedding method based on walk sequences in graph representation learning. In real-world traffic scenarios, traffic flow direction is often non-random, and vehicles exhibit certain tendencies in path selection during travel. This path tendency represents the dynamic transitivity of traffic flow. By improving the traditional walk sequence process, a dynamic traffic flow direction walk path sequence that matches historical traffic flow transition probabilities is obtained to extract the dynamic transitivity sequence characteristics of traffic flow. Another core concept of this model is to use a graph sampling aggregation model, utilizing node sampling and message passing mechanisms to obtain regional dynamic and static features. These features are then combined with dynamic transitivity sequence characteristics to identify key segments. The model's results on both public and private datasets have demonstrated high accuracy in segment selection.

[0004] Although relatively good results have been achieved, the Path2vec model still has the following problems to be solved:

[0005] (1) In the traditional graph sampling aggregation model, node sampling is achieved by uniformly randomly sampling a fixed number of neighboring nodes. This lacks specificity for complex traffic networks and is prone to include noise nodes, which affects the effect of the aggregation model.

[0006] (2) High-dimensional vector node features may introduce redundancy or noise, resulting in low model training efficiency and unsuitable for lightweight requirements in practical applications. Summary of the Invention

[0007] In view of the above problems existing in the prior art, the technical problem to be solved by the present invention is: how to improve the accuracy and efficiency of locating key road sections.

[0008] To solve the above technical problems, the present invention adopts the following technical solution: a key road section identification method based on road network dynamic feature sampling and aggregation, comprising the following steps:

[0009] Select a traffic database X for a certain area:

[0010] S100: Using the networkx library and the pandas library in Python to process data X, obtain a traffic data information table corresponding to X. The traffic data information table includes a network topology diagram and traffic characteristic information;

[0011] Convert the traffic data information table corresponding to X into a matrix to obtain the traffic network line diagram G;

[0012] The expression of G is as follows:

[0013] G=(V,E,S T ,F T ,W T )

[0014] Where V=(v1,v2,…,v n ) is a set of nodes, v n represents the nth node in G, n represents the total number of nodes in G, E=(v i ,v j ) represents the edge set in G, (v i ,v j ) represents v in G i With v j An edge between, i,j∈[1,n]; S T =(s1,s2,…,s n ) represents the set of average speeds of each node in G within a specific time T, s n represents the average speed of vehicles passing through the nth node in G; F T =(f1,f2,…,f n ) represents the set of traffic flows at each node in G within a specific time T, f n represents the traffic flow passing through the nth node in G; Q T =(q1,q2,…,q n ) represents the set of lengths of the actual road sections corresponding to the nodes, q n Indicates the length of the actual road segment corresponding to the nth node in G;

[0015] S200: constructing a key road identification model M, M including a data preprocessing module, a GraphDFSAGE module and a key list output module;

[0016] The data preprocessing module is used to preprocess G according to S in G T ,F T ,Q T , get the initial multi-dimensional regional feature vector set of all nodes in G

[0017] The GraphDFSAGE module is an improved GraphSAGE layer model, which changes the random sampling method of GraphSAGE into a sampling method based on traffic flow and with a bias; Input GraphDFSAGE module with G, and output the embedding feature set of all nodes

[0018] The key list output module is used to As input, the output is the final critical road list;

[0019] S300: Select the traffic database of the area Y to be tested, repeat S100-S200, and obtain a list of key roads of Y.

[0020] As an example, the initial multi-dimensional regional feature vector set of all nodes in G is obtained in S200. The calculation process is as follows:

[0021] S210: Calculate v i The initial eigenvector of The calculation expression is as follows:

[0022]

[0023] Among them, CONCAT(·) represents the connection operation, s i ∈S T , f i ∈F T ,q i ∈Q T , i∈[1,n];

[0024] S211: Calculate the initial feature vectors of all nodes in V and stack the initial multi-dimensional regional feature vectors of all nodes in rows to form an initial multi-dimensional regional feature vector set The calculation expression is as follows:

[0025]

[0026] in, represents the initial feature vector of node v1 in the l=1 sampling layer, Represents node v n Initial feature vector at sampling level l=1.

[0027] As an example, the node embedding vector is output in S200 The steps are as follows:

[0028] Set the number of sampling layers of the GraphDFSAGE module to K, set the edge sampling weight value of each node, and set the number of neighborhood samples of each node to n;

[0029] S220: Calculate the i-th node v i Neighborhood N(v i ), N(v i ) all nodes in it are v i Neighbor nodes, N(v i )The calculation expression is as follows:

[0030] N(v i )={u i,j :D(u,v i )≤K,u i,j ∈G}

[0031] Among them, D(u i,j ,v i ) means v i with u i,j The distance between i,j Indicates v i The jth neighbor node of

[0032] S221: Calculate v i Sampling dynamic neighborhood DFSN (v i ), the expression is as follows:

[0033]

[0034] in, Indicates v i with u i,j The traffic volume between, Top(n) represents v i with u i,j When the traffic volume is arranged in descending order, the top n traffic volume values;

[0035] S222: Aggregate all u in layer l i,j The embedding features of , the calculation expression is as follows:

[0036]

[0037] in, Represents node v i Neighbor node u i,j The embedded features at layer l, Indicates u i,j Embedded features in the l-1 layer, l represents the l-th sampling layer, l∈[1,K];

[0038] S223: Aggregation and Get v i The final embedding features The calculation expression is as follows:

[0039]

[0040] in, represents any neighbor node u i,j , Indicates v i In the embedded features of the l-1 layer, σ is the nonlinear activation function, and Wl is the learnable weight matrix of the l layer;

[0041] S224: Repeat S220-S223 to obtain the final embedded features of all nodes, and stack the final embedded features of all nodes by row to obtain the final embedded feature set of all nodes in V The calculation expression is as follows:

[0042]

[0043] in, Represents node v n The final embedding features at the Kth layer.

[0044] Preferably, in S200, the clustering algorithm is the kmeans algorithm, and λ is set to 3.

[0045] Preferably, the step of obtaining the final key road list in S200 is as follows:

[0046] S230: Set clustering algorithm parameter λ and use clustering algorithm to Clustering is performed to obtain λ-class traffic sections;

[0047] S231: Calculate the average value of traffic demand index of all nodes in each type of traffic section obtained in S230 The calculation formula is as follows:

[0048]

[0049] in, h is The total number of nodes in the traffic segment q h , represents the length of the actual road section corresponding to the hth node, f h represents the traffic flow passing through the hth node, s h represents the average speed of the vehicle passing the hth node;

[0050] S232: Select the category with the highest average value of traffic demand index as the key category, sort the traffic demand indexes of all nodes in the key category in descending order, select the actual road sections corresponding to the traffic demand indexes of the first W nodes as key road sections, output the W key road sections, and obtain the final key road list.

[0051] Compared with the prior art, the present invention has at least the following advantages:

[0052] 1. The technical methods described in this invention can more accurately extract traffic information from road networks. Traditional GraphSAGE uses a random sampling approach, where all neighbors are sampled with equal probability. At each sampling level, for each central node, the central node's neighbors are selected as candidate nodes. A certain number of nodes are randomly selected from these candidate nodes as sampling nodes. The sampled nodes are then aggregated to represent the central node through feature aggregation. GraphDFSAGE, on the other hand, transforms GraphSAGE's random sampling approach into a traffic-based, biased sampling approach. Specifically, at each sampling level, for each central node, all neighbors are first collected and sorted by the traffic between them. Nodes with the highest traffic rankings are selected as candidate sampling nodes. The nodes obtained through this sampling approach are then aggregated to represent the central node. This improved node sampling method improves the quality of information aggregation. Points with high traffic weights typically represent stronger interactions with the central node. Aggregating these filtered neighbors avoids noise interference and generates higher-quality node embedding vectors. By selectively filtering neighbors with high traffic weights, gradient updates for invalid data are reduced, while accelerating model convergence. This allows the model to learn key patterns more quickly, reducing the number of model iterations. The random sampling method in traditional methods is likely to sample neighboring nodes with low correlation. Such low-correlation nodes have low value and affect the generalization of the model. However, the method of the present invention filters out low-correlation nodes, allowing the model to focus on strong semantic correlation. The application of semantic correlation in traffic road networks is manifested as high-traffic roads having a higher correlation with surrounding roads.

[0053] 2. To address the problems of low model training efficiency caused by single feature extraction and lack of targeted neighbor sampling in traditional graph sampling aggregation models, the present invention innovatively proposes a road network dynamic feature sampling aggregation model GraphDFSAGE, which achieves biased dynamic sampling of node neighborhoods to obtain multi-dimensional regional dynamic and static features, and clusters the obtained embedding vectors to obtain key road sections in the network; the present invention uses a Python library in the data calling process, and will directly output the trained results when calling the Python library.

[0054] 3. From the perspective of the dynamic and static characteristics of the traffic network, the accuracy of the selected road sections is evaluated based on node network efficiency, network survivability, and dynamic contagion indicators. Compared with traditional indicator evaluation methods and graph representation learning methods, the method proposed in this invention has higher accuracy and is significantly superior to the key road sections obtained by traditional methods in terms of network efficiency indicators. It has better performance in terms of the number of subgraphs and the ratio of the number of nodes, and has a huge advantage over other methods in terms of network efficiency indicators. This shows that the model can fully extract and aggregate road network node features, has a greater impact on the efficiency of information exchange in the network and the integrity of the road network, and enhances its influence on the dynamic and static indicators of the network.

[0055] 4. This invention improves upon the traditional GraphSAGE, which has three sampling layers and five samples per layer. However, the improved GraphDFSAGE in this invention reduces the number of sampling layers and the number of samples per layer to two, reducing the number of variables in the sampling layers and number of samples. Furthermore, the initial node features only include flow, speed, and length, resulting in a reduced dimensionality. This reduces sampling complexity and, consequently, noise in the sampling. A line graph mapping method based on a traffic network is used, where road links are represented as nodes and their connections as edges. The traffic flow, speed, and length attributes of a road segment are integrated into the node attributes of the traffic network, thereby achieving a lightweight representation of node features.

[0056] 5. While ensuring the improvement of dynamic and static indicators, the number of layers and neighboring nodes in the model graph sampling neighborhood is reduced, effectively reducing the complexity of sampling, improving the embedding quality of the model, and increasing the running speed of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 Schematic diagram of the method architecture of the present invention;

[0058] Figure 2 Schematic diagram of the dynamic feature sampling layer structure;

[0059] Figure 3 This is the network efficiency change diagram of PEMS08 during the lunchtime peak period;

[0060] Figure 4 This is the change diagram of node ratio during the noon peak of PEMS08;

[0061] Figure 5 This is the change diagram of the number of sub-graphs during the noon peak of PEMS08;

[0062] Figure 6 This is the change diagram of network efficiency during evening peak hour at Chongqing West Railway Station;

[0063] Figure 7 is the total network efficiency value of Chongqing West Railway Station during evening peak hours;

[0064] Figure 8 This is the change diagram of the node ratio during the evening peak hour at Chongqing West Railway Station;

[0065] Figure 9 This is the total node ratio map of Chongqing West Railway Station during the evening peak hour;

[0066] Figure 10 This is the change diagram of the number of sub-graphs during the evening peak hour at Chongqing West Railway Station;

[0067] Figure 11 This is the total number of sub-graphs at the evening peak of Chongqing West Railway Station;

[0068] Figure 12 The changing trend of congestion nodes during evening rush hour at Chongqing West Railway Station;

[0069] Figure 13 This is the road network location map of the top 15 key road sections in the area. DETAILED DESCRIPTION

[0070] The present invention is described in further detail below.

[0071] Select a traffic database X for a certain area:

[0072] S100: Using the networkx library and the pandas library in Python to process data X, obtain a traffic data information table corresponding to X. The traffic data information table includes a network topology diagram and traffic characteristic information;

[0073] The traffic data information table corresponding to X is converted into a matrix to obtain the traffic network line graph G; the networkx library and the pandas library are existing technologies;

[0074] The expression of G is as follows:

[0075] G=(V,E,S T ,F T ,W T )

[0076] Where V=(v1,v2,…,v n ) is a set of nodes, v n represents the nth node in G, n represents the total number of nodes in G, E=(v i ,v j ) represents the edge set in G, (v i ,v j ) represents v in G i With v j An edge between, i,j∈[1,n]; S T =(s1,s2,…,s n ) represents the set of average speeds of each node in G within a specific time T, s nrepresents the average speed of vehicles passing through the nth node in G; F T =(f1,f2,…,f n ) represents the set of traffic flows at each node in G within a specific time T, f n represents the traffic flow passing through the nth node in G; Q T =(q1,q2,…,q n ) represents the set of lengths of the actual road sections corresponding to the nodes, q n Indicates the length of the actual road segment corresponding to the nth node in G;

[0077] S200: constructing a key road identification model M, M including a data preprocessing module, a GraphDFSAGE module and a key list output module;

[0078] The data preprocessing module is used to preprocess G according to S in G T ,F T ,Q T , get the initial multi-dimensional regional feature vector set of all nodes in G

[0079] The GraphDFSAGE module is an improved GraphSAGE layer model, which changes the random sampling method of GraphSAGE into a sampling method based on traffic flow and with a bias; Input GraphDFSAGE module with G, and output the embedding feature set of all nodes The GraphDFSAGE module, also known as the dynamic feature sampling layer, primarily acquires dynamic neighborhoods and aggregates domain features. In acquiring the dynamic neighborhood, the top n neighbors in terms of traffic volume are selected from each layer of the central node's neighbors, based on the number of sampling layers, K, and the number of samples, n. In each layer of domain feature aggregation, for the central node, the features of the nodes in the dynamic neighborhood of that layer are aggregated according to the selected aggregation function, concatenated with the node's embedded features in the previous layer, and linearly represented as the final embedding vector using an activation function.

[0080] The key list output module is used to As input, the output is the final critical road list;

[0081] In S200, the initial multi-dimensional regional feature vector set of all nodes in G is obtained. The calculation process is as follows:

[0082] S210: Calculate v i The initial eigenvector of The calculation expression is as follows:

[0083]

[0084] Among them, CONCAT(·) represents the connection operation, s i ∈S T , f i ∈F T ,q i ∈Q T , i∈[1,n];

[0085] S211: Calculate the initial feature vectors of all nodes in V and stack the initial multi-dimensional regional feature vectors of all nodes in rows to form an initial multi-dimensional regional feature vector set The calculation expression is as follows:

[0086]

[0087] in, represents the initial feature vector of node v1 in the l=1 sampling layer, Represents node v n Initial feature vector at sampling level l=1.

[0088] The node embedding vector is obtained by outputting in S200 The steps are as follows:

[0089] Set the number of sampling layers of the GraphDFSAGE module to K, set the edge sampling weight value of each node, and set the number of neighborhood samples of each node to n;

[0090] S220: Calculate the i-th node v i Neighborhood N(v i ), N(v i ) all nodes in it are v i Neighbor nodes, N(v i )The calculation expression is as follows:

[0091] N(v i )={u i,j :D(u,v i )≤K,u i,j ∈G}

[0092] Among them, D(u i,j ,v i ) means v i with u i,j The distance between i with u i,j There are multiple paths between them, and the shortest path is v i with u i,j The distance between i,j Indicates v i The jth neighbor node of

[0093] S221: Calculate v i Sampling dynamic neighborhood DFSN (v i ), the expression is as follows:

[0094]

[0095] in, Indicates v i with u i,j The traffic volume between, Top(n) represents v i with u i,j When the traffic volume is arranged in descending order, the top n traffic volume values;

[0096] S222: Aggregate all u in layer l i,j The embedding features of , the calculation expression is as follows:

[0097]

[0098] in, Represents node v i Neighbor node u i,j The embedded features at layer l, Indicates u i,j Embedded features in the l-1 layer, l represents the l-th sampling layer, l∈[1,K]; It is v i The initial value of a neighbor node is equal to its initial value when it is its own node;

[0099] S223: Aggregation and Get v i The final embedding features The calculation expression is as follows:

[0100]

[0101] in, represents any neighbor node u i,j , Indicates v i In the embedded features of the l-1 layer, σ is the nonlinear activation function, W l is the learnable weight matrix of the lth layer; when l = 1, Take the calculated value in step S310;

[0102] S224: Repeat S220-S223 to obtain the final embedded features of all nodes, and stack the final embedded features of all nodes by row to obtain the final embedded feature set of all nodes in V The calculation expression is as follows:

[0103]

[0104] in, Represents node v n The final embedding features at the Kth layer.

[0105] In S200 , the clustering algorithm is the kmeans algorithm, and λ is set to 3.

[0106] The steps of obtaining the final key road list in S200 are as follows:

[0107] S230: Set clustering algorithm parameter λ and use clustering algorithm to Clustering is performed to obtain λ-class traffic sections;

[0108] S231: Calculate the average value of traffic demand index of all nodes in each type of traffic section obtained in S230 The calculation formula is as follows:

[0109]

[0110] in, h is The total number of nodes in the traffic segment q h , represents the length of the actual road section corresponding to the hth node, f h represents the traffic flow passing through the hth node, s h represents the average speed of the vehicle passing the hth node;

[0111] S232: Select the category with the highest average value of traffic demand index as the key category, sort the traffic demand indexes of all nodes in the key category in descending order, select the actual road sections corresponding to the traffic demand indexes of the first W nodes as key road sections, output the W key road sections, and obtain the final key road list.

[0112] S300: Select the traffic database of the area Y to be tested, repeat S100-S200, and obtain a list of key roads of Y.

[0113] Experiment and analysis:

[0114] 1. Dataset Description

[0115] The present invention conducts experiments on the PEMS08 public dataset and the Chongqing West Railway Station private dataset to obtain the traffic flow data, average road speed data, road length data and other traffic characteristics required by the traffic network line diagram for each external road detection device during the morning peak hours 8:00-9:00, the lunch peak hours 12:00-13:00 and the evening peak hours 17:00-18:00.

[0116] PEMS08 dataset: PEMS08 is a PeMS traffic flow dataset for the California freeway network. PEMS08 collects data from 170 detectors every 5 minutes for 62 days. This dataset includes traffic flow, link connectivity, and link length for the California freeway network. After constructing a traffic network diagram, K is set to 30, which is an appropriate number of key links.

[0117] Chongqing West Railway Station Dataset: The same experimental verification was conducted on a private dataset from the West Railway Station. This dataset contains raw traffic data from July 1, 2024, to July 31, 2024, collected by external road detection equipment in the road network surrounding Chongqing West Railway Station. After data cleaning and processing, traffic flow data, average road speed data, and road length data were obtained for each external road detection device during peak hours in the morning, afternoon, and evening. K was set to 15, indicating an appropriate number of key road sections.

[0118] 2. Comparison Model

[0119] Comparing the key section selection method based on the road network dynamic feature sampling aggregation model, the key section selection method based on the centrality index, and the key section selection method based on graph representation learning proposed in the present invention, from the perspective of overall network efficiency, two indicators of network invulnerability are the number of subgraphs existing in the network after deleting the node, the proportion of nodes in the network's largest connected subgraph after deleting the node, and the accuracy of the selected key sections evaluated by the network infectious index. The lower the overall network efficiency index, the greater the impact of the node on the overall network efficiency. The more subgraphs that exist in the network after deleting the node, the more dispersed the node divides the network, and the more important the node. The lower the proportion of nodes in the network's largest connected subgraph after deleting the node, the greater the impact of the node on the global traffic capacity of the network, and the more important the node. The higher the network infectious index, the more congested sections there are in the road network, and the greater the influence of the node on the spread of the congestion state of the global road network.

[0120] Nine models were selected as comparison models for the comparative experiments. These models are very representative, including key link identification models based on flow or centrality selection, key link identification models based on random walks in graph representation learning, and key link identification models based on graph neural network methods in graph representation learning. The following is a brief introduction to these baseline models:

[0121] (1) Onlyflow: This method uses only traffic flow as an indicator to determine the importance of traffic sections. Key sections are selected based on the traffic flow ranking. Traffic flow is an important data in traffic flow. Traffic sections with high traffic flow often have a greater impact on the traffic network. This method selects the top K traffic sections in the network.

[0122] (2) Only Degree: This method uses only node degree as an indicator to judge the importance of traffic sections. Key sections are selected based on the degree ranking. Node degree is one of the indicators of network node centrality. Degree reflects the number of connections to nodes in the network. A higher node degree indicates more edges connected to the node. When the node fails, more nodes may be affected. This method selects the top K traffic sections in the network.

[0123] (3) OnlyBetweenness: This method uses only node betweenness as an indicator of the importance of a traffic segment, and selects key segments based on the ranking of betweenness. Node betweenness is one of the indicators of network node centrality. Nodes with high betweenness are often bridges in a network, connecting different communities or groups, such as information dissemination intermediaries in social networks. This method selects the top K traffic segments ranked by network betweenness.

[0124] (4) Only Closeness: This means that only node proximity is used as an indicator to judge the importance of traffic sections, and key sections are selected based on the proximity ranking. Node proximity is one of the indicators of network node centrality, representing the inverse of the average shortest distance from a node to all other nodes, reflecting the efficiency of node information dissemination. Nodes with high proximity can quickly reach other nodes. This method selects the top K traffic sections in the network with the highest proximity ranking.

[0125] (5) Deepwalk: It is a purely random walk model. The traffic network line diagram is input into Deepwalk. A random walk is performed on each node to obtain a walk path sequence. The word vector model is used to learn the embedding vector. The important road sections are obtained through clustering, and the top K traffic sections are selected from them.

[0126] (6) Node2vec: Biased random walk model. The traffic network line graph is input into Node2vec. A biased random walk is performed on each node to obtain the walk path sequence. The word vector model is used to learn the embedding vector. The important road sections are obtained through clustering, and the top K traffic sections are selected from them.

[0127] (7) Path2vec: The dynamic traffic flow wandering model proposed in Chapter 3 obtains a dynamic traffic flow wandering path sequence that matches the historical traffic flow transfer probability. The Skip-gram model is then used to learn the low-dimensional vector representation of the node, capturing the dynamic transfer sequence characteristics of the traffic flow and embedding the vector. The vectors are clustered to obtain important road sections, from which the top K traffic sections are selected.

[0128] (8) OnlyGraphSAGE: A graph sampling and aggregation model that integrates dynamic attributes such as node traffic and speed characteristics with static attributes such as node degree into a feature matrix. By sampling a certain number of neighboring nodes and aggregating them, the node embedding vector is obtained. After clustering, important road sections are obtained, and the top K traffic sections are selected from them.

[0129] (9) GraphDFSAGE: This paper proposes a road network dynamic feature sampling and aggregation model that combines the dynamic transitivity sequence characteristics of dynamic traffic flow with biased multi-dimensional regional characteristics. It samples more important nodes through dynamic graph sampling, obtains embedding vectors, and obtains key road segments through clustering. The top K traffic road segments are selected from these.

[0130] 3. Model training settings

[0131] The parameters of each model in the experiment are as follows: K is set to 30 in the PEMS08 dataset and to 15 in the West Station dataset, where K represents the number of key road sections selected. The embedding vector lengths of Deepwalk, Node2vec, and Path2vec are 128, the walk sequence length is 80, and the number of walks is 10. The embedding vector length of GraphSAGE is 128, the number of model sampling layers is 3, and the number of sampled neighbors is 5. The infection rate is set to 0.8, the recovery rate is 0.2, the time step is set to 1, and the total time T is 15 to simulate the continuous spread of high infection rate and low recovery rate. The road network dynamic feature sampling aggregation model GraphDFSAGE proposed in this invention has 2 sampling layers and 3 sampling nodes in the experiment. Compared with the sampling of GraphSAGE, the number of layers is lower and the number of nodes is smaller.

[0132] 4. Experimental Results

[0133] After identifying key segments during the morning rush hour (8:00-9:00 AM), the afternoon rush hour (12:00-1:00 PM), and the evening rush hour (17:00-18:00 PM), using traditional flow or centrality metrics models, traditional graph representation learning models, and the proposed aggregation model method, the authors calculated the network efficiency after the key segments were disabled, the number of subgraphs remaining in the network after node deletion, the ratio of the number of nodes in the largest connected component of the network after node deletion to the number of nodes in the original graph, and the number of infected nodes in the network simulated by an infectious disease model. The average values of these metrics for each result were calculated, as well as the total values of these metrics after 100 experiments. The traffic network diagram constructed based on the PEMS08 dataset contains 277 segment nodes. In actual traffic control and planning, excessive nodes are not selected as key nodes for cost considerations. Therefore, K was set to 30 in the PEMS08 dataset model as an appropriate number of key segments. The same experiment was conducted on the Chongqing West Railway Station private dataset. The following section briefly describes the experimental results on the public dataset, followed by a detailed description of the experimental results on the private dataset.

[0134] (1) Experimental results on the public dataset PEMS08

[0135] Figure 3-Figure 5 This paper demonstrates the changing trends of the network efficiency, subgraph node ratio, and number of network subgraphs during the morning peak period of the public dataset PEMS08 using the road network dynamic feature sampling and aggregation model over 100 experiments. The network efficiency of the road network dynamic feature sampling and aggregation model is significantly lower than that of traditional indicator methods and traditional graph representation learning methods, indicating that the key road segments obtained by the road network dynamic feature sampling and aggregation model have a greater impact on the network transmission efficiency and global traffic capacity of the entire network. Overall, the road network dynamic feature sampling and aggregation model proposed in this paper significantly outperforms other comparison methods in all evaluation indicators on the public dataset, demonstrating that the model can effectively solve the key road segment classification problem in standard application scenarios.

[0136] Table 1 Performance indicators of each model of PEMS08

[0137]

[0138]

[0139] Table 1 shows the performance of various models on the PEMS08 dataset. To enhance the clarity of the results, the values for each metric are the sum of the results from 100 simulated trials. The proposed GraphDFSAGE model outperforms all other baseline methods in both dynamic and static metrics for key sections. Compared to the dynamic traffic flow model presented in Chapter 3, the ratio of the number of nodes in the largest connected component of the network to the number of nodes in the original graph after node deletion improves, achieving the best performance among the baseline methods.

[0140] (2) Experimental results of private dataset Chongqing West Railway Station

[0141] The GraphDFSAGE model proposed in this paper has further improved the performance of indicators on the Chongqing West Railway Station dataset. Figure 6 , Figure 7 The network efficiency indicators of Chongqing West Railway Station during the evening rush hour from 17:00 to 18:00 are shown. The network efficiency after the failure of key sections obtained by the GraphDFSAGE model proposed in this invention is the lowest, followed by the GraphSAGE model and the Path2vec model. The network efficiency results after the failure of sections based on flow and centrality selection methods are poor. Compared with the sections selected by flow, GraphDFSAGE has a 20% increase in the ability to influence network efficiency indicators, which is significantly better than the other eight methods. This shows that the key sections obtained by the method proposed in this invention can better affect the efficiency of network information exchange and have a greater global influence.

[0142] Figures 8-11 The results show the indestructibility index of Chongqing West Railway Station during the evening rush hour from 17:00 to 18:00. Compared with the model proposed in Chapter 3, the GraphDFSAGE model has a more obvious improvement in the node ratio index, and the node ratio is further reduced. The effect of the method proposed in Chapter 3 is close to that of the method proposed in Chapter 3 in terms of the number of subgraphs. The node number ratio index of the GraphDFSAGE model proposed in this invention is better than that of the other 8 methods, indicating that the method proposed in this invention takes into account both local connection patterns and global path dependencies, and more comprehensively evaluates the function of nodes in complex transportation networks. At the same time, the model proposed in this invention can obtain that the network cannot maintain a high integrity after the deletion of key sections, and has a greater ability to damage network connectivity. It stably obtains key sections with higher network segmentation capabilities, and has a stronger ability to damage the integrity of the road network after node failure.

[0143] Figure 12The changing trend of congestion nodes in the traffic congestion infection model during the evening peak period is shown. The blue one is the GraphDFSAGE model proposed by the present invention. From the changing trend of the congested nodes, it can be seen that the node with the highest degree is the first to experience the outbreak of congestion in the infection model, the congestion time is earlier, and the number of congestion is higher. The node with the highest degree is connected to the largest number of edges, so it conforms to the logic of rapid infection and high number of infections in the traffic congestion infection model. The infection peak time of the GraphDFSAGE model is later than that of the node with the highest degree, and the number of infections at the peak is slightly lower than that of the node with the highest degree, but it is also much higher than the nodes selected by other baseline methods. From the trend of the number of infected nodes, it can be seen that the method proposed by the present invention, although not the node with the highest degree, the nodes selected by the present invention are closely related to the nodes with the highest degree. Once the node of the present invention is infected, it will quickly spread to the nodes with high degrees, and then its continuous infection ability for the entire road network is also better than that of the nodes with high degrees. This shows that the method of the present invention has successfully selected traffic sections with strong dynamic infection capabilities.

[0144] Figure 13 Figure 3 shows the locations of the top 15 key road segments in real traffic during the evening rush hour at Chongqing West Railway Station, as determined by the GraphDFSAGE model proposed in this invention, and the locations of the top 15 key road segments obtained using traffic flow. The present invention uses red lines in the figure. Key road segments in a real road network may be located along main roads, bridges, tunnels, or connecting sections around transportation hubs. The results show that key road segments based on traffic flow only include main roads with high traffic volume, and sections with less impact on the actual road network, such as Blue 11 and Blue 26. However, the key road segments obtained by the GraphDFSAGE method proposed in this invention include not only main roads but also important intersecting areas, such as Red 9 and Blue 2, and drop-off areas, such as Red 4 and Red 5. This method better aligns with commonly encountered congested road segments and is more consistent with the empirical location of traffic segments. However, the road segments obtained using traffic flow are less capable of identifying complex intersecting areas and important road segments connecting the road network, primarily focusing on identifying main roads with high traffic volume.

[0145] Table 2 Performance indicators of various models of Chongqing West Railway Station

[0146]

[0147] Table 2 shows the performance indicators of the Chongqing West Railway Station model during three peak periods. It can be seen that the GraphDFSAGE model proposed in this invention has further improved network efficiency, network anti-destruction, and network dynamic infection compared with the model proposed in Chapter 3, especially the network efficiency index. Consistent with the sampling logic of the dynamic feature sampling model, the dynamic feature sampling process simulates the transmission trend of traffic flow in an area, and indirectly reflects the process of information exchange in the network, thus bringing about a large improvement in network efficiency indicators. By analyzing the data, the model proposed in this invention fully obtains the dynamic transmission sequence characteristics of traffic flow, and has the highest network infection index. The model focuses on aggregating the dynamic and static regional characteristics of the road network. Compared with the traditional graph sampling aggregation method, while reducing the complexity of sampling aggregation, it improves the influence on the dynamic and static indicators of the network, and the selected key sections are more accurate.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for identifying key road sections based on road network dynamic feature sampling and aggregation, characterized in that: The steps include: Select a traffic database X for a certain area: S100: Using the networkx library and the pandas library in Python to process data X, obtain a traffic data information table corresponding to X. The traffic data information table includes a network topology diagram and traffic characteristic information; Convert the traffic data information table corresponding to X into a matrix to obtain the traffic network line diagram G; The expression of G is as follows: G=(V,E,S T ,F T ,Q T ) Where V=(v1,v2,…,v n ) is a set of nodes, v n represents the nth node in G, n represents the total number of nodes in G, E=(v i ,v j ) represents the edge set in G, (v i ,v j ) represents v in G i With v j An edge between, i,j∈[1,n]; S T =(s1,s2,…,s n ) represents the set of average speeds of each node in G within a specific time T, s n represents the average speed of vehicles passing through the nth node in G; F T =(f1,f2,…,f n ) represents the set of traffic flows at each node in G within a specific time T, f n represents the traffic flow passing through the nth node in G; Q T =(q1,q2,…,q n ) represents the set of lengths of the actual road sections corresponding to the nodes, q n Indicates the length of the actual road segment corresponding to the nth node in G; S200: constructing a key road identification model M, M including a data preprocessing module, a GraphDFSAGE module and a key list output module; The data preprocessing module is used to preprocess G according to S in G T ,F T ,Q T , get the initial multi-dimensional regional feature vector set of all nodes in G The GraphDFSAGE module is an improved GraphSAGE layer model, which changes the random sampling method of GraphSAGE into a sampling method based on traffic flow and with a bias; Input GraphDFSAGE module with G, and output the embedding feature set of all nodes The key list output module is used to As input, the output is the final critical road list; S300: Select the traffic database of the area Y to be tested, repeat S100-S200, and obtain a list of key roads of Y.

2. The method for identifying key road sections based on road network dynamic feature sampling and aggregation according to claim 1, characterized in that: In S200, the initial multi-dimensional regional feature vector set of all nodes in G is obtained. The calculation process is as follows: S210: Calculate v i The initial eigenvector of The calculation expression is as follows: Among them, CONCAT(·) represents the connection operation, s i ∈S T , f i ∈F T ,q i ∈Q T , i∈[1,n]; S211: Calculate the initial feature vectors of all nodes in V and stack the initial multi-dimensional regional feature vectors of all nodes in rows to form an initial multi-dimensional regional feature vector set The calculation expression is as follows: in, represents the initial feature vector of node v1 in the l=1 sampling layer, Represents node v n Initial feature vector at sampling level l=1.

3. The method for identifying key road sections based on road network dynamic feature sampling and aggregation according to claim 2, characterized in that: The node embedding vector is obtained by outputting in S200 The steps are as follows: Set the number of sampling layers of the GraphDFSAGE module to K, set the edge sampling weight value of each node, and set the number of neighborhood samples of each node to n; S220: Calculate the i-th node v i Neighborhood N(v i ), N(v i ) all nodes in it are v i Neighbor nodes, N(v i )The calculation expression is as follows: N(v i )={u i,j :D(u,v i )≤K,u i,j ∈G} Among them, D(u i,j ,v i ) means v i with u i,j The distance between i,j Indicates v i The jth neighbor node of S221: Calculate v i Sampling dynamic neighborhood DFSN (v i ), the expression is as follows: in, Indicates v i with u i,j The traffic volume between, Top(n) represents v i with u ,j When the traffic volume is arranged in descending order, the top n traffic volume values; S222: Aggregate all u in layer l ,j The embedding features of , the calculation expression is as follows: in, Represents node v i Neighbor node u i,j The embedded features at layer l, Indicates u ,j Embedded features in the l-1 layer, l represents the l-th sampling layer, l∈[1,K]; S223: Aggregation and Get v i The final embedding features The calculation expression is as follows: in, represents any neighbor node u i,j , Indicates v i In the embedded features of the l-1 layer, σ is the nonlinear activation function, W l is the learnable weight matrix of layer l; S224: Repeat S220-S223 to obtain the final embedded features of all nodes, and stack the final embedded features of all nodes by row to obtain the final embedded feature set of all nodes in V The calculation expression is as follows: in, Represents node v n The final embedding features at the Kth layer.

4. The method for identifying key road sections based on road network dynamic feature sampling and aggregation according to claim 3, characterized in that: In S200 , the clustering algorithm is the kmeans algorithm, and λ is set to 3.

5. The method for identifying key road sections based on road network dynamic feature sampling and aggregation according to claim 4, characterized in that: The steps of obtaining the final key road list in S200 are as follows: S230: Set clustering algorithm parameter λ and use clustering algorithm to Clustering is performed to obtain λ-class traffic sections; S231: Calculate the average value of traffic demand index of all nodes in each type of traffic section obtained in S230 The calculation formula is as follows: in, h is The total number of nodes in the traffic segment q h , represents the length of the actual road section corresponding to the hth node, f h represents the traffic flow passing through the hth node, s h represents the average speed of the vehicle passing the hth node; S232: Select the category with the highest average value of traffic demand index as the key category, sort the traffic demand indexes of all nodes in the key category in descending order, select the actual road sections corresponding to the traffic demand indexes of the first W nodes as key road sections, output the W key road sections, and obtain the final key road list.

Images