Signaling road network matching method and system based on heterogeneous graph convolutional network and attention mechanism

Abstract

The invention discloses a signaling road network matching method based on a heterogeneous graph convolutional network and an attention mechanism, and belongs to the crossing field of mobile communication and intelligent traffic technologies. According to the method, a multi-relation graph containing a road segment set and a base station sampling point set is constructed, wherein an edge set covers associated edges of base station sampling points and road segments, topological edges of the road segments and time edges of the base station sampling points; performing feature extraction on the multi-relation graph by adopting a 3-5-layer heterogeneous graph convolutional network to obtain low-dimensional vector representation of nodes; dynamically calculating an observation probability and a transition probability based on an attention mechanism; and performing path search by using an improved Viterbi algorithm in combination with a backtracking mechanism and a pruning strategy, and outputting a path with the highest score as a matching result. The method can effectively solve the problems of poor adaptability to complex road networks, insufficient robustness to high positioning error signaling data, low precision caused by matching probability immobilization and the like in the prior art, significantly improves the accuracy and robustness of signaling road network matching in complex scenes, and can be applied to the fields of traffic flow monitoring, travel mode analysis, urban planning and the like.

Description

Technical Field

[0001] This application relates to the fields of mobile communication and intelligent transportation technology, and in particular to a signaling road network matching method based on heterogeneous graph convolutional networks and attention mechanisms. Background Technology

[0002] With the rapid development of mobile communication technology, user movement trajectory data (signaling data) is widely used in traffic flow monitoring, travel pattern analysis, urban planning, and other fields. Signaling-road network matching, as a core technology, requires aligning signaling trajectories with the actual road network to obtain the true travel path.

[0003] Existing technologies have significant limitations in traditional road network matching methods: nearest neighbor matching algorithms based on geometric distance rely solely on spatial distance and ignore road network topology, making them prone to matching errors in complex road networks (such as areas with dense intersections); Dijkstra's algorithm based on topological information relies excessively on road network connectivity and has poor adaptability to signaling data with high positioning errors; although Hidden Markov Models (HMMs) combine geometric and topological information, the observation probability and transition probability are mostly fixed functions set manually, making it difficult to capture the dynamic relationship between signaling and the road network, resulting in a significant decrease in matching accuracy in scenarios with sparse signaling sampling or complex road network structures.

[0004] Therefore, a signaling-road network matching method that can adaptively learn signaling and road network features and dynamically optimize matching probabilities is needed to improve matching accuracy and robustness in complex scenarios. Summary of the Invention

[0005] This application aims to address the technical problems existing in signaling road network matching methods, such as poor adaptability to complex road networks, insufficient robustness to signaling data with high positioning errors, and low accuracy caused by fixed matching probabilities.

[0006] To achieve the above objectives, this application provides a signaling road network matching method based on heterogeneous graph convolutional networks and attention mechanisms, comprising the following steps:

[0007] Step 1: Construct a multi-relationship graph , where the set of nodes , This is a collection of road segments, encompassing all road segments within the urban road network. This is the set of base station sampling points, which includes all signaling sampling points generated during user movement.

[0008] The edge set It contains three types of directed edges: Type 1 is the association edge between the base station sampling point and the road segment, connecting... With possible matches Type 2 consists of topological edges between road segments, connecting adjacent road segments. Type 3 consists of time edges for base station sampling points, connected in chronological order of sampling time. .

[0009] Step 2: Node embedding based on Heterogeneous Graph Convolutional Network (HGCN). HGCN is used to extract features from the multi-relation graph G, and node vector representations are learned through multi-layer message passing.

[0010] Furthermore, regarding the first Layer nodes , its first Layer features The calculation formula is:

[0011] Where R = {type1, type2, type3} is the set of relation types; For nodes In relation The set of neighbors below; This is a normalization constant; , The weight matrix is ​​a learnable weight matrix; This is the ReLU activation function.

[0012] By stacking 3-5 layers of HGCN, a low-dimensional vector representation (e.g., 64-dimensional) of all nodes is output, denoted as... (Characteristics of base station sampling points) and (Road segment characteristics).

[0013] Step 3: Probability calculation based on attention mechanism;

[0014] 3.1 Observation Overview : Indicates the base station sampling point Matching road segments probability

[0015] in: for The set of candidate road segments (in terms of (The road section within a 500-meter radius of the center). , This is the attention weight matrix (optimized through training).

[0016] 3.2 Transition Probability :based on arrive The movement, from the road section Transferred to probability

[0017] in: for The set of candidate road segments; , This is the attention weight matrix.

[0018] Step 4: Path search based on the improved Viterbi algorithm;

[0019] 4.1 Candidate Path Generation: For each base station sampling point ,reserve The top 5 road segments with the highest observed probabilities were selected as candidate matching road segments.

[0020] 4.2 Path Score Calculation: For the path Its score for:

[0021] 4.3 Improvement strategies, including backtracking mechanisms and pruning strategies;

[0022] The backtracking mechanism involves backtracking to the previous sampling point to reselect candidate road segments when the score of a certain path is lower than 1 / 3 of the score of the current best path; the pruning strategy involves retaining only the top 10 candidate paths and deleting the rest of the low-scoring paths.

[0023] 4.4 Output the path P with the highest score as the matching result.

[0024] This application also provides a signaling matching system based on heterogeneous convolutional networks and attention mechanisms, the system comprising: The data storage module receives base station sampling point data and road network data, and performs format verification and preprocessing. The multi-relationship graph construction module constructs a multi-relationship graph based on the input data, which includes the location vector representation of base station sampling points and road segments. The heterogeneous graph convolutional network module extracts features from multiple relationship graphs and generates low-dimensional vector representations of base station sampling points and road segments; The attention probability calculation module utilizes node features to calculate the observation probability and transition probability through an attention mechanism.

[0025] The path search module generates and filters candidate paths based on probability values ​​using an improved Viterbi algorithm, and outputs the matching path with the highest score.

[0026] Result output module: Receives the optimal path and outputs the matching result in a preset format.

[0027] This application has at least the following beneficial effects: it achieves fusion modeling of the association between base station sampling points and road segments, the topological connection of road segments, and the temporal relationship of sampling points based on the construction of multi-relationship graphs, breaking through the limitations of traditional single-relationship modeling; it achieves deep extraction of heterogeneous node features based on Heterogeneous Graph Convolutional Network (HGCN), solving the problem of dependence on manual features and improving feature representation capabilities; it dynamically optimizes observation probability and transition probability based on attention mechanism, enhancing adaptability to signaling data with high positioning error; and it improves path search efficiency while ensuring matching accuracy based on the improved Viterbi algorithm combined with backtracking and pruning strategies, significantly improving the accuracy and robustness of signaling road network matching in complex scenarios. Attached Figure Description

[0028] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 is a flowchart of the steps of a signaling road network matching method based on heterogeneous graph convolutional network and attention mechanism provided in an embodiment of this application;

[0030] Figure 2 is a schematic diagram of the multi-relationship diagram structure provided in an embodiment of this application;

[0031] Figure 3 is a schematic diagram of the structure of a signaling road network matching system based on heterogeneous graph convolutional network and attention mechanism provided in an embodiment of this application; Detailed Implementation

[0032] To further illustrate the technical means and effects adopted by this application to achieve the intended inventive purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the signaling road network matching method and system based on heterogeneous graph convolutional networks and attention mechanisms proposed in this application. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this invention.

[0033] The road network topology was obtained from OpenStreetMap, containing three road segments of different levels, namely... (Main Road A, 2000 meters long, 6 lanes in both directions) (Secondary arterial road B, 1500 meters long, 4 lanes in both directions) (Branch road C, 800 meters long, two lanes in both directions). Among them, and At the intersection The connection is achieved through traffic light control. and At the intersection The points are directly connected, forming the basic road network topology of the area. A user drives within this area, and their mobile terminal communicates with base stations along the route, generating three base station sampling points, namely... ( =8:00, latitude and longitude coordinates are ), ( =8:05, latitude and longitude coordinates are ), (t3=8:10, latitude and longitude coordinates are) Due to factors such as base station signal coverage and environmental interference, the signaling positioning error ranges from 50 to 100 meters, which poses a certain challenge to accurate road network matching.

[0034] Step 1: This step aims to construct a multi-relationship graph that can comprehensively reflect the various relationships between signaling sampling points and the road network. The specific process is as follows;

[0035] Based on the road network data and user signaling data of this area, a node set is generated: , ;

[0036] Based on the straight-line distance between the base station sampling points and each road segment, and combined with the actual distribution of the road segments, the associated edges between the base station sampling points and the road segments are determined: , ; , ; ;

[0037] Based on the road network topology within the region, determine the topological edges between road segments: , ;

[0038] Based on the order of sampling times, the time edge of the base station sampling points is determined: , .

[0039] Step Two: This step uses a heterogeneous graph convolutional network to extract features from nodes in a multi-relationship graph to obtain vector representations that effectively characterize node features. The specific process is as follows:

[0040] Initial feature input: For road segment nodes, the initial features include attributes such as road segment length, number of lanes, and speed limit information. These attributes are normalized and then used as the initial input. For base station sampling point nodes, the initial features are the normalized values ​​of latitude and longitude coordinates.

[0041] Node characteristics are calculated using a 3-layer HGCN:

[0042] Layer 1 HGCN computation: For each node, feature fusion is performed based on its neighboring nodes with different relationship types. (Based on nodes...) For example, its neighbors at the road segment node are and The neighbors of the base station sampling point node are and In the first layer of calculation, It aggregates the initial features of these neighboring nodes and combines them with its own initial features, using a weight matrix. (where rel corresponds to type 1 and type 2 respectively) and The calculation is performed, and then processed by the ReLU activation function to obtain... Feature representation at layer 1.

[0043] Layer 2 HGCN computation: Based on the node features obtained in Layer 1, a new round of feature fusion is performed. At this time, nodes not only aggregate the features of their direct neighbors, but are also indirectly influenced by the features of more distant nodes, resulting in richer feature representations.

[0044] Layer 3 Output: Following the feature extraction of the first two layers, the third layer further refines and abstracts the features. The final output is a 64-dimensional low-dimensional vector representation of all nodes, where... (64-dimensional vector). (64-dimensional vector).

[0045] Step 3: This step uses an attention mechanism to calculate the observation probability and transition probability respectively, so as to dynamically reflect the correlation strength between signaling sampling points and road segments, as well as between road segments. The specific process is as follows:

[0046] 3.1 Observation probability: For base station sampling points Its candidate road segment set First, calculate separately. and , and The attention value, i.e. and ,in This is a 64×64 attention weight matrix obtained through training and optimization. Then, the observation probabilities are calculated using the softmax function; in this embodiment, the result is... and This indicates More likely to match .

[0047] 3.2 Transition Probability: From arrive During the movement, consider Candidate road sections and The transfer situation between candidate road segments. The candidate road sections are and , The candidate road sections are and Therefore, possible transfer paths include → , → (because and (If not directly connected, the path is invalid.) → , → For valid paths → and → Calculate its attention value. and ;

[0048] Calculate the transition probability using the softmax function: , .

[0049] Step 4: This step uses an improved Viterbi algorithm to find the optimal matching path among the candidate paths. The specific process is as follows:

[0050] Based on the magnitude of the observation probability, for each base station sampling point, the top 5 road segments are retained as candidate matching road segments.

[0051] Candidate path 1 in this embodiment: → → Calculate its score as ;

[0052] Candidate path 2 in this embodiment: → → ,Score ; The scores of other candidate paths were calculated in a similar way, and their scores were all lower than those of the two paths mentioned above.

[0053] Backtracking mechanism: During the calculation process, when a certain path (such as...) is reached, the backtracking mechanism is activated. → → When the score of the current optimal path is lower than 1 / 3 of the current optimal path score (0.1512) (i.e., 0.0504), the backtracking mechanism is triggered, and the path is backtracked to... Reselect candidate road segments, due to The probability of observation is low, and this path is no longer considered as a candidate path after reselection.

[0054] Pruning strategy: Only the top 10 candidate paths with the highest scores are retained. In this example, only the two candidate paths mentioned above are retained after filtering.

[0055] After backtracking and pruning, path 1 has the highest score, and the output is the matching result, which is the user's actual path. → → .

[0056] The above embodiments illustrate the implementation process of the present invention in detail. In practical applications, parameters such as the number of candidate road segments, the number of HGCN layers, and the dimension of the attention weight matrix can be adjusted according to the specific road network complexity (such as the number of road segments, the density of intersections, etc.) and the amount of signaling data (such as the number of sampling points, the sampling frequency, etc.) to achieve a better matching effect.

[0057] Based on the same inventive concept as the above method, this application also provides a signaling matching system based on heterogeneous convolutional networks and attention mechanisms. This system includes: a data storage module, a multi-relationship graph construction module, a heterogeneous graph convolutional network module, an attention probability calculation module, a path search module, and a result output module. It should be noted that each module in the above system corresponds to a specific step of the method provided in this application embodiment, possessing the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in this application embodiment.

[0058] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the principles of this application should be included within the protection scope of this application.

Claims

1. A signaling road network matching method, characterized in that, Includes the following steps: Step 1: Construct a multi-relationship graph , where the set of nodes , For all road segments, The set of all base station sampling points; edge set It includes the edges representing the association between base station sampling points and road segments, the topological connection between road segments, and the temporal relationship between base station sampling points; Step 2: Use a Heterogeneous Graph Convolutional Network (HGCN) to process the multi-relationship graph. Node embedding is performed to obtain node feature representations. The message passing process of the HGCN is as follows: ,in For nodes In the Layer feature representation, For activation function, and For learnable weight matrix, The normalization constant is For nodes In relation The set of neighbors below; Step 3: Calculate the observation probability using the attention mechanism. and transition probability ,in For base station sampling points, , For road sections; Step 4: Perform path search using an improved Viterbi algorithm, which includes a backtracking mechanism and a pruning strategy. The path with the highest score is output as the matching result. The path score is calculated as follows: 。 2. The method according to claim 1, characterized in that, The attention mechanism calculates the observation probability. The process is as follows: ,in Sampling points for base stations Feature representation, For road section Feature representation, Sampling points for base stations The set of candidate road segments.

3. The method according to claim 1, characterized in that, The attention mechanism calculates the transition probability. , The process is as follows: ,in Sampling points for base stations The set of candidate road segments.

4. The method according to claim 1, characterized in that, The backtracking mechanism is as follows: when the current candidate path score is detected to be lower than a preset threshold, the path search is performed again by backtracking to the candidate road segment corresponding to the previous base station sampling point.

5. The method according to claim 1, characterized in that, The pruning strategy specifically involves setting a path score threshold and removing candidate paths with scores below that threshold.

6. The method according to claim 1, characterized in that, The activation function This is the ReLU function.

7. A signaling network matching device, characterized in that, include: The graph building module is used to build multi-relationship graphs. ; The feature extraction module is used to embed nodes into multi-relationship graphs using a heterogeneous graph convolutional network (HGCN); The probability calculation module is used to calculate the observation probability and transition probability by combining the attention mechanism; The path matching module is used to perform path search using an improved Viterbi algorithm and output matching results.

8. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the method according to any one of claims 1-6.

Images