New road evaluation and selection method and system based on graph neural network

Abstract

The invention discloses a new road evaluation and selection method and system based on a graph neural network, and relates to the field of road design, and the method comprises the steps: obtaining a plurality of new road construction schemes, and each scheme comprises the layout, connection mode and design parameters of a new road; modeling each newly-built road and the existing road together to form a graph structure; endowing the nodes of each graph structure with features; and utilizing the trained graph neural network evaluation model to predict the accident risk of each road section, calculating a global risk index according to the predicted accident risk, and selecting a scheme with the lowest risk. The method solves the problems that a traditional new road scheme is low in evaluation efficiency and insufficient in accuracy.

Description

Technical Field

[0001] The present invention relates to the field of road design, and more specifically, to a new road evaluation and selection method and system based on graph neural network. Background Art

[0002] Road construction occupies an important position in urban planning and infrastructure development. Scientific and reasonable road design and layout have a direct impact on traffic safety, operational efficiency and the overall development of the city. In the actual design process of new roads, it is necessary not only to consider the constraints of the starting point, end point and environment, but also to comprehensively evaluate the connection relationship with the existing road network to ensure that the new roads can be efficiently integrated with the existing road system. For example, the design parameters such as the width, radius of curvature and slope of the new road often need to be coordinated with the existing roads to avoid traffic accidents and improve the operational efficiency of the overall road network. However, the traditional road design evaluation process mostly relies on experience judgment and manual analysis, lacks systematic and automated methods, and is difficult to quickly, accurately and comprehensively evaluate multiple road plans. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a new road evaluation and selection method and system based on graph neural network to solve the problems mentioned in the background technology.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions: A new road construction plan evaluation method based on graph neural network includes the following steps: Obtain multiple new road construction plans, each plan including the layout, connection method and design parameters of the new road; Model each new road construction plan as a graph structure ; in Represents a set of road segments, including new road segments and existing road segments; Indicates the connection relationship between road segments; Assign features to each node of the graph structure. The node features of existing roads include existing features of the roads, and the node features of newly built roads include design features of newly built roads. Using the trained graph neural network evaluation model, each graph structure is processed to predict the accident risk of each road section. Based on the predicted accident risk, the global risk index of each new road construction plan is calculated; Based on the global risk index, multiple new road construction plans are compared, and the plan with the lowest risk is selected as the optimal new road construction plan.

[0005] Optionally, the method for generating a new road construction plan includes: The starting point constraint, end point constraint, environmental constraint and existing road network graph structure of road construction are used as input information; Using a graph neural network encoder to jointly encode the input information to obtain a latent space representation vector; The latent space representation vector is input into a graph structure decoder to generate multiple candidate new road plans that meet constraint conditions and can be connected to the existing road network, each plan being represented by a graph structure.

[0006] Optionally, the graph structure decoder is a graph variational autoencoder, specifically comprising: The encoding module is used to jointly encode the starting point constraint condition, the end point constraint condition, the environmental constraint condition and the existing road structure into Gaussian distribution parameters in the latent space, and use the Gaussian distribution parameters to sample the latent space representation vector; The decoding module gradually generates the nodes and connection relationships of the new roads through the latent space representation vector, forming a new road graph structure that can be topologically connected to the existing road network.

[0007] Optionally, the method further includes fusing the generated new road graph structure with the existing road graph structure into a combined graph; Using the graph neural network evaluation model to predict and evaluate the risk of the combined graph; The prediction results are used as feedback information to optimize the latent space sampling strategy or graph structure decoder parameters in the structure generation phase.

[0008] Optionally, the starting point constraint, the end point constraint and the environmental constraint include at least one or more of the following: The geographical coordinates of the starting and ending points; Maximum permissible slope; Avoidance area; Minimum turning radius; Regional planning layer information.

[0009] Optionally, the node features of the newly built road include any one or more of the following: The virtual road image is obtained by: constructing a three-dimensional road model according to road design parameters and terrain elevation data; and generating a virtual image of the road by rendering based on the three-dimensional road model; Road characteristics, including road width, road length, curvature radius, and slope.

[0010] Optionally, the node features of the existing road section include image features, road features, historical traffic flow data and historical accident data.

[0011] Optionally, the calculation formula of the global risk index is: ; in, For graph structure G The global risk indicator For graph structure G The number of nodes in For the The accident risk prediction value of each node, is the weight of the corresponding node, which is determined by the length of the road section, traffic flow and accident severity.

[0012] The present invention also discloses a new road evaluation and selection system based on graph neural network, comprising: A scheme acquisition module is used to acquire multiple new road construction schemes, each of which includes the layout, connection mode and design parameters of the new road; A graph building module, which is used to model each new road construction plan as a graph structure ; in Represents a set of road segments, including new road segments and existing road segments; Indicates the connection relationship between road segments; A feature assignment module is used to assign features to each node of the graph structure. The node features of an existing road include the existing features of the road, and the node features of a newly built road include the design features of the newly built road. The evaluation module includes a trained graph neural network evaluation model, which is used to process each graph structure, predict the accident risk of each road section, and calculate the global risk index of each new road construction plan based on the predicted accident risk; The comparison module is used to compare multiple new road construction plans based on the global risk index and select the plan with the lowest risk as the optimal new road construction plan.

[0013] Optionally, the system further includes: A structure generation module is used to generate multiple candidate new road structure plan diagrams based on the starting point constraint conditions, end point constraint conditions, environmental constraint conditions and existing road structure information of road construction; The feedback optimization module is used to use the global risk index output by the evaluation module as a feedback signal to optimize and update the latent space sampling strategy, generation model parameters or path selection strategy of the structure generation module, thereby realizing closed-loop self-optimization of structure generation and evaluation.

[0014] The advantage of the present invention over the prior art is that it effectively solves the problems of low efficiency and strong subjectivity of manual evaluation in the traditional road evaluation process. By combining the newly built roads with the existing roads to form a graph structure, and assigning specific design and existing feature information to each node (road section), the trained graph neural network model is used to automatically predict the accident risk of each road section and calculate the global risk index, which can realize the rapid comparison and evaluation of multiple new road construction plans. The present invention not only takes into account the safety of the newly built roads, but also fully considers the impact of the newly built roads on the existing roads after they are connected to the existing roads.

[0015] Furthermore, the present invention also provides a method for generating a new road structure, which uses a graph neural network to encode the starting point, end point and environmental constraints, and combines the existing road network structure information to jointly encode this information into the latent space, and automatically generates multiple candidate solutions for selection through a graph structure decoder, which greatly reduces manpower and material resources. In addition, the present invention also implements a closed-loop feedback mechanism between the structure generation and evaluation process, and feeds the evaluation results back to the generation model to optimize the sampling strategy or model parameters of the latent space, and continuously improves the rationality and practicality of solution generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a schematic diagram of the method of the present invention; Figure 2 It is a schematic diagram of a newly constructed road plan automatically generated by the method of the present invention; Figure 3 It is a simplified diagram of the structure of the present invention. DETAILED DESCRIPTION

[0017] The specific implementation of the present invention will be described below in conjunction with the accompanying drawings.

[0018] like Figure 1 The schematic diagram of the method of the present invention shown comprises the following steps: Obtain multiple new road construction plans, each plan including the layout, connection method and design parameters of the new road; Model each new road construction plan as a graph structure ; in Represents a set of road segments, including new road segments and existing road segments; Indicates the connection relationship between road segments; Each node of the graph structure is given a feature. The node features of the existing road include the existing features of the road, and the node features of the newly built road include the design features of the newly built road; Figure 3 A simplified diagram of the graph structure is shown.

[0019] Using the trained graph neural network evaluation model, each graph structure is processed to predict the accident risk of each road section. Based on the predicted accident risk, the global risk index of each new road construction plan is calculated; Based on the global risk index, multiple new road construction plans are compared, and the plan with the lowest risk is selected as the optimal new road construction plan.

[0020] Among them, the graph neural network evaluation model uses graph neural networks, which are a type of neural network model specifically used to process graph structure data. Its core idea is to propagate information and aggregate features through the connection relationship between nodes, so that each node can fuse the feature information from neighboring nodes after several layers of propagation. In the road structure evaluation application, the graph neural network first models the entire road network as a graph structure, where each node represents a road section or intersection. The node features can include road width, traffic flow, pedestrian density, historical accident data, surrounding environment image features, etc., and the edges represent the connection relationship between road sections. During the training process, the system inputs a large number of historical road networks and their corresponding accident rates as supervision labels into the model. The model learns how to predict the accident risk probability of a node based on its own characteristics and the characteristics of its neighboring nodes through the forward propagation process, and continuously adjusts the model parameters through back propagation and optimization algorithms to minimize the error between the predicted accident rate and the actual accident rate. After sufficient training, the graph neural network has a strong generalization ability and can accurately predict the accident risk of each node on the new road structure graph. Therefore, when a new road is added, the accident risk can also be re-predicted by the graph neural network due to the overall change in the road structure.

[0021] In the above embodiment, the new road construction plan can be manually selected and determined, but in other embodiments, the new road construction plan can also be automatically generated by a neural network.

[0022] like Figure 2 As shown, specifically, the method for generating a new road construction plan includes: The starting point constraint, end point constraint, environmental constraint and existing road network graph structure of road construction are used as input information; Using a graph neural network encoder to jointly encode the input information to obtain a latent space representation vector; The latent space representation vector is input into a graph structure decoder to generate multiple candidate new road plans that meet the constraint conditions and can be connected to the existing road network, each plan is represented by a graph structure, and the graph structure only contains the new road part.

[0023] Among them, in a preferred embodiment, the graph structure decoder is a graph variational autoencoder, specifically comprising: The encoding module is used to jointly encode the starting point constraint condition, the end point constraint condition, the environmental constraint condition and the existing road structure into Gaussian distribution parameters in the latent space, and use the Gaussian distribution parameters to sample the latent space representation vector; The decoding module gradually generates the nodes and connection relationships of the new roads through the latent space representation vector, forming a new road graph structure that can be topologically connected to the existing road network.

[0024] Among them, the design of the graph neural network encoder is based on the graph structure data modeling and information propagation mechanism. In the graph structure, nodes and edges together constitute an irregular data structure. Different from the regular grid input in the traditional convolutional neural network, the graph neural network relies on an adjacency-aware calculation method: each node continuously aggregates the information of its neighbor nodes during the forward propagation process to iteratively update its own representation. This process can effectively capture the spatial topological characteristics of the local structure in the graph and the complex dependencies between nodes. In the present invention, the road network naturally has graph structure characteristics. The nodes can represent the intersections or end points of the road sections, and the edges represent the connection relationship of the sections. The starting point constraints, end point constraints and environmental constraints can be expressed by embedded expressions in node features or edge weights, so it is very suitable to use the graph neural network encoder for representation learning. After sending this information into the graph neural network encoder, the encoder effectively integrates the local and global structural information through multi-layer propagation operations, thereby forming a potential space vector representation that can comprehensively reflect the road design requirements, geographical conditions and existing network structures.

[0025] In order to give the model the ability to generate diversified structures, the method of the present invention further introduces the framework of graph variational autoencoders. Unlike ordinary autoencoders that directly map to a deterministic vector, the encoding module of the graph variational autoencoder maps the input graph structure to Gaussian distribution parameters in the latent space - mean and variance vectors, from which sampling is performed to obtain the latent space representation vector. This mechanism introduces probabilistic and generalization capabilities, so that the same input conditions can also generate multiple candidate solutions that meet the requirements but have different structures. This latent vector serves as the input of the decoder and controls the overall style and distribution range of the subsequent generated graph structure.

[0026] In the decoding stage, the decoder gradually grows a graph structure that only contains the newly built road part based on the latent vector. The generation process follows the principle of topological consistency. The decoder can use recursion or attention mechanism to determine the generation position of the next node and the connection method with the existing nodes until a complete graph is formed. Since the key constraints in the input graph structure have been modeled during the training process of the latent space, the structure generated by the decoder naturally meets the connection requirements of the starting point and the end point, and will not violate the input environmental constraints. And through feature consistency, it is guaranteed to be topologically connectable with the existing road network. This idea based on probabilistic graph generation provides powerful model freedom and structural expression capabilities in principle, making the road plan design for constraint optimization not only structurally feasible, but also with good generalization ability and generation efficiency. When generating, in principle, only the newly built road part needs to be generated to improve efficiency.

[0027] In a further embodiment, the method of the present invention further comprises fusing the generated new road graph structure with the existing road graph structure into a combined graph; Using the graph neural network evaluation model to predict and evaluate the risk of the combined graph; The prediction results are used as feedback information to optimize the latent space sampling strategy or graph structure decoder parameters in the structure generation phase.

[0028] In the above embodiment, the predicted risk assessment results are used as feedback in the structure generation stage, which not only enables the generation model to learn to avoid high-risk areas and connection methods in each iteration, but also guides the generation model to output road plans with reasonable structure, stable topology and lower risk by fine-tuning the latent space sampling strategy and decoder parameters. This self-optimization mechanism that integrates generation and evaluation breaks the passive process of traditional separation of design and verification, enables the solution generation to have the ability of active optimization oriented to the goal, and constructs an intelligent closed loop of perception-generation-feedback-regeneration in principle, which significantly improves the adaptability and generation quality of the system in complex traffic planning tasks.

[0029] In a further embodiment, the starting point constraint, the end point constraint and the environmental constraint include at least one or more of the following: The geographical coordinates of the starting and ending points; Maximum permissible slope; Avoidance area; Minimum turning radius; Regional planning layer information.

[0030] In a further embodiment, the node features of the newly built road include any one or more of the following: The virtual road image is obtained by: constructing a three-dimensional road model according to road design parameters and terrain elevation data; and generating a virtual image of the road by rendering based on the three-dimensional road model; Road characteristics, including road width, road length, curvature radius, and slope.

[0031] In a further embodiment, the node features of the existing road section include image features, road features, historical traffic flow data and historical accident data.

[0032] In a further embodiment, the calculation formula of the global risk index is: ; in, For graph structure G The global risk indicator For graph structure G The number of nodes in For the The accident risk prediction value of each node, is the weight of the corresponding node. The weight is determined by the length of the road section, traffic flow and accident severity, and can also be assigned manually.

[0033] The present invention also discloses a new road evaluation and selection system based on graph neural network, comprising: A scheme acquisition module is used to acquire multiple new road construction schemes, each of which includes the layout, connection mode and design parameters of the new road; A graph building module, which is used to model each new road construction plan as a graph structure ; in Represents a set of road segments, including new road segments and existing road segments; Indicates the connection relationship between road segments; A feature assignment module is used to assign features to each node of the graph structure. The node features of an existing road include the existing features of the road, and the node features of a newly built road include the design features of the newly built road. The evaluation module includes a trained graph neural network evaluation model, which is used to process each graph structure, predict the accident risk of each road section, and calculate the global risk index of each new road construction plan based on the predicted accident risk; The comparison module is used to compare multiple new road construction plans based on global risk indicators and select the plan with the lowest risk as the optimal new road construction plan.

[0034] In an optional embodiment, the system further includes: A structure generation module is used to generate multiple candidate new road structure plan diagrams based on the starting point constraint conditions, end point constraint conditions, environmental constraint conditions and existing road structure information of road construction; The feedback optimization module is used to use the global risk index output by the evaluation module as a feedback signal to optimize and update the latent space sampling strategy, generation model parameters or path selection strategy of the structure generation module, thereby achieving closed-loop self-optimization of structure generation and evaluation.

[0035] In terms of hardware platform, the system can use high-performance computing devices, such as servers or dedicated workstations, including high-performance CPUs and high-performance GPUs to efficiently support the training and reasoning calculations of graph neural networks. In addition, the server must be equipped with sufficient network interfaces to enable data interaction with external data platforms or design departments.

[0036] In terms of software implementation, this system can be deployed based on the Linux system to ensure system stability and efficiency. The system's software architecture design mainly includes data acquisition and management subsystem, road structure modeling subsystem, graph neural network evaluation subsystem, structure scheme generation subsystem and closed-loop feedback optimization subsystem.

[0037] Specifically, the software implementation of the solution acquisition module can be developed in Python, and combined with a database to effectively store, manage and retrieve multiple new road solution data. Data sources may include planning and design documents, CAD files, GIS vector data, DEM data, etc. These data are imported in batches through the Web API interface or local files.

[0038] In terms of software implementation of the mapping module, Python language can be used in combination with graph computing libraries such as NetworkX or DGL to quickly convert road plan data into a graph structure, complete the construction of nodes and edges, and express topological relationships.

[0039] The feature assignment module is also implemented in Python. Pandas, NumPy and GeoPandas libraries can be integrated for data processing to complete the extraction and standardization of node feature data. Computer vision libraries such as OpenCV and PyTorch can be integrated for visual image feature processing to extract and process image features. Data cleaning and normalization techniques are used for preprocessing of historical traffic flow data and accident data to form a unified data input format.

[0040] The software implementation of the evaluation module can use PyTorch or TensorFlow framework to develop, train and deploy graph neural network models. After pre-training and tuning, the model uses CUDA technology to achieve efficient reasoning on the GPU, quickly calculating the accident risk of each node and the global risk index. The model evaluation results are efficiently called through the RESTful API interface.

[0041] The software implementation of the comparison module is carried out in Python, specifically using the Pandas library to sort and compare risk indicators and automatically select the plan with the lowest risk. The output of this module directly generates easy-to-interpret interactive charts (such as those generated by the Plotly library), which facilitates decision makers to conduct intuitive comparative analysis.

[0042] The structure generation module can be implemented using a deep learning framework based on a graph variational autoencoder (Graph-VAE), using the Python language with the PyTorch Geometric library to automatically generate a new road graph structure through an encoder and decoder structure.

[0043] The feedback optimization module is implemented as a feedback control module written in Python. It monitors the risk results output by the assessment module and uses reinforcement learning algorithms (such as DDPG or PPO) or optimization algorithms (such as Bayesian optimization) to dynamically adjust the latent space sampling strategy, generate model parameters or path selection strategy, thereby achieving closed-loop adaptive optimization of the model. When the system is running, the feedback optimization module regularly performs automatic parameter calibration to achieve gradual optimization and risk reduction of the generated structure.

[0044] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A new road evaluation and selection method based on graph neural network, characterized in that: The steps include: Obtain multiple new road construction plans, each plan including the layout, connection method and design parameters of the new road; Model each new road construction plan as a graph structure ; in Represents a set of road segments, including new road segments and existing road segments; Indicates the connection relationship between road segments; Assign features to each node of the graph structure. The node features of existing roads include existing features of the roads, and the node features of newly built roads include design features of newly built roads. Using the trained graph neural network evaluation model, each graph structure is processed to predict the accident risk of each road section. Based on the predicted accident risk, the global risk index of each new road construction plan is calculated; Based on the global risk index, multiple new road construction plans are compared, and the plan with the lowest risk is selected as the optimal new road construction plan.

2. The method according to claim 1, characterized in that The methods for generating new road construction plans include: The starting point constraint, end point constraint, environmental constraint and existing road network graph structure of road construction are used as input information; Using a graph neural network encoder to jointly encode the input information to obtain a latent space representation vector; The latent space representation vector is input into a graph structure decoder to generate multiple candidate new road plans that meet constraint conditions and can be connected to the existing road network, each plan being represented by a graph structure.

3. The method according to claim 2, characterized in that The graph structure decoder is a graph variational autoencoder, specifically comprising: The encoding module is used to jointly encode the starting point constraint condition, the end point constraint condition, the environmental constraint condition and the existing road structure into Gaussian distribution parameters in the latent space, and use the Gaussian distribution parameters to sample the latent space representation vector; The decoding module gradually generates the nodes and connection relationships of the new roads through the latent space representation vector, forming a new road graph structure that can be topologically connected to the existing road network.

4. The method according to claim 2, characterized in that: The method further includes fusing the generated newly constructed road graph structure with the existing road graph structure into a combined graph; Using the graph neural network evaluation model to predict and evaluate the risk of the combined graph; The prediction results are used as feedback information to optimize the latent space sampling strategy or graph structure decoder parameters in the structure generation phase.

5. The method according to claim 4, characterized in that The starting point constraint, end point constraint and environmental constraint include at least one or more of the following: The geographical coordinates of the starting and ending points; Maximum permissible slope; Avoidance area; Minimum turning radius; Regional planning layer information.

6. The method according to claim 1, characterized in that The node features of the newly built road include any one or more of the following: The virtual road image is obtained by: constructing a three-dimensional road model according to road design parameters and terrain elevation data; Generate a virtual image of the road based on the rendering of the road 3D model; Road characteristics, including road width, road length, curvature radius, and slope.

7. The method according to claim 1, characterized in that The node features of the existing road section include image features, road features, historical traffic flow data and historical accident data.

8. The method according to claim 1, characterized in that The calculation formula of the global risk index is: ； in, For graph structure G The global risk indicator For graph structure G The number of nodes in For the The accident risk prediction value of each node, is the weight of the corresponding node.

9. A new road evaluation and selection system based on graph neural network, characterized in that: include: A scheme acquisition module is used to acquire multiple new road construction schemes, each of which includes the layout, connection mode and design parameters of the new road; A graph building module, which is used to model each new road construction plan as a graph structure ; in Represents a set of road segments, including new road segments and existing road segments; Indicates the connection relationship between road segments; A feature assignment module is used to assign features to each node of the graph structure. The node features of an existing road include the existing features of the road, and the node features of a newly built road include the design features of the newly built road. The evaluation module includes a trained graph neural network evaluation model, which is used to process each graph structure, predict the accident risk of each road section, and calculate the global risk index of each new road construction plan based on the predicted accident risk; The comparison module is used to compare multiple new road construction plans based on the global risk index and select the plan with the lowest risk as the optimal new road construction plan.

10. The system according to claim 9, characterized in that The system further comprises: A structure generation module is used to generate multiple candidate new road structure plan diagrams based on the starting point constraint conditions, end point constraint conditions, environmental constraint conditions and existing road structure information of road construction; The feedback optimization module is used to use the global risk index output by the evaluation module as a feedback signal to optimize and update the latent space sampling strategy, generation model parameters or path selection strategy of the structure generation module, thereby realizing closed-loop self-optimization of structure generation and evaluation.

Images