An Emergency Logistics Route Planning Method and System Based on Edge-Enhanced Graph Attention

Abstract

This invention discloses an emergency logistics route planning method and system based on edge-enhanced graph attention, relating to the field of emergency logistics route planning. The method includes: dividing road segments into multiple micro-segments and representing them based on micro-traffic flow data and topological connectivity data; constructing a segment association graph for each road segment, with segments as nodes and traffic flow interaction relationships as edges; fusing conflict information using a graph convolutional network layer to update segment features; subsequently, calculating risk indicators for the road segments based on the segment feature set and dynamically optimizing edges with abnormal indicators; finally, determining the computational granularity mode of each edge in the graph attention network based on the optimized indicators, and modulating the attention mechanism accordingly to generate the final emergency logistics route. This invention achieves refined perception and dynamic modeling of risk heterogeneity within road segments, improving the reliability and timeliness of emergency logistics route planning.

Description

Technical Field

[0001] This invention relates to the field of emergency logistics route planning technology, and more specifically, this application relates to an emergency logistics route planning method and system based on edge-enhanced graph attention. Background Technology

[0002] Emergency logistics route planning, a crucial component of modern logistics and disaster response, aims to deliver supplies to their destination quickly and safely. In recent years, graph convolutional network layers, particularly graph attention networks, have been widely applied to traffic network representation and route decision-making due to their powerful topological relationship modeling capabilities. Existing methods typically abstract road networks as graph structures composed of nodes and edges, and evaluate the overall traffic flow status of a route by aggregating node information.

[0003] However, most current research tends to focus on information aggregation at intersection nodes or treat the entire road segment as a homogeneous whole. While this approach offers efficiency advantages at the macro level, it implicitly assumes a crucial premise: that traffic conditions and risk distribution within a road segment are uniform. In reality, under actual emergency scenarios, due to the dynamic changes and complex interactions of micro-traffic flows, traffic density, speed, and the degree of traffic convergence conflicts often vary significantly across different sections of a single road segment. This heterogeneity of internal states means that the traditional method of treating the entire road segment as a single feature vector inevitably masks potential local high-risk areas within it.

[0004] Specifically, if the uneven distribution of risks within road segments is ignored, the edge weights or features used in the route planning model will only be a rough average. This could lead the system to choose a route that is generally acceptable but contains severe congestion or conflict bottlenecks, causing emergency vehicles to experience unpredictable local delays in actual traffic, ultimately deviating from the stringent time reliability requirements of emergency logistics. Therefore, the key issue is how to break through the inherent paradigm of treating road segments as overall analytical units and design a modeling mechanism that can accurately perceive and quantify the heterogeneity of micro-risks within road segments and effectively integrate it into global route decision-making. Summary of the Invention

[0005] To address the aforementioned technical problems, this paper provides an emergency logistics route planning method and system based on edge-enhanced graph attention. This technical solution solves the problems mentioned in the background section.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] Firstly, this application provides an emergency logistics route planning method based on edge-enhanced graph attention, including:

[0008] Acquire micro-level traffic flow data and topological connectivity data for each edge of the target logistics network;

[0009] Based on the micro-traffic flow data and topological connection data, each edge is divided into segments to obtain and output multiple micro-segments, and the initial feature vector of each micro-segment is calculated.

[0010] A paragraph association graph is constructed based on the micro-segments, with micro-segments as nodes and traffic flow interaction relationships as association edges, and association edge features characterizing the intensity of traffic flow conflicts are calculated;

[0011] The paragraph association graph and associated edge features are input into the graph convolutional network layer to update the initial feature vector and output the paragraph feature vector that incorporates conflict information.

[0012] Construct a set of paragraph feature vectors based on the paragraph feature vectors of each side and calculate the risk index;

[0013] When the risk index is greater than or equal to the preset risk threshold, the optimization sub-process is executed: identify abnormal paragraph clusters and refine them to obtain refined paragraphs, retain the paragraph feature vectors of the micro paragraphs and calculate the initial feature vectors of the refined paragraphs, locally reconstruct the paragraph association graph and perform a limited number of rounds of feature propagation to recalculate the risk index.

[0014] The recalculated risk index is compared with the risk threshold. If it is still greater than or equal to the threshold, the optimization sub-process is repeated within the maximum number of iterations. Otherwise, the final risk index is output.

[0015] The granularity mode of graph attention network computation is determined based on the final risk indicators of each side, and the attention computation is modulated according to the granularity mode to output the emergency logistics route planning results.

[0016] Secondly, this application provides an emergency logistics route planning system based on edge-enhanced graph attention, used to implement the aforementioned emergency logistics route planning method based on edge-enhanced graph attention, including:

[0017] The data acquisition module is used to acquire micro-traffic flow data and topology connection data of each side in the target logistics network;

[0018] The segmentation and representation module is used to segment each edge based on the micro traffic flow data and topological connection data, obtain and output multiple micro segments, and calculate the initial feature vector of each micro segment.

[0019] The association graph construction module is used to construct a paragraph association graph based on the micro-segments, with micro-segments as nodes and traffic flow interaction relationships as association edges, and to calculate association edge features that characterize the intensity of traffic flow conflicts.

[0020] The feature fusion and update module is used to input the paragraph association graph and the associated edge features into the graph convolutional network layer, update the initial feature vector, and output the paragraph feature vector that fuses conflict information.

[0021] The risk analysis module is used to construct a set of paragraph feature vectors based on the paragraph feature vectors of each side and to calculate risk indicators;

[0022] The dynamic optimization module is used to execute an optimization sub-process when the risk index is greater than or equal to the preset risk threshold: identify abnormal paragraph clusters and refine them into refined paragraphs, retain the paragraph feature vectors of the micro paragraphs and calculate the initial feature vectors of the refined paragraphs, locally reconstruct the paragraph association graph and perform a limited number of rounds of feature propagation to recalculate the risk index.

[0023] It is also used to compare the recalculated risk index with the risk threshold. If it is still greater than or equal to the risk threshold, the optimization sub-process is repeated within the maximum number of iterations; otherwise, the final risk index is output.

[0024] The granular decision and path planning module is used to determine the granularity mode of graph attention network calculation based on the final risk index of each edge, and modulate the attention calculation according to the granularity mode to output the emergency logistics path planning result.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] This application divides road segments into internal segments using micro-traffic flow data and constructs a segment association graph with segments as nodes and traffic flow interactions as associations. It then uses graph convolutional network layers to fuse conflict information, solving the problem of treating the entire road segment as a homogeneous unit and masking the heterogeneity of its internal risk distribution. This enables refined perception and representation of the micro-traffic status and potential conflict bottlenecks within the road segment.

[0027] This application calculates risk indicators based on the set of paragraph feature vectors and optimizes the sub-process for edge triggering of abnormal indicators. By identifying and refining abnormal paragraph clusters, locally reconstructing the association graph and performing feature updates in a limited number of rounds, it solves the modeling bias caused by the indiscriminate uniform analysis of complex road sections in traditional methods, realizes dynamic, adaptive and accurate modeling of high-risk road sections, and improves the reliability of status assessment. Attached Figure Description

[0028] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Wherein:

[0029] Figure 1 This is a flowchart of the emergency logistics route planning method based on edge-enhanced graph attention proposed in this invention;

[0030] Figure 2This is a schematic diagram of the emergency logistics route planning system based on edge-enhanced graph attention in this invention. Detailed Implementation

[0031] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0032] In existing technologies, emergency logistics route planning suffers from the problem of treating the entire route as a homogeneous whole and ignoring the heterogeneity of risk distribution within it. Traditional methods typically make decisions based on the average characteristics of the entire route, often failing to perceive and quantify high-risk bottlenecks such as congestion and conflicts in localized areas within the route. This results in planned routes that perform well in macro-level assessments but may experience severe delays in actual travel due to localized congestion, making it difficult to meet the stringent time reliability requirements in emergency scenarios.

[0033] To address the aforementioned issues, this application divides road segments into internal segments, constructs a segment association graph with segments as nodes and traffic flow interactions as associated edges, and utilizes graph convolutional network layers to fuse conflict information, achieving refined perception of the heterogeneity of micro-risks within each road segment. Furthermore, risk indicators for road segments are calculated based on segment feature sets, and an optimization sub-process is dynamically executed for edges with abnormal indicators: identifying and refining anomalous segment clusters within them, and performing a limited number of rounds of feature updates and indicator recalculation based on the locally reconstructed segment association graph, achieving focused and adaptive modeling of high-risk road segments. Finally, the computational granularity is determined based on the final risk indicators of each edge, and a graph attention network is modulated to generate path planning results. By introducing segment-level analysis, dynamic refinement of anomalous regions, and adaptive granularity modulation, this application solves the problem of the "black box" handling of risk distribution within road segments in traditional methods, improving the accuracy and reliability of emergency logistics path planning.

[0034] Example 1:

[0035] like Figure 1 As shown, this application introduces an emergency logistics route planning method based on edge-enhanced graph attention, including:

[0036] S1. Obtain micro-level traffic flow data and topological connection data for each edge of the target logistics network;

[0037] The target logistics network is abstracted as a set of nodes. Sum of edges The graph structure formed , where an edge represents a road segment connecting two nodes.

[0038] Parallel acquisition of two types of edge data:

[0039] (1) Micro-traffic flow data: refers to data reflecting the traffic flow of a road segment during the planning period. Real-time dynamic operational status data. This data is collected and transmitted to a data processing center via roadside traffic detection equipment (such as inductive loops and microwave radar) or mobile detection sources (such as floating car GPS). For any side... Its micro-traffic flow data at least includes the edge during the cycle Information on the average vehicle speed, vehicle density, and variance of vehicle acceleration within the area.

[0040] (2) Topology connectivity data: refers to static data describing the inherent geometric structure and connectivity relationships of the road network. It is obtained from digital map databases or road infrastructure management systems. For any side... Its topology connection data includes at least: the length of the edge, the number of lanes, and a series of key point information distributed along the edge; key points include intersections, lane number change points, ramp merging points and diverging points; it also includes the connection relationship of the edge in the network, that is, the identifiers of its upstream and downstream adjacent edges.

[0041] After obtaining the raw data, perform the following preprocessing to form a regularized input:

[0042] Data cleaning: Identify and remove outliers in micro-traffic flow data, and imput missing values ​​using the average of adjacent time periods or the average of the same period in history.

[0043] Spatiotemporal alignment: Unifying all micro-level traffic flow data to the same planning period. They are associated and bound to the corresponding topology connection data records at the same time granularity through edge identifiers.

[0044] The output is a series of structured data records. Each record corresponds to an edge. It includes its associated, preprocessed micro-traffic flow data objects and topological connectivity data objects.

[0045] S2. Based on micro-traffic flow data and topological connection data, each edge is divided into segments to obtain and output multiple micro-segments, and the initial feature vector of each micro-segment is calculated.

[0046] The process of step S2 specifically includes:

[0047] Obtain key point information of road structure in topology connection data. Key point information includes intersection location, lane number change point, ramp merging point and diverging point;

[0048] Based on key point information, each edge is divided into multiple continuous micro segments, so that the road structure inside each micro segment remains homogeneous and the boundary of the micro segment is located at the key point.

[0049] For each micro-segment, statistical features, including the average vehicle speed, vehicle density, and vehicle acceleration variance of that micro-segment, are extracted from the micro-traffic flow data flowing through that micro-segment to form an initial feature vector.

[0050] From the current edge The key point sequence is read from the topology connection data. The key points include intersections, lane number change points, ramp merging points and diverging points, and the distance of each key point from the starting point of the edge is recorded.

[0051] Using adjacent key points as boundaries, the edges are... It is divided into multiple continuous micro-segments. For example, if the keypoint distance sequence is... Then we get Paragraph 1, No. The scope of the paragraph is After dividing the data, the length of each segment is determined; if a segment is too short (e.g., less than 50 meters), it is merged with the adjacent segment to ensure the stability of subsequent calculations.

[0052] For each micro-segment obtained from the segmentation, traffic flow statistical features flowing through that spatial range are extracted from the micro-traffic flow data associated with that edge, forming an initial feature vector. Statistical features should include at least: average vehicle speed, vehicle density, and variance of vehicle acceleration within the segment. Feature extraction requires mapping the micro-traffic flow data to the corresponding segment spatial range based on the detector location or vehicle trajectory points for calculation.

[0053] Finally, output the current edge. The set of micro-segments and the initial feature vector corresponding to each segment.

[0054] By using a micro-segmentation method based on key road points, this approach addresses the problem of neglecting internal structural heterogeneity in traditional traffic analysis due to the overall treatment of the entire road. It achieves accurate segmentation of homogeneous road segments and provides a fine-grained, highly representative data foundation for subsequent traffic modeling and optimization by extracting multi-dimensional traffic flow statistical features.

[0055] S3. Construct a paragraph association graph based on micro-segments, with micro-segments as nodes and traffic flow interaction relationships as association edges, and calculate the association edge features that characterize the intensity of traffic flow conflicts.

[0056] The process of calculating the associated edge features that characterize the intensity of traffic flow conflicts specifically includes:

[0057] For each association edge in the paragraph association graph, it connects an upstream paragraph node and a downstream paragraph node;

[0058] Obtain the output traffic flow of the upstream segment node and the remaining traffic capacity of the downstream segment node;

[0059] The formula for calculating the intensity of traffic flow conflict is: ;

[0060] in, Indicates the intensity of traffic conflict. This indicates the output traffic flow. Indicates the remaining throughput capacity. It is a very small normal value used to prevent division by zero errors.

[0061] The set of micro-segments of each side and their initial feature vectors output from step S2 are received as input.

[0062] For the current edge Construct a paragraph association graph. Treat each micro-segment within the edge as a node. Based on traffic flow direction and spatial adjacency, establish directed association edges between nodes: if paragraphs... downstream endpoints and paragraphs The upstream endpoints coincide, and the traffic flow is from Flow direction Then create a path consisting of nodes Pointing to node The associated edges.

[0063] Calculate the traffic conflict intensity characteristics for each associated edge. For nodes... Pointing to node For the associated edges, perform the following calculation:

[0064] Get upstream node Output traffic volume of the corresponding segment Traffic flow is extracted or estimated from the micro-traffic flow data associated with that edge.

[0065] Get downstream paragraphs Remaining throughput capacity It calculates the theoretical throughput of the section. With current input traffic flow The difference is obtained, that is The theoretical traffic capacity is determined based on the number of lanes and road grade information for that section. Indicates downstream node The input traffic flow for the corresponding paragraph;

[0066] Based on the above parameters, the conflict strength of the associated edge is calculated using the following formula. : ;

[0067] in, It is a very small normal value used to prevent division by zero errors.

[0068] Output the current edge The corresponding paragraph relationship graph. This graph uses each micro-segment as a node and the traffic flow interaction between segments as directed edges. Each relationship edge is accompanied by a calculated conflict intensity value. .

[0069] By constructing a segment association graph with micro-segments as nodes and traffic flow interaction relationships as edges, this application solves the problem of difficulty in quantifying and assessing the intensity of traffic flow conflicts in traditional traffic analysis. For the dynamic interaction between upstream and downstream segments, it reflects the source of traffic pressure through the positive correlation of upstream output traffic flow and characterizes the capacity limitation by combining the negative correlation of downstream remaining capacity. At the same time, based on the comprehensive calculation of flow and capacity, it ensures the accurate characterization of conflict intensity and provides data support for traffic flow optimization and congestion early warning.

[0070] S4. Input the paragraph association graph and the associated edge features into the graph convolutional network layer, update the initial feature vector, and output the paragraph feature vector that incorporates conflict information.

[0071] The process of step S4 specifically includes:

[0072] The initial feature vectors of each node in the paragraph association graph and the features of each associated edge are input into a graph convolutional network layer.

[0073] Through graph convolutional network layers, the initial feature vector of each node is aggregated and nonlinearly transformed with the initial feature vectors of its neighboring nodes and the associated edge features of the connecting edges;

[0074] Output the updated initial feature vector of each node as the paragraph feature vector for fusing conflict information.

[0075] The input consists of the initial feature vectors of each segment output from step S2 and the segment association graph and associated edge features output from step S3. The purpose is to use graph convolutional network layers to enable each segment's features to not only include its own traffic state, but also to fuse traffic conflict information from adjacent segments and connecting edges, thereby obtaining more comprehensive and representative segment features.

[0076] For the edge being processed, obtain its paragraph association graph. Each node in the diagram (corresponding to a micro-segment) is accompanied by an initial feature vector from step S2. Each directed associated edge is attached with the traffic conflict intensity from step S3. .

[0077] The above graph data is input into a Graph Convolutional Network (GCN) layer for feature propagation and updating. The specific operation is as follows:

[0078] Feature aggregation: For each node in the graph (Corresponding paragraph) This layer aggregates its own features, the features of all direct neighbor nodes (i.e., upstream and downstream segments directly connected by associated edges), and the conflict intensity features of the associated edges connecting these neighbors.

[0079] Transformation and Activation: The aggregated information is linearly transformed using a learnable weight matrix, and a non-linear activation function (such as ReLU) is introduced to generate the updated feature representation for that node. This process can be abstractly represented by the following formula:

[0080] ;

[0081] in, Represents a node Updated feature vectors This represents the activation function. Represents the weight parameters. Represents a node The initial feature vector, Represents a node The initial feature vector, Represents a node With nodes The intensity of traffic conflict between the related edges. Represents a node The set of neighboring nodes, This represents aggregate functions (such as mean aggregation, weighted summation, etc.).

[0082] During this process, the norm or numerical range of the updated feature vector can be checked. If extreme values ​​(such as gradient explosion) occur, stabilization techniques such as gradient clipping can be enabled.

[0083] Finally, the graph convolutional layer outputs an updated feature vector for each node in the graph. This vector is the segment feature vector that integrates conflict information. It encodes the traffic state of the segment itself and the conflict impact between it and adjacent segments caused by traffic flow interaction.

[0084] By utilizing segment association graphs and traffic conflict intensity features, this application addresses the problem in traditional methods of isolated road segment traffic features and the failure to effectively integrate traffic conflict and topological relationship information. Through graph convolutional network layers, it aggregates the features of each segment with those of adjacent segments and comprehensively models the traffic conflict intensity along connecting edges. Furthermore, nonlinear transformations achieve deep information fusion, ultimately yielding segment feature vectors that comprehensively reflect the road segment's own state and its dynamic impact from upstream and downstream traffic conflicts. This enables effective modeling of the dynamic interaction effects of road segments, enhancing the feature representation ability and predictive value.

[0085] S5. Construct a set of paragraph feature vectors based on the paragraph feature vectors of each side and calculate the risk index;

[0086] The process of calculating the risk indicators for each side specifically includes:

[0087] Based on the paragraph feature vectors of all micro-segments within the edge, calculate the local risk score for each micro-segment;

[0088] The mean of the local risk scores within the edge is calculated based on the local risk scores, and each local risk score is normalized.

[0089] The risk index of an edge is calculated based on the dispersion of the normalized local risk score. The risk index is used to characterize the unevenness of the risk distribution within the edge.

[0090] The set of updated feature vectors for all segments within each edge, output from step S4, is received as input.

[0091] For the current edge First, the local risk score of each paragraph is calculated based on the updated feature vectors of each paragraph. The score is obtained through a mapping function, such as calculating the norm of the eigenvectors, to obtain a scalar value that characterizes the relative risk status of the paragraph.

[0092] The specific implementation of the rating mapping function is as follows: Vector norm-based mapping function: The "intensity" or "activity" of the risk status is indirectly represented by calculating the norm (length) of the paragraph feature vector.

[0093] Local risk scoring is calculated using the Euclidean norm, and the formula is as follows:

[0094] ;

[0095] in, This indicates a local risk score. Let represent the paragraph feature vector of micro-paragraph j, and D represent the dimension of the paragraph feature vector. This represents the d-th element of the feature vector of paragraph j.

[0096] Calculate the mean of the local risk scores, and normalize each local risk score based on historical values ​​using the min-max method;

[0097] Risk indicators The calculation formula is: ;

[0098] in, This represents the total number of segments contained within the edge. This represents the j-th normalized local risk score. This represents a risk indicator used to characterize the degree of unevenness in the risk distribution within the edge.

[0099] By calculating the risk index of edges based on the feature vector set of segments, this application solves the problem that traditional risk assessment methods only focus on the overall risk level of road segments while ignoring their internal risks. For the risk differences within each edge, it quantifies the relative risk of each segment through local risk scoring and comprehensively assesses the risk situation within all roads using risk indices, providing a scientific basis for accurately identifying high-risk road segments and optimizing traffic management strategies.

[0100] S6. When the risk index is greater than or equal to the preset risk threshold, execute the optimization sub-process: identify abnormal paragraph clusters and refine them into refined paragraphs, retain the paragraph feature vectors of the micro paragraphs and calculate the initial feature vectors of the refined paragraphs, locally reconstruct the paragraph association graph and perform a limited number of rounds of feature propagation to recalculate the risk index.

[0101] The recalculated risk index is compared with the risk threshold. If it is still greater than or equal to the threshold, the subprocess is repeatedly optimized within the maximum number of iterations. Otherwise, the final risk index is output.

[0102] The process of identifying anomalous paragraph clusters and refining them into detailed paragraphs specifically includes:

[0103] Micro-segments with local risk scores greater than or equal to a preset local risk threshold are marked as high-risk segments.

[0104] Multiple high-risk paragraphs that are spatially adjacent are merged to form an anomalous paragraph cluster;

[0105] For aberrant paragraph clusters, secondary divisions are performed within their corresponding spatial ranges to generate multiple refined paragraphs with smaller spatial scales, forming a refined paragraph set.

[0106] The process of partially reconstructing the paragraph association graph and performing a limited number of rounds of feature propagation to recalculate risk indicators specifically includes:

[0107] Replace the nodes belonging to the abnormal paragraph clusters in the original paragraph association graph with the nodes in the corresponding refined paragraph set;

[0108] Based on the topological connection data, refined connection edges are established between paragraph nodes and between them and other retained nodes in the original paragraph connection graph, forming a locally reconstructed paragraph connection graph.

[0109] Using the initial feature vector of the refined paragraph and the paragraph feature vector that preserves the micro-paragraph as input, the graph convolutional network layer feature propagation update is performed no more than a preset number of times on the locally reconstructed paragraph association graph;

[0110] Based on the updated paragraph feature vector, the local risk score of all paragraphs on that edge is recalculated, and a new risk index for that edge is calculated accordingly.

[0111] Based on the risk indicators of each side calculated in step S5 Perform the operation. For those that meet the requirements... The edge ( If the internal risk is determined to be significant (based on a preset threshold), the following optimization process is triggered:

[0112] Identify and refine anomalous paragraph clusters: based on local risk scores for each paragraph along the edge. The score will be greater than the preset local risk threshold. Paragraphs are marked as high-risk paragraphs. Spatially adjacent high-risk paragraphs are merged to form anomalous paragraph clusters. Subsequently, a finer-grained secondary division is performed within the spatial range corresponding to each anomalous paragraph cluster, generating multiple refined paragraphs that constitute a refined paragraph subset; the remaining paragraphs constitute the original paragraph set.

[0113] Local reconstruction and feature recalculation:

[0114] Graph Reconstruction: Replace the nodes of the corresponding abnormal clusters in the original paragraph association graph with nodes from the refined paragraph subset, and reconstruct the association edges between nodes according to the topological connection relationship to form a locally reconstructed paragraph association graph.

[0115] Feature preparation and update: For refined segments, features are extracted from the original micro-traffic flow data based on their spatial range, and their initial feature vectors are calculated. The initial feature vectors of the refined segments are input together with the existing feature vectors of the original segments, and a limited number of rounds (e.g., 1-2 times) of graph convolutional network layer feature propagation are performed on the reconstructed graph to obtain the updated feature vectors of all segments.

[0116] Indicator Recalculation: Based on the updated paragraph feature vector, recalculate the local risk scores of all paragraphs along the edge, and derive new risk indicators accordingly. .

[0117] The decision to continue with the next round of refinement and recalculation is based on the recalculated risk index. The iterative optimization of the target edge terminates and the final risk index for that target edge is output when any of the following termination conditions are met:

[0118] Termination condition met: When the recalculated risk index is less than the preset risk threshold, the risk distribution of the target edge is considered to have met the preset requirements, and the iterative optimization is terminated.

[0119] Iteration round limit termination condition: When the number of iterations reaches the preset maximum number, the iteration optimization is terminated to avoid uncontrollable computational overhead caused by infinite refinement;

[0120] After terminating the iterative optimization, the risk indicator corresponding to the current round is used as the final risk indicator for the target edge.

[0121] By designing optimized sub-processes for high-risk road sections, this application addresses the problems of traditional methods, such as coarse-grained analysis of high-risk areas and difficulty in accurately locating internal risk sources. For identified clusters of abnormal segments, it not only locks down core risk areas by merging adjacent high-risk segments but also implements secondary segmentation within these areas to improve analytical resolution. Simultaneously, through local reconstruction of the correlation graph and limited-round feature updates, it achieves refined modeling and reassessment of high-risk areas while preserving the original analytical framework. This ensures the accuracy of risk indicators and provides a reliable basis for developing precise local traffic intervention measures.

[0122] The process of setting preset risk thresholds specifically includes:

[0123] Obtain the topological importance data for each edge in the target logistics network. The topological importance data includes betweenness centrality and historical average flow carrying capacity.

[0124] Based on topological importance data, calculate the personalized threshold corresponding to each edge;

[0125] The calculated personalized threshold is used as the preset risk threshold for the corresponding edge.

[0126] Set a personalized risk indicator threshold (i.e., a preset threshold) for each edge in the logistics network. This threshold is used to dynamically determine whether to trigger an optimization subprocess. It is not globally fixed, but rather adaptively adjusted based on the network topology importance of the edges.

[0127] Topological importance data acquisition: For each edge in the target logistics network Obtain its two topological importance data:

[0128] Betweenness centrality This metric measures the importance of edges in network connectivity. It calculates the shortest path between all node pairs by analyzing the topological connectivity data (relationships between nodes and edges) of the entire road network and statistically analyzes the edges. The number of times the path is traversed by these shortest paths, after normalization, is obtained. . The higher the value, the more critical the edge is on.

[0129] Historical average flow carrying capacity ratio This metric measures the historical busyness of an edge. Edges are obtained from a historical traffic flow database. Average traffic volume at multiple historical time periods (such as the morning rush hour during the same period in the past month) And in combination with the design capacity of this side (The ratio can be calculated from the number of lanes, lane class, etc. in the topology data.) . The higher the value, the heavier the historical load on that side.

[0130] Calculate personalized thresholds:

[0131] Normalization: due to and The units and scope may differ; firstly, consider all edges in the network. and Normalize each value to ensure it falls within the interval [0, 1], and obtain... and .

[0132] Overall Importance Score: Calculate the overall topological importance score for each edge. One feasible approach is to take the weighted maximum or average of the two, for example: or ,in , To adjust the weights (all can be set to 1).

[0133] Mapping to personalized thresholds: Based on the principle that "the more important the edge, the lower the threshold should be for more sensitive monitoring," the overall importance score is... Mapped to a personalized threshold Set a global baseline threshold range. (e.g., [0.8, 1.5]). The mapping relationship can use linear or nonlinear decay functions, for example: .in this way, The higher the value (the more important the edge), the more accurate the calculation. The lower.

[0134] Check all calculated Are all values ​​within a reasonable range (e.g., greater than 0 and not exceeding the theoretical maximum value)? Correct outliers, for example, by truncating to boundary values.

[0135] Output a mapping table or data structure containing each edge in the network. Unique identifiers and their corresponding personalized preset thresholds .

[0136] By dynamically setting personalized thresholds based on topological importance data, this application solves the resource misallocation and risk sensitivity imbalance problems caused by the use of uniform static thresholds in traditional methods. Addressing the differences in importance among different road segments in the logistics network, it reflects the criticality of the global structure through betweenness centrality and characterizes the actual operational load by combining historical average traffic carrying capacity ratio. Furthermore, it dynamically calculates thresholds based on the rule that higher importance corresponds to lower thresholds, setting stricter risk trigger standards for core road segments. This enables focused monitoring and priority optimization of critical road segments, ensuring efficient and accurate allocation of risk management resources.

[0137] S7. Determine the granularity mode of graph attention network calculation based on the final risk indicators of each side, and modulate the attention calculation according to the granularity mode to output the emergency logistics route planning results.

[0138] Based on the final risk index of each edge, the specific granularity mode for calculating graph attention network for each edge includes:

[0139] Based on the final risk indicators of each side This determines the computational granularity pattern of the graph attention network in subsequent processes.

[0140] In practice, two thresholds are preset. and ( ),in accordance with The value divides each edge into three computational granularity modes:

[0141] Fine-grained mode: If This indicates that the risk within the edge is significant, and the decision is made by treating each micro-segment of the edge as an independent unit to participate in the subsequent graph attention calculation.

[0142] Medium-grained mode: If The decision is to cluster the segments within the edge and use sub-clusters (each sub-cluster contains several adjacent segments) as representative units to participate in the calculation.

[0143] Coarse-grained mode: If This indicates that the risk within the edge is low, and the decision is to aggregate the features of all segments of the edge into a comprehensive feature vector of the edge as a representative unit for calculation.

[0144] Output a granular pattern decision list, specifying the pattern used for each edge and its corresponding feature data set.

[0145] By employing a granular computation mode based on a dynamic decision graph attention network of risk indicators, this application addresses the problem that traditional fixed-granularity graph computation models cannot adapt to the differences in risk within road segments. For edges with different risk levels, a finer-grained computation mode is used for high-risk road segments to capture local high-risk details, while a coarser-grained mode is used for low-risk road segments to improve overall computational efficiency. At the same time, through adaptive switching of granularity modes, the allocation of computational resources in the model is ensured to accurately match the actual risk of the road segments, thereby optimizing global computational performance while ensuring analytical accuracy.

[0146] After determining the granularity pattern, and before modulating the attention computation according to the granularity pattern, the following steps are also included:

[0147] Obtain the risk indicators for each edge over multiple consecutive planning periods in history to form a historical indicator sequence;

[0148] Based on historical indicator sequences, predict the expected values ​​of indicators for each side in the current period, and calculate the deviation between the risk indicators and the expected values ​​of indicators for each side in the current period.

[0149] If the deviation of a certain edge exceeds a set range, a decay factor is introduced for that edge when modulating the graph attention network. The decay factor is used to reduce the weight of the feature vector of the current period segment of that edge in the attention calculation.

[0150] For each edge To obtain its historical continuity Risk indicator sequence within a planning cycle and obtain the current cycle indicator. (Right now Based on historical data, predict the expected value of the current cycle indicator. For example, a weighted moving average method can be used for calculation. Then, the deviation between the current actual indicator and the expected value is calculated. .

[0151] If the deviation of a certain edge Exceeding the preset threshold If the state is abnormal, the decision is to introduce a decay factor less than 1 onto that edge. (e.g., 0.5); otherwise, set =1.0. The decay factor is used to reduce the weight of the current periodic eigenvector of this edge in the graph attention calculation.

[0152] By introducing a dynamic deviation detection and attenuation mechanism based on historical indicator sequences, this application addresses the problem of abnormal road segments caused by sudden changes in traffic flow states excessively influencing the graph attention network and interfering with the overall stability of the model. For the temporal fluctuations of road segment risk indicators, it both predicts reasonable expected values ​​for the current period using historical data to identify abnormal deviations and applies attenuation factors to road segments with excessive deviations to suppress their weight in attention calculations. This dynamic modulation process is integrated into the calculation flow after granular decision-making, ensuring that the graph attention network can capture normal inter-road segment interactions while effectively filtering sudden abnormal interferences in complex dynamic traffic environments, thereby improving the model's robustness and prediction accuracy.

[0153] The attention calculation process of the granularity pattern modulated graph attention network is followed, and emergency logistics route planning results are generated based on its output, specifically including:

[0154] The input includes: a granularity pattern decision list, which defines the computational unit (segment, sub-cluster, or whole edge) for each edge in the graph attention network and its corresponding set of feature vectors; and a macroscopic topology diagram of the target logistics network. , where nodes Represents an intersection, edge Representative road section.

[0155] Based on the computational granularity pattern of each side, the input graph structure and attention computation process of the modulated graph attention network are adjusted to ultimately generate the optimal or reliable emergency logistics path from the origin of the supplies to the destination.

[0156] Load the granularity pattern decision list and the decay factor mapping table for each original edge in the network. Based on its granularity pattern, obtain the feature representation of its participation in the graph attention network computation:

[0157] Fine-grained mode: Obtain the feature vectors of all paragraphs under this edge. .

[0158] Medium-granularity mode: Obtain the representative feature vectors of each sub-cluster divided by this edge. .

[0159] Coarse-grained mode: Obtain the comprehensive feature vector of this edge. .

[0160] Construct the modulated network graph:

[0161] Graph structure modulation: Based on the granularity pattern, the original macroscopic topological graph... By performing "expansion" or "contraction" processing, a modulated graph is constructed for graph attention network computation. .

[0162] For edges using the coarse-grained mode, in The edge remains as a connection to the original terminating node, and its associated feature is the composite feature vector. .

[0163] For edges using fine-grained or medium-grained patterns, "virtual nodes" need to be introduced inside the edge. Specifically, in the original graph... Between the two nodes connected by the edge, a new set of virtual nodes is inserted to represent its segment or sub-cluster. New edges are established between these virtual nodes and between them and the original endpoints based on the spatial adjacency relationships of the original segment / sub-cluster, forming a "micro-path". The features attached to these virtual nodes are the feature vectors of their segments or sub-clusters. This is equivalent to unfolding the original edge into a micro-subgraph in the computational graph.

[0164] Modulation attention computation and path generation:

[0165] Attention calculation: the modulated image The features of nodes (including virtual nodes) are input into a multi-layer graph attention network. The attention mechanism in the network learns to compute the attention weights between any two connected nodes in the graph. For nodes from the same original edge but in different granularity modes, the network adaptively assigns different importance based on their features and connectivity.

[0166] Edge weight / node cost derivation: The final layer of a graph attention network outputs a hidden state representing the passage status or risk for each node (including virtual nodes). For macroscopic topological graphs... Each original edge in Its final passage cost is aggregated by its corresponding The hidden states of all nodes (i.e., their internal virtual nodes and their endpoints) are obtained. For example, the weighted sum or maximum value of the hidden states of all nodes on the path can be taken.

[0167] Path search: Based on all calculated original edges (the cost of passage), in the macro topology diagram Run a path planning algorithm (e.g., the K-shortest path algorithm that takes into account time reliability or the random most reliable path algorithm) to calculate one or more optimal emergency logistics paths from a specified origin to an end point.

[0168] Check whether the attention weight distribution of the graph attention network output is reasonable (e.g., whether there are edges or nodes with excessively concentrated weights), and whether the final calculated path cost is within a reasonable historical range. If an anomaly is found, log it and consider enabling an alternative routing strategy.

[0169] Output the final emergency logistics route planning results. The results should include at least: the recommended optimal route sequence, such as: [start point, segment e1, intersection V1, segment e2, ..., end point]; the overall estimated travel time or comprehensive risk score for the route.

[0170] By adaptively modulating the attention calculation process of the graph attention network based on the granularity pattern of dynamic decision-making, this application solves the problem of traditional static path planning models struggling to balance computational accuracy and efficiency when dealing with complex and non-uniform road risk environments. Targeting the risk distribution characteristics of different road segments, the application modulates the attention calculation process to enable the network to focus on the road segments and interaction information most relevant to the current granularity. Furthermore, it generates the final emergency logistics route based on the network's accurate output, achieving a balance between in-depth analysis of road network risk details and optimized global planning efficiency. This provides highly accurate and feasible optimal route planning results for emergency logistics.

[0171] Example 2:

[0172] like Figure 2 As shown, this application provides an emergency logistics route planning system based on edge-enhanced graph attention, used to implement the aforementioned emergency logistics route planning method based on edge-enhanced graph attention, including:

[0173] The data acquisition module 100 is used to acquire micro-traffic flow data and topology connection data of each side in the target logistics network;

[0174] The segmentation and representation module 200 is used to segment each edge based on micro traffic flow data and topological connection data, obtain and output multiple micro segments, and calculate the initial feature vector of each micro segment.

[0175] The association graph construction module 300 is used to construct a paragraph association graph based on micro-segments, with micro-segments as nodes and traffic flow interaction relationships as association edges, and to calculate association edge features that characterize the intensity of traffic flow conflicts.

[0176] The feature fusion and update module 400 is used to input the paragraph association graph and the associated edge features into the graph convolutional network layer, update the initial feature vector, and output the paragraph feature vector that integrates conflict information.

[0177] Risk analysis module 500 is used to construct a set of paragraph feature vectors based on the paragraph feature vectors of each side and calculate risk indicators;

[0178] The dynamic optimization module 600 is used to execute an optimization sub-process when the risk index is greater than or equal to the preset risk threshold: identify abnormal paragraph clusters and refine them into refined paragraphs, retain the paragraph feature vectors of the micro paragraphs and calculate the initial feature vectors of the refined paragraphs, locally reconstruct the paragraph association graph and perform a limited number of rounds of feature propagation to recalculate the risk index.

[0179] It is also used to compare the recalculated risk index with the risk threshold. If it is still greater than or equal to the risk threshold, the subprocess is repeatedly optimized within the maximum number of iterations. Otherwise, the final risk index is output.

[0180] The granular decision and path planning module 700 is used to determine the granularity mode of graph attention network calculation based on the final risk index of each edge, and modulate the attention calculation according to the granularity mode to output the emergency logistics path planning results.

[0181] The system proposed in this embodiment belongs to the same inventive concept as the method proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0182] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. An emergency logistics route planning method based on edge-enhanced graph attention, characterized in that, include: Acquire micro-level traffic flow data and topological connectivity data for each edge of the target logistics network; Based on the micro-traffic flow data and topological connection data, each edge is divided into segments to obtain and output multiple micro-segments, and the initial feature vector of each micro-segment is calculated. A paragraph association graph is constructed based on the micro-segments, with micro-segments as nodes and traffic flow interaction relationships as association edges, and association edge features characterizing the intensity of traffic flow conflicts are calculated; The paragraph association graph and the associated edge features are input into the graph convolutional network layer to update the initial feature vector and output the paragraph feature vector that incorporates conflict information. Construct a set of paragraph feature vectors based on the paragraph feature vectors of each side and calculate the risk index; When the risk index is greater than or equal to the preset risk threshold, the optimization sub-process is executed: identify abnormal paragraph clusters and refine them to obtain refined paragraphs, retain the paragraph feature vectors of the micro paragraphs and calculate the initial feature vectors of the refined paragraphs, locally reconstruct the paragraph association graph and perform a limited number of rounds of feature propagation to recalculate the risk index. The recalculated risk index is compared with the risk threshold. If it is still greater than or equal to the threshold, the optimization sub-process is repeated within the maximum number of iterations. Otherwise, the final risk index is output. The granularity mode of graph attention network computation is determined based on the final risk indicators of each side, and the attention computation is modulated according to the granularity mode to output the emergency logistics route planning results.

2. The emergency logistics route planning method based on edge-enhanced graph attention according to claim 1, characterized in that, The process of dividing each edge into segments based on the micro-traffic flow data and topological connectivity data, obtaining and outputting multiple micro-segments, and calculating the initial feature vector of each micro-segment specifically includes: Obtain key point information of road structure in the topology connection data, including intersection location, lane number change point, ramp merging point and diverging point; Based on the key point information, each edge is divided into multiple continuous micro segments, so that the road structure inside each micro segment remains homogeneous, and the boundary of the micro segment is located at the key point. For each micro-segment obtained from the division, statistical features, including the average vehicle speed, vehicle density, and vehicle acceleration variance of the micro-traffic flow data flowing through the micro-segment, are extracted to form the initial feature vector.

3. The emergency logistics route planning method based on edge-enhanced graph attention according to claim 1, characterized in that, The process of calculating the associated edge features that characterize the intensity of traffic flow conflicts specifically includes: For each association edge in the paragraph association graph, it connects an upstream paragraph node and a downstream paragraph node; Obtain the output traffic flow of the upstream segment node and the remaining traffic capacity of the downstream segment node; The formula for calculating the intensity of traffic flow conflict is: ; in, Indicates the intensity of traffic conflict. This indicates the output traffic flow. Indicates the remaining throughput capacity. It is a very small normal value used to prevent division by zero errors.

4. The emergency logistics route planning method based on edge-enhanced graph attention according to claim 1, characterized in that, The process of inputting the paragraph association graph and association edge features into the graph convolutional network layer, updating the initial feature vector, and outputting the paragraph feature vector that incorporates conflict information specifically includes: The initial feature vectors of each node in the paragraph association graph and the association edge features of each association edge are input into a graph convolutional network layer. The graph convolutional network layer aggregates and nonlinearly transforms the initial feature vector of each node with the initial feature vectors of its neighboring nodes and the associated edge features of the connecting edges. The updated initial feature vector of each node is output as the paragraph feature vector of the fused conflict information.

5. The emergency logistics route planning method based on edge-enhanced graph attention according to claim 1, characterized in that, The process of calculating the aforementioned risk indicators specifically includes: Based on the paragraph feature vectors of all micro-segments within the edge, calculate the local risk score for each micro-segment; The mean of the local risk scores within the edge is calculated based on the local risk scores, and each local risk score is normalized. The risk index of an edge is calculated based on the dispersion of the normalized local risk score. The risk index is used to characterize the unevenness of the risk distribution within the edge.

6. The emergency logistics route planning method based on edge-enhanced graph attention according to claim 5, characterized in that, The process of identifying anomalous paragraph clusters and refining them into detailed paragraphs specifically includes: Micro-segments with local risk scores within the edge that are greater than or equal to a preset local risk threshold are marked as high-risk segments. Multiple high-risk paragraphs that are spatially adjacent are merged to form an anomalous paragraph cluster; For the abnormal paragraph cluster, a secondary division is performed within its corresponding spatial range to generate multiple refined paragraphs with smaller spatial scales, forming a refined paragraph set.

7. The emergency logistics route planning method based on edge-enhanced graph attention according to claim 6, characterized in that, The process of partially reconstructing the paragraph association graph and performing a limited number of rounds of feature propagation to recalculate risk indicators specifically includes: Replace the nodes belonging to the abnormal paragraph cluster in the original paragraph association graph with the corresponding nodes in the refined paragraph set; Based on the topological connection data, establish the connection edges between the refined paragraph nodes and between them and other retained nodes in the original paragraph association graph to form a locally reconstructed paragraph association graph. Using the initial feature vector of the refined paragraph and the paragraph feature vector of the retained micro-paragraph as input, graph convolutional network layer feature propagation update is performed no more than a preset number of times on the locally reconstructed paragraph association graph; Based on the updated paragraph feature vectors, the local risk scores of all paragraphs along the edge are recalculated, and a new risk index for the edge is calculated accordingly.

8. The emergency logistics route planning method based on edge-enhanced graph attention according to claim 1, characterized in that, The process of setting the preset risk threshold specifically includes: Obtain the topological importance data of each edge in the target logistics network, wherein the topological importance data includes betweenness centrality and historical average flow carrying ratio; Based on the topological importance data, calculate the personalized threshold corresponding to each edge; The calculated personalized threshold is used as the preset risk threshold for the corresponding edge.

9. The emergency logistics route planning method based on edge-enhanced graph attention according to claim 1, characterized in that, After determining the granularity pattern and before modulating the attention calculation according to the granularity pattern, the following steps are also included: Obtain the risk indicators for each edge over multiple consecutive planning periods in history to form a historical indicator sequence; Based on the historical indicator sequence, predict the expected values ​​of the indicators for each side of the current planning period, and calculate the deviation between the risk indicators for each side of the current period and the expected values ​​of the indicators. If the deviation of a certain edge exceeds a set range, then when modulating the graph attention network, a decay factor is introduced for that edge. The decay factor is used to reduce the weight of the segment feature vector in the attention calculation within the current planning period of that edge.

10. An emergency logistics route planning system based on edge-enhanced graph attention, characterized in that, The method for implementing the emergency logistics route planning based on edge-enhanced graph attention as described in any one of claims 1-9 includes: The data acquisition module is used to acquire micro-traffic flow data and topology connection data of each side in the target logistics network; The segmentation and representation module is used to segment each edge based on the micro traffic flow data and topological connection data, obtain and output multiple micro segments, and calculate the initial feature vector of each micro segment. The association graph construction module is used to construct a paragraph association graph based on the micro-segments, with micro-segments as nodes and traffic flow interaction relationships as association edges, and to calculate association edge features that characterize the intensity of traffic flow conflicts. The feature fusion and update module is used to input the paragraph association graph and the associated edge features into the graph convolutional network layer, update the initial feature vector, and output the paragraph feature vector that fuses conflict information. The risk analysis module is used to construct a set of paragraph feature vectors based on the paragraph feature vectors of each side and to calculate risk indicators; The dynamic optimization module is used to execute an optimization sub-process when the risk index is greater than or equal to the preset risk threshold: identify abnormal paragraph clusters and refine them into refined paragraphs, retain the paragraph feature vectors of the micro paragraphs and calculate the initial feature vectors of the refined paragraphs, locally reconstruct the paragraph association graph and perform a limited number of rounds of feature propagation to recalculate the risk index. It is also used to compare the recalculated risk index with the risk threshold. If it is still greater than or equal to the risk threshold, the optimization sub-process is repeated within the maximum number of iterations; otherwise, the final risk index is output. The granular decision and path planning module is used to determine the granularity mode of graph attention network calculation based on the final risk index of each edge, and modulate the attention calculation according to the granularity mode to output the emergency logistics path planning result.

Images