Intelligent line selection architecture and method for large transportation
By combining a layered service architecture with graph neural networks and large language models, the problem of relying on human experience in the planning of heavy cargo transportation routes is solved. This enables an intelligent route selection method that is automated, dynamically adjusted, and optimized for multiple objectives, thereby improving the planning efficiency and decision-making quality of heavy cargo transportation.
Patent Information
- Application Number
- CN202511424136.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-30
- Publication Date
- 2026-01-13
AI Technical Summary
The planning of routes for large-item transportation relies on human experience, which is inefficient and prone to errors. Multiple constraints are difficult to balance in a coordinated manner, traditional optimization algorithms cannot be dynamically adjusted, and existing models lack the fusion of multi-source heterogeneous features and domain knowledge, resulting in rigid planning schemes.
By adopting a hierarchical service architecture, combining graph neural networks and large language models, and using a hybrid inference architecture and reinforcement learning optimizer, we can generate and optimize the transportation routes for large items, and achieve multimodal feature fusion and real-time dynamic adjustment.
It automates and efficiently responds to large-item transportation route planning, enabling rapid response to emergencies, optimizing transportation costs, improving decision-making quality, and supporting efficient concurrent requests.
Smart Images
Figure CN121329264A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of intelligent transportation and artificial intelligence, in particular to a large piece transportation intelligent route selection architecture and method. BACKGROUND
[0002] Large piece transportation (such as wind power equipment, chemical devices, etc.) faces complex challenges far beyond ordinary logistics in route planning due to inherent characteristics such as size over-limit, heavy weight, and high turning radius requirements. Currently, the following technical bottlenecks exist in this field: First, the planning process highly depends on manual experience. Engineers need to manually check a large amount of static data such as road height limit, width limit, and bearing capacity, and determine the feasibility of the route based on personal experience. This method is not only inefficient, but also prone to planning failure due to insufficient experience or outdated information, leading to transportation accidents. Second, it is difficult to balance multiple constraints; an optimal route needs to meet multiple objectives such as safety (avoiding accident risks), economy (controlling costs), timeliness (ensuring efficiency), and stability (complying with regulations). Traditional optimization algorithms often use fixed weights for decision-making, which cannot be dynamically adjusted according to factors such as cargo type and real-time traffic conditions (such as sudden traffic accidents and temporary traffic controls), resulting in rigid planning solutions that are difficult to adapt to complex and changing real-world transportation environments.
[0003] Currently, although some studies have attempted to apply graph neural networks (GNN) to model road networks, they often only capture basic topological structures and lack sufficient fusion of multi-source heterogeneous features (such as dynamic events and spatial constraints) on nodes and edges, resulting in limited model representation capabilities. Directly invoking large language models (LLM) for route generation can easily produce absurd results that violate physical limitations due to the lack of professional domain knowledge (such as road engineering constraints). SUMMARY
[0004] The present application aims to address the current reliance on manual route selection for large piece transportation, which is inefficient and prone to errors. To address this deficiency, a large piece transportation intelligent route selection architecture and method are proposed.
[0005] To achieve the above objectives, the present application employs the following technical solutions: A large piece transportation intelligent route selection architecture includes a front-end request access layer, a back-end intelligent analysis layer, and a result return and feedback layer. The front-end request access layer is used to receive client transportation requests and intelligently distribute them through an API gateway, forwarding real-time transportation requests to a graph structure construction module. The back-end intelligent analysis layer is used to analyze client transportation requests, generate a candidate route set through a hybrid inference architecture, and filter out the optimal route using a reinforcement learning optimizer. The result return and feedback layer is used for receiving the cached optimal route, returning the optimal route scheme to the client, saving the transportation data to the historical database after the transportation is completed, and dynamically updating the optimization weight by using the actual transportation feedback.
[0006] As a further preferred embodiment of the present application, the backend intelligent analysis layer comprises: a graph structure construction module for generating a transportation network topology in real time; a feature fusion module for integrating task parameters and graph semantic vectors; a hybrid inference architecture for generating candidate routes; a reinforcement learning optimizer for analyzing a client transportation request, generating a candidate route set by the hybrid inference architecture, and screening an optimal route by using the reinforcement learning optimizer.
[0007] An intelligent route selection method for large-size transportation, comprising the following steps: a backend intelligent analysis layer establishes a transportation intelligent route selection model; the transportation intelligent route selection model comprises a real-time transportation network topology, a historical transportation network topology, and historical transportation data; a front-end request access layer receives a client transportation request and intelligently distributes the request through an API gateway, and forwards the real-time transportation request to the backend intelligent analysis layer; the backend intelligent analysis layer analyzes the transportation request, extracts multi-dimensional task parameters, and matches a historical closest request based on the transportation intelligent route selection model; the multi-dimensional task parameters include task type, cargo parameters, historical case features, and road network topology information; when the multi-dimensional task parameter type of the transportation request deviates from at least 3 parameters in the multi-dimensional parameters of the historical transportation data by ±5% or less, the historical transportation data is the historical closest request; a hybrid inference architecture generates a candidate route set based on multi-modal input, compresses graph structure information by a graph neural network, and aligns the information with a large language model prompt word space; a reinforcement learning optimizer dynamically evaluates the safety, economy, timeliness, and stability of the candidate route set by a reinforcement learning framework, and outputs an optimal route; the transportation intelligent route selection model is based on the historical closest request, combines the multi-dimensional task parameters of the current transportation request, combines the graph attention network compression graph structure and the large language model prompt word, realizes graph structure compression, multi-modal fusion, and prompt word space alignment, and generates a candidate route; a reinforcement learning optimizer dynamically evaluates the candidate route by a reinforcement learning framework, and outputs an optimal route to a result return and feedback layer; the result return and feedback layer caches and accelerates the response to the optimal route, and returns the final route scheme to the client; after the transportation is completed, the transportation data is saved to the historical database, and the optimization weight is dynamically updated by using the actual transportation feedback.
[0008] As a further preferred embodiment of the present application, the back-end intelligent analysis layer establishes a transportation intelligent route selection model, comprising: The graph structure construction module constructs a transportation network graph structure, which includes a node system and an edge weight system, the node system including transportation origin-destination points, over-limit detection stations, highway toll stations, traffic hubs, road intersections and dynamic obstacle points, and the edge weight system including road static attributes and dynamic attributes; The feature fusion module extracts and fuses task types, cargo parameters, historical case features and road network topology information to form multi-modal input; The hybrid inference architecture, based on multi-modal input, compresses graph structure information through graph neural network and aligns with large language model prompt word space to generate a candidate route set; The reinforcement learning optimizer dynamically evaluates the safety, economy, timeliness and stability of the candidate routes through a reinforcement learning framework, and outputs the optimal route to the result return and feedback layer.
[0009] As a further preferred embodiment of the present application, the node system is structured and coded by geographic coordinates, maximum load capacity and spatial constraints, and the edge weight system is adapted to road state changes by fusion of static attributes and dynamic attributes.
[0010] As a further preferred embodiment of the present application, the historical case features are encoded by a pre-trained language model and weighted processed in combination with a geographic similarity matrix.
[0011] The large transportation intelligent route selection architecture and method proposed by the present application have the following advantages compared with the prior art: 1. The present application receives transportation requests, performs intelligent analysis and returns optimized routes through a hierarchical service architecture, while achieving data sedimentation and feedback update; 2. The present application realizes full-process automation from data integration, feature fusion, route generation to optimization decision, greatly reducing the dependence on human experience, significantly improving the efficiency of route planning, and quickly responding to transportation requests; 3. The present application enhances the real-time response capability of the system, dynamically adjusts the route to avoid traffic accidents and other sudden situations; 4. The present application reduces transportation costs by multi-objective trade-off of toll fees and time costs; 5. The present application uses historical cases for continuous learning to improve the intelligent level and decision quality of the system; 6. The present application constructs an end-to-end intelligent route selection closed-loop system to support large-scale concurrent requests and efficient response. BRIEF DESCRIPTION OF DRAWINGS
[0012] Figure 1 Large transportation intelligent route selection hierarchical service architecture diagram. DETAILED DESCRIPTION
[0013] The application will be specifically introduced below in combination with the drawings and specific embodiments.
[0014] The application is directed to large piece transportation scenarios, and through multi-modal feature fusion, graph structure semantic compression and large model inference optimization, intelligent generation and dynamic decision of large piece transportation routes are realized, which is suitable for highway transportation route planning of large pieces of materials such as wind power equipment and chemical devices.
[0015] The architecture involved in the application realizes full-process connection through strict hierarchical design, the front-end request access layer focuses on task receiving and shunting, the back-end intelligent analysis layer integrates core technologies such as graph structure construction, feature fusion, large model inference and reinforcement learning, the result returning and feedback layer guarantees route delivery and system continuous optimization, and together they constitute a complete intelligent route selection service system.
[0016] The application relates to an intelligent route selection architecture and method for large piece transportation, which accurately models the physical constraints of the road network by fusing graph neural networks (GNN), combines the knowledge reasoning ability of large language models (LLM), and constructs a decision framework that cooperates data driving and knowledge guidance. Through the following processes, an intelligent route selection method for large piece transportation is obtained, and an intelligent route selection system is built: (1) data modeling and feature extraction engineering; (2) multi-modal input modeling; (3) design of hybrid inference architecture; (4) verification and optimization of candidate routes; (5) deployment of hierarchical service architecture.
[0017] Embodiment one, in combination Figure 1 An intelligent route selection architecture for large piece transportation includes a front-end request access layer, a back-end intelligent analysis layer, and a result returning and feedback layer.
[0018] The front-end request access layer is configured to receive client transportation requests and intelligently shunt them through an API gateway, forwarding real-time transportation requests to a graph structure construction module.
[0019] The back-end intelligent analysis layer is configured to analyze client transportation requests, generate a candidate route set through a hybrid inference architecture, and filter out the optimal route using a reinforcement learning optimizer. The back-end intelligent analysis layer includes: a graph structure construction module for generating a transportation network topology in real time; a feature fusion module for integrating task parameters and graph semantic vectors; a hybrid inference architecture for generating candidate routes; and a reinforcement learning optimizer for analyzing client transportation requests, generating a candidate route set through the hybrid inference architecture, and filtering out the optimal route using the reinforcement learning optimizer.
[0020] The result returning and feedback layer is configured to receive the cached optimal route, return the optimal route solution to the client, save the transportation data to a historical database after the transportation is completed, and dynamically update the optimization weights using actual transportation feedback.
[0021] Embodiment two: an intelligent route selection method for large item transportation, comprising the following steps: 101. The backend intelligent analysis layer establishes a transportation intelligent route selection model; the transportation intelligent route selection model includes real-time transportation network topology, historical transportation network topology, and historical transportation data.
[0022] The graph structure construction module analyzes the historical transportation routes and historical waybill data of large items to transform into a graph structure including a node system and an edge weight system. The node system is structured and coded through geographic coordinates, maximum load capacity, and spatial constraints, and the edge weight system is adapted to road state changes through the fusion of static attributes and dynamic attributes.
[0023] The geographic coordinates, maximum load capacity, and spatial constraints of the road are fused through Cartesian product coding; the spatial constraints include road width limit, road height limit, and minimum turning radius of the road.
[0024] The node system is formed in combination with transportation route data; the transportation route data includes transportation origin and destination, over-limit monitoring station, highway toll station, transportation hub, road intersection, and dynamic obstacle point.
[0025] The static attributes and dynamic attributes of road traffic are distinguished through Hadamard product. The static attributes of road traffic include traffic distance, road grade, road height limit, road width limit, and historical accident frequency of the section; the dynamic attributes of road traffic include:
[0026] Based on the node system and the edge weight system, a cross-modal attention mechanism is used for dynamic fusion to construct the transportation network graph structure.
[0027] The feature fusion module extracts and fuses task type, cargo parameter, historical case feature, and road network topology information to form multi-modal input.
[0028] The task type is used to identify the transportation category and drive the preference of subsequent route planning.
[0029] The cargo parameters include cargo length, cargo width, cargo height, cargo weight, vehicle minimum turning radius, and vehicle maximum allowable inclination angle, which are used to determine whether the road is passable.
[0030] The historical cases are encoded by a pre-trained language model and are processed by combining a geographical similarity matrix weighting to fuse text semantics and geographical spatial similarity, so that the system can automatically filter and strengthen the historical cases highly related to the current transportation task, preferentially recommend historical successful routes, and avoid known risk points.
[0031] The road network topology is used to dynamically aggregate the features of the adjacent nodes and edges through the message passing mechanism of the GNN, to realize accurate calculation of route feasibility, and to ensure the collaborative decision of the physical constraints of large-scale transportation and dynamic road conditions.
[0032] The hybrid reasoning architecture is based on multi-modal input, compresses graph structure information through a graph neural network, and aligns with the large language model prompt word space to generate a candidate route set; The reinforcement learning optimizer dynamically evaluates the safety, economy, timeliness and stability of the candidate routes through a reinforcement learning framework, and outputs the optimal route.
[0033] 102、The front-end request access layer receives the client transportation request and intelligently distributes it through the API gateway, forwarding the real-time transportation request to the back-end intelligent analysis layer.
[0034] 103、The back-end intelligent analysis layer analyzes the transportation request, extracts multi-dimensional task parameters, and matches the closest historical request based on the transportation intelligent route selection model. The multi-dimensional task parameters include task type, cargo parameters, historical case features, and road network topology information. When the multi-dimensional task parameter type of the transportation request deviates from at least 3 parameters in the multi-dimensional parameters of the historical transportation data by ±5% or less, the historical transportation data is the closest historical request; 104、The hybrid reasoning architecture is based on multi-modal input, compresses graph structure information through a graph neural network, and aligns with the large language model prompt word space to generate a candidate route set; 105、The reinforcement learning optimizer dynamically evaluates the safety, economy, timeliness and stability of the candidate route set through a reinforcement learning framework, and outputs the optimal route. 106、The transportation intelligent route selection model is based on the closest historical request, combined with the multi-dimensional task parameters of the current transportation request, combined with the graph attention network to compress the graph structure and the large language model prompt word, to realize graph structure compression, multi-modal fusion and prompt word space alignment, and to generate candidate routes; 107、The reinforcement learning optimizer dynamically evaluates the candidate routes through a reinforcement learning framework, and outputs the optimal route to the result return and feedback layer; 108、The result return and feedback layer caches the optimal route to accelerate the response, and returns the final route scheme to the client. After the transportation is completed, the transportation data is saved to the historical database, and the actual transportation feedback is used to dynamically update the optimization weight.
[0035] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the above examples do not limit the present application in any form, and any technical solution obtained by equivalent substitution or equivalent transformation falls within the scope of the present application.
Claims
1. An intelligent route selection architecture for oversize load transportation, characterized by, The front-end request access layer, the back-end intelligent analysis layer, and the result returning and feedback layer are included. The front-end request access layer is configured to receive a client transportation request and intelligently distribute the request through an API gateway, and forward the real-time transportation request to a graph structure construction module. The back-end intelligent analysis layer is configured to analyze the client transportation request, generate a candidate route set through a hybrid inference architecture, and filter out an optimal route using a reinforcement learning optimizer. The result returning and feedback layer is configured to receive the cached optimal route, return the optimal route solution to the client, save the transportation data to a historical database after the transportation is completed, and dynamically update the optimization weight using actual transportation feedback.
2. The intelligent route selection architecture for oversize load transportation of claim 1, wherein, The back-end intelligent analysis layer includes: The graph structure construction module is configured to generate a transportation network topology in real time. The feature fusion module is configured to integrate task parameters and graph semantic vectors. The hybrid inference architecture is configured to generate a candidate route. The reinforcement learning optimizer is configured to analyze the client transportation request, generate a candidate route set through the hybrid inference architecture, and filter out an optimal route using the reinforcement learning optimizer.
3. The intelligent route selection method for oversize cargo transportation according to claim 1 or 2, characterized in that, The method includes the following steps: The back-end intelligent analysis layer establishes a transportation intelligent route selection model, which includes a real-time transportation network topology, a historical transportation network topology, and historical transportation data. The front-end request access layer receives a client transportation request and intelligently distributes the request through an API gateway, and forwards the real-time transportation request to the back-end intelligent analysis layer. The back-end intelligent analysis layer analyzes the transportation request, extracts multi-dimensional task parameters, and matches a historical closest request based on the transportation intelligent route selection model. The multi-dimensional task parameters include task type, cargo parameters, historical case features, and road network topology information. When the multi-dimensional task parameter type of the transportation request deviates from at least 3 parameters in the multi-dimensional parameters of the historical transportation data by within ±5%, the historical transportation data is the historical closest request. The hybrid inference architecture generates a candidate route set based on multi-modal input, compresses graph structure information through a graph neural network, and aligns with a large language model prompt word space. The reinforcement learning optimizer dynamically evaluates the safety, economy, timeliness, and stability of the candidate route set through a reinforcement learning framework, and outputs an optimal route. The transportation intelligent route selection model is based on the historical closest request, combined with the multi-dimensional task parameters of the current transportation request, combined with the graph attention network compression graph structure and the large language model prompt word, realizes graph structure compression, multi-modal fusion, and prompt word space alignment, and generates a candidate route. The reinforcement learning optimizer dynamically evaluates the candidate route through a reinforcement learning framework, and outputs an optimal route to the result returning and feedback layer. The result returning and feedback layer caches and accelerates the response to the optimal route, and returns the final route solution to the client. When the transportation is completed, the transportation data is saved to the historical database, and the optimization weight is dynamically updated using actual transportation feedback.
4. The intelligent route selection method for oversize cargo transportation according to claim 3, characterized in that, The back-end intelligent analysis layer establishes a transportation intelligent route selection model, which includes: The graph structure construction module constructs a transportation network graph structure, which includes a node system and an edge weight system, the node system includes transportation origin-destination points, over-limit detection stations, highway toll stations, traffic hubs, road intersections and dynamic obstacle points, and the edge weight system includes road static attributes and dynamic attributes; The feature fusion module extracts and fuses task types, cargo parameters, historical case features and road network topology information to form multi-modal input; The hybrid reasoning architecture compresses graph structure information through a graph neural network and aligns it with a large language model prompt word space based on multi-modal input to generate a candidate route set; The reinforcement learning optimizer dynamically evaluates the safety, economy, timeliness and stability of the candidate routes through a reinforcement learning framework and outputs an optimal route to the result return and feedback layer.
5. The intelligent route selection method for oversize cargo transportation according to claim 3, wherein, The node system is structured and coded through geographic coordinates, maximum load and spatial constraints, and the edge weight system is adapted to road state changes through the fusion of static attributes and dynamic attributes.
6. The intelligent route selection method for oversize cargo transportation according to claim 3, wherein, The historical case features are encoded by a pre-trained language model and are weighted and processed in combination with a geographic similarity matrix.