Intelligent logistics scheduling system and method based on deep learning

Abstract

The invention discloses an intelligent logistics scheduling system and method based on deep learning, and the method comprises the following steps: S1, collecting logistics data, and carrying out the preprocessing of the logistics data; s2, constructing a logistics task graph; s3, constructing a hierarchical scheduling structure, and setting a high-layer policy network and a low-layer policy network; s4, constructing a multi-agent mapping relation according to the task allocation scheme; s5, performing strategy optimization on the logistics task subgraph, and outputting specific scheduling actions of all transport vehicles; s6, updating the real-time state of each vehicle, and calculating the task completion state and the resource use ratio in the current scheduling round; and S7, taking a data comparison result as training feedback, updating strategy parameters of the high-layer strategy network and the low-layer strategy network, and entering a next round of scheduling decision. According to the invention, intelligent division, accurate matching and dynamic scheduling of large-scale logistics tasks can be realized, and the scheduling efficiency and the resource utilization rate are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent logistics and intelligent scheduling, and particularly relates to an intelligent logistics scheduling system and method based on deep learning. BACKGROUND

[0002] In the existing intelligent logistics scheduling technology, a rule engine, heuristic algorithm or traditional operational optimization model is generally used for scheduling and path planning of transportation resources. Such methods usually rely on manually set scheduling rules or static cost functions, and are difficult to cope with dynamic changes in the structure of logistics tasks and real-time decision-making problems in a multi-dimensional data environment. At the same time, in the face of complex distribution scenarios with large-scale multi-vehicle and multi-task, traditional methods lack efficient intelligent strategy generation capabilities in task division, priority determination and resource allocation, resulting in low scheduling efficiency, insufficient resource utilization, and frequent response lags.

[0003] Some studies introduce deep reinforcement learning methods to realize scheduling optimization, but most of them are only applied to single-agent control, lack the ability to model multi-agent collaborative behavior, and lack effective representation and utilization of task graph structure characteristics, which cannot realize efficient collaboration of regional division, task decomposition and strategy scheduling, limiting the application value of the algorithm in the actual intelligent logistics environment.

[0004] Therefore, how to provide an intelligent logistics scheduling system and method based on deep learning is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide an intelligent logistics scheduling system and method based on deep learning. The present application combines graph neural networks, multi-agent reinforcement learning and hierarchical strategy structure to construct a high-low layer collaborative task scheduling mechanism. The whole process from task graph construction, regional division, priority sorting, agent mapping to scheduling action generation and feedback update is described in detail, and the present application has the advantages of fast scheduling response speed, high resource utilization efficiency and strong adaptability to dynamic scenarios.

[0006] According to an embodiment of the present application, an intelligent logistics scheduling method based on deep learning comprises the following steps:

[0007] S1, collecting logistics data and performing preprocessing;

[0008] S2, based on the preprocessed logistics data, constructing a logistics task graph, wherein the nodes of the logistics task graph represent distribution tasks and transportation vehicles, and the edges represent the scheduling relationship and distance cost between the distribution tasks and the transportation vehicles;

[0009] S3, construct a hierarchical scheduling structure, set a high-level strategy network and a low-level strategy network, the high-level strategy network performs regional division and task priority sorting based on the logistics task graph, and generates a task allocation scheme;

[0010] S4, according to the task allocation scheme, construct a multi-agent mapping relationship, map the transport vehicle agent to the corresponding region, and extract the logistics task subgraph in the region as the input of the low-level strategy network;

[0011] S5, in each region, the low-level strategy network performs strategy optimization on the logistics task subgraph, and outputs specific scheduling actions of all transport vehicles, including departure time decision, path sequence planning and task execution order;

[0012] S6, according to the specific scheduling action, update the real-time state of each vehicle, and calculate the task completion state and resource usage ratio in the current scheduling round, and write into the system database for data comparison;

[0013] S7, the data comparison result is used as training feedback to update the strategy parameters of the high-level strategy network and the low-level strategy network, and enter the next round of scheduling decision until all distribution tasks are completed.

[0014] Optionally, the logistics data includes distribution task data, vehicle state data and traffic condition data.

[0015] Optionally, the preprocessing includes data deduplication, format standardization, missing value filling and outlier removal.

[0016] Optionally, the S2 specifically includes:

[0017] S21, extract the distribution task information and the transport vehicle state information from the preprocessed logistics data, set one task node for each distribution task, and one vehicle node for each transport vehicle;

[0018] S22, according to the spatial position, time requirement and path accessibility between the distribution task and the transport vehicle, construct the connection edge between the nodes, the connection edge is used to represent the scheduling relationship between the distribution task and the transport vehicle;

[0019] S23, extract traffic condition information from preprocessed logistics data, calculate corresponding distance cost for each connection edge, the distance cost is generated based on path length and traffic congestion degree weighting, used to quantify the scheduling cost of the vehicle executing the corresponding task;

[0020] S24, all nodes and corresponding edges are uniformly constructed into a weighted directed graph structure to form a logistics task graph, the logistics task graph is used to express the structured relationship and cost constraint of the distribution task and the transport vehicle in the scheduling process;

[0021] S25, structure optimization processing is performed on the logistics task graph, and the sparsity and scheduling correlation of the graph structure representation are improved by screening low-cost paths, removing unreachable edges, and limiting the maximum number of connections.

[0022] Optionally, the S3 specifically includes:

[0023] 31. Based on the logistics task graph, the task nodes are clustered using the OPTICS clustering algorithm, and the task nodes with close spatial distance and high task intensity in the same region are divided into the same scheduling region. Meanwhile, each vehicle node is initially attributed according to the clustering region to which the current position belongs;

[0024] S32, a hierarchical scheduling structure is constructed, in each scheduling region, a high-level strategy network is constructed based on a graph attention mechanism, and embedding calculation is performed on all task nodes in the scheduling region, the structural features and attribute features of each task node are extracted, the dependency relationship between nodes is modeled through an attention weight mechanism, and a task priority sequence is formed;

[0025] S33, according to the task priority sequence and the initial attribution relationship of the transportation vehicles in the region, a task allocation scheme is generated using the principle of local minimum scheduling cost, high-priority tasks are matched to vehicles with close distance and minimum current load, and a mapping relationship between tasks and vehicles in the scheduling region is formed;

[0026] S34, the task allocation scheme generated in each scheduling region is summarized to form a global task allocation scheme, and the global task allocation scheme is output as a high-level strategy network.

[0027] Optionally, the S4 specifically includes:

[0028] S41, according to the mapping relationship between tasks and vehicles in the scheduling region formed by the high-level strategy network, a set of corresponding task nodes is established for each transportation vehicle Agent, and a multi-agent mapping configuration in the region is generated, which is used to identify the scheduling region to which each Agent belongs and the set of tasks to be executed by each Agent;

[0029] S42, according to the multi-agent mapping configuration, a logistics task subgraph is constructed for each scheduling region, the logistics task subgraph is composed of task nodes, vehicle nodes and corresponding edges, and the connection relationship of the edges is determined by the scheduling path in the task allocation scheme;

[0030] S43, structure pruning processing is performed on each logistics task subgraph, redundant nodes not participating in task assignment are removed, vehicle nodes and edge structures having effective connections with allocated tasks are retained, and are used for a low-level strategy network.

[0031] Optionally, the multi-agent mapping configuration is based on a global task allocation scheme output by a high-level policy network, adopts a regional attention aggregation mechanism to calculate scheduling association strength between each transport vehicle agent and a task set to which the transport vehicle agent belongs, forms a mapping structure, constructs a priority assignment sequence of each task node in the task set to the corresponding transport vehicle agent, and introduces position sequence coding and task time window coding in the mapping structure.

[0032] Optionally, the S5 specifically includes:

[0033] S51, inputting a logistics task subgraph into a low-level policy network in each scheduling area, and the low-level policy network adopts a graph convolution structure and a regional attention aggregation mechanism to perform feature extraction and policy scoring on a scheduling path between a task node and a vehicle node;

[0034] S52, generating a scheduling action sequence for each transport vehicle according to extracted graph structure features and sequence features, the scheduling action sequence including a departure time decision of the vehicle, an arrangement order of path nodes passed through, and an execution order of corresponding tasks, and the scheduling order is sorted according to a task time window urgency and a path cumulative cost;

[0035] S53, outputting the scheduling action sequence generated by each transport vehicle to an instruction execution layer to constitute a scheduling result of this round, and the scheduling result is used as an optimal execution scheme under the current policy and is used for scheduling state updating.

[0036] Optionally, the S6 specifically includes:

[0037] S61, sequentially executing a departure, driving and stopping action of each transport vehicle according to the scheduling action sequence output by the low-level policy network, updating a current position of the vehicle, a task completion identifier and a vehicle load state after each action node to form a scheduling real-time state table;

[0038] S62, after each scheduling round ends, counting completion conditions of all tasks, calculating a number of completed tasks, a number of uncompleted tasks and an average completion time in the scheduling round, and calculating a resource usage proportion according to the logistics task graph and the global task allocation scheme;

[0039] S63, writing the scheduling real-time state table, the task completion state and the resource usage proportion as scheduling data of the current round into a designated data area of a database, and establishing an index relationship according to a task number and a vehicle number, for data comparison and network training feedback.

[0040] According to the embodiment of the application, a smart logistics scheduling system based on deep learning includes:

[0041] A data processing module is configured to collect logistics data and perform preprocessing.

[0042] a logistics task module for constructing a logistics task graph, nodes of the logistics task graph representing delivery tasks and transport vehicles, and edges representing scheduling relationships and distance costs between the delivery tasks and the transport vehicles;

[0043] a hierarchical scheduling module for constructing a hierarchical scheduling structure, setting a high-level strategy network and a low-level strategy network, the high-level strategy network performing regional division and task priority sorting based on the logistics task graph and generating a task allocation scheme;

[0044] a multi-agent mapping module for constructing a multi-agent mapping relationship, mapping transport vehicle agents to corresponding regions, and extracting a logistics task subgraph in each region as an input of the low-level strategy network;

[0045] a strategy optimization module for, in each region, the low-level strategy network performing strategy optimization on the logistics task subgraph and outputting specific scheduling actions of all transport vehicles, including departure time decision, path sequence planning, and task execution order;

[0046] an update calculation module for updating real-time states of the vehicles, calculating task completion states and resource usage proportions in the current scheduling round, and writing the states and proportions into a system database for data comparison;

[0047] a training feedback module for taking the data comparison results as training feedback, updating strategy parameters of the high-level strategy network and the low-level strategy network, and entering a next round of scheduling decision until all delivery tasks are completed.

[0048] The present application has the following advantages:

[0049] Firstly, the present application realizes regional management and priority control of large-scale delivery tasks by constructing a logistics task graph and introducing a hierarchical scheduling strategy based on a graph structure, effectively solving the problems of unreasonable task division and coarse scheduling granularity in traditional methods, and improving the spatial organization efficiency of scheduling.

[0050] Secondly, the present application realizes fine-grained path planning and task order optimization within each scheduling region by adopting a multi-agent mapping mechanism and a low-level strategy network, significantly improving the rationality of vehicle scheduling and the timeliness of task execution.

[0051] Finally, the present application realizes dynamic optimization and continuous iteration of the strategy through real-time state updating, resource usage statistics, and feedback mechanisms in the scheduling process, enhancing the adaptive ability and overall scheduling stability of the system in complex and variable logistics environments. BRIEF DESCRIPTION OF DRAWINGS

[0052] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and are meant to explain the application without limiting the application to the exact embodiments shown. In the drawings:

[0053] Figure 1 A flow chart of a smart logistics scheduling method based on deep learning proposed in the application;

[0054] Figure 2 A logistics structure diagram under a hierarchical scheduling structure of a smart logistics scheduling method based on deep learning proposed in the application;

[0055] Figure 3 A module structure diagram of a smart logistics scheduling system based on deep learning proposed in the application. DETAILED DESCRIPTION

[0056] The application will now be described in further detail with reference to the drawings. These drawings show only the essential features of the application and are therefore to be regarded only as a schematic illustration of the basic structure of the application.

[0057] REFERENCE Figures 1-2 A smart logistics scheduling method based on deep learning, comprising the following steps:

[0058] S1, collecting logistics data and performing preprocessing;

[0059] S2, constructing a logistics task graph based on the preprocessed logistics data, wherein the nodes of the logistics task graph represent distribution tasks and transportation vehicles, and the edges represent scheduling relationships and distance costs between the distribution tasks and the transportation vehicles;

[0060] S3, constructing a hierarchical scheduling structure, setting a high-level strategy network and a low-level strategy network, wherein the high-level strategy network performs regional division and task priority sorting based on the logistics task graph and generates a task allocation scheme;

[0061] S4, constructing a multi-agent mapping relationship according to the task allocation scheme, mapping the transportation vehicle agents to the corresponding regions, and extracting the logistics task subgraph in the region as the input of the low-level strategy network;

[0062] S5, in each region, the low-level strategy network performs strategy optimization on the logistics task subgraph, and outputs specific scheduling actions of all transportation vehicles, including departure time decision, path sequence planning and task execution order;

[0063] S6, updating the real-time state of each vehicle according to the specific scheduling action, and calculating the task completion state and resource usage ratio in the current scheduling round, and writing into the system database for data comparison;

[0064] S7, update the policy parameters of the high-level policy network and the low-level policy network as training feedback of the data comparison result, and enter the next round of scheduling decision until all distribution tasks are completed.

[0065] The present application realizes intelligent control of the whole process from task extraction to task completion by introducing a hierarchical strategy structure and combining logistics task graph modeling and deep learning scheduling strategy, and improves the scheduling response efficiency and task completion quality in large-scale distribution scenarios.

[0066] In the embodiment, the logistics data includes distribution task data, vehicle state data and traffic condition data.

[0067] The present application enhances the adaptability of the scheduling model to real-time environmental changes by introducing distribution task data, vehicle state data and traffic condition data as scheduling basis, and improves the accuracy and effectiveness of scheduling.

[0068] In the embodiment, the preprocessing includes data deduplication, format standardization, missing value filling and outlier removal.

[0069] The present application improves the quality of input data by performing deduplication, standardization, missing value filling and outlier removal on logistics data, and provides a stable foundation for subsequent graph structure construction and strategy learning.

[0070] In the embodiment, S2 specifically includes:

[0071] S21, extract distribution task information and transportation vehicle state information from preprocessed logistics data, set one task node corresponding to each distribution task, and one vehicle node corresponding to each transportation vehicle;

[0072] S22, according to the spatial position, time requirement and path accessibility between the distribution task and the transportation vehicle, construct the connection edge between the nodes, and the connection edge is used to represent the scheduling relationship between the distribution task and the transportation vehicle;

[0073] S23, extract traffic condition information from preprocessed logistics data, calculate the corresponding distance cost for each connection edge, and the distance cost is generated based on the weighted path length and traffic congestion degree, which is used to quantify the scheduling cost of the vehicle executing the corresponding task;

[0074] S24, all nodes and corresponding edges are uniformly constructed into a weighted directed graph structure to form a logistics task graph, and the logistics task graph is used to express the structured relationship and cost constraint of the distribution task and the transportation vehicle in the scheduling process;

[0075] S25, perform structure optimization processing on the logistics task graph, and improve the sparsity and scheduling relevance of the graph structure expression by filtering low-cost paths, removing unreachable edges and limiting the maximum number of connections.

[0076] The application constructs a logistics task graph with a weighted directed graph structure, and optimizes the structure in combination with path cost and reachability, so that the scheduling process has a clear structure, a complete task-resource relationship, and improved accuracy and expression ability of scheduling modeling.

[0077] In the embodiment, the S3 specifically includes:

[0078] 31. Based on the logistics task graph, the task nodes are clustered using the OPTICS clustering algorithm, and the task nodes with close spatial distance and high task intensity in the same region are divided into the same scheduling region, and each vehicle node is initially attributed according to the clustering region to which the current position belongs;

[0079] S32, a hierarchical scheduling structure is constructed, in each scheduling region, a high-level strategy network is constructed based on a graph attention mechanism, and embedding calculation is performed on all task nodes in the scheduling region, the structural features and attribute features of each task node are extracted, the dependency relationship between nodes is modeled through the attention weight mechanism, and a task priority sequence is formed;

[0080] S33, according to the task priority sequence and the initial attribution relationship of the transportation vehicle in the region, a task allocation scheme is generated using the principle of local minimum scheduling cost, high-priority tasks are matched to vehicles with close distance and minimum current load, and a mapping relationship between tasks and vehicles in the scheduling region is formed;

[0081] S34, the task allocation scheme generated in each scheduling region is summarized to form a global task allocation scheme, and the global task allocation scheme is output as a high-level strategy network.

[0082] The application introduces the OPTICS clustering algorithm and the graph attention mechanism based on the task graph, realizes the organic combination of scheduling region division and task priority determination, and improves the accuracy of task assignment and the rationality of regional scheduling.

[0083] In the embodiment, the S4 specifically includes:

[0084] S41, according to the mapping relationship between tasks and vehicles in the scheduling region formed by the high-level strategy network, a task node set corresponding to each transportation vehicle Agent is established, and a multi-agent mapping configuration in the region is generated, which is used to identify the scheduling region to which each Agent belongs and the task set to be executed by each Agent;

[0085] S42, according to the multi-agent mapping configuration, a logistics task subgraph is constructed for each scheduling region, the logistics task subgraph is composed of task nodes, vehicle nodes and corresponding edges, and the connection relationship of the edges is determined by the scheduling path in the task allocation scheme;

[0086] S43, performing structural pruning processing on each logistics task subgraph, removing redundant nodes not participating in task assignment, retaining vehicle node and edge structure having effective connection with assigned tasks, and used for low-level strategy network.

[0087] The application improves the organization efficiency and execution effectiveness of the multi-agent scheduling system by constructing a multi-agent mapping relationship, landing the scheduling scheme into an agent configuration structure, and pruning the generated optimized logistics task subgraph.

[0088] In the embodiment, the multi-agent mapping configuration is based on the global task allocation scheme output by the high-level strategy network, uses a regional attention aggregation mechanism to calculate the scheduling correlation strength between each transport vehicle agent and the task set to which it belongs, forms a mapping structure, constructs a priority assignment sequence of each task node in the task set to the corresponding transport vehicle agent, and introduces position order encoding and task time window encoding in the mapping structure.

[0089] The application enhances the assignment expression capability between tasks and agents in multi-agent mapping by introducing a regional attention aggregation mechanism and task time window encoding, and realizes structured and high-precision input of task allocation.

[0090] In the embodiment, the S5 specifically includes:

[0091] S51, inputting the logistics task subgraph into the low-level strategy network in each scheduling region, the low-level strategy network using a graph convolution structure and a regional attention aggregation mechanism to extract features and score strategies of the scheduling path between the task nodes and the vehicle nodes;

[0092] S52, generating a scheduling action sequence for each transport vehicle according to the extracted graph structure features and sequence features, the scheduling action sequence including the vehicle's departure time decision, the arrangement order of the path nodes and the execution order of the corresponding tasks, the scheduling order being sorted according to the task time window urgency and the path cumulative cost;

[0093] S53, outputting the scheduling action sequence generated by each transport vehicle to the instruction execution layer to constitute the scheduling result of this round, the scheduling result being the optimal execution scheme under the current strategy and used for scheduling state update.

[0094] The application realizes the linkage optimization of departure time, path planning and task sorting by the low-level strategy network for graph convolution and scheduling strategy fusion calculation of the logistics task subgraph, and improves the overall coordination level of the transport vehicle scheduling behavior.

[0095] In the embodiment, the S6 specifically includes:

[0096] S61, according to the scheduling action sequence output by the low-layer strategy network, sequentially performing the departure, driving and stopping actions of each transport vehicle, updating the current position of the vehicle, the task completion identifier and the vehicle load state after each action node, and forming a scheduling real-time state table;

[0097] S62, after each scheduling round ends, the completion of all tasks is counted, the number of completed tasks, the number of uncompleted tasks and the average completion time in the scheduling round are calculated, and the resource usage proportion is calculated according to the logistics task graph and the global task allocation scheme;

[0098] S63, the scheduling real-time state table, the task completion state and the resource usage proportion are written into the database designated data area as the scheduling data of the current round, and an index relationship is established according to the task number and the vehicle number, which is used for data comparison and network training feedback.

[0099] The present application updates the real-time state and counts the resource proportion of the scheduling action execution result, establishes a scheduling data index mechanism, provides a closed-loop basis for feedback training, and improves the iterative optimization capability of the scheduling system.

[0100] Reference Figure 3 A smart logistics scheduling system based on deep learning, comprising:

[0101] A data processing module is used for collecting logistics data and performing preprocessing;

[0102] A logistics task module is used for constructing a logistics task graph, wherein the nodes of the logistics task graph represent distribution tasks and transport vehicles, and the edges represent the scheduling relationship and distance cost between the distribution tasks and the transport vehicles;

[0103] A hierarchical scheduling module is used for constructing a hierarchical scheduling structure, setting a high-layer strategy network and a low-layer strategy network, and the high-layer strategy network is used for performing regional division and task priority sorting based on the logistics task graph, and generating a task allocation scheme;

[0104] A multi-agent mapping module is used for constructing a multi-agent mapping relationship, mapping the transport vehicle Agent to the corresponding region, and extracting the logistics task subgraph in the region as the input of the low-layer strategy network;

[0105] A strategy optimization module is used for performing strategy optimization on the logistics task subgraph in each region by the low-layer strategy network, outputting specific scheduling actions of all transport vehicles, including departure time decision, path sequence planning and task execution order;

[0106] An update calculation module is used for updating the real-time state of each vehicle, calculating the task completion state and the resource usage proportion in the current scheduling round, and writing into the system database for data comparison;

[0107] a training feedback module configured to update the policy parameters of the high-level policy network and the low-level policy network with the data comparison result as the training feedback, and enter a next round of scheduling decision until all the delivery tasks are completed.

[0108] The intelligent logistics scheduling system constructed by the present application has full-chain functional modules such as data processing, task modeling, policy execution and feedback optimization, ensures the automatic and intelligent operation of the scheduling process, and improves the overall operation efficiency and task execution quality of the system.

[0109] Embodiment 1

[0110] In order to verify the feasibility of the present application in implementation, the present application is applied to the daily multi-region and multi-vehicle logistics scheduling process of a large intelligent distribution platform. The platform processes more than 15000 delivery tasks per day, has more than 700 delivery vehicles, covers multiple geographic areas, has uneven task intensity distribution, and has complex traffic conditions and fast dynamic changes in order distribution during peak hours. The traditional scheduling system sets the scheduling strategy based on fixed rules or manual experience, which is difficult to effectively respond to sudden situations such as dynamic order insertion and traffic congestion, often leading to idle resources in some areas and tight transportation capacity in some areas, the overall delivery delay rate is maintained at more than 12%, the customer complaint rate is high, and an intelligent and adaptive scheduling optimization mechanism is urgently needed to replace it.

[0111] In this embodiment, first, the historical logistics data of the platform is collected and preprocessed, and standardized input data including order location, time requirement, vehicle current state, traffic conditions and other features are constructed. By constructing a logistics task graph, all orders to be processed and schedulable vehicle nodes are included in the graph structure, and based on a spatial clustering algorithm, the tasks are regionally divided, and a scheduling region and task priority sequence are generated in combination with the task priority index. For each region, a mapping relationship between vehicles and tasks is generated, and a task subgraph is constructed and input into the low-level policy network to perform path planning, task sorting and time scheduling, and output a detailed scheduling action sequence. After execution, the system real-time state table records the task completion, average completion time, vehicle empty running rate and resource utilization of each round.

[0112] From the actual scheduling performance, the scheduling efficiency of the system after the implementation of the present application is significantly improved. Compared with the fixed strategy system before the implementation, the average task completion time is reduced from 54 minutes to 42 minutes, the average vehicle waiting time is reduced from 18 minutes to 9 minutes, the task delay rate is reduced from 12.4% to 3.7%, and the overall resource utilization is improved by 17.2%. In periods of dynamic fluctuations in traffic conditions, the system automatically adjusts the scheduling strategy through the feedback training mechanism, maintains the average delay within ±2 minutes, and significantly enhances the stability of the system.

[0113] The following is the scheduling performance improvement data statistics result in the embodiment:

[0114] Table 1 Comparison table of intelligent logistics scheduling performance data

[0115]

[0116]

[0117] From the above implementation results, it can be seen that the intelligent logistics scheduling method based on deep learning proposed in the present application can efficiently operate in a complex multi-task and multi-vehicle scheduling scene, not only improves the task execution efficiency and resource utilization capability, but also significantly reduces the uncertainty and execution risk in the scheduling process, and has very high practical application value.

[0118] The above is only the preferred specific implementation of the present application, but the protection scope of the present application is not limited to this, any person skilled in the art within the technical range disclosed in the present application, according to the technical scheme and the inventive concept of the present application, equivalent replacement or change, should be covered in the protection scope of the present application.

Claims

1. A smart logistics scheduling method based on deep learning, characterized by: The steps include: S1. Collect logistics data and pre-process it; S2. Construct a logistics task graph based on the preprocessed logistics data, where the nodes of the logistics task graph represent delivery tasks and transport vehicles, and the edges represent the scheduling relationship and distance cost between the delivery tasks and transport vehicles; S3. Build a hierarchical scheduling structure and set up a high-level policy network and a low-level policy network. The high-level policy network performs regional division and task priority sorting based on the logistics task graph and generates a task allocation plan. S4. Based on the task allocation plan, a multi-agent mapping relationship is constructed, the transport vehicle agent is mapped to the corresponding area, and the logistics task subgraph within the area is extracted as the input of the low-level policy network; S5. In each region, the low-level policy network performs policy optimization on the logistics task subgraph and outputs the specific scheduling actions for all transport vehicles, including departure time decisions, path sequence planning, and task execution order. S6. Update the real-time status of each vehicle based on the specific scheduling action, calculate the task completion status and resource usage ratio in the current scheduling round, and write it into the system database for data comparison; S7. Use the data comparison results as training feedback to update the policy parameters of the high-level policy network and the low-level policy network, and enter the next round of scheduling decision-making until all delivery tasks are completed.

2. The intelligent logistics scheduling method based on deep learning according to claim 1 is characterized in that: The logistics data includes delivery task data, vehicle status data and traffic condition data.

3. The intelligent logistics scheduling method based on deep learning according to claim 1 is characterized in that: The preprocessing includes data deduplication, format standardization, missing value filling and outlier removal.

4. The intelligent logistics scheduling method based on deep learning according to claim 1 is characterized in that: The S2 specifically includes: S21, extracting delivery task information and transport vehicle status information from the pre-processed logistics data, setting each delivery task to correspond to a task node, and each transport vehicle to correspond to a vehicle node; S22. Constructing connecting edges between nodes based on the spatial location, time requirements, and path accessibility between the delivery task and the transport vehicle, wherein the connecting edges are used to represent the scheduling relationship between the delivery task and the transport vehicle; S23. Extract traffic condition information from the pre-processed logistics data and calculate the corresponding distance cost for each connecting edge. The distance cost is generated based on the weighted path length and traffic congestion level to quantify the scheduling cost of the vehicle performing the corresponding task; S24. All nodes and corresponding edges are uniformly constructed into a weighted directed graph structure to form a logistics task graph, which is used to express the structured relationship and cost constraints between delivery tasks and transport vehicles in the scheduling process; S25. Perform structural optimization on the logistics task graph to improve the sparsity and scheduling relevance of the graph structure expression by screening low-cost paths, removing unreachable edges, and limiting the maximum number of connections.

5. The intelligent logistics scheduling method based on deep learning according to claim 1 is characterized in that: The S3 specifically includes:

31. Based on the logistics task graph, the OPTICS clustering algorithm is used to cluster the task nodes. Task nodes with close spatial distance and high task density in the same area are divided into the same scheduling area. At the same time, each vehicle node is initially assigned to the cluster area according to its current location. S32. Build a hierarchical scheduling structure. In each scheduling area, build a high-level policy network based on the graph attention mechanism, perform embedding calculations on all task nodes in the scheduling area, extract the structural and attribute features of each task node, and model the dependencies between nodes through the attention weight mechanism to form a task priority sequence. S33. Based on the task priority sequence and the initial ownership relationship of the transport vehicles in the area, a task allocation plan is generated using the local minimum scheduling cost principle, and high-priority tasks are preferentially matched to vehicle nodes that are closer and have the smallest current load, thereby forming an assignment mapping relationship between tasks and vehicles within the scheduling area; S34. Summarize the task allocation plans generated in each scheduling area to form a global task allocation plan, and output the global task allocation plan as a high-level policy network.

6. The intelligent logistics scheduling method based on deep learning according to claim 1 is characterized in that: The S4 specifically includes: S41. Based on the task-to-vehicle assignment mapping relationship formed by the high-level policy network within the scheduling area, a corresponding task node set is established for each transport vehicle agent, and a multi-agent mapping configuration within the area is generated to identify the scheduling area to which each agent belongs and its set of tasks to be executed; S42. Construct a logistics task subgraph for each scheduling area based on the multi-agent mapping configuration. The logistics task subgraph consists of task nodes, vehicle nodes, and corresponding edges. The connection relationship of the edges is determined by the scheduling path in the task allocation plan. S43. Perform structural pruning on each logistics task subgraph, remove redundant nodes that do not participate in task assignment, retain vehicle nodes and edge structures that have valid connections with assigned tasks, and use them for the low-level strategy network.

7. The deep learning-based intelligent logistics scheduling method according to claim 6 is characterized in that: The multi-agent mapping configuration is based on the global task allocation scheme output by the high-level policy network. It adopts the regional attention aggregation mechanism to calculate the scheduling association strength between each transport vehicle agent and its corresponding task set to form a mapping structure. It constructs a priority assignment sequence from each task node in the task set to the corresponding transport vehicle agent, and introduces position sequence coding and task time window coding into the mapping structure.

8. The intelligent logistics scheduling method based on deep learning according to claim 1 is characterized in that: The S5 specifically includes: S51. In each scheduling area, the logistics task subgraph is input into the low-level policy network. The low-level policy network uses a graph convolution structure and regional attention aggregation mechanism to extract features and perform policy scoring on the scheduling path between the task node and the vehicle node. S52. Generate a scheduling action sequence for each transport vehicle based on the extracted graph structure features and sequence features. The scheduling action sequence includes the vehicle's departure time decision, the order of the path nodes it passes through, and the execution order of the corresponding tasks. The scheduling sequence is sorted based on the urgency of the task time window and the cumulative cost of the path. S53: Output the scheduling action sequence generated for each transport vehicle to the instruction execution layer to form the scheduling result of this round. The scheduling result is used as the optimal execution plan under the current strategy and is used to update the scheduling status.

9. The intelligent logistics scheduling method based on deep learning according to claim 1 is characterized in that: The S6 specifically includes: S61. Execute the departure, travel, and stop actions of each transport vehicle in sequence according to the scheduling action sequence output by the low-level policy network. Update the vehicle's current position, task completion indicator, and vehicle load status after each action node to form a real-time scheduling status table. S62. After each scheduling round, count the completion status of all tasks, calculate the number of completed tasks, the number of uncompleted tasks, and the average completion time in the scheduling round, and calculate the resource utilization ratio based on the logistics task map and the global task allocation plan; S63. Write the real-time scheduling status table, task completion status, and resource usage ratio as the scheduling data for the current round into the designated data area of ​​the database, and establish an index relationship between the task number and the vehicle number for data comparison and network training feedback.

10. A smart logistics scheduling system based on deep learning, executing a smart logistics scheduling method based on deep learning according to any one of claims 1 to 9, characterized in that: include: Data processing module, used to collect logistics data and perform pre-processing; The logistics task module is used to construct a logistics task graph, in which the nodes represent delivery tasks and transport vehicles, and the edges represent the scheduling relationship and distance cost between delivery tasks and transport vehicles; A hierarchical scheduling module is used to build a hierarchical scheduling structure, set up a high-level policy network and a low-level policy network. The high-level policy network divides regions and prioritizes tasks based on the logistics task graph, and generates a task allocation plan. The multi-agent mapping module is used to build a multi-agent mapping relationship, map the transport vehicle agent to the corresponding area, and extract the logistics task subgraph within the area as the input of the low-level policy network; The policy optimization module is used to optimize the logistics task subgraph in each region using the low-level policy network, outputting the specific scheduling actions for all transport vehicles, including departure time decisions, path sequence planning, and task execution order. The update calculation module is used to update the real-time status of each vehicle and calculate the task completion status and resource usage ratio in the current scheduling round, and write it into the system database for data comparison; The training feedback module is used to use the data comparison results as training feedback, update the policy parameters of the high-level policy network and the low-level policy network, and enter the next round of scheduling decisions until all delivery tasks are completed.

Images