Unmanned aerial vehicle dynamic logistics planning method based on large model and workflow

Abstract

The invention discloses an unmanned aerial vehicle dynamic logistics planning method based on a large model and workflow, and the method comprises the steps: 1, carrying out the data collection, and obtaining logistics situation data; and step 2, constructing a logistics situation knowledge graph. And step 3, analyzing and initializing a natural language task. And step 4, sorting target priorities. And 5, generating a strategy feasibility matrix. And step 6, calculating logistics distribution. And 7, processing the logistics planning scheme. And step 8, triggering and executing dynamic re-planning. Through deep fusion of a large model and a workflow engine, real-time perception and quick response to a high-dynamic logistics environment are realized, and a planning scheme can be dynamically adjusted according to a real-time situation; a dynamically updated logistics situation knowledge graph is constructed by introducing a GraphRAG technology, strategy rules and physical constraints are used as hard constraints to be pre-embedded into an optimization process, and the executability and compliance of a generation scheme are ensured; according to the invention, the requirements of the modern efficient logistics environment on the scheduling real-time performance and the scheme reliability can be met.

Description

TECHNICAL FIELD

[0001] The present application relates to a large model and workflow-based unmanned aerial vehicle dynamic logistics planning method, belonging to the field of logistics planning and artificial intelligence. BACKGROUND

[0002] In modern logistics distribution systems, dynamic scheduling is the core link of resource allocation and task planning. Its core task is to scientifically allocate distribution resources under the premise of meeting multiple operation strategies and physical constraints, so as to maximize the overall distribution efficiency. Although there has been a lot of research in this field, the existing methods still have some limitations, which are difficult to adapt to the harsh requirements of real-time, flexibility and reliability in high-dynamic business environment.

[0003] Traditional logistics scheduling systems mostly rely on predefined static rule bases or offline trained optimization models, which lack flexibility and cannot effectively respond to rapid changes in the logistics environment. For example, when orders surge or unmanned aerial vehicle status updates, the effective distribution range of unmanned aerial vehicles, task time limit and other key parameters will fluctuate significantly, and systems based on historical data or fixed rules cannot adjust the scheme in real time, resulting in a high proportion of mismatch between scheduling plans and actual situations, and scheme failure. In addition, in the face of new unmanned aerial vehicle equipment or special order demand, such systems lack immediate reasoning ability and need to rely on manual intervention to reconstruct rules, with a response delay of several hours, which is difficult to support real-time scheduling decisions.

[0004] Another outstanding problem is the separation of operation strategies, physical constraints and optimization algorithms. Most systems check strategy rules outside the optimization process, resulting in schemes that often violate basic physical principles, such as assigning unmanned aerial vehicles with insufficient range to handle long-distance distribution tasks, or ignoring load matching relationships, resulting in much lower execution efficiency than theoretical values. Although methods such as multi-agent reinforcement learning have been introduced to enhance dynamic allocation capabilities, they still have obvious defects: agents are guided by individual rewards, lack of global resource coordination mechanism, and are prone to cause key resources to be exhausted in local areas; at the same time, multi-agent reinforcement learning schemes do not have built-in physical rule checkers, and need additional manual review and correction, which increases the decision-making delay.

[0005] At its root, existing research has focused on single-point improvements, lacking a full-process integrated design from knowledge modeling, dynamic reasoning, real-time verification to scheme execution. Unmanned aerial vehicle logistics scheduling is a complex decision-making process that couples knowledge-driven and data-driven, and urgently needs to build a next-generation architecture that integrates dynamic knowledge base and real-time constraint verification. The system needs to embed physical limitations such as distribution range, load matching, power constraints, etc., to ensure that the generated scheme has high executability; at the same time, a closed-loop mechanism for real-time sensing, decision-making, execution and re-planning must be established, so as to simultaneously guarantee scheduling speed, scheme compliance and resource utilization efficiency in a highly uncertain business environment. SUMMARY

[0006] The present application aims at the defects and deficiencies of the prior art, and proposes a large model and workflow-based unmanned aerial vehicle dynamic logistics planning method. The method constructs a queryable logistics situation knowledge graph through the GraphRAG technology; the logic reasoning ability of the large model, the retrieval ability of the knowledge graph, the precise calculation ability of the traditional optimization algorithm, and the constraint review ability of the rule base are organically integrated by using the workflow engine to form an automated, closed-loop, and interpretable logistics planning intelligent system, which effectively improves the accuracy, compliance, and resource utilization efficiency of the distribution scheme. The method can dynamically respond to changes in the logistics environment, adjust the distribution scheme in real time, and ensure that the planning result meets the operation strategy and resource utilization requirements.

[0007] The technical solution adopted by the present application to solve its technical problems is: a large model and workflow-based unmanned aerial vehicle dynamic logistics planning method. The specific steps of the method include:

[0008] Step 1: Data collection to obtain logistics situation data.

[0009] Unmanned aerial vehicle unit data: unmanned aerial vehicle unit name, coordinates, empty power, load power coefficient, maximum endurance range, maximum load, cruising speed, remaining power, cost coefficient.

[0010] Distribution target data: target name, distribution coordinates, material weight, time efficiency urgency, landing site safety level, business value.

[0011] The workflow engine starts the data collection service, and obtains the logistics situation data in real time to provide basic data support for subsequent knowledge graph construction and large model reasoning.

[0012] Step 2: Construct a logistics situation knowledge graph.

[0013] (21) Data access and processing: process structured, semi-structured, and unstructured data from multiple sources.

[0014] (22) Entity and relationship extraction: use large models and natural language processing techniques to extract entities, relationships, and attributes from text.

[0015] (23) Knowledge fusion and alignment: disambiguate and integrate extracted knowledge to ensure consistency.

[0016] (24) Vector index construction: encode entities, relationships, and text evidence into vectors, build graph index, and realize fast query and traceability.

[0017] (25) Graph storage and query service: store the graph in a graph database and provide efficient query interface to support subsequent reasoning and planning.

[0018] Step 3: Natural Language Task Parsing and Initialization.

[0019] (31) LLM Semantic Parsing: The large language model receives user instructions, parses their core intent, key areas, target types, and expected effects, and defines clear goals and scope for the task.

[0020] (32) GraphRAG Context Enhancement: LLM queries the knowledge graph through GraphRAG to retrieve entities and relationships related to the instruction, forming a rich decision-making context and improving the accuracy and timeliness of decision-making.

[0021] (33) Structured Output: LLM converts ambiguous language instructions into structured JSON format task descriptions, serving as input parameters for the workflow, ensuring that task execution is based on precise and clear instructions.

[0022] Step 4: Target Priority Ranking.

[0023] (41) Construct Pairwise Comparison Matrix: According to the urgency, landing site safety level, and business value criteria, construct a pairwise comparison matrix based on the order and expert scores, and perform consistency testing to ensure the relative importance of the criteria is credible.

[0024] (42) Calculate Weights: Under the premise of consistency, calculate the weight coefficients of each criterion according to the pairwise comparison matrix, obtain w1, w2, w3, and satisfy

[0025] (43) Target Priority Ranking: Rank the target priority according to the following formula

[0026] a j = w1L j + w2S j + w3ρ j

[0027] L j is the urgency of target j, S j is the landing site safety level of target j, ρ j is the business value of target j, w1, w2, w3 are weight coefficients. The final output of the target priority list provides input basis for subsequent steps 5 and 6.

[0028] Step 5: Generate Strategy Feasibility Matrix.

[0029] (51) Preliminary Strategy Feasibility Matrix Generation:

[0030] The workflow engine calls the large model inference service to generate a binary strategy feasibility matrix according to the task instruction, the situation data, and the strategy rule. The matrix is an n*m preliminary strategy feasibility matrix F, where n represents the number of unmanned aerial vehicle units, m represents the number of target quantities of the distribution party, and the matrix elements are 0 or 1. Each element F ij in the matrix indicates whether the target j can be distributed by the unmanned aerial vehicle unit i. An example of the matrix F is as follows:

[0031]

[0032] 1: indicates that the target can be distributed by the unmanned aerial vehicle unit.

[0033] 0: indicates that the target cannot be distributed by the unmanned aerial vehicle unit.

[0034] The strategy rule includes distance feasibility rules, load capacity feasibility rules, business matching rules, and other multi-dimensional rules. The purpose is to filter out obviously infeasible "unmanned aerial vehicle unit-target" pairs in advance before the optimization algorithm runs, thereby significantly reducing the search space and improving the efficiency of the algorithm.

[0035] (52) Preprocessing of the strategy feasibility matrix:

[0036] This step identifies and isolates infeasible targets, mainly for targets that cannot be distributed due to insufficient unmanned aerial vehicle capacity.

[0037] (52.1) Identification: For target j, if the value of all elements in the corresponding matrix column vector is 0, the target is marked as an "infeasible target".

[0038] (52.2) Isolation: All marked "infeasible targets" will be removed from the optimization target set of this logistics planning and added to a list of infeasible targets for temporary storage, which will be input to step 7.

[0039] (52.3) Reason tracing: Query the root cause of the infeasible target through the knowledge graph and rule base.

[0040] (52.4) Update matrix: The updated matrix F' contains only the columns corresponding to feasible targets.

[0041] (53) Resource conflict detection and dynamic arbitration:

[0042] After ensuring global feasibility in step (52), this step aims to solve the problem of exclusive resource conflicts between remaining feasible targets, i.e., the scenario where multiple targets compete for the same unique available unmanned aerial vehicle unit. This step simulates the decision-making process of the delivery personnel through automated arbitration based on target priority.

[0043] (53.1) Conflict detection: The workflow engine inputs the matrix F' outputted in step (52) into the conflict detection module, and detects the conflicts between targets and UAV units by loop.

[0044] (53.2) Intelligent arbitration: For each conflict UAV unit i and its corresponding conflict target set C i , the workflow engine starts the arbitrator, calls the large model service, and the arbitrator adopts the target priority arbitration strategy to select the target j * with the highest priority from the conflict target set C i , and assigns the UAV unit i to the target j * .

[0045] (53.3) Matrix update: Update the matrix F' according to the arbitration result, remove the conflict targets and convert them to "to-be-planned targets", input the to-be-planned target list to step 7, and finally the workflow engine will input the complete strategy feasibility matrix F' into the optimization algorithm.

[0046] Step 6: Calculate logistics allocation.

[0047] (61) Input necessary parameters:

[0048] Strategy feasibility matrix, UAV unit attributes, and delivery target attributes.

[0049] (62) Initialize ant colony:

[0050] Each ant represents a "UAV unit-target" resource allocation scheme, and the position of the ant represents the pairing combination of the target and the UAV unit. Define the allocation vector J = (j1, j2, j3, …, j n ) generated by the ant. j i is the target j assigned to the UAV unit i, and the index is shown as follows:

[0051] J = [1, 3, 5, 2, 4, 1]

[0052] This means that "UAV unit-1" transports "target-1", "UAV unit-2" transports "target-3", and so on. The ants randomly generate an initial resource allocation scheme as the starting point of the search.

[0053] (63) Constraints:

[0054] Strategy feasibility matrix as a hard constraint: In the calculation process of the ant colony algorithm, all delivery target and UAV unit pairings must meet the legality requirements in the matrix. That is, only the position of F' ij = 1 can be selected.

[0055] Each UAV unit can only deliver one target: each UAV unit can only deliver one target in a task. This can be expressed by the following constraint expression:

[0056]

[0057] where x ij is the decision variable, indicating whether UAV unit i delivers target j.

[0058] (64) Time and cost calculation:

[0059] Let the planned path length of UAV i and delivery target j be L ij , considering the round-trip path length, the cruising speed is v i . The single round-trip transportation duration calculation formula is as follows:

[0060]

[0061] The power consumption calculation formula is as follows:

[0062]

[0063] s.t.E ij <R i

[0064] P i is the power of the UAV when empty, W j is the weight of the target material, k is the power coefficient of the UAV load, R i is the remaining power of the UAV.

[0065] The total duration calculation formula is as follows:

[0066]

[0067] The total cost calculation formula is as follows:

[0068]

[0069] c i is the unit time cost coefficient of the UAV. d i is the unit power cost coefficient of the UAV. L ij is the unit planned path length of the UAV. E ij is the unit power consumption of the UAV.

[0070] (65) Fitness function design:

[0071] Calculate the fitness of each ant, and the fitness function includes the following factors.

[0072] Delivery time: the optimization goal is the total delivery time, to ensure that the order is completed on time.

[0073] Delivery cost: The optimization goal is to minimize the total delivery cost.

[0074] Target priority: Ensure that high-priority targets are delivered first.

[0075] The fitness function is as follows:

[0076]

[0077] Where: a j : Target priority of target j. λ: Weight coefficient.

[0078] (66) State transition probability and pheromone update:

[0079] State transition probability calculation formula:

[0080]

[0081] Where:

[0082] Probability of UAV unit i selecting target j.

[0083] τ ij : Pheromone concentration, indicating the importance of UAV unit i to target j, higher pheromone means more inclined to choose this path.

[0084] η ij : Heuristic information, indicating η ij = a j , the higher the priority of the target, the greater the heuristic information, and the more inclined the ant is to choose this target for delivery.

[0085] F [i][j] : Strategy feasibility matrix, indicating whether UAV unit i delivers target j, i.e., whether it meets the strategy rules.

[0086] α,β: Importance weight of pheromone and importance weight of heuristic information, respectively. By adjusting these two parameters, the relative influence of pheromone and heuristic information in the selection process can be controlled.

[0087] Pheromone update:

[0088] Ants update the pheromone concentration through fitness, and paths with higher fitness will obtain more pheromone, guiding subsequent ants to preferentially choose this path. Pheromone update formula:

[0089] τ ij (t+1) = (1-ρ)*τ ij (t) + △τ ij

[0090] ρ: Volatility rate of pheromone.

[0091] Δτ ij : Fitness contribution calculated according to pheromone update formula, the increased amount of pheromone is usually proportional to the fitness of the path, Δτ ij = Q*Fit, where Q is a constant and Fit is the fitness value of the ant.

[0092] (67) Iteration and convergence:

[0093] The ant colony algorithm gradually updates the pheromone concentration through multiple iterations. Paths with higher fitness will attract more ants to choose, eventually converging to the optimal pairing solution J, outputting the optimal allocation scheme.

[0094] Step 7: Process the logistics planning scheme.

[0095] (71) Scheme analysis and semantic mapping:

[0096] The system receives and analyzes the optimal allocation scheme output in step 6, which includes the allocation vector J, and the large model performs deep semantic understanding of the scheme to identify all valid "drone unit-target" delivery pairs.

[0097] (72) Knowledge graph query and instruction enhancement:

[0098] For each valid delivery pair, the LLM queries the GraphRAG logistics situation knowledge graph to obtain relevant context information and generates structured JSON format logistics delivery instructions to ensure that each parameter is traceable.

[0099] (73) Output:

[0100] Logistics delivery instructions: Based on the aforementioned scheme and query results, generate structured and executable logistics delivery instructions for downstream automated warehouse systems or unmanned aerial vehicle flight control systems to execute.

[0101] Infeasible target list: Contains targets that cannot be delivered due to resource condition restrictions, recording the reasons why they cannot be delivered.

[0102] To-be-planned target list: Records targets that cannot be allocated due to resource conflicts or strategy adjustments, indicating that these targets will be re-evaluated in the next planning cycle.

[0103] Step 8: Dynamic re-planning trigger and execution.

[0104] (81) Trigger conditions:

[0105] When one of the following conditions is met, a new logistics planning cycle will be triggered:

[0106] Planning instruction trigger: the user issues a new logistics planning instruction, starting a new planning cycle.

[0107] Resource change trigger: changes in logistics situation, such as completion of tasks by unmanned aerial vehicles, refresh of power supply or update of resources, trigger re-evaluation of the to-be-planned targets and unfeasible targets.

[0108] (82) Perform re-planning:

[0109] For the new targets, first complete the incremental warehousing and graph updating according to steps 1 and 2; then merge the new targets with the to-be-planned targets and unfeasible targets output in step 7, and send them to step 3 to start a new planning instance. When planning, the to-be-planned targets are preferentially allocated; the unfeasible targets are re-evaluated based on the latest resources and situation, and if the conditions are met, they are included in the allocation. The successfully allocated targets are removed from the original list and enter execution; the new targets with incomplete data are marked as "to-be-warehoused" and do not participate in this round.

[0110] Beneficial effects:

[0111] 1. The present application realizes real-time perception and rapid response to high-dynamic logistics environment through deep integration of large models and workflow engines, and can dynamically adjust the planning scheme according to real-time situation.

[0112] 2. The present application constructs a dynamically updated logistics situation knowledge graph by introducing GraphRAG technology, and embeds strategy rules and physical constraints as hard constraints in the optimization process in advance, which fundamentally ensures the execution and compliance of the generated scheme.

[0113] 3. The present application constructs a closed-loop dynamic re-planning mechanism of "perception-decision-execution-re-planning", and the system can adjust itself in real time according to the resource state and order changes during task execution, thereby significantly improving the resource utilization efficiency and decision response speed, and meeting the stringent requirements of modern efficient logistics environment on scheduling real-time and scheme reliability. BRIEF DESCRIPTION OF DRAWINGS

[0114] Figure 1 Workflow diagram of the present application.

[0115] Figure 2 Step diagram for generating strategy feasibility matrix of the present application.

[0116] Figure 3 Algorithm step diagram of the present application.

[0117] Figure 4 Re-planning step diagram of the present application.

[0118] Figure 5 Target priority distribution diagram of the present application.

[0119] Figure 6The strategy feasibility matrix of the application.

[0120] Figure 7 The algorithm convergence graph of the application.

[0121] Figure 8 The distribution allocation result graph of the application.

[0122] Figure 9 The logistics allocation instruction effect graph of the application.

[0123] Figure 10 The dynamic re-planning effect graph of the application. DETAILED DESCRIPTION

[0124] To make the purpose, technical solutions and advantages of the application clearer and more apparent, the application will be further described in detail below with reference to the drawings.

[0125] As Figure 1 shown, the application provides a large model and workflow-based unmanned aerial vehicle dynamic logistics planning method, and the specific implementation steps are as follows:

[0126] Step 1: Data collection is performed to obtain logistics situation data.

[0127] Unmanned aerial vehicle unit data: unmanned aerial vehicle unit name, coordinates, empty power, load power coefficient, maximum endurance mileage, maximum load, cruising speed, remaining power, cost coefficient.

[0128] Distribution target data: target name, distribution coordinates, material weight, time efficiency urgency, landing site safety level, business value.

[0129] The workflow engine starts the data collection service, obtains real-time logistics situation data, and provides basic data support for subsequent knowledge graph construction and large model reasoning.

[0130] Step 2: Construction of logistics situation knowledge graph.

[0131] (21) Data access and processing: processing structured, semi-structured and unstructured data from multiple sources.

[0132] (22) Entity and relationship extraction: using large models and natural language processing technology to extract entities, relationships and attributes from text.

[0133] (23) Knowledge fusion and alignment: disambiguate and integrate extracted knowledge to ensure consistency.

[0134] (24) Vector index construction: encode entities, relationships and text evidence into vectors, build graph index, realize fast query and traceability.

[0135] (25) Graph Storage and Query Service: Store graphs in a graph database and provide efficient query interfaces to support subsequent reasoning and planning.

[0136] Step 3: Natural Language Task Parsing and Initialization.

[0137] (31) LLM Semantic Analysis: The large language model receives user instructions, analyzes their core intent, key areas, target types, and expected effects, and defines clear goals and scope for the task.

[0138] (32) GraphRAG Context Enhancement: LLM queries the knowledge graph through GraphRAG to retrieve entities and relationships related to the instruction, forming a rich decision-making context and improving the accuracy and timeliness of decisions.

[0139] (33) Structured Output: LLM converts ambiguous language instructions into structured JSON format task descriptions, serving as input parameters for the workflow, ensuring that task execution is based on precise and clear instructions.

[0140] Step 4: Target Priority Ranking.

[0141] This step calls code service nodes through the workflow engine and calculates target priorities based on the analytic hierarchy process.

[0142] (41) Construct Pairwise Comparison Matrix: According to the urgency of time, the safety level of landing site, and business value, construct a pairwise comparison matrix based on the order and expert scores, and perform consistency check to ensure the relative importance of the criteria is credible.

[0143] (42) Calculate Weights: Under the premise of consistency, calculate the weight coefficients of each criterion according to the pairwise comparison matrix, get w1, w2, w3, and satisfy

[0144] (43) Target Priority Ranking: Rank the target priorities according to the following formula

[0145] a j = w1L j + w2S j + w3ρ j

[0146] L j is the urgency of time for target j, S j is the safety level of landing site for target j, ρ j is the business value of target j, w1, w2, w3 are weight coefficients. The final output of the target priority list provides input basis for subsequent steps 5 and 6.

[0147] Step 5: Generate Strategy Feasibility Matrix.

[0148] Step 5 generates a strategy feasibility matrix, which is the key to precise resource screening. The detailed execution logic is shown in Figure 2 .

[0149] (51) Preliminary strategy feasibility matrix generation:

[0150] The workflow engine calls the large model inference service to generate a binary strategy feasibility matrix according to the task instructions, situation data, and strategy rules. This matrix is an n*m preliminary strategy feasibility matrix F, where n represents the number of UAV units and m represents the number of target quantities. The matrix elements are 0 or 1, and each element F ij in the matrix represents whether target j can be delivered by UAV unit i. An example of the matrix F is as follows:

[0151]

[0152] 1: indicates that the target can be delivered by the UAV unit.

[0153] 0: indicates that the target cannot be delivered by the UAV unit.

[0154] The strategy rules include distance feasibility rules, load capacity feasibility rules, business matching rules, and other multi-dimensional rules. The purpose is to pre-filter out obviously infeasible "UAV unit-target" pairs before running the optimization algorithm, thereby significantly reducing the search space and improving algorithm efficiency.

[0155] (52) Preprocessing strategy feasibility matrix:

[0156] This step identifies and isolates infeasible targets, mainly targeting targets that cannot be delivered due to insufficient UAV capabilities.

[0157] (52.1) Identification: For target j, if all elements in its corresponding matrix column vector are 0, the target is marked as an "infeasible target".

[0158] (52.2) Isolation: All marked "infeasible targets" will be removed from the optimization target set of this logistics planning and added to a list called infeasible targets for temporary storage. This list will be input to Step 7.

[0159] (52.3) Reason tracing: Query the root cause of the infeasible target through the knowledge graph and rule base.

[0160] (52.4) Update matrix: The updated matrix F', only contains the columns corresponding to feasible targets.

[0161] (53) Resource conflict detection and dynamic arbitration:

[0162] After ensuring global feasibility in step (52), this step aims to resolve the problem of exclusive resource conflicts between remaining feasible targets, i.e., the scenario where multiple targets compete for the same unique available drone unit. This step simulates the decision-making process of the delivery personnel through automated arbitration based on target priority.

[0163] (53.1) Conflict detection: The workflow engine inputs the matrix F' output from step (52) into the conflict detection module, which detects target and drone unit conflicts through a loop.

[0164] (53.2) Intelligent arbitration: For each conflicting drone unit i and its corresponding conflict target set C i , the workflow engine starts the arbitrator, calls the large model service, and the arbitrator adopts the target priority arbitration strategy to select the highest priority target j i from the conflict target set C * and assigns drone unit i to target j * .

[0165] (53.3) Matrix update: Update the matrix F' according to the arbitration result, remove the conflict target and convert it to "to-be-planned target", input the to-be-planned target list to step 7, and finally the workflow engine will input the complete strategy feasibility matrix F' into the optimization algorithm.

[0166] Step 6: Calculate logistics distribution.

[0167] After completing the feasible pairing screening, step 6 calculates the optimal distribution through the ant colony algorithm with rule enhancement, the core steps of which are shown in Figure 3 .

[0168] (61) Input necessary parameters:

[0169] Strategy feasibility matrix, drone unit attributes, delivery target attributes.

[0170] (62) Initialize the ant colony:

[0171] Each ant represents a "drone unit-target" resource allocation scheme, and the position of the ant represents the pairing combination of the target and the drone unit. Define the allocation vector J = (j1, j2, j3, …, j n ) generated by the ant. j i is the target j assigned to drone unit i, and the index is shown as follows:

[0172] J = [1, 3, 5, 2, 4, 1]

[0173] This means "drone unit-1" transports "target-1", "drone unit-2" transports "target-3", and so on. Ants randomly generate an initial resource allocation scheme as the starting point of the search.

[0174] (63) Constraints:

[0175] Strategy feasibility matrix as a hard constraint: In the calculation process of the ant colony algorithm, all the distribution targets and drone units must meet the legality requirements in the matrix. That is, only the position of F' ij = 1 can be selected.

[0176] Each drone unit can only distribute one target: Each drone unit can only distribute one target in a task. This can be expressed by the following constraint expression:

[0177]

[0178] Where x ij is the decision variable, indicating whether drone unit i distributes target j.

[0179] (64) Time and cost calculation:

[0180] Let the planned path length of drone i and distribution target j be L ij , considering the round-trip path length, and the cruising speed be v i . The single round transportation duration calculation formula is as follows:

[0181]

[0182] The power consumption calculation formula is as follows:

[0183]

[0184] s.t. E ij <R i

[0185] P i is the power of the drone when empty, W j is the weight of the target material, k is the power coefficient of the drone load, and R i is the remaining power of the drone.

[0186] The total duration calculation formula is as follows:

[0187]

[0188] The total cost calculation formula is as follows:

[0189]

[0190] ci is the cost coefficient of the UAV per unit of time. i is the cost coefficient of the UAV per unit of electrical energy. ij is the length of the planned path per unit of the UAV. ij is the electrical energy consumed per unit of the UAV.

[0191] (65) Fitness function design:

[0192] The fitness of each ant is calculated, and the fitness function includes the following factors.

[0193] Delivery time: The optimization goal is the total delivery time to ensure that orders are completed on time.

[0194] Delivery cost: The optimization goal is the total delivery cost to avoid unnecessary waste of resources.

[0195] Target priority: Ensure that high-priority targets are delivered first.

[0196] The fitness function is as follows:

[0197]

[0198] where: j is the target priority of target j. λ is the weight coefficient.

[0199] (66) State transition probability and pheromone update:

[0200] State transition probability calculation formula:

[0201]

[0202] where:

[0203] is the probability of UAV unit i selecting target j.

[0204] ij is the pheromone concentration, indicating the importance of UAV unit i to target j, and the higher the pheromone means the more inclined to choose this path.

[0205] ij is the heuristic information, indicating that η ij = a j , the higher the priority of the target, the greater the heuristic information, and the ant is more inclined to choose this target for delivery.

[0206] F [i][j] : Strategy feasibility matrix, indicating whether UAV unit i delivers target j, i.e., whether it meets the strategy rules.

[0207] ​​α, β: the importance weight of pheromone and the importance weight of heuristic information, respectively. By adjusting these two parameters, the relative influence of pheromone and heuristic information in the selection process can be controlled.

[0208] Pheromone update:

[0209] Ants update pheromone concentration by fitness, paths with higher fitness will get more pheromone, guiding subsequent ants to preferentially select this path. Pheromone update formula:

[0210] τ ij (t+1) = (1-ρ)*τ ij (t) +△τ ij

[0211] ρ: pheromone evaporation rate.

[0212] △τ ij : fitness contribution calculated according to the pheromone update formula, the amount of pheromone added is usually proportional to the fitness of the path, △τ ij = Q*Fit, where Q is a constant and Fit is the fitness value of this ant.

[0213] (67) Iteration and convergence:

[0214] Ant colony algorithm gradually updates pheromone concentration through multiple iterations. Paths with higher fitness will attract more ants to select, eventually converging to the optimal pairing scheme J, outputting the optimal allocation scheme.

[0215] Step 7: Process logistics planning scheme.

[0216] (71) Scheme analysis and semantic mapping:

[0217] The system receives and analyzes the optimal allocation scheme output in step 6, which includes the allocation vector J, and the large model performs deep semantic understanding of the scheme to identify all valid "drone unit-target" delivery pairs.

[0218] (72) Knowledge graph query and instruction enhancement:

[0219] For each valid delivery pair, the LLM queries the GraphRAG logistics situation knowledge graph to obtain relevant context information and generates structured JSON format logistics delivery instructions to ensure that each parameter is traceable.

[0220] (73) Output:

[0221] Logistics delivery instructions: Based on the aforementioned scheme and query results, structured and executable logistics delivery instructions are generated for downstream automated warehouse systems or unmanned aerial vehicle flight control systems to execute.

[0222] Unfeasible Target List: Contains targets that cannot be dispatched due to resource constraints, recording the current reasons for unfeasibility.

[0223] To-be-planned Target List: Records targets that cannot be allocated due to resource conflicts or policy adjustments, indicating that these targets will be re-evaluated in the next planning cycle.

[0224] Step 8: Dynamic re-planning trigger and execution.

[0225] To cope with the dynamic changes in the logistics environment, the present invention designs a re-planning mechanism as shown in Figure 4 to ensure that the system can continuously respond to real-time status.

[0226] (81) Trigger conditions:

[0227] When one of the following conditions is met, a new logistics planning cycle will be triggered:

[0228] Planning instruction trigger: The user issues a new logistics planning instruction, starting a new planning cycle.

[0229] Resource change trigger: Changes in logistics situation, such as completion of tasks by drones, refresh of power, or updates of resources, trigger re-evaluation of to-be-planned targets and unfeasible targets.

[0230] (82) Execute re-planning:

[0231] First, complete incremental storage and graph update according to steps 1 and 2 for new targets; then merge new targets with to-be-planned targets and unfeasible targets output in step 7, and send them to step 3 to start a new planning instance. Prioritize allocation of to-be-planned targets during planning; simultaneously re-evaluate unfeasible targets based on the latest resources and situation, and include them in allocation if conditions are met. Successfully allocated targets are removed from the original list and enter execution; data-incomplete new targets are marked as "to-be-stored" and do not participate in this round.

[0232] Example analysis:

[0233] In a preferred embodiment of the present invention, 50 drones and 50 delivery targets are configured, and a data collection service is started through a workflow engine to obtain real-time logistics situation data. The drones are distributed in the range of 3.12-9.95 kg according to their actual load capacity, with an average load of 6.53 kg, a range of 40.47-79.57 km, and an average range of 59.34 km. These drones are deployed in three distribution centers: 20 in the west distribution center, 17 in the central distribution center, and 13 in the east distribution center, forming a delivery network covering the entire area. The delivery targets include five types: hospital emergency points, detection sites, field hospitals, isolation points, and community delivery points, each type having different location distribution characteristics and priority attributes.

[0234] The system first constructs a logistics situation knowledge graph, integrates multi-source data through GraphRAG technology, and forms a knowledge graph. After receiving the user's natural language instruction "please start logistics planning based on existing data", the large language model analyzes the instruction, and clearly defines the distribution related targets, priority criteria and business value, etc. As shown in the figure, Figure 5 The target priority calculation result based on the analytic hierarchy process shows that the average priority of hospital emergency targets is 0.909, the detection site is 0.815, the shelter hospital is 0.737, the isolation point is 0.630, and the community distribution point is 0.525, forming a clear task priority gradient.

[0235] As shown in the figure, Figure 6 In the strategy feasibility matrix generation process, 1860 unfeasible pairs are filtered out from the original 2500 "drone-target" pairs, and finally 640 feasible pairs are retained, with a filtering ratio of 74.4%. Among them, 5 targets are completely unassignable due to distance or weight constraints, which reflects the system's precise screening ability under complex constraints. As shown in the figure, Figure 7 The rule-enhanced ant colony algorithm is used for optimization search, and after 200 iterations, the optimal fitness reaches 0.4499, showing good convergence characteristics.

[0236] The final allocation scheme realizes 41 effective distribution pairs, covering 82% of the distribution targets, and the spatial allocation result is shown in the figure. Figure 8 Specifically, the west distribution center completes 11 distribution tasks with a total cost of 595.91 yuan and an average distribution distance of 15.92 km; the central distribution center completes 17 distribution tasks with a total cost of 782.00 yuan and an average distribution distance of 11.63 km; the east distribution center completes 13 distribution tasks with a total cost of 492.47 yuan and an average distribution distance of 10.79 km. The total distribution cost of the system is 1870.38 yuan, and the total distribution time is 68.41 hours. As shown in the figure, Figure 9 The system generates structured logistics distribution instructions based on the scheme, ensuring the executability of the scheme.

[0237] This example demonstrates the significant performance advantages of the present application: in the high-priority target protection aspect, the hospital emergency target allocation rate reaches 88.9%, and the detection site reaches 90%; in the resource utilization aspect, the drone utilization rate reaches 82%; in the spatial optimization aspect, the allocation scheme fully utilizes the location advantages of the three distribution centers, achieving balanced spatial coverage. As shown in the figure, Figure 10 The system exhibits efficient dynamic re-planning capability under state update, and can quickly respond to environmental changes.

[0238] The implementation of the present application should not be limited to the above-mentioned embodiments, and anyone should know that any technical solutions made under the inspiration of the present application and having the same or similar technical solutions as the present application fall within the protection scope of the present application. The present example proves the technical advantages of the present application in emergency material distribution and provides reliable technical support for intelligent logistics scheduling.

Claims

1. A large model and workflow-based unmanned aerial vehicle dynamic logistics planning method, characterized in that, The method steps are as follows: Step 1: data collection, obtaining logistics situation data; Step 2: building a logistics situation knowledge graph; Step 3: natural language task analysis and initialization; Step 4: target priority sorting; Step 5: generating a strategy feasibility matrix; Step 6: calculating logistics distribution; Step 7: processing logistics planning schemes; Step 8: dynamic re-planning triggering and execution.

2. The large model and workflow-based dynamic UAV logistics planning method according to claim 1, characterized in that, The division steps of step 2 are as follows: Step (21) data access and processing: processing structured, semi-structured and unstructured data from multiple sources; Step (22) entity and relationship extraction: using large models and natural language processing techniques to extract entities, relationships and attributes from text; Step (23) knowledge fusion and alignment: disambiguating and integrating extracted knowledge to ensure consistency; Step (24) vector index construction: encoding entities, relationships and text evidence into vectors, building graph index for fast query and traceability; Step (25) graph storage and query service: storing the graph in a graph database and providing efficient query interface to support subsequent reasoning and planning.

3. The method of claim 1, wherein, The division steps of step 3 are as follows: Step (31) LLM semantic analysis: large language model receives user instructions, analyzes its core intent, key areas, target type and expected effect, defines clear target and scope for the task; Step (32) GraphRAG context enhancement: LLM queries knowledge graph through GraphRAG, retrieves entities and relationships related to instructions, forms rich decision-making context, improves decision-making accuracy and timeliness; Step (33) structured output: LLM converts fuzzy language instructions into structured JSON format task description as input parameters for workflow, ensures task execution based on accurate and clear instructions.

4. The method of claim 1, wherein, The division steps of step 4 are as follows: Call code service nodes through workflow engine, calculate target priority according to AHP; Step (41) build pairwise comparison matrix: according to the order of time urgency, landing site safety level and business value, build pairwise comparison matrix according to orders and expert scoring, and conduct consistency test to ensure the credibility of the relative importance of the criteria; Step (42) calculates the weight: under the premise of consistency passing, the weight coefficient of each criterion is calculated according to the pair comparison matrix, w1, w2, w3 are obtained, and satisfy Step (43) target priority sorting: target priority sorting according to the following formula a j = w1L j + w2S j + w3p j L j aging urgency of target j, S j landing site safety level of target j, p j business value of target j, w1, w2, w3 are weight coefficients; the final output target priority list provides input basis for subsequent steps 5 and step 6.

5. The large model and workflow-based dynamic UAV logistics planning method according to claim 1, wherein, The division steps of step 5 are as follows: Step (51) initial strategy feasibility matrix generation: The workflow engine calls the large model inference service to generate a binary strategy feasibility matrix according to the task instruction, the situation data, and the strategy rule. The matrix is an n*m preliminary strategy feasibility matrix F, where n represents the number of unmanned aerial vehicle units, m represents the number of target quantities of the distribution party, and the matrix elements are 0 or 1. Each element F ij in the matrix indicates whether the target j can be distributed by the unmanned aerial vehicle unit i; for example, the matrix F is as follows: 1: indicates that the target can be delivered by the UAV unit; 0: indicates that the target cannot be delivered by the UAV unit; Strategy rules include distance feasibility rules, load capacity feasibility rules, business matching rules and other multi-dimensional rules, the purpose is to filter out obviously infeasible "UAV unit-target" pairs in advance through a series of clear strategy rules before optimization algorithm runs, thereby significantly reducing search space and improving algorithm efficiency; Step (52) preprocessing strategy feasibility matrix: Identify and isolate infeasible targets, mainly for targets that cannot be delivered due to insufficient UAV capacity; Step (52.1) Identification: For target j, if all the elements in its corresponding column vector of matrix F are 0, then the target is marked as "unfeasible target"; Step (52.2) Isolation: All the marked "unfeasible targets" will be removed from the optimization target set of this logistics planning and added to a list called "unfeasible targets" for temporary storage, which will be input into Step 7; Step (52.3) Cause tracing: Query the root cause of the unfeasible target through the knowledge graph and rule base; Step (52.4) Update matrix: The updated matrix F' contains only the columns corresponding to the feasible targets; Step (53) Resource conflict detection and dynamic arbitration: After ensuring global feasibility in Step (52), this step aims to solve the problem of exclusive resource conflict between the remaining feasible targets, i.e., the scenario where multiple targets compete for the same unique available drone unit; this step simulates the decision-making process of the delivery personnel through automated arbitration based on target priority; Step (53.1) Conflict detection: The workflow engine inputs the matrix F' output from Step (52) into the conflict detection module, which detects target and drone unit conflicts through a loop; Step (53.2) Intelligent Arbitration: For each conflict UAV unit i and its corresponding conflict target set C i , the workflow engine starts an arbitrator, calls a large model service, and the arbitrator selects the target j * with the highest priority from the conflict target set C i using the target priority arbitration strategy, and assigns the UAV unit i to the target j * ; Step (53.3) Matrix update: Update the matrix F' according to the arbitration results, remove the conflict targets and convert them to "to-be-planned targets", input the to-be-planned target list into Step 7, and finally the workflow engine will input the complete strategy feasibility matrix F' into the optimization algorithm.

6. The large model and workflow-based dynamic UAV logistics planning method of claim 1, wherein: The division step of Step 6 is as follows: Step (61) Input necessary parameters: Strategy feasibility matrix, drone unit attributes, delivery target attributes Step (62) Initialize ants: Each ant represents a "drone unit-target" resource allocation scheme, and the position of the ant represents the paired combination of target and drone unit; the allocation vector J = (j1, j2, j3,..., jn generated by the ant is defined as follows: ) ; j i Target j assigned to drone unit i, with index shown as follows: J＝[1，3，5，2，4，1] This means that "drone unit-1" transports "target-1", "drone unit-2" transports "target-3", and so on; ants randomly generate an initial resource allocation scheme as the starting point for search; Step (63) Constraint conditions: Strategy feasibility matrix as a hard constraint: in the process of ant colony algorithm calculation, all the distribution targets and UAV unit pairs must meet the legal requirements in the matrix; that is, only the position of F′ ij = 1 can be selected; Each drone unit can only deliver one target: each drone unit can only deliver one target in a task; this can be expressed by the following constraint expression: where x ij is the decision variable, representing whether the UAV unit i delivers the target j; Step (64) Time and cost calculation: Let the planned path length of the UAV i and the delivery target j be L ij , considering the round-trip path length, the cruising speed is v i ; the single-round transportation duration calculation formula is as follows: The formula for calculating the power consumption is as follows: s.t.F ij <R i P i W is the power of the UAV in the air, W j k is the power coefficient of the UAV, R i The remaining power for the drone; The formula for calculating the total duration is as follows: The formula for calculating the total cost is as follows: c i is the cost coefficient of the UAV per unit time;d i is the cost coefficient of the UAV per unit electric energy;L ij is the length of the planned path of the UAV per unit;E ij is the electric energy consumed by the UAV per unit. Step (65) Design of fitness function: Calculate the fitness of each ant, and the fitness function includes the following factors; Delivery time: the optimization target is the total delivery time to ensure that orders are completed on time; Delivery cost: the optimization target is the total delivery cost to avoid unnecessary resource waste; Target priority: ensure that high-priority targets are delivered first; The fitness function is as follows: wherein: a j : target priority of target j; λ: weight coefficient; Step (66) State transition probability and pheromone update: State transition probability calculation formula: Where: probability that drone unit i selects target j; τ ij : pheromone concentration, representing the selection importance of the UAV unit i to the target j, the higher pheromone means the more inclined to select this path; η ij : heuristic information, representing η ij = a j , the higher the priority of the target, the greater the heuristic information, and the more the ant tends to select this target for delivery; F [i][j] : strategy feasibility matrix, indicating whether the UAV unit i delivers the target j, i.e., whether it meets the strategy rule; α, β: importance weight of pheromone and importance weight of heuristic information respectively; by adjusting these two parameters, the relative influence of pheromone and heuristic information in the selection process can be controlled; Pheromone update: Ants update the pheromone concentration through fitness, and paths with higher fitness will obtain more pheromone, guiding subsequent ants to preferentially select this path; the pheromone update formula is as follows: τ ij (t+1) = (1 - p) * τ ij (t) + Δτ ij ρ: pheromone evaporation rate; Δτ ij : The fitness contribution calculated according to the pheromone update formula, the amount of pheromone increased is usually proportional to the fitness of the path, Δτ ij = Q * Fit, where Q is a constant and Fit is the fitness value of the ant; Step (67) iteration and convergence: The ant colony algorithm gradually updates the pheromone concentration through multiple iterations; paths with higher fitness will attract more ants to choose, and eventually converge to the optimal matching scheme J, outputting the optimal allocation scheme.

7. The large model and workflow-based dynamic UAV logistics planning method of claim 1, wherein, The division step of step 7 is specifically: Step (71) scheme analysis and semantic mapping: The system receives and analyzes the optimal allocation scheme output by step 6, which includes the allocation vector J, and the large model performs deep semantic understanding of the scheme to identify all valid "drone unit-target" delivery pairs; Step (72) knowledge graph query and instruction enhancement: For each valid delivery pair, the LLM queries the GraphRAG logistics situation knowledge graph to obtain relevant context information and generates structured JSON format logistics delivery instructions to ensure that each parameter is traceable; Step (73) output: Logistics delivery instructions: Based on the aforementioned scheme and query results, generate structured and executable logistics delivery instructions for downstream automated warehouse systems or unmanned aerial vehicle flight control systems to execute; Infeasible target list: Contains targets that cannot be delivered due to resource condition restrictions, recording the reasons why they cannot be delivered; To-be-planned target list: Records targets that cannot be allocated due to resource conflicts or strategy adjustments, indicating that these targets will be re-evaluated in the next planning cycle.

8. The large model and workflow-based dynamic UAV logistics planning method of claim 1, wherein, The division step of step 8 is specifically: Step (81) trigger condition: When one of the following conditions is met, a new logistics planning cycle will be triggered: Planning instruction trigger: The user issues a new logistics planning instruction, starting a new planning cycle; Resource change trigger: Changes in logistics situation, such as drones completing tasks, battery refresh, or resource updates, trigger re-evaluation of to-be-planned targets and infeasible targets; Step (82) execute re-planning: Newly added targets are first processed according to steps 1 and 2 to complete incremental warehousing and graph updates; then, the newly added targets and the to-be-planned targets and infeasible targets output by step 7 are combined and sent to step 3 to start a new planning instance; prioritize allocating to-be-planned targets during planning; Simultaneously re-evaluate infeasible targets based on the latest resources and situation; if the conditions are met, they will be included in the allocation; successfully allocated targets are removed from the original list and enter execution; data-incomplete new targets are marked as "to-be-warehoused" and do not participate in this round.

Images