Intelligent logistics management system based on large language model

Abstract

The invention provides an intelligent logistics management system based on a large language model, and relates to the field of intelligent logistics management. According to the method, a low-rank adaptation technology general large language model is adopted in advance for fine tuning, and an external LLM service obtained through fine tuning is innovatively and deeply fused into a task decision process, so that deep understanding of user intentions and accurate analysis of complex environment situations can be realized, and an optimal distribution scheme is autonomously generated based on the deep understanding of the user intentions and the accurate analysis of the complex environment situations. Besides, the urban environment management module, the logistics management scheduling module, the simulation engine module and the large language model interface module included in the system do not work independently, but closely cooperate through an event-driven and message-passing mechanism, so that a complete unmanned aerial vehicle logistics management closed loop from environment perception to decision making and execution to event feedback is formed. The architecture can reflect and control the distribution task of the unmanned aerial vehicle in real time, and carries out self-correction on a set scheme, thereby finally remarkably improving the intelligent level and operation efficiency of logistics management.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent logistics management, and particularly relates to an intelligent logistics management system based on a large language model. BACKGROUND

[0002] The application prospect of unmanned aerial vehicles (UAVs) in the logistics industry is very broad, and urban UAV distribution systems are gradually becoming an important force driving the transformation of the modern logistics industry.

[0003] In the prior art, urban UAV distribution systems mainly use the following two types for task scheduling and resource allocation:

[0004] Rule-based UAV distribution expert systems, which generally perform static scheduling of UAV task execution through pre-set rules. For example, the paper "Research on Urban UAV and Public Transportation Collaborative Transportation System" (Li Dong, Li Hongtao, 2023) proposes a rule-driven campus scenario distribution model, which realizes the automation of path planning and task execution through the setting of trigger mechanisms.

[0005] Optimization path scheduling systems based on mathematical modeling, which generally construct graph structure models based on urban road networks, consider multiple constraint conditions such as delivery time window, power limit, load capacity, traffic prediction, etc., and use linear programming, integer programming, genetic algorithm, etc. to solve the optimal path solution. For example, the paper "Research on Vehicle Routing Optimization for Express Delivery" (He Linlin et al., 2021) constructs a city express delivery model based on multiple constraints and uses a genetic algorithm to optimize the path solution.

[0006] However, the above-mentioned rule-based expert systems and mathematical optimization-based decision systems have inherent defects in understanding and responding to complex and variable urban environments. Rule-based systems, due to their hard-coded decision logic, exhibit rigidity and poor adaptability when faced with unforeseen or new environmental conditions (e.g., sudden weather changes, temporary flight restricted areas, dynamic traffic congestion, new UAV interference), and the cost of rule maintenance is high. On the other hand, mathematical optimization-based systems are highly dependent on explicit assumptions and parameters, and their effectiveness significantly decreases when the actual situation deviates from the model, and they are difficult to effectively handle unstructured and noisy environmental information (such as sudden event information on social media, fuzzy real-time traffic images), resulting in insufficient robustness in complex and variable real-world environments. SUMMARY

[0007] (I) Technical problems solved

[0008] In view of the deficiencies of the prior art, the present application provides an intelligent logistics management system based on a large language model, which solves the technical problem of how to improve the comprehensive grasp and adaptive decision-making ability of complex and dynamic urban environments.

[0009] (ii) Technical Solution

[0010] To achieve the above object, the present application is implemented by the following technical solutions:

[0011] An intelligent logistics management system based on a large language model, comprising a city environment management module, a logistics management scheduling module, a simulation engine module, and a large language model interface module;

[0012] The city environment management module is configured to store and maintain in real time environment situation data including all unmanned aerial vehicles, all order tasks, and historical events, and format the data as delivery data.

[0013] The logistics management scheduling module is configured to execute a current decision-making process, including:

[0014] The user instruction data and the delivery data are fused to construct structured LLM request data, a request is sent to an external LLM service through the large language model interface module, and an unstructured text response returned by the external LLM service is received, the unstructured text response is parsed into LLM decision data, the LLM decision is translated into a series of delivery instruction data, and the delivery instruction data is sent to the simulation engine module, wherein a general large language model is fine-tuned in advance using a low-rank adaptation technique to obtain the external LLM service.

[0015] The simulation engine module is configured to execute delivery instructions, simulate the physical behavior and task execution process of unmanned aerial vehicles, and trigger corresponding events, feed the events back to the city environment management module to update the environment situation data, and feed the events back to the logistics management scheduling module to trigger a new round of decision-making process.

[0016] Preferably, the intelligent logistics management system comprises a visualization and command interaction module, the visualization and command interaction module comprising a simulation monitoring window and a command decision window.

[0017] The simulation monitoring window is configured to display at least the unmanned aerial vehicle positions, order statuses, and event timelines in the environment situation data in real time.

[0018] The command decision window is configured to receive user instruction data and send the data to the logistics management scheduling module, display input prompt words and responses of the external LLM service in real time, and display warning information in real time.

[0019] Preferably, the logistics management scheduling module is configured to

[0020] When the unstructured text response is parsed into LLM decision data, a preset resource intelligent allocation logic is executed, including:

[0021] Based on the distribution data, a global resource pre-check is performed. If the total resources of available drones are insufficient, subsequent order task allocation is stopped and a global resource shortage event is recorded. If the total resources are sufficient, the global resource pre-check is passed and the optimal solution attempt stage is entered.

[0022] In the optimal solution attempt stage, all available drones are traversed to attempt to find a single drone that can independently and completely satisfy the current order task demand. If successful, the current order task is allocated to the single drone. Otherwise, the collaborative solution attempt stage is entered.

[0023] In the collaborative solution attempt stage, the current order task demand is decomposed to attempt to search for a combination of two or more available drones to collaboratively satisfy the current order task demand. If successful, the current order task is allocated to the combination of two or more available drones. Otherwise, the allocation of the next order task is entered.

[0024] After all the above allocation attempts, the finishing stage is entered, in which the successfully allocated order tasks participate in the translation of the distribution instruction data, and for all finally allocated failed order tasks, their state is updated to resource insufficient waiting for allocation, and a distribution failure event and its failure reason are recorded.

[0025] The visualization and command interaction module displays warning information of global resource shortage events and / or distribution failure events and their failure reasons.

[0026] Preferably, after receiving the LLM request data, the external LLM service identifies the decision type of the current task, including:

[0027] When the user instruction data is first input or the system is requested to schedule autonomously, the external LLM service identifies it as an initial scheduling decision; when periodic monitoring shows that the task is not completed and the drone has not arrived at the destination, the external LLM service identifies it as a continuous task allocation or endurance supply decision; when the drone is disturbed, the external LLM service identifies it as a disturbance response decision and formulates a new scheduling strategy based on the disturbance situation.

[0028] Preferably, the logistics management scheduling module is used for

[0029] After the unstructured text response is parsed into LLM decision data, corresponding system prompts and user prompts are constructed based on the decision type.

[0030] Preferably, the event includes a disturbance event, and the simulation engine module is used for

[0031] In the process of simulating the task execution of the unmanned aerial vehicle, the probability of the unmanned aerial vehicle being disturbed is calculated based on the distance between the current position of the unmanned aerial vehicle and the interference source; wherein the interference source refers to a physical or electronic signal source that produces interference to the unmanned aerial vehicle individual, and the interference probability is represented by a decay function:

[0032]

[0033] Wherein, P interference (d interf ) is the probability of the unmanned aerial vehicle being disturbed; d interf is the distance between the current position of the unmanned aerial vehicle and the interference source; P max is the maximum interference probability that the unmanned aerial vehicle can reach near the interference source; d max is the maximum action distance of the interference source, indicating that the unmanned aerial vehicle will not be disturbed beyond this distance; and β is the decay factor of the interference probability, used to adjust the speed of the interference intensity decay with distance.

[0034] When d interf is less than d max , it is considered that the unmanned aerial vehicle is within the interference range and the corresponding interference probability P interference (d interf ) is calculated, and random sampling judgment is made based on the probability value, including:

[0035] A pseudo-random number Random Number uniformly distributed in the interval [0, 1] is generated, and if the Random Number is greater than P interference (d interf ), it is determined that the unmanned aerial vehicle is not disturbed; otherwise, it is determined that the unmanned aerial vehicle is disturbed, the flight state of the unmanned aerial vehicle is set to the disturbed state, and the interference event corresponding to the interference source is triggered.

[0036] Preferably, the event includes a task completion event, a task incomplete event and a task failure event, and the simulation engine module is used to

[0037] Based on the flight distance, average speed and environmental factors of the unmanned aerial vehicle, the corrected actual delivery time is obtained, which is represented as:

[0038]

[0039] EnvironmentalFactor=C1·WindSpeed+C2·LoadRatio+C3·TerrainComplexity

[0040] Wherein, T(d flight ) represents the actual delivery time of the order task; d flightis the actual flight distance of the UAV; v is the average flight speed of the UAV; EnvironmentalFactor is a comprehensive correction factor of the environmental factors on the time consumption; C1, C2, and C3 are weight coefficients of the corresponding terms, respectively; WindSpeed is the current wind speed; LoadRatio is the ratio of the current load of the UAV to the maximum load; and TerrainComplexity is the terrain complexity of the flight path.

[0041] based on whether the order task has been disturbed and the corrected actual delivery time consumption T(d flight ) and the delivery time window of the order task, the following judgment process is performed:

[0042] If the UAV has not been disturbed and T(d flight ) is within or before the delivery time window, the order task state is updated to completed, and a task completion event containing success information is triggered.

[0043] If the UAV has not been disturbed and T(d flight ) is after the delivery time window, the order task state is updated to timeout, and a task completion event is triggered.

[0044] If the UAV has been disturbed and the task is aborted or other simulation conditions fail, the order task state is updated to failure, and a task failure event containing the failure reason is triggered.

[0045] Preferably, the delivery data is a structured snapshot of the environmental situation data, and the structured snapshot refers to a complete data set containing the environmental situation data in a predefined format.

[0046] Preferably, the logistics management scheduling module is used to run the LLM work thread in a separate thread.

[0047] Preferably, the logistics management scheduling module is used to maintain a task tracking list to manage all ongoing decision-making processes.

[0048] (Three) beneficial effects

[0049] The present application provides an intelligent logistics management system based on a large language model. Compared with the prior art, the following beneficial effects are achieved:

[0050] The application adopts a low-rank adaptation technology to fine-tune a general large language model in advance, and innovatively deeply integrates the external LLM service obtained by fine-tuning into a task decision-making process, so that it can realize deep understanding of user intent, accurate analysis of complex environmental situation, and autonomous generation of an optimal distribution scheme based thereon. In addition, the city environment management module, logistics management and scheduling module, simulation engine module and large language model interface module included in the system do not work independently, but closely cooperate through an event-driven and message passing mechanism, thereby forming a complete unmanned aerial vehicle logistics management closed loop from environment perception to decision-making, execution and event feedback. This architecture can reflect and control the unmanned aerial vehicle distribution task in real time, self-correct the established scheme, and thus significantly improve the intelligent level and operation efficiency of logistics management. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0052] Figure 1 A structural block diagram of an intelligent logistics management system based on a large language model provided by an embodiment of the present application;

[0053] Figure 2 A data flow transfer process schematic diagram of an automatic closed-loop decision-making and feedback mechanism provided by an embodiment of the present application. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application are described clearly and completely. Obviously, the described embodiments are some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0055] The embodiments of the present application provide an intelligent logistics management system based on a large language model, which solves the technical problem of how to improve the overall grasp of complex dynamic urban environment and self-adaptive decision-making capability.

[0056] The technical solutions in the embodiments of the present application are as follows to solve the above technical problems:

[0057] Embodiments of the present application aim to overcome the shortcomings of existing unmanned aerial vehicle logistics management systems in environmental adaptability (comprehensive understanding of complex dynamic urban environment and adaptive decision-making ability), human-computer interaction, decision transparency and system iterative maintenance, thereby constructing an intelligent unmanned aerial vehicle logistics management system based on a localized fine-tuned large language model (LLM). The system fine-tunes the LLM to accurately grasp the knowledge of unmanned aerial vehicle scheduling and task planning. The core purpose is to improve real-time environmental perception and task response ability, shorten the observe-orient-decide-act (OODA) cycle; realize intelligent scheduling and path planning, support natural language instructions and automatically generate standardized scheduling solutions; ensure resource optimization and risk management, and as far as possible to ensure order satisfaction rate, and intelligently select the optimal unmanned aerial vehicle combination and path through the LLM.

[0058] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in conjunction with the drawings in the specification and specific embodiments.

[0059] Embodiment 1:

[0060] Embodiments of the present application provide an intelligent logistics management system based on a large language model, as shown in Figure 1 The city environment management module is used to store and maintain in real time the environmental situation data including all unmanned aerial vehicles, all order tasks and historical events, and format them into delivery data.

[0061] The city environment management module is used to store and maintain in real time the environmental situation data including all unmanned aerial vehicles, all order tasks and historical events, and format them into delivery data.

[0062] The logistics management scheduling module is used to execute the current decision-making process, including:

[0063] The user instruction data and the delivery data are fused to construct structured LLM request data; the external LLM service is sent a request through the large language model interface module, and an unstructured text response returned by the external LLM service is received; and the unstructured text response is parsed into LLM decision data, the LLM decision is translated into a series of delivery instruction data, and the delivery instruction data is issued to the simulation engine module; wherein a general large language model is fine-tuned in advance using low-rank adaptation technology to obtain the external LLM service;

[0064] The simulation engine module is used to execute the delivery instructions, simulate the physical behavior and task execution process of the unmanned aerial vehicle, and trigger the corresponding events, feed the events back to the city environment management module to update the environmental situation data, and feed the events back to the logistics management scheduling module to trigger a new round of decision-making process.

[0065] The system closely integrates the LLM with the actual environment, realizes closed-loop feedback and dynamic adjustment, effectively deals with unexpected situations, and ultimately reduces system cost, significantly improves the intelligent level and operation efficiency of unmanned aerial vehicle logistics management.

[0066] In an optional embodiment, the intelligent logistics management system provided by the embodiment of the present application, as shown in Figure 1 The visualization and command interaction module includes a simulation monitoring window and a command decision window.

[0067] The simulation monitoring window is configured to display at least the positions of the unmanned aerial vehicles, the order states, and the event timelines in the environmental situation data in real time.

[0068] The command decision window is configured to receive user instruction data and send the user instruction data to the logistics management scheduling module, display the input prompt words and responses of the external LLM service in real time, and display warning information in real time.

[0069] It can be understood that the visualization and command interaction module enhances the transparency of command interaction and decision-making. Information transparency enables users to intuitively understand the input basis and output results of AI decision-making, thereby improving the trust of human-machine collaboration and promoting efficient logistics management.

[0070] The system architecture provided by the embodiment of the present application is precise, and forms a complete unmanned aerial vehicle logistics management closed loop (i.e., an 'automated closed-loop decision and feedback mechanism') from environmental perception to decision-making, execution, and event feedback. The core of the automated closed-loop decision and feedback mechanism is a continuous, event-driven data flow. The complete flow process of the data flow is as shown in Figure 2 The data flow can be divided into the following core stages:

[0071] First stage: environmental perception and instruction input

[0072] This stage is the starting link of the process, and the core is to collect all the context information required for decision-making. The process can be started by receiving a user instruction data by the visualization and command interaction module. The user instruction data carries the user's high-level intention. In the embodiment of the present application, the 'high-level intention' defines the business target expected to be achieved (for example, 'prioritize delivery of all VIP orders'), rather than specific, machine-executable steps (for example, 'command A unmanned aerial vehicle to fly to B coordinate point'). This target-oriented instruction requires the system to have semantic understanding and strategy planning capabilities.

[0073] This user instruction data, bearing the high-level intent, is transmitted to the logistics management scheduling module, which serves as the "control hub". To translate this intent into action, the module requests delivery data from the city environment management module, which serves as the "digital twin" data center. This delivery data is a structured snapshot of the system's real-time state at a specific point in time, i.e., environmental situation data. By "structured snapshot", we mean a complete data set organized in a predefined format (such as JSON) that contains the status of all key entities of the system at that instant (such as all drones, orders, and environmental information), ensuring the temporal consistency and data integrity of the basis for decision-making.

[0074] The city environment management module then provides this delivery data and transmits it back to the logistics management scheduling module. At the same time, the delivery data is also transmitted in parallel to the visualization and command interaction module for interface updates.

[0075] Second stage: intelligent decision-making and decision analysis, instruction issuance

[0076] After grasping the user instructions and the situation, the process enters the intelligent decision-making stage. The logistics management scheduling module fuses the two, constructing a structured LLM request data. This LLM request data is a specially formatted text prompt (Prompt) for large language models. This "text prompt" not only contains the task data to be processed, but also may contain specific instructions, rules, and examples, with the purpose of guiding the LLM to conduct specific direction reasoning and generate specific format output. This request is sent to the large language model interface module.

[0077] The large language model interface module, as a "cognitive bridge", initiates requests to external LLM services and receives responses. Then, it parses the unstructured text response returned by the LLM into a LLM decision data. This LLM decision data is a structured strategy instruction (for example, indicating the allocation relationship between specific drones and tasks) that is executable within the system. This decision is returned to the logistics management scheduling module.

[0078] Third stage: instruction execution simulation

[0079] This stage is to translate intelligent decisions into physical actions in the simulated world. After receiving the LLM decision, the logistics management scheduling module translates it into a series of delivery instruction data. This delivery instruction data is a bottom-level, machine-executable command set (for example, containing specific parameters such as takeoff, waypoints, and landing).

[0080] The delivery instruction is sent to the simulation engine module. According to the instruction, the module accurately simulates the physical behavior and task process of the UAV, and continuously generates new delivery event feedback data in the process. This delivery event feedback data is a discrete and structured message encapsulating specific events (such as "task completed", "encounter interference"), and is the key signal to realize the closed loop.

[0081] Fourth stage: closed loop feedback adjustment

[0082] This stage is the convergence and restart of the closed loop. The newly generated delivery event feedback data is sent out through two parallel paths:

[0083] One is the decision adjustment: the delivery event feedback data is sent back to the logistics management scheduling module. This data stream enables the module to trigger a new round of decision-making process (return to the second stage) according to the real-time execution results, forming a fast and adaptive decision-making closed loop.

[0084] The second is the situation update: the delivery event feedback data is sent to the city environment management module for situation update. In the embodiment of the present application, the "situation update" refers to the process of modifying or covering the state data in the "digital twin" data center according to the event feedback generated by the simulation engine, which represents the real changes of the physical world. This data stream ensures that the "digital twin" state of the system is always consistent with the latest situation, providing a reliable data basis for all subsequent decision-making, thus forming a fundamental closed loop that guarantees data integrity.

[0085] Through the above four stages of data flow, the embodiment of the present application completely realizes the whole life cycle closed loop management from environment perception to decision-making, execution and event feedback.

[0086] Next, the various modules of the system will be introduced in detail:

[0087] For the city environment management module, it is used to store and maintain the environmental situation data including all UAVs, all order tasks and historical events in real time, and format it into delivery data.

[0088] The module is configured as the core data layer in the system, responsible for maintaining and providing the real-time "digital twin" state of the entire city environment and UAV delivery task.

[0089] The environmental situation data is a highly structured data set containing all drones (including ID, payload, battery level, status), all drones in flight, all order tasks (including ID, destination, cargo information, priority, time window), and all historical events occurring during the simulation. This module ensures the freshness and accuracy of task information by simulating real-time injection of external environment data (such as weather, temporary flight restrictions, etc.) and drone state data (such as battery consumption, payload changes). It also manages detailed characteristics of drones (such as payload, endurance, speed) and aggregates available resources of all drones, formatting them into a resource status string that is easily understood by the LLM, providing the necessary resource parameters for the LLM to make decisions.

[0090] In embodiments of the present application, the "resource status string" refers to a specially designed and formatted human-readable text for large language models (LLM), which serves to aggregate and condense the complex, scattered, and dynamically changing available resource data within the system into a concise and clear text summary. This formatted text is designed to cater to the characteristics of LLMs, which rely on text-based reasoning. By precisely controlling the input information, the LLM can be guided to make more efficient and reasonable decisions, while optimizing the cost and efficiency of API calls.

[0091] The logistics management and scheduling module is used to execute the current decision-making process, including:

[0092] Fusing user instruction data with the distribution data to construct structured LLM request data, sending a request to an external LLM service through the large language model interface module, and receiving an unstructured text response returned by the external LLM service, and parsing the unstructured text response into LLM decision data, translating the LLM decision into a series of distribution instruction data, and issuing it to the simulation engine module; wherein a general large language model is fine-tuned using low-rank adaptation technology to obtain the external LLM service.

[0093] This module is configured as the "control center" in the system. It drives the entire decision-making and feedback process by coordinating and calling the city environment management module, the large language model interface module, and the simulation engine module. Its specific functions are as follows:

[0094] It drives the simulation step, receives user instruction data, schedules the large language model interface module for decision-making, analyzes and executes the scheduling instructions generated by the LLM, and feeds back the key events generated by the simulation engine module (especially the UAV interference and task completion / uncompletion / failure) to the LLM for decision adjustment in real time. Optionally, the logistics management scheduling module also runs the LLM worker in a separate thread to ensure that the asynchronous execution of LLM requests does not block the main UI thread, and to maintain the task tracking list. In the embodiment of the present application, the main UI thread refers to the main thread in the system that is responsible for updating the user interface (UI) and responding to user operations. In order to avoid the blocking of the main UI thread caused by time-consuming network requests to the external large language model (LLM) service (manifested as interface lag or non-response), such request tasks are handed over to an "LLM Worker" running in a separate background thread. The logistics management scheduling module assigns decision-making requests to the Worker, which performs network communication in the background independently, and then notifies the main thread of the results for subsequent processing. This design of separating time-consuming operations from interface responses ensures the smoothness and usability of the system. At the same time, the logistics management scheduling module is used to maintain a task tracking list to manage all ongoing decision-making processes.

[0095] The logistics management scheduling module realizes the automatic decision type triggering and management: the system provides real-time distribution data (such as order task status, UAV position, interference event) to the LLM, and the LLM intelligently summarizes and judges the decision type of the current task according to the knowledge learned in its training. That is, the external LLM service identifies the decision type of the current task after receiving the LLM request data, which specifically includes:

[0096] 1) When the user instruction data is first input or the system is requested to schedule autonomously, the external LLM service identifies it as an initial scheduling decision;

[0097] 2) When the periodic monitoring shows that the task is not completed and the UAV has not arrived at the destination, the external LLM service identifies it as a continuous task allocation or endurance supply decision;

[0098] 3) When the UAV is interfered, the external LLM service identifies it as an interference response decision, and formulates a new scheduling strategy based on the interference situation.

[0099] On this basis, the logistics management and scheduling module converts the structured environment data and user intentions within the system into natural language prompts that the LLM can understand, sends them to the external LLM service through a network request, and receives and preliminarily analyzes the decision or feedback text returned by the LLM. This module dynamically constructs system prompts and user prompts according to different decision task types (such as initial scheduling, continuous task allocation, and interference response), accurately guiding the reasoning direction and output format of the external LLM service. For example, when dealing with interference response decisions, it will inject detailed "drone interference response rules" as prompts to the external LLM service to guide the LLM to develop targeted response strategies.

[0100] Further, the logistics management and scheduling module performs a preset resource intelligent allocation logic when parsing the unstructured text response into LLM decision data to cope with complex task requirements and multi-drone collaboration, specifically including:

[0101] Based on the distribution data, perform a global resource pre-check. If the total resources of available drones are insufficient, stop subsequent order task allocation and record it as a global resource shortage event. If the total resources are sufficient, pass the global resource pre-check and enter the optimal solution attempt stage.

[0102] In the optimal solution attempt stage, iterate through all available drones to try to find a single drone that can independently and completely meet the current order task requirements. If successful, assign the current order task to this single drone. Otherwise, enter the collaborative solution attempt stage.

[0103] In the collaborative solution attempt stage, decompose the current order task requirements and try to search for a combination of two or more available drones to meet the current order task requirements. If successful, assign the current order task to the combination of two or more available drones. Otherwise, proceed to the allocation of the next order task.

[0104] After all the above allocation attempts, enter the finishing stage. Translate the successfully allocated order tasks into the distribution instruction data, and for all order tasks that fail to be allocated, update their status to "resource insufficient, waiting for allocation" and record them as allocation failure events and their failure reasons.

[0105] Display warning information for global resource shortage events and / or allocation failure events and their failure reasons through the visualization and command interaction module.

[0106] The simulation engine module is used to execute distribution instructions, simulate the physical behavior and task execution process of drones, and trigger corresponding events. The events are fed back to the city environment management module to update the environment situation data, and to the logistics management and scheduling module to trigger a new round of decision-making process.

[0107] The module is configured as a "physical engine" in the system, responsible for accurately simulating the physical behavior of the UAV and the task execution process, and triggering corresponding events. It is the evolution core of the UAV flight physics and task behavior, and acts as the "physical engine" of the entire system. It accurately simulates the UAV's takeoff, flight, battery consumption, load transportation, sudden interference (such as signal interference, communication interruption) determination, order completion determination, delivery time calculation, and dynamic updating of the UAV's battery level and load status. The UAV flight path is planned according to the start and end points, and the flight time is calculated based on the flight speed and distance.

[0108] In an optional embodiment, the event includes an interference event, a task completion event, a task failure event, and a task failure event.

[0109] (1) Interference event

[0110] Specifically, the simulation engine module is used to calculate the probability of interference of the UAV based on the distance between the current position of the UAV and the interference source during the simulation of the task execution process of the UAV; wherein the interference source refers to a physical or electronic signal source that interferes with the UAV individual (such as a high-power radio tower, a complex electromagnetic environment near a communication base station, or an illegal UAV jammer, etc.), and the interference probability is represented by a decay function:

[0111]

[0112] wherein P interference (d interf ) is the probability of interference with the UAV individual; d interf is the distance between the current position of the UAV and the interference source; P max is the maximum interference probability that the UAV can reach near the interference source; d max is the maximum action distance of the interference source, indicating that the UAV will not be interfered beyond this distance; β is the decay factor of the interference probability, used to adjust the speed of the interference intensity decay with distance;

[0113] When d interf <d max , it is considered that the UAV is within the interference range and the corresponding interference probability P interference (d interf ) is calculated, and random sampling is performed based on the probability value, including:

[0114] Generate a pseudo-random number Random Number uniformly distributed in the interval [0, 1], if Random Number is greater than P interference (d interf), it is determined that the UAV is not interfered; otherwise, it is determined that the UAV is interfered, the flight state of the UAV is set as an interfered state (such as communication interruption, signal loss), and an interference event corresponding to the interference source is triggered.

[0115] (2) Task completion event, task uncompleted event and task failure event

[0116] Specifically, the simulation engine module is configured to obtain a corrected actual delivery time consumption based on the flight distance, average speed and environmental factors of the UAV, and the corrected actual delivery time consumption is expressed as:

[0117]

[0118] EnvironmentalFactor = C1·WindSpeed + C2·LoadRatio + C3·TerrainComplexity

[0119] wherein, T(d flight ) represents the actual delivery time consumption of the order task; d flight is the actual flight distance of the UAV; v is the average flight speed of the UAV; EnvironmentalFactor is a comprehensive correction factor of environmental factors (such as weather, wind speed, load) on time consumption, which is positive when the time consumption increases, negative when the time consumption decreases, and zero when there is no correction; C1, C2, and C3 are weight coefficients of the corresponding terms respectively; WindSpeed is the current wind speed; LoadRatio is the ratio of the current load of the UAV to the maximum load; and TerrainComplexity is the terrain complexity of the flight path (such as mountainous area, urban building dense area, etc.).

[0120] It should be noted that the environmental factors here (different from the above interference sources, which are constraints that affect the entire operating environment and are usually predictable, mainly used for correcting planning and cost calculation, rather than triggering sudden abnormalities.

[0121] Based on whether the order task has been interfered, and the corrected actual delivery time consumption T(d flight ) and the delivery time window of the order task, the following judgment process is performed:

[0122] If the UAV has not been interfered and T(d flight ) is within or before the delivery time window, the order task state is updated to completed, and a task completion event containing success information is triggered;

[0123] If the UAV has not been interfered and T(d flight) if the delivery time window has passed, then the order task status is updated to timeout, and a task not completed event is triggered;

[0124] If the UAV has been interfered and the task is aborted or other simulation conditions fail, the order task status is updated to failure, and a task failure event containing the failure reason is triggered.

[0125] For the large language model interface module, it serves as a "cognitive bridge" between the system and external large language model services, responsible for semantic-level interaction. It is called "cognitive bridge" because it connects the structured data world inside the system and the cognitive reasoning world of external LLM.

[0126] For the visualization and command interaction module, it is configured as a graphical user interface (GUI) in the system for human-computer interaction. It mainly includes a simulation monitoring window and a command decision window:

[0127] The simulation monitoring window is used to at least display the UAV position, order status and event timeline in the environment situation data in real time;

[0128] The command decision window is used to receive user instruction data and send it to the logistics management and scheduling module, display the input prompt words and responses of the external LLM service in real time, and display warning information in real time.

[0129] In addition, it is necessary to emphasize that the core provided by the embodiment of the application is the deep application of LLM, and its successful implementation depends on the following key technical mechanisms:

[0130] 1) Large language model fine-tuning

[0131] The general large language model (LLM) is fine-tuned in advance using the low-rank adaptation (LoRA) technology, so that it can understand complex logistics scenario data, perform scheduling reasoning, and generate accurate task instructions. This is the core cornerstone of the intelligent system. Specifically, LoRA adds two smaller, trainable low-rank matrices and to bypass the large weight matrix W0 of the pre-trained model (such as the Q, K, V projection matrix in the attention mechanism) to realize incremental update, where rank r << min(d, k), i.e. W = W0 + BA, only A and B are trained during fine-tuning. This method significantly reduces the number of trainable parameters from d x k of the original model to r x (d + k), while maintaining the integrity of the original knowledge of the model, significantly reducing training costs, improving training efficiency, and reducing memory usage during deployment.

[0132] 2) UAV resource fine-grained intelligent allocation and visual management

[0133] In terms of unit-level resource management and global aggregation, each UAV maintains its independent load and battery level information. The city environment management module is responsible for aggregating the available resources of all delivery UAVs and formatting them into a "resource state" string that the LLM can understand, serving as an important input for LLM decision-making, ensuring that LLM does not exceed the actual resource limit when planning tasks. In terms of intelligent allocation logic, the logistics management scheduling module implements intelligent allocation logic when analyzing LLM scheduling decisions to address complex task requirements and multi-UAV collaboration. This intelligent allocation logic is a progressive process that clearly defines and handles different levels of "resource shortage" situations. At the same time, the resource state and allocation warnings are displayed in real-time through a visual interface, ensuring the optimization and transparency of UAV resource utilization efficiency.

[0134] In summary, compared with the prior art, the following beneficial effects are achieved:

[0135] 1. Environmental understanding and adaptive ability are greatly improved: The embodiments of the present application break through the dependence of traditional rule systems on preset scenarios and the limitations of mathematical models on precise modeling. Through the LLM fine-tuned by LoRA, the system can learn complex nonlinear environmental patterns and deep scheduling experience from large-scale, high-fidelity simulation data, thereby understanding and effectively responding to unanticipated, dynamically changing urban environments (such as sudden weather, temporary flight restricted areas, and dynamic obstacles). When faced with complex unknown scenarios, the system can automatically generate more reasonable and efficient response strategies than traditional systems, significantly improving the flexibility and generalization of decision-making, which effectively reduces the cost and time of manually maintaining and updating scheduling rules, and improves the flexibility and efficiency of UAV operations.

[0136] 2. Human-computer interaction efficiency and naturalness are significantly improved: The embodiments of the present application realize intuitive and natural natural language interaction between users and the system. Users do not need to learn complex command syntax or code, and can clearly express task intentions directly using daily language instructions. The system can also provide feedback on decision-making logic and execution in an easily understandable natural language format. The operation convenience is significantly improved, which effectively reduces the user's cognitive burden and operation complexity, allowing them to focus more on macro scheduling strategies and risk analysis rather than specific scheduling instruction details.

[0137] 3. Accurate order satisfaction rate and optimal operation cost: The system provided by the embodiment of the present application can accurately select and combine the optimal unmanned aerial vehicle type, quantity and path based on the intelligent reasoning of the LLM and the accurate environment and unmanned aerial vehicle data to achieve the goal of the lowest total operation cost while maintaining the core service requirement of keeping the specified order satisfaction rate at a high level, effectively avoiding the resource waste caused by traditional manual scheduling or simple rules. Under the same order quantity and resources, the system can more efficiently complete the delivery task, or optimize more order delivery without additional resource consumption, significantly improving resource utilization efficiency, which significantly improves the overall benefit and market competitiveness of logistics operation.

[0138] 4. Robust operation and significantly enhanced task continuity: The embodiment of the present application realizes a real-time, closed-loop environment event feedback mechanism, especially for sudden situations such as communication interruption, signal shielding and the like during the journey of the unmanned aerial vehicle. The system can instantly perceive the interference event, and the LLM quickly generates a supplementary scheduling strategy according to the "unmanned aerial vehicle interference response rule". Under the simulated interference environment, the task completion rate is effectively guaranteed. Compared with the decision interruption or delay that may occur in the traditional system under interference, the present application can quickly adjust the scheduling to ensure the continuity and success rate of the delivery task, thereby enhancing the operation adaptability of the system in complex urban environments.

[0139] 5. System iteration and maintenance cost is greatly reduced: The embodiment of the present application uses the LoRA fine-tuning technology, so that the model training parameter quantity is extremely small and the training efficiency is extremely high. When there are new service demands, unmanned aerial vehicle model upgrades or environmental changes, only the LLM needs to be quickly fine-tuned through incremental data, without the need to retrain the entire large model or make substantial modifications to the hard-coded rules. This greatly shortens the model update iteration cycle and significantly reduces system maintenance costs, so that the system can quickly adapt to the rapidly changing logistics industry demands and technological development, maintaining its advanced nature and operational competitiveness.

[0140] It should be noted that, in this text, relational terms such as first and second are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment. Without more limitations, the element defined by the statement "includes a" does not exclude the presence of another identical element in the process, method, article or equipment including the element.

[0141] The above examples are only used to illustrate the technical solutions of the present application, but not to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those ordinarily skilled in the art should understand: the technical solutions recorded in the foregoing examples can still be modified, or some technical features can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An intelligent logistics management system based on a large language model, characterized in that: It includes urban environment management module, logistics management and scheduling module, simulation engine module and large language model interface module; The urban environment management module is responsible for storing and maintaining in real time the environmental situation data of all drones, all order tasks and historical events, and formatting it into delivery data; The logistics management and scheduling module is used to execute the current decision-making process, including: The method further comprises fusing user instruction data with the delivery data to construct structured LLM request data; sending a request to an external LLM service through the large language model interface module and receiving an unstructured text response returned by the external LLM service; parsing the unstructured text response into LLM decision data, translating the LLM decision into a series of delivery instruction data, and sending the data to the simulation engine module; wherein the general large language model is pre-adapted using low-rank adaptation technology to obtain the external LLM service; The simulation engine module is used to execute delivery instructions, simulate the physical behavior and task execution process of the drone, and trigger corresponding events, feed the events back to the urban environment management module to update the environmental situation data, and feed the events back to the logistics management scheduling module to trigger a new round of decision-making process.

2. The intelligent logistics management system according to claim 1, characterized in that: It includes a visualization and command interaction module, which includes a simulation monitoring window and a command decision window; The simulation monitoring window is used to display at least the drone position, order status and event timeline in the environmental situation data in real time; The command decision window is used to receive user instruction data and send it to the logistics management scheduling module, display the input prompt words and responses of the external LLM service in real time, and display warning information in real time.

3. The intelligent logistics management system according to claim 2, characterized in that: The logistics management scheduling module is used to When parsing the unstructured text response into LLM decision data, the preset resource intelligent allocation logic is executed, including: Perform a global resource pre-check based on the delivery data. If the total resources of available drones are insufficient, subsequent order task allocation is stopped and recorded as a global resource shortage event. If the total resources are sufficient, the global resource pre-check is passed and the optimal solution attempt phase is entered. In the optimal solution attempt phase, all available drones are traversed to try to find a single drone that can independently and fully meet the requirements of the current order task. If successful, the current order task is assigned to the single drone, otherwise it enters the collaborative solution attempt phase; In the collaborative solution attempt phase, the current order task requirements are decomposed, and an attempt is made to search for a combination of two or more available drones to collaboratively meet the current order task requirements. If successful, the current order task is assigned to the combination of two or more available drones, otherwise the next order task is assigned. After all the above allocation attempts, the closing phase begins, in which the successfully allocated order tasks are included in the translation of the delivery instruction data. For all order tasks that ultimately fail to be allocated, their status is updated to insufficient resources waiting for allocation, and an allocation failure event and the reason for the failure are recorded. Warning information of global resource shortage events and / or allocation failure events and their failure causes is displayed through the visualization and command interaction module.

4. The intelligent logistics management system according to claim 1, characterized in that: After receiving the LLM request data, the external LLM service identifies the decision type of the current task, including: When user instruction data is input for the first time or the system is requested to perform autonomous scheduling, the external LLM service identifies it as an initial scheduling decision; when periodic monitoring shows that the task is not completed and the drone has not arrived at the destination, the external LLM service identifies it as a continuous task allocation or endurance supply decision; when the drone is interfered with, the external LLM service identifies it as an interference response decision and formulates a new scheduling strategy based on the interference situation.

5. The intelligent logistics management system according to claim 4, characterized in that: The logistics management scheduling module is used to After parsing the unstructured text response into LLM decision data, corresponding system prompts and user prompts are constructed based on the decision type.

6. The intelligent logistics management system according to claim 1, characterized in that: The event includes an interference event, and the simulation engine module is used to During the simulated drone mission execution process, the probability of the drone being interfered with is calculated based on the distance between the drone's current position and the interference source. The interference source refers to the physical or electronic signal source that interferes with the individual drone. The interference probability is expressed using an attenuation function: Among them, P interference (d interf ) is the probability of interference to individual drones; d interf is the distance between the current position of the UAV and the interference source; P max is the maximum interference probability that the UAV can achieve near the interference source; d max is the maximum effective distance of the interference source, indicating that the UAV will not be interfered with beyond this distance; β is the attenuation factor of the interference probability, which is used to adjust the speed at which the interference intensity decays with distance; When d interf <d max When the UAV is considered to be within the interference range, the corresponding interference probability P is calculated. interference (d interf ), and make random sampling decisions based on the probability value, including: Generate a pseudo-random number Random Number uniformly distributed in the interval [0,1]. If Random Number is greater than P interference (d interf ), the drone is determined to be not interfered with; otherwise, the drone is determined to be interfered with, the flight state of the drone is set to the interfered state, and an interference event corresponding to the interference source is triggered.

7. The intelligent logistics management system according to claim 6, characterized in that: The events include task completion events, task incomplete events and task failure events. The simulation engine module is used to Based on the flight distance, average speed, and environmental factors of the drone, the actual delivery time after correction is obtained, which is expressed as: EnvironmentalFactor=C1·WindSpeed+C2·LoadRatio+C3·TerrainComplexity Among them, T(d flight ) represents the actual delivery time of the order task; d flight is the actual flight distance of the UAV; v is the average flight speed of the UAV; EnvironmentalFactor is the comprehensive correction factor of environmental factors on flight time; C1, C2, C3 are the weight coefficients of the corresponding items, WindSpeed ​​is the current wind speed, LoadRatio is the ratio of the current load of the UAV to the maximum load, and TerrainComplexity is the terrain complexity of the flight path; Based on whether the order task has been interrupted and the corrected actual delivery time T(d flight ) and the delivery time window of the order task, perform the following judgment process: If the drone has not been interfered with and T(d flight ) If the order is within or before the delivery time window, the order task status is updated to completed and a task completion event containing success information is triggered; If the drone has not been interfered with and T(d flight ) After the delivery time window, the order task status is updated to timeout and a task uncompleted event is triggered; If the drone has been interfered with and the mission has been aborted or other simulation conditions have caused it to fail, the order mission status will be updated to failure, and a mission failure event containing the failure reason will be triggered.

8. The intelligent logistics management system according to any one of claims 1 to 7, characterized in that: The distribution data is a structured snapshot of the environmental situation data, and the structured snapshot refers to a complete data set containing the environmental situation data in a predefined format.

9. The intelligent logistics management system according to any one of claims 1 to 7, characterized in that: The logistics management scheduling module is used to run the LLM working thread in an independent thread.

10. The intelligent logistics management system according to any one of claims 1 to 7, characterized in that: The logistics management scheduling module is used to maintain a task tracking list to manage all ongoing decision-making processes.

Images