A Method and Apparatus for Optimizing Drone Rescue Dispatch Based on Semantic Translation Modeling

Abstract

This invention discloses a method and apparatus for optimizing UAV rescue scheduling based on semantic translation modeling, relating to the field of intelligent scheduling technology for UAV emergency rescue. The method includes: acquiring emergency rescue task description information; converting natural language rules in the task description information into standardized constraint expressions, and identifying response time targets, coverage targets, resource parameters, and network structure parameters; constructing an initial scheduling model, solving the initial scheduling model using an improved multi-objective optimization algorithm to obtain an initial UAV scheduling scheme; when interference events are detected during the execution of the initial scheduling scheme, constructing a dynamic re-optimization model based on the initial scheduling scheme as the disturbance benchmark, solving the dynamic re-optimization model with the scheme disturbance cost and arrival delay time as optimization objectives, and outputting a repair scheduling scheme and interpretable decision results, significantly improving the accuracy, real-time performance, and decision credibility of emergency rescue scheduling.

Description

Technical Field

[0001] This invention relates to the field of intelligent scheduling technology for UAV emergency rescue, and in particular to a method and apparatus for optimizing UAV rescue scheduling based on semantic translation modeling. Background Technology

[0002] In recent years, natural disasters and emergencies have occurred frequently. As an important piece of equipment for emergency rescue, the rationality of the dispatch plan for drones directly affects the efficiency of rescue response and resource utilization.

[0003] Existing drone dispatching technologies have significant drawbacks: First, they are difficult to directly adapt to task descriptions in natural language, requiring manual conversion of natural language rules into algorithm-recognizable constraints, which is inefficient and prone to logical conflicts due to human error. Second, they lack the ability to handle dynamic interference; most solutions can only generate an initial dispatching plan. When faced with sudden interference such as drone malfunctions, environmental changes, or new demand points, they cannot quickly build a dynamic re-optimization model based on the initial plan after the interference occurs, making it difficult to achieve a controllable trade-off between plan stability and rescue timeliness. Third, the decision results lack interpretability; the dispatching plans output by traditional optimization algorithms are only specific execution instructions and cannot clearly present the decision-making logic, which is not conducive to rescue personnel's rapid understanding and execution.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a method and apparatus for optimizing drone rescue scheduling based on semantic translation modeling. This aims to solve the technical problems of existing technologies being unable to adapt to natural language task descriptions, cope with dynamic interference, and provide interpretable decisions, resulting in insufficient efficiency, accuracy, and reliability of drone emergency rescue scheduling.

[0006] To achieve the above objectives, this invention provides a drone rescue scheduling optimization method based on semantic translation modeling, which includes the following steps:

[0007] Obtain emergency rescue mission description information;

[0008] The task description information is semantically parsed using a large language model, and the natural language rules in the task description information are converted into standardized constraint expressions, and response time targets, coverage targets, resource parameters and network structure parameters are identified.

[0009] An initial scheduling model is constructed based on the response time target, the coverage target, the standardized constraint expression, the resource parameters, and the network structure parameters.

[0010] The initial scheduling model is solved by an improved multi-objective optimization algorithm to obtain the initial scheduling scheme for the UAV.

[0011] When an interference event is detected during the execution of the initial scheduling scheme, a dynamic re-optimization model is constructed using the initial scheduling scheme as the perturbation benchmark.

[0012] The dynamic re-optimization model is solved with the scheme disturbance cost and arrival delay time as optimization objectives to obtain the repair scheduling scheme;

[0013] Output the repair scheduling scheme and the interpretable decision results generated based on semantic translation modeling.

[0014] Furthermore, to achieve the above objectives, this invention also proposes a drone rescue scheduling optimization device based on semantic translation modeling, the device comprising:

[0015] The task information acquisition module is used to acquire emergency rescue task description information;

[0016] The semantic parsing processing module is used to perform semantic parsing on the task description information through a large language model, convert the natural language rules in the task description information into standardized constraint expressions, and identify response time targets, coverage targets, resource parameters, and network structure parameters.

[0017] The initial scheduling solution module is used to construct an initial scheduling model based on the response time objective, the coverage objective, the standardized constraint expression, the resource parameters, and the network structure parameters; and to solve the initial scheduling model using an improved multi-objective optimization algorithm to obtain the initial scheduling scheme for the UAV.

[0018] The dynamic re-optimization module is used to construct a dynamic re-optimization model based on the initial scheduling scheme when an interference event is detected during the execution of the initial scheduling scheme; solve the dynamic re-optimization model with the scheme interference cost and arrival delay time as optimization objectives to obtain a repaired scheduling scheme; and output the repaired scheduling scheme and interpretable decision results generated based on semantic translation modeling.

[0019] One or more technical solutions proposed in this application have at least the following technical effects:

[0020] By employing semantic parsing and semantic translation modeling using a large language model, the direct conversion of emergency rescue task descriptions in natural language form into standardized constraint expressions is achieved without manual intervention. This avoids logical conflicts caused by human error and significantly improves the efficiency of adapting task requirements to the scheduling model. An initial scheduling model is constructed based on an improved multi-objective optimization algorithm. Combined with a dynamic re-optimization mechanism, the initial scheduling scheme is used as a perturbation benchmark to cope with sudden disturbances. This allows for the rapid generation of repair scheduling schemes while balancing the cost of scheme perturbation with the timeliness of rescue arrival, effectively addressing the problem of insufficient dynamic adaptability in existing technologies. Simultaneously, interpretable decision results are generated through semantic translation modeling, coupled with output content such as early warning information and compliance prompts, making the scheduling decision logic clear and traceable. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating an embodiment of the UAV rescue scheduling optimization method based on semantic translation modeling in this application.

[0024] Figure 2 This is a schematic diagram of the form semantic modeling workflow provided in Embodiment 1 of the UAV rescue dispatch optimization method based on semantic translation modeling in this application;

[0025] Figure 3 This is a schematic diagram of the technical route for form-assisted optimization modeling provided in Embodiment 1 of the UAV rescue dispatch optimization method based on semantic translation modeling in this application;

[0026] Figure 4 This is a flowchart illustrating the working principle of the form-assisted model solving method provided in Embodiment 1 of the UAV rescue scheduling optimization method based on semantic translation modeling in this application.

[0027] Figure 5 This is a schematic diagram of the interference management mechanism provided in Embodiment 1 of the UAV rescue dispatch optimization method based on semantic translation modeling in this application;

[0028] Figure 6 This is a functional diagram of the result interpretation visualization module provided in Embodiment 1 of the UAV rescue dispatch optimization method based on semantic translation modeling in this application;

[0029] Figure 7 This is a flowchart illustrating Embodiment 2 of the UAV rescue scheduling optimization method based on semantic translation modeling in this application;

[0030] Figure 8 The flowchart of the improved NSGA-II algorithm provided in Embodiment 2 of the UAV rescue scheduling optimization method based on semantic translation modeling in this application is shown below.

[0031] Figure 9 This is a flowchart illustrating Embodiment 3 of the UAV rescue scheduling optimization method based on semantic translation modeling in this application.

[0032] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0033] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0034] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0035] Based on this, embodiments of this application provide a method for optimizing drone rescue scheduling based on semantic translation modeling, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the drone rescue scheduling optimization method based on semantic translation modeling. This embodiment is applied to the initial scheduling model and the dynamic re-optimization model under interference scenarios.

[0036] In this embodiment, the UAV rescue scheduling optimization method based on semantic translation modeling includes steps S10~S70:

[0037] Step S10: Obtain and integrate emergency rescue mission description information, which includes structured mission information and natural language mission supplementary information;

[0038] This invention proposes an optimization method for drone emergency rescue scheduling based on form-based semantic modeling. The specific form workflow is as follows: Figure 2As shown, the constructed form enables human-computer interaction and prompt engineering. Guided prompt engineering allows users to create role cards, select semantic guidance templates, and adjust preset fields, significantly improving the efficiency of natural language input and shortening the number of human-computer dialogue rounds. This supports natural language parsing and recognition alignment functions driven by a large model. Then, prior knowledge is used to learn rules and achieve semantic compilation under domain knowledge enhancement. The form-based semantic modeling function assists in combining and optimizing the problem generation model and solution algorithm, then calls external tools for solving, and finally returns the solution results. The form enables the interpretation and visualization of the solution results. Prior knowledge input can be in the form of tables, images, or text. After semantic parsing by the large language model, variables and constraints are extracted, and standardized constraint expressions are generated. Next, an initial scheduling optimization model is constructed, and the optimal initial solution is obtained through an improved NSGA-II algorithm. Finally, through the verification and output module, tools such as Python / Matlab and Gurobi / CPLEX are called for model generation and verification, algorithm recommendation, and interpretation and visualization, including modifying graphs / Gantt charts and semantic-model-solution causal chains.

[0039] The core idea is to transform the natural language task description of emergency experts into an initial scheduling model through form-based input templates and semantic parsing mechanisms; then, to use a multi-layer optimization framework to achieve integrated optimization of initial task scheduling and dynamic re-optimization under interference scenarios after task start.

[0040] The drone emergency rescue dispatch optimization framework consists of two layers:

[0041] The first layer is the initial scheduling model. Upon receiving an emergency task, the system automatically generates an UAV scheduling optimization model through structured form input (modules for objectives, constraints, resources, and regions). The model is generated by a semantic modeling system and the optimal initial solution is obtained by solving the problem using an improved NSGA-II algorithm.

[0042] The second layer is a dynamic re-optimization model for interference scenarios. When interference occurs during task execution (such as drone malfunction, new requirements, or severe weather), the system automatically detects the event and triggers the construction of the dynamic re-optimization model. While maintaining the stability of the output solution from the first layer, the dynamic re-optimization model achieves optimal scheduling for the remaining tasks and outputs a repair scheduling solution. The entire process is driven by a form-based semantic engine, allowing form users to update constraints and priorities using natural language descriptions, achieving low-barrier, highly adaptive scheduling optimization.

[0043] like Figure 3As shown, the system of this invention is divided into five core modules. The process begins with the form input module, where task objectives, constraints, and resources are input. Next, the semantic parsing and parameter generation module identifies entities and generates variables and constraints. Then, an initial scheduling optimization model is constructed, aiming to minimize the shortest response time and maximize coverage. Afterwards, the disturbance management and dynamic re-optimization module is triggered upon detecting disturbance events, performing minimum disturbance re-optimization. Finally, the results visualization and interaction module generates Gantt charts, path graphs, and decision trees to display the optimization results.

[0044] The form input module provides task templates, including target settings (shortest response time, maximum coverage), constraints (range, battery power, payload, time window), and resource lists (number of drones, base location).

[0045] The semantic parsing module uses a large language model to parse natural language descriptions, identify entities (drones, bases, mission points), relationships (priority, constraint types), and parameters, and automatically generate corresponding mathematical variables and constraint formulas.

[0046] The initial scheduling optimization module (first layer) constructs a multi-objective mixed integer programming model: the objective functions include "shortest overall response time" and "maximum rescue coverage under limited equipment conditions".

[0047] Constraints include base opening and closing, drone take-off and arrival consistency, flow conservation, payload limits, energy consumption constraints, and service coverage thresholds.

[0048] The improved NSGA-II algorithm (constructive population initialization, local refinement, and bounded archive maintenance) is used to solve the problem.

[0049] The interference event management and dynamic re-optimization module (second layer) monitors the UAV status, mission progress, and environmental data in real time during execution. When an interference event is detected (e.g., UAV malfunction, new mission point, path blockage), it automatically triggers local re-optimization. The new model minimizes the disturbance cost while maintaining the stability of the original scheduling scheme.

[0050] The results visualization and interaction module displays the evolution of the solution using Gantt charts, path topology diagrams, and task coverage diagrams; users can readjust the priority through a form (e.g., "Prioritize Area A"), and the system will recalculate and output a differentiated report in real time.

[0051] This solution can be applied in the following scenarios: earthquake or flood relief, where multiple drones can be rapidly deployed to different disaster sites, and the system automatically generates the optimal route. If a drone loses contact or a new rescue area is added, the system automatically reconfigures the route. In mountainous communication restoration scenarios, if a drone's operation is interrupted due to insufficient power or signal loss, the system automatically allocates a backup drone. In cross-regional collaborative rescue scenarios, multiple bases and different types of drones can work together to execute tasks, and resource allocation ratios can be adjusted in real time via form input.

[0052] In one feasible implementation, step S10 includes steps A11 to A14:

[0053] Step A11: Collect structured task information through the emergency task input form. The structured task information includes the rescue area, the number of drones, and the location of the rescue base.

[0054] It should be noted that the emergency task input form refers to an electronic data entry interface with fixed preset field formats. Its purpose is to standardize user input behavior and ensure the uniformity of data format. Structured task information refers to data entered in a uniform format according to preset field specifications, which can be directly recognized and processed without additional conversion. The rescue area refers to the geographical scope where emergency rescue needs to be carried out, i.e., the specific administrative region or the area defined by latitude and longitude. The number of drones refers to the total number of drones that can be deployed in this rescue mission, which is the core basic data for resource allocation. The location of the rescue base refers to the geographical coordinates of the fixed location used for parking drones, storing rescue supplies, and dispatching and commanding operations; it is the core node for drone take-off, landing, and resupply.

[0055] Understandably, by designing an emergency task input form that includes core fields such as the number of drones in the rescue area and the location of the rescue base, users are guided to enter information according to the field requirements, thereby directly collecting structured task information with a uniform format. The form input instructions yielded the following problem description:

[0056] like Figure 4 As shown, for the optimization problem of drone emergency rescue dispatch, the following problem description was obtained through form input instructions: Consider a typical multi-base, multi-drone emergency rescue scenario. The post-disaster road-air hybrid access network can be abstracted as a weighted directed graph. The node set consists of two types: a set of rescue bases (which can be located / activated). With demand point set ,then Edge set , represents a directed connection between any two points that are traversable; each arc Flight distance With flight time (can be (Converted to a uniform cruising speed) and energy consumption coefficient .

[0057] Rescue system equipment integration This indicates the number of drones that can be deployed. Each drone With maximum load Battery capacity limit / maximum power consumption With maximum range Resource constraints; its departure and recovery must be carried out from a certain operational base. Completed. All bases. Available material supply With available takeoff and landing slots / maximum number of drones available simultaneously (optional constraint). Various requirements. Demand Compared to minimum service ratio (satisfy If the task only requires "coverage upon arrival", then it can be set as follows: .

[0058] Step A12: Receive supplementary natural language task information input by the user. This information includes a description of rescue priorities and a description of special scenario constraints. It should be noted that this supplementary natural language information refers to task-related information input by the user in everyday conversational language; it does not need to adhere to a fixed field format and can be freely expressed. The rescue priority description indicates the user's explicit ranking of the urgency of different rescue needs and is the core basis for dispatch priority allocation. The description of special scenario constraints refers to unconventional constraints mentioned by the user, i.e., special restrictions or requirements that exceed basic dispatch requirements.

[0059] Step A13: Perform integrity checks on the structured task information and natural language task supplementary information, remove information that does not conform to the preset format, and obtain the structured task information and natural language task supplementary information that have passed the checks. It should be noted that integrity checks refer to the process of checking whether the structured task information has filled all the required fields and whether the natural language task supplementary information contains valid semantics.

[0060] Step A14: Integrate the validated structured task information and the validated natural language task supplementary information to obtain complete emergency rescue task description information. It should be noted that complete emergency rescue task description information refers to the full amount of task data, including structured basic configurations and unstructured personalized requirements, and is the direct processing object for subsequent semantic parsing.

[0061] Step S20: Semantic parsing of the task description information is performed using a large language model, converting the natural language rules in the task description information into standardized constraint expressions, and identifying response time targets, coverage targets, resource parameters, and network structure parameters.

[0062] Understandably, calling a large language model to perform semantic analysis on the complete description information of the emergency rescue mission first identifies key entities and relationships, extracts natural language rules and transforms them into standardized constraint expressions, and then extracts response time, target coverage, target resource parameters and network structure parameters from them.

[0063] In one feasible implementation, step S20 includes steps A21 to A26:

[0064] Step A21: Identify key entities in the emergency rescue mission description information using a large language model. These key entities include drones, rescue demand points, rescue bases, and mission execution periods. The purpose of this step is to extract core participants and time boundaries from the mission description, providing foundational entity support for subsequent relational extraction and rule refinement. This involves preprocessing and optimizing the mission description information, adapting it to a large language model using an entity dictionary in the emergency rescue domain, and then performing sentence-by-sentence type classification and consistency verification to accurately extract key entities such as drone rescue demand points, rescue bases, and mission execution periods.

[0065] Further, this step includes: preprocessing the emergency rescue mission description information to remove redundant characters and invalid expressions, resulting in preprocessed emergency rescue mission description information; loading an emergency rescue domain entity dictionary and inputting it into a large language model for adaptation and optimization, resulting in a domain-adapted model for entity recognition; calling the entity recognition optimization model adapted to the emergency rescue scenario to perform sentence-by-sentence entity recognition on the preprocessed emergency rescue mission description information and marking candidate entities; classifying candidate entities by type to distinguish between drone-type, rescue target-type, resource-type, and time-type entities; and performing consistency verification on the classified candidate entities, eliminating duplicate recognized entities to obtain the final key entities.

[0066] It should be noted that the emergency rescue mission description information is a collection of information related to resource allocation and constraints for rescue dispatch. Text preprocessing is a preliminary process to standardize the original text to improve the accuracy of entity recognition. The preprocessed emergency rescue mission description information is the semantically clear core text after removing redundant and invalid content. The emergency rescue domain entity dictionary is a vocabulary collection that includes professional terms in fields such as drones, rescue needs, and no-fly zones. Adaptation and optimization is the process of integrating this dictionary into a large language model and adjusting parameters to accurately identify professional entities. Entity recognition, in essence, involves loading the emergency rescue domain entity dictionary to adapt and optimize the large language model, enabling... It has the ability to identify professional terms such as drones, rescue bases, and demand points; the domain-adapted model used for entity recognition is a model that can accurately identify relevant entities after domain adaptation; sentence-by-sentence entity recognition is the process of detecting and identifying entities in text according to sentence order; candidate entities are domain key entity words or phrases initially determined by the model; type classification is the process of classifying candidate entities into drones, rescue targets, resources, and time categories according to semantic attributes and functions; consistency verification is the process of checking whether candidate entities are duplicated or semantically conflicting; and the final key entities are the core entity set that has been verified to be without duplication or conflict and can support subsequent processes.

[0067] Step A22: Extract the relationships between key entities. These relationships include the hierarchical relationship between drones and rescue bases, and the correspondence between rescue needs and priorities. It should be noted that these relationships refer to the logical connections or correspondences between different key entities, forming the core of the rescue mission's logical framework. The hierarchical relationship between drones and rescue bases refers to the correspondence where a drone belongs to a specific rescue base; that is, a particular drone needs to take off, land, and resupply from a designated rescue base. The correspondence between rescue needs and priorities refers to the relationship where each rescue need corresponds to a clear ranking of urgency, serving as the core basis for scheduling priority allocation.

[0068] Step A23: By performing entity recognition and logical relation extraction on the emergency rescue mission description information, natural language rules are extracted. These rules are then converted into standardized constraint expressions through operator normalization. It should be noted that logical relation extraction refers to the process of extracting logical constraints between entities from the mission description, which is the foundation for refining natural language rules. Natural language rules refer to the constraints on rescue dispatch expressed in everyday language, representing specific limitations imposed by users on mission execution. Operator normalization refers to the process of converting the constraint logic in the natural language rules into standardized operator expressions, ensuring the consistency and readability of the constraint expressions. Operators refer to symbols or combinations of symbols representing operations or logical relationships. Standardized constraint expressions refer to standardized constraint expressions that conform to mathematical logic or computer syntax, which can be directly recognized and parsed by the dispatch model and are the core constraint basis for model construction.

[0069] Step A24: Identify the response time objective and coverage objective from the emergency rescue mission description information. The response time objective is to minimize the total rescue response time, and the coverage objective is to maximize the coverage ratio of rescue request points. It should be noted that the response time objective refers to the core optimization requirement for time efficiency in the rescue mission and is a key indicator for measuring the timeliness of the dispatch plan. Minimizing the total rescue response time means minimizing the total time from issuing a request to receiving rescue for all rescue request points, reflecting the speed of rescue. The coverage objective refers to the core optimization requirement for the breadth of coverage in the rescue mission and is a key indicator for measuring the comprehensiveness of the dispatch plan.

[0070] Step A25: Extract the resource parameters and network structure parameters corresponding to key entities. Resource parameters include the drone's maximum battery capacity and maximum payload. Network structure parameters include the location coordinates and connectivity between the rescue base and the demand point. It should be noted that resource parameters refer to quantitative indicators related to rescue resources, forming the resource constraint basis for model construction and directly affecting the feasibility of the scheduling scheme. The maximum drone battery capacity refers to the maximum battery power a drone can consume in a single mission, determining its maximum range. The maximum payload capacity refers to the maximum weight or volume of supplies a drone can carry in a single mission, determining its supply transportation capacity. Network structure parameters are quantitative indicators describing the geographical layout and connectivity of the rescue scenario, providing a geographical basis for route planning. The location coordinates of the rescue base and the demand point refer to the specific geographic location data of the rescue base and the demand point, serving as the core basis for calculating distances and paths. Connectivity refers to whether there is a passable path between the rescue base and the demand point, determining the drone's flight route selection.

[0071] Step A26: Combine prior knowledge in the field of emergency rescue to perform logical verification on standardized constraint expressions, resource parameters and network structure parameters, correct logical conflicts, and complete semantic parsing, standardized constraint expression transformation and target and parameter identification.

[0072] The purpose of this step is to verify and correct logical conflicts in the parameters and constraint expressions, ensuring the consistency and validity of all input data and providing reliable data support for subsequent model construction. It should be noted that prior knowledge in the emergency rescue field refers to professional common sense and practical experience accumulated in emergency rescue scenarios, which serves as the basis for judging the logical rationality.

[0073] Step S30: Construct an initial scheduling model based on response time target, coverage target, standardized constraint expression, resource parameters and network structure parameters;

[0074] The purpose of this step is to integrate all core input data, construct a solvable scheduling model, and obtain an initial execution plan through an algorithm, so as to provide a basic scheduling basis for the rescue mission.

[0075] Step S40: Solve the initial scheduling model using an improved multi-objective optimization algorithm to obtain the initial scheduling scheme for the UAV;

[0076] Understandably, with response time and coverage objectives as the optimization directions, an initial scheduling model is constructed by combining standardized constraint expressions, resource parameters, and network structure parameters. An improved multi-objective optimization algorithm is then called to solve the model, ultimately yielding the initial scheduling scheme for the UAV.

[0077] Step S50: When an interference event is detected during the execution of the initial scheduling scheme, a dynamic re-optimization model is constructed using the initial scheduling scheme as the disturbance reference.

[0078] It should be noted that disturbance events refer to unexpected situations that occur during the execution of the initial scheduling plan, causing the original plan to become infeasible or the optimization effect to decrease. The disturbance baseline refers to ensuring that the difference between the repaired scheduling plan and the original plan is minimized, based on the initial scheduling plan. The dynamic re-optimization model refers to the mathematical model constructed for generating repaired scheduling plans in response to disturbance events, which can quickly adapt to unexpected situations.

[0079] Step S60: Solve the dynamic re-optimization model with the scheme disturbance cost and arrival delay time as optimization objectives to obtain the repair scheduling scheme;

[0080] It should be noted that the disturbance cost refers to the degree of difference between the repaired scheduling scheme and the initial scheduling scheme; the smaller the difference, the lower the disturbance cost. Arrival delay time refers to the time affected by interference; it is the time it takes for the drone to reach the rescue request point compared to the original plan, and a shorter delay is better. The repaired scheduling scheme refers to the new scheduling scheme adapted to the interference scenario obtained by solving the dynamic re-optimization model, which can ensure the continuous operation of the rescue mission.

[0081] Step S70: Output the repair scheduling scheme and the corresponding interpretable decision results. The interpretable decision results include semantic constraint result comparison relationship, decision tree causal chain and difference index.

[0082] The purpose of this step is to address unexpected interference during the rescue process, quickly generate a repair and scheduling plan that adapts to the interference scenario, and provide an explanation of the decision-making logic to ensure the continuous and efficient conduct of the rescue mission.

[0083] Understandably, the execution process of the initial scheduling scheme is monitored in real time. Upon detecting an interference event, a dynamic re-optimization model is constructed using the initial scheme as the perturbation baseline. The model is solved with the perturbation cost and arrival delay time as optimization objectives. After obtaining the repaired scheduling scheme, the repaired scheduling scheme and interpretable decision results generated based on semantic translation modeling are output. For example... Figure 5 As shown, Figure 5This is a schematic diagram of an interference management mechanism, illustrating the interference detection and handling process during a drone rescue mission. Figure 6 The diagram illustrates the workflow of an optimization result visualization module, which aims to transform the drone rescue dispatch scheme obtained through optimization algorithms into an intuitive visual output.

[0084] This embodiment provides a drone rescue scheduling optimization method based on semantic translation modeling. It transforms natural language input into a mathematical optimization model through a semantic mapping system, including entity recognition, logic extraction, and operator normalization, automatically generating computable constraints. A "initial scheduling—interruption and re-optimization" model is employed to dynamically adjust drone missions, quantifying and minimizing the cost of plan disturbances and quickly repairing scheduling schemes. A large language model and multi-objective optimization algorithm are linked, and algorithm parameters are continuously optimized through a form interface. Results are visually displayed through Gantt charts, path graphs, and decision trees, achieving transparency in AI decision-making.

[0085] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 7 This embodiment applies to the initial scheduling model. Step S30 includes steps S301 to S303, and step S40 includes steps S401 to S403:

[0086] Step S301: With the optimization objectives of minimizing response time and maximizing coverage, an initial scheduling model is constructed by combining standardized constraint expressions, resource parameters, and network structure parameters.

[0087] It's important to clarify that minimizing response time aims to minimize the total time from when a request for assistance is issued to when assistance is received, which is the core manifestation of timely assistance. Maximizing coverage aims to ensure that as many assistance requests as possible receive assistance services, which is the core manifestation of comprehensive assistance. Optimization objectives refer to the priority optimization directions that the scheduling scheme must satisfy, determining the model's construction direction. In essence, by using minimizing response time and maximizing coverage as the core optimization directions, and incorporating standardized constraint expressions, resource parameters, and network structure parameters, a preliminary scheduling model is built.

[0088] Step S302: Add single cargo flow constraints to the basic multi-objective initial scheduling model, define the coupling relationship between flow variables and routing decisions, set flow conservation constraints at the node level and load consistency constraints at the departure segment, and obtain an initial scheduling model with a complete constraint system.

[0089] It should be noted that single-cargo flow constraints refer to the constraints set for the flow process of a single relief supply, used to ensure the rationality of the supply transportation logic. Flow variables refer to variables describing the quantity and direction of supply flow, and are the core parameters for quantifying the supply transportation process. Routing decisions refer to the planning decisions for UAV flight routes, directly affecting the efficiency of supply transportation. Coupling relationships refer to the interrelationship and mutual influence between flow variables and routing decisions; that is, routing decisions determine the values ​​of flow variables, and flow variables, in turn, constrain routing decisions. Node-level flow conservation constraints refer to the constraint that the inflow and outflow of supplies remain balanced at each relief node, ensuring that supplies do not increase or decrease unnecessarily. Departure segment load consistency constraints refer to the constraint that the load of the UAV departing from the relief base remains consistent with the planned weight of transported supplies, avoiding load exceeding or falling short of limits. The initial scheduling model with a complete constraint system refers to a comprehensive constraint model that includes basic constraints and refined constraints, which can fully guarantee the feasibility of the scheduling scheme.

[0090] Step S303: Calculate the time linkage linearization constant, and integrate the time linkage linearization constant into the target initial scheduling model with complete constraint system to obtain the target initial scheduling model;

[0091] It should be noted that the time-linked linearization constant is a constant used to transform time-dependent nonlinear relationships into linear relationships, and it is the core parameter for achieving model linearization. Time linkage refers to the interrelationship and mutual influence of the time periods of various stages in the scheduling process, such as the relationship between flight time and payload weight. In essence, the time-linked linearization constant is first calculated, and then this constant is incorporated into a model containing a complete constraint system to complete the transformation from a nonlinear model to a linear model, resulting in an initial scheduling model that can be solved efficiently.

[0092] Step S401: Based on the constraints of the target initial scheduling model and the set of operating parameters of the improved multi-objective optimization algorithm, a constructive seed is used to initialize the initial population.

[0093] It should be noted that constructive seed initialization refers to the process of generating initial solutions that meet the constraints based on model constraints and the algorithm parameter set, ensuring the feasibility of the initial solutions. The initial population refers to the initial set of solutions for algorithm iterations; each solution corresponds to a potential scheduling scheme and forms the basis for subsequent optimization. The runtime parameter set refers to the set of various parameters required for the improved multi-objective optimization algorithm to run, including key parameters such as population size, crossover probability, and mutation probability.

[0094] In one feasible implementation, S401 includes steps A31-A36:

[0095] Step A31: Set the population size value based on the priority of the response time objective. It should be noted that the priority of the response time objective refers to its ranking in terms of importance among all optimization objectives; higher priority objectives require priority protection. The population size value refers to the number of solutions included in the initial population of the improved multi-objective optimization algorithm. The population size directly affects the algorithm's search range and solution efficiency.

[0096] Step A32: Set the crossover probability value based on the constraint strength of the standardized constraint expression. It should be noted that constraint strength refers to the strictness of the constraints imposed by the standardized constraint expression on the scheduling scheme; the higher the constraint strength, the smaller the feasible region of the scheme. The crossover probability value refers to the probability of two parent solutions performing a crossover operation in the improved multi-objective optimization algorithm, directly affecting the generation efficiency of new solutions and population diversity.

[0097] Step A33: Set the mutation probability value based on the complexity of the network structure parameters. It should be noted that the complexity of the network structure parameters refers to the complexity of the location distribution and connectivity between the rescue base and the demand points; the higher the complexity, the greater the difficulty of path planning. The mutation probability value refers to the probability of a solution undergoing mutation in the improved multi-objective optimization algorithm, used to increase population diversity and avoid the algorithm getting trapped in local optima.

[0098] Step A34: Set an iteration threshold based on the accuracy requirements of the coverage target. It should be noted that the accuracy requirements of the coverage target refer to the precise degree of coverage required for the proportion of rescue demand points; higher accuracy requirements necessitate a more thorough algorithm search. The iteration threshold refers to the maximum number of iterations allowed by the improved multi-objective optimization algorithm and is one of the important criteria for stopping the algorithm's iteration.

[0099] Step A35: Based on the balance requirement between response time and coverage objectives, set constraint deviation values ​​through a deviation constraint strategy. It should be noted that the balance requirement refers to finding the optimal equilibrium point between the response time and coverage objectives, ensuring both rescue speed and coverage breadth. The deviation constraint strategy refers to adjusting the optimization weights of the two objectives by setting an allowable deviation range. The constraint deviation value refers to the allowable deviation range of the objective function value, used to flexibly adjust the priority of the two objectives.

[0100] Step A36: Integrate the population size, crossover probability, mutation probability, iteration threshold, and constraint bias to obtain the operating parameter set of the improved multi-objective optimization algorithm. It should be noted that integration refers to arranging and compiling the parameters such as population size and crossover probability set from each step according to the algorithm's operational requirements, forming a logically coherent and directly callable parameter set. The operating parameter set of the improved multi-objective optimization algorithm refers to the complete set of parameters required for algorithm execution and is the core foundation for the algorithm's normal operation.

[0101] Step S402 involves performing non-dominated sorting, crowding distance calculation, crossover and mutation operations, and local refinement on the initial population. Non-dominated solutions are maintained during the iteration process using a bounded external archive until the preset convergence condition is met. It should be noted that non-dominated sorting refers to the process of dividing solutions in the population into different dominance levels based on the objective function value of the solution, used to distinguish the quality of solutions. Non-dominated means that a solution is not inferior to other solutions in all objectives and is superior to other solutions in at least one objective. Crowding distance calculation refers to calculating the crowding degree of each solution in its respective level, used to maintain population diversity. Crowding distance refers to the density of other solutions around a given solution. Crossover operation refers to exchanging some genes between two parent solutions to generate offspring solutions; this is the core method for generating new solutions. Mutation operation refers to randomly changing some genes of a solution to generate offspring solutions, used to increase population diversity and avoid local optima. Local refinement refers to the operation of locally optimizing and adjusting the generated offspring solutions, which can improve the quality of the solutions. The bounded external archive refers to the set used to store the best non-dominated solutions during the iteration process, with a capacity limit to ensure that all stored solutions are high-quality solutions. Preset convergence criteria refer to the criteria for determining when the algorithm stops iterating, such as when the number of iterations reaches a threshold or when the optimal solution is no longer being improved.

[0102] Understandably, the initial population is subjected to non-dominated sorting, crowding distance calculation, crossover and mutation operations, and local refinement. The optimal non-dominated solution is maintained through a bounded external archive until the preset convergence condition is met and the iteration stops.

[0103] In one feasible implementation, step S402 includes:

[0104] Fitness is calculated for each individual in the initial population using an objective function, which is constructed based on response time and coverage objectives. The initial population is then non-dominated based on the fitness results, resulting in sets of individuals with different dominance levels. Crowding distances within each set are calculated, and individuals with crowding distances greater than a preset crowding threshold are selected. These selected individuals are then crossbred using a single-point crossover strategy. The crossbred individuals are then mutated using a random mutation strategy to obtain offspring. Finally, a two-point exchange optimization operation is performed. The offspring individuals undergo local refinement, optimizing the path planning and adjusting the master-slave allocation relationship between the drone and the rescue base, resulting in refined offspring individuals. These refined offspring individuals are then merged with the selected individuals to obtain a merged set. A removal operation is performed on the merged set to eliminate duplicate individuals and infeasible individuals that do not conform to the standardized constraint expressions, resulting in a deduplicated subset of individuals. This deduplicated subset is then subjected to non-dominated solution filtering to obtain a set of non-dominated solutions that meets the requirements. The non-dominated solution set is input into a bounded external archive to determine if it meets the preset convergence conditions. If the preset convergence conditions are met, the iteration stops.

[0105] It should be noted that by integrating the advantages of the parent scheme through single-point cross-fertilization and introducing new scheduling possibilities through random mutation, after generating the offspring scheme, local refinement is performed through two-point exchange optimization, including 2-opt path adjustment and master-slave allocation adjustment between UAVs and bases. After merging the refined offspring and high-quality parent schemes, duplicate individuals and infeasible individuals that do not meet the standardization constraints are eliminated. Finally, the set of non-dominated solutions, i.e., the optimal candidate scheme for the current iteration, is selected. By maintaining the diversity and quality of solutions through bounded external archives, a high-quality scheduling scheme that takes into account both objectives is gradually refined.

[0106] To verify the rationality and completeness of the model design output by the form-based semantic modeling module, it is first necessary to check whether the model structure and constraints faithfully reflect real-world semantics and business logic. Then, it is checked whether the model's constraint system is complete, the feasible domain is non-empty, and there are no internal conflicts or logical overlaps. Specifically, the division of the node set (base nodes, demand nodes) accurately corresponds to the starting point and mission objective of emergency rescue; path constraints and flow conservation equations reflect mission continuity and UAV closed-loop scheduling logic; and capacity, payload, and service coverage constraints faithfully reflect the mission execution conditions under resource-constrained scenarios.

[0107] To ensure that the optimized model automatically generated by form-based semantic modeling is not only "semantically correct" but also "mathematically complete," we systematically verify the constraint system:

[0108] By having each drone take off from any base and immediately return without visiting any demand points, all core constraints can be satisfied, and only the coverage-related flags need to be set to 0. This "empty journey" construction proves that the feasible region of the model is not empty. In real-world scenarios, as long as the total supply can cover the minimum demand and the drones have effective resources, a feasible solution that meets the minimum coverage requirement can be formed, and the model still maintains a non-empty feasible region under normal parameters.

[0109] The constraint system ensures logical coherence through collaborative design: on the one hand, it achieves route closure, access consistency, and flow conservation, effectively avoiding disconnected paths and sub-loops, while tightening the feasible region; on the other hand, it avoids constraint conflicts and redundancy in dimensions such as supply control and resource constraints, and some related constraints can be selectively enabled to reduce solution errors.

[0110] Time-dependent constraints are implemented through linearization with a large constant, thereby... The structure is based on an upper bound of the sum of the worst path length and the service duration plus a small margin, thus avoiding both excessive relaxation and excessively small relaxation that would make it infeasible.

[0111] Will Set to:

[0112]

[0113] It consists of three parts: the first part is the upper limit of flight, the middle part is the upper limit of service time, and the last part is... To ensure numerical safety margin, linearization should not produce spurious feasible solutions and should avoid excessively large values. This leads to numerical instability. No feasible solutions violating the time topology were observed in the experiments. It is not feasible because it is too small.

[0114] In summary, the model's constraint set is complete and consistent at both the semantic and mathematical levels, the feasible region is non-empty, and there are no significant internal conflicts or redundant couplings. The form semantic input-model automatic generation mechanism ensures the model's interpretability and business consistency at the semantic level, laying a solid foundation for subsequent mathematical solutions.

[0115] The validity verification aims to evaluate whether the model driven by form semantic modeling can converge to a stable, interpretable solution within a finite time and exhibit the expected performance on problems of different scales. This approach selects 12 examples from the Barreto benchmark library for batch solving.

[0116] The UAV emergency rescue scheduling optimization model constructed in this paper should primarily consider rescue response time and demand satisfaction under the premise of limited transportation capacity and supply materials. Decision-makers can continue to use the form to input demand preferences under different scenarios, continuously adjusting the specific parameters and variables of the algorithm designed for the model. For example, input: Please help me design a scientific and efficient algorithm to solve the bi-objective optimization model you helped me build. Then the form system will suggest two types of algorithm solutions: exact solution and heuristic algorithm solution. Because in large-scale real-world problems, using solvers such as CPLEX and Gurobi to solve mixed-integer linear programming (MILP) models is quite time-consuming and not suitable for the bi-objective scheduling optimization model proposed in this paper, the NSGA-II multi-objective algorithm solution model recommended by the form system is selected here. The hypervolume indicator (HV) and coverage indicator (C-metric) of the Pareto solution set obtained by the traditional NSGA-II algorithm initially recommended by the form system and the continuously interactively optimized and improved NSGA-II algorithm are shown in Table 1:

[0117] Table 1. HV and C-metric indices of the two algorithms in the solution example set.

[0118]

[0119] Based on Table 1, the following conclusions can be drawn:

[0120] In the solution results for all instances, > This indicates that the improved algorithm achieves a larger non-dominated coverage in both target spaces, resulting in better overall quality and distribution of the solution set. Regarding the C-metric, most instances show... Significantly higher than This means that a large proportion of the solutions obtained using the traditional NSGA-II algorithm are dominated by solutions obtained using the improved NSGA-II algorithm; the reverse is rare, demonstrating the advantage of the improved algorithm in terms of dominance relationships.

[0121] To deeply analyze the underlying mechanisms behind the performance improvement of the improved NSGA-II algorithm, this paper summarizes three key improvement measures proposed for the traditional NSGA-II algorithm during the continuous interactive iteration process of the form system. For detailed procedures, please refer to [link to relevant documentation]. Figure 8This figure illustrates an optimization process for drone rescue scheduling based on an improved NSGA-II algorithm. The process begins with initializing the population and calculating fitness (response time and coverage). Next, non-dominated sorting and crowding distance calculations are performed, followed by the generation of a new population through selection, crossover, and mutation operations. Local search improvements are then implemented, including 2-opt path optimization and base master-slave adjustment. If convergence is not achieved, bounded external archive updates are performed. Once convergence is achieved, the Pareto front solution set is output, and the solution is visualized and analyzed.

[0122] First, constructive population initialization: This aims to generate several highly feasible and high-quality "seed solutions" through greedy or heuristic strategies, and mix them with random individuals to form the initial population. This effectively reduces the proportion of infeasible solutions in early iterations, improving the algorithm's convergence starting point and solution diversity. Second, local improvement: In each generation or every other generation, lightweight local searches such as 2-opt operations and neighborhood rearrangements are performed on some individuals to improve the quality of the Pareto front with lower computational cost. Finally, a bounded archive mechanism: Through periodic archiving and crowding distance maintenance, the diversity and distribution balance of the solution set are maintained, while preventing high-quality solutions from being lost during evolution, balancing global exploration capability and local boundary fidelity.

[0123] The synergistic effect of the three mechanisms enables the improved algorithm to significantly outperform the traditional NSGA-II in terms of solution feasibility, convergence performance, and Pareto front quality. To further analyze the performance of the improved algorithm obtained through continuous form interaction, a detailed analysis is conducted using the small example coordGaspelle3.dat and the large example coordDas150.dat.

[0124] The improved NSGA-II derived from continuous form interaction outperforms the traditional NSGA-II based on initial form suggestions in both examples. Combined with the analysis in Table 1, we can conclude that:

[0125] In the small example coordGaspelle3, the HV values ​​are 13.733 and 11.559 respectively, and This indicates that the Pareto solution set obtained by the improved NSGA-II algorithm almost completely covers the non-dominated boundary of the traditional NSGA-II algorithm, with a more "convex" front and a more coherent distribution. This verifies that "constructive seed + lightweight local improvement" can effectively accelerate convergence and improve boundary quality on small-scale, geometrically regular instances.

[0126] On the large example coordDas150, the HV values ​​are 7127.15 and 2500.40, respectively, and In high-dimensional combinatorial spaces, the improved NSGA-II algorithm's external archive and diversity maintenance play a crucial role, significantly outperforming the traditional NSGA-II algorithm in wide-area search and Pareto boundary expansion; the traditional NSGA-II algorithm exhibits a typical "boundary collapse / sparse coverage" phenomenon in its Pareto solution set. This comparison demonstrates the stability advantage of the improved mechanism in large-scale and difficult geometric cases.

[0127] Step S403: Select the optimal solution for the target from the bounded external archive to obtain an initial scheduling scheme that includes UAV route allocation, mission execution sequence and material delivery plan.

[0128] It should be noted that the optimal solution refers to the solution that achieves the best balance between response time and coverage objectives, and is the optimal solution that meets the rescue needs. Drone routing allocation refers to specifying the flight route of each drone, that is, the specific path from the rescue base to each rescue point of need. Mission execution sequencing refers to specifying the order and time nodes for each drone to perform various tasks, ensuring the orderly conduct of missions. The material delivery plan specifies the type and quantity of materials carried by each drone and the amount to be delivered to each need point, ensuring accurate material delivery.

[0129] Understandably, based on the actual needs of the rescue mission, the optimal solution is selected from the non-dominated solutions in the bounded external archive, and an initial scheduling scheme including drone routing, mission execution timing, and material delivery plan is obtained analytically. In the specific implementation, the model building process under multi-round form training and inference includes: determining the objective function, where objective 1 is to minimize the overall response time (summation type): Objective 2 is to maximize coverage:

[0130] Define the constraints, including base activation / deactivation and drone assignment:

[0131]

[0132]

[0133] Drones depart from and are recovered from the base (start-end constraints):

[0134]

[0135]

[0136] Access and flow conservation:

[0137]

[0138]

[0139]

[0140] Single cargo flow and capacity (including sub-loop elimination):

[0141]

[0142]

[0143]

[0144] Supply aggregation and capacity limits:

[0145]

[0146]

[0147] Unmanned aerial vehicle (UAV) payload and load consistency (optional tightening):

[0148]

[0149] Coverage and Demand Fulfillment:

[0150]

[0151] Energy consumption and range:

[0152]

[0153]

[0154] Response time is linked to arrival time; assuming a certain activation base... If the takeoff time is 0, then:

[0155]

[0156]

[0157]

[0158] The parameters include:

[0159] :node The flight distance (calculated from coordinates). :node Flight time (if the speed is uniform, it can be made) ). :node Energy consumption (configurable) ). Drones The maximum load. Drones Maximum energy consumption. Drones Maximum range (optional). :base The available supply of materials. Demand Points The demand. Demand Points Minimum service ratio threshold (reaching) (Consider it as coverage). : A sufficiently large constant (Big-M, used for time and logic linkage).

[0160] Decision variables include: Drones Is it an arc? . Drones Do you need to access the request point? . Drones Whether by the base Initiate and reclaim (base master-slave assignment). :base Enable or disable? Demand Points Is it covered (meeting the minimum service ratio)? Demand Points Actual amount of supplies obtained (partially available for supply). Drones In the arc The flow of goods carried on the platform (single commodity flow). Drones Reaching the demand point At that moment. Demand Points Arrival time (across drones for responding to targets).

[0161] This embodiment provides a method for optimizing UAV rescue scheduling based on semantic translation modeling. It optimizes UAV rescue scheduling by constructing an initial scheduling model in a hierarchical manner. A basic framework is established, incorporating single-cargo flow constraints and time-linked processing to ensure model integrity and solvability. The rationality of algorithm initialization is improved by adjusting parameters such as population size and crossover probability. The iterative optimization process includes fitness calculation, non-dominated sorting, crossover mutation, and local refinement, using bounded external archives to maintain the optimal solution, ensuring efficient and high-quality solutions. Two-point exchange optimization is applied to improve the rationality of path planning, ensuring a balanced response time and coverage for the scheduling scheme. This provides a scientific and feasible execution basis for rescue missions, enhancing the intelligence and precision of UAV scheduling.

[0162] Based on the first and second embodiments of this application, in the third embodiment of this application, the content that is the same as or similar to the above embodiments can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 9 This embodiment applies to a dynamic re-optimization model under interference scenarios. Step S40 includes steps S501~S503, step S60 includes steps S601~S602, and step S70 includes steps S701~S702.

[0163] Step S501: Monitor the execution process of the initial scheduling plan in real time and identify whether there are any interference events. Interference events include at least UAV failure, sudden environmental changes, new demand points, and changes to the demand order.

[0164] It should be noted that drone malfunction refers to abnormal equipment conditions such as mechanical failures, electrical failures, or control system malfunctions that occur during drone flight or mission execution, preventing the drone from completing the mission normally. Sudden environmental changes refer to unexpected changes in the natural environment of the rescue area, such as severe weather like strong winds, heavy rain, or dense fog, or geographical changes like terrain collapses, which can affect drone flight safety and mission execution efficiency. New demand points refer to newly emerging locations requiring rescue support during the execution of the dispatch plan, i.e., newly added disaster-stricken population gathering points or areas lacking supplies. Changes to demand orders refer to adjustments in existing rescue demand information, such as changes in the type of rescue supplies or adjustments to rescue priorities.

[0165] Understandably, by using a real-time monitoring system to track the execution status of the initial scheduling plan, the system can dynamically investigate the drone's operating status, the rescue area environment, and rescue needs, and identify any interference events such as drone malfunctions or sudden environmental changes.

[0166] Step S502: When an interference event is detected, the core parameters of the initial scheduling scheme are extracted as the disturbance benchmark. It should be noted that the core parameters refer to the key data in the initial scheduling scheme that plays a decisive role in the scheduling logic and are the foundation for ensuring the core functions of the scheduling scheme. In other words, when an interference event is detected, core parameters such as UAV route allocation, task execution sequence, and material delivery plan are extracted from the initial scheduling scheme and used as the disturbance benchmark for subsequent dynamic re-optimization.

[0167] Step S503 involves constructing a dynamic re-optimization model based on the disturbance baseline, standardized constraint expressions, and actual operating condition parameters after disturbance. It should be noted that actual operating condition parameters refer to the actual state parameters of the rescue scenario after the disturbance event occurs, such as the number of faulty drones, the altered environmental conditions, and the location of newly added demand points—data reflecting the current real-world rescue scenario. In essence, this involves using the extracted disturbance baseline as a foundation, combining it with the original standardized constraint expressions, and incorporating the actual operating condition parameters after disturbance to construct a dynamic re-optimization model adapted to the current rescue scenario.

[0168] Step S601: Set the optimization objectives of the dynamic re-optimization model as minimizing the scheme perturbation cost and minimizing the arrival delay time. Solve the dynamic re-optimization model using an improved multi-objective optimization algorithm. Maintain the non-dominated solution in the iteration process through bounded external archives until the preset convergence condition is met. It should be noted that the improved multi-objective optimization algorithm refers to an intelligent algorithm that improves upon the traditional multi-objective optimization algorithm and can better balance multiple conflicting objectives.

[0169] Understandably, minimizing the perturbation cost and arrival delay time of the scheme are taken as the core optimization objectives of the dynamic re-optimization model. An improved multi-objective optimization algorithm is called to solve the model. The optimal non-dominated solution in the iteration process is maintained by a bounded external archive until the preset convergence condition is met and the iteration stops.

[0170] Step S602: Select the optimal solution for the target from the bounded external archive to obtain the repair scheduling scheme; it can be understood that, based on the actual needs of the rescue mission, the optimal solution for the target is selected from the non-dominated solutions stored in the bounded external archive, and the repair scheduling scheme adapted to the current interference scenario is obtained by parsing.

[0171] Step S701: Based on the semantic translation modeling method, the decision logic of the repair scheduling scheme is translated to generate an interpretable decision result. The interpretable decision result includes the semantic constraint result comparison relationship, the decision tree causal chain, and the difference index.

[0172] It should be noted that interpretability information refers to the auxiliary decision-making understanding data output by the system, which is used to present the complete logical chain from task input to solution generation, making scheduling decisions transparent; semantic constraint result mapping relationship refers to the mapping table between the input natural language rules and the automatically generated constraint expressions and final scheduling actions; decision tree causal chain refers to the causal logic of decision nodes such as interference event triggering, goal trade-offs, and solution repair, which are visualized in the form of a decision tree; difference index quantifies the adjustment range of the repair solution compared to the initial solution, including the number of path changes, the number of assignment changes, total delay time, and energy consumption increment.

[0173] Understandably, the result translation method is used to analyze and translate the algorithmic decision-making logic behind the repair scheduling scheme, transforming the complex technical logic into plain natural language and generating interpretable decision results.

[0174] Step S702: Output the repair scheduling plan and interpretable decision results to the display terminal.

[0175] It should be noted that the display terminal refers to the electronic device used to display dispatch plans and decision results, such as the display screen in the rescue command center and the mobile terminals of staff. It is the core carrier for rescue personnel to obtain dispatch information. Output refers to the process of presenting the repair dispatch plan and interpretable decision results in a visual form on the display terminal, ensuring that rescue personnel can quickly obtain and view relevant information.

[0176] Understandably, the generated repair and scheduling plan and interpretable decision results are processed in a preset visualization format and output to the display screen of the rescue command center or the mobile terminal of the staff for rescue personnel to refer to and implement.

[0177] In one feasible implementation, step S702 includes steps A41-A45:

[0178] Step A41: Generate early warning information based on infeasible scenarios encountered during the dynamic re-optimization model solution process;

[0179] It should be noted that infeasible scenarios refer to scenarios identified during the dynamic re-optimization model solving process where the rescue mission cannot be completed with existing resources and constraints, such as insufficient total resources to cover all demand points or infeasible flight routes under extreme conditions.

[0180] Step A42: Perform batch repair processing on cascading failure scenarios, divide repair priorities and generate step-by-step execution plans;

[0181] It's important to note that a cascading failure scenario refers to a situation where an initial fault triggers a chain reaction, leading to malfunctions in multiple devices or systems. For example, a malfunction in one drone might cause other drones to experience mission overload and subsequently malfunction. Batch repair refers to dividing multiple repair tasks within a cascading failure scenario into different batches based on priority, thus avoiding chaotic repair processes. Repair priority refers to ranking the repair tasks by their urgency, prioritizing the repair of faults affecting core rescue missions. A phased execution plan clearly defines the execution sequence, timelines, and responsible parties for each batch of repair tasks, ensuring the orderly progress of the repair work.

[0182] Step A43: Output compliance prompts through red and yellow light indicators, and mark the corresponding warning signs for situations such as failure to avoid no-fly zones, exceeding resource limits, and exceeding the threshold for lateness;

[0183] It should be noted that the red and yellow light system uses red and yellow lights to distinguish different warning levels. A red light indicates a serious violation or high risk, while a yellow light indicates a general violation or low risk. The compliance alert informs rescue personnel whether the repair and dispatch plan meets preset constraints. No-fly avoidance failure refers to a situation where the planned flight route in the repair and dispatch plan fails to avoid no-fly zones. Resource exceeding limits refers to situations where the drone's payload, battery power, or other resources exceed their limits. Delay exceeding the threshold refers to situations where the drone's arrival time at the required point exceeds a preset threshold. Warning signs are distinctive indicators set for specific violations, used to quickly distinguish different types of violations.

[0184] Step A44: Integrate the remediation scheduling plan, interpretable decision results, early warning information, step-by-step execution plan and compliance prompts to generate a standardized output report;

[0185] It should be noted that integration refers to the process of summarizing and organizing the repair and dispatch plan, interpretable decision results, early warning information, step-by-step execution plan, and compliance tips according to a preset format to form a logically coherent and complete overall document. Standardized output reports refer to output documents that conform to a unified format specification, containing all the information required for rescue execution, facilitating unified viewing and use by rescue personnel.

[0186] Step A45: Convert the standardized output report into a preset format and output it to the display terminal.

[0187] It should be noted that the preset format refers to a pre-defined format adapted to the display terminal, such as a combination of text and images, a table format, etc., to ensure that the report is clearly readable on the display terminal. Conversion refers to the process of adjusting the original format of the standardized output report to the preset display format to adapt to the display requirements of different display terminals.

[0188] In practical implementation, during the execution of UAV emergency rescue missions, interference events such as UAV malfunctions, new mission locations, route blockages / no-fly zones, and sudden weather changes may occur. To ensure the stability of the solution and rapid recovery of performance, a dynamic re-optimization model based on "minimum disturbance—rapid feasibility—interpretability" is constructed. This model is based on the mission execution time... To optimize the starting point, considering remaining resources, network accessibility, and new demands, a dual objective is adopted: Objective A (stability priority) minimizes the cost of scheme disturbances (path changes, access assignment changes, and energy consumption increments); Objective B (timeliness priority) minimizes arrival delay time (including incomplete points and new points).

[0189] After the interference occurred, in the original network Based on the network formed by the first layer model, still let To adapt to interference scenarios, the requirements are divided into a set of completed requirements. Incomplete collection With the new set And make the set of requirements to be optimized At the same time, the arc set is divided into a set of hindered arcs. With available arc set The drone ensemble remains Parameters are divided into two categories: static parameters and execution-time state / preference parameters. Static parameters are common to the first-layer model and include:

[0190] :arc Flight distance; :arc Flight time (which can be converted to a uniform speed); :arc Energy consumption coefficient; Drones The maximum load; Drones The upper limit of energy consumption; Drones Maximum range (optional); :base Available supply; Demand Points Demand; Demand Points The minimum service ratio threshold; Large constants (used for time-linked linearization). Execution-time state and preference parameters include: The system moment that triggers re-optimization; :time drones The current location; , , Corresponding drone Remaining energy consumption / range / payload; : Remaining supplies at the base; The baseline scheme is the suffix of the plan before the interference occurred (the original route / access / arrival time that was to be executed when no interference occurred). Arc modification and access assignment modification budget; Disturbance cost weights (path change, delay time, energy consumption increment); New points The expected / commitment deadline (if not given, it can be filled in from the rule base).

[0191] The decision variable system covers three core categories: routing and access, logistics and supply, and timing and start / stop. At the same time, a new variable for describing disturbances has been added.

[0192] Routing and access include: Drones Is it an arc? ; Drones Access ; : Has the coverage threshold been reached?

[0193] Logistics and supply include: : In the arc The capacity of the data stream; From the base pass service The amount; .

[0194] Timing and on / off include: : arrive The moment; : Arrival time (cross drone aggregation); : Whether by the base Start and end; :base Enable or disable?

[0195] The perturbation characterization includes: : Has the route changed relative to the baseline? : Whether the access assignment has changed relative to the baseline; : Energy consumption increment relative to the benchmark; : Number of late arrivals.

[0196] If a fixed recycling site is required, the corresponding The baseline value is fixed; if a drone malfunctions and exits, its remaining resources can be set to 0 or only a return-to-home feasible path can be allowed.

[0197] The objective function includes:

[0198] Objective A: Minimize the cost of the disturbance.

[0199]

[0200] Objective B: Minimize total lateness:

[0201] The following seven types of constraints work together to ensure the feasibility, stability, and timeliness of the solution:

[0202] (1) The prohibited area is aligned with the starting point:

[0203]

[0204] (2) Completed tasks are frozen:

[0205]

[0206] (3) Access consistency and path closure:

[0207]

[0208]

[0209]

[0210] (4) Capacity and current conservation (including sub-loop suppression):

[0211]

[0212]

[0213]

[0214]

[0215] (5) Supply aggregation and coverage threshold:

[0216]

[0217]

[0218] (6) Energy consumption / range and time linkage:

[0219]

[0220]

[0221]

[0222] (7) Definition of lateness and variance budget:

[0223]

[0224]

[0225]

[0226]

[0227]

[0228] variable: .

[0229] In the actual implementation, select the interference type and enter / confirm. Add point attributes Blocked area (automatically mapped to) Disruption budget With weight Entity identification and relation extraction generate "field-constraint-variable" mapping entries; for default... Perform rule completion and consistency verification; automatically generate optimized models; if "Lock recycling base / base master-slave" is selected, the system automatically fixes relevant variables or includes them in the disturbance budget; by default, warm start is performed (using the baseline solution as the initial solution and seed neighborhood), and the search is conducted within the "rolling time domain + local feasible domain" to ensure minute-level recalculation; output the repaired path, time sequence, and supply allocation; synchronously provide... , , , Equal difference indicators and comparison tables (semantic-constraint-result), and displayed using Gantt charts / path diagrams / coverage heatmaps; continue execution under the new scheme; if interference occurs again, repeat the steps to form a closed loop.

[0230] Algorithm configuration and recommendations include:

[0231] Multi-objective strategy: Default lexicographical order (stability first, then...) back );when For hard budgets, ε-constraints are recommended.

[0232] Improved NSGA-II: Constructed seed initialization (including baseline neighborhood), lightweight local search (2-opt / insertion / late-driven fine-tuning), bounded external files and feasibility repair.

[0233] Parameter recommendations:

[0234] Small disturbances ( Population 40–60, iterations 80–150;

[0235] Medium-sized disturbances ( Population 80–120, iterations 150–300;

[0236] For large disturbances: first tighten the search domain (increase the lock rate and prohibit changing the recycling base), then appropriately relax the iteration; if time is tight, the fast mode can be activated (reduce iteration and increase the local search frequency).

[0237] To adapt to scheduling requirements with rigid budgets or time commitments, this implementation provides an ε-constraint solution strategy. The basic idea is to retain one objective for optimization while transforming the other objectives into threshold constraints. Feasible optimal solutions with different preferences are obtained by adjusting the thresholds. For this model, two commonly used configurations are provided:

[0238] (1) Rigid budget and time-sensitive priority: Add the following to the constraint set:

[0239]

[0240] and with For the goal.

[0241] (2) Time-sensitive and stability-first: Add new constraints to the constraint set and with The target is to set the thresholds. All the thresholds mentioned above can be input parameterized via a form; the system supports scanning multiple sets of thresholds and outputting non-dominated solution sets that conform to different command preferences.

[0242] The solution process can directly adopt the aforementioned multi-objective configuration: under the warm start and local feasible region search mechanism, combined with constructed seed initialization, lightweight local search, and bounded external file maintenance, dynamic re-optimization can be achieved in minutes. Parameter value ranges can be automatically recommended through a form and fine-tuned online, without the need for additional offline parameter tuning.

[0243] Output items and exception handling include:

[0244] The report shows the number of arcs changed. Change the number of assignments Always late Energy consumption increment List of affected tasks and recovery path.

[0245] Infeasibility alert: When insufficient budget or tight deadlines render the task infeasible, the system automatically provides minimum allowance suggestions. or (the extent of the relaxation).

[0246] When multiple machines exit in succession, the cascading failure strategy recommends incremental batch repair (small disturbances in stages) to reduce the impact on execution.

[0247] Compliance and safety alerts provide red / yellow light warnings for reasons such as failure to avoid no-fly zones, exceeding resource limits, or being late and exceeding thresholds.

[0248] The model does not require reconstructing the overall solution and achieves feasible repair within minutes with minimal perturbation. Its core effects are reflected in four aspects: First, warm start-up and local feasible domain search significantly shorten recalculation time; second, explicit modeling of perturbation budgets and differential variables makes stability and timeliness measurable and balancing; third, the causal chain of semantics-constraints-results and differentiated reports support review and verification and transparent command; and fourth, the form fields and constraint templates are universal and can be reused in other emergency scenarios and combinatorial optimization tasks, with good transferability.

[0249] To illustrate, consider a scenario with 2 bases, 3 drones, and 12 demand points: At time τ, a new demand point is added, and a no-fly zone appears. By setting Larc=5, Lasg=2, and α:β:γ=1:2:1 in a form, and employing an ε-constraint configuration prioritizing "budget rigidity and timeliness," the system performs re-optimization based on a warm start, outputting a repair scheduling scheme with no more than 5 modified arcs and no more than 2 assignment changes, while simultaneously generating differentiated reports on total delays, energy consumption increments, and path changes. This example verifies the feasibility of the model for rapid and minimal-disturbance recovery from disturbance events under a limited disturbance budget.

[0250] This embodiment provides a method for optimizing UAV rescue scheduling based on semantic translation modeling. It ensures data standardization and completeness through structured forms and natural language information collection. Semantic parsing and constraint transformation are performed using an adapted large language model, reducing the difficulty of processing unstructured information and improving parameter extraction accuracy. A multi-objective initial scheduling model is constructed and solved using an improved optimization algorithm to achieve an optimal balance between response time and coverage, ensuring the scientific validity and feasibility of the initial scheduling. A dynamic re-optimization mechanism for interference minimizes adjustment costs, rapidly generates repair scheduling schemes, enhances the system's anti-interference capability and emergency response speed, and improves the efficiency and reliability of UAV emergency scheduling.

[0251] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the UAV rescue scheduling optimization method based on semantic translation modeling in this application. Any simple transformations based on this technical concept are within the protection scope of this application.

Claims

1. A method for optimizing the dispatch of unmanned aerial vehicles (UAVs) for rescue based on semantic translation modeling, characterized in that, The method comprises: obtaining emergency rescue task description information; performing semantic analysis on the task description information by a large language model, converting natural language rules in the task description information into a standardized constraint expression, and identifying a response time target, a coverage range target, resource parameters, and network structure parameters; constructing an initial scheduling model based on the response time target, the coverage range target, the standardized constraint expression, the resource parameters, and the network structure parameters; solving the initial scheduling model by an improved multi-objective optimization algorithm to obtain an initial scheduling scheme for unmanned aerial vehicles; when an interference event is detected during execution of the initial scheduling scheme, constructing a dynamic re-optimization model with the initial scheduling scheme as a disturbance reference; solving the dynamic re-optimization model with scheme disturbance cost and arrival delay time as optimization objectives to obtain a repaired scheduling scheme; outputting the repaired scheduling scheme and an interpretable decision result corresponding to the repaired scheduling scheme, the interpretable decision result including a semantic constraint result comparison relationship, a decision tree causal chain, and a difference index; The step of performing semantic analysis on the task description information by a large language model, converting natural language rules in the task description information into a standardized constraint expression, and identifying a response time target, a coverage range target, resource parameters, and network structure parameters comprises: identifying key entities in the emergency rescue task description information by a large language model, wherein the key entities include unmanned aerial vehicles, rescue demand points, rescue bases, and task execution time periods; extracting association relationships between the key entities, wherein the association relationships include a subordinate relationship between unmanned aerial vehicles and rescue bases, and a corresponding relationship between rescue demand points and priorities; extracting natural language rules through entity recognition and logical relationship extraction on the emergency rescue task description information, and converting the natural language rules into a standardized constraint expression through operator normalization processing; identifying the response time target and the coverage range target from the emergency rescue task description information, wherein the response time target is to minimize total rescue response time, and the coverage range target is to maximize rescue demand point coverage ratio; extracting the resource parameters and the network structure parameters corresponding to the key entities, wherein the resource parameters include an upper limit of unmanned aerial vehicle power and an upper limit of load, and the network structure parameters include location coordinates and traffic connection relationships between rescue bases and demand points; performing logical verification on the standardized constraint expression, the resource parameters, and the network structure parameters in combination with prior knowledge in the emergency rescue field, correcting logical conflict content, and completing semantic analysis, standardized constraint expression conversion, and target and parameter identification; The step of constructing a dynamic re-optimization model with the initial scheduling scheme as a disturbance reference when an interference event is detected during execution of the initial scheduling scheme comprises: real-time monitoring of the execution process of the initial scheduling scheme to identify whether there is an interference event, wherein the interference event at least includes unmanned aerial vehicle failure, environmental mutation, demand point addition, and demand change; When the interference event is detected, core parameters of the initial scheduling scheme are extracted as a disturbance reference; Based on the disturbance reference, the standardized constraint expression and the actual working condition parameters after interference, a dynamic re-optimization model is constructed; The step of solving the dynamic re-optimization model with the scheme disturbance cost and the arrival delay time as the optimization objectives to obtain a repair scheduling scheme, comprising: The optimization objectives of the dynamic re-optimization model are set as minimization of the scheme disturbance cost and minimization of the arrival delay time, and the dynamic re-optimization model is solved by the improved multi-objective optimization algorithm, and the non-dominated solutions in the iteration process are maintained through the bounded external archive until the preset convergence condition is met; The target optimal solution is selected from the bounded external archive to obtain a repair scheduling scheme; The step of outputting the repair scheduling scheme and the interpretable decision result corresponding to the repair scheduling scheme, wherein the interpretable decision result comprises a semantic constraint result correspondence relationship, a decision tree causal chain and a difference index, comprising: Based on the semantic translation modeling method, the decision logic of the repair scheduling scheme is result-translated to generate an interpretable decision result, which comprises a semantic constraint result correspondence relationship, a decision tree causal chain and a difference index; The repair scheduling scheme and the interpretable decision result are output to a display terminal. 2.The unmanned aerial vehicle rescue scheduling optimization method based on semantic translation modeling according to claim 1, wherein, The step of obtaining emergency rescue task description information, comprising: Structured task information is collected through an emergency task input form, and the structured task information includes a rescue area, a number of unmanned aerial vehicles and a rescue base location; Natural language task supplement information input by a user is received, and the natural language task supplement information includes rescue priority instructions and special scene constraint descriptions; The structured task information and the natural language task supplement information are subjected to integrity checking, and information that does not conform to a preset format is removed to obtain checked structured task information and checked natural language task supplement information; The checked structured task information and the checked natural language task supplement information are integrated to obtain complete emergency rescue task description information. 3.The unmanned aerial vehicle rescue scheduling optimization method based on semantic translation modeling according to claim 1, wherein, The step of identifying key entities in the emergency rescue task description information by using a large language model, comprising: Text preprocessing is performed on the emergency rescue task description information to remove redundant characters and invalid expressions to obtain preprocessed emergency rescue task description information; An emergency rescue field entity dictionary is loaded, and the emergency rescue field entity dictionary is input into the large language model for adaptation optimization to obtain a field adaptation model for entity recognition; The entity recognition optimization model adapted to the emergency rescue scene is called to perform sentence-by-sentence entity recognition on the preprocessed emergency rescue task description information to mark candidate entities; The candidate entities are classified by type to distinguish unmanned aerial vehicle entities, rescue target entities, resource entities and time entities; The classified candidate entities are subjected to consistency checking to remove repeatedly identified entities to obtain final key entities. 4.The method of claim 1, wherein, The step of constructing an initial scheduling model based on the response time target, the coverage range target, the standardized constraint expression, the resource parameter and the network structure parameter comprises: An initial scheduling model is constructed with the optimization objectives of minimizing the response time target and maximizing the coverage range target, combined with the standardized constraint expression, the resource parameter and the network structure parameter; Single-goods flow constraints are added to the initial scheduling model to define the coupling relationship between flow variables and routing decisions, set node-level flow conservation constraints and departure section load consistency constraints, and obtain an initial scheduling model with a complete constraint system; A time linkage linearization constant is calculated, and the time linkage linearization constant is integrated into the target initial scheduling model with a complete constraint system to obtain a target initial scheduling model; The step of solving the initial scheduling model by using the improved multi-objective optimization algorithm to obtain an initial scheduling scheme of the unmanned aerial vehicle comprises: An initial population is obtained by constructingively initializing seeds based on the constraint requirements of the target initial scheduling model and the operating parameter set of the improved multi-objective optimization algorithm; Non-dominated sorting, crowded distance calculation, crossover and mutation operations and local refinement processing are performed on the initial population, and non-dominated solutions in an iteration process are maintained through a bounded external archive until a preset convergence condition is met; A target optimal solution is selected from the bounded external archive to obtain an initial scheduling scheme including unmanned aerial vehicle routing allocation, task execution timing and material delivery planning.

5. The method of claim 4, wherein, The step of performing non-dominated sorting, crowded distance calculation, crossover and mutation operations and local refinement processing on the initial population, and maintaining non-dominated solutions in an iteration process through a bounded external archive until a preset convergence condition is met comprises: Fitness calculation is performed on each individual in the initial population through a target function to obtain a fitness calculation result, wherein the target function is constructed based on the response time target and the coverage range target; The initial population is non-dominated sorted based on the fitness calculation result to divide the initial population into individual sets of different dominance levels; The crowded distances of individuals in each individual set are calculated, and individuals with crowded distances greater than a preset crowded threshold are screened based on the crowded distances to obtain screened individuals; The screened individuals are subjected to crossover operation through a single-point crossover strategy to obtain post-crossover individuals; The post-crossover individuals are subjected to mutation operation through a random mutation strategy to obtain offspring individuals; The offspring individuals are subjected to local refinement processing through a two-point exchange optimization operation to optimize path planning and adjust the master-slave allocation relationship between the unmanned aerial vehicle and the rescue base, and obtain refined offspring individuals; The refined offspring individuals and the screened individuals are combined to obtain a combined individual set; The combined individual set is subjected to a rejection operation to reject duplicate individuals and infeasible individuals that do not meet the standardized constraint expression, and obtain a subset of de-duplicated and screened individuals; The subset of de-duplicated and screened individuals is subjected to non-dominated solution screening to obtain a set of required non-dominated solutions. Input the non-dominated solution set into a bounded external file, determine whether a preset convergence condition is met, and stop iteration when the preset convergence condition is met.

6. The method of claim 4, wherein, The method further comprises, before the method: Setting a population size value based on a priority of the response time target; Setting a crossover probability value based on a constraint strength of the standardized constraint expression; Setting a mutation probability value based on a complexity of the network structure parameter; Setting an iteration number threshold value based on an accuracy requirement of the coverage range target; Setting a constraint deviation value through a deviation constraint strategy based on a balance requirement of the response time target and the coverage range target; Integrating the population size value, the crossover probability value, the mutation probability value, the iteration number threshold value, and the constraint deviation value to obtain a set of running parameters of the improved multi-objective optimization algorithm.

7. The method of claim 1, wherein, The step of outputting the repair scheduling scheme and the interpretable decision result to a display terminal comprises: generating early warning information according to an infeasible scenario in a solving process of the dynamic re-optimization model; performing batch repair processing on a cascading failure scenario, dividing repair priorities, and generating a step-by-step execution plan; outputting compliance prompts through red-yellow light identifiers, and marking corresponding warning identifiers for situations of failure of flight ban avoidance, resource over-limit, and delay over threshold; integrating the repair scheduling scheme, the interpretable decision result, the early warning information, the step-by-step execution plan, and the compliance prompts to generate a standardized output report; converting the standardized output report into a preset format and outputting the report to a display terminal.

8. An unmanned aerial vehicle rescue dispatch optimization device based on semantic translation modeling, characterized in that, The device comprises: a task information acquisition module configured to acquire emergency rescue task description information; a semantic analysis processing module configured to perform semantic analysis on the task description information through a large language model, convert natural language rules in the task description information into standardized constraint expressions, and identify a response time target, a coverage range target, resource parameters, and network structure parameters; the semantic analysis processing module is further configured to identify key entities in the emergency rescue task description information through the large language model, wherein the key entities include unmanned aerial vehicles, rescue demand points, rescue bases, and task execution time periods; extract the association relationships between the key entities, wherein the association relationships include a subordinate relationship between unmanned aerial vehicles and rescue bases and a corresponding relationship between rescue demand points and priorities; extract natural language rules through entity recognition and logical relationship extraction on the emergency rescue task description information, and convert the natural language rules into standardized constraint expressions through operator normalization processing; identify the response time target and the coverage range target from the emergency rescue task description information, wherein the response time target is to minimize total rescue response time, and the coverage range target is to maximize rescue demand point coverage ratio; extract the resource parameters and the network structure parameters corresponding to the key entities, wherein the resource parameters include an upper limit of unmanned aerial vehicle power and an upper limit of load, and the network structure parameters include location coordinates and traffic connection relationships of rescue bases and demand points; and a repair scheduling module configured to generate a repair scheduling scheme based on the standardized constraint expression, the response time target, the coverage range target, the resource parameters, and the network structure parameters. The device comprises: a task information acquisition module configured to acquire emergency rescue task description information; a semantic analysis processing module configured to perform semantic analysis on the task description information through a large language model, convert natural language rules in the task description information into standardized constraint expressions, and identify a response time target, a coverage range target, resource parameters, and network structure parameters; the semantic analysis processing module is further configured to identify key entities in the emergency rescue task description information through the large language model, wherein the key entities include unmanned aerial vehicles, rescue demand points, rescue bases, and task execution time periods; extract the association relationships between the key entities, wherein the association relationships include a subordinate relationship between unmanned aerial vehicles and rescue bases and a corresponding relationship between rescue demand points and priorities; extract natural language rules through entity recognition and logical relationship extraction on the emergency rescue task description information, and convert the natural language rules into standardized constraint expressions through operator normalization processing; identify the response time target and the coverage range target from the emergency rescue task description information, wherein the response time target is to minimize total rescue response time, and the coverage range target is to maximize rescue demand point coverage ratio; extract the resource parameters and the network structure parameters corresponding to the key entities, wherein the resource parameters include an upper limit of unmanned aerial vehicle power and an upper limit of load, and the network structure parameters include location coordinates and traffic connection relationships of rescue bases and demand points; and a repair scheduling module configured to generate a repair scheduling scheme based on the standardized constraint expression, the response time target, the coverage range target, the resource parameters, and the network structure parameters. The standardization constraint expression, the resource parameter, and the network structure parameter are logically checked in combination with prior knowledge in the emergency rescue field, logical conflict contents are revised, and semantic analysis, standardization constraint expression conversion, and target and parameter identification are completed; An initial scheduling solving module is configured to construct an initial scheduling model based on the response time target, the coverage range target, the standardization constraint expression, the resource parameter, and the network structure parameter; and solve the initial scheduling model by using an improved multi-objective optimization algorithm to obtain an initial scheduling scheme of the UAV. A dynamic re-optimization module is configured to construct a dynamic re-optimization model based on the initial scheduling scheme as a disturbance reference when detecting an interference event in an execution process of the initial scheduling scheme; solve the dynamic re-optimization model to obtain a repair scheduling scheme, with a scheme disturbance cost and an arrival delay time as optimization targets; and output the repair scheduling scheme and an interpretable decision result corresponding to the repair scheduling scheme, the interpretable decision result including a semantic constraint result comparison relationship, a decision tree causal chain, and a difference index. The dynamic re-optimization module is further configured to monitor the execution process of the initial scheduling scheme in real time, and identify whether there is an interference event, wherein the interference event at least includes UAV failure, environmental mutation, demand point addition, and demand single change. When detecting the interference event, the core parameter of the initial scheduling scheme is extracted as a disturbance reference. A dynamic re-optimization model is constructed based on the disturbance reference, the standardization constraint expression, and actual working condition parameters after the interference. The dynamic re-optimization module is further configured to set an optimization target of the dynamic re-optimization model as minimization of a scheme disturbance cost and minimization of an arrival delay time, solve the dynamic re-optimization model by using the improved multi-objective optimization algorithm, and maintain non-dominated solutions in an iteration process through a bounded external file until a preset convergence condition is met. A target optimal solution is selected from the bounded external file to obtain a repair scheduling scheme. The dynamic re-optimization module is further configured to translate a decision logic of the repair scheduling scheme based on a semantic translation modeling method to generate an interpretable decision result, the interpretable decision result including a semantic constraint result comparison relationship, a decision tree causal chain, and a difference index. The repair scheduling scheme and the interpretable decision result are output to a display terminal.

Images