An intelligent task allocation method for land-air unmanned system

Abstract

The present application relates to the technical field of land-air unmanned system task allocation, and particularly relates to a land-air unmanned system intelligent task allocation method, comprising: establishing a heuristic task allocation database based on a plurality of task environments and corresponding task allocation results; constructing a question and answer content, fine-tuning an existing pre-training model based on the question and answer content to obtain a fine-tuned model; updating parameters of the fine-tuned model based on a plurality of output answers recording task allocation results to obtain a strategy model; establishing an environment modeling expert, a task allocation expert and a scheme evaluation expert based on the strategy model; for a to-be-processed task environment, processing the to-be-processed task environment through a gating module and the environment modeling expert, the task allocation expert and the scheme evaluation expert to obtain a task allocation result of the to-be-processed task environment; and the present application can improve the efficiency and quality of land-air unmanned system task allocation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of land-air unmanned system task allocation, in particular to an intelligent task allocation method for land-air unmanned system. BACKGROUND

[0002] Civil unmanned aerial vehicles are widely used in low-altitude economic scenarios, including logistics distribution, infrastructure inspection, aerial photography and mapping, emergency search and rescue, and communication relay, due to their ease of operation and environmental adaptability. However, in long-term community monitoring and continuous express delivery scenarios that require long-time or phased completion, the traditional single-machine operation mode is difficult to meet the actual needs due to the hard constraints of endurance time and flight distance.

[0003] To break through this technical bottleneck, a land-air collaborative scheme has been favored by the industry: by deploying ground base stations and unmanned aerial vehicles to form a collaborative network, energy dynamic replenishment and task relay execution are achieved. In the scenario of multi-machine rotation operation, how to scientifically plan the task timing and spatial distribution of each unmanned platform is the core problem of improving the overall efficiency of the system. Since multi-machine collaborative task allocation is a typical NP-hard problem, efficient algorithms are needed to obtain a near-optimal solution.

[0004] Existing task allocation research mainly focuses on task modeling and solution strategies, gradually incorporating time windows, environmental uncertainties, and energy constraints into model design, and introducing various optimization mechanisms in multi-unmanned aerial vehicle collaboration and ground base station location selection scenarios, but the energy recovery modeling in multiple rounds of tasks is still insufficient. In terms of algorithms, heuristic and deep learning methods have been applied, and hybrid algorithms have shown potential, but they mainly focus on single task allocation and lack multi-round correlation and historical feedback mechanisms. Land-air collaborative tasks generally use hierarchical solution and two-stage framework, but real-time calculation efficiency is insufficient in long-period tasks. Some work has focused on task re-allocation under environmental changes, but there is still room for improvement in multi-machine heterogeneous collaboration and multi-objective optimization guided by fusion learning.

[0005] The existing land-air unmanned system collaborative shift mode provides an effective solution for long-term unmanned surveillance tasks, but in actual deployment, key elements such as endurance capability limitations, dynamic obstacle avoidance requirements, and real-time re-planning capabilities still need to be considered. For the problem of multi-round task allocation, systematically integrating dynamic environment modeling, hierarchical optimization algorithms, and efficient solution mechanisms is an important direction to improve the operational efficiency of unmanned aerial vehicle clusters in low-altitude economic scenarios. SUMMARY

[0006] In view of the above problems, the present application provides an intelligent task allocation method for land-air unmanned system, which solves the technical problem of insufficient task allocation efficiency and quality of land-air unmanned system in the prior art.

[0007] The application provides an intelligent task allocation method for a land-air unmanned system, comprising the following steps:

[0008] Step S1, a plurality of task environments are randomly generated, each task environment is pre-allocated, a task allocation result is obtained, and a heuristic task allocation database is established based on the plurality of task environments and the corresponding task allocation results;

[0009] Step S2, a question and answer content is constructed using the task environment and the task allocation result in the heuristic task allocation database, an existing pre-trained model is fine-tuned based on the question and answer content, and a fine-tuned model is obtained;

[0010] Step S3, a plurality of output answers recording the task allocation result are obtained from the fine-tuned model, a target function is calculated based on each output answer, and the parameters of the fine-tuned model are iteratively updated by maximizing the target function to obtain a strategy model;

[0011] Step S4, an environment modeling expert, a task allocation expert and a scheme evaluation expert are established by the strategy model;

[0012] Step S5, for a to-be-processed task environment, the task allocation result of the to-be-processed task environment is obtained through the processing of the gating module and the environment modeling expert, the task allocation expert and the scheme evaluation expert.

[0013] Preferably, in step S1, the task environment includes the location and size of the no-fly zone and the no-go zone, the location of the unmanned aerial vehicle, the location of the ground base station and the location of the waypoint; the task allocation result includes the task path of the unmanned aerial vehicle, the task stay time list of the unmanned aerial vehicle, the task path of the ground base station and the task stay time list of the ground base station.

[0014] Preferably, step S2 comprises:

[0015] Step S2-1, structured data established by the task environment is used as the question content, and structured data established by the task allocation result is used as the answer content to construct a fine-tuning data set;

[0016] Step S2-2, load the existing pre-trained model and set the LoRA parameter;

[0017] Step S2-3, fine-tune the existing pre-trained model based on the fine-tuning data set and the LoRA parameter to obtain a fine-tuned model.

[0018] Preferably, step S2-1 specifically comprises:

[0019] Question and answer pairs in the form of "user":"problem" and "assistant":"answer" are constructed, including:

[0020] The no-fly zone and no-go zone position and size in the task environment in the heuristic task allocation database, the UAV, the waypoint and the ground base station position are converted into json format data as the problem content;

[0021] The task path of the UAV in the task allocation result in the heuristic task allocation database, the task stay time list of the UAV, the task path of the ground base station, the task stay time list of the ground base station are converted into json format data and added to the problem content to form the answer content;

[0022] A plurality of question and answer pairs are constructed, and the set of all question and answer pairs is taken as the fine-tuning data set.

[0023] Preferably, step S3 comprises:

[0024] Step S3-1, input the question content in the fine-tuning data set into the fine-tuning model, and obtain a plurality of output answers recording the task allocation result;

[0025] Step S3-2, calculate the objective function value according to the task duration of each output answer corresponding to the UAV; update the parameters of the fine-tuning model based on the objective function value;

[0026] Step S3-3, return to step S3-1 until the objective function value is greater than a preset threshold, and the training is completed.

[0027] Preferably, in step S3-2, the calculation expression of the objective function value is:

[0028]

[0029]

[0030]

[0031] wherein, denotes the UAV numbered , is the waypoint sequence visited by the UAV , is the th waypoint in the waypoint sequence , denotes the total number of waypoints in the waypoint sequence, is the duration of the task performed by the UAV , is the path distance between the waypoints of the UAV, is the flight speed of the UAV , is the flight speed of the UAV at the waypoint time of charging, for the objective function, representing the calculation of the maximum value.

[0032] Preferably, step S4 specifically comprises:

[0033] The strategy model is respectively given corresponding prompt words for environment modeling, task allocation and scheme evaluation, and environment modeling experts, task allocation experts and scheme evaluation experts are constructed, wherein:

[0034] The prompt words of the environment modeling expert are used to analyze the waypoint positions and obstacle information, and generate a weight matrix of the distances between the waypoints;

[0035] The prompt words of the task allocation expert are used to analyze the weight matrix of the distances between the waypoints, analyze the number and performance of the ground base stations and the unmanned aerial vehicles, and generate or modify the task paths and stay times of the unmanned aerial vehicles and the ground base stations on the basis of the original, so that the total task time is as short as possible;

[0036] The prompt words of the scheme evaluation expert are used to exchange or change the task paths and stay times of the unmanned aerial vehicles and the ground base stations, and determine whether a shorter total task time is obtained.

[0037] Preferably, step S5 specifically comprises:

[0038] Step S5-1, the gating module sends the input question to the environment modeling expert, the environment modeling expert generates a weight matrix, synthesizes the weight matrix with the input question, obtains the synthesized instruction and inputs the gating module again;

[0039] Step S5-2, the gating module sends the synthesized instruction to the task allocation expert, generates a task allocation result including the task paths and stay times of the unmanned aerial vehicles and the ground base stations, and sends the task allocation result to the scheme evaluation expert;

[0040] Step S5-3, the scheme evaluation expert determines whether the task allocation result reaches an ideal situation, if not, returns to step S5-2, and if so, outputs the task allocation result.

[0041] Preferably, in step S2-2, the setting of the LoRA parameters comprises:

[0042] Setting the rank of the LoRA fine-tuning, the fine-tuning target module, the scaling factor of the LoRA fine-tuning, the sparsity regularization parameter, the trigger bias, the use of the gradient checkpoint flag, the random number seed, the use of the rslora flag, and the loftq setting flag;

[0043] The process of fine-tuning the existing pre-trained model in step S2-3 includes setting fine-tuning parameters, including: training batch size, accumulated gradient step number, learning rate warm-up step number, learning rate, training round, whether to use 16-bit floating point, training log recording step interval, optimizer type, weight decay coefficient, learning rate scheduler type, model output saving directory, whether to remove unused columns in the dataset, field name of the text data in the dataset, whether to skip automatic preprocessing of the dataset, specified number of parallel processes for data loading and preprocessing, and maximum sequence length of model input.

[0044] Preferably, step S2 further includes:

[0045] Step S2-4, testing whether the fine-tuned model meets the test requirements, specifically includes:

[0046] Setting parameters of the fine-tuned model, including setting: generated text content form, maximum token number, whether to use cache, generation diversity, and minimum probability threshold of generated text.

[0047] Inputting the question content in the fine-tuning dataset in step S2-1 into the fine-tuned model to obtain the answer content, and evaluating the quality of the test answer content by manual reading judgment or automatic index evaluation to determine whether the fine-tuned model meets the test requirements.

[0048] Compared with the prior art, the present application has at least the following beneficial effects:

[0049] (1) The present application constructs a structured and high-quality heuristic task allocation database by randomly generating multiple task environments and using existing task allocation algorithms to obtain high-quality allocation results. This database covers a variety of task environments and allocation schemes, and can enrich the training dataset, providing data basis for the effective learning and generalization ability improvement of subsequent models.

[0050] (2) The present application uses the heuristic task allocation database to fine-tune the existing pre-trained model, and also updates the model parameters through reward mechanism and objective function optimization iterative iteration. This training method fully integrates expert experience and data-driven methods, so that the final obtained strategy model can realize more efficient and intelligent scheme output in complex and variable land-air unmanned system task allocation scenarios, effectively improving the quality and efficiency of task allocation.

[0051] (3) The application assigns the strategy model to three roles of environment modeling experts, task allocation experts and scheme evaluation experts respectively, and realizes the division of labor and cooperation among the multiple experts through the efficient scheduling of the gate module. The intelligent experts fully exert their respective field expertise and collectively participate in the task allocation decision, effectively improving the scientific nature of the system decision and the robustness of the task completion, solving the problem of limited decision-making ability of a single expert, and significantly enhancing the comprehensive performance of the system. BRIEF DESCRIPTION OF DRAWINGS

[0052] The accompanying drawings are included to provide a further understanding of the application, and are incorporated in and constitute apart of this specification. The drawings illustrate embodiments of the application and, together with the description, serve to explain the principles of the application.

[0053] Figure 1 The intelligent task allocation method flow chart of the land-air unmanned system provided by the application.

[0054] Figure 2 The land-air cooperative task allocation scene schematic diagram provided by the application.

[0055] Figure 3 The land-air cooperative heuristic task allocation flowchart

[0056] Figure 4 The multi-expert collaborative large model architecture schematic diagram for intelligent task allocation provided by the application. DETAILED DESCRIPTION

[0057] In order to more clearly understand the above-mentioned purposes, features and advantages of the application, the application will be further described in detail below in combination with the drawings and specific embodiments. It should be noted that the embodiments of the application and the features in the embodiments can be combined with each other without conflict. In addition, the application can also be implemented in other ways different from those described herein, therefore, the protection scope of the application is not limited by the specific embodiments disclosed below.

[0058] The land-air unmanned system intelligent task allocation method provided by the application is based on large language model fine-tuning training, and the intelligent task allocation of the system in a complex dynamic environment is greatly enhanced through low-rank adaptive fine-tuning (Low-Rank Adaptation, LoRA), group relative policy optimization (Group Relative Policy Optimization, GRPO), and reinforcement learning and multi-expert collaborative architecture. The establishment of the heuristic task allocation database provides high-quality simulation data support for large model training and decision-making. On the premise of retaining the original system data basis and decision logic, precise modeling and dynamic optimization of various collaborative task scenarios are realized through deep reinforcement learning, which greatly improves the allocation efficiency and resource utilization, thereby significantly making up for the performance bottleneck of the traditional collaborative system in processing high-dimensional and variable task environments, and effectively realizing the intelligent collaborative operation of land-air integration.

[0059] To illustrate the effectiveness of the method of the application, the above technical solutions of the application are described in detail below through a specific embodiment. As shown in Figure 1 , a land-air unmanned system intelligent task allocation method is disclosed, and the specific implementation steps are as follows:

[0060] Step S1, a plurality of task environments are randomly generated, each task environment is pre-allocated, a task allocation result is obtained, and a heuristic task allocation database is established based on the plurality of task environments and the corresponding task allocation results.

[0061] The task environment includes the location and size of the no-fly zone and the no-go zone, the location of the unmanned aerial vehicle, the waypoint, and the ground base station; the task allocation result includes the task path of the unmanned aerial vehicle, the task stay time list of the unmanned aerial vehicle, the task path of the ground base station, and the task stay time list of the ground base station.

[0062] The land-air collaborative task of the application refers to the deployment of a ground base station and an unmanned aerial vehicle to form a collaborative network in processes such as long-term community monitoring and continuous express transportation, the unmanned aerial vehicle can land on the ground base station, move with the ground base station, realize dynamic energy replenishment and task relay execution, and complete multi-machine rotation operation.

[0063] As shown in Figure 2 , a map of the land-air collaborative task environment is shown, and the map is divided into a plurality of grids. In some embodiments, each grid can have coordinates for positioning the grid position.

[0064] The map contains a no-fly zone, a no-drive zone, a UAV and a ground base station. The no-fly zone represents an area where the UAV is not allowed to fly. The no-drive zone represents an area where the ground base station is not allowed to drive. The map can contain multiple UAVs and ground base stations for performing tasks. Each ground base station can drive on the ground to transport the UAV.

[0065] The map also includes waypoints, which are points that the UAV and the ground base station need to pass through. In some embodiments, the waypoints can be determined by an optimization algorithm based on the locations of the no-fly zone, the no-drive zone, etc. in the map to obtain the optimal waypoint positions, which are then recorded in the task environment map.

[0066] In this step, multiple task environments can be randomly generated, each of which can contain different map sizes, grid numbers, locations and sizes of no-fly zones and no-drive zones, numbers and locations of UAVs and ground base stations, etc.

[0067] For the randomly generated task environments, the present application can use manual annotation or existing task allocation algorithms to pre-allocate each task environment and obtain the task allocation results. For example, Figure 3 The above is an existing task allocation algorithm, a land-air collaborative heuristic task allocation method.

[0068] In some embodiments, the land-air collaborative heuristic task allocation method can construct constraints based on the no-fly zone and the no-drive zone, plan the motion path of the UAV and the ground base station, calculate the transfer path between the waypoints using the RRT algorithm, perform task allocation based on the transfer path and the waypoints, and the task allocation algorithm can include: using a peak clustering algorithm to spatially cluster the waypoints to form independent sub-regional clusters, and introducing an improved genetic algorithm for collaborative optimization to minimize the weighted amount of the total task time and the interval between the end times of each group of tasks as the collaborative optimization objective.

[0069] Finally, the task allocation algorithm obtains the task path of the UAV corresponding to the current task environment, the task stay time list of the UAV, the task path of the ground base station, and the task stay time list of the ground base station as the task allocation result. The task path of the UAV records the sequence of the UAV visiting the waypoints obtained by optimization, which is the path sequence of the task executed by the UAV. The task stay time list of the UAV records the time the UAV stays during the execution of the task when visiting the waypoints. The task path of the ground base station records the sequence of the ground base station visiting the waypoints obtained by optimization, and the task stay time of the ground base station records the time the ground base station stays during the execution of the task when visiting the waypoints.

[0070] The present application collects pairs of randomly generated different task environments and corresponding task allocation results to form a heuristic task allocation database, which serves as a data set for subsequent training of a large language model.

[0071] Step S2, constructing a question and answer content by using the task environment and task allocation result in the heuristic task allocation database, fine-tuning an existing pre-trained model by the question and answer content, and obtaining a fine-tuned model;

[0072] The process of fine-tuning the existing pre-trained model in the application adopts the LoRA method. LoRA is a parameter-efficient fine-tuning method, and the principle of LoRA is inspired by the concept of intrinsic dimension. That is, when a pre-trained model adapts to a specific task, the over-parameterized model actually has a lower intrinsic dimension. The core idea is to freeze the parameters of the pre-trained model, introduce two low-rank matrices through low-rank decomposition, and indirectly simulate the effect of full-parameter fine-tuning through optimization of the low-rank matrices.

[0073] LoRA will insert a low-rank matrix at a specific location in the pre-trained model (such as a self-attention module or a feedforward network) to form a "bypass". Only a small number of parameters in the low-rank matrix need to be updated during the training phase. After training is completed, the low-rank matrices are multiplied and combined with the pre-trained model parameters to serve as the fine-tuned model parameters. In some embodiments, the unsloth fine-tuning tool can be used for LoRA fine-tuning.

[0074] The fine-tuning process of the application includes three parts: preparing a fine-tuning data set, loading a pre-trained model and setting LoRA parameters, and fine-tuning a model, which are described in detail as follows:

[0075] Step S2-1, structured data established by the task environment as the question content, and structured data established by the task allocation result as the answer content, to construct a fine-tuning data set;

[0076] Any group of data in the fine-tuning data set of the application can be composed of a question and answer pair in the form of "user":"problem" and "assistant":"answer", wherein the content of the problem and answer question and answer pair adopts structured data as the content.

[0077] In some embodiments, the structured data is json format data, which records the task environment and the corresponding task allocation result information. A typical json format data recording the task environment and the corresponding task allocation result information is shown in Table 1:

[0078] Table 1

[0079]

[0080] The json format data records the task environment and the corresponding task allocation result, wherein the task environment information is as follows:

[0081] NavPoints represents the individual waypoints, type is list, each element in the list contains two values, as [x, y] coordinates, representing a waypoint in the environment.

[0082] Obs represents the no-fly zone and no-go zone information, type is list, each element in the list represents an obstacle (i.e. no-fly zone or no-go zone), each element contains three values, as [x, y] coordinates and obstacle radius.

[0083] UGV represents the ground base station, type is list, each element in the list contains the position coordinates of the ground base station; in some embodiments, each element in the list can also contain the speed of the ground base station.

[0084] UAV represents the UAV, type is list, each element in the list contains the position coordinates of the UAV; in some embodiments, each element in the list can also contain the speed and range of the UAV.

[0085] The task allocation result information is as follows:

[0086] Seq_UGV represents the task path of the ground base station, type is list, each element of the list represents a sequence of access waypoints of a ground base station, in the example, there are two sequences of UGVs, the first UGV sequence is from waypoint 1 to waypoint 2; the second UGV sequence is from waypoint 1 to waypoint 4 and then to waypoint 1.

[0087] Seq_UAV represents the task path of the UAV, type is list, each element of the list represents a sequence of access waypoints of a UAV, in the example, there are three UAVs visiting waypoints in the sequence, for example, the first one is visiting waypoints in the order of waypoint 1, waypoint 3, waypoint 2, waypoint 5, and waypoint 1.

[0088] Mis_UGV represents the task stay time list of the ground base station, each element of the list corresponds to a stay time sequence of a ground base station. The numbers here are the stay time or task execution time of the ground base station at the corresponding waypoint in its sequence. In the example, there are two UGVs with stay time sequences, the first UGV has a stay time of 0 at the first waypoint in its sequence and a stay time of 0.3 at the second waypoint.

[0089] Mis_UAV represents the task stay time list of the UAV, each element of the list corresponds to a stay time sequence of a UAV. In the example, there are 3 UAVs with stay time sequences, the first UAV has stay times of 0, 0, 0.2, 0, and 0 at the waypoints in its sequence.

[0090] Some intermediate data information is also recorded in the json format data, which can be modified by the pre-trained model.

[0091] MapEnv represents the weight matrix of the environment model. The transition overheads such as time or energy consumption between waypoints are recorded. In some embodiments, it can be set to null, or generated or modified by a pre-trained model.

[0092] Obj_FUN represents the value of the optimization objective function. In some embodiments, it can be set to null, or generated or modified by a pre-trained model.

[0093] flag represents the state of the current task allocation scheme or the control signal of the optimization process, which can be generated or modified by a pre-trained model.

[0094] The above describes the structured data format of the question and answer in the answer content in detail. When constructing the question and answer, the no-fly zone and no-go zone position and size in the task environment in the heuristic task allocation database, the positions of the unmanned aerial vehicle, the waypoint and the ground base station can be converted into the NavPoints, Obs, UGV, UAV items in the structured data, and other items can be set to null as the problem content in "user":"problem".

[0095] On the basis of the problem content, the task path of the unmanned aerial vehicle, the task stay time list of the unmanned aerial vehicle, the task path of the ground base station, and the task stay time list of the ground base station in the task allocation result in the heuristic task allocation database can be converted into Seq_UGV, Seq_UAV, Mis_UGV, Mis_UAV in the structured data and added to the problem content to form the answer content in "assistant":"answer". Finally, a set of question and answer pairs are formed.

[0096] The set of all question and answer pairs is used as the fine-tuning data set for subsequent model fine-tuning.

[0097] Step S2-2, load the existing pre-trained model and set the LoRA parameter;

[0098] The present application selects an existing pre-trained model as the fine-tuning object. In some embodiments, models such as Qwen2.5-VL-7B-Insrtuct can be used. There are various pre-trained models in the prior art, which can be selected as needed. The present application does not limit the specific pre-trained model used.

[0099] In some embodiments, the pre-trained model parameters can be set, including the context window length of the model, the parameter type of the model, and whether to use 4-bit quantization for fine-tuning.

[0100] In some embodiments, the LoRA parameters can be set, including: the rank of LoRA fine-tuning, the fine-tuning target module, the scaling factor of LoRA fine-tuning, the sparsity regularization parameter, the trigger bias, the use of gradient checkpoint markers, the random number seed, the use of rslora markers, loftq setting markers, etc.

[0101] The specific settings of the pre-trained model parameters and the LoRA parameters are shown in Table 2:

[0102] Table 2

[0103]

[0104] Step S2-3, fine-tuning the existing pre-trained model based on the fine-tuning dataset and the LoRA parameters to obtain a fine-tuned model;

[0105] In some embodiments, the fine-tuning parameters can be set, including: the training batch size, the number of accumulated gradient steps, the learning rate warm-up step number, the learning rate, the training round, whether to use 16-bit floating point numbers, the step interval for recording training logs, the optimizer type, the weight decay coefficient, the learning rate scheduler type, the saving directory of model output (such as checkpoints, logs), whether to remove unused columns in the dataset, the field name of the text data in the dataset, whether to skip automatic preprocessing of the dataset, the number of parallel processes specified for data loading and preprocessing, the maximum sequence length of the model input, etc.

[0106] The specific settings of the fine-tuning parameters are shown in Table 3:

[0107] Table 3

[0108]

[0109] In some embodiments, step S2 can include step S2-4, testing whether the fine-tuned model meets the test requirements.

[0110] After fine-tuning is completed, the fine-tuned model is tested, and the parameters of the fine-tuned model are set, including: the generated text content form, the maximum number of tokens, whether to use caching, the generation diversity, the minimum probability threshold of the generated text, etc.

[0111] After setting the parameters of the fine-tuned model, the question content in step S2-1 is input into the fine-tuned model to obtain the answer content. The test answer content can be evaluated and tested using manual reading and judgment or automatic indicators to determine whether the fine-tuned model meets the test requirements. If the test requirements are not met, the parameters can be adjusted and fine-tuned again.

[0112] Step S3, obtaining a plurality of output answers of record task allocation results from the fine-tuning model, calculating a target function based on each output answer, and updating the parameters of the fine-tuning model by maximizing the target function to obtain a strategy model;

[0113] The fine-tuning model is trained by the GRPO method for reinforcement learning, and the GRPO method can update the model parameters by relatively comparing a group of model outputs without a separate critic network.

[0114] The specific steps include:

[0115] Step S3-1, inputting the question content in the fine-tuning dataset into the fine-tuning model to obtain a plurality of output answers of record task allocation results;

[0116] In some embodiments, 8 candidate answers are generated for each question content, which are used for subsequent calculation of the target function.

[0117] Step S3-2, calculating a target function value according to the unmanned aerial vehicle task duration corresponding to each output answer, and updating the parameters of the fine-tuning model based on the target function value;

[0118] In some embodiments, the target function value can be calculated from the unmanned aerial vehicle task duration calculated according to the waypoint sequence visited by the unmanned aerial vehicle in the candidate answer.

[0119] (Formula 1)

[0120] (Formula 2)

[0121] (Formula 3)

[0122] The unmanned aerial vehicle numbered is referred to as unmanned aerial vehicle . is the waypoint sequence visited by the unmanned aerial vehicle , is the i-th waypoint in the waypoint sequence , denotes the total number of waypoints in the waypoint sequence, is the duration of the task performed by the unmanned aerial vehicle , is the path distance of the unmanned aerial vehicle between waypoints, is the flight speed of the unmanned aerial vehicle , is the flight speed of the unmanned aerial vehicle , is the flight speed of the unmanned aerial vehicle The time of charging, formula 2 gives the calculation method of the task time of each UAV, that is, the task time of the UAV is the time of the UAV transferring between the waypoints plus the time of the UAV charging at the waypoints. For the target function, formula 3 gives the calculation method of the target function, that is, the maximum value of the task time of all UAVs.

[0123] Step S3-3, return to step S3-1 until the target function value is greater than the preset threshold, and the training is completed.

[0124] Step S4, establishing an environment modeling expert, a task allocation expert and a scheme evaluation expert by the policy model;

[0125] After obtaining the policy model, the present application designs corresponding prompt words for environment modeling, task allocation and scheme evaluation, and constructs an environment modeling expert, a task allocation expert and a scheme evaluation expert.

[0126] Among them, the prompt words of the environment modeling expert are used to analyze the waypoint positions and obstacle information, and generate a weight matrix of the distances between the waypoints;

[0127] The prompt words of the task allocation expert are used to analyze the weight matrix of the distances between the waypoints, analyze the number and performance of the ground base stations and the UAVs, and generate or modify the task paths and the staying time of the UAVs and the ground base stations on the basis of the original, so that the total task time is as short as possible.

[0128] The prompt words of the scheme evaluation expert are used to exchange or change the task paths and the staying time of the UAVs and the ground base stations, and judge whether a shorter total task time is obtained.

[0129] Table 4, table 5 and table 6 respectively show the prompt words of the environment modeling expert, the task allocation expert and the scheme evaluation expert.

[0130] Table 4

[0131]

[0132] Table 5

[0133]

[0134] Table 6

[0135]

[0136] In the above manner, the present application sets the optimized policy model as three kinds of experts, that is, an environment modeling expert, a task allocation expert and a scheme evaluation expert, each of which has a corresponding expertise and can play a role in cooperation to improve the quality of scheme generation.

[0137] Step S5, for the task environment to be processed, through the processing of the gating module and the environment modeling expert, the task allocation expert and the scheme evaluation expert, the task allocation result of the task environment to be processed is obtained.

[0138] As shown in the figure, for the problem of task allocation for the task environment to be processed, the application establishes a multi-expert collaborative large model architecture to process it. The specific description is as follows: Figure 4

[0139] Step S5-1, the gating module sends the input problem to the environment modeling expert, the environment modeling expert generates a weight matrix, and the weight matrix is combined with the input problem to obtain a synthesized instruction and input the gating module again;

[0140] Step S5-2, the gating module sends the synthesized instruction to the task allocation expert, generates a task allocation result including the task path of the unmanned aerial vehicle and the ground base station and the stay time, and sends it to the scheme evaluation expert;

[0141] Step S5-3, the scheme evaluation expert judges whether the task allocation result reaches the ideal situation, if not, returns to step S5-2, and if so, outputs the task allocation result.

[0142] The intelligent task allocation method of the land-air unmanned system provided by the application can realize efficiency order of magnitude improvement in typical urban low-altitude scenes (such as logistics distribution and community monitoring) compared with the traditional single machine shift mode. Through simulation testing and entity system demonstration, the significant advantages of the algorithm in global optimization capability are verified. The method deeply integrates the endurance constraint modeling, dynamic task shift and adaptive allocation strategy, and provides a complete system solution for long-term unmanned operation in the low-altitude economic scene.

[0143] Although the detailed implementation of the present application describes each action or step in a specific order, it should be understood that such actions or steps are required to be performed in the specific order shown or in a sequential order, or all the illustrated actions or steps should be performed to achieve the desired results. In certain circumstances, multitasking and parallel processing can be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be interpreted as limiting the scope of the present disclosure. Certain features described in the context of separate embodiments can also be combined in a single implementation. Conversely, various features described in the context of a single implementation can also be implemented separately or in any suitable subcombination. The above description is only the preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any changes or replacements within the technical scope disclosed by the present application can be easily thought of by those skilled in the art, which should be covered within the protection scope of the present application. ​

[0144] The above description is only the preferred embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. An intelligent task allocation method for land-air unmanned systems, characterized in that, The method comprises the following steps: Step S1, randomly generating a plurality of task environments, pre-allocating each task environment, obtaining a task allocation result, and establishing a heuristic task allocation database based on the plurality of task environments and the corresponding task allocation results; Step S2, constructing a question and answer content using the task environment and the task allocation result in the heuristic task allocation database, fine-tuning an existing pre-trained model based on the question and answer content, and obtaining a fine-tuned model; Step S2 comprises: Step S2-1, constructing a fine-tuning data set by using structured data established by the task environment as question content and structured data established by the task allocation result as answer content; Step S2-2, loading an existing pre-trained model and setting LoRA parameters; specifically including: Setting the pre-trained model parameters, including: setting the context window length of the model, loading the parameter type of the model according to the GPU, and selecting whether to use 4-bit quantization for fine-tuning according to the size of the video memory; The setting of the LoRA parameters specifically includes: Setting the rank of the LoRA fine-tuning to any one of 8, 16, 32, 64, and 128, setting the fine-tuning target module to ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj",], setting the scaling factor of the LoRA fine-tuning to 16, setting the sparsity regularization parameter to between 0 and 0.1, setting the trigger bias to "none", setting the use gradient checkpoint flag to "unsloth", setting the random number seed, setting the use rslora flag to False, and setting the loftq flag to None; Step S2-3, fine-tuning the existing pre-trained model based on the fine-tuning data set and the LoRA parameters to obtain a fine-tuned model; Step S2-1 specifically includes: Constructing a question and answer pair in the form of "user":"problem" and "assistant":"answer", including: Converting the location and size of the no-fly zone and the no-go zone in the task environment in the heuristic task allocation database, the location of the unmanned aerial vehicle, the location of the waypoint, and the location of the ground base station into json format data as the problem content; Converting the task path of the unmanned aerial vehicle, the task stay time list of the unmanned aerial vehicle, the task path of the ground base station, and the task stay time list of the ground base station in the task allocation result in the heuristic task allocation database into json format data and adding them to the problem content to form the answer content; Constructing a plurality of question and answer pairs, and taking the set of all question and answer pairs as the fine-tuning data set; Step S3, obtaining a plurality of output answers recording the task allocation result from the fine-tuned model, calculating an objective function based on each output answer, and updating the parameters of the fine-tuned model through maximizing the objective function to obtain a strategy model; The calculation expression of the objective function is: wherein, denotes a drone numbered , is a sequence of waypoints visited by the drone , is a sequence of waypoints , denotes the th waypoint in the sequence of waypoints, denotes a total number of waypoints in the sequence of waypoints, denotes a duration of time for which the drone travels along the path, denotes a flight speed of the drone , denotes a time for which the drone is charging at the waypoint , is an objective function, denotes computing a maximum value; Step S4, establishing an environment modeling expert, a task allocation expert, and a scheme evaluation expert based on the strategy model; The step S4 specifically comprises: The strategy model is respectively given corresponding prompt words for environment modeling, task allocation and scheme evaluation, and environment modeling experts, task allocation experts and scheme evaluation experts are constructed, wherein: The prompt words of the environment modeling expert are used to analyze the waypoint positions and obstacle information, and generate a weight matrix of the distances between the waypoints; The prompt words of the task allocation expert are used to analyze the weight matrix of the distances between the waypoints, analyze the number and performance of the ground base stations and the unmanned aerial vehicles, and generate or modify the task paths and stay times of the unmanned aerial vehicles and the ground base stations on the basis of the original ones, so that the total time of task completion is as short as possible; The prompt words of the scheme evaluation expert are used to exchange or change the task paths and stay times of the unmanned aerial vehicles and the ground base stations, and judge whether a shorter total task time is obtained; The step S5, for the task environment to be processed, the task allocation result of the task environment to be processed is obtained through the processing of the gating module and the environment modeling expert, the task allocation expert and the scheme evaluation expert.

2. The intelligent task allocation method for land-air unmanned system according to claim 1, wherein, In step S1, the task environment includes the positions and sizes of the no-fly zones and no-go zones, the positions of the unmanned aerial vehicles, the positions of the ground base stations and the positions of the waypoints; the task allocation result includes the task paths of the unmanned aerial vehicles, the task stay time list of the unmanned aerial vehicles, the task paths of the ground base stations and the task stay time list of the ground base stations.

3. The intelligent task assignment method for land-air unmanned system according to claim 2, wherein, The step S3 comprises: Step S3-1, input the question content in the fine-tuning data set into the fine-tuning model, and obtain a plurality of output answers recording the task allocation results; Step S3-2, calculate the objective function value according to the task time of each output answer; update the parameters of the fine-tuning model based on the objective function value; Step S3-3, return to step S3-1 until the objective function value is greater than a preset threshold, and the training is completed.

4. The intelligent task allocation method for land-air unmanned system according to claim 3, wherein, The step S5 specifically comprises: Step S5-1, the gating module sends the input question to the environment modeling expert, the environment modeling expert generates a weight matrix, and the weight matrix is combined with the input question to obtain a combined instruction and input the combined instruction to the gating module again; Step S5-2, the gating module sends the combined instruction to the task allocation expert, generates a task allocation result including the task paths and stay times of the unmanned aerial vehicles and the ground base stations, and sends the task allocation result to the scheme evaluation expert; Step S5-3, the scheme evaluation expert judges whether the task allocation result reaches an ideal situation, if not, returns to step S5-2, and if so, outputs the task allocation result.

5. The intelligent task assignment method for land-air unmanned system according to claim 4, wherein, In step S2-2, the setting of the LoRA parameters comprises: Setting the rank of LoRA fine-tuning, the fine-tuning target module, the scaling factor of LoRA fine-tuning, the sparsity regularization parameter, the trigger bias, the gradient checkpoint marker, the random number seed, the rslora marker and the loftq setting marker; The process of fine-tuning the pre-trained model according to step S2-3 includes setting fine-tuning parameters, including: training batch size, accumulated gradient step number, learning rate warm-up step number, learning rate, training round, whether to use 16-bit floating point, training log recording step interval, optimizer type, weight decay coefficient, learning rate scheduler type, model output saving directory, whether to remove unused columns in the dataset, field name of the text data in the dataset, whether to skip automatic preprocessing of the dataset, specified number of parallel processes for data loading and preprocessing, and maximum sequence length of model input.

6. The intelligent task assignment method for land-air unmanned system according to claim 5, wherein, Step S2 further includes: Step S2-4, testing whether the fine-tuned model meets the test requirements, specifically including: Setting parameters of the fine-tuned model, including setting: generated text content form, maximum token number, whether to use cache, generation diversity, and minimum probability threshold of generated text; Input the question content in the fine-tuning dataset in step S2-1 into the fine-tuning model to obtain the answer content, and evaluate the quality of the test answer content by artificial reading judgment or automatic index to determine whether the fine-tuned model meets the test requirements.

Images