CN-122022086-A - Unmanned plane cluster collaborative task planning method and system based on large language model and environment feedback mechanism

CN122022086ACN 122022086 ACN122022086 ACN 122022086ACN-122022086-A

Abstract

The application discloses an unmanned aerial vehicle cluster collaborative task planning method and system based on a large language model and an environment feedback mechanism, wherein the unmanned aerial vehicle cluster collaborative task planning method comprises the steps of carrying out knowledge extraction and text segmentation based on an unmanned aerial vehicle combat expert knowledge base to obtain an expert knowledge text fragment set; the method comprises the steps of carrying out semantic coding and similarity screening on a text fragment set according to expert knowledge to construct strategy prompt words, inputting the strategy prompt words into a large language model to generate an initial unmanned aerial vehicle cluster collaborative task planning strategy, carrying out unmanned aerial vehicle simulation deduction on the basis of the initial unmanned aerial vehicle cluster collaborative task planning strategy to obtain multi-section track data, constructing an external track knowledge base on the basis of the track data and screening a target track text on the basis of track freshness and text similarity, constructing strategy evolution prompt words on the basis of the target track text, historical dialogue information and preset prompt words, and inputting the strategy prompt words into the large language model to generate an optimized unmanned aerial vehicle cluster collaborative task planning strategy. The application can improve the adaptability, rationality and reliability of task planning.

Inventors

HAO YIXUE
HU LONG
LI XIANZHI
WANG RUI
YIN JIAJIE

Assignees

华中科技大学

Dates

Publication Date: 20260512
Application Date: 20251219

Claims (10)

1. The unmanned aerial vehicle cluster collaborative task planning method based on a large language model and an environment feedback mechanism is characterized by comprising the following steps of: Constructing an unmanned aerial vehicle fighter expert knowledge base, and carrying out knowledge extraction and text segmentation based on the unmanned aerial vehicle fighter expert knowledge base to obtain an expert knowledge text fragment set; carrying out semantic coding and similarity screening according to the expert knowledge text fragment set, and constructing strategy prompt words; inputting the strategy prompt word into a large language model for strategy reasoning, and generating an initial unmanned aerial vehicle cluster collaborative task planning strategy; Performing unmanned aerial vehicle simulation deduction based on the initial unmanned aerial vehicle cluster collaborative task planning strategy to obtain multi-section track data; Constructing an external track knowledge base based on the track data, and screening track texts in the external track knowledge base based on track freshness and text similarity to obtain target track texts; And constructing a strategy evolution prompt word based on the target track text, the historical dialogue information of the large language model and a preset prompt word, inputting the strategy evolution prompt word into the large language model again for strategy optimization, and generating an optimized unmanned plane cluster collaborative task planning strategy.
2. The method of claim 1, wherein the constructing the unmanned aerial vehicle combat expert knowledge base, performing knowledge extraction and text segmentation based on the unmanned aerial vehicle combat expert knowledge base, to obtain a set of expert knowledge text segments, comprises: Collecting external unmanned aerial vehicle cluster combat data, and constructing an unmanned aerial vehicle combat expert knowledge base based on the external unmanned aerial vehicle cluster combat data; Carrying out text knowledge extraction on the data in the unmanned aerial vehicle combat expert knowledge base according to the editing type to obtain an original expert knowledge text; And recursively dividing the original expert knowledge text according to the priority segmenter, the maximum block length and the overlap length by using a recursion character text splitter to obtain an expert knowledge text fragment set.
3. The method of claim 1, wherein the performing semantic coding and similarity screening according to the expert knowledge text segment set to construct a policy hint word comprises: Acquiring simulation environment information, and constructing an initial prompt word based on the simulation environment information and a thinking chain structure template; Encoding the initial prompt word and the expert knowledge text segment set through a text encoder respectively to obtain a prompt vector and an expert knowledge text vector set; Calculating cosine similarity of each expert knowledge text vector in the prompt vector and the expert knowledge text vector set, and sorting the expert knowledge text vectors in the expert knowledge text vector set in a descending order according to the cosine similarity to obtain a sorted expert knowledge text vector set; taking expert knowledge text fragments corresponding to the expert knowledge text vectors of the preset quantity in the sequenced expert knowledge text vector set as target expert knowledge text fragments; and splicing the target expert knowledge text segment with the initial prompt word to obtain a strategy prompt word.
4. The method of claim 1, wherein the inputting the policy hint words into a large language model for policy reasoning generates an initial unmanned cluster collaborative mission planning policy, comprising: setting the role of a large language model as an unmanned aerial vehicle cluster commander, and loading simulation environment information in the strategy prompt words, wherein the simulation environment information comprises the number and the type of unmanned aerial vehicles on the my side, the number and the position of enemy defense units, a three-dimensional battlefield space boundary and a task target; The large language model is guided to analyze battlefield situations based on the simulation environment information, and an environment analysis result is output, wherein the environment analysis result comprises the relative distance between each unmanned aerial vehicle and each enemy target, the functional characteristic difference of the unmanned aerial vehicle and the current actionable area constraint; Guiding the large language model to allocate an enemy target index to each unmanned aerial vehicle based on the environmental analysis result to form a target allocation mapping table, wherein the same enemy target can be attacked by multiple unmanned aerial vehicles in a cooperative manner; Guiding the large language model to output corresponding discrete task actions according to each pair of the unmanned aerial vehicle-enemy target allocation pairs in the target allocation mapping table and the type and the relative position of the unmanned aerial vehicle, wherein the discrete task actions are selected from a preset action set; and generating an initial unmanned aerial vehicle cluster collaborative task planning strategy according to the enemy target indexes of all unmanned aerial vehicles and the corresponding discrete task actions.
5. The method of claim 1, wherein the performing unmanned aerial vehicle simulation deduction based on the initial unmanned aerial vehicle cluster collaborative mission planning strategy to obtain multi-segment trajectory data comprises: mapping discrete task actions in the initial unmanned aerial vehicle cluster collaborative task planning strategy into continuous physical control quantities executable by an unmanned aerial vehicle, and converting the continuous physical control quantities into physical control instructions, wherein the discrete task actions at least comprise attack, hover and distance, and the continuous physical control quantities comprise yaw angle variation and pitch angle variation; Inputting the physical control instruction into a simulation environment for execution, and collecting first simulation data with a fixed sampling frequency in the execution process, wherein the first simulation data comprises a current simulation step length, battlefield environment observation data, unmanned aerial vehicle operation data and a task completion rate; When the occurrence of a key event is monitored, collecting second simulation data in a preset time window after the triggering moment of the key event, wherein the key event comprises an unmanned aerial vehicle matrix event or an enemy target destroyed event; and carrying out fusion processing on the first simulation data and the second simulation data to obtain multi-section track data.
6. The method of claim 5, wherein mapping discrete task actions in the initial drone cluster collaborative task planning strategy to continuous physical control quantities executable by a drone and converting the continuous physical control quantities to physical control instructions comprises: Analyzing the initial unmanned aerial vehicle cluster cooperative task planning strategy, extracting discrete task actions of each unmanned aerial vehicle in the initial unmanned aerial vehicle cluster cooperative task planning strategy, and acquiring enemy target indexes corresponding to the discrete task actions; based on the enemy target index, acquiring a first position coordinate of a corresponding enemy target in a current simulation step length from a simulation environment, and acquiring a second position coordinate of the unmanned aerial vehicle in the current simulation step length; according to the discrete task actions, the first position coordinates and the second position coordinates, calculating the expected yaw angle variation and the expected pitch angle variation of the unmanned aerial vehicle under the self coordinate system; and taking the expected yaw angle variation and the expected pitch angle variation as continuous physical control quantities executable by the unmanned aerial vehicle, and converting the continuous physical control quantities into physical control instructions through a preset control instruction conversion rule.
7. The method of claim 1, wherein the constructing an external track knowledge base based on the track data, and screening track text in the external track knowledge base based on track freshness and text similarity, to obtain target track text, comprises: Constructing an external track knowledge base based on the track data, and acquiring track freshness of each track data in the external track knowledge base, wherein the track freshness is determined according to iterative rounds generated by the corresponding track data; Sorting track data in the external track knowledge base in a descending order according to the track freshness to obtain a sorted track list; converting the track data in the ordered track list into a candidate track text described by a corresponding natural language to obtain a candidate track text list; Traversing the candidate track text list, and respectively calculating the text similarity between the track text in the candidate track text list and all track texts in the initial track list; under the condition that the text similarity reaches a preset similarity threshold and the number of track texts in the initial track list does not reach a preset number threshold, adding the candidate track texts into the initial track list to obtain an updated track list; and taking the track text in the updated track list as a target track text after the number of the track texts in the initial track list reaches a preset number threshold or the candidate track text list is traversed.
8. An unmanned aerial vehicle cluster collaborative mission planning system based on a large language model and an environment feedback mechanism, which is characterized by comprising: The segmentation module is used for constructing an unmanned aerial vehicle combat expert knowledge base, and carrying out knowledge extraction and text segmentation based on the unmanned aerial vehicle combat expert knowledge base to obtain an expert knowledge text segment set; the screening module is used for carrying out semantic coding and similarity screening according to the expert knowledge text fragment set and constructing strategy prompt words; The reasoning module is used for inputting the strategy prompt words into the large language model to carry out strategy reasoning and generating an initial unmanned aerial vehicle cluster collaborative task planning strategy; The deduction module is used for carrying out unmanned aerial vehicle simulation deduction based on the initial unmanned aerial vehicle cluster collaborative task planning strategy to obtain multi-section track data; The screening module is further used for constructing an external track knowledge base based on the track data, and screening track texts in the external track knowledge base based on track freshness and text similarity to obtain target track texts; The optimization module is used for constructing strategy evolution prompt words based on the target track text, the historical dialogue information of the large language model and preset prompt words, inputting the strategy evolution prompt words into the large language model again for strategy optimization, and generating an optimized unmanned plane cluster collaborative task planning strategy.
9. An unmanned aerial vehicle cluster collaborative mission planning apparatus based on a large language model and an environmental feedback mechanism, characterized in that the apparatus comprises a memory, a processor and a computer program stored on the memory and executable on the processor, the computer program being configured to implement the steps of the unmanned aerial vehicle cluster collaborative mission planning method based on a large language model and an environmental feedback mechanism as claimed in any one of claims 1 to 7.
10. A storage medium, wherein a large language model and environment feedback mechanism-based unmanned aerial vehicle cluster cooperative task planning program is stored on the storage medium, and the large language model and environment feedback mechanism-based unmanned aerial vehicle cluster cooperative task planning program is executed by a processor to realize the large language model and environment feedback mechanism-based unmanned aerial vehicle cluster cooperative task planning method according to any one of claims 1 to 7.

Description

Unmanned plane cluster collaborative task planning method and system based on large language model and environment feedback mechanism Technical Field The application relates to the technical field of unmanned aerial vehicle cluster control, in particular to an unmanned aerial vehicle cluster collaborative task planning method and system based on a large language model and an environment feedback mechanism. Background With the rapid development of unmanned aerial vehicle technology, unmanned aerial vehicle cluster collaborative combat has become an important form of modern battlefield. However, dynamic mission planning in complex battlefield environments is a very challenging problem. The traditional methods mainly depend on heuristic optimization algorithms (such as genetic algorithms, particle swarm algorithms and the like), and when the methods are used for processing heterogeneous unmanned aerial vehicle loads, multitask constraints and high dynamic environments, the problems of difficult modeling, weak adaptability, high calculation overhead, lack of interpretability and the like exist. In recent years, large language models exhibit powerful semantic understanding and logical reasoning capabilities. There are research attempts to introduce the method into the decision-making field, but the existing method is mostly dependent on one-time strategy generation of a large language model, lacks deep interaction and closed loop feedback with a specific simulation environment, is easy to generate a 'illusion' problem, and the generated strategy is difficult to map directly to a control instruction conforming to physical dynamics, so that 'decision-execution' is split. Disclosure of Invention The application mainly aims to provide an unmanned aerial vehicle cluster collaborative task planning method and system based on a large language model and an environment feedback mechanism, and aims to solve the technical problems of poor adaptability, disjoint decision and execution and lack of continuous evolution capability in unmanned aerial vehicle cluster task planning in the prior art. In order to achieve the above purpose, the present application provides an unmanned aerial vehicle cluster collaborative task planning method based on a large language model and an environmental feedback mechanism, the unmanned aerial vehicle cluster collaborative task planning method based on the large language model and the environmental feedback mechanism comprising: Constructing an unmanned aerial vehicle fighter expert knowledge base, and carrying out knowledge extraction and text segmentation based on the unmanned aerial vehicle fighter expert knowledge base to obtain an expert knowledge text fragment set; carrying out semantic coding and similarity screening according to the expert knowledge text fragment set, and constructing strategy prompt words; inputting the strategy prompt word into a large language model for strategy reasoning, and generating an initial unmanned aerial vehicle cluster collaborative task planning strategy; Performing unmanned aerial vehicle simulation deduction based on the initial unmanned aerial vehicle cluster collaborative task planning strategy to obtain multi-section track data; Constructing an external track knowledge base based on the track data, and screening track texts in the external track knowledge base based on track freshness and text similarity to obtain target track texts; And constructing a strategy evolution prompt word based on the target track text, the historical dialogue information of the large language model and a preset prompt word, inputting the strategy evolution prompt word into the large language model again for strategy optimization, and generating an optimized unmanned plane cluster collaborative task planning strategy. In addition, in order to achieve the above purpose, the application also provides an unmanned aerial vehicle cluster collaborative task planning system based on a large language model and an environment feedback mechanism, the unmanned aerial vehicle cluster collaborative task planning system based on the large language model and the environment feedback mechanism comprises: The segmentation module is used for constructing an unmanned aerial vehicle combat expert knowledge base, and carrying out knowledge extraction and text segmentation based on the unmanned aerial vehicle combat expert knowledge base to obtain an expert knowledge text segment set; the screening module is used for carrying out semantic coding and similarity screening according to the expert knowledge text fragment set and constructing strategy prompt words; The reasoning module is used for inputting the strategy prompt words into the large language model to carry out strategy reasoning and generating an initial unmanned aerial vehicle cluster collaborative task planning strategy; The deduction module is used for carrying out unmanned aerial vehicle simulation deduction based on the initial