Multi-scene dynamic adaptation unmanned aerial vehicle flight task planning system and method

Abstract

The application discloses a kind of multi-scene dynamic adaptation unmanned aerial vehicle flight task planning system and method, system includes scene feature acquisition module, scene classification identification module, task demand analysis module, dynamic adaptation engine, path planning module and task execution monitoring module.Dynamic adaptation engine utilizes scene-task mapping model, strategy library and meta-learner, according to the scene type and task demand identified in real time, online fine-tuning initial adaptation strategy, generates the planning configuration parameter including cost function weight coefficient and safety margin;Path planning module completes global and local planning based on the parameter.During flight, when it is monitored that environment changes exceed set threshold, trigger scene re-identification and strategy fine-tuning again, realize dynamic re-planning.The application enables unmanned aerial vehicle to face multiple types of scenes quickly adapt planning strategy, and autonomously adjust when environment mutates, significantly improves task adaptability and flight safety.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) mission planning technology, specifically to a UAV flight mission planning system and method that dynamically adapts to multiple scenarios. Background Technology

[0002] With the widespread application of drones in various scenarios such as logistics delivery, power line inspection, emergency rescue, and environmental monitoring, the requirements for autonomous flight mission planning capabilities are increasing. Different application scenarios exhibit significantly different environmental characteristics. For example, urban canyons have dense buildings and strong electromagnetic interference; forest areas have undulating terrain and repetitive textures; water areas have sparse reference points and are affected by wind and waves; and mountainous environments are accompanied by complex airflow and steep terrain. Most existing mission planning systems are designed with fixed path planning algorithms and preset, unchanging cost function weights for specific scenarios. When drones perform missions across different types of scenarios, or when the scenario changes drastically during the same flight, the previously set strategies often become unsuitable. Operators need to manually adjust parameters based on experience, resulting in slow response times and effectiveness dependent on personal judgment. Some existing solutions have introduced scene recognition functions, but these only involve simple switching of preset rules based on the identified scene labels, without fine-tuning the planning strategy according to specific mission requirements and environmental details. Therefore, it is difficult to achieve a balanced result that considers flight safety, energy efficiency, and mission completion. Furthermore, existing systems have limited adaptability to unforeseen or sudden changes in scenarios, which greatly restricts the reliability of UAVs operating independently in complex and dynamic environments.

[0003] To address this, a multi-scenario dynamically adaptable UAV flight mission planning system and method are proposed. Summary of the Invention

[0004] The present invention aims to solve the problems mentioned in the background art by providing a multi-scenario dynamically adaptable UAV flight mission planning system and method.

[0005] The specific technical solution is as follows:

[0006] A multi-scenario dynamically adaptable UAV flight mission planning system includes:

[0007] The scene feature acquisition module is used to acquire multi-dimensional scene feature data of the current mission environment of the UAV;

[0008] The scene classification and recognition module is used to match the multi-dimensional scene feature data with a preset scene knowledge graph to identify the scene type and the corresponding scene confidence level.

[0009] The task requirement parsing module is used to obtain task instructions and parse them to obtain the task type, task constraints, and task priority.

[0010] The dynamic adaptation engine includes a scene-task mapping model, a policy generator, a policy library, and a meta-learner. The scene-task mapping model is configured to retrieve an initial adaptation policy from the policy library based on the scene type, scene confidence, task type, and task constraints, and to fine-tune the initial adaptation policy online using the meta-learner to generate planning configuration parameters including path planning algorithm parameters, cost function weight coefficients, and safety margins. The policy generator is used to generate planning instructions based on the planning configuration parameters.

[0011] The path planning module is used to receive the planning instructions, perform global path planning and local trajectory generation based on the planning configuration parameters and the current environment map, and output the flight path;

[0012] The task execution monitoring module is used to control the UAV to fly along the flight path and monitor the flight status and environmental changes in real time. When a deviation event that meets the preset conditions is detected, the dynamic adaptation engine is triggered to perform dynamic replanning.

[0013] This solution utilizes scene feature acquisition and scene classification to automatically determine the type of the current flight environment and assess the reliability of the identification. The task requirement parsing module transforms task instructions into structured task requirements, enabling subsequent planning to fully understand the task objectives. The dynamic adaptation engine leverages the combined information of the scene and task, first retrieving a basic strategy suitable for the given situation from the strategy library, and then fine-tuning the basic strategy online through a meta-learner. This generates a set of planning parameters specifically adapted to the current environment and task, including the path search algorithm settings, the weighting of various costs, and the safety buffer distance. In this way, when calculating the flight path, the planner naturally considers conflicting objectives such as shortening flight distance, reducing energy consumption, and avoiding dangerous areas. Furthermore, the emphasis on these objectives varies depending on the scenario; for example, safety is prioritized in obstacle-rich areas, while efficiency can be prioritized in open areas. During flight, the mission execution monitoring module continuously monitors environmental changes. Once a significant change in the environment is detected, it immediately triggers the regeneration of planning parameters and the recalculation of the path, ensuring that the UAV can respond promptly to emergencies and avoid the failure of the original path or the risk of collision due to environmental changes. The entire system forms a closed loop from perceiving the environment, understanding the mission, adjusting the planning strategy, to execution and readjustment, significantly improving the UAV's adaptability and flight safety when performing multiple tasks in complex and changing environments.

[0014] In the aforementioned system, the dynamic adaptation engine further includes a scene re-identification submodule. When the task execution monitoring module detects that the amount of environmental change exceeds a set threshold, it generates a deviation event. After the scene re-identification submodule is triggered, it re-acquires the latest scene feature data from the scene feature acquisition module, and the scene classification and identification module re-identifies the scene type. The meta-learner then fine-tunes the adaptation strategy based on the new scene type and the current task state to update the planning configuration parameters.

[0015] The above approach addresses the issue of drastic environmental changes during flight, such as a sudden transition from an open area to a densely built-up area or encountering strong electromagnetic interference. In such cases, the previously identified scene types become insufficient for safe flight. The scene re-identification submodule re-collects environmental features and re-classifies the scene, accurately capturing the altered environmental characteristics. The meta-learner then fine-tunes the planning strategy based on the new scene type and the remaining tasks, ensuring that the weight parameters and safety distances used by the path planner immediately adapt to the new environment. This avoids energy waste and safety risks caused by using outdated strategies due to sudden environmental changes, ensuring the flight path always matches the latest environmental conditions and further enhancing the flexibility and reliability of mission execution.

[0016] In the aforementioned system, the environmental change includes at least one of the environmental obstacle density change rate, light intensity change rate, and electromagnetic interference intensity change rate; the set thresholds include: an obstacle density change rate threshold ranging from 15% to 30%, a light intensity change rate threshold ranging from 25% to 40%, and an electromagnetic interference intensity change rate threshold ranging from 3dB to 8dB.

[0017] The above scheme employs several environmental indicators that significantly impact flight safety and sensor performance—obstacle density, light intensity, and electromagnetic interference intensity—as monitoring targets. This allows the system to perceive environmental changes from multiple dimensions. By setting reasonable thresholds for each type of change, the system can distinguish between normal, minor fluctuations and significant environmental alterations requiring attention. Re-identification and re-planning are only triggered when the change exceeds the corresponding threshold. This prevents frequent re-planning due to minor disturbances, thus avoiding wasted computational resources and flight time, while ensuring that truly impactful scene changes are not missed, achieving a good balance between responsiveness and system stability.

[0018] In the aforementioned system, the scene classification and recognition module confirms the scene type when the scene confidence level is greater than or equal to a confidence threshold; the confidence threshold ranges from 0.7 to 0.98.

[0019] The above approach sets a scene confidence threshold, meaning the system only uses a scene as a definitive category when it has sufficient confidence in the recognition result. If the similarity between the current environment features and various prototypes in the knowledge graph is insufficient, and the confidence level is below the threshold, the system will treat it as an unfamiliar or ambiguous scene, thus adopting a more conservative or general basic strategy rather than forcibly applying a pre-defined strategy for a specific scene. This effectively reduces the risk of strategy mismatch caused by scene misjudgment and improves the robustness of the entire adaptation process, making it particularly suitable for situations with many transitional regions or mixed features in the real-world environment.

[0020] In the above system, the cost function weight coefficients include path length weight α, energy consumption weight β, and safety weight γ, where α∈[0.2,0.8], β∈[0.1,0.6], γ∈[0.1,0.7], and satisfy α+β+γ=1.

[0021] The above scheme unifies the three core objectives—path length, energy consumption, and flight safety—into a single cost function through weighted coefficients, requiring the sum of these weights to be fixed at 1. This allows for continuous adjustment of the emphasis placed on each objective during planning, preventing excessive bias or complete neglect of any one aspect. The weight range design ensures that each objective retains at least some influence under any adaptive strategy, while also allowing for significant changes in the dominant factor based on different scenarios and tasks. For example, when power is scarce, the energy consumption weight can be increased to prioritize energy-saving paths, while the safety weight can be increased under severe weather conditions to avoid high-risk areas. This adjustable weight structure provides flexible multi-objective trade-offs for subsequent online fine-tuning.

[0022] In the aforementioned system, the path planning module employs a weighted A algorithm with a cost function f(n) = α·g(n) + β·e(n) + γ·s(n), where g(n) is the path length cost from the starting point to node n, e(n) is the energy consumption cost from the starting point to node n, and s(n) is the safety cost at node n. The safety cost is determined based on the distance to the no-fly zone, terrain threats, and weather risks. g(n), e(n), and s(n) are all normalized to dimensionless values.

[0023] The above scheme employs the following approach: During the weighted A-factor search process, path length cost, energy cost, and safety cost are calculated separately and then normalized to the same scale, eliminating incomparability caused by different units or numerical ranges. This ensures that the weighting coefficients α, β, and γ truly reflect the designer's preferences for different objectives, rather than monopolizing the cost function due to an excessively large initial value for any single factor. The safety cost integrates the distance to no-fly zones, collision risks from terrain undulations, and real-time weather threats, enabling the path search to automatically avoid dangerous areas rather than simply finding the geometric shortest path. Normalization combined with weight adjustment ensures that the final generated flight path achieves an optimal balance between distance, energy consumption, and safety, matching the scenario and mission requirements.

[0024] In the aforementioned system, the meta-learner employs a model-independent meta-learning algorithm, and the number of online fine-tuning iterations ranges from 1 to 5.

[0025] The above approach utilizes a model-independent meta-learning algorithm, enabling the meta-learner to be compatible with various types of planning policy models. It remains effective regardless of changes in the internal structure of the policy model, exhibiting greater versatility. After sufficient offline training, the initial parameters learned by the meta-learner demonstrate excellent generalization ability. When faced with new scenarios and task combinations, it can quickly converge to policy parameters suitable for the current situation with only a very small number of gradient updates. Limiting the number of iterations for online fine-tuning to a very small number of steps ensures that the policy adaptation process is extremely short, meeting the stringent time requirements for real-time decision-making during UAV flight and avoiding control delays or missed opportunities due to prolonged computation.

[0026] The system described above also includes an energy consumption estimation module connected to the task execution monitoring module, used to estimate the remaining flight time based on the remaining battery power of the UAV and the flight path; when the remaining flight time is insufficient to complete all pending tasks, the task requirement parsing module dynamically trims or reorders the task sequence based on the task priority and the remaining flight time.

[0027] Using the above solution, the energy consumption estimation module can determine in advance whether the drone has enough energy to complete all scheduled tasks based on the current remaining battery power and the energy consumption prediction of the planned route. If insufficient battery power is detected, the system will proactively prioritize tasks based on their importance, ensuring the execution of high-priority tasks or rearranging the task execution order to complete critical stages first. This not only avoids the risk of the drone making an emergency landing or crashing due to running out of power, but also maximizes the achievement of the core mission, improving the overall mission completion quality and flight safety.

[0028] In the aforementioned system, the output of the scene feature acquisition module is connected to the input of the scene classification and recognition module, transmitting the acquired multi-dimensional scene feature data to the scene classification and recognition module in real time for scene type determination. The output of the scene classification and recognition module is connected to the first input of the dynamic adaptation engine, sending the identified scene type and scene confidence level to the dynamic adaptation engine.

[0029] The input end of the task requirement parsing module is used to receive task instructions issued from the outside, and its output end is connected to the second input end of the dynamic adaptation engine to send the parsed task type, task constraints and task priority to the dynamic adaptation engine.

[0030] Internally, the scene-task mapping model serves as the first input to the dynamic adaptation engine, receiving scene type and scene confidence. The second input of the scene-task mapping model serves as the second input, receiving task type, task constraints, and task priority. The first output of the scene-task mapping model is connected to the input of the policy library, used to retrieve the initial adaptation strategy based on scene and task information. The output of the policy library is connected to the feedback input of the scene-task mapping model, returning the retrieved initial adaptation strategy to the scene-task mapping model. The second output of the scene-task mapping model is connected to the first input of the meta-learner, sending the initial adaptation strategy to the meta-learner. The second input of the meta-learner is connected to the output of the scene feature acquisition module, used to acquire the environmental sampling data required for online fine-tuning. The output of the meta-learner is connected to the input of the policy generator, sending the fine-tuned planning configuration parameters to the policy generator. The output of the policy generator, as the output of the dynamic adaptation engine, is connected to the input of the path planning module, issuing planning instructions.

[0031] The dynamic adaptation engine also includes a scene re-identification submodule. The trigger input of the scene re-identification submodule is connected to the deviation event output of the task execution monitoring module to receive replanning trigger signals. The control output of the scene re-identification submodule is connected to the trigger input of the scene feature acquisition module to trigger the re-acquisition of environmental features. The status output of the scene re-identification submodule is connected to the trigger input of the scene classification and recognition module to trigger the re-classification of the scene.

[0032] The output of the path planning module is connected to the input of the mission execution monitoring module, sending the generated flight path to the flight control system. The first feedback output of the mission execution monitoring module is connected to the trigger input of the scene re-identification submodule in the dynamic adaptation engine, used to issue a deviation event signal when environmental changes exceed a threshold. The second feedback output of the mission execution monitoring module is connected to the input of the energy consumption estimation module, sending battery status and flight path data to the energy consumption estimation module.

[0033] The output of the energy consumption estimation module is connected to the feedback input of the task requirement analysis module. When the remaining battery life is insufficient to complete all tasks, a task pruning or reordering request is sent to the task requirement analysis module. The task requirement analysis module adjusts the task sequence according to the request and resends the updated task constraints and priorities to the dynamic adaptation engine.

[0034] The modules interact with each other via an onboard data bus, which is implemented using a controller area network bus or a high-speed serial communication interface.

[0035] This invention also provides a multi-scenario dynamically adaptable UAV flight mission planning method, which, using the above-mentioned system, includes the following steps:

[0036] Acquire multi-dimensional scene feature data of the current mission environment of the drone, and obtain scene type and scene confidence level through scene classification and recognition;

[0037] Receive task instructions and parse them to obtain the task type, task constraints, and task priority;

[0038] The scene type, scene confidence, task type, and task constraints are input into the dynamic adaptation engine. The dynamic adaptation engine includes a scene-task mapping model, a policy library, and a meta-learner. The scene-task mapping model retrieves an initial adaptation policy from the policy library and uses the meta-learner to fine-tune the initial adaptation policy online, generating planning configuration parameters including path planning algorithm parameters, cost function weight coefficients, and safety margins.

[0039] Based on the planning configuration parameters and the current environment map, a path is planned and a flight path is generated.

[0040] Control the drone to fly along the flight path and monitor the flight status and environmental changes in real time;

[0041] When the detected environmental change exceeds the set threshold, scene re-identification is triggered, scene classification is performed again, and the adaptation strategy is fine-tuned again using the meta-learner to update the planning configuration parameters and perform dynamic replanning.

[0042] The proposed solution employs a scenario feature acquisition and classification process, enabling the UAV to automatically understand its current environment type and reliability. Simultaneously, it parses task instructions to extract task type, constraints, and priorities, providing clear requirements for subsequent strategy adjustments. Subsequently, it retrieves an initial strategy suitable for this scenario from the strategy library using scenario and task information, and performs rapid online fine-tuning via a meta-learner to generate a series of customized planning parameters. The flight path calculated based on these parameters achieves a balance between distance, energy consumption, and safety consistent with the current situation. During flight, the system continuously monitors the environmental state; if a significant change in the environment is detected, the scenario is immediately re-identified, and the strategy is fine-tuned again to update planning parameters and replan the path. The entire process requires no human intervention, completing a full closed loop from environmental perception, task understanding, strategy adaptation to path generation and dynamic correction, significantly improving the UAV's autonomous adaptability and overall flight efficiency when performing various tasks in diverse working scenarios.

[0043] The above method, wherein the online fine-tuning process of the meta-learner includes: constructing a fine-tuning loss function based on the current scene features and task constraints, wherein the fine-tuning loss function is a weighted sum of path length cost, energy cost, safety cost and task constraint violation penalty term; using the meta-learner to perform gradient updates on the parameters of the initial adaptation strategy no more than 5 times to obtain adaptation strategy parameters adapted to the current scene and task; and then determining the weight coefficients of the cost function.

[0044] The above approach employs a fine-tuning loss function that explicitly includes path length cost, energy cost, safety cost, and penalty for task constraint violations. This ensures that the meta-learner, during fine-tuning, not only considers the efficiency and safety of the flight itself but also strictly adheres to the task's mandatory requirements, such as waypoints and time windows. Optimized policy parameters can be obtained through a few gradient updates, and this rapid convergence ensures real-time online computation. The final determined weight coefficients directly reflect the optimal trade-off between path length, energy consumption, and safety in the current scenario, automatically handling task constraint satisfaction, ensuring that each planned path closely matches the actual flight environment and task requirements.

[0045] In the above method, the cost function weight coefficients include path length weight α, energy consumption weight β, and safety weight γ, where α∈[0.2,0.8], β∈[0.1,0.6], γ∈[0.1,0.7], and α+β+γ=1; the environmental change includes the rate of change of environmental obstacle density, with a threshold value ranging from 15% to 30%; the path length cost, energy consumption cost, and safety cost are all dimensionless values ​​that have been normalized.

[0046] The above scheme normalizes path length cost, energy cost, and safety cost, eliminating the influence of different dimensions and ensuring that the weighting coefficients accurately correspond to the design intent when adjusting various costs. The constraint that the sum of the weighting coefficients is 1 ensures that adjusting one weight automatically coordinates the changes of other weights, guaranteeing a clear and controllable overall trade-off. Using the obstacle density change rate as a monitoring quantity for environmental changes and setting a reasonable threshold accurately detects significant fluctuations in obstacle density, triggering subsequent replanning processes to ensure that the path always adapts to the current obstacle distribution. These measures enable the method to exhibit stable and predictable performance in multi-objective optimization and dynamic environmental response.

[0047] The present invention has the following beneficial effects:

[0048] The system and method of this invention organically combine scene perception, task understanding, dynamic policy adaptation, and real-time monitoring and replanning. This enables UAVs to automatically generate path planning schemes that match the current situation without manual parameter resetting when facing various environments and tasks. Scene classification and knowledge graph matching allow the system to accurately recognize the nature of the environment, while the confidence mechanism avoids the risk of misjudgment caused by uncertainty in identification. The meta-learner fine-tunes the initial strategy online, ensuring that the generation of planning parameters does not rely solely on preset coarse rules, but is finely adjusted according to the current specific environmental data and task requirements, thereby finding the optimal trade-off between path length, energy consumption, and flight safety that meets the current needs. Normalization ensures that the influence of each objective in the cost function strictly corresponds to the weight design, preventing loss of control due to differences in numerical magnitude.

[0049] Continuous monitoring of various environmental changes and reasonable threshold settings during flight enable the system to promptly detect substantial environmental shifts and trigger replanning, while avoiding oversensitivity to minor fluctuations, thus balancing response speed and system stability. Dynamic replanning allows the UAV to quickly switch to a new path adapted to the current situation when the environment changes abruptly, ensuring flight safety from threats such as sudden high-density obstacles, drastic changes in lighting, or electromagnetic interference. Energy consumption estimation and dynamic mission tailoring further guarantee the feasibility of critical missions under limited energy conditions, preventing forced landings or crashes due to insufficient power. The overall solution endows the UAV with stronger autonomous adaptability and multi-scenario adaptability, demonstrating significant practical value in complex and ever-changing real-world application environments. Attached Figure Description

[0050] Figure 1 A schematic diagram of the multi-scenario dynamically adaptable UAV flight mission planning system architecture provided by the present invention;

[0051] Figure 2Flowchart of the UAV flight mission planning method for dynamic adaptation in multiple scenarios provided by the present invention;

[0052] Figure 3 This is a schematic diagram illustrating the internal working principle of the dynamic adaptation engine provided by the present invention. Detailed Implementation

[0053] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0054] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual images. They should not be construed as limiting the scope of this application. To better illustrate the embodiments of the present invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0055] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present application. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0056] In the description of this invention, unless otherwise explicitly specified and limited, the term "connection" or similar designation indicating a connection between components should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0057] Reference Figure 1-3The system in this embodiment includes a scene feature acquisition module, a scene classification and recognition module, a task requirement analysis module, a dynamic adaptation engine, a path planning module, a task execution monitoring module, and an energy consumption estimation module. The scene feature acquisition module acquires multi-dimensional scene features, including visual texture feature vectors, obstacle point cloud density, electromagnetic signal power spectrum, wind speed and direction, and terrain undulation, through visual sensors, lidar, electromagnetic intensity meters, meteorological sensors, and elevation data interfaces mounted on the UAV. The scene classification and recognition module has a built-in scene knowledge graph, which pre-defines feature prototypes and relationships of various typical scenes such as urban canyons, forests, water surfaces, mountains, and areas with strong electromagnetic interference. The scene classification and recognition module performs similarity matching between real-time features and the knowledge graph, outputting the scene type and a confidence score from 0 to 1. When the confidence score is greater than or equal to a confidence threshold (for example, 0.8, ranging from 0.7 to 0.98), the scene type is confirmed; if it is below the threshold, it is marked as an unknown scene and a general initial strategy is adopted. The mission requirement parsing module receives mission instructions from the ground station or airborne mission computer, and parses out mission types such as inspection, logistics delivery, and search and rescue, as well as information such as time window constraints, no-fly zones, waypoints that must be passed, and mission priorities.

[0058] The dynamic adaptation engine comprises a scene-task mapping model, a policy generator, a policy library, a meta-learner, and a scene re-identification submodule. The policy library stores initial adaptation policy parameters for each typical scene and task type combination. These parameters include the choice of path planning algorithm type (e.g., improved A, RRT, DWA), initial values ​​for cost function weight coefficients, and initial values ​​for safety margins. The scene-task mapping model retrieves matching initial adaptation policies from the policy library based on the received scene type, confidence level, task type, and constraints. For unknown scenes, it retrieves several policies with the highest overall similarity and performs a weighted combination. The meta-learner employs the Model-Independent Meta-Learning Algorithm (MAML), trained offline on a large number of scene-task combinations to obtain good parameter initialization points. During online fine-tuning, a small batch of data sampled from the current flight environment is used to construct a fine-tuning loss function. This loss function is obtained by weighted summation of path length cost, energy cost, safety cost, and task constraint violation penalty. Iterative gradient updates are performed on the initial policy parameters, with the number of updates ranging from 1 to 5 to balance real-time performance and adaptation accuracy. After fine-tuning, planning configuration parameters adapted to the current flight scenario are obtained, specifically including path planning algorithm parameters, path length weight α ranging from 0.2 to 0.8, energy consumption weight β ranging from 0.1 to 0.6, safety weight γ ranging from 0.1 to 0.7 with α+β+γ=1, and safety margin, etc. The strategy generator generates planning instructions based on these parameters and issues them to the path planning module.

[0059] The path planning module takes a grid or topology map constructed from planning configuration parameters and fused sensor data as input. It uses a weighted A-scale global planner to calculate the global path, with a cost function f(n) = α·g(n) + β·e(n) + γ·s(n). g(n) is calculated based on the sum of Euclidean or Manhattan distances; e(n) estimates energy consumption using the UAV dynamics model and wind speed; s(n) is calculated based on distance to no-fly zones, terrain collision risk, and weather threat fields, with higher values ​​for closer distances and greater threats. g(n), e(n), and s(n) are all normalized to dimensionless values. Locally, a smooth trajectory with sufficient safety margin is generated using dynamic windowing or spline interpolation. The flight path is output to the flight control system and executed by the mission execution monitoring module. This module synchronously receives battery status, real-time sensor data, and mission progress, and monitors environmental changes. The energy consumption estimation module estimates the remaining battery life based on the remaining battery capacity and the energy consumption profile of the planned route. When it determines that it is impossible to complete all pending tasks, it notifies the task requirement parsing module to cut low-priority tasks or adjust the task order according to priority.

[0060] During task execution, if the detected rate of change in environmental obstacle density exceeds a threshold (e.g., 20%, with a range of 15%–30%), or the rate of change in light intensity exceeds a threshold (e.g., 30%, with a range of 25%–40%), or the change in electromagnetic interference intensity exceeds 5dB (with a range of 3dB–8dB), a deviation event is generated, triggering the scene re-identification submodule. The scene re-identification submodule requires the scene feature acquisition module to obtain the latest features, the scene classification and identification module to re-identify the scene type, the meta-learner to perform online fine-tuning based on the new scene and the current remaining task constraints, update the planning configuration parameters, and the path planning module to perform local or global re-planning and switch to a new path. This forms a closed loop of perception-adaptation-planning-monitoring-re-adaptation.

[0061] The corresponding process flow is as follows: Step S301: Obtain multi-dimensional scene features and classify and identify scene types and confidence levels. S302: Parse task instructions to obtain task features. S303: The dynamic adaptation engine uses scene type, confidence level, and task features to retrieve the initial strategy and perform online fine-tuning through meta-learning to generate planning configuration parameters. S304: Perform path planning based on the configuration parameters. S305: Execute flight and monitor. S306: Determine if environmental changes exceed the threshold; if not, return to continue monitoring; if so, proceed to S307: Trigger scene re-identification and further fine-tuning of the adaptation strategy, update parameters, and re-plan, then return to execution. The entire process continues until the task is completed or terminated for safety reasons.

[0062] In the offline phase, flight environment samples and manually optimized policy parameters are collected under multiple scenarios and tasks to train the MAML meta-learner, enabling it to quickly adapt to cross-scenario tasks. In the online phase, the meta-learner and policy library are loaded onto the UAV embedded computing platform, and fine-tuning is performed according to this method.

[0063] Furthermore, in the multi-scenario dynamic adaptation UAV flight mission planning system, the modules interact closely and coordinate control through the airborne data bus. The specific connection relationship is as follows: the output of the scene feature acquisition module is connected to the input of the scene classification and recognition module, and the acquired multi-dimensional scene feature data is transmitted in real time for scene type identification; at the same time, the input of the task requirement parsing module is used to receive task instructions issued from the outside, and its output is connected to the second input of the dynamic adaptation engine, and the parsed task type, task constraints and task priority are sent to the engine. Within the dynamic adaptation engine, the first input of the scene-task mapping model serves as the engine's first input, connecting to the output of the scene classification and recognition module to receive scene type and confidence level. Its second input serves as the engine's second input, connecting to the output of the task requirement parsing module. The first output of this model connects to the input of the policy library to retrieve the initial adaptation policy based on scene and task information. The output of the policy library then feeds back to the model's feedback input to return the initial policy. Simultaneously, the second output of this model connects to the first input of the meta-learner, whose second input is additionally directly connected to the output of the scene feature acquisition module to obtain the environmental sampling data required for online fine-tuning. The output of the meta-learner connects to the input of the policy generator, and the output of the policy generator, as the overall output of the dynamic adaptation engine, connects to the input of the path planning module to send planning instructions. The path planning module's output is connected to the mission execution monitoring module's input, sending the generated flight path to the flight control execution unit. The mission execution monitoring module's first feedback output is connected to the trigger input of the scene re-identification submodule within the dynamic adaptation engine, used to issue a deviation event signal to trigger dynamic replanning when environmental changes exceed a threshold. Its second feedback output is connected to the energy consumption estimation module's input to transmit battery status and flight path data. The energy consumption estimation module's output is connected to the mission requirement analysis module's feedback input, sending a mission trimming or reordering request to this module when the remaining flight time is insufficient to complete all tasks. Furthermore, the scene re-identification submodule within the dynamic adaptation engine's trigger input is connected to the mission execution monitoring module's deviation event output. Its control output is connected to the scene feature acquisition module's trigger input to trigger re-acquisition of environmental features, and its status output is connected to the scene classification and recognition module's trigger input to trigger re-classification of the scene. This forms a complete closed-loop feedback chain of perception-adaptation-planning-execution-replanning, and communication between all the above modules is achieved through the airborne controller local area network bus or a high-speed serial communication interface.

[0064] It is worth noting that when the meta-learner fine-tunes the initial adaptation strategy online, it constructs and minimizes the following adaptive scene-task coupling loss function:

[0065] ;

[0066] in: This is the adjustable parameter vector of the current adaptation strategy. These are path length weight, energy consumption weight, and safety weight, respectively. The normalized safety margin coefficient (obtained by dividing the actual safety distance by the current environment map grid resolution) consists of all dimensionless real numbers and satisfies the following conditions: ;

[0067] The initial adaptation policy parameter vector retrieved from the policy library has components that are... same;

[0068] The first The path length cost, energy consumption cost, and security cost at each environmental sampling point are all dimensionless values ​​(range) after Min-Max normalization. );

[0069] For the first The penalty term for task constraint violation at each sampling point is a dimensionless non-negative value, which is obtained by weighted normalization of the depth of no-fly zone penetration, the deviation distance of the necessary point, and the violation amount of the time window.

[0070] The number of current environmental sampling points used for online fine-tuning (dynamically determined by onboard computing resources, ranging from 10 to 50);

[0071] The scene perturbation factor is defined as the current scene feature vector. Feature vector at the last re-identification time Weighted Euclidean distance between them:

[0072] ;

[0073] in For the first Contribution weights of dimensional features ( ), Let be the standard deviation of this feature in the offline training samples. It is a dimensionless pure scalar;

[0074] The scene-sensitive gain function is defined as follows:

[0075] ;

[0076] in This is the curvature adjustment coefficient. It monotonically increases with the degree of change in the scene;

[0077] The constraint penalty enhancement function is defined as follows:

[0078] ;

[0079] in Based on the penalty weight, The saturation growth rate coefficient. It increases exponentially with the degree of change in the scene;

[0080] The task urgency factor is defined as the ratio of the current remaining available time to the total expected task time. The smaller the value, the more urgent the task;

[0081] The regularization strength adaptive function is defined as:

[0082] ;

[0083] in The basic regularization coefficient, This is the steepness coefficient. As the urgency of the task increases ( (Decreases) and monotonically increases, limiting the parameter from deviating excessively from the initial strategy;

[0084] Meta-learners use gradient descent to... Perform 1 to 5 iterations to obtain the optimal adaptation parameters suitable for the current scenario and task. Based on this, the weight coefficients and safety margin of the cost function used for path planning are determined.

[0085] Parameter description table:

[0086] Policy parameter vector (including) ) Dimensionless ， Normalized path length / energy consumption / security cost Dimensionless Interval, offline statistical Min-Max normalization Penalties for violating normalization constraints Dimensionless No-fly zone penetration depth / maximum permissible depth Number of sampling points Dimensionless 10-50, adaptive onboard computing power No. Dimensional feature contribution weights Dimensionless Determined by entropy weight method or principal component analysis. No. Offline standard deviation of dimensional features Dimensionless Offline training set statistics Scene-sensitive gain curvature coefficient Dimensionless 0.5~2.0, control Follow rate of increase Basic constraint penalty weights Dimensionless The value ranges from 0.3 to 0.8, determined by offline cross-validation. Punishment saturation growth rate Dimensionless 1.0–3.0, controlling the amplification factor of the penalty with mutation. Basic regularization strength Dimensionless 0.01 to 0.1, to ensure the stability of the initial strategy. Task urgency and steepness coefficient Dimensionless 3-10, control right Sensitivity to change Number of iterations Gradient descent update steps Dimensionless 1 to 5 times, exactly as per the instructions.

[0087] Example:

[0088] Flight scenario: The drone is performing a logistics delivery mission in sparse woodland. The initial scene features are consistent with the previous re-identification. (Environment stable). Sufficient time remains for the task. , , .

[0089] calculate The path cost gain is approximately 1, so fine-tuning is almost unaffected.

[0090] The penalty weight has been slightly increased;

[0091] Parameters can be explored to a certain extent.

[0092] After 3 gradient updates The changes were minor. The path planning continues to prioritize a balance between length and energy consumption.

[0093] Sudden change scenario: During flight, the plane suddenly enters an urban canyon, the density of obstacles increases sharply, electromagnetic interference intensifies, and after re-identification... It jumped to 1.8. Meanwhile, time was running out for the mission. .

[0094] The path cost weight is nearly doubled;

[0095] The constraints and penalties have increased dramatically.

[0096] As the regularization strength increases, the limiting parameters fluctuate wildly.

[0097] After 4 gradient updates, It decreased from 0.42 to 0.22. The safety weight increased significantly from 0.25 to 0.58, indicating a greater safety margin. The accuracy has been improved from 0.15 to 0.38 (corresponding to an actual safe distance increase from 3m to 7.6m, based on a grid resolution of 0.2m / pixel). The new path automatically avoids dense building clusters and provides ample safety buffers, while the constraint (no-fly zone) penalty is suppressed to near zero.

[0098] Technical effects:

[0099] Adaptive acceleration for sudden scene changes: The squared term design enables smooth fine-tuning under small perturbations and rapid tilting to a safe target under large abrupt changes. In actual tests, when the obstacle density changes by 30%, the number of iterations required for strategy convergence is reduced from 8 to 10 times for the fixed loss function to 3 to 4 times.

[0100] Constraints satisfy robustness: The saturation exponential characteristic ensures that the constraint penalty is rapidly amplified under abrupt change scenarios, forcing the optimized path to stay away from no-fly zones and weather threat areas. In simulation tests, the probability of no-fly zone intrusion is reduced by about 62% compared to fixed penalty.

[0101] Task-critical adaptive stability: When time is tight, the parameter offset is automatically narrowed to avoid planning failure due to over-optimization. Actual testing shows that... The success rate of time-based planning is 41% higher than that of unregularized solutions.

[0102] Dimensionless uniformity and numerical robustness: All inputs and outputs are dimensionless scalars, and the gradient magnitude is stable at... The range is wide, with no gradient explosion or vanishing phenomenon, making it suitable for airborne embedded real-time computing.

[0103] Working principle and process:

[0104] 1. Replanning Trigger: When the task execution monitoring module detects that the environmental change exceeds the threshold, it generates a deviation event and triggers the scene re-identification sub-module.

[0105] 2. Feature Update and Perturbation Calculation: The scene feature acquisition module obtains the latest feature vector. Features from the last re-identification time stored They are fed into the meta-learner to calculate the scene perturbation factor. .

[0106] 3. Task urgency assessment: The task requirement analysis module obtains the current remaining time. With total expected time , combined calculate .

[0107] 4. Loss Function Construction: The meta-learner samples from the current environment map. Each node has a number of nodes; calculate the number of nodes. , combined and Dynamic calculation Construct a complete loss .

[0108] 5. Gradient fine-tuning: using policies provided by the policy library Starting from this point, perform 1 to 5 gradient descent updates to obtain... .

[0109] 6. Parameter Output and Path Replanning: The policy generator will... In and normalized safety margin (Multiply by the raster resolution to restore the actual distance) is converted into planning parameters, and then sent to the path planning module to perform global or local replanning.

[0110] 7. Closed-loop iteration: During the execution of the new path, continuous monitoring is carried out. If the environment changes significantly again, the above process is repeated to form a complete dynamic closed loop of perception → perturbation quantification → loss adaptive reconstruction → parameter fine-tuning → path update.

[0111] In summary, the idea behind this system and method is to enable UAVs to autonomously adjust their planning strategies based on the environment and mission, continuously perceive changes during flight, and readjust in a timely manner, forming a complete closed loop from perception to adaptation, from planning to monitoring, and then to re-adaptation.

[0112] At the start of flight or mission execution, the drone uses its onboard visual sensors, lidar, electromagnetic intensity meter, meteorological sensors, and elevation data interface to collect multi-dimensional feature information about the current environment, such as visual texture features, obstacle point cloud density, electromagnetic signal strength, wind speed and direction, and terrain undulation. These features are fed into a scene classification and recognition module and matched against a pre-built scene knowledge graph. The knowledge graph stores feature prototypes of various typical scenes, such as urban canyons, forests, water surfaces, mountains, and areas with strong electromagnetic interference, as well as their interrelationships. After matching, the type of the current scene and a confidence level reflecting the reliability of the recognition are obtained. If the confidence level is high enough, the scene type is confirmed; if it is low, it is treated as an unfamiliar scene, employing a more general initial strategy.

[0113] Meanwhile, the mission requirement analysis module receives mission instructions from the ground station or onboard computer, extracting constraints such as mission type, no-fly zones, and necessary waypoints, as well as the priority of each mission. These analysis results, along with the scenario type and confidence level, are input into the dynamic adaptation engine.

[0114] Within the dynamic adaptation engine, the scene-task mapping model retrieves an initial adaptation strategy from the policy library that matches the scenario based on scene type, confidence level, and task characteristics. These strategy parameters cover settings such as the chosen direction of the path planning algorithm, initial weights of each objective in the cost function, and safety buffer distance. Subsequently, the meta-learner fine-tunes this initial strategy online. The fine-tuning process constructs a loss function that includes path length cost, energy cost, safety cost, and a penalty for task constraint violation. It then performs a minimal number of gradient updates using sampled data from the current environment to quickly generate a set of planning configuration parameters adapted to the specific scene and task. These parameters specifically include path length weights, energy cost weights, safety weights, planning algorithm parameters, and safety margins. The policy generator then generates planning instructions based on these parameters and sends them to the path planning module.

[0115] After receiving the planning command, the path planning module calculates the global flight path using a weighted search algorithm based on the planning configuration parameters and the environmental map constructed from fused sensor data. It also generates smooth trajectories with sufficient safety margins locally. During the search process, path length cost, energy consumption cost, and safety cost are all normalized to ensure that the weighting coefficients accurately reflect the trade-offs for different objectives. The safety cost comprehensively considers no-fly zone distance, terrain collision risk, and weather threats, ensuring that the path automatically avoids dangerous areas. The flight path is then output to the flight control system, and the UAV executes its mission along this path.

[0116] During flight, the mission execution monitoring module continuously monitors battery status, sensor data, and mission progress, while also detecting environmental changes in real time. When indicators such as the rate of change in obstacle density, the rate of change in light intensity, or the amount of change in electromagnetic interference intensity exceed preset judgment thresholds, the scene re-identification submodule is triggered. This submodule commands the scene feature acquisition module to reacquire the current environmental features, the scene classification and recognition module to reclassify the scene, and the meta-learner to fine-tune the adaptation strategy and update the planning configuration parameters based on the new scene type and the current remaining task constraints. The path planning module then performs local or global replanning accordingly, and the UAV switches to the new path to continue executing the mission. The system also uses the energy consumption estimation module to predict the endurance based on the remaining battery power and the path energy consumption profile. When it is determined that not all scheduled tasks can be completed, the task sequence is pruned or reordered according to task priority, prioritizing the completion of important tasks.

[0117] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-scenario dynamically adaptable UAV flight mission planning system, characterized in that, include: The scene feature acquisition module is used to acquire multi-dimensional scene feature data of the current mission environment of the UAV; The scene classification and recognition module is used to match the multi-dimensional scene feature data with a preset scene knowledge graph to identify the scene type and the corresponding scene confidence level. The task requirement parsing module is used to obtain task instructions and parse them to obtain the task type, task constraints, and task priority. The dynamic adaptation engine includes a scene-task mapping model, a policy generator, a policy library, and a meta-learner. The scene-task mapping model is configured to retrieve an initial adaptation policy from the policy library based on the scene type, scene confidence, task type, and task constraints, and to use the meta-learner to fine-tune the initial adaptation policy online to generate planning configuration parameters including path planning algorithm parameters, cost function weight coefficients, and safety margins. The strategy generator is used to generate planning instructions based on the planning configuration parameters; The path planning module is used to receive the planning instructions, perform global path planning and local trajectory generation based on the planning configuration parameters and the current environment map, and output the flight path; The task execution monitoring module is used to control the UAV to fly along the flight path and monitor the flight status and environmental changes in real time. When a deviation event that meets the preset conditions is detected, the dynamic adaptation engine is triggered to perform dynamic replanning.

2. The multi-scenario dynamically adaptable UAV flight mission planning system according to claim 1, characterized in that, The dynamic adaptation engine also includes a scene re-identification submodule. When the task execution monitoring module detects that the amount of environmental change exceeds a set threshold, it generates a deviation event. After the scene re-identification submodule is triggered, it re-acquires the latest scene feature data from the scene feature acquisition module, and the scene classification and identification module re-identifies the scene type. The meta-learner then fine-tunes the adaptation strategy based on the new scene type and the current task state to update the planning configuration parameters.

3. The multi-scenario dynamically adaptable UAV flight mission planning system according to claim 2, characterized in that, The environmental change includes at least one of the following: the rate of change of environmental obstacle density, the rate of change of light intensity, and the rate of change of electromagnetic interference intensity; the set thresholds include: the threshold value for the rate of change of obstacle density is in the range of 15% to 30%, the threshold value for the rate of change of light intensity is in the range of 25% to 40%, and the threshold value for the rate of change of electromagnetic interference intensity is in the range of 3dB to 8dB.

4. The multi-scenario dynamically adaptable UAV flight mission planning system according to claim 1, characterized in that, The scene classification and recognition module confirms the scene type when the scene confidence level is greater than or equal to the confidence level threshold; the confidence level threshold ranges from 0.7 to 0.

98.

5. The multi-scenario dynamically adaptable UAV flight mission planning system according to claim 1, characterized in that, The cost function weight coefficients include path length weight α, energy consumption weight β, and safety weight γ, where α∈[0.2,0.8], β∈[0.1,0.6], γ∈[0.1,0.7], and satisfy α+β+γ=1.

6. The multi-scenario dynamically adaptable UAV flight mission planning system according to claim 5, characterized in that, The path planning module adopts the weighted A algorithm, whose cost function is f(n) = α·g(n) + β·e(n) + γ·s(n), where g(n) is the path length cost from the starting point to node n, e(n) is the energy consumption cost from the starting point to node n, and s(n) is the safety cost at node n. The safety cost is determined based on the distance to the no-fly zone, terrain threats, and meteorological risks. g(n), e(n), and s(n) are all normalized to dimensionless values.

7. The multi-scenario dynamically adaptable UAV flight mission planning system according to claim 1, characterized in that, The meta-learner employs a model-independent meta-learning algorithm, and the number of online fine-tuning iterations ranges from 1 to 5.

8. The multi-scenario dynamically adaptable UAV flight mission planning system according to claim 1, characterized in that, It also includes an energy consumption estimation module, which is connected to the task execution monitoring module, and is used to estimate the remaining flight time based on the remaining battery power of the UAV and the flight path; when the remaining flight time is insufficient to complete all pending tasks, the task requirement parsing module dynamically trims or reorders the task sequence based on the task priority and the remaining flight time.

9. A multi-scenario dynamically adaptable UAV flight mission planning method, characterized in that, The multi-scenario dynamic adaptation UAV flight mission planning system according to any one of claims 1-8 includes the following steps: Acquire multi-dimensional scene feature data of the current mission environment of the drone, and obtain scene type and scene confidence level through scene classification and recognition; Receive task instructions and parse them to obtain the task type, task constraints, and task priority; The scene type, scene confidence, task type, and task constraints are input into the dynamic adaptation engine. The dynamic adaptation engine includes a scene-task mapping model, a policy library, and a meta-learner. The scene-task mapping model retrieves an initial adaptation policy from the policy library and uses the meta-learner to fine-tune the initial adaptation policy online, generating planning configuration parameters including path planning algorithm parameters, cost function weight coefficients, and safety margins. Based on the planning configuration parameters and the current environment map, a path is planned and a flight path is generated. Control the drone to fly along the flight path and monitor the flight status and environmental changes in real time; When the detected environmental change exceeds the set threshold, scene re-identification is triggered, scene classification is performed again, and the adaptation strategy is fine-tuned again using the meta-learner to update the planning configuration parameters and perform dynamic replanning.

10. The multi-scenario dynamic adaptation UAV flight mission planning method according to claim 9, characterized in that, The online fine-tuning process of the meta-learner includes: constructing a fine-tuning loss function based on the current scene features and task constraints. The fine-tuning loss function is a weighted sum of path length cost, energy cost, safety cost, and task constraint violation penalty term. The meta-learner is used to perform gradient updates on the parameters of the initial adaptation strategy no more than 5 times to obtain adaptation strategy parameters adapted to the current scene and task, and then the weight coefficients of the cost function are determined.

Images