Artificial Intelligence-Based Unmanned Aerial Vehicle Mission Planning Method

Abstract

This invention relates to the field of unmanned aerial vehicle (UAV) mission planning technology, specifically an artificial intelligence-based UAV mission planning method. The method includes: receiving mission description text through a mission interface and parsing it to generate a structured list of mission objectives; acquiring a digital map of the target area and real-time environmental parameters, and filtering key geographical elements and dynamic environmental factors relevant to the mission; constructing a spatiotemporal constraint network, treating each objective, geographical element, and dynamic environmental factor as nodes, and transforming the UAV's flight capabilities into constraints between nodes; inputting the network into a neural network solver, and generating a preliminary trajectory sequence through simulated annealing; and calling a physical simulator to perform dynamics and energy consumption simulations, correcting illegal trajectory segments to form a feasible trajectory sequence. This method improves the accuracy and reliability of UAV mission planning and adapts to complex missions and dynamic environmental requirements.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) mission planning technology, and in particular to an artificial intelligence-based UAV mission planning method. Background Technology

[0002] Mission planning for unmanned aerial vehicles (UAVs) is a core technology ensuring their autonomous mission completion. Current technologies primarily rely on manual interpretation of mission requirements and the determination of mission objectives. Following this, basic geographic information of the target area is combined with traditional path search algorithms to generate flight paths. Some technologies incorporate environmental parameters for planning assistance, but these are all static environmental parameters. This fails to effectively integrate mission objectives, geographic features, and dynamic environmental factors. Furthermore, the vague requirements described in the mission description are difficult to translate into precise, structured objectives, making it impossible to accurately match the actual mission scenario.

[0003] Existing technologies have shortcomings. They lack a constraint system that integrates mission objectives, geographical features, dynamic environments, and the aircraft's own capabilities. This leads to the oversight of the impact of UAV flight capabilities, mission logic, and real-time environmental changes during trajectory generation, resulting in trajectories that are disconnected from actual requirements. Furthermore, the lack of effective physical verification after initial trajectory generation, relying solely on theoretical calculations to assess feasibility, can easily lead to violations of physical laws and energy consumption exceeding the aircraft's capacity, rendering the trajectory unfeasible and requiring repeated adjustments. Therefore, a UAV mission planning solution is needed that can accurately analyze mission requirements, integrate multiple constraints, and possess physical verification capabilities to address the aforementioned problems of existing technologies. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an artificial intelligence-based unmanned aerial vehicle mission planning method.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an artificial intelligence-based unmanned aerial vehicle mission planning method, comprising: The task description text is received through the task interface, and natural language parsing is performed on the task description text to generate a structured list of task objectives. Digital maps and real-time environmental parameters of the target area are obtained from the environmental database. Based on the structured task target list, key geographic elements and dynamic environmental factors related to the task are selected. Construct a spatiotemporal constraint network, taking each target in the structured task target list, the key geographic features, and the dynamic environmental factors as nodes, and transforming the flight capabilities of the unmanned aerial vehicle, the logical relationships of the tasks, and the external constraints into connections and constraints between the nodes; The spatiotemporal constraint network is input into a neural network solver, which searches the solution space of the spatiotemporal constraint network using a simulated annealing strategy to generate a preliminary mission trajectory sequence. A physical simulator is invoked to perform dynamics and energy consumption simulations on the preliminary mission trajectory sequence. The physical simulator is based on the aerodynamic model and dynamic system model of the unmanned aerial vehicle. Based on the results of the dynamics and energy consumption simulation, the track segments in the preliminary mission track sequence that violate physical laws or energy constraints are corrected to form a feasible track sequence.

[0006] As a further aspect of the present invention, a digital map and real-time environmental parameters of the target area are obtained from an environmental database. Based on the structured task target list, key geographic elements and dynamic environmental factors related to the task are selected, including: Parse the structured task target list and extract the geographic location attribute, time window attribute, and task type attribute of all task targets; Based on the geographic location attributes, a region of interest containing all task objectives is delineated on the digital map; Within the region of interest, all terrain obstacles exceeding a preset height threshold, all protected building structures, and all known communication blind spots are identified and extracted as key geographic elements. Based on the time window attribute, the predicted wind speed field, predicted precipitation area, and light intensity change curve of the region of interest within the corresponding time period are obtained from the real-time environmental parameters and used as the dynamic environmental factor.

[0007] As a further aspect of the present invention, a spatiotemporal constrained network is constructed, comprising: Each task objective in the structured task objective list is used as a task node; Each terrain obstacle, each protected building structure, and each communication blind spot in the aforementioned key geographic features is designated as a prohibited node or a cost node. The state of the dynamic environmental factors at different time slices is used as the environmental state node; Establish temporal dependency edges between task nodes, whereby the temporal dependency edges represent the required order of task execution; Establish spatial exclusion edges between task nodes and prohibited or cost nodes. The spatial exclusion edges indicate that the flight path needs to avoid a spatial region or pay an additional cost when it approaches it. Establish fitness edges between task nodes and environment state nodes, where the weight of the fitness edge represents the difficulty of performing the task under the environment state. The network consisting of all nodes and edges is defined as the spatiotemporal constrained network.

[0008] As a further aspect of the present invention, the spatiotemporal constraint network is input into a neural network solver, which searches the solution space of the spatiotemporal constraint network using a simulated annealing strategy to generate a preliminary mission trajectory sequence, including: The topology and node attributes of the spatiotemporal constrained network are encoded into feature vectors and input into the encoder part of the neural network solver; The decoder part of the neural network solver outputs the access sequence of a task node in an autoregressive manner, and the access sequence is a candidate solution. Calculate the total cost of the candidate solution under the spatiotemporal constraint network. The total cost includes the penalty cost for violating spatial exclusion edges, the execution cost for low environmental fitness, and the efficiency cost for track length. Using the simulated annealing strategy, with the goal of minimizing the total cost, the neural network solver is controlled to generate multiple candidate solutions and perform iterative optimization. From the optimized candidate solutions, the candidate solution with the lowest total cost is selected and mapped to the preliminary mission track sequence containing spatial coordinates and timestamps.

[0009] As a further aspect of the present invention, the simulated annealing strategy is employed, with the goal of minimizing the total cost, to control the neural network solver to generate multiple candidate solutions and perform iterative optimization, including: Set the initial temperature parameters, temperature decay coefficient, and termination temperature threshold for the simulated annealing algorithm, take the initial candidate solution generated by the neural network solver as the current solution, and calculate the total cost of the current solution. In each iteration, based on the current solution, the neural network solver performs random perturbation in the solution space to generate a new neighboring candidate solution and calculates the total cost of the new candidate solution; The new candidate solution is accepted based on the simulated annealing criterion. If the total cost of the new candidate solution is lower than the total cost of the current solution, the new candidate solution is accepted as the current solution. If the total cost of the new candidate solution is higher than the total cost of the current solution, the new candidate solution is accepted as the current solution with a probability related to the difference between the current temperature and the total cost. Each time a new candidate solution is accepted, the current temperature is reduced according to the temperature decay coefficient until the current temperature reaches the termination temperature threshold, thus ending the iterative optimization process. The final current solution is the optimized candidate solution.

[0010] As a further aspect of the present invention, a physical simulator is invoked to perform dynamics and energy consumption simulations on the preliminary mission trajectory sequence, including: The initial mission track sequence is discretized into a series of dense track points and line segments connecting the track points; For each flight segment, based on the speed, direction and curvature of its starting point, and combined with the aerodynamic model of the unmanned aerial vehicle, the output of each power unit required to maintain the flight state is calculated. Based on the output of each power unit and the efficiency curve of the power system model, calculate the instantaneous energy consumption and total energy consumption during the flight segment. At the same time, the aerodynamic loads and structural stresses experienced by the unmanned aerial vehicle during flight are calculated. The dynamics and energy consumption simulation is performed segment by segment until the entire preliminary mission trajectory sequence is traversed.

[0011] As a further aspect of the present invention, the step of calculating the instantaneous energy consumption and total energy consumption during the flight segment based on the output of each power unit and the efficiency curve of the power system model includes: Based on the flight state parameters of the discretized track segments of the preliminary mission track sequence, determine the output power of each power unit required to maintain the flight state; In the efficiency curve database of the power system model, the corresponding instantaneous energy conversion efficiency value is obtained by querying and interpolating based on the current output power of each power unit; Divide the output power of each power unit by its corresponding instantaneous energy conversion efficiency value to obtain the instantaneous input energy consumption of each power unit. Sum the instantaneous input energy consumption of each power unit to obtain the comprehensive instantaneous energy consumption of the flight path segment. The total instantaneous energy consumption of the flight segment is calculated by integrating the total instantaneous energy consumption of the flight segment along the expected flight time of the flight segment. The total energy consumption of the entire mission sequence is obtained by summing up the total energy consumption of all track segments in the preliminary mission track sequence. As a further aspect of the present invention, based on the results of the dynamics and energy consumption simulation, the trajectory segments in the preliminary mission trajectory sequence that violate physical laws or energy constraints are corrected, including: The flight path segments in the simulation results where the aerodynamic load exceeds the structural strength threshold are identified and marked as high-risk segments. The simulation results show a flight path segment where the instantaneous energy consumption exceeds the maximum output power of the power unit, which is then marked as an overload segment. Identify situations where the total energy consumption of the entire sequence exceeds the total energy carried by the unmanned aerial vehicle; For the high-risk segment, in the space near the track segment, a local sampling and gradient descent method is used to search for an alternative path with lower aerodynamic load; For the overload segment, the execution time of the trajectory segment is extended in the time dimension, and a smoother transition curve between speed and acceleration is redesigned. If the total energy consumption exceeds the limit, the track subsequence with the highest energy consumption will be identified in the preliminary mission track sequence, and the track subsequence will be globally replanned to reduce energy consumption. The replaced, smoothed, or replanned track segments are integrated into the original sequence to form the feasible track sequence.

[0012] As a further aspect of the present invention, after forming the feasible track sequence, the method further includes: Based on the feasible flight path sequence, a corresponding control command sequence is generated, which includes flight attitude commands, waypoint commands, and mission payload operation commands. Construct a test environment containing multiple disturbance sources, including a gust model, a sensor noise model, and an actuator delay model; In the test environment, the digital twin of the unmanned aerial vehicle was driven by the control command sequence to perform Monte Carlo simulation operation, and a large amount of process data of the simulation operation was collected. Statistical analysis is performed on the process data to evaluate the robustness of the feasible flight path sequence under disturbances, and the mission success rate index is calculated.

[0013] As a further aspect of the present invention, the method further includes a pre-training step for the neural network solver: Collect a large amount of historical successful mission data, which includes mission descriptions, environmental data, final execution track sequences, and mission results; For each piece of historical data, construct a corresponding spatiotemporal constraint network and a labeled track sequence; Using the spatiotemporal constraint network as input and the labeled trajectory sequence as training target, the neural network solver is subjected to supervised training so that it learns the mapping relationship from task constraints to feasible trajectories.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: By constructing a spatiotemporal constraint network using each objective in the structured task objective list, key geographical features related to the task, and dynamic environmental factors as nodes, the flight capabilities of the unmanned aerial vehicle (UAV), the logical relationships of the task, and external constraints are transformed into connections and constraints between nodes. This spatiotemporal constraint network is then input into a neural network solver, which uses simulated annealing to search the solution space and generate a preliminary task trajectory sequence. This approach achieves a comprehensive integration of task requirements, geographical environment, and the UAV's own capabilities, ensuring that the preliminary trajectory sequence fully adapts to task logic and real-time environmental changes. This avoids the problem of trajectory decoupling from task objectives and environmental conditions in traditional planning methods. Furthermore, the simulated annealing strategy efficiently searches for optimal solutions, reduces the generation of invalid trajectories, and improves the accuracy and efficiency of trajectory planning.

[0015] A physical simulator based on the aerodynamic and propulsion system models of the unmanned aerial vehicle (UAV) is used to simulate the dynamics and energy consumption of the initial mission trajectory sequence. Based on the simulation results, trajectory segments in the initial sequence that violate physical laws or energy constraints are corrected to form a feasible trajectory sequence. This approach allows for physical feasibility verification during the trajectory generation phase, eliminating the need for repeated subsequent adjustments. It solves the problem of traditional planning methods where trajectory feasibility relies on actual flight testing and is prone to becoming unexecutable. Furthermore, through precise energy consumption and dynamic simulations, the corrected trajectory matches the aircraft's power capacity, reducing the risk of energy depletion and attitude loss during flight, and improving the reliability and executability of the trajectory. Attached Figure Description

[0016] Figure 1 This is a flowchart of the artificial intelligence-based unmanned aerial vehicle mission planning method described in this invention; Figure 2 A flowchart for screening key geographic elements and dynamic environmental factors; Figure 3 A flowchart for constructing a spatiotemporal constrained network. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the 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, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0019] See Figure 1This invention provides an artificial intelligence-based unmanned aerial vehicle (UAV) mission planning method. The method includes: receiving mission description text through a mission interface; parsing the mission description text using natural language to generate a structured mission target list; obtaining a digital map and real-time environmental parameters of the target area from an environmental database; and filtering out key geographic elements and dynamic environmental factors related to the mission based on the structured mission target list; constructing a spatiotemporal constraint network, using each target, key geographic element, and dynamic environmental factor in the structured mission target list as a node, and transforming the UAV's flight capabilities, mission logic relationships, and external constraints into connections and constraints between nodes; inputting the spatiotemporal constraint network into a neural network solver, which searches the solution space of the spatiotemporal constraint network using a simulated annealing strategy to generate a preliminary mission trajectory sequence; calling a physical simulator to perform dynamics and energy consumption simulations on the preliminary mission trajectory sequence, with the physical simulator performing calculations based on the UAV's aerodynamic and dynamic system models; and correcting trajectory segments in the preliminary mission trajectory sequence that violate physical laws or energy constraints based on the results of the dynamics and energy consumption simulations to form a feasible trajectory sequence.

[0020] In one embodiment of the present invention, see [reference] Figure 2 The system parses a structured list of task objectives, extracting the geographic location, time window, and task type attributes of all objectives. Based on the geographic location attributes, a region of interest (ROI) encompassing all objectives is defined on a digital map. Within the ROI, all terrain obstacles exceeding a preset height threshold, all protected building structures, and all known communication blind spots are identified and extracted as key geographic features. Based on the time window attributes, predicted wind speed fields, predicted precipitation areas, and light intensity variation curves for the ROI within the corresponding time period are obtained from real-time environmental parameters, serving as dynamic environmental factors.

[0021] In a specific implementation, an exemplary application scenario of this invention is that an unmanned aerial vehicle (UAV) performs a power line inspection task in a mountainous area. The operator inputs the following task description text through the task interface: "Between 14:00 and 16:00 today, inspect transmission towers No. 1, No. 2, and No. 3 located in Valley A in sequence, and take high-resolution photos of the insulators on the top of the towers. Then, arrive at the relay station on Ridge B before 16:30." The natural language parsing module processes this task description text, and the generated structured task target list contains four task target entries. Each entry parses and encapsulates a geographic location attribute, a time window attribute, and a task type attribute. For example, for the target "inspecting transmission tower No. 1," its geographic location attribute is parsed as a set of GPS coordinates, its time window attribute is parsed as a sub-time period within the time interval [14:00, 16:00], and its task type attribute is identified as "inspection and photography."

[0022] In some embodiments, the parsing process extracts the geographic location attributes of all task objectives from a structured list of task objectives. These attributes are typically represented as a series of geographic coordinate points. The system calculates a polygonal region on a digital map that encompasses all these points based on these coordinate points and defines this region as the region of interest (ROI) for this task. For example, based on the coordinates of the aforementioned three transmission towers and one relay station, the ROI is defined as a rectangular geographic area covering a portion of the area from Valley A to Ridge B. In a specific implementation, the system scans the digital map elevation data layer of this rectangular region, identifying and extracting all terrain features whose altitude exceeds a preset height threshold. The preset height threshold can be set based on the basic safe flight altitude of the unmanned aerial vehicle (UAV). For example, if the preset height threshold is set to 150 meters, and there are three mountain peaks in the region with heights of 200 meters, 180 meters, and 120 meters, the first two peaks would be identified as terrain obstacles among the key geographic features. Simultaneously, the system extracts all protected building structures located within the ROI from a dedicated layer of the digital map, such as a protected historical building, and extracts the coordinate ranges of all known communication blind spots within the region from the communication assurance database; these will all be considered as key geographic features.

[0023] In its implementation, the system further analyzes the time window attribute of each task objective in the structured task objective list. For example, "between 14:00 and 16:00 today" defines the total time window for the inspection task, while "before 16:30" defines the deadline for reaching the relay station. Based on these time attributes, the system retrieves predicted meteorological data for the corresponding time period from the real-time environmental parameter database. The acquired predicted wind speed field can be a four-dimensional grid data containing wind speed and direction information at different altitudes. The predicted precipitation area is a probability distribution map extrapolated from radar echoes, and the light intensity variation curve is generated by an astronomical calculation model combined with date, time, and geographical location. These data together constitute dynamic environmental factors. For example, the system obtains that between 14:30 and 15:00, a weak precipitation area is predicted to occur in Valley A, and the wind direction will change from westerly to southwesterly.

[0024] Optionally, the preset altitude threshold is a configurable parameter whose value is related to the UAV model and its mission type. For UAVs performing low-altitude precision inspections, this threshold may be set to a vertical distance of 100 meters from the planned flight path. Protected building structure information can be derived from specific attribute layers in a geographic information system, while communication blind spot data may come from historical signal test reports or terrain occlusion analysis models. In specific implementations, after extracting time window attributes from the structured task target list, the system typically discretizes continuous time windows into a series of equally spaced time slices, such as one time slice every 5 minutes, and obtains a snapshot of the corresponding dynamic environmental factors for each time slice, so that they can be correlated when constructing the spatiotemporal constraint network later. It can be understood that this embodiment, by parsing the structured task target list, extracts three types of core attributes, and based on these attributes, filters highly relevant key geographic elements and dynamic environmental factors from massive amounts of data in the environmental database, effectively narrowing the range of environmental information that the subsequent planning algorithm needs to process, and laying a data foundation for constructing an accurate spatiotemporal constraint network.

[0025] In one embodiment of the present invention, see [reference] Figure 3 Each task objective in the structured task objective list is used as a task node. Each terrain obstacle, each protected building structure, and each communication blind spot in the key geographic features is used as a forbidden node or a cost node. The state of dynamic environmental factors in different time slices is used as an environment state node. Temporal dependency edges are established between task nodes, representing the required order of task execution. Spatial exclusion edges are established between task nodes and forbidden or cost nodes, representing the spatial areas that the flight path must avoid or the additional cost of approaching. Fitness edges are established between task nodes and environment state nodes, with the weight of the fitness edge representing the difficulty of performing the task under the environmental state. The network formed by all nodes and edges is defined as a spatiotemporal constraint network.

[0026] In specific implementation, the process of constructing a spatiotemporal constraint network based on the structured task target list, key geographic features, and dynamic environmental factors obtained in the above embodiment is illustrated using the mountain power line inspection task as an example. The structured task target list includes four task targets: "Inspecting Transmission Tower No. 1", "Inspecting Transmission Tower No. 2", "Inspecting Transmission Tower No. 3", and "Arriving at the B Ridge Relay Station". Each task target will serve as a task node in the spatiotemporal constraint network. Each task node is assigned a set of attributes, including its precise geographic coordinates, the time window in which the task must be executed, and the task type code. Two high mountains (with heights exceeding a preset height threshold) identified among the key geographic features are defined as forbidden nodes, whose attributes include the three-dimensional spatial envelope of the mountain and an extremely high intrusion penalty value. Protected building structures are also defined as forbidden nodes. Communication blind spots are defined as cost nodes, whose attributes include the spatial range of the blind spot and a crossing cost weight related to the degree of communication signal attenuation.

[0027] In some embodiments, the states of dynamic environmental factors in different time slices are instantiated as environmental state nodes. For example, the total task time window [14:00, 16:30] is discretized at 5-minute intervals, and an environmental state node is created for each time slice. Each environmental state node is associated with a snapshot of the predicted wind speed, wind direction, precipitation probability, and light intensity data for the corresponding time slice. For the task node "Inspecting Transmission Tower No. 1", its assigned expected execution time window is 14:20 to 14:30. The system will establish fitness edges between this task node and the environmental state nodes corresponding to the two time slices "14:20-14:25" and "14:25-14:30". The weight of the fitness edge is used to quantify the difficulty of performing the inspection task under the specific environmental state. The weight calculation can be based on an evaluation model, such as considering the impact of wind speed on shooting stability, the impact of precipitation on the clarity of the optical sensor, and the impact of light on image quality.

[0028] In practical implementation, temporal dependency edges are established between task nodes to encode the logical relationships between tasks. For example, there is a temporal dependency edge between the task node "Inspecting Transmission Tower 1" and the task node "Inspecting Transmission Tower 2". This means that the task "Inspecting Transmission Tower 2" must start after the task "Inspecting Transmission Tower 1" is completed. Similarly, the task "Inspecting Transmission Tower 3" must be completed after the task "Inspecting Transmission Tower 2". All inspection tasks must be completed before the task "Arriving at Relay Station B on Ridge". The temporal dependency edge itself can have a minimum time interval constraint. Spatial exclusion edges are established between task nodes and prohibited or cost nodes. For example, there is a spatial exclusion edge between each task node and a prohibited node representing a high mountain. This edge means that when the planned trajectory approaches or enters the spatial range defined by the prohibited node, a very high penalty cost will be included in the total cost. The spatial exclusion edge between a task node and a communication blind zone cost node means that the trajectory crossing this area will cause a decrease in communication quality, thus generating a cost proportional to the crossing distance.

[0029] Optionally, the weights of the fitness edges can be calculated using a quantization formula, expressed as:

[0030] In the formula, the symbol This represents the overall weight value of the fitness edges; a higher value indicates greater difficulty or cost in performing the task under that environmental condition. (Symbol) It's about wind speed A function used to evaluate the impact of wind on mission performance (such as flight stability and camera shake). Symbol It's about the probability of precipitation. A function is used to assess the negative impact of precipitation on sensor performance. (Symbol) It is about light intensity A function used to evaluate how well illumination conditions meet the quality requirements of optical observation tasks. (Symbol) , and This is the corresponding impact factor coefficient, the value of which is adjusted according to the task type attribute. For example, for a "taking a photo" task, The value of may be relatively large. It can be understood that, through the above method, the spatiotemporal constraint network formally expresses a complex planning problem containing multi-dimensional constraints such as space, time, environment, and rules as a network graph structure composed of nodes and weighted edges. The task nodes, forbidden nodes, cost nodes, and environment state nodes in the network are abstractions of entities and constraints, while the temporal dependency edges, spatial exclusion edges, and fitness edges specifically define the interaction rules between nodes and the cost quantification method.

[0031] In one embodiment of the present invention, a large amount of historical successful task data is collected, including task descriptions, environmental data, final execution trajectory sequences, and task results. A corresponding spatiotemporal constraint network and labeled trajectory sequences are constructed for each piece of historical data. Using the spatiotemporal constraint network as input and the labeled trajectory sequences as training targets, a neural network solver is subjected to supervised training, enabling it to learn the mapping relationship from task constraints to feasible trajectories. During the planning process, the topology and node attributes of the spatiotemporal constraint network are encoded as feature vectors and input to the encoder part of the neural network solver. The decoder part of the neural network solver outputs the access sequence of a task node sequentially in an autoregressive manner; the access sequence is a candidate solution. The total cost of the candidate solution under the spatiotemporal constraint network is calculated. The total cost includes the penalty cost for violating spatial exclusion edges, the execution cost for low environmental fitness, and the efficiency cost for trajectory length. A simulated annealing strategy is adopted, with the goal of minimizing the total cost, to control the neural network solver to generate multiple candidate solutions and perform iterative optimization. The simulated annealing algorithm is configured with initial temperature parameters, a temperature decay coefficient, and a termination temperature threshold. The initial candidate solutions generated by the neural network solver are used as the current solution, and the total cost of the current solution is calculated. In each iteration, based on the current solution, the neural network solver randomly perturbs the solution space to generate a new, neighboring candidate solution, and calculates the total cost of the new candidate solution. The simulated annealing criterion determines whether to accept the new candidate solution: if the total cost of the new candidate solution is lower than the total cost of the current solution, it is accepted; otherwise, if the total cost of the new candidate solution is higher than the total cost of the current solution, it is accepted with a probability related to the difference between the current temperature and the total cost. After each acceptance of a new candidate solution, the current temperature is reduced according to the temperature decay coefficient until it reaches the termination temperature threshold, ending the iterative optimization process. The final current solution is the optimized candidate solution. From the optimized candidate solutions, the candidate solution with the lowest total cost is selected and mapped to a preliminary mission track sequence containing spatial coordinates and timestamps.

[0032] In practical implementation, pre-training the neural network solver is a crucial preliminary step. This pre-training process requires collecting a large amount of historical successful task data. This data includes task description text, environmental data during the corresponding task execution, the actual flight path sequence executed by the UAV, and evaluations of the task completion results. For example, from the historical task records of a drone logistics company, thousands of point-to-point delivery tasks within a city can be extracted. Each data point includes order text, daily weather data, pilot-confirmed flight path, and a tag indicating whether delivery was on time. For each historical data point, based on its task description and environmental data, a spatiotemporal constraint network corresponding to the task execution needs to be reconstructed. The actual flight path sequence is then transformed into an ordered sequence of accessing task nodes, serving as the labeled flight path sequence corresponding to that spatiotemporal constraint network. Using the spatiotemporal constraint network as input features and the labeled flight path sequence as the training objective, the neural network solver undergoes supervised training, enabling it to learn the mapping relationship from complex task constraints to feasible flight path sequences. A well-trained neural network solver can generate high-quality initial candidate solutions more efficiently when faced with new spatiotemporal constrained networks.

[0033] In practical implementation, after constructing a spatiotemporal constraint network for a new task, the planning system encodes the network's topology and the attribute information of each node into a fixed-dimensional feature vector. This encoding process transforms the connections between nodes, edge weights, and node attributes such as coordinates, time, and type into tensor formats that can be processed by the encoder part of the neural network solver. The encoder part of the neural network solver is typically a graph neural network. It receives the encoded feature vector, passes it through multiple layers of information transmission and aggregation, and finally outputs a context vector representing the hidden state of the entire spatiotemporal constraint network. The decoder part of the neural network solver then uses an autoregressive approach, conditioned on the context vector, to sequentially output a sequence of task node visits. Each output step is the probability distribution of the next most likely visited task node under the conditions of the currently generated sequence. By sampling or taking the maximum value, a node is selected from this distribution, thus gradually generating a complete sequence of candidate solutions. For example, in the spatiotemporal constrained network of a power line inspection task, the decoder may first generate "Visit transmission tower node 1", then, given that the first node is transmission tower 1, generate "Visit transmission tower node 2", and so on, until a sequence containing all task nodes is generated.

[0034] In some embodiments, calculating the total cost of a candidate solution under a spatiotemporally constrained network is a deterministic evaluation process, and the total cost mainly consists of three parts. The first part is the penalty cost for violating spatial exclusion edges; that is, if the trajectory planned by the candidate solution crosses the spatial range of a forbidden node or gets too close to a cost node, a numerical cost is calculated according to a preset penalty function. The second part is the execution cost due to low environmental fitness, which is calculated by accumulating the weights of the fitness edges between each task node in the candidate solution sequence and the environmental state node corresponding to its execution time. The third part is the efficiency cost due to the trajectory length, which is usually proportional to the total length of the flight path implied by the candidate solution sequence. The total cost of a candidate solution is the weighted sum of these three costs, specifically calculated as follows:

[0035] Where: symbol The total cost of a candidate solution is represented by the symbol. Represents the cost of space penalty, symbol Represents the cost of environmental adaptation, symbol Represents the efficiency cost of track length, symbol , and These are weighting coefficients used to balance the importance of the three cost components. A simulated annealing strategy is employed to control the neural network solver for iterative optimization, with the goal of minimizing the total cost. The system sets the initial temperature parameters for the simulated annealing algorithm. Temperature decay coefficient and termination temperature threshold The simulated annealing strategy first generates a candidate solution as the current solution and calculates its total cost. In each iteration of the simulated annealing strategy, based on the current solution, the neural network solver performs random perturbation in the solution space to generate a new neighboring candidate solution. This random perturbation can be implemented by having the decoder sample based on a hidden state that introduces random noise, potentially resulting in a different node access order. The total cost of the new candidate solution is then calculated, and the simulated annealing criterion is used to determine whether to accept it. The simulated annealing criterion states that if the total cost of the new candidate solution is lower than the total cost of the current solution, the new candidate solution is accepted as the current solution. If the total cost of the new candidate solution is higher than the total cost of the current solution, it is rejected with a certain probability. Accepting a new candidate solution as the current solution, with probability The calculation formula is:

[0036] in: The difference in total cost between the old and new solutions. The current temperature parameter is used. Each time a new candidate solution is received, the current temperature is decreased according to the temperature decay coefficient, i.e., execution is performed. The iterative optimization process continues until the current temperature reaches the termination temperature threshold, at which point the current solution is obtained, which is the optimized candidate solution. Optionally, the historical successful task data collected during the pre-training phase can cover various task types to enhance the generalization ability of the neural network solver. See Table 1, which shows a simplified historical successful task dataset used to construct training samples.

[0037] Table 1: Historical Successful Task Data Table

[0038] From the optimized candidate solutions, the one with the lowest total cost is selected and mapped into a preliminary task trajectory sequence containing spatial coordinates and timestamps. The mapping process, based on the access order of task nodes in the candidate solutions, queries the geographical coordinates stored in the attributes of each task node and connects these coordinate points with straight lines or curves satisfying the minimum turning radius to form a spatial path. Simultaneously, based on the association between the time window attributes of the task nodes and the environmental state nodes, an appropriate timestamp is assigned to each waypoint on the path, thus forming the preliminary task trajectory sequence. It can be understood that this embodiment provides prior knowledge to the neural network solver through pre-training, processes the graph-based constraint network through an encoder-decoder structure, manages the solution space search process through simulated annealing, and combines multi-angle cost evaluation to finally output a preliminary trajectory sequence that achieves a balance between constraint satisfaction and cost minimization.

[0039] In one embodiment of the present invention, the initial mission trajectory sequence is discretized into a series of dense waypoints and line segments connecting the waypoints. For each line segment, based on the velocity, direction, and curvature of its starting point, and in conjunction with the aerodynamic model of the unmanned aerial vehicle, the output of each power unit required to maintain the flight state is calculated. Based on the output of each power unit and the efficiency curve of the power system model, the instantaneous energy consumption and total energy consumption during the flight of the line segment are calculated. Based on the flight state parameters of the line segment after the initial mission trajectory sequence is discretized, the output power of each power unit required to maintain the flight state is determined. In the efficiency curve database of the power system model, the corresponding instantaneous energy conversion efficiency value is obtained by querying and interpolating based on the current output power of each power unit. The output power of each power unit is divided by its corresponding instantaneous energy conversion efficiency value to obtain the instantaneous input energy consumption of each power unit. The instantaneous input energy consumption of each power unit is summed to obtain the comprehensive instantaneous energy consumption of the line segment. The comprehensive instantaneous energy consumption of the line segment is integrated along the expected flight time of the line segment to calculate the total flight energy consumption of the line segment. The total energy consumption of all flight segments in the initial mission trajectory sequence is summed to obtain the total energy consumption of the entire mission sequence. Simultaneously, the aerodynamic loads and structural stresses experienced by the unmanned aerial vehicle during flight are calculated. Dynamic and energy consumption simulations are performed segment by segment until the entire initial mission trajectory sequence is traversed.

[0040] In practical implementation, the core function of the physical simulator is to simulate the dynamics and energy consumption of the initial mission trajectory sequence. Taking the initial mission trajectory sequence of a power line inspection mission as an example, this sequence includes spatial coordinates and timestamps from the takeoff point, three transmission tower inspection points, to the relay station. The physical simulator first discretizes this continuous initial mission trajectory sequence into a series of dense track points and line segments connecting the track points. The granularity of discretization can be set according to the simulation accuracy requirements, for example, setting one track point every 0.1 seconds or every 0.5 meters, and adjacent track points constitute a line segment. For each line segment, the physical simulator calculates the output of each power unit required to maintain the flight state based on the velocity vector and heading angle of its starting track point, as well as the curvature of the line segment calculated from the three consecutive track points, combined with the aerodynamic model of the UAV. For example, for a quadcopter UAV, the physical simulator needs to calculate the thrust and torque required by each of the four rotors to maintain a specific speed, overcome wind resistance, and complete a turning maneuver.

[0041] In some embodiments, calculating the instantaneous energy consumption and total energy consumption during flight segments relies on the efficiency curves of the power system model. Based on the flight state parameters of each flight segment after discretization of the initial mission trajectory sequence, including airspeed, rate of climb, and attitude angle, the physical simulator can determine the output power of each power unit required to maintain that flight state. The unmanned aerial vehicle's power system model contains an efficiency curve database, which describes the energy conversion efficiency of the entire power chain (motor, ESC, propeller) under different output power levels. The physical simulator queries the efficiency curve database based on the current output power of each power unit and calculates the corresponding instantaneous energy conversion efficiency value through interpolation. Table 2 shows a simplified set of power unit efficiency curve data.

[0042] Table 2: Efficiency Curves of the Power Unit of a Certain Type of Quadrotor Unmanned Aerial Vehicle

[0043] In practical implementation, after obtaining the instantaneous energy conversion efficiency value of each power unit, the physical simulator divides the output power of each power unit by its corresponding instantaneous energy conversion efficiency value to obtain the instantaneous input energy consumption of each power unit, i.e., the power actually obtained from the battery. The instantaneous input energy consumption of all power units is summed to obtain the comprehensive instantaneous energy consumption of the flight segment at this moment. The physical simulator integrates the comprehensive instantaneous energy consumption along the expected flight time of the flight segment to calculate the total flight energy consumption for traversing the flight segment. The total flight energy consumption of all flight segments in the preliminary mission flight sequence is accumulated to obtain the predicted total energy consumption for executing the entire mission sequence. The calculation of energy consumption can be formally expressed as follows: the flight segment at time... Comprehensive instantaneous energy consumption The calculation formula is:

[0044] Where: symbol Indicates in The total instantaneous input energy consumption of the entire unmanned aerial vehicle's power system at any given moment, symbol This indicates the number of power units; for a quadcopter, ,symbol Indicates in Time of the first The output power of each power unit, symbol Indicates the first The output power of each power unit is The instantaneous energy conversion efficiency value is obtained by querying the efficiency curve and interpolating. Optionally, while calculating energy consumption, the physical simulator will simultaneously calculate the aerodynamic loads and structural stresses experienced by the UAV in flight. The aerodynamic load calculation is based on the current flight speed, attitude angle, and force and moment coefficients provided by the aerodynamic model, while the structural stress calculation is based on the distribution of aerodynamic loads on the airframe structure and the UAV's structural finite element model. The physical simulator performs the above dynamics and energy consumption simulations segment by segment and time step by step until it traverses the entire preliminary mission trajectory sequence, thereby generating a complete simulation data sequence containing multiple dimensions such as time, position, speed, energy consumption, and load. It can be understood that this embodiment, through high-fidelity physical simulation, transforms the preliminary mission trajectory sequence from geometric and temporal planning into quantifiable physical quantities and energy consumption predictions, which provides accurate data basis for subsequent verification of the physical feasibility of the trajectory and its correction.

[0045] In one embodiment of the invention, trajectory segments in the simulation results where aerodynamic loads exceed the structural strength threshold are identified and marked as high-risk segments. Trajectory segments in the simulation results where instantaneous energy consumption exceeds the maximum output power of the power unit are identified and marked as overload segments. Situations where the total energy consumption of the entire sequence exceeds the total energy carried by the unmanned aerial vehicle are also identified. For high-risk segments, a local sampling and gradient descent method is used in the space near the trajectory segment to search for an alternative path with lower aerodynamic loads. For overload segments, the execution time of the trajectory segment is extended in the time dimension, and a smoother transition curve for velocity and acceleration is replanned. If the total energy consumption exceeds the limit, the trajectory subsequence with the highest energy consumption is identified in the initial mission trajectory sequence, and the trajectory subsequence is globally replanned to reduce energy consumption. The trajectory segments that have been replaced, smoothed, or replanned are integrated into the original sequence to form a feasible trajectory sequence. Based on the feasible trajectory sequence, a corresponding control command sequence is generated, which includes flight attitude commands, waypoint commands, and mission payload operation commands. A test environment was constructed containing multiple disturbance sources, including a gust model, a sensor noise model, and an actuator delay model. Within this test environment, a Monte Carlo simulation was performed on a digital twin of the unmanned aerial vehicle (UAV) driven by a sequence of control commands, collecting a large amount of process data. Statistical analysis was performed on the process data to evaluate the robustness of feasible flight path sequences under disturbances and to calculate the mission success rate.

[0046] In practice, after the physical simulator completes the dynamics and energy consumption simulation of the initial mission trajectory sequence, the system corrects the trajectory sequence based on the physical constraint violations identified in the simulation results. Trajectory segments where aerodynamic loads exceed structural strength thresholds are marked as high-risk segments. For example, a trajectory requiring the UAV to make high-speed turns in strong crosswinds, where the calculated stress at the wing root exceeds 80% of the material's yield strength, is marked as a high-risk segment. Trajectory segments where instantaneous energy consumption exceeds the maximum output power of the power unit are marked as overload segments. For example, a trajectory requiring a rapid climb, where the calculated instantaneous power demand reaches 120% of the maximum continuous power of a single motor, is marked as an overload segment. The system also determines whether the total energy consumption of the entire sequence exceeds the total energy capacity carried by the UAV. For example, if the battery's nominal capacity is 20,000 mAh, but the simulated total energy consumption is equivalent to 21,000 mAh, then the energy consumption is deemed excessive.

[0047] In some embodiments, for identified high-risk segments, the correction method involves searching for an alternative path with lower aerodynamic loads in the space near the original track segment using local sampling and gradient descent. Local sampling generates multiple candidate offset track points in the three-dimensional space around the high-risk track segment with a certain step size. The gradient descent method is then used to iteratively adjust the positions of these track points to minimize a cost function with aerodynamic load as the main component, which guides the search process.

[0048] Where: symbol The symbol represents the optimization cost of the alternative path. Represents a vector consisting of the coordinates of candidate waypoints, with the symbol... It is a function that calculates the predicted aerodynamic loads on the path defined by the sequence of waypoints. (Symbol: ...) It is a function that calculates the geometric deviation of the new path from the original high-risk section path, with the sign... and This involves adjusting the weighting coefficients for the importance of load and path deviation. For identified overload segments, the correction method is to extend the execution time of the trajectory segment in the time dimension and reduce the instantaneous power demand by replanning a smoother transition curve for speed and acceleration. For example, a 60-degree turn that is required to be completed within 2 seconds is relaxed to be completed within 4 seconds, and the trajectory of its angle change over time is replanned using a fifth-order polynomial curve, thereby limiting the angular acceleration within the motor's response capability.

[0049] In practice, if total energy consumption exceeds the limit, the system identifies the most energy-intensive subsequence from the initial mission trajectory sequence. For example, analysis reveals that the segment from transmission tower 2 to transmission tower 3, due to headwinds and a winding route, consumes 40% of the total energy; this subsequence is selected as the target for global replanning. Global replanning of the high-energy-consuming subsequence reduces energy consumption. Replanning may involve selecting a more energy-efficient path under spatiotemporal constraints or adjusting the flight speed curve while meeting time windows. The replaced, smoothed, or replanned trajectory segments are then integrated into the original sequence, replacing the problematic segments, thus forming a physically feasible and energy-tolerant feasible trajectory sequence.

[0050] Optionally, based on feasible flight path sequences, the system generates corresponding control command sequences. These sequences include flight attitude commands accurate to the control cycle, a series of navigation waypoint commands, and mission payload operation commands triggered at specific waypoints, such as "activate the infrared camera and take a picture at coordinates (X,Y,Z)". A test environment containing multiple disturbance sources is constructed for robustness verification. These sources include a gust model simulating sudden wind shear, a sensor noise model simulating sensor measurement errors, and an actuator delay model simulating actuator response lag. In the test environment, the generated control command sequences drive a digital twin of the UAV in Monte Carlo simulations. During each run, the parameters of the gust model, sensor noise model, and actuator delay model are randomly sampled within preset ranges to simulate real-world uncertainties. The system collects a large amount of process data from the simulation, including deviations between the flight trajectory and the command trajectory, mission completion status, and remaining energy in each simulation. Statistical analysis is performed on the process data to evaluate the robustness of feasible trajectory sequences under disturbances and to calculate the mission success rate index. For example, if 950 out of 1000 Monte Carlo simulations successfully reach all target points and complete the task, the mission success rate is evaluated as 95%. It can be understood that this embodiment identifies and corrects parts of the initial plan that do not conform to physical laws or energy constraints through specific methods, and conducts extensive testing by introducing random disturbances in a high-fidelity simulation environment, thereby outputting a final trajectory scheme that is not only feasible under ideal conditions but also highly robust under uncertain environments.

[0051] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An artificial intelligence-based unmanned aerial vehicle mission planning method, characterized in that, The method includes: The task description text is received through the task interface, and natural language parsing is performed on the task description text to generate a structured list of task objectives. Digital maps and real-time environmental parameters of the target area are obtained from the environmental database. Based on the structured task target list, key geographic elements and dynamic environmental factors related to the task are selected. Construct a spatiotemporal constraint network, taking each target in the structured task target list, the key geographic features, and the dynamic environmental factors as nodes, and transforming the flight capabilities of the unmanned aerial vehicle, the logical relationships of the tasks, and the external constraints into connections and constraints between the nodes; The spatiotemporal constraint network is input into a neural network solver, which searches the solution space of the spatiotemporal constraint network using a simulated annealing strategy to generate a preliminary mission trajectory sequence. A physical simulator is invoked to perform dynamics and energy consumption simulations on the preliminary mission trajectory sequence. The physical simulator is based on the aerodynamic model and dynamic system model of the unmanned aerial vehicle. Based on the results of the dynamics and energy consumption simulation, the track segments in the preliminary mission track sequence that violate physical laws or energy constraints are corrected to form a feasible track sequence.

2. The artificial intelligence-based unmanned aerial vehicle mission planning method as described in claim 1, characterized in that, Obtain digital maps and real-time environmental parameters of the target area from the environmental database. Based on the structured task objective list, filter out key geographic features and dynamic environmental factors relevant to the task, including: Parse the structured task target list and extract the geographic location attribute, time window attribute, and task type attribute of all task targets; Based on the geographic location attributes, a region of interest containing all task objectives is delineated on the digital map; Within the region of interest, all terrain obstacles exceeding a preset height threshold, all protected building structures, and all known communication blind spots are identified and extracted as key geographic elements. Based on the time window attribute, the predicted wind speed field, predicted precipitation area, and light intensity change curve of the region of interest within the corresponding time period are obtained from the real-time environmental parameters and used as the dynamic environmental factor.

3. The artificial intelligence-based unmanned aerial vehicle mission planning method as described in claim 2, characterized in that, Construct a spatiotemporally constrained network, including: Each task objective in the structured task objective list is used as a task node; Each terrain obstacle, each protected building structure, and each communication blind spot in the aforementioned key geographic features is designated as a prohibited node or a cost node. The state of the dynamic environmental factors at different time slices is used as the environmental state node; Establish temporal dependency edges between task nodes, whereby the temporal dependency edges represent the required order of task execution; Establish spatial exclusion edges between task nodes and prohibited or cost nodes. The spatial exclusion edges indicate that the flight path needs to avoid a spatial region or pay an additional cost when it approaches it. Establish fitness edges between task nodes and environment state nodes, where the weight of the fitness edge represents the difficulty of performing the task under the environment state. The network consisting of all nodes and edges is defined as the spatiotemporal constrained network.

4. The artificial intelligence-based unmanned aerial vehicle mission planning method as described in claim 3, characterized in that, The spatiotemporal constraint network is input into a neural network solver, which searches the solution space of the spatiotemporal constraint network using a simulated annealing strategy to generate a preliminary mission trajectory sequence, including: The topology and node attributes of the spatiotemporal constrained network are encoded into feature vectors and input into the encoder part of the neural network solver; The decoder part of the neural network solver outputs the access sequence of a task node in an autoregressive manner, and the access sequence is a candidate solution. Calculate the total cost of the candidate solution under the spatiotemporal constraint network. The total cost includes the penalty cost for violating spatial exclusion edges, the execution cost for low environmental fitness, and the efficiency cost for track length. Using the simulated annealing strategy, with the goal of minimizing the total cost, the neural network solver is controlled to generate multiple candidate solutions and perform iterative optimization. From the optimized candidate solutions, the candidate solution with the lowest total cost is selected and mapped to the preliminary mission track sequence containing spatial coordinates and timestamps.

5. The artificial intelligence-based unmanned aerial vehicle mission planning method as described in claim 4, characterized in that, Using the simulated annealing strategy, with the objective of minimizing the total cost, the neural network solver is controlled to generate multiple candidate solutions and perform iterative optimization, including: Set the initial temperature parameters, temperature decay coefficient, and termination temperature threshold for the simulated annealing algorithm, take the initial candidate solution generated by the neural network solver as the current solution, and calculate the total cost of the current solution. In each iteration, based on the current solution, the neural network solver performs random perturbation in the solution space to generate a new neighboring candidate solution and calculates the total cost of the new candidate solution; The new candidate solution is accepted based on the simulated annealing criterion. If the total cost of the new candidate solution is lower than the total cost of the current solution, the new candidate solution is accepted as the current solution. If the total cost of the new candidate solution is higher than the total cost of the current solution, the new candidate solution is accepted as the current solution with a probability related to the difference between the current temperature and the total cost. Each time a new candidate solution is accepted, the current temperature is reduced according to the temperature decay coefficient until the current temperature reaches the termination temperature threshold, thus ending the iterative optimization process. The final current solution is the optimized candidate solution.

6. The artificial intelligence-based unmanned aerial vehicle mission planning method as described in claim 4, characterized in that, A physics simulator is invoked to perform dynamics and energy consumption simulations on the preliminary mission trajectory sequence, including: The initial mission track sequence is discretized into a series of dense track points and line segments connecting the track points; For each flight segment, based on the speed, direction and curvature of its starting point, and combined with the aerodynamic model of the unmanned aerial vehicle, the output of each power unit required to maintain the flight state is calculated. Based on the output of each power unit and the efficiency curve of the power system model, calculate the instantaneous energy consumption and total energy consumption during the flight segment. At the same time, the aerodynamic loads and structural stresses experienced by the unmanned aerial vehicle during flight are calculated. The dynamics and energy consumption simulation is performed segment by segment until the entire preliminary mission trajectory sequence is traversed.

7. The artificial intelligence-based unmanned aerial vehicle mission planning method as described in claim 6, characterized in that, The calculation of instantaneous energy consumption and total energy consumption during the flight segment, based on the output of each power unit and the efficiency curve of the power system model, includes: Based on the flight state parameters of the discretized track segments of the preliminary mission track sequence, determine the output power of each power unit required to maintain the flight state; In the efficiency curve database of the power system model, the corresponding instantaneous energy conversion efficiency value is obtained by querying and interpolating based on the current output power of each power unit; Divide the output power of each power unit by its corresponding instantaneous energy conversion efficiency value to obtain the instantaneous input energy consumption of each power unit. Sum the instantaneous input energy consumption of each power unit to obtain the comprehensive instantaneous energy consumption of the flight path segment. The total instantaneous energy consumption of the flight segment is calculated by integrating the total instantaneous energy consumption of the flight segment along the expected flight time of the flight segment. The total energy consumption of the entire mission sequence is obtained by summing up the total energy consumption of all flight segments in the preliminary mission trajectory sequence.

8. The artificial intelligence-based unmanned aerial vehicle mission planning method as described in claim 7, characterized in that, Based on the results of the dynamics and energy consumption simulation, the track segments in the preliminary mission track sequence that violate physical laws or energy constraints are corrected, including: The flight path segments in the simulation results where the aerodynamic load exceeds the structural strength threshold are identified and marked as high-risk segments. The simulation results show a flight path segment where the instantaneous energy consumption exceeds the maximum output power of the power unit, which is then marked as an overload segment. Identify situations where the total energy consumption of the entire sequence exceeds the total energy carried by the unmanned aerial vehicle; For the high-risk segment, in the space near the track segment, a local sampling and gradient descent method is used to search for an alternative path with lower aerodynamic load; For the overload segment, the execution time of the trajectory segment is extended in the time dimension, and a smoother transition curve between speed and acceleration is redesigned. If the total energy consumption exceeds the limit, the track subsequence with the highest energy consumption will be identified in the preliminary mission track sequence, and the track subsequence will be globally replanned to reduce energy consumption. The replaced, smoothed, or replanned track segments are integrated into the original sequence to form the feasible track sequence.

9. The artificial intelligence-based unmanned aerial vehicle mission planning method as described in claim 8, characterized in that, After forming the feasible track sequence, the method further includes: Based on the feasible flight path sequence, a corresponding control command sequence is generated, which includes flight attitude commands, waypoint commands, and mission payload operation commands. Construct a test environment containing multiple disturbance sources, including a gust model, a sensor noise model, and an actuator delay model; In the test environment, the digital twin of the unmanned aerial vehicle was driven by the control command sequence to perform Monte Carlo simulation operation, and a large amount of process data of the simulation operation was collected. Statistical analysis is performed on the process data to evaluate the robustness of the feasible flight path sequence under disturbances, and the mission success rate index is calculated.

10. The artificial intelligence-based unmanned aerial vehicle mission planning method as described in claim 1, characterized in that, The method also includes a pre-training step for the neural network solver: Collect a large amount of historical successful mission data, which includes mission descriptions, environmental data, final execution track sequences, and mission results; For each piece of historical data, construct a corresponding spatiotemporal constraint network and a labeled track sequence; Using the spatiotemporal constraint network as input and the labeled trajectory sequence as training target, the neural network solver is subjected to supervised training so that it learns the mapping relationship from task constraints to feasible trajectories.

Images