Modular integrated control platform for complex mission unmanned aerial vehicles

Abstract

The application belongs to the field of unmanned aerial vehicle control, and particularly relates to a complex task-oriented unmanned aerial vehicle modular integrated control platform, which comprises: a collection module that collects and processes real-time flight state, complex task text, meteorological data and other information of unmanned aerial vehicle groups; a task analysis module that decomposes complex tasks into a single subtask set and an association matrix, and evaluates the task complexity for resource allocation; an atmospheric analysis module that inverses a real-time flight association adjustment graph based on the corrected meteorological data, flight state and the like, in combination with a target optimization algorithm and an optimization function with the minimum resource quantity and the maximum task completion probability as targets; a record analysis module that generates a performance transition matrix by using a hidden Markov algorithm; and a flight control module that corrects the flight state in response to the adjustment graph, and ensures that the task completion probability meets a preset safety threshold. The platform realizes efficient cooperative control and resource optimized allocation of unmanned aerial vehicle groups under complex tasks through modular integration.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of unmanned aerial vehicle control, and particularly relates to a modular integrated control platform for unmanned aerial vehicles facing complex tasks. BACKGROUND

[0002] As the core carrier for executing complex tasks, unmanned aerial vehicle clusters are widely used in combat and reconnaissance integration, multi-aircraft cooperative reconnaissance, target range test evaluation, emergency material delivery, etc. The complex tasks executed by the unmanned aerial vehicle clusters usually include multi-target and multi-stage cooperative requirements, and need to rely on the cooperative support of real-time flight state perception, accurate environment adaptation, dynamic task scheduling and reliable performance monitoring. However, the existing technology has significant deficiencies in the fusion and connection of these key links and the overall performance guarantee. On the one hand, in the task processing and cluster adaptation link, the complex task can only be decomposed into independent single subtasks, and neither the association logic between the single subtasks nor the initialization configuration of the unmanned aerial vehicle cluster based on the single subtask requirements is built, resulting in the innate mismatch between the single subtask allocation and the cluster capability, which lays a foundation for the deviation of subsequent task execution. On the other hand, the atmospheric environment data processing lacks general adaptability, although it can collect flight area meteorological data such as airspeed, air pressure height and attack angle, it relies on fixed correction parameters and cannot cope with the differences in aerodynamic characteristics of different models of unmanned aerial vehicles, hardware deviations of different sensors and airflow disturbance of different installation positions, resulting in insufficient accuracy of collected data and difficulty in providing accurate actual environment state support for subsequent links. At the same time, the performance monitoring link is limited to fault detection or output performance recording of a single unmanned aerial vehicle, and does not quantify the performance evolution law of the unmanned aerial vehicle cluster in the task execution process through a probabilistic model, making it difficult to grasp the completion probability of the single subtask and the overall complex task in real time and to predict the task risk. The flight control link only adjusts based on the real-time flight state error of a single unmanned aerial vehicle, and can only guarantee the attitude stability of a single unmanned aerial vehicle, without combining the environment adaptation requirements of the single subtask association or integrating the cluster performance probability feedback, so that the data and control logic of each link are disconnected and cannot form effective cooperation, ultimately resulting in the difficulty of the overall completion probability of the complex task to meet the preset safety requirements and the inability to support the reliable execution of complex tasks by the unmanned aerial vehicle cluster. SUMMARY

[0003] To address the shortcomings of existing technologies, this invention proposes a modular integrated control platform for unmanned aerial vehicles (UAVs) designed for complex tasks. This platform includes a data acquisition module, a task analysis module, an atmospheric analysis module, a flight control module, and a recording and analysis module. The data acquisition module collects and processes real-time flight status of the UAV swarm, complex task text, meteorological data, and other information. The task analysis module decomposes the complex task into a set of individual subtasks and an association matrix, assessing the task complexity for resource allocation. The atmospheric analysis module, based on corrected meteorological data and flight status, combines an objective optimization algorithm with an optimization function aiming to minimize resource consumption and maximize task completion probability to invert the real-time flight association adjustment graph. The recording and analysis module uses a hidden Markov algorithm to generate a performance transition matrix. The flight control module responds to the adjustment graph to correct the flight status, ensuring that the task completion probability meets a preset safety threshold. This platform, through modular integration, achieves efficient collaborative control and optimized resource allocation of UAV swarms under complex tasks.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A modular integrated control platform for unmanned aerial vehicles (UAVs) designed for complex missions includes: a data acquisition module, a mission analysis module, an atmospheric analysis module, a flight control module, and a record analysis module.

[0006] The acquisition module is used to acquire real-time flight status parameters, complex mission text, meteorological data of the flight area, output performance data and fault detection data of the configured UAV swarm, and perform timestamp alignment and filtering.

[0007] The task parsing module is used to obtain a single subtask set and a single subtask association matrix that can be constructed by a single subtask that can be completed by at most one drone, based on the complex task text and the task parsing algorithm, and to obtain the task complexity for allocating resources by combining the task complexity evaluation algorithm.

[0008] The atmospheric analysis module is used to invert the real-time flight association adjustment map of the UAV swarm by combining the meteorological data of the flight area after correction by the general correction coefficient, the real-time flight status parameters of each UAV, the single sub-task association matrix and the task complexity, the target optimization algorithm and the preset target optimization function; the target optimization function is constructed by the allocation of resources for the complex task, the probability of completing the single sub-task and the probability of completing the complex task.

[0009] The record parsing module is used to obtain the performance transition matrix of the UAV swarm by combining the output performance data and fault detection data with the Hidden Markov Algorithm.

[0010] The flight control module is used to respond to the real-time flight correlation adjustment diagram to make real-time corrections to the real-time flight status of at least one UAV, so that the probability of completing a single subtask and the probability of completing a complex task in the performance transfer matrix of the UAV swarm meet the preset task safe completion probability threshold in real time.

[0011] Specifically, the flight state parameters include at least flight attitude, flight heading, flight position, and speed; the meteorological data of the flight area includes at least indicated airspeed, vacuum speed, barometric altitude, climb rate, angle of attack, and sideslip angle; the universal correction coefficient is obtained by combining the aerodynamic characteristics of the aircraft model, sensor hardware characteristics data, and sensor installation position parameters with experimental controlled variable simulation experiments, and is used to correct the meteorological data of the flight area collected by different models of UAVs, sensors, and sensor installation positions in real time; the UAV swarm performance transfer matrix is ​​constructed by the output performance index, failure probability, probability of completing a single sub-task, and probability of completing a complex task of each UAV at each time point; the flight correlation adjustment map is constructed by combining the real-time position of each UAV and the flight correlation adjustment parameters at the corresponding position with a graph neural network.

[0012] Specifically, the task complexity for allocating resources includes:

[0013] Based on a pre-trained natural language processing model, named entity recognition and semantic role labeling are performed on complex task texts. Combined with a pre-built UAV task domain knowledge base, element standardization and disambiguation are performed to obtain a structured set of task elements.

[0014] The set of structured task elements is input into a rule-based decomposition engine. Based on the single drone executable principle defined by the upper limit of the output performance of the configured drone swarm, the complex task is decomposed into a single subtask and a single subtask set composed of the single subtasks is constructed.

[0015] Based on the correlation analysis algorithm combining temporal dependency, spatial proximity, and resource sharing corresponding to a single subtask set, a single subtask correlation matrix is ​​obtained. Temporal dependency is determined based on the execution order constraint of the single subtask, spatial proximity is calculated based on the inverse proportional function of the Euclidean distance of the target position of the single subtask, and resource sharing is determined based on the degree of satisfaction between the type of computational resources required by the single subtask and the amount of resources allocated for the corresponding type of resources.

[0016] Specifically, obtaining the task complexity used to allocate resource quantities also includes:

[0017] Based on the historical single subtask complexity and corresponding computational resource amount, a task complexity-computational resource mapping function is constructed. At the same time, based on the historical meteorological data of the corresponding flight area at the corresponding time point, an environment-adjustment resource mapping function is constructed to adjust the computational resource amount required to correct the flight status at the corresponding time point.

[0018] Each individual subtask is mapped to a vertex, and the task complexity-computation resource mapping function and the environment-adjustment resource mapping function are mapped to the corresponding vertices. At the same time, weighted edges are constructed based on the non-zero correlation degree in the single subtask correlation matrix. A weighted undirected graph model is constructed to represent the collaborative computational complexity between single subtasks and the expected computational complexity of real-time correction of the UAV's flight status during flight.

[0019] Specifically, obtaining the task complexity used to allocate resource quantities also includes:

[0020] Based on the weighted undirected graph model, combined with the maximum clique search algorithm and the temporal dependency of each individual subtask, the maximum fully cooperative subgraph in the weighted undirected graph at each time point is obtained, so as to determine the instantaneous cooperative computing resource requirements at each time point.

[0021] Based on the instantaneous collaborative computing resource requirements and the corresponding available computing resource types at each time point, the task computational complexity at each time point is calculated.

[0022] A weighted average algorithm is used to obtain the task complexity for allocating resources, based on the expected completion time of the complex task and the computational complexity of the task at each time point.

[0023] Specifically, the inversion yields a real-time flight correlation adjustment map of the drone swarm, including:

[0024] Based on the single subtask set and the single subtask association matrix, the world path of all UAVs during the mission execution is constructed, and the position point corresponding to each time point on the world path is used as the world origin of the world coordinate system of all UAVs at the corresponding time point.

[0025] The weighted undirected graph model is mapped onto the world coordinate system using the timestamp of each world origin and the execution timestamp of the single subtask corresponding to each vertex, to obtain the dynamic flight graph space of the UAV swarm.

[0026] Specifically, the process of acquiring the dynamic flight map space of a drone swarm includes:

[0027] Based on the preset flight path corresponding to each vertex of the weighted undirected graph model and the single UAV executable principle, the timestamp corresponding to each world origin on the world path and the estimated timestamp corresponding to each position point in the preset flight path of each vertex are discretized to the plane coordinate system corresponding to the world origin of the corresponding timestamp to obtain the target position point of each UAV in each plane coordinate system.

[0028] Simultaneously, the non-zero correlation degree in the single subtask correlation matrix is ​​used to construct weighted edges. According to the timestamp of each world origin and the temporal dependency, spatial proximity and resource sharing of the single subtask at the corresponding timestamp, it is decomposed into weighted edges between all UAV target position points in the two-dimensional plane coordinate system under the corresponding timestamp.

[0029] Based on all target location points in the plane coordinate system corresponding to each world origin and the weighted edges between all target location points, a static flight diagram of the UAV swarm under each world origin is obtained.

[0030] Specifically, the process of acquiring the dynamic flight map space of a drone swarm also includes:

[0031] Based on the static flight map of the UAV swarm at each world origin, combined with the world path and the preset flight path corresponding to each UAV, a dynamic flight map space of the UAV swarm is obtained.

[0032] Based on the single sub-task complexity of each UAV from the current target location to the next target location in the dynamic flight map space of the UAV swarm, combined with the corrected meteorological data of the flight area, the task complexity-computational resource mapping function and the environment-adjustment resource mapping function under the corresponding single sub-task, and the resource sharing of the corresponding time interval from the current target location to the next target location, a local resource optimization function and a global resource optimization function and a first resource constraint space are constructed. The global resource optimization function is constructed from the local resource optimization function and the local collaborative resource optimization function, and is used to characterize the computational resource requirements of the internal control parameters of each UAV from the current target location to the next target location, as well as the collaborative computational resource requirements of all UAVs to maintain a safe distance and correctly execute the corresponding single sub-task.

[0033] Specifically, the inversion yields a real-time flight correlation adjustment map of the drone swarm, which also includes:

[0034] Based on the probability of each UAV completing a single subtask within the corresponding time period from the current target location to the next target location in the dynamic flight map space of the UAV swarm, and the probability of all UAVs completing a single subtask within the corresponding time period, a local task completion probability function is constructed. By utilizing the temporal dependency and spatial proximity of the corresponding time period from the current target location to the next target location for each UAV, a second spatiotemporal constraint condition between the UAV task execution priority and the flight space distance is constructed.

[0035] Based on the resource optimization function corresponding to each UAV, the first resource constraint, the local task completion probability function, and the second spatiotemporal constraint, combined with the flight state parameters of each UAV at each target location and the Hidden Markov algorithm, the flight state probability transition matrix for each UAV to complete the corresponding task, the conditional probability function for successfully completing the corresponding single sub-task, and the global probability function for completing the complex task are obtained.

[0036] Based on local and global resource optimization functions, local task completion probability functions, flight state probability transition matrix for each UAV to complete a corresponding task, conditional probability function for successfully completing a single sub-task, and global probability function for completing complex tasks, a global multi-objective optimization function is constructed. A global constraint space is constructed using the first resource constraint condition and the second spatiotemporal constraint condition.

[0037] Specifically, the inversion yields a real-time flight correlation adjustment map of the drone swarm, which also includes:

[0038] Based on the global multi-objective optimization function and global constraint space, combined with a preset flight state adjustment coefficient library, and combined with multi-objective optimization search algorithm and twin simulation algorithm, the optimization search is performed with the objectives of minimizing the local resource optimization function and the global resource optimization function, as well as maximizing the local task completion probability function, the conditional probability function of each UAV successfully completing the corresponding single sub-task, and the global probability function of completing the complex task, in combination with the global constraint space. The optimal flight state correction adjustment coefficients of all UAVs in the UAV swarm dynamic flight map space from the current target position point to the next target position point are obtained.

[0039] The optimal flight state correction adjustment coefficient for each UAV from the current target location to the next target location is mapped to the local relative path connection between the corresponding UAV from the current target location to the next target location. At the same time, the complexity of the corresponding local relative path adjustment process is calculated based on the optimal flight state correction adjustment coefficient in the local relative path connection.

[0040] Based on the complexity of the corresponding local relative path adjustment process combined with the HSV color space, the color depth of the corresponding local relative path adjustment complexity is obtained.

[0041] Based on the local relative path of each UAV and the corresponding optimal flight state correction adjustment coefficient, and the color depth combined graph algorithm of the adjustment complexity mapping of the corresponding local relative path, a real-time flight association adjustment graph of all UAV swarms between adjacent world origins is constructed.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] This invention addresses the shortcomings of existing technologies by using a task analysis module to extract structured elements and construct correlation matrices for complex tasks. Combined with a task complexity assessment algorithm, it achieves precise decomposition of individual subtasks and adapts them to swarm capabilities, ensuring an inherent match between individual subtask allocation and UAV swarm performance. By constructing universal correction coefficients based on aircraft aerodynamic characteristics, sensor hardware characteristics, and installation location parameters, it improves the accuracy of meteorological data acquisition in different scenarios, providing accurate environmental condition support for subsequent stages. The invention utilizes a hidden Markov algorithm in the record analysis module to generate a UAV swarm performance transition matrix, quantifying the evolution of swarm performance, and enabling real-time monitoring of task completion probability and risk prediction. Furthermore, the invention constructs a dynamic flight graph space for the UAV swarm through an atmospheric analysis module, integrating multi-objective optimization functions and spatiotemporal constraints to obtain optimal flight state correction coefficients. Combined with a flight control module, it achieves real-time correction of the swarm's flight state, enabling deep collaboration between data and control logic at each stage. This effectively improves the overall probability of completing complex tasks, meets preset safety thresholds, and provides strong support for the reliable execution of complex tasks by UAV swarms. Attached Figure Description

[0044] Figure 1 This is a flowchart of the modular integrated control platform for unmanned aerial vehicles (UAVs) designed for complex tasks, as described in Embodiment 1 of the present invention.

[0045] Figure 2 This is a simplified schematic diagram of the flight static image of the unmanned aerial vehicle swarm in Embodiment 1 of the present invention;

[0046] Figure 3 This is a simplified schematic diagram of the dynamic flight pattern of an unmanned aerial vehicle (UAV) swarm according to Embodiment 1 of the present invention. Detailed Implementation

[0047] Example 1

[0048] Please see Figure 1 The present invention provides an embodiment of a modular integrated control platform for unmanned aerial vehicles (UAVs) oriented to complex tasks, comprising: a data acquisition module, a task analysis module, an atmospheric analysis module, a flight control module, and a record analysis module;

[0049] The acquisition module is used to acquire real-time flight status parameters, complex mission text, meteorological data of the flight area, output performance data and fault detection data of the configured UAV swarm, and perform timestamp alignment and filtering processing; the flight status parameters include at least flight attitude, flight heading, flight position and speed;

[0050] The task parsing module is used to obtain a single subtask set and a single subtask association matrix that can be constructed by a single subtask that can be completed by at most one drone, based on the complex task text and the task parsing algorithm, and to obtain the task complexity for allocating resources by combining the task complexity evaluation algorithm.

[0051] The atmospheric analysis module is used to invert the real-time flight association adjustment map of the UAV swarm by combining the meteorological data of the flight area after correction by the general correction coefficient, the real-time flight status parameters of each UAV, the single sub-task association matrix and the task complexity, the target optimization algorithm and the preset target optimization function.

[0052] It should be further explained that the objective optimization function in this embodiment is constructed from the resource allocation amount of the complex task, the probability of completing a single subtask, and the probability of completing the complex task. It is used to solve the real-time flight correlation adjustment map corresponding to the minimum resource allocation amount, the probability of completing a single subtask, and the maximum probability of completing the complex task.

[0053] It should be further explained that the meteorological data for the flight area in this embodiment includes at least indicated airspeed, vacuum speed, barometric altitude, climb rate, angle of attack, and sideslip angle. It should also be explained that the general correction coefficient in this embodiment is obtained through simulation experiments using the experimental control variable method, combining the aerodynamic characteristics of the aircraft, sensor hardware characteristics, and sensor installation location parameters. This coefficient is used to correct the meteorological data for the flight area collected by different types of UAVs, sensors, and sensor installation locations in real time. Furthermore, it should be explained that the flight correlation adjustment diagram in this embodiment is constructed by combining the real-time position of each UAV and the corresponding flight correlation adjustment parameters with a graph neural network. The flight correlation adjustment parameters in this embodiment include at least attitude control corrections for adjusting the aircraft's pitch, roll, and yaw angles; power output correction coefficients for adjusting the throttle percentage or motor speed of each power unit; a heading trajectory correction vector for indicating the local heading adjustment angle and distance; cooperative synchronization parameters for maintaining relative speed, position offset, and formation phase among multiple UAVs; and resource allocation weights for dynamically allocating computational resources between flight control, task calculation, and cooperative communication. These parameters together constitute a complete quantitative description of the real-time flight status adjustment of the UAV at a specific location.

[0054] It should be further explained that the meteorological data of the flight area in this embodiment is acquired through the configured SA-DADC atmospheric data computer. This device can calculate atmospheric data such as indicated airspeed, vacuum speed, barometric altitude, climb rate, angle of attack, and sideslip angle by collecting relevant signals from the matching atmospheric data sensor. It can also realize the heating and de-icing control and monitoring of the atmospheric wind vane sensor. This model is a general-purpose off-the-shelf product. By configuring different correction coefficients, it can be adapted to the use of matching atmospheric data sensors on different aircraft models and in different installation locations. Its main application scenarios include supersonic UAVs, target drones, and wind tunnel test benchmarking instruments.

[0055] The record parsing module is used to obtain the performance transition matrix of the UAV swarm by combining the output performance data and fault detection data with the Hidden Markov Algorithm.

[0056] It should be further explained that the UAV swarm performance transfer matrix in this embodiment is constructed from the output performance index, failure probability, probability of completing a single subtask, and probability of completing a complex task of each UAV at each time point. It should also be explained that the output performance index of the UAV at each time point in this embodiment is obtained by the acquisition module collecting real-time operational data such as the computing power utilization, power output efficiency, and navigation accuracy of the UAV. The failure probability is derived by the record parsing module by inputting the failure detection data obtained by the acquisition module into the Hidden Markov Algorithm and combining it with historical failure data and real-time operating status. It is used to construct the UAV swarm performance transfer matrix together with the probability of completing a single subtask and the probability of completing a complex task. This provides the flight control module with a quantitative basis for the real-time performance status and failure risk of the UAV, supports the flight control module in responding to the real-time flight correlation adjustment diagram to correct the flight status, and ensures that the probability of completing a single subtask and the probability of completing a complex task meet the preset task safety completion probability threshold.

[0057] The flight control module is used to respond to the real-time flight correlation adjustment diagram to make real-time corrections to the real-time flight status of at least one UAV, so that the probability of completing a single subtask and the probability of completing a complex task in the performance transfer matrix of the UAV swarm meet the preset task safe completion probability threshold in real time.

[0058] It is important to further explain that during the actual data collection process of UAVs, the differences in aerodynamic characteristics between different UAV models can lead to deviations in the meteorological data collected by sensors under the same flight conditions due to varying airflow field influences. Simultaneously, the inherent differences in the hardware characteristics of different sensors, such as sensitivity and temperature drift, can cause discrepancies between the collected data and the true values. Furthermore, the varying degrees of obstruction by the fuselage structure and airflow disturbances at different sensor installation locations can also result in deviations in the collected meteorological data. If these deviations are not corrected, the meteorological data of the flight area will fail to accurately reflect the actual atmospheric environment, thus affecting the accuracy of the real-time flight correlation adjustment map generated by the atmospheric analysis module. This interferes with the real-time correction of the UAV flight status by the flight control module, and ultimately may cause the completion probability of a single subtask and the completion probability of complex tasks in the UAV swarm performance transfer matrix to fail to meet the preset safe completion probability threshold, affecting the successful completion of complex tasks. Therefore, it is necessary to use a universal correction coefficient obtained from simulation experiments using the experimental control variable method, based on the aerodynamic characteristics of the UAV model, sensor hardware characteristics, and sensor installation location parameters, to perform real-time correction of the meteorological data collected from different UAV models, sensors, and sensor installation locations, in order to eliminate deviations and ensure data accuracy. It should be further explained that the process of obtaining the general correction coefficient in this embodiment includes:

[0059] A101. Obtain a complete set of parameters affecting the deviation of sensor data acquisition, including: obtaining airfoil parameters, aspect ratio parameters, and mass distribution parameters from the target UAV model, and integrating them into model parameters; obtaining sensitivity parameters and temperature drift characteristic parameters from the supporting air data sensor, and integrating them into sensor characteristic parameters; obtaining three-dimensional installation coordinate parameters and orientation angle parameters from the sensor installation configuration data, and integrating them into installation position parameters; performing standardization processing on all the above parameters to form a standardized set of acquisition correlation parameters.

[0060] A102. Build a standard wind tunnel experimental environment and set benchmark atmospheric conditions including standard temperature, standard pressure, and standard air flow velocity; based on the set of acquisition correlation parameters, conduct systematic simulation experiments using the control variable method, specifically including: keeping other parameters unchanged, sequentially changing the different curvature values of the airfoil parameters, different ratios of the aspect ratio parameters, different spatial positions of the sensor three-dimensional installation coordinate parameters, and different angle values of the orientation angle parameters, and simultaneously recording the indicated airspeed, true airspeed, barometric altitude, angle of attack, and sideslip angle meteorological data output by the sensor under each parameter combination.

[0061] A103. During the simulation experiment, synchronously collect the true values of each meteorological data through the standard measuring instruments supporting the wind tunnel test bench; for each parameter combination, calculate the difference between the sensor output value and the corresponding true value to obtain the deviation amount of each meteorological data channel, and the meteorological data channels at least include the indicated airspeed channel, true airspeed channel, and angle of attack channel; establish a mapping relationship database containing the corresponding relationship between parameter combinations, meteorological data channels, and deviation amounts.

[0062] A104. Based on the mapping relationship database, perform model fitting using the multiple regression algorithm to construct an atmospheric parameter correction model; the atmospheric parameter correction model uses model parameters, sensor characteristic parameters, and installation position parameters as independent variables and the deviation correction amounts of each meteorological data channel as dependent variables; by solving the regression coefficients of the model, obtain general correction coefficients applicable to different parameter combinations.

[0063] A105. According to the combination rules of UAV model, sensor model, sensor three-dimensional installation coordinates, and sensor orientation angle, perform unique coding on the general correction coefficients; construct a general correction coefficient library to store all the unique encodings, general correction coefficients, and the association relationships of meteorological data channels; the general correction coefficient library supports real-time querying and calling of the corresponding general correction coefficients by inputting UAV model, sensor model, and sensor installation position parameters. It should be further noted that the general correction coefficient library in this embodiment is connected to the atmospheric analysis module through a data interface to provide real-time correction services for the meteorological data of flight areas collected by different UAV models, different sensors, and different installation positions.

[0064] Next, a specific complete example is used to illustrate the whole process. The example is only for demonstrating the feasibility at the calculation level and does not represent the actual values. The specific value-taking method can be determined by those skilled in the art through simulation experiments or physical experiments. Exemplarily, in the UAV cluster for mountain geological disaster monitoring on the platform of this application, when carrying out the correction of sensor data deviation for mountain transmission line inspection, first extract the airfoil parameters, aspect ratio parameters, and mass distribution parameters from the configured multi-rotor inspection UAV models and integrate them into the model parameters. Extract the sensitivity parameter of 0.02 kPa per volt and the temperature drift characteristic parameter of 0.018 V per degree Celsius from the supporting atmospheric data sensors and integrate them into the sensor characteristic parameters. Extract the three-dimensional installation coordinate parameters of X0\Y0\Z8 cm and the orientation angle parameter of 90 degrees from the sensor installation configuration data and integrate them into the installation position parameters. Perform normalization and standardization processing on all parameters to form a standardized set of acquisition-related parameters. Subsequently, build a standard wind tunnel experimental environment, set the reference atmospheric conditions of a standard temperature of 25 degrees Celsius, a standard pressure of 1013 hPa, and a standard air flow velocity of 18 m / s, and carry out a systematic controlled variable simulation experiment based on the set of acquisition-related parameters strictly following the principle of single variable. First, keep the sensor characteristic parameters and the installation position parameters completely constant, and sequentially change the airfoil curvature values of the UAV to 0.03, 0.05, and 0.07, and simultaneously record the indicated airspeed of 22 m / s, 25 m / s, and 27 m / s, the true airspeed of 20 m / s, 23 m / s, and 25 m / s, the barometric altitude of 320 m\350 m\370 m, the angle of attack of 3 degrees\5 degrees\7 degrees, and the sideslip angle of 2 degrees\3 degrees\4 degrees of the complete meteorological data output by the sensor at each curvature value, then keep the model parameters and the sensor characteristic parameters constant, and sequentially adjust the aspect ratio parameter ratio to 2.5, 3.5, and 4.5, and simultaneously record the full set of meteorological data for the corresponding parameter combinations. Subsequently, fix the model parameters and the sensor characteristic parameters, and sequentially change the sensor three-dimensional installation coordinate parameters to X6\Y0\Z8 cm, X0\Y6\Z8 cm, and X6\Y6\Z8 cm, and simultaneously collect the sensor output data at each spatial point. Finally, keep the model parameters and the sensor characteristic parameters unchanged, and sequentially adjust the sensor orientation angle parameter to 75 degrees and 105 degrees, and completely record the meteorological data results corresponding to each angle value. Throughout the simulation experiment, the true values of each meteorological data are simultaneously collected through the standard speedometer, high-precision barometer, and angle measuring instrument supporting the wind tunnel test bench. For each parameter combination, calculate the difference between the sensor output value and the true value one by one, and accurately obtain the deviation amounts of each meteorological data channel, namely, the deviation amount of the indicated airspeed channel of 2 m / s, 1.5 m / s, and 1 m / s, the deviation amount of the true airspeed channel of 1.8 m / s, 1.2 m / s, and 0.9 m / s, and the deviation amount of the angle of attack channel of 0.6 degrees, 0.4 degrees, and 0.2 degrees, and completely establish a mapping relationship database that includes the one-to-one correspondence relationship among the parameter combination, meteorological data, and channel deviation amount.Based on this mapping database, a multivariate regression algorithm is used to fit the model and build an atmospheric parameter correction model. The model takes a three-dimensional independent variable matrix consisting of standardized aircraft parameters, sensor characteristic parameters, and installation location parameters as the core input, and a one-dimensional dependent variable matrix consisting of the deviation corrections of the indicated airspeed channel, vacuum speed channel, and angle of attack channel as the output. The core logic of the model execution is to normalize the three types of independent variable parameters into a multivariate linear regression calculation matrix according to the dimensions, take the deviation of each meteorological data channel as the regression target value, and solve the model regression coefficients iteratively by the least squares method. Finally, a universal correction coefficient suitable for different parameter combinations is obtained. Among them, the parameter combination of airfoil curvature 0.05, aspect ratio 3.5, sensor three-dimensional coordinates X0 / Y0 / Z 8 cm, and heading angle 90 degrees corresponds to an indicated airspeed correction coefficient of 1.02, a vacuum speed correction coefficient of 1.01, and an angle of attack correction coefficient of 0.98. Following a fixed combination rule of UAV model, sensor, 3D installation coordinates, and sensor orientation angle, all general correction coefficients are uniquely coded. A general correction coefficient library is built to fully store the relationships between all unique codes, general correction coefficients, and meteorological data channels. This library is connected in real time to the atmospheric analysis module via a dedicated data interface. When conducting actual inspections of power transmission lines in mountainous areas, the sensor model, actual 3D installation coordinates, and orientation angle parameters of the UAV performing the inspection task are input into the coefficient library. The corresponding general correction coefficients can then be queried and called in real time to accurately correct the meteorological data collected by the UAV in the flight area. This effectively offsets the acquisition errors caused by differences in aerodynamic characteristics of different UAV models, deviations in sensor hardware characteristics, and different installation locations, ensuring that the corrected meteorological data accurately reflects the actual atmospheric environment of the inspection area.

[0065] It should be further explained that the task complexity for obtaining the amount of resources to be allocated in this embodiment includes:

[0066] B101. Based on a pre-trained Chinese BERT model, named entity recognition and semantic role annotation are performed on complex task texts. Key actions, spatial locations, time constraints, target objects, and performance indicators are extracted. Combined with a pre-built UAV task domain knowledge base, element standardization and disambiguation are performed to obtain a structured task element set. The training dataset of the pre-trained Chinese BERT model in this embodiment comes from text data such as task instructions, execution reports, and technical documents accumulated from historical mountain geological disaster monitoring tasks. Key actions, spatial locations, time constraints, and other entity and semantic role information are manually annotated by domain experts to form an annotated sample set. The training method is supervised fine-tuning based on the BERT-Base pre-trained model. The loss function is the cross-entropy loss function. The training parameters are set as batch size 32, learning rate 2e-5, training epochs 10, and weight decay coefficient 0.01.

[0067] B102. The structured task element set is input into a rule-based decomposition engine. Based on the single drone executable principle defined by the upper limit of the output performance of the configured drone swarm, the complex task is decomposed into a single subtask, and a single subtask set composed of single subtasks is constructed. In this embodiment, the decision tree classification algorithm used to extract core requirement parameters has a training dataset that is a sample set of task element features and core requirement parameters corresponding to historical single subtasks. The samples include features such as task actions and performance indicators, as well as requirement parameters such as matching target load and expected execution time. The annotation is completed by technicians based on their task execution experience. The training method is supervised training. The information gain criterion is used for feature selection. The loss function is the logarithmic loss function. The training parameters are set as follows: maximum depth 8, minimum number of sample segments 20, minimum number of sample leaf nodes 10, and feature sampling ratio 1.0.

[0068] It should be further explained that the process of decomposing a complex task into a single subtask based on the single drone executable principle defined by the upper limit of the output performance of the configured drone swarm in this embodiment includes:

[0069] Based on the hardware specification documents, factory performance test reports, and real-time status monitoring data of the configured drone swarm, the upper limit parameters of the output performance of each drone are obtained. The Min-Max normalization algorithm is used to normalize the upper limit parameters of the output performance, resulting in a set of drone performance thresholds with uniform dimensions. The upper limit parameters of the output performance include maximum payload, endurance, flight speed range, operating radius, peak computing power of computing unit, sensor detection accuracy, and number of concurrent task processing.

[0070] Based on the set of performance thresholds for a single UAV, the executable principles for a single UAV are defined, forming a system of executable principles for a single UAV that includes quantitative constraints. The executable principles for a single UAV are as follows: the target load of a single subtask does not exceed the maximum payload of the UAV; the execution time does not exceed the flight time; the operating distance does not exceed the operating radius; the required computing power does not exceed the peak computing power of the computing unit; the parallelism of the task does not exceed the number of concurrent processing tasks; and it does not exceed the preset minimum granularity threshold for a single subtask to avoid excessive splitting of a single subtask leading to a surge in collaborative overhead. Next, a specific complete example will be used to illustrate the entire process. The example is only to illustrate the feasibility at the computational level and does not represent the actual values. The specific values ​​can be determined by those skilled in the art through simulation experiments or physical experiments. For example, for the drone swarm operation scenario of inspecting power transmission lines in mountainous areas, the technicians first extract and process the single-unit performance threshold set by the Min-Max standardization algorithm based on the hardware specification document, factory performance test report and real-time status monitoring data of the multi-rotor drone. This set serves as the core input for determining the executable principle of a single drone. The specific values ​​are: maximum payload of 2 kg, flight time of 120 minutes, operating radius of 8 km, peak computing power of computing unit of 100 GFLOPS, and number of concurrent tasks of 3. At the same time, combined with the task coordination efficiency requirements, the preset minimum granularity threshold of a single sub-task is set to cover a power transmission line length of not less than 5 km. The execution logic structure of this principle system is based on the set of performance thresholds for a single UAV as the judgment benchmark and the core requirement parameters of a single subtask as the judgment object. Compliance verification is completed by matching quantitative constraints one by one. The specific process is to extract the core requirement parameters of the single subtask to be judged, such as the target load, execution time, operating distance, required computing power, task parallelism, and coverage line length. Then, it is verified in turn whether the target load does not exceed the maximum payload, whether the execution time does not exceed the endurance time, whether the operating distance does not exceed the operating radius, whether the required computing power does not exceed the peak computing power of the computing unit, whether the task parallelism does not exceed the number of concurrent processing tasks, and finally, whether the coverage line length of the subtask is not less than the preset minimum granularity threshold. Only when all constraints are met can the single subtask be judged to meet the executable principle and the judgment result is output. Otherwise, the judgment is not passed.Taking the sub-task of inspecting transmission line towers in section A of the mountainous area as an example, its core requirements are: target load of 1.5 kg, execution time of 80 minutes, working distance of 6 km, required computing power of 60 GFLOPS, parallel task quantity of 2, and coverage line length of 6 km. After inputting these parameters into the judgment system, it was found that none of the requirements exceeded the performance threshold of the UAV, and the coverage line length met the minimum granularity threshold requirement. The final judgment result was passed, and this sub-task can be executed independently by a single UAV. However, another sub-task with a coverage line length of only 3 km, even though its target load of 1.2 kg, execution time of 60 minutes, and other requirements met the threshold requirements, failed to meet the minimum granularity threshold. The final judgment result was failed, and it needs to be merged with the adjacent sub-tasks and re-judged.

[0071] Based on the structured task element set, a decision tree classification algorithm is used to extract the core requirement parameters corresponding to each task element. The core requirement parameters are then mapped one-to-one with the parameter types of the single UAV performance threshold set to construct a task requirement-performance threshold mapping table. The core requirement parameters include task load, expected execution time, target operation distance, computing resource requirements, sensor accuracy requirements, and parallel execution requirements. The decision tree classification algorithm achieves accurate extraction and classification by learning the correspondence between task element features and core requirement parameters.

[0072] The structured task element set and task requirement-performance threshold mapping table are input into a rule-based decomposition engine. The decomposition engine has a built-in recursive segmentation algorithm, which first splits the complex task into several first-level task units according to the key action dimension. Then, based on the task requirement-performance threshold mapping table, the rule matching and verification algorithm verifies whether the core requirement parameters of each first-level task unit meet the single UAV executable principle. The recursive segmentation algorithm is used to realize the hierarchical decomposition of the task, and the rule matching and verification algorithm completes the compliance judgment by comparing the core requirement parameters with the quantitative constraints.

[0073] For first-level task units whose core requirement parameters, verified by the rule-matching verification algorithm, all meet the single-UAV executable principle, they are directly identified as single subtasks. For first-level task units whose core requirement parameters exceed the single-UAV executable principle, they are further divided into second-level task units according to spatial location or execution dimension. The rule-matching verification algorithm is then used to verify whether the core requirement parameters of the second-level task units meet the single-UAV executable principle. This splitting-verification process is repeated until all split task units meet the single-UAV executable principle. In this embodiment, the execution dimension refers to the sequence of steps or processes in the execution of a task. Specific technical means include: based on the action sequence description in the task elements, a method of constructing a directed acyclic graph is used to analyze and identify the logical dependencies and sequential relationships between actions; based on a preset task process template, linear temporal logic is used to formally model complex actions and split them into atomic operation steps with a strict sequential order; through a recursive segmentation algorithm, based on the identified action dependencies, complex task units that exceed performance constraints are decomposed layer by layer into a sequence of sub-steps that meet the sequential execution requirements.

[0074] Based on all the task units that meet the single UAV executable principle obtained after decomposition, the target object, time constraints, spatial location, performance indicators, and resource requirement attribute information of each task unit are extracted. The attribute format is unified by a data standardization alignment algorithm, and a single subtask set composed of all single subtasks is constructed. The single subtask set contains the unique identifier, attribute information, and association relationship of each single subtask with the original complex task. The data standardization alignment algorithm realizes the unification of attribute formats of different task units through preset attribute templates.

[0075] B103. Based on the temporal dependency, spatial proximity, and resource sharing of a single subtask set, a correlation matrix for a single subtask is obtained using a correlation analysis algorithm. Temporal dependency is determined based on the execution order constraints of the single subtask; spatial proximity is calculated based on the inverse proportional function of the Euclidean distance between the target locations of the single subtasks; and resource sharing is determined based on the degree to which the required computing resource types of a single subtask satisfy the allocation of corresponding resource types. It should be further noted that in this embodiment, the degree of resource sharing satisfaction is quantitatively defined using a fuzzy comprehensive evaluation method. Specific technical means include: based on various types of computing resources... The system calculates the total global allocation and real-time occupied resources, and uses a resource monitoring subsystem to obtain the remaining available resources. For any two individual subtasks that require overlapping sets of resource types, a resource type matching algorithm based on set theory is used for identification. For each overlapping resource type, a ratio analysis method is used to calculate the satisfaction level of that resource type, i.e., the ratio of the remaining available resources to the total demand for that resource type. Based on the preset importance weight of each overlapping resource type in task collaborative execution, a weighted average algorithm is used to fuse the satisfaction levels of each resource type, generating a comprehensive quantitative value representing the strength of resource sharing between two individual subtasks. The preset importance weights in this embodiment are constructed using the contribution values ​​of the corresponding individual subtasks to the complex task. The association analysis algorithm used in this embodiment to construct the association matrix of a single subtask uses a training dataset consisting of historical records of temporal constraints, spatial locations, and resource requirements for a single subtask. This dataset covers data such as subtask execution order, geographical coordinates, and resource consumption. The training method is unsupervised, learning association rules between subtasks through frequent itemset mining. The loss function is a confidence loss function, and the training parameters are set to minimum support of 0.2, minimum confidence of 0.7, and maximum itemset length of 5. The fuzzy comprehensive evaluation method used in this embodiment to quantify the degree of resource sharing satisfaction uses a training dataset consisting of historical resource allocation effect evaluation samples. This dataset includes corresponding data on resource type, remaining available allocation, total demand, and comprehensive evaluation results of sharing. The training method is semi-supervised, combining... Expert experience is used to define the fuzzy membership function, and the loss function is the fuzzy proximity loss function. The training parameters are set as follows: the membership function type is trapezoidal, the fuzzy evaluation matrix dimension is 5×3, and the weight iteration count is 50. In this embodiment, the resource type matching algorithm based on set theory for identifying overlapping resource types does not require additional training and is directly constructed based on set theory rules. The effectiveness of the matching rules is verified only by using the resource type list dataset of historical subtasks. The ratio analysis method for calculating the satisfaction of a single type of resource and the weighted average algorithm for integrating the overall satisfaction are both deterministic algorithms that do not require training. They only need to use the preset resource importance weights obtained from the statistical analysis of the contribution values ​​of historical single subtasks to complex tasks as parameters to ensure that the calculation results fit the actual needs of the task.

[0076] It should be further explained that the single subtask association matrix is ​​constructed in this embodiment as follows:

[0077] S301. Based on the preset complete completion time of a complex task, a time discretization algorithm is used to divide it into a continuous sequence of time points.

[0078] S302. Based on a structured task element set, a temporal rule matching engine is used to parse the execution order constraints of each individual subtask and generate temporal dependency weights. Non-zero temporal dependency weights are assigned when there are mandatory sequential execution constraints, and zero weights are assigned when there are no sequential constraints. The training dataset for the temporal rule matching engine in this embodiment comes from a set of time-constraint annotation samples from historical mountain geological disaster monitoring and collaborative mapping tasks. This dataset includes instruction texts for sequential and parallel execution of subtasks, task flow documents, and manually annotated temporal constraint type labels. The training method is supervised fine-tuning, iteratively optimizing the engine based on a preset temporal rule library. The cross-entropy loss function is used. Training parameters are set to a batch size of 32, a learning rate of 2e-5, 10 training epochs, and a weight decay coefficient of 0.01. For example, in a landslide monitoring task, the temporal dependency weight of subtask A on B is assigned as 0.5.

[0079] S303. Based on the geographic coordinates of the target location of each single subtask in the structured task element set, the Euclidean distance algorithm is used to calculate the spatial distance between any two single subtasks, and the spatial correlation strength value is mapped to the target location using a preset inverse proportional function.

[0080] S304. Based on the task complexity-computation resource mapping table, extract the types of computational resources required for each single subtask, and calculate the resource sharing association strength value between any two single subtasks using a resource satisfaction evaluation algorithm. It should be further explained that the construction method of the task complexity-computation resource mapping table in this embodiment includes the following steps: Based on historical task execution records, extract the complexity feature vectors of each single subtask, and use the K-means clustering algorithm to perform cluster analysis on the complexity feature vectors, dividing historical single subtasks into several categories with similar resource consumption patterns; the complexity feature vectors include action type, number of target objects, spatial span, precision requirements, and time urgency; for each task category, use a multiple linear regression algorithm, with the complexity feature vector as the independent variable and the actual consumed CPU computing power, memory usage, communication bandwidth, and processing latency as the dependent variable, to train a resource demand prediction model for that category; integrate the prediction models of all categories to construct a mapping relationship table from task complexity to multi-dimensional computational resource requirements. This table supports quick querying and outputting the estimated computational resource values ​​for each dimension by inputting the complexity features of a new single subtask. In this embodiment, the training dataset for the multiple linear regression algorithm comes from two sources: first, the complexity feature vectors and resource consumption data of each subtask after clustering; and second, the complexity feature vectors and corresponding computational resource data of historical single subtasks, such as the feature vector of the LiDAR scanning task and the resource consumption of 8.75 gigaflops per second. The training method is supervised training. The loss function is the mean squared error loss function. The training parameters are set to a regularization coefficient of 0.01, 50 training epochs, and a learning rate of 0.001. For example, when constructing the mapping function, the intercept term coefficient is set to 1.2, and the coefficients for each feature dimension are 0.5, 0.8, 3.0, and 0.1. The training dataset for the resource satisfaction evaluation algorithm in this embodiment comes from the historical total resource allocation and real-time occupied resource data collected by the resource monitoring subsystem, as well as the contribution value data of a single subtask to the complex task. For example, in the landslide monitoring task, the remaining available computing power is 70 gigaflops per second, the total demand is 86 gigaflops per second, and the subtask contribution value weight is 0.5. The training method is a no-training-process approach, directly calculating based on preset rules. There is no loss function. Training parameters are set to resource importance weights, for example, a weight of 0.5 for a weighted average.

[0081] S305. For any two single subtasks in a single subtask set, at each time point, the task coordination computational complexity is calculated using a weighted summation algorithm based on the temporal dependency weight, spatial association strength value, and resource sharing association strength value.

[0082] S306. Arrange the task coordination computational complexity of all pairs of single subtasks in the single subtask set at each time point in the order of single subtask pairs to construct the single subtask association matrix.

[0083] Next, a specific complete example will be used to illustrate the entire process. This example is only to illustrate the feasibility at the computational level and does not represent actual values. Specific values ​​can be determined by those skilled in the art through simulation experiments or physical experiments. For example, in the assumed geological disaster landslide monitoring task of this embodiment, three single sub-tasks are set: sub-task A is to perform multispectral imaging of the landslide body, sub-task B is to perform lidar scanning of the crack area, and sub-task C is to transmit data in real-time throughout the entire period. First, based on the preset total task duration of 2 hours, a time discretization algorithm is used to divide the time into 24 consecutive time point sequences at 5-minute intervals. Its time sequence rule matching engine determines that sub-task A must complete data acquisition before sub-task B begins; therefore, the time sequence dependency weight of A on B is assigned to 0.5, and the weight between other tasks without a mandatory order is 0.

[0084] In the spatial correlation calculation, based on the target coordinates of subtask A (118.75°E, 32.07°N) and subtask B (118.76°E, 32.06°N), the Euclidean distance algorithm was used to calculate the ground distance between the two points as 1.2 km. This distance was then mapped to a spatial correlation strength value using a preset inverse proportional function, with the proportionality coefficient set to 1. The calculated strength value was 0.833.

[0085] For resource sharing assessment, a task complexity-computation resource mapping table is first constructed based on historical task records. The K-means clustering algorithm is used to cluster historical subtasks according to feature vectors such as action type and spatial span, with a cluster size of 3. Subsequently, a multiple linear regression algorithm is used to train a prediction model for each category, with a model regularization coefficient set to 0.01. According to this table, the total image processing computing power requirement for the three current subtasks is 86 gigaflops per second, while the remaining available computing power displayed by the resource monitoring subsystem is 70 gigaflops per second. The ratio analysis method is used to calculate the single-category resource satisfaction as 70 divided by 86, approximately equal to 0.814. Based on the contributions of subtasks A and B to the overall monitoring (0.6 and 0.4 respectively), a preset importance weight of 0.5 is calculated. Then, a weighted average algorithm is used to fuse the results to obtain a resource sharing association strength value of 0.814.

[0086] Finally, at each 5-minute time point, for any two subtasks, the temporal dependency weight, spatial association strength value, and resource sharing association strength value are weighted and summed according to preset weights of 0.4, 0.3, and 0.3 to obtain the task coordination computational complexity at that moment. The subtasks are then arranged in order to construct a single subtask association matrix with a dimension of 3×3×24.

[0087] B104. Based on the single subtask association matrix, extract the task coordination computational complexity of each pair of single subtasks at each time point, and combine the preset time point weights to calculate the total task coordination computational complexity using a weighted summation algorithm.

[0088] B105. Based on the task complexity-computation resource mapping function, extract the flight control computation requirements of each single subtask at each time point, and obtain the control complexity of a single subtask through an accumulation algorithm.

[0089] B106. Summing up the control complexity of all individual subtasks yields the total control complexity. Summing up the total coordination computation complexity of all pairs of individual subtasks yields the total coordination computation complexity. Adding the two together yields the computational complexity of the complex task.

[0090] It should be further explained that the total control complexity in this embodiment refers to the total amount of basic computing resources required for the independent execution of all individual subtasks, used to quantify the total basic computing power requirements supporting the autonomous flight and mission execution of each UAV. It is implemented through an accumulation algorithm. The specific technical means are as follows: based on the task complexity-computation resource mapping function, the flight control computational requirements of each individual subtask at each time point are extracted; the control complexity of each individual subtask is obtained through the accumulation algorithm; and then the control complexities of all individual subtasks are summed.

[0091] It should be further explained that the total coordination computational complexity in this embodiment refers to the total amount of additional computing resources generated by the collaborative association between all pairs of individual subtasks. It is used to quantify the collaborative computational overhead added by the UAV swarm to maintain the collaborative cooperation between tasks (such as data interaction, state synchronization, and conflict resolution). It is implemented through a weighted summation algorithm. The specific technical means is as follows: based on the single subtask association matrix, the total task coordination computational complexity of each pair of individual subtasks in the entire task cycle is extracted, and then the total task coordination computational complexity of all pairs of individual subtasks is summed.

[0092] It should be further explained that the computational complexity of the complex task in this embodiment is the sum of the total control complexity and the total coordination computational complexity. This sum represents the overall load of all computing resources required to complete the entire complex task, providing a core quantitative decision-making basis for subsequent global resource optimization allocation, scheduling strategy formulation, and system capacity planning. This is achieved through addition, specifically by directly adding the summed total control complexity to the total coordination computational complexity.

[0093] The timing rule matching engine determines the type of execution order constraint between individual subtasks by parsing the time constraint information and logical association description in the structured task element set. This includes tasks that must be executed serially in a specific order, tasks that can be executed in parallel without explicit order constraints, and tasks that are triggered after certain conditions are met.

[0094] The resource satisfaction evaluation algorithm is implemented through the following technical means: obtaining the total allocation and occupied amount of each type of computing resources, calculating the remaining available allocation amount, comparing the resource type overlap of any two single subtasks, and combining the resource type overlap degree with the degree of satisfaction of remaining resources to quantify and generate a resource sharing association strength value.

[0095] It should be further explained that the role of timing dependency in this embodiment is to clarify the logical sequence of execution between individual subtasks, avoiding problems such as task flow breakage and failure to achieve objectives due to disordered execution order of individual subtasks. It also provides a core basis for subsequent task scheduling, time window allocation, and resource timing optimization for the UAV swarm, ensuring the coherence and logic of the overall execution of complex tasks. The specific technical means to achieve this is as follows: first, extract the time constraint information and logical relationship description corresponding to each individual subtask from the structured task element set, such as descriptions like initiating a strike mission after completing target area reconnaissance or simultaneously conducting signal monitoring in three areas; then, match these descriptions using preset timing rules. The engine analyzes this information to determine the type of execution order constraints between individual subtasks, including those that must be executed sequentially in a specific order, those that can be executed in parallel without explicit order constraints, and those that are triggered after meeting specific conditions. Among these, the cases where execution must be sequentially in a specific order include situations where the output of a preceding individual subtask is a prerequisite for the execution of a subsequent individual subtask. Finally, the determination results are quantified into numerical representations that can be incorporated into the correlation analysis algorithm. When there are mandatory sequential execution constraints, corresponding temporal dependency weight values ​​are assigned to reflect the correlation strength. When there are no order constraints, the weight values ​​are set to zero, thereby accurately capturing the temporal dependencies between individual subtasks and providing core temporal dimension support for the construction of the correlation matrix of individual subtasks.

[0096] It should be further explained that in this embodiment, the geographic coordinates of the target location of each single subtask are extracted based on the structured task element set. The Euclidean distance between the target locations of any two single subtasks is calculated using a geographic information processing algorithm. This distance is then substituted into a preset inverse proportional function to generate a spatial association strength value. The types of computing resources required for each single subtask are extracted from the structured task element set and the task complexity-computation resource mapping table. The total allocation and occupied amount of each type of computing resource are obtained, and the remaining available allocation amount is calculated. The overlap of resource types between any two single subtasks is compared to determine the degree of satisfaction of the remaining amount of overlapping resources. The resource sharing association strength value is generated by combining the resource type overlap and satisfaction degree. The spatial association strength value and the resource sharing association strength value are respectively incorporated into the association analysis algorithm to provide spatial and resource dimension numerical support for the association degree calculation of corresponding single subtask pairs in the single subtask association matrix.

[0097] Next, a specific complete example will be used to illustrate the entire process. This example is only to illustrate the feasibility at the computational level and does not represent actual values. Specific values ​​can be determined by those skilled in the art through simulation or physical experiments. For example, in the mountainous collaborative mapping task addressed in this embodiment, the system processes two single sub-tasks: sub-task A performs orthophoto acquisition and sub-task B performs lidar scanning. First, based on a preset sequence of three time points, the weights of each time point are set to 0.3, 0.4, and 0.3, respectively. According to the single sub-task correlation matrix, the computational complexity of task coordination between sub-task A and sub-task B at the three time points is extracted as 1.2, 1.5, and 0.8, respectively. Using a weighted summation algorithm, the complexity of each time point is multiplied by its corresponding weight and then summed to obtain the total computational complexity of task coordination between sub-task A and B as 1.2. Subsequently, based on the task complexity to computational resource mapping function, the flight control computational requirements of sub-task A at the three time points are extracted as 2.0, 2.5, and 1.8, respectively, and for sub-task B as 1.5, 2.0, and 1.2, respectively. The control complexity of subtask A is obtained as 6.3 and that of subtask B as 4.7 using an accumulation algorithm. Adding these two together yields a total control complexity of 11.0. Finally, the total control complexity of 11.0 is added to the total coordination computational complexity of 1.2, resulting in a computational complexity of 12.2 for the complex task. This computational complexity characterizes the overall computational resource load required to complete the entire collaborative mapping task in the mountainous area.

[0098] B107. Based on the historical single subtask complexity and corresponding computational resource quantity, construct a task complexity-computational resource mapping function. At the same time, based on the historical meteorological data of the corresponding flight area at the corresponding time point, construct an environment-adjustment resource mapping function to adjust the computational resource quantity required to correct the flight status at the corresponding time point.

[0099] It should be further explained that when the UAV swarm in this embodiment performs complex tasks, the computational resource requirements and complexity of a single subtask exhibit historical regularity, while the computational resources required for flight status adjustments fluctuate dynamically with changes in meteorological data in the flight area. Both directly affect the accurate allocation of platform computational resources and the stability of task execution. Therefore, it is necessary to construct two mapping functions to achieve advance prediction and dynamic adaptation of resource requirements. Specifically, a task complexity-computational resource mapping function is constructed based on the historical complexity of a single subtask and the corresponding amount of computational resources. This function utilizes the correlation between historical task complexity and resource consumption to establish a quantitative correspondence between the complexity characteristics of a single subtask and the required amount of computational resources, enabling rapid prediction of new tasks. This system addresses the fundamental computing resource requirements of tasks, preventing insufficient supply or waste due to inaccurate resource forecasting. It also constructs an environment-adjustment resource mapping function based on historical meteorological data of flight areas at corresponding time points and the amount of computing resources required for flight status correction. This function captures the inherent correlation between changes in meteorological data and the resource consumption for flight status adjustments, enabling the prediction of additional computing resource requirements for flight status correction under different weather conditions. This avoids untimely adjustments due to resource shortages caused by weather fluctuations. Together, these two aspects provide a quantitative basis for the advance planning and dynamic scheduling of platform computing resources, ensuring that computing resources can meet the basic requirements of a single sub-task while adapting to additional adjustment needs brought about by weather changes, guaranteeing the smooth execution of complex tasks and flight control.

[0100] Next, a specific complete example will be used to illustrate the entire process. This example is only to illustrate the feasibility at the computational level and does not represent actual values. Specific values ​​can be determined through simulation or physical experiments conducted by those skilled in the art. For example, in the hypothetical geological disaster monitoring task scenario, a task complexity-computational resource mapping function is constructed. This function is based on historical records and implemented using a multiple linear regression algorithm. Specifically, the input is set as the complexity feature vector of a single subtask, which includes four dimensions: action type, target area area, resolution requirement, and terrain undulation. For example, the input vector for a lidar scanning task is the scanning action type code value 2, an area of ​​5 square kilometers, a resolution requirement of 0.05 meters, and a terrain undulation of 30 degrees. The algorithm model coefficients are set with an intercept term of 1.2, and the coefficients for the four dimensions are 0.5, 0.8, 3.0, and 0.1, respectively. Through linear calculation, the estimated basic computational resources for this subtask are 1.2 + 0.5 × 2 + 0.8 × 5 + 3.0 × 0.05 + 0.1 × 30 = 8.75 gigaflops per second. Simultaneously, an environment-adjustment resource mapping function was constructed. This function employs a feedforward neural network with a single hidden layer (8 dimensions) and ReLU activation function. The input is a vector of meteorological data for the flight area, including three dimensions: wind speed (10 m / s), turbulence intensity (0.2), and temperature gradient (0.5 degrees Celsius per 100 meters). The weight matrix trained on this network maps the data to the hidden layer before outputting it, ultimately yielding an additional 1.8 gigaflops per second of computational resource adjustment required by the flight control module to maintain attitude stability under the current meteorological conditions. Together, these two functions provide the platform with a total of 10.55 gigaflops per second of resource allocation data when planning scanning tasks.

[0101] B108. Map each single subtask to a vertex, and map the task complexity-computation resource mapping function and the environment-adjustment resource mapping function to the corresponding vertex. At the same time, construct weighted edges based on the non-zero correlation degree in the correlation matrix to construct a weighted undirected graph model that represents the expected computational complexity of collaborative computation between single subtasks and the real-time correction of the UAV's flight status during flight.

[0102] It should be further explained that in this embodiment, a single subtask is the basic unit for executing a complex task. Mapping it to a vertex enables a structured abstraction of the task entity. The task complexity-computational resource mapping function quantifies the basic computational resources required by the single subtask itself, while the environment-adjustment resource mapping function captures the additional computational resources required for flight status corrections caused by changes in meteorological data. Both are embedded in the vertex, allowing each vertex to fully represent the comprehensive computational resource characteristics of a single subtask under static requirements and dynamic environments. Simultaneously, the collaborative computational complexity between single subtasks is reflected by the non-zero correlation degree in the correlation matrix. Based on this, weighted edges are constructed to quantify the collaborative correlation strength in temporal, spatial, and resource dimensions. Finally, through the combination of vertices and weighted edges, the dispersed resource requirements of single subtasks and their collaborative... The relationships are integrated into a unified weighted undirected graph model, enabling a panoramic quantitative representation of the computational requirements of a single subtask, the collaborative computational consumption between subtasks, and the computational overhead of dynamic environmental adjustments. The motivation lies in the fact that existing methods struggle to simultaneously integrate the complex relationships between the basic resources of a single subtask, dynamic environmental resources, and collaboratively related resources. This leads to problems such as a disconnect between static estimates and dynamic requirements, and the neglect of collaborative consumption in computational resource allocation. By constructing this model, multi-dimensional computational complexity can be transformed into intuitive graph structure features, providing a unified quantitative analysis carrier for subsequent resource optimization allocation, load balancing scheduling, and risk prediction. This ensures that computational resources can not only meet the basic requirements of a single subtask but also adapt to the additional consumption brought about by collaborative interactions and environmental changes, thereby improving the stability and efficiency of complex task execution.

[0103] B109. Based on the weighted undirected graph model, combined with the maximum clique search algorithm and the temporal dependency of each individual subtask, the maximum fully cooperative subgraph in the weighted undirected graph at each time point is obtained to determine the instantaneous cooperative computing resource requirement at each time point; the instantaneous cooperative computing resource requirement includes the flight control computing requirement corresponding to each individual subtask and the coordination control computing requirement between individual subtasks corresponding to non-zero weight edges in the weighted undirected graph.

[0104] It should be further explained that the process of determining the instantaneous collaborative computing resource requirements at each point in time in this embodiment includes:

[0105] S601. Based on the time constraint information and logical association description of each individual subtask in the structured task element set, configure an execution time window for each individual subtask.

[0106] S602. The weighted undirected graph model is processed using the maximum clique search algorithm to identify the maximum complete subgraphs that overlap within the execution time window and whose weights of all edges between vertices are greater than a preset correlation threshold. These subgraphs are then used as the maximum complete cooperative subgraphs at the corresponding time points. It should be further noted that the correlation threshold in this embodiment is set by those skilled in the art based on the accuracy requirements of the maximum clique search algorithm.

[0107] S603. For each vertex in the maximum fully cooperative subgraph, call the task complexity-computation resource mapping function it carries to obtain the basic flight control computation requirements, and perform real-time environment correction based on the environment-adjustment resource mapping function to obtain the corrected flight control computation requirements.

[0108] S604. Based on the weight values ​​of each edge in the maximum fully cooperative subgraph, and combined with the data interaction frequency and state synchronization requirement parameters between individual subtasks, a weighted product algorithm is used to calculate the coordination control computation amount between individual subtasks.

[0109] S605. Sum the corrected flight control computational requirements of all vertices in the maximum fully cooperative subgraph with the coordination control computational requirements of all edges to obtain the instantaneous cooperative computational resource requirements at the corresponding time point.

[0110] S606. Repeat S602 to S605 for all time points within the execution time range of the complex task to obtain the instantaneous collaborative computing resource demand sequence within the complete task cycle.

[0111] The configuration of the execution time window is achieved based on the following technical means: extracting the start time, end time and sequential execution constraints of each single subtask from the structured task element set, and allocating a determined time interval to each single subtask through a time window allocation algorithm, ensuring that the time windows of single subtasks with sequential execution constraints do not overlap, while the time windows of single subtasks without sequential constraints are allowed to overlap.

[0112] Specifically, the maximum clique search algorithm employs the Bron-Kerbosch algorithm. This algorithm recursively backtracks through a weighted undirected graph to search for all complete subgraphs that satisfy the condition that all edge weights are greater than a preset correlation threshold and that the execution time windows corresponding to the vertices overlap. The subgraph with the most vertices is then selected as the maximum complete collaborative subgraph. A specific complete example illustrates the entire process. This example is only to demonstrate the feasibility at the computational level and does not represent actual values. Specific values ​​can be determined through simulation or physical experiments conducted by those skilled in the art. For instance, in the collaborative task for geological disaster monitoring in this application, four single subtasks are set, corresponding to four vertices V1, V2, V3, and V4. Each vertex has a corresponding execution time window; for example, V1 is from 9:00 to 10:00, V2 from 9:30 to 10:30, V3 from 9:00 to 9:30, and V4 from 10:00 to 10:30. The edge weights in the constructed weighted undirected graph are 0.7 between V1 and V2, 0.8 between V1 and V3, 0.6 between V2 and V3, 0.4 between V1 and V4, 0.5 between V2 and V4, and 0.3 between V3 and V4. The correlation threshold is set to 0.5. The Bron-Kerbosch algorithm is used, which recursively backtracks and searches with three sets R, P, and X as parameters. Initially, R is an empty set, P is the set of all vertices, and X is an empty set. During the recursion, vertices in P are continuously added to R, and P is updated to be the set of vertices that are connected to all vertices in R by edges with an edge weight greater than 0.5. At the same time, the overlap between the time windows of the newly added vertices and the time windows of all vertices in R is checked. For example, at 9:15 AM, the time windows of vertices V1, V2, and V3 overlap, and the edge weights between each pair of them are all greater than 0.5. Therefore, the algorithm finds the largest fully cooperative subgraph composed of vertices V1, V2, and V3 through recursive search. This largest clique is output for subsequent calculation of the instantaneous cooperative computing resource requirements at this time point.

[0113] The calculation of the coordination control computation is specifically achieved through the following formula:

[0114] Coordination and control computational load = Σ(edge ​​weight × data interaction frequency coefficient × state synchronization requirement coefficient);

[0115] Among them, the data interaction frequency coefficient is based on the preset data interaction frequency between individual subtasks, and the state synchronization requirement coefficient is based on the preset state synchronization accuracy requirement between individual subtasks.

[0116] B110. Based on the instantaneous collaborative computing resource requirements and the corresponding available computing resource types at each time point, the task computational complexity at each time point is calculated.

[0117] It should be further explained that, in this embodiment, the type of available computing resources at a given time point refers to the collection of various computing resources that the UAV swarm platform can call in real time at that time point, including CPU computing power, memory capacity, data processing bandwidth, real-time communication channel resources, and available computing power of dedicated processing modules. This is calculated by summarizing and calculating the load data of each UAV computing unit, resource allocation logs, and real-time monitoring data from the cluster resource manager, which are obtained in real time by the acquisition module. Specifically, it is the total available resources minus the resource usage already allocated to other tasks. The technical means for calculating the task computational complexity at each time point is: the instantaneous collaborative computing resource requirement at that time point is calculated as follows: Resource types are decomposed into corresponding demand sub-items; the ratio of each demand sub-item to the remaining available resources in the corresponding allocable resource class is calculated to obtain the supply and demand tension coefficient of a single resource type, where a coefficient greater than or equal to 1 indicates a resource shortage, and a coefficient less than 1 indicates sufficient resources; based on the preset weights of each resource type in task execution, the supply and demand tension coefficients of a single resource type are weighted and summed to obtain the comprehensive supply and demand coefficient at that time point; the comprehensive supply and demand coefficient is multiplied by the total amount of instantaneous collaborative calculation resource demand at that time point, and the result is the task computation complexity at that time point. The higher the value, the stronger the resource constraints and the greater the computational pressure faced by the task execution at that time point.

[0118] B111. Based on the expected completion time of the complex task and the computational complexity of the task at each time point, a weighted average algorithm is used to obtain the task complexity for allocating resources.

[0119] The motivation and reason for this approach in this embodiment is that the execution of a complex task is a continuous process over time, and the computational complexity of the task varies at different points in time. For example, the resource pressure during the execution of a critical single subtask is significantly higher than that during other stages. Relying solely on the complexity at a single point in time cannot reflect the overall resource demand intensity throughout the entire task cycle. On the other hand, simply averaging the complexity at each point in time would mask the differences in resource priority at critical points, leading to excessive resource allocation bias towards non-critical stages or neglect of overall needs. By combining the expected completion time of the complex task, the complexity at each point in time can be incorporated into a unified time framework, covering the entire cycle from task initiation to completion. Simultaneously, a weighted average algorithm is used to weight and integrate the time points based on their importance in task execution. For example, stages with closely related single subtasks and stages with complex weather conditions are given higher weights. This not only quantifies the overall resource demand intensity throughout the entire task cycle but also highlights the resource priority at critical points in time. This provides a quantitative basis for resource allocation that balances the overall and local aspects, as well as static planning and dynamic adaptation, ensuring that the allocation of computational resources not only meets the total needs of the entire task cycle but also prioritizes critical stages, avoiding a decrease in task execution efficiency or failure of critical single subtasks due to unbalanced resource allocation.

[0120] It should be further explained that the real-time flight correlation adjustment map of the UAV swarm obtained by inversion in this embodiment includes:

[0121] S101. Based on the single subtask set and the single subtask association matrix, construct the world path of all UAVs during the mission execution process, and use the position point corresponding to each time point on the world path as the world origin of the world coordinate system of all UAVs at the corresponding time point.

[0122] S102. The weighted undirected graph model is mapped to the world coordinate system using the timestamp of each world origin and the execution timestamp of the single subtask corresponding to each vertex, to obtain the dynamic flight graph space of the UAV swarm. This embodiment achieves unified mapping and calculation of tasks and UAV flight states through the following technical means: First, based on the configuration of the output performance upper limit parameter of the UAV swarm, a unified performance threshold set is generated using the Min-Max normalization algorithm, and a single UAV executable principle containing quantitative constraints is defined accordingly; Next, the structured task element set obtained by parsing the complex task text using the BERT model is input into a rule-based decomposition engine, and decomposed into a single subtask set that corresponds one-to-one with a single UAV through a recursive segmentation algorithm and a rule matching verification algorithm; Then, a system is constructed with the single subtask as the vertex, A weighted undirected graph model is used, with the correlation between tasks as the edge weights. For each vertex, a task complexity-computation resource mapping function and an environment-adjustment resource mapping function are associated. Then, based on the planned world path, the continuous task time is discretized into a time point sequence. The path position corresponding to each time point is used as the origin of the world coordinate system. The weighted undirected graph model is discretized to the world coordinate system at the corresponding time point, so that the state of each vertex (subtask) at each time stamp is completely aligned with the real-time working state of the UAV at the corresponding time point. Finally, at each discrete time point, the two mapping functions carried by each vertex are called in parallel to calculate the basic computing resource requirements of the subtask and the state correction resource requirements under the current weather conditions. This achieves the complete integration and collaborative calculation of task logic requirements, physical flight state and environmental influencing factors in a unified spatiotemporal framework.

[0123] It should be further explained that the process of obtaining the dynamic flight map space of the UAV swarm in this embodiment includes:

[0124] S1021. Based on the preset flight path corresponding to each vertex of the weighted undirected graph model and the single UAV executable principle, the timestamp corresponding to each world origin on the world path and the estimated timestamp corresponding to each position point in the preset flight path corresponding to each vertex are discretized to the plane coordinate system corresponding to the world origin of the corresponding timestamp to obtain the target position point of each UAV in each plane coordinate system.

[0125] It should be further noted that each planar coordinate system is a two-dimensional planar coordinate system constructed using the world origin corresponding to each point in time; please refer to [link / reference needed]. Figure 2 In the diagram, A represents the world path, and t1, t2, ..., tn represent n consecutive world origins. These world origins correspond one-to-one with each timestamp during the flight of all UAVs. That is, the world origins in the diagram are continuous. To illustrate this, they are discretized. w21, w22, ..., w2m represent m UAV swarms located in the plane coordinate system corresponding to the world origin t2. B represents the weighted undirected graph model of the UAV swarm at a single time point in the plane coordinate system corresponding to the world origin t2.

[0126] S1022. Simultaneously, construct weighted edges from the non-zero correlation in the single subtask correlation matrix, and decompose them into weighted edges between all UAV target position points in the two-dimensional plane coordinate system under the corresponding time stamp, according to the timestamp of each world origin and the temporal dependency, spatial proximity and resource sharing of the single subtask at the corresponding timestamp.

[0127] It should be further explained that the reason why the weighted edges constructed from the non-zero correlation in the single subtask correlation matrix in this embodiment are decomposed into weighted edges between all UAV target position points in the two-dimensional plane coordinate system under the corresponding timestamp is because the non-zero correlation in the single subtask correlation matrix is ​​itself constructed based on the temporal dependency, spatial proximity, and resource sharing among single subtasks, and each single subtask corresponds one-to-one with a specific UAV, and its correlation can directly characterize the strength of the collaborative correlation between the corresponding UAVs; at the same time, the timestamp of each world origin is precisely aligned with the execution time window of the single subtask, which can filter out single subtasks that are active at the same timestamp. The system ensures that only drones requiring collaboration within the same timeframe are associated. The two-dimensional plane coordinate system, based on the world origin of the timestamp, unifies the spatial coordinates of all active drones' corresponding target locations. Temporal dependency limits the collaborative logic of drones at the same timestamp, spatial proximity is quantified by the relative distance between target locations, and resource sharing reflects the collaborative needs of drones for computing resources. The strength of these dimensions of association can be visualized in the two-dimensional plane through weighted edges between target locations, and the weights of these edges retain the original non-zero correlation characteristics. Therefore, it can be decomposed into weighted edges between all drones' corresponding target locations within the two-dimensional plane coordinate system at the corresponding timestamp. The training dataset for this embodiment's fuzzy comprehensive evaluation comes from historical resource allocation records, including data on remaining available computing power, memory, and other resources, along with manually labeled comprehensive evaluation results of resource sharing. The training method is semi-supervised training, combining expert experience to set the fuzzy membership function. The loss function uses the fuzzy proximity loss function. The training parameters are set as follows: trapezoidal membership function type, 5×3 dimension of the fuzzy evaluation matrix, and 50 iterations of the weights.

[0128] S1023. Based on all target location points in the plane coordinate system corresponding to each world origin and the weighted edges between all target location points, obtain a static flight diagram of the UAV swarm under each world origin.

[0129] Next, a specific and complete calculation example will be used to illustrate the entire process. This example is only to illustrate the feasibility at the computational level and does not represent actual values. The specific values ​​can be determined by those skilled in the art through simulation experiments or physical experiments. For example, in the complex task of monitoring landslides in a configured test mountain area, a two-dimensional plane coordinate system is constructed for the world origin corresponding to the timestamp t2. This world origin is an intermediate monitoring point on the preset landslide monitoring world path. First, based on the hardware specification documents, factory performance test reports, and real-time status monitoring data of the configured five multi-rotor UAVs, the original performance parameters such as the maximum payload, endurance, and operating radius of the UAVs are input using the Min-Max normalization algorithm, and a set of performance thresholds with uniform dimensions is output. Based on this, the executable principle of a single UAV is defined. The structured task element set obtained by parsing complex task text using the BERT model is then input into a rule-based decomposition engine. Through a recursive segmentation algorithm and a rule-matching verification algorithm, it is decomposed into five single image acquisition sub-tasks, establishing a one-to-one correspondence between each sub-task and a drone. A weighted undirected graph model is constructed, with the five single sub-tasks as vertices and the inter-task correlation as edge weights. For each vertex, a task complexity-computation resource mapping function and an environment-adjustment resource mapping function are associated. Next, based on a preset flight path, the estimated timestamp of the drone's position corresponding to each single sub-task is precisely aligned with timestamp t2 and discretized into the plane coordinate system corresponding to t2, yielding the target position points of the five drones: drone 1 corresponds to P21, drone 2 to P22, drone 3 to P23, drone 4 to P24, and drone 5 to P25. Subsequently, non-zero correlation degrees were extracted from the correlation matrix of individual subtasks. These correlation degrees were determined based on temporal dependency, spatial proximity, and resource sharing. The correlation degree between subtask 1 and subtask 2 was 0.7, between subtask 2 and subtask 3 was 0.6, and between subtask 4 and subtask 5 was 0.5. The correlation degree between the remaining subtasks was zero. Temporal dependency was determined when all subtasks were actively executing at timet2, with no sequential constraints. Spatial proximity was quantified using the Euclidean distance algorithm, inputting the target location coordinates and outputting the relative distance. Resource sharing was quantified using the fuzzy comprehensive evaluation method, inputting the remaining available computing power data and outputting the overall satisfaction level. These non-zero correlation degrees were constructed into weighted edges. Based on the temporal constraints of timet2, drones corresponding to active subtasks within the same time period were selected. The weights of the weighted edges retained their original correlation degree values, connecting the target location points corresponding to drones 1 and 2, 2 and 3, and 4 and 5, respectively.Finally, based on the five target location points in the world origin plane coordinate system of timestamp t2 and the corresponding weighted edges, a static map of the UAV swarm flight under the world origin is obtained. In this static map, the nodes are the target location points of the UAVs, the weights of the weighted edges represent the cooperative association strength between the corresponding UAVs, and all nodes and weighted edges are in a unified two-dimensional plane coordinate system, realizing the accurate mapping between a single subtask and the UAV flight status under the timestamp.

[0130] S1024. Based on the static flight graph of the UAV swarm at each world origin, combined with the world path and the preset flight path corresponding to each UAV, a dynamic flight graph space of the UAV swarm is obtained; wherein the dynamic flight graph space of the UAV swarm includes the static flight graph of the UAV swarm corresponding to consecutive time points, which is used to characterize the flight status of all UAVs at each time point and the correlation between space and computational complexity.

[0131] It should be further explained that the preset flight path for each UAV in this embodiment is constructed by a path optimization algorithm based on a single subtask and its correlation matrix, combined with particle swarm optimization. Specifically, based on the target position coordinates and correlation matrix of the single subtask, the initial flight path is generated using the standard particle swarm optimization algorithm. The first optimization objective is to minimize the total Euclidean length of the path, and the second is to maximize the satisfaction of temporal dependency constraints. A dual-objective fitness function is constructed using a linear weighting method for optimization. Based on the spatial correlation strength value extracted from the single subtask correlation matrix, a fifth-order polynomial interpolation path smoothing algorithm is used to process the heading angle sequence of the initial path to eliminate abrupt changes in heading. To improve trajectory smoothness, based on an environmental map containing both static and dynamic obstacles, and pre-defined dynamic constraints on the maximum turning rate and climb rate of the UAV, an improved A* algorithm considering dynamic constraints is used for local fine-grained path planning to obtain an obstacle avoidance path coordinate sequence that meets flightability requirements. Based on the planned path set of all UAVs, a 3D bounding box collision detection algorithm is used for cross-validation between paths. Path points are iteratively adjusted to ensure that the minimum Euclidean distance between any two paths meets a pre-defined safety threshold. Based on the final optimized and safe path coordinate sequence, a cubic uniform B-spline curve parameterization method is used for fitting, and a complete executable flight trajectory containing continuous 3D position coordinates, velocity vectors, and high-precision timestamps is generated through first-order difference calculation. The training dataset for the particle swarm optimization algorithm in this embodiment is historical UAV flight path planning data, including indicators such as total path length and temporal constraint satisfaction. The training method is unsupervised optimization training. The loss function is a weighted loss function of path length and constraint satisfaction. The training parameters are set as follows: population size 50, maximum number of iterations 100, inertia weight 0.8, and learning factor 1.5. The training dataset for the improved A* algorithm in this embodiment comes from static and dynamic obstacle coordinate data of the environmental map and UAV dynamic constraint parameters, such as maximum turning rate and climb rate. The training method is no training procedure; path search is performed directly. No loss function is used. The training parameters are set to a heuristic function weight coefficient of 1.2 and a search depth threshold of 50. The training parameters for the 3D bounding box collision detection algorithm in this embodiment are set to a UAV safe distance threshold of 30 meters.

[0132] S1025. Based on the single sub-task complexity of each UAV from the current target location to the next target location in the dynamic flight map space of the UAV swarm, combined with the corrected flight area meteorological data, the task complexity-computation resource mapping function and the environment-adjustment resource mapping function under the corresponding single sub-task, and the resource sharing of the corresponding time interval from the current target location to the next target location, a local resource optimization function and a global resource optimization function and a first resource constraint space are constructed. In this embodiment, the global resource optimization function is constructed from the local resource optimization function and the local collaborative resource optimization function, and is used to characterize the flight control computation requirements of each UAV from the current target location to the next target location and the collaborative computation requirements required for at least two UAVs to maintain a safe flight state at the same timestamp. The local resource optimization function is used to characterize the flight control computation requirements of a single UAV. For example, in this embodiment, the flight... Flight control computational requirements refer to the computational resources needed for a single UAV to perform basic flight control (such as attitude stabilization, trajectory tracking, and obstacle avoidance). For example, when a UAV is flying in a straight line, the processor computing power required by its flight control system to solve the attitude control law. Cooperative computational resource requirements refer to the additional computational overhead incurred by multiple UAVs to maintain a cooperative and safe state (such as formation keeping, collision avoidance, and formation changes). For example, when three UAVs are making synchronized turns, the communication and coordination computing power required to exchange position data in real time and solve relative distance constraints. Local resource optimization functions quantify the computational requirements of a single UAV's flight control, such as using the processor utilization rate of a certain UAV as the optimization objective. Local cooperative resource optimization functions quantify the computational resource requirements of multiple UAVs' cooperative operations, such as using the data exchange latency and coordination computational load during formation keeping as the optimization objective. The two are combined by weighting to form a global resource optimization function, which is used to optimize the resource consumption of a single UAV and the cooperative overhead of the cluster while satisfying safety and efficiency constraints.

[0133] S103. Based on the probability of each UAV completing a single subtask within the corresponding time period from the current target location to the next target location in the dynamic flight map space of the UAV swarm, and the probability of all UAVs completing a single subtask within the corresponding time period, a local task completion probability function is constructed. Utilizing the temporal dependency and spatial proximity of the corresponding time period from the current target location to the next target location for each UAV, a second spatiotemporal constraint condition is constructed between the UAV task execution priority and the flight space distance. The probability of each UAV completing a single subtask within the corresponding time period from the current target location to the next target location in the dynamic flight map space of the UAV swarm is calculated using a Bayesian neural network model, based on the local task completion probability function corresponding to the current target location and the corresponding failure probability between the current target location and the next target location.

[0134] Traditional single subtask completion probability estimation does not fully consider the interference of UAV flight zone failure risks on mission execution, easily leading to deviations between probability values ​​and actual scenarios. This, in turn, affects the rationality of mission constraint construction and resource optimization allocation. Combining failure probability can accurately quantify the possibility of failure causing mission interruption or efficiency reduction within this range, improving the accuracy of probability estimation. Based on the local mission completion probability function corresponding to the current target position of the UAV, the failure probability between the current target position and the next target position is incorporated. A Bayesian neural network model is used to mine the inherent correlation between the two types of data and fuse them for calculation, outputting a single subtask completion probability that fits the reality. Then, a local mission completion probability function is constructed based on this probability, and a second spatiotemporal constraint is constructed by combining temporal dependency and spatial proximity. Based on this, a hierarchical optimization framework is corrected. The core design principle lies in the fact that the local resource optimization function focuses on the internal control resource requirements of a single UAV executing a single sub-task, which are affected by task complexity, environmental factors, and failure probability. The local collaborative resource optimization function addresses the interactive resource overhead generated in multi-UAV collaborative flight to maintain formation safety, task synchronization, and collaborative adjustments to cope with failures. By organically combining the two to construct a global resource optimization function, it ensures both the independent resource guarantee and failure response resource supply for a single UAV executing a single sub-task, while also taking into account the additional resource consumption brought about by swarm collaboration. Under the premise of satisfying the first resource constraint and the second spatiotemporal constraint, it achieves full-dimensional resource optimization allocation from individual control to swarm collaboration, while also taking into account failure response, effectively improving the overall resource utilization efficiency and task execution reliability of UAV swarms in complex task environments. The training dataset of the Bayesian neural network model in this embodiment comes from the failure probability data within the historical UAV flight range and the local task completion probability data at the current position. The training method is supervised training, iteratively updating the network weights based on historical label data. The loss function adopts the mean squared error loss function. The training parameters are set as follows: hidden layer dimension 64, activation function ReLU, number of training epochs 100, batch size 16, and learning rate 0.001.

[0135] S104. Based on the resource optimization function and the first resource constraint, the local task completion probability function and the second spatiotemporal constraint for each UAV, combined with the flight state parameters of each UAV at each target location and the Hidden Markov Algorithm, the flight state probability transition matrix for each UAV to complete the corresponding task, the conditional probability function for successfully completing the corresponding single sub-task, and the global probability function for completing the complex task are obtained. It should be further noted that one specific implementation of the Hidden Markov Algorithm in this embodiment is as follows:

[0136] S1041. Based on the correspondence between the flight state parameters of each UAV in the world coordinate system and the world origin, define the state space of the Hidden Markov Model; the state space includes the three-dimensional position coordinates, velocity vector, attitude angle, energy status and mission execution progress index of the UAV, wherein the state transition is synchronized with the change of the world origin.

[0137] S1042. Based on the real-time flight status parameters acquired by the acquisition module, the corrected meteorological data of the flight area, and the mission execution data, the observation space of the Hidden Markov Model is defined. The observation space includes the position deviation, attitude error, meteorological influence coefficient, resource consumption rate, and mission completion index measured by the sensors. The position deviation is obtained through the fusion of differential GPS and inertial navigation system data. The attitude error comes from the difference in attitude calculation between the gyroscope and the visual sensor. The meteorological influence coefficient is calculated by comparing the real-time measurement value of the atmospheric data computer with the standard atmospheric parameters. The resource consumption rate is statistically analyzed in real time by monitoring the load of the computing unit and the energy management system. The mission completion index is determined by comparing the degree of conformity between the actual execution progress and the preset mission plan. These observation indicators together constitute a multi-dimensional data source for UAV status monitoring, providing real-time input for the Hidden Markov Model to accurately evaluate flight status stability, environmental adaptability, and mission execution efficiency, thereby supporting state transition probability calculation, mission completion probability prediction, and resource scheduling decisions.

[0138] S1043. The Baum-Welch algorithm is used to train the model on historical flight state sequences and mission completion records to estimate the parameters of the Hidden Markov Model (HMM). The state transition probability matrix represents the conditional probability of the UAV transitioning from one spatial state point to the next in the world coordinate system, reflecting the dynamic evolution characteristics of state parameters such as flight attitude and position coordinates in the spatiotemporal dimension. The observation probability matrix represents the probability distribution of various observation data acquired by sensors at each time point in the corresponding flight state, establishing a statistical correlation between the internal changes in flight state and external sensor measurements. The initial state probability distribution represents the prior probability of the UAV being in each possible state at the start of the mission, providing an initial condition benchmark for the entire probabilistic evolution process. These three parameters together constitute a complete probabilistic description framework for the spatiotemporal evolution of the UAV flight state, laying a mathematical foundation for subsequent state prediction and mission probability calculation. In this embodiment, the training dataset for the Baum-Welch algorithm comes from historical UAV flight state sequence data and mission completion records, including observation data such as position deviation and attitude error, as well as state labels. The training method is unsupervised training, iteratively estimating the parameters of the HMM. The log-likelihood loss function is used. The training parameters were set to 100 iterations and 0.001 convergence threshold.

[0139] S1044. Based on the trained Hidden Markov Model, the probability of each UAV being in each state under a given observation sequence is calculated by the forward algorithm, and a conditional probability function for successfully completing the corresponding single sub-task is constructed by combining the task success determination rules; the conditional probability function takes the current flight state and environmental parameters as input and outputs the task completion probability.

[0140] S1045. Based on the conditional probability functions of all UAVs and the temporal dependencies in the association matrices of individual subtasks, a global probability function for completing complex tasks is constructed using a probabilistic graphical model fusion algorithm. This global probability function comprehensively considers the joint distribution of the completion probabilities of each individual subtask and the dependencies between tasks. The specific implementation process of constructing the global probability function using a Bayesian network in this embodiment includes: constructing a directed acyclic graph structure with each individual subtask as a node based on the temporal dependency weights and spatial association strength values ​​in the association matrices of individual subtasks; configuring a conditional probability table with spatiotemporal constraints for each node using the conditional probability function output by the Hidden Markov Model and the task priority parameters; calculating the joint probability distribution of all node states by fusing the temporal constraints and resource coordination constraints of the task execution sequence based on the chain rule of the probabilistic graphical model; and using a belief propagation algorithm to perform iterative message passing and probabilistic reasoning on the constructed probabilistic network. By calculating marginal probabilities, a global probability function that comprehensively considers task dependencies, dynamic resource allocation, and environmental interference factors is obtained, thereby achieving an accurate quantitative assessment of the probability of completing complex tasks and providing a probabilistic basis for the collaborative control of UAV swarms. In this embodiment, the Bayesian network and belief propagation algorithm are trained using supervised training to optimize the conditional probability table of network nodes. The cross-entropy loss function is used. The training parameters are set to 10 iterations and a confidence convergence threshold of 0.005.

[0141] S1046. The most likely state sequence is decoded using the Viterbi algorithm, and the flight state probability transition matrix of each UAV is updated in real time to predict the state evolution and mission completion trend at future time points.

[0142] This process synchronously correlates flight state evolution with changes in the world coordinate system using a Hidden Markov Model, effectively quantifying the impact of factors such as weather interference and resource constraints on mission execution in dynamic environments. Based on conditional probability functions, it accurately assesses the probability of completing a single sub-task, and combines a probabilistic graphical model to integrate task dependencies and construct a global probability function, achieving full-dimensional probability assessment from individual states to group collaboration. Finally, through real-time state sequence decoding and probability transition matrix updates, it provides data-driven decision-making basis for dynamic resource allocation, risk warning, and mission replanning, significantly improving the adaptability and mission success rate of UAV swarms in complex mission environments.

[0143] Next, a specific and complete example will be used to illustrate the entire process. This example is only to illustrate the feasibility at the computational level and does not represent actual values. The specific values ​​can be determined by those skilled in the art through simulation experiments or physical experiments. For example, in the landslide monitoring task of the Qinglong Mountains in this application, a Hidden Markov Model is constructed for the image acquisition subtask performed by UAV No. 1. The model state space is defined as five states: normal flight state S1, attitude slight deviation state S2, resource shortage state S3, meteorological disturbance state S4, and task completion state S5. The observation space is defined based on real-time data as four observation dimensions: position deviation level 0 to 5, attitude error level 0 to 4, meteorological influence coefficient level 0 to 3, and resource consumption rate level 0 to 3. The model parameters are obtained by training 100 historical flight data using the Baum-Welch algorithm. The number of training iterations is set to 100, and the convergence threshold is set to 0.001. Finally, the state transition probability matrix A is obtained, where the probability of transitioning from S1 to S1 is set to 0.75, and the probability of transitioning to S2 is set to 0. .15. The probability of transferring to S3 is set to 0.05, the probability of transferring to S4 is set to 0.05, and the probability of transferring from S2 to S1 is set to 0.6. The probability of transferring to S2 is set to 0.3, the probability of transferring to S3 is set to 0.05, the probability of transferring to S4 is set to 0.05, the probability of transferring from S3 to S1 is set to 0.4, the probability of transferring to S3 is set to 0.5, the probability of transferring to S4 is set to 0.1, the probability of transferring from S4 to S1 is set to 0.3, the probability of transferring to S2 is set to 0.2, and the probability of transferring to S4 is set to 0.5. The probability of S5 is set to 0.5, the probability of transitioning from S5 to S5 is set to 1.0, the probability of observing a vector combination of position deviation level 2, attitude error level 1, meteorological influence coefficient level 0, and resource consumption rate level 1 in state S1 in the observation probability matrix B is set to 0.12, the probability of observing a vector combination of position deviation level 0, attitude error level 0, meteorological influence coefficient level 0, and resource consumption rate level 1 in state S1 is set to 0.25, and the initial state distribution π is set to 0.8, 0.1, 0.05, 0.05, and 0. Based on the trained model input, the observation sequence of the most recent 10 time points is used. This sequence consists of 10 sets of data, including position deviation level 2, attitude error level 1, meteorological influence coefficient level 0, and resource consumption rate level 1, and so on. The probability of being in each state is calculated using a forward algorithm. When the forward algorithm is executed, the forward probability of each state at the first time point is calculated using the forward probability initialization formula, where S1 is 0.096, S2 is 0.008, S3 is 0.0025, S4 is 0.003, and S5 is 0. Then, the forward probability of each state at each subsequent time point is calculated sequentially using the forward probability recursion formula. Finally, the forward probabilities of each state at the last time point are accumulated to obtain the probability of being in each state.Based on the temporal dependencies between the three drone subtasks, the completion of drone subtask 1 is set as a prerequisite for the initiation of drone subtask 2, and the completion of drone subtask 2 is a prerequisite for the initiation of drone subtask 3. A global probability function is constructed using a Bayesian network, and a belief propagation algorithm is used for 10 iterations. In each iteration, the confidence value of each node is updated using a confidence update formula. The initial confidence values ​​are set to 0.82 for the completion probability of drone subtask 1, 0.78 for drone subtask 2, and 0.75 for drone subtask 3. After 10 iterations, the confidence values ​​of each node converge, and the global probability of the entire complex task is calculated using a Bayesian neural network. Then, the Viterbi algorithm is used to decode the most likely state sequence at the next 5 time points. When the Viterbi algorithm is executed, the initial Viterbi probability and path pointer of each state are initialized first using the Viterbi probability initialization formula. Then, the Viterbi probability and path pointer of each state at each future time point are calculated sequentially using the Viterbi probability recursion formula. Finally, the most likely state sequence is obtained by backtracking using the path backtracking formula, which is S1, S1, S2, S1, S5. At the same time, based on the new state sequence data, the state transition probability matrix is ​​updated in real time using the state transition probability update formula, updating the probability of S1 to S3 to 0.76 and the probability of S1 to S2 to 0.14, which are used for subsequent state prediction and decision-making.

[0144] S105. Based on the local resource optimization function and the global resource optimization function, the local task completion probability function, the flight state probability transition matrix of each UAV to complete the corresponding task, the conditional probability function of successfully completing the corresponding single sub-task, and the global probability function of completing the complex task, a global multi-objective optimization function is constructed, and a global constraint space is constructed with the first resource constraint condition and the second spatiotemporal constraint condition.

[0145] It should be further explained that the local resource optimization function in this embodiment is used to quantify the amount of flight control computing resources required for a single UAV to perform its corresponding single sub-task. Its calculation is based on the task complexity-computation resource mapping function and the environment-adjustment resource mapping function, with the goal of minimizing the resource consumption of a single platform.

[0146] It should be further explained that the global resource optimization function in this embodiment is used to quantify the total computing resource requirements generated by the entire UAV swarm when collaboratively executing complex tasks. It is constructed by a linear weighted sum method of the local resource optimization function and the local collaborative resource optimization function, aiming to optimize the individual consumption and the cluster collaborative overhead in a coordinated manner.

[0147] It should be further explained that the local task completion probability function in this embodiment is used to characterize the probability that a single UAV can successfully complete its corresponding single sub-task within a specific time and space segment. Its calculation integrates state prediction based on a hidden Markov model and real-time failure probability, and is usually obtained by data fusion through a Bayesian network.

[0148] It should be further explained that the conditional probability function for the successful completion of a single subtask in this embodiment is used to accurately calculate the instantaneous probability of the single subtask being successfully completed under the given current flight state and environmental observation sequence, and is solved by the forward algorithm of the Hidden Markov Model.

[0149] It should be further explained that the global probability function for completing the complex task in this embodiment is used to evaluate the overall probability of the entire complex task being successfully completed. It integrates the conditional probabilities of all individual subtasks and their temporal dependencies, and is calculated using a Bayesian network-based probabilistic graphical model through a belief propagation algorithm.

[0150] It should be further explained that the first resource constraint in this embodiment defines the upper limit of the real-time available total amount of various types of computing resources (such as computing power, memory, and bandwidth) of the UAV swarm, which is used to constrain the solution space of the resource optimization function.

[0151] It should be further explained that the second spatiotemporal constraint condition in this embodiment defines the temporal dependency, spatial proximity and task priority relationship that the task execution must satisfy, in order to ensure the logical correctness and spatiotemporal feasibility of task decomposition and execution.

[0152] The aforementioned functions and conditions together constitute the global multi-objective optimization function and global constraint space, providing clear teaching objectives and boundaries for subsequent multi-objective optimization search algorithms (such as NSGA-II), thereby solving for the optimal flight control strategy that minimizes resource consumption and maximizes mission success rate.

[0153] This embodiment uses a Hidden Markov Model (HMM) to synchronously model the flight state transitions of UAVs with changes in the world coordinate system. Its core motivation lies in addressing the uncertainty quantification problem between flight state and mission completion in dynamic environments. Traditional methods struggle to accurately capture the probabilistic characteristics of flight state affected by weather interference, resource constraints, and spatiotemporal changes. HMMs, however, utilize state transition matrices to characterize the statistical laws governing state evolution between world origins, quantify the probability of completing individual sub-tasks through conditional probability functions, and then integrate global task dependencies through probabilistic graphical models, ultimately achieving a complete closed loop from individual state prediction to swarm mission probability assessment. This probabilistic modeling approach provides data-driven decision-making support for resource scheduling, risk warning, and mission replanning, significantly improving the robustness and success rate of UAV swarm systems under complex tasks.

[0154] Next, a specific and complete example will be used to illustrate the entire process. This example is only to illustrate the feasibility at the computational level and does not represent actual values. The specific values ​​can be determined by those skilled in the art through simulation experiments or physical experiments. For example, in the specific implementation of the collaborative monitoring task in the test mountain area set up in this application, three UAVs are set to perform image acquisition, laser scanning, and data transmission sub-tasks respectively. The local resource optimization function is constructed based on the task complexity to computing resource mapping function. Its input is the sub-task complexity feature vector. For example, the vector value of the image acquisition task is an area of ​​4 square kilometers and a resolution of 0.05 meters. The basic computing power requirement of 3.2 gigaflops per second is calculated by using linear coefficients of 0.8 and 3.0. Combined with the environmental adjustment resource mapping function, the output correction amount of 0.4 under the condition of a wind speed of 5 meters per second is used to finally obtain the local resource optimization function value of 3.6 gigaflops per second. The global resource optimization function uses preset weights of 0.7 and 0.3 to weight the sum of the local optimization values ​​of the three UAVs and the cooperative communication overhead of 0.9 gigaflops per second, resulting in a global value of 10.5 gigaflops per second. The local task completion probability function uses a Bayesian network to fuse the current attitude stability (0.9) and energy sufficiency (0.95), outputting a probability of 0.855. Based on the state transition matrix of the Hidden Markov Model and the current observation sequence, the conditional probability function value for the image acquisition subtask is calculated to be 0.82. Then, by fusing the conditional probabilities and temporal dependencies of the three subtasks using a Bayesian network and iteratively calculating using the belief propagation algorithm, the global probability function value for completing the entire monitoring task is obtained as 0.75. The first resource constraint is set as a maximum total available computing power of 15 gigaflops per second and a maximum communication bandwidth of 100 megabits per second. The second spatiotemporal constraint requires that image acquisition must be completed before scanning begins, and the minimum distance between the target points of each UAV must be greater than 30 meters. These functions and constraints together constitute a global multi-objective optimization problem with the objectives of minimizing global resource consumption of 10.5 and maximizing global task success rate of 0.75, providing clear input for the subsequent NSGA-II optimization algorithm.

[0155] S106. Based on the global multi-objective optimization function and global constraint space, combined with a preset flight state adjustment coefficient library, and integrating a multi-objective optimization search algorithm and a twin simulation algorithm, the optimization search is performed using the global constraint space. The goal is to minimize the local resource optimization function and the global resource optimization function, and to maximize the local task completion probability function, the conditional probability function of each UAV successfully completing its corresponding single sub-task, and the global probability function of completing a complex task. This yields the optimal flight state correction adjustment coefficients for the local relative paths of all UAVs in the UAV swarm's dynamic flight map space from the current target position to the next target position. Please refer to [link to relevant documentation]. Figure 3Where q11, q12, ..., q1m correspond to the local relative paths of m UAVs from the world origin t1 to the world origin t2. In this embodiment, the flight state of each UAV in the corresponding m local relative paths at each time point between n world origins is adjusted in real time based on the collected meteorological data of the flight area. Here, q(i-1)1, q(i-1)2, ..., q(i-1)m represent the set of m local relative paths between the world origin t(i-1) and the world origin ti. Based on the m local relative paths at each time point between n world origins, the dynamic flight graph space of the UAV swarm in this application is constructed. It should be further noted that the flight state adjustment coefficient library preset in this embodiment is a knowledge base that stores a variety of verified and reusable combinations of flight state adjustment parameters. It is used to provide a high-quality initial candidate solution set for the multi-objective optimization search algorithm, thereby effectively guiding the search direction, accelerating the optimization convergence process and improving the quality of the final solution. The construction method includes the following technical means: Based on historical successful mission records and high-fidelity simulation data, data mining technology is used to extract effective flight state adjustment coefficient combinations under different mission scenarios, meteorological conditions, and UAV models to form an original parameter set; the extracted original parameter set is subjected to feature engineering processing, and a feature dimensionality reduction method based on principal component analysis and the Min-Max standardization algorithm are used to generate standardized adjustment coefficient feature vectors; based on the standardized feature vectors, a hierarchical clustering algorithm is used to divide them into several categories with similar adjustment patterns, and a representative baseline adjustment coefficient combination is generated for each category; for each baseline coefficient combination, based on the constructed high-fidelity digital twin simulation environment, a systematic simulation verification is performed using the control variable method to evaluate its effectiveness and robustness under various boundary conditions, and the verified coefficient combinations and their associated scenario feature labels are stored in a relational database; finally, a flight state adjustment coefficient library that supports fast retrieval under multi-dimensional conditions is constructed using database indexing technology. This library supports real-time querying and returning a set of related alternative adjustment coefficients by inputting mission scenario features, UAV models, and meteorological condition parameters, which serve as components of the initial population in the multi-objective optimization search algorithm.

[0156] The optimal flight state correction adjustment coefficients obtained in this embodiment include: attitude control correction coefficients for adjusting the response sensitivity of the aircraft's pitch, roll, and yaw; power output correction coefficients for optimizing the balance between motor thrust and energy consumption; heading trajectory correction coefficients for adjusting the accuracy and smoothness of path tracking; cooperative synchronization correction coefficients for maintaining the relative position and speed synchronization among multiple UAVs; and resource allocation correction coefficients for dynamically adjusting the allocation of computing resources between flight control and mission execution.

[0157] It should be further explained that the specific implementation process of the multi-objective optimization search algorithm and the twin simulation algorithm in this embodiment includes:

[0158] S1061. Based on the global multi-objective optimization function and the global constraint space, the non-dominated sorting genetic algorithm NSGA-II with elitist strategy is adopted as the multi-objective optimization search algorithm. An initial population containing multiple combinations of flight state correction adjustment coefficients is initialized based on the preset population size parameter, wherein each individual represents a complete set of coefficient configuration schemes through real number encoding.

[0159] S1062. Based on the UAV dynamics model, atmospheric environment model and sensor characteristic parameters, a high-fidelity digital twin simulation environment is constructed. Based on the twin simulation algorithm, parallel simulation calculations are performed on each individual in the population. Multiple simulation instances are run simultaneously using multi-threading technology to obtain the performance index data of each coefficient combination within the complete task cycle.

[0160] S1063. Based on a multiphysics coupled computing engine, the cooperative flight state of the UAV swarm under each coefficient combination is calculated in real time during the twin simulation. The kinematic equations are solved using a numerical integration method. Random disturbances are injected based on a sensor noise model, and data on local resource consumption, global cooperative efficiency, task completion, and probability indicators are collected. The training dataset for the numerical integration algorithm in this embodiment comes from the parameters of the UAV dynamics model and the atmospheric environment model. The training method is a no-training-process approach, directly solving the kinematic equations. There is no loss function. The training parameters are set to an integration step size of 0.01 seconds and a total integration time equal to the task cycle.

[0161] S1064. Based on the simulation output dataset, the fitness value of each individual in the global multi-objective optimization function is calculated using a multi-objective fitness function. The goal achievement degree of the resource optimization function is calculated using a weighted summation method. The degree of achievement of the task probability function is calculated using a probability product rule. The degree of constraint violation is calculated using a constraint violation quantification method. The training dataset for the multi-objective fitness function calculation algorithm in this embodiment comes from the performance index data such as local resource consumption, global collaborative efficiency, and task completion degree output by the twin simulation. The training parameters are set to a resource optimization goal weight of 0.7 and a task probability goal weight of 0.3.

[0162] S1065. The population individuals are hierarchically sorted based on the non-dominated sorting algorithm, the distribution density of individuals in the target space is evaluated based on the crowding calculation operator, high-quality individuals are selected based on the tournament selection operator, and a new generation of population is generated based on the simulated binary crossover operator and the polynomial mutation operator; in this embodiment, the training parameter of the simulated binary crossover operator is set to the crossover distribution index of 20.

[0163] S1066. Based on the convergence criterion, the optimization iteration process is repeatedly executed. When the preset maximum number of iterations or the convergence threshold of the solution is reached, the combination of flight state correction adjustment coefficients with the best overall performance is selected from the optimal solution set based on the Pareto dominance relationship.

[0164] S1067. The robustness of the optimal coefficient combination is verified based on the sensitivity analysis method. Monte Carlo tests are conducted by injecting random perturbation parameters into the twin simulation environment. The stability index of the coefficients under different perturbation conditions is evaluated based on the statistical analysis method. The training dataset of the sensitivity analysis and Monte Carlo test algorithms in this embodiment comes from the twin simulation performance index data and random perturbation parameter data of the optimal coefficient combination. The training parameters are set to 1000 perturbation times and ±5% perturbation amplitude.

[0165] S1068. Based on the verified set of optimal flight state correction adjustment coefficients, the coefficient configuration parameters are injected into the UAV swarm control system through the communication interface, and the closed-loop conversion from simulation optimization to actual application is completed based on the feedback control mechanism.

[0166] Next, a specific complete example will be used to illustrate the entire process. This example is only to illustrate the feasibility at the computational level and does not represent actual values. The specific values ​​can be determined by those skilled in the art through simulation experiments or physical experiments. For example, in the test mountain landslide collaborative monitoring task set in this embodiment, the non-dominated sorting genetic algorithm NSGA-II with an elitist strategy is used to optimize the flight state correction adjustment coefficients of three UAVs. The algorithm sets the population size to 100 individuals, and each individual uses a real number encoding method to represent attitude control, power output, heading trajectory, collaborative synchronization, and resources. The source allocation uses adjustment coefficients for five dimensions. The maximum number of iterations is set to 300 generations. The crossover probability is set to 0.9, the mutation probability to 0.1, and both the crossover and mutation distribution indices are set to 20. The algorithm input is a fitness evaluation set containing global resource optimization function values, global task probability function values, and constraint violation degrees. A randomly selected individual from the initial population corresponds to a resource consumption of 15.2 gigaflops per second, a task success probability of 0.72, and a constraint violation degree of 0. The adjustment coefficients for the five dimensions corresponding to this individual are: attitude control coefficient 0.62, power output coefficient... The algorithm uses a high-fidelity digital twin simulation environment to perform parallel simulation calculations on all individuals within each generation of the population, with coefficients of 0.95, 0.78, 0.55, and 0.88. After each generation, individuals are first stratified and sorted using a non-dominated sorting operation, and then density assessments are performed on individuals at the same level using a crowding calculation operation. An elite retention strategy is then used to merge the parent and offspring populations, selecting individuals with better fitness to form a new generation. After 300 generations of iterative calculations, the algorithm outputs a Pareto front containing 15 non-dominated solutions. Based on the resource supply and execution requirements of the actual monitoring task, the technicians selected a balanced solution. The five-dimensional optimization coefficient combination corresponding to this balanced solution is: attitude control coefficient 0.85, power output coefficient 1.2, heading trajectory coefficient 0.93, cooperative synchronization coefficient 0.78, and resource allocation coefficient 1.05. The resource consumption corresponding to this coefficient combination is 18.6 gigabit floating-point operations per second, the task success probability is 0.79, and the constraint violation degree is 0, which meets the requirement of a total computing power constraint of 20 gigabit floating-point operations per second. This optimization coefficient combination is injected into the actual UAV flight control system for real-time correction and adjustment of flight status.

[0167] S107. Map the optimal flight state correction adjustment coefficient of each UAV from the current target position to the next target position to the local relative path connection between the corresponding UAV from the current target position to the next target position. At the same time, calculate the complexity of the corresponding local relative path adjustment process based on the optimal flight state correction adjustment coefficient in the local relative path connection.

[0168] S108. Based on the adjustment process complexity of the corresponding local relative path and combined with the HSV color space, obtain the adjustment complexity mapping color depth of the corresponding local relative path.

[0169] S109. Based on the local relative path of each UAV and the corresponding optimal flight state correction adjustment coefficient and the corresponding local relative path adjustment complexity mapping color depth combined graph algorithm, a real-time flight association adjustment graph of all UAV groups between adjacent world origins is constructed.

[0170] S110. Map the real-time flight correlation adjustment map of all UAV swarms between each adjacent world origin to the UAV swarm dynamic flight map space. Combine the online learning algorithm with the real-time general correction coefficient-corrected flight area meteorological data for online fine-tuning training, so that the global multi-objective optimization function value of the UAV swarm between each adjacent world origin satisfies the corresponding preset function value in real time. The online fine-tuning training process in this embodiment includes: acquiring real-time flight area meteorological data corrected by the general correction coefficient, combining the real-time flight correlation adjustment map between adjacent world origins in the UAV swarm dynamic flight map space, the local relative paths of each UAV, and the... The optimal flight state correction adjustment coefficient is determined by employing an online gradient descent algorithm as the core online learning algorithm. Training samples are extracted based on real-time flight data of UAV swarms between adjacent world origins and the associated adjustment graph. The model parameters are continuously updated through incremental learning. The global multi-objective optimization function value of the UAV swarm between adjacent world origins is calculated in real time and compared with the preset value. If the preset value is not met, the local relative path parameters and flight state correction adjustment coefficients of the UAVs are dynamically adjusted. The data acquisition, model update, function value verification, and parameter adjustment process is iteratively executed until the global multi-objective optimization function value of the UAV swarm between adjacent world origins meets the corresponding preset function value in real time.

[0171] Next, a specific complete example will be used to illustrate the entire process. The example is only to illustrate the feasibility at the computational level and does not represent the actual values. The specific values ​​can be determined by those skilled in the art through simulation experiments or physical experiments. For example, in the test mountain landslide collaborative monitoring task set up in this application, the optimal flight state correction adjustment coefficients obtained by optimization through a non-dominated sorting genetic algorithm with an elite strategy are injected into the local relative path connection of each UAV from the current waypoint to the next waypoint. The specific coefficients include attitude control coefficient 0.85, power output coefficient 1.2, heading trajectory coefficient 0.93, collaborative synchronization coefficient 0.78, and resource allocation coefficient 1.05. The range of the flight segment covered by this local relative path is delineated based on the terrain features of the landslide monitoring area. The control commands of each axis on the path include three types: pitch axis, roll axis, and heading axis. Based on the change rate of each axis control command, the complexity of its adjustment process is calculated to be 0.65 through a complexity evaluation method. The complexity value was then processed using the HSV color space mapping algorithm. The algorithm fixed the saturation (S) at 0.9 and the brightness (V) at 0.8, mapping the complexity value to a hue (H) value via a linear mapping. Complexity 0 corresponds to an H value of 0 degrees, and complexity 1 corresponds to an H value of 255 degrees. Thus, complexity 0.65 was mapped to a hue H value of 130 degrees, generating a corresponding dark green depth identifier. This identifier visually represents the complexity of the path adjustment. Based on the color-coded local paths of all participating UAVs and their corresponding correction coefficients, a force-directed algorithm was used to construct a real-time flight correlation adjustment graph between adjacent world origins. The algorithm set the repulsion coefficient to -100 and the attraction coefficient to 0.1. Nodes in the graph represent the real-time position of the UAVs, and node attributes include the UAV number, current waypoint coordinates, and remaining computing power. Colored edges in the graph represent the UAV's adjustment path, and the edge color depth is positively correlated with the path adjustment complexity. Finally, this real-time flight correlation adjustment map was mapped to the dynamic flight map space, and the path parameters were fine-tuned online using an online gradient descent algorithm combined with real-time corrected wind speed and turbulence data. The algorithm was set with a learning rate of 0.01 and an iteration limit of 50 times. The real-time corrected wind speed data was 3.2 m / s, and the turbulence data was 0.5. The fine-tuned path parameters included waypoint spacing, heading angle correction, and flight altitude correction. After 15 iterations, the global multi-objective optimization function value of the current flight segment was adjusted from the initial 12.5 to the preset threshold of 10.0, meeting the dual requirements of path accuracy and collaborative efficiency for landslide collaborative monitoring tasks.

[0172] S111. For an inertial navigation system based on SAfINS inertial satellite ensemble, the navigation support process in the event of satellite communication failure in the dynamic flight map space of an unmanned aerial vehicle (UAV) swarm includes:

[0173] S1111: Based on the configured closed-loop fiber optic gyroscope, accelerometer, and high-performance satellite navigation receiver board, a compactly integrated navigation algorithm is adopted when the satellite signal is valid. The algorithm fuses GNSS pseudorange, carrier phase observations, and inertial measurement unit data through a Kalman filter to calculate the pitch angle, roll angle, heading angle, three-dimensional position coordinates, velocity vector, and precise timestamp of each UAV in the world coordinate system in real time. The compactly integrated navigation algorithm adopts an error-state Kalman filter architecture, and the state vector includes position error, velocity error, attitude error, gyroscope zero bias, and accelerometer zero bias. It should be further explained that this embodiment adopts a strapdown inertial navigation solution process. Attitude update is based on quaternion differential equations and Picard approximation to achieve real-time attitude angle calculation. Velocity update integrates the Earth's gravity field model and Coriolis force compensation formula to calculate three-dimensional velocity. Position update outputs position through a geocentric inertial frame to geocentric Earth-solid frame coordinate transformation algorithm. The test scenario is set to an atmospheric environment with normal to wide temperature range and normal humidity. The UAV's flight altitude covers low to medium altitude and the speed is within the normal flight range. The hardware specifications are based on the low zero-bias stability of fiber optic gyroscopes and the low zero-bias of accelerometers. Error compensation uses extended Kalman filtering to estimate sensor zero bias, scaling factor and installation error in real time. Combined with the multi-position calibration method in the initial alignment stage, attitude error is corrected. The feedforward compensation algorithm eliminates the influence of Earth curvature and meridian convergence angle on position calculation, ensuring that the positioning accuracy is maintained at the meter level in pure inertial mode for a short time, the pitch and roll angle error is within the high accuracy range, and the heading angle error is kept within the controllable range for a short time. The training datasets for the compactly combined navigation algorithm and the Kalman filter algorithm in this embodiment are GNSS pseudorange, carrier phase observations, and inertial measurement unit data. The training parameters are set to a Kalman filter state vector dimension of 15, a process noise covariance matrix diagonal element of 0.01, and an observation noise covariance matrix diagonal element of 0.005. The training parameters for the quaternion differential equation and Picard approximation method in this embodiment are set to 3 Picard approximation iterations.

[0174] S1112. Based on the satellite signal quality monitoring algorithm, when the carrier noise density ratio is lower than the preset threshold or the number of visible satellites is insufficient, it automatically switches to pure inertial navigation mode, uses the quaternion method for attitude calculation, solves the inertial navigation differential equation by numerical integration using the Runge-Kutta method, maintains a pitch and roll angle accuracy of 0.1 degrees, a heading angle accuracy of 1 degree within 120 seconds, and a position calculation accuracy of 3 meters within 60 seconds; the training dataset for the coordinate transformation algorithm in this embodiment comes from the geocentric inertial coordinate data of the UAV, and the training parameters are set as Earth ellipsoid model parameters, such as WGS-84 coordinate system parameters.

[0175] S1113. Based on the sensor error compensation parameters in the universal correction coefficient library, the zero-bias instability of the fiber optic gyroscope and the scale factor error of the accelerometer are corrected in real time. The navigation state of each UAV is uniformly transformed to a global spatiotemporal coordinate system with the world origin as the reference through a coordinate system transformation matrix. It should be further explained that the universal correction coefficient library in this embodiment is configured with corresponding correction coefficients for the position deviation, attitude angle error, and velocity error of the navigation data. The correction method for the navigation data is that the corrected data is equal to the original navigation data multiplied by a proportional correction coefficient plus a fixed compensation coefficient. The correction coefficients are matched and called from the universal correction coefficient library according to the combination conditions of UAV model, sensor model, and sensor installation location. The correction order prioritizes the correction of attitude angle error affected by the aerodynamic characteristics of the aircraft model, and then corrects position error and velocity error in turn. The training dataset of the satellite signal quality monitoring algorithm in this embodiment comes from the carrier noise density ratio and satellite visibility data output by the satellite navigation receiver board. The training method is no training process, directly executing signal quality judgment without loss function. The training parameters are set to a carrier noise density ratio threshold of 40dB-Hz and a satellite visibility threshold of 4.

[0176] S1114. Based on the state transition probability matrix of the Hidden Markov Model, a forward prediction algorithm is used to estimate the state evolution trend of the UAV during the navigation mode switching process. The optimal state sequence is obtained through the Viterbi decoder to ensure the continuity and consistency of navigation data in the spatiotemporal dimension.

[0177] S1115. Based on the cooperative positioning mechanism in the dynamic flight graph space of the UAV swarm, during the period of loss of satellite navigation signal, the extended Kalman filter algorithm is adopted to fuse ultra-wideband ranging data and visual relative measurement information. The navigation error of individual UAVs is compensated through a distributed fusion architecture to maintain the navigation accuracy of the swarm. The cooperative positioning mechanism constructs a relative measurement graph model, transforms the ranging information and visual observation between UAVs into a graph optimization problem, uses the Gauss-Newton method to solve the maximum a posteriori estimate, and maps the solution to the local relative path between the corresponding world origin in the dynamic flight graph space of the UAV swarm, ensuring that flight control can be performed for a preset short period of time even when satellite signal is lost.

[0178] S1116. Construct a navigation performance verification platform based on a digital twin simulation environment. Simulate various satellite signal obstruction scenarios using the Monte Carlo method. Employ statistical hypothesis testing methods to evaluate the maintenance accuracy and recovery capability of the inertial navigation system, ensuring that it meets the requirements for complex mission execution.

[0179] S1117. Based on the status data output by multi-mode navigation, the flight status correction coefficient is adjusted in real time through the status feedback controller in the flight control module. The incremental PID control algorithm is used to dynamically optimize the control parameters to ensure flight stability and mission execution efficiency in any navigation mode.

[0180] This embodiment utilizes the cost-effective sensor combination and advanced algorithms of the SAfINS system to achieve precise fusion and seamless switching of navigation data within the dynamic flight map spatial framework of UAV swarms. This ensures high-precision navigation when satellites are available and short-term accuracy maintenance when signals fail, providing reliable navigation support for collaborative UAV swarm missions in complex environments.

[0181] Next, a specific and complete example will be used to illustrate the entire process. This example is only to illustrate the feasibility at the computational level and does not represent actual values. The specific values ​​can be determined by those skilled in the art through simulation experiments or physical experiments. For example, in the complex task of monitoring landslides in a designated test area, the navigation system adopts a compact combination navigation algorithm. The input of this algorithm is the pseudorange rate data output by the satellite navigation receiver and the angular rate and acceleration data output by the inertial measurement unit. The core relies on the error state Kalman filter to achieve data fusion. The state vector of the filter contains 15 dimensions, including position error, velocity error, attitude error, gyroscope bias, and accelerometer bias. The process noise covariance matrix Q is set as a diagonal matrix with all diagonal elements of 0.01, and the observation noise covariance matrix is ​​set as a diagonal matrix with all diagonal elements of 0.005. The output fused navigation state data includes position, velocity, and attitude information. When the carrier-to-noise ratio is below 40 dB-Hz, the system automatically switches to pure inertial navigation mode. In this mode, the fourth-order Runge-Kutta method is used to solve the navigation differential equations. The algorithm input is the angular rate and acceleration data output in real time by the inertial measurement unit, and the solution frequency is set to 100 Hz. At the same time, a general correction coefficient library is called. The current UAV model and fiber optic gyroscope model are input, and the corresponding compensation parameters are output. A fixed compensation of -0.02 degrees per hour is applied to the gyroscope zero bias, and a fixed compensation of 0.01 meters per square second is applied to the accelerometer zero bias. Then, the attitude is calculated using the quaternion method. The input of this method is the compensated gyroscope angular rate data, and the output is an attitude quaternion used to characterize the real-time attitude of the UAV. Simultaneously, navigation state prediction is performed based on a Hidden Markov Model (HMM). The model input is historical navigation state data, and the state space is defined as normal navigation state, inertial navigation transition state, and abnormal navigation state. The state transition matrix is ​​trained based on 100 sets of historical navigation data, where the probability of switching from a normal navigation state to its own state is 0.95, the probability of switching to an abnormal state is 0.05, and the probability of switching from an abnormal state to a normal navigation state is 0.1. The predicted probability of each navigation state is output. When a single drone in the cluster loses satellite signal, the cooperative positioning mechanism is activated, using an extended Kalman filter to fuse ultra-wideband ranging data and inertial navigation data. The input of this filtering algorithm is the ranging data between the ultra-wideband base station and the drone, and the attitude and velocity data output by the drone's inertial navigation. The ranging noise variance is set to 0.1 square meters, and the process noise covariance matrix is ​​set as a diagonal matrix with all diagonal elements of 0.02. The output is the fused relative position estimate. Subsequently, the relative position is solved iteratively using the Gauss-Newton method. The input of this method is the ranging residual and the initial relative position estimate, and the upper limit of the number of iterations is set to 20. The output is the accurate relative position coordinates.These navigation status data, after multi-source fusion and compensation, are ultimately injected into the world coordinate system of the UAV swarm's dynamic flight map space. The input consists of latitude, longitude, elevation coordinates, and attitude information from the navigation status data, while the output is the UAV's spatiotemporal coordinates in the world coordinate system, providing a reliable spatiotemporal reference for subsequent trajectory correction and swarm collaborative control.

[0182] The modular integrated control platform for complex missions provided in this embodiment achieves end-to-end optimization from mission analysis and environmental perception to flight control through a multi-module collaborative technical architecture. Specifically, by acquiring multi-source data through the acquisition module and using a universal correction coefficient to correct meteorological data deviations in real time, the accuracy of environmental perception is ensured. Secondly, the mission analysis module, based on a pre-trained language model and a rule decomposition engine, transforms complex missions into a set of single sub-missions that conform to the capabilities of a single UAV. Furthermore, by constructing a weighted undirected graph model to quantify the collaborative complexity of missions, the platform achieves refined and executable mission planning. Finally, by establishing flight state and... The probabilistic correlation of mission completion enables quantitative assessment and prediction of uncertainties in the mission execution process, providing a theoretical basis for optimal resource allocation. Furthermore, the flight state correction coefficients obtained based on multi-objective optimization algorithms and digital twin technology, combined with an online learning mechanism, achieve dynamic optimization of control parameters, ensuring the adaptability and stability of the UAV swarm in complex environments. Finally, the integration of the SAfINS navigation system's multi-mode seamless switching capability ensures navigation continuity under various signal conditions, forming a complete closed loop with flight control, significantly improving mission success probability and overall system efficiency, and providing reliable technical support for the application of UAV swarms in complex scenarios such as reconnaissance, logistics, and emergency rescue.

[0183] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

Claims

1. A modular integrated control platform for unmanned aerial vehicles (UAVs) designed for complex tasks, characterized in that: include: Data acquisition module, mission analysis module, atmospheric analysis module, flight control module, and record analysis module; The acquisition module is used to acquire real-time flight status parameters, complex mission text, meteorological data of the flight area, output performance data and fault detection data of the configured UAV swarm, and perform timestamp alignment and filtering. The task parsing module is used to obtain a single subtask set and a single subtask association matrix that can be constructed by a single subtask that can be completed by at most one drone, based on the complex task text and the task parsing algorithm, and to obtain the task complexity for allocating resources by combining the task complexity evaluation algorithm. The atmospheric analysis module is used to invert the real-time flight association adjustment map of the UAV swarm by combining the meteorological data of the flight area after correction by the general correction coefficient, the real-time flight status parameters of each UAV, the single sub-task association matrix and the task complexity, the target optimization algorithm and the preset target optimization function; the target optimization function is constructed by the allocation of resources for the complex task, the probability of completing the single sub-task and the probability of completing the complex task. The record parsing module is used to obtain the performance transition matrix of the UAV swarm by combining the output performance data and fault detection data with the Hidden Markov Algorithm. The flight control module is used to respond to the real-time flight correlation adjustment diagram to make real-time corrections to the real-time flight status of at least one UAV, so that the probability of completing a single subtask and the probability of completing a complex task in the performance transfer matrix of the UAV swarm meet the preset task safe completion probability threshold in real time. The task complexity for obtaining the amount of resources to allocate includes: Based on the historical single subtask complexity and corresponding computational resource amount, a task complexity-computational resource mapping function is constructed. At the same time, based on the historical meteorological data of the corresponding flight area at the corresponding time point, an environment-adjustment resource mapping function is constructed to adjust the computational resource amount required to correct the flight status at the corresponding time point. Each individual subtask is mapped to a vertex, and the task complexity-computation resource mapping function and the environment-adjustment resource mapping function are mapped to the corresponding vertices. At the same time, weighted edges are constructed based on the non-zero correlation degree in the single subtask correlation matrix. A weighted undirected graph model is constructed to represent the collaborative computational complexity between single subtasks and the expected computational complexity of real-time correction of the UAV's flight status during flight.

2. The modular integrated control platform for unmanned aerial vehicles (UAVs) oriented towards complex tasks as described in claim 1, characterized in that, The flight state parameters include at least flight attitude, flight heading, flight position, and speed; the meteorological data of the flight area includes at least indicated airspeed, vacuum speed, barometric altitude, climb rate, angle of attack, and sideslip angle; the universal correction coefficient is obtained by combining the aerodynamic characteristics of the aircraft, sensor hardware characteristics, and sensor installation position parameters with experimental controlled variable simulation experiments, and is used to correct the meteorological data of the flight area collected by different types of UAVs, sensors, and sensor installation positions in real time; the UAV swarm performance transfer matrix is ​​constructed by the output performance index, failure probability, probability of completing a single sub-task, and probability of completing a complex task of each UAV at each time point; the flight correlation adjustment map is constructed by combining the real-time position of each UAV and the flight correlation adjustment parameters at the corresponding position with a graph neural network.

3. The modular integrated control platform for unmanned aerial vehicles (UAVs) oriented towards complex tasks as described in claim 2, characterized in that, The method of obtaining the task complexity for allocating resources also includes: Based on a pre-trained natural language processing model, named entity recognition and semantic role labeling are performed on complex task texts. Combined with a pre-built UAV task domain knowledge base, element standardization and disambiguation are performed to obtain a structured set of task elements. The set of structured task elements is input into a rule-based decomposition engine. Based on the single drone executable principle defined by the upper limit of the output performance of the configured drone swarm, the complex task is decomposed into a single subtask and a single subtask set composed of the single subtasks is constructed. Based on the correlation analysis algorithm combining temporal dependency, spatial proximity, and resource sharing corresponding to a single subtask set, a single subtask correlation matrix is ​​obtained. Temporal dependency is determined based on the execution order constraint of the single subtask, spatial proximity is calculated based on the inverse proportional function of the Euclidean distance of the target position of the single subtask, and resource sharing is determined based on the degree of satisfaction between the type of computational resources required by the single subtask and the amount of resources allocated for the corresponding type of resources.

4. The modular integrated control platform for unmanned aerial vehicles (UAVs) oriented towards complex tasks as described in claim 3, characterized in that, The method of obtaining the task complexity for allocating resources also includes: Based on the weighted undirected graph model, combined with the maximum clique search algorithm and the temporal dependency of each individual subtask, the maximum fully cooperative subgraph in the weighted undirected graph at each time point is obtained, so as to determine the instantaneous cooperative computing resource requirements at each time point. Based on the instantaneous collaborative computing resource requirements and the corresponding available computing resource types at each time point, the task computational complexity at each time point is calculated. A weighted average algorithm is used to obtain the task complexity for allocating resources, based on the expected completion time of the complex task and the computational complexity of the task at each time point.

5. The modular integrated control platform for unmanned aerial vehicles (UAVs) oriented towards complex tasks as described in claim 4, characterized in that, The inversion yields a real-time flight correlation adjustment map of the drone swarm, including: Based on the single subtask set and the single subtask association matrix, the world path of all UAVs during the mission execution is constructed, and the position point corresponding to each time point on the world path is used as the world origin of the world coordinate system of all UAVs at the corresponding time point. The weighted undirected graph model is mapped onto the world coordinate system using the timestamp of each world origin and the execution timestamp of the single subtask corresponding to each vertex, to obtain the dynamic flight graph space of the UAV swarm.

6. The modular integrated control platform for unmanned aerial vehicles (UAVs) oriented towards complex tasks as described in claim 5, characterized in that, The process of acquiring the dynamic flight map space of the UAV swarm includes: Based on the preset flight path corresponding to each vertex of the weighted undirected graph model and the single UAV executable principle, the timestamp corresponding to each world origin on the world path and the estimated timestamp corresponding to each position point in the preset flight path of each vertex are discretized to the plane coordinate system corresponding to the world origin of the corresponding timestamp to obtain the target position point of each UAV in each plane coordinate system. Simultaneously, the non-zero correlation degree in the single subtask correlation matrix is ​​used to construct weighted edges. According to the timestamp of each world origin and the temporal dependency, spatial proximity and resource sharing of the single subtask at the corresponding timestamp, it is decomposed into weighted edges between all UAV target position points in the two-dimensional plane coordinate system under the corresponding timestamp. Based on all target location points in the plane coordinate system corresponding to each world origin, and combined with the weighted edges between all target location points, a static flight diagram of the UAV swarm under each world origin is obtained.

7. The modular integrated control platform for unmanned aerial vehicles (UAVs) oriented towards complex tasks as described in claim 6, characterized in that, The process of acquiring the dynamic flight map space of the drone swarm also includes: Based on the static flight map of the UAV swarm at each world origin, combined with the world path and the preset flight path corresponding to each UAV, a dynamic flight map space of the UAV swarm is obtained. Based on the single sub-task complexity of each UAV from the current target location to the next target location in the dynamic flight map space of the UAV swarm, combined with the corrected meteorological data of the flight area, the task complexity-computational resource mapping function and the environment-adjustment resource mapping function under the corresponding single sub-task, and the resource sharing of the corresponding time interval from the current target location to the next target location, a local resource optimization function and a global resource optimization function and a first resource constraint space are constructed. The global resource optimization function is constructed from the local resource optimization function and the local collaborative resource optimization function, and is used to characterize the computational resource requirements of the internal control parameters of each UAV from the current target location to the next target location, as well as the collaborative computational resource requirements of all UAVs to maintain a safe distance and correctly execute the corresponding single sub-task.

8. The modular integrated control platform for unmanned aerial vehicles (UAVs) oriented towards complex tasks as described in claim 7, characterized in that, The inversion yields a real-time flight correlation adjustment map of the drone swarm, which also includes: Based on the probability of each UAV completing a single subtask within the corresponding time period from the current target location to the next target location in the dynamic flight map space of the UAV swarm, and the probability of all UAVs completing a single subtask within the corresponding time period, a local task completion probability function is constructed. By utilizing the temporal dependency and spatial proximity of the corresponding time period from the current target location to the next target location for each UAV, a second spatiotemporal constraint condition between the UAV task execution priority and the flight space distance is constructed. Based on the resource optimization function corresponding to each UAV, the first resource constraint, the local task completion probability function, and the second spatiotemporal constraint, combined with the flight state parameters of each UAV at each target location and the Hidden Markov algorithm, the flight state probability transition matrix for each UAV to complete the corresponding task, the conditional probability function for successfully completing the corresponding single sub-task, and the global probability function for completing the complex task are obtained. Based on local and global resource optimization functions, local task completion probability functions, flight state probability transition matrix for each UAV to complete a corresponding task, conditional probability function for successfully completing a single sub-task, and global probability function for completing complex tasks, a global multi-objective optimization function is constructed. A global constraint space is constructed using the first resource constraint condition and the second spatiotemporal constraint condition.

9. The modular integrated control platform for unmanned aerial vehicles (UAVs) oriented towards complex tasks as described in claim 8, characterized in that, The inversion yields a real-time flight correlation adjustment map of the drone swarm, which also includes: Based on the global multi-objective optimization function and global constraint space, combined with a preset flight state adjustment coefficient library, and combined with multi-objective optimization search algorithm and twin simulation algorithm, the optimization search is performed with the objectives of minimizing the local resource optimization function and the global resource optimization function, as well as maximizing the local task completion probability function, the conditional probability function of each UAV successfully completing the corresponding single sub-task, and the global probability function of completing the complex task, in combination with the global constraint space. The optimal flight state correction adjustment coefficients of all UAVs in the UAV swarm dynamic flight map space from the current target position point to the next target position point are obtained. The optimal flight state correction adjustment coefficient for each UAV from the current target location to the next target location is mapped to the local relative path connection between the corresponding UAV from the current target location to the next target location. At the same time, the complexity of the corresponding local relative path adjustment process is calculated based on the optimal flight state correction adjustment coefficient in the local relative path connection. Based on the complexity of the corresponding local relative path adjustment process combined with the HSV color space, the color depth of the corresponding local relative path adjustment complexity is obtained. Based on the local relative path of each UAV and the corresponding optimal flight state correction adjustment coefficient, and the color depth combined graph algorithm of the adjustment complexity mapping of the corresponding local relative path, a real-time flight association adjustment graph of all UAV swarms between adjacent world origins is constructed.

Images