An unmanned aerial vehicle control system based on low-altitude satellite and cellular network fusion

Abstract

The application discloses a kind of unmanned plane control systems based on low-altitude satellite and cellular network fusion, it is related to unmanned plane cluster intelligent control technical field, including: attitude cooperative navigation module, for using end cloud cooperation's intelligent sensing system and ant colony algorithm, by deploying multimodal sensor in unmanned plane end, real-time acquisition unmanned plane multimodal attitude and environmental data, and in cloud based on ant colony algorithm dynamic optimization cluster flight attitude and path.The application is realized through end cloud cooperation's intelligent sensing system and ant colony algorithm, real-time attitude synchronization and collision avoidance of multiple unmanned planes in dynamic airspace, unmanned plane end multimodal sensor acquisition high-frequency attitude and environmental data, cloud is based on ant colony algorithm dynamic optimization cluster path, and combined with real-time obstacle information adjustment flight spacing and direction, effectively improve the flight safety and cooperative control precision under high-density airspace, adapt to the flight demand of complex terrain such as urban canyon.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for unmanned aerial vehicle (UAV) swarms, specifically to a UAV control system based on the fusion of low-altitude satellite and cellular networks. Background Technology

[0002] With the improvement of drone hardware performance and the advancement of sensor technology, the functions of control systems are becoming more and more powerful, enabling precise flight control, navigation and positioning, and data transmission. Drone control systems typically include modules such as navigation, communication, autopilot, and mission execution, and can process large amounts of data in real time and make rapid responses.

[0003] For example, the UAV swarm control system in Chinese Patent Publication No. CN120353240A can accurately acquire mission information and equipment information of unmanned equipment, and determine the equipment requirement scheme for each mission based on the equipment information and mission information.

[0004] Existing technologies lack a high-precision control mechanism for real-time attitude coordination among multiple UAVs in dynamic airspace, making it difficult to adapt to attitude interference and collision risks caused by complex terrains such as urban canyons and dense building clusters. Furthermore, in achieving basic attitude coordination, it is difficult to effectively integrate low-altitude satellite navigation signals and cellular network positioning data, resulting in insufficient 3D positioning accuracy and decreased path planning reliability in areas with signal blockage or weak signals. Simultaneously, the lack of forward-looking fault prediction and adaptive recovery capabilities based on multi-source real-time data and historical mission knowledge makes it difficult to achieve seamless mission continuity and cluster self-healing scheduling when UAV performance degrades or experiences sudden anomalies. Therefore, this paper proposes a UAV control system based on the fusion of low-altitude satellite and cellular networks to address the aforementioned problems. Summary of the Invention

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a UAV control system based on the fusion of low-altitude satellite and cellular network, comprising the following modules: The attitude cooperative navigation module is used to adopt an intelligent sensing system and ant colony algorithm in collaboration between the end and the cloud. By deploying multimodal sensors on the UAV, it collects multimodal attitude and environmental data of the UAV in real time, and dynamically optimizes the swarm flight attitude and path in the cloud based on the ant colony algorithm to achieve attitude correction and collision avoidance. The fusion positioning decision module is used to build a knowledge graph-based fusion positioning model. Combined with the 5GMEC (Mobile Edge Computing) collaborative algorithm, it uses the knowledge graph to associate satellite signal strength, cellular base station topology and historical positioning data. Through mobile edge computing, it calculates the optimal fusion positioning weight in real time, performs 3D position correction and dynamic path replanning for UAVs, and improves positioning accuracy in weak signal areas. The spatiotemporal state analysis module is used to introduce a spatiotemporal graph neural network to analyze the spatiotemporal correlation between UAV state, environmental data and historical mission data, in order to identify UAV performance degradation trends and abnormal patterns, and accurately identify UAV performance degradation and abnormal behavior. The fault prediction and early warning module trains a fault prediction model based on federated learning while protecting data privacy, performs early warning and autonomous scheduling switching before faults occur, ensures the continuous and stable operation of the system in complex airspace environments, and realizes forward-looking fault warning and intelligent scheduling switching. The autonomous scheduling and recovery module combines fault warning results with real-time task status to automatically trigger task migration or equipment switching mechanisms when equipment fails or performance degrades, ensuring the continuity and integrity of task execution. Based on real-time health status and task priority, it dynamically adjusts the allocation of UAV tasks within the cluster to achieve cluster-level fault recovery and resource optimization scheduling.

[0006] Preferably, the attitude cooperative navigation module includes an edge-cloud attitude perception unit and a dynamic collision avoidance unit; The edge-cloud attitude perception unit is used to deploy multimodal sensors on the UAV to form an intelligent sensing system, collect attitude and environmental data in real time, and transmit the data to the cloud to form a real-time attitude state matrix of the UAV, providing a high-frequency, high-precision real-time attitude data stream, and providing a reliable perception basis for cloud decision-making and control. The dynamic collision avoidance unit dynamically optimizes the cluster cooperative path based on the ant colony algorithm, generates the optimal flight attitude and path of the cluster in real time, and dynamically adjusts the spacing and flight direction of the UAVs by combining terrain and obstacle information, so as to achieve attitude synchronization and conflict resolution in high-density airspace.

[0007] Preferably, the edge-cloud attitude sensing unit specifically includes: Deploying multimodal sensors on the drone side constitutes an intelligent sensing system, including an inertial measurement unit, a visual sensor, and a lidar, which collects the attitude angle, angular velocity, acceleration, and distance information of surrounding obstacles of each drone in real time. Real-time data collection can enhance the accuracy of perception of its own attitude and the surrounding environment. Through low-altitude satellite communication links, the raw data collected by multimodal sensors is transmitted in real time to the attitude processing platform in the cloud. At the same time, timestamps and UAV identification are marked to construct a time-series attitude data stream, so as to realize reliable data transmission with low latency over long distances and traceability of identity. The received time-series attitude data stream is time-aligned and noise-filtered in the cloud, and then fused to form a high-confidence real-time attitude state matrix of the UAV, thereby improving data quality and outputting unified, accurate and reliable cluster attitude state information.

[0008] Preferably, the dynamic collision avoidance unit specifically includes: Based on the received real-time attitude state matrix of the UAV and environmental obstacle information, a dynamic obstacle map of the current airspace is constructed in the cloud. Combined with the mission target points of each UAV, the pheromone distribution and path nodes of the ant colony algorithm are initialized, providing a real-time environmental modeling basis for dynamic path optimization and improving the accuracy and timeliness of obstacle avoidance decisions. The ant colony algorithm is applied for iterative optimization. Based on the relative distance between drones, attitude consistency index and obstacle proximity, the pheromone concentration is dynamically updated, and the attitude adjustment vector and path point sequence of each drone in the next moment are generated in real time. This enables intelligent optimization of multi-drone collaborative paths, effectively reducing collision risk and maintaining formation stability. Based on the optimized output path and attitude commands, the commands are sent to each UAV execution unit through a low-latency control link to dynamically adjust the flight direction, speed and relative distance between UAVs. This enables collaborative collision avoidance and attitude synchronization in high-density dynamic airspace, ensuring fast and reliable execution of commands and guaranteeing real-time synchronization and safe flight of the cluster in complex airspace.

[0009] Preferably, the fusion positioning decision module includes a multi-source positioning fusion unit and a dynamic path replanning unit; The multi-source positioning fusion unit is used to establish a knowledge graph, associate multi-source information such as satellite signal strength, cellular base station topology and historical positioning data, construct a fusion positioning model, and calculate the optimal fusion positioning weight in real time through the 5GMEC collaborative algorithm to perform dynamic correction of the UAV's three-dimensional position, thereby realizing intelligent fusion and real-time correction of multi-source positioning data and improving the accuracy and robustness of position information. The dynamic path replanning unit, based on the fusion positioning results and combined with environmental changes and mission requirements, optimizes the UAV's flight path in real time and adaptively adjusts the flight path in areas with signal obstruction, ensuring mission continuity and reliability.

[0010] Preferably, the multi-source localization fusion unit specifically includes: A knowledge graph-based fusion positioning model is constructed, with nodes including satellite signal strength, cellular base station identifiers, historical positioning trajectories and environmental features, and edges representing signal propagation relationships and spatiotemporal association rules. Semantic associations between multi-source positioning elements are established to support intelligent data matching and inference. By using 5GMEC nodes to obtain real-time information on the satellite navigation signal quality, cellular base station reception strength, and topology of the current UAV location, and combining this with historical positioning knowledge graphs for matching and confidence assessment, the system achieves real-time comparison between observation data and historical knowledge, thereby improving the reliability of positioning results in complex environments. An adaptive weighted fusion algorithm is adopted to calculate the optimal fusion weight of satellite and cellular positioning data in real time at the MEC edge, dynamically correct the UAV's three-dimensional position, and output high-precision fusion positioning results. This effectively suppresses single source error and can maintain continuous and reliable three-dimensional positioning capability even in areas with signal obstruction.

[0011] Preferably, the dynamic path replanning unit specifically includes: Based on the fusion of positioning results and real-time environmental perception data, signal blockage areas, dynamic obstacles, and airspace control areas on the current and predicted future flight paths are identified, improving the ability to anticipate environmental threats and enhancing the safety and reliability of flight paths. Combining mission requirements and UAV performance constraints, a rolling optimization method is used to generate local path replanning schemes in real time on the MEC side. Flight corridors with good signal coverage and low collision risk are prioritized to optimize flight trajectories, reduce signal interruption and collision risks, and ensure continuous and stable mission execution. The system dynamically updates the waypoint sequence and flight parameters of each UAV in the next stage based on the path replanning results, and ensures the consistency of the path within the cluster through a collaborative communication mechanism. This enables continuous and reliable flight in complex terrain, ensures coordinated flight of the cluster, and improves the adaptability and robustness of the overall mission in complex environments.

[0012] Preferably, the spatiotemporal state analysis module specifically includes: Collect time-series data on the status of drones in historical missions, including power consumption curves, motor vibration spectrum, positioning drift sequences and environmental interference records, construct a spatiotemporal status dataset, form a long-term traceable health record, and provide a data foundation for trend analysis and anomaly tracing; Spatiotemporal graph neural networks are used to model state data. Nodes represent the state vector of the UAV at each time step, and edges represent the spatiotemporal dependencies between states. Deep spatiotemporal features are extracted through graph convolution and temporal convolution to capture the spatiotemporal correlation of state evolution and extract high-level features with discriminative power to improve representation capabilities. Based on the extracted deep spatiotemporal features, the performance degradation trend, abnormal behavior patterns and their correlation with external environmental factors of UAVs are identified, providing a structured state representation for fault prediction, realizing early risk identification and root cause correlation, and supporting accurate prediction and proactive maintenance decisions.

[0013] Preferably, the fault prediction and early warning module specifically includes: By adopting a federated learning framework, fault prediction sub-models based on spatiotemporal features are trained locally on each drone. Only the model parameters are uploaded to the cloud for aggregation, realizing collaborative learning under the protection of data privacy, effectively avoiding the risk of raw data leakage, and enhancing data security and privacy protection. The cloud aggregates various fault prediction sub-models to generate a global fault prediction model, receives spatiotemporal state feature inputs from each UAV in real time, and outputs fault probability and anomaly type labels for a future period of time, significantly improving the model's generalization ability and overall accuracy of fault prediction. When the predicted failure probability exceeds the set threshold, an early warning signal is automatically triggered, and a backup drone scheduling plan is pre-generated to achieve forward-looking autonomous health management and scheduling switchover preparation.

[0014] Preferably, the autonomous scheduling and recovery module specifically includes: Real-time monitoring of the health status and mission progress of each UAV, combined with fault warning results, dynamically assesses the overall mission completion risk and resource availability of the cluster, significantly improving the early identification and response capabilities of cluster mission risks; When a drone is determined to be about to fail or has insufficient performance, a healthy drone is automatically selected from the reserve queue, and a task migration or equipment takeover mechanism is triggered based on task priority and resource matching to ensure seamless and rapid task switching in the event of failure. During mission migration, the remaining UAV mission allocation and path planning within the cluster are dynamically adjusted to ensure continuous mission execution and optimized utilization of airspace resources. This enables seamless handover and cluster self-healing scheduling in case of failure, ensuring cluster-level mission continuity and overall operational safety.

[0015] This invention provides a drone control system based on the fusion of low-altitude satellite and cellular networks. It has the following beneficial effects: (i) This UAV control system based on the fusion of low-altitude satellite and cellular network achieves real-time attitude synchronization and collision avoidance of multiple UAVs in dynamic airspace through an intelligent sensing system and ant colony algorithm with end-cloud collaboration. The multi-modal sensors on the UAV end collect high-frequency attitude and environmental data, and the cloud dynamically optimizes the cluster path based on the ant colony algorithm and adjusts the flight spacing and direction in combination with real-time obstacle information, effectively improving flight safety and collaborative control accuracy in high-density airspace and adapting to the flight needs of complex terrains such as urban canyons.

[0016] (II) This UAV control system based on the fusion of low-altitude satellite and cellular network constructs a knowledge graph-based fusion positioning model. Combined with the 5GMEC collaborative algorithm, it fuses low-altitude satellite navigation signals and cellular network positioning data in real time. By associating historical positioning data with environmental features through the knowledge graph, it dynamically calculates the optimal fusion weight, thereby achieving high-precision three-dimensional position correction and dynamic path replanning in areas with signal obstruction or weak signal, significantly improving the robustness and mission continuity of the positioning system.

[0017] (III) This UAV control system based on the fusion of low-altitude satellite and cellular network adopts a federated learning framework. It trains fault prediction sub-models locally on each UAV and only uploads model parameters for cloud aggregation. This protects the privacy of the original data and realizes the collaborative optimization of the global model. It can output fault probability and anomaly type in real time based on spatiotemporal characteristics, provide early warning of potential faults, support autonomous scheduling and switching, and improve the system's forward-looking health management capabilities and operational stability. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the workflow of an unmanned aerial vehicle (UAV) control system based on the fusion of low-altitude satellites and cellular networks according to the present invention. Figure 2 This is a data flow diagram of an unmanned aerial vehicle (UAV) control system based on the fusion of low-altitude satellites and cellular networks, according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: a drone control system based on the fusion of low-altitude satellite and cellular network, comprising the following modules: The attitude cooperative navigation module is used to employ an edge-cloud collaborative intelligent sensing system and ant colony algorithm. By deploying multimodal sensors on the UAV, it collects multimodal attitude and environmental data of the UAV in real time, and dynamically optimizes the swarm flight attitude and path based on the ant colony algorithm in the cloud. This enables attitude correction and collision avoidance, achieving high-precision attitude synchronization and real-time collision avoidance, significantly improving the safety and stability of the swarm's cooperative flight in complex airspace. The attitude cooperative navigation module includes an edge-cloud attitude perception unit and a dynamic collision avoidance unit. The edge-cloud attitude perception unit is used to deploy multimodal sensors on the UAV to form an intelligent sensing system. It collects attitude and environmental data in real time and transmits the data to the cloud to form a real-time attitude state matrix of the UAV. It provides a high-frequency, high-precision real-time attitude data stream, providing a reliable perception basis for cloud decision-making and control. The multimodal sensors deployed on the UAV constitute an intelligent sensing system, including an inertial measurement unit, a vision sensor, and a lidar. It collects the attitude angle, angular velocity, acceleration, and distance information of surrounding obstacles of each UAV in real time. Real-time collection can enhance the perception accuracy of its own attitude and the surrounding environment. Through a low-altitude satellite communication link, the raw data collected by the multimodal sensors is transmitted to the attitude processing platform in the cloud in real time. At the same time, timestamps and UAV identification are marked to construct a time-series attitude data stream, realizing reliable long-distance, low-latency data transmission and identity traceability. The received time-series attitude data stream is time-aligned and noise-filtered in the cloud, and fused to form a high-confidence real-time attitude state matrix of the UAV, improving data quality and outputting unified, accurate, and reliable cluster attitude state information. It should be noted that the UAV integrates a multimodal intelligent sensing system consisting of an inertial measurement unit (IMU), a visual sensor, and a lidar. The IMU collects attitude angle, angular velocity, and three-axis acceleration data at a frequency of no less than 200Hz. The visual sensor acquires images of the surrounding environment at a frame rate of 30fps. The lidar scans surrounding obstacles at a frequency of 10Hz and outputs distance information. The data from each sensor undergoes preliminary timestamp alignment and encapsulation by an onboard embedded processor, which assigns a unique UAV identifier to form a structured data packet. This data packet is then transmitted uplink to the attitude processing platform in the cloud via a low-altitude satellite communication link, with the transmission link latency controlled within 100 milliseconds. The data transmission bandwidth is no less than 2Mbps to ensure real-time performance and integrity. After receiving the time-series attitude data stream, the cloud performs millisecond-level time synchronization of the multi-source data based on the network time protocol to eliminate the timing deviation caused by transmission delay. Subsequently, the Kalman filter algorithm is used to filter the noise of the angular velocity and acceleration data output by the inertial measurement unit. Visual data is used for motion compensation through feature point matching. The lidar point cloud data is processed by clustering and fitting to extract obstacle distance information. During the fusion process, the extended Kalman filter is used to unify the multimodal data to the same coordinate system and output a high-confidence real-time attitude state matrix containing three-dimensional attitude angles, angular velocities, accelerations and distances to surrounding obstacles. The dynamic collision avoidance unit dynamically optimizes the swarm's collaborative path based on the ant colony algorithm, generating the optimal flight attitude and path for the swarm in real time. Combining terrain and obstacle information, it dynamically adjusts the spacing and flight direction of the drones, achieving attitude synchronization and conflict resolution in high-density airspace. Through intelligent path planning and real-time adjustments, it effectively avoids collisions between drones and with environmental obstacles. Based on the received real-time attitude state matrix of the drones and environmental obstacle information, it constructs a dynamic obstacle map of the current airspace in the cloud. Combining the mission target points of each drone, it initializes the pheromone distribution and path nodes of the ant colony algorithm, providing a real-time environmental modeling foundation for dynamic path optimization and improving the accuracy of obstacle avoidance decisions. In terms of timeliness, the ant colony algorithm is applied for iterative optimization. Based on the relative distance between UAVs, attitude consistency index and obstacle proximity, the pheromone concentration is dynamically updated, and the attitude adjustment vector and path point sequence of each UAV in the next moment are generated in real time. This enables intelligent optimization of multi-UAV collaborative paths, effectively reducing collision risks and maintaining formation stability. Based on the optimized path and attitude commands, they are sent to each UAV execution unit through a low-latency control link to dynamically adjust the flight direction, speed and relative distance between UAVs. This enables collaborative collision avoidance and attitude synchronization in high-density dynamic airspace, ensuring fast and reliable execution of commands and guaranteeing real-time synchronization and safe flight of the cluster in complex airspace. It should be noted that after receiving the real-time attitude state matrix from each UAV in the cloud, a dynamic obstacle map of the current airspace is constructed based on the acquired data and pre-set airspace geographic information. This map represents the airspace in a grid format with a grid resolution of 0.5m × 0.5m × 0.5m. Each grid node stores the probability of obstacle presence, type, and dynamic change trend. Simultaneously, based on the preset mission target point (GPS coordinates) of each UAV, the airspace grid nodes between the current position of each UAV and the target point are marked as a set of passable path nodes. On this basis, the pheromone distribution of the ant colony algorithm is initialized: an initial pheromone concentration τ0 = 0.1 is assigned to each path node, and the current position of each UAV is set as the starting node of the ant colony. The task objective is set as the food source node. The number of ants in the algorithm is set to be consistent with the number of drones in the current airspace. Each ant represents a virtual path exploration agent for one drone. The iterative optimization process of the ant colony algorithm is initiated, with each iteration cycle being 100 milliseconds. In each iteration, each ant (representing one drone) selects the next path node to move to according to the state transition probability formula. The calculation of the state transition probability comprehensively considers: the relative distance between drones (obtained through the real-time attitude state matrix, requiring the distance to be no less than 3 meters during cooperative flight), the attitude consistency index (calculated based on the Euler angle difference of the attitude angles of each drone, with a target difference value of less than 5°), and the proximity of the current node to obstacles in the dynamic obstacle map. The path selection is measured by the Euclidean distance from the node to the nearest obstacle mesh (with a safety threshold of 2 meters). After path selection, the pheromone concentration on relevant nodes is dynamically updated based on path quality (safety margin). The volatile factor ρ is set to 0.2, and the pheromone enhancement coefficient Q is set to 1.0. After a preset 50 iterations or convergence determination, an optimized path from the current position to the next stage target point is output for each UAV. This path is represented by a series of continuous path node sequences, and corresponding attitude adjustment vectors are generated synchronously. The optimized UAV path point sequences (each path point includes latitude, longitude, and elevation coordinates and a preset timestamp for arrival at the point) and attitude adjustment vectors are encapsulated into standard MAVLink command messages and transmitted via a low-latency control link (using...). Using the UDP protocol (with an end-to-end latency requirement of less than 50 milliseconds and a data packet transmission frequency of 20Hz), commands are sent to the flight control execution units of each UAV. Based on the received commands and its own real-time status, the UAV's onboard flight controller calculates the underlying execution quantities of engine throttle and control surface deflection through a PID controller, dynamically adjusting the flight direction, speed, and relative distance to other UAVs in the cluster. Throughout the flight process, a closed-loop process from perception, map building, optimization to command issuance is continuously executed to achieve millisecond-level collaborative collision avoidance and attitude synchronization control of multiple UAVs in a high-density dynamic airspace. At the same time, a monitoring mechanism is set up. If a UAV does not receive a valid command or reports an abnormal status for two consecutive cycles, the collision avoidance hovering plan is automatically triggered. The expression for calculating the state transition probability is as follows: ; ; In the formula: For at any time ,Ant From node Transfer to node The probability of; For the edge The pheromone concentration on the surface is updated iteratively over time; For heuristic functions, retrieve nodes. The reciprocal of the Euclidean distance to the target node is used to guide the ant towards the target. The weight of the pheromone-inspired factor is set to 1.0 to control the influence of pheromone concentration on selection. The distance heuristic factor weight, which controls the influence of the distance heuristic term, is set to 2.0; To coordinate the weighting of safety factors and control their weighting in path selection, the weighting is set to 3.0; For ants The set of nodes that are allowed to be accessed in the current state; The variable to be summed represents a node among all allowed access nodes; To assess the safety of collaborative flight among drones, a comprehensive evaluation is conducted, including distance maintenance, attitude consistency, and obstacle avoidance. The weights for the relative distance term, attitude consistency term, and obstacle proximity term are 0.4, 0.3, and 0.3, respectively, satisfying the following conditions: ; The real-time relative distance between the drone and other drones in the cluster is obtained from the attitude state matrix; The minimum safe distance threshold between drones is set to 3 meters to avoid collisions. This is the Euler angle difference norm between the current drone and the cluster's desired attitude, reflecting the degree of attitude synchronization. The target attitude difference should be less than 5°. If it exceeds this, the value of this term is reduced to suppress the probability of movement in that direction. The Euclidean distance from the current node to the nearest obstacle is retrieved from the dynamic obstacle map; The minimum safe distance threshold from obstacles is set to 2 meters to ensure flight safety. Indicates when A positive penalty term is generated if the condition is met; otherwise, the value is 0. The fusion positioning decision module is used to build a knowledge graph-based fusion positioning model. Combined with the 5GMEC (Mobile Edge Computing) collaborative algorithm, it uses the knowledge graph to associate satellite signal strength, cellular base station topology and historical positioning data. Through mobile edge computing, it calculates the optimal fusion positioning weight in real time, performs 3D position correction and dynamic path replanning for UAVs, improves positioning accuracy in weak signal areas, significantly improves positioning accuracy and path reliability in signal-obstructed areas, and ensures continuous mission execution. The spatiotemporal state analysis module is used to introduce spatiotemporal graph neural networks to analyze the spatiotemporal correlation between UAV state, environmental data and historical mission data, in order to identify UAV performance degradation trends and abnormal patterns, accurately identify UAV performance degradation and abnormal behavior, and provide data support for fault early warning and health management. The fault prediction and early warning module trains a fault prediction model based on federated learning while protecting data privacy. It performs early warning and autonomous scheduling switching before faults occur, ensuring the continuous and stable operation of the system in complex airspace environments. It realizes forward-looking fault warning and intelligent scheduling switching, and improves the system's self-healing ability and operational continuity under abnormal conditions. The autonomous scheduling and recovery module combines fault warning results with real-time task status to automatically trigger task migration or equipment switching mechanisms when equipment fails or performance degrades, ensuring the continuity and integrity of task execution. Based on real-time health status and task priority, it dynamically adjusts the allocation of UAV tasks within the cluster, realizing cluster-level fault recovery and resource optimization scheduling, enabling rapid replacement of faulty UAVs and seamless task continuation, and ensuring the overall task completion rate and resource utilization efficiency of the cluster.

[0021] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: the fusion positioning decision module includes a multi-source positioning fusion unit and a dynamic path replanning unit; The multi-source positioning fusion unit is used to establish a knowledge graph, linking multi-source information such as satellite signal strength, cellular base station topology, and historical positioning data to construct a fusion positioning model. It then uses the 5GMEC collaborative algorithm to calculate the optimal fusion positioning weights in real time, dynamically correcting the UAV's 3D position. This achieves intelligent fusion and real-time correction of multi-source positioning data, improving the accuracy and robustness of location information. The knowledge graph-based fusion positioning model is constructed, with nodes including satellite signal strength, cellular base station identifiers, historical positioning trajectories, and environmental features. Edges represent signal propagation relationships and spatiotemporal association rules, establishing semantic relationships between multi-source positioning elements to support... Intelligent data matching and inference are achieved by using 5GMEC nodes to obtain the satellite navigation signal quality, cellular base station reception strength and topology of the current UAV location in real time. This data is then combined with historical positioning knowledge graphs for matching and confidence assessment, enabling real-time comparison between observed data and historical knowledge. This improves the reliability of positioning results in complex environments. An adaptive weighted fusion algorithm is used to calculate the optimal fusion weight of satellite and cellular positioning data in real time at the MEC edge, dynamically correcting the UAV's 3D position and outputting high-precision fusion positioning results. This effectively suppresses single source errors and maintains continuous and reliable 3D positioning capabilities even in areas with signal obstruction. It should be noted that the fusion positioning knowledge graph, built upon a graph structure, digitally models the positioning elements of the physical world and their relationships. The graph nodes comprise four main types of entities: satellite signal strength nodes, storing the carrier-to-noise ratio, elevation angle, and multipath error estimates for each navigation satellite at a specific spatiotemporal point; cellular base station identifier nodes, recording the base station's global cell identifier, geographic coordinates, operating frequency band, and transmit power; historical positioning trajectory nodes, sequentially linking the UAV's past positioning coordinates, speed, and heading; and environmental feature nodes, representing static geographic information such as terrain elevation, building outlines, and signal obstruction areas. Edges between nodes define the association rules between entities: satellites and base stations are connected via edge connections based on signal propagation relationships. Its attributes include line-of-sight and non-line-of-sight propagation model parameters; positioning trajectory points are connected by spatiotemporal correlation edges, and transition probabilities are established based on temporal proximity and kinematic constraints; the map is stored and indexed through a graph database; during the positioning process, 5GMEC nodes deployed at the network edge aggregate multi-source observation data from the UAV in real time. Specifically, the MEC obtains the raw measurement values ​​reported by the UAV through the air interface: for satellite navigation, this includes pseudorange, carrier phase, Doppler shift, and corresponding ephemeris data of visible satellites, and calculates the position accuracy factor and integrity risk value at a frequency of 1Hz; for cellular networks, it collects the reference signal received power, time difference of arrival, and base station data of the serving cell and at least three neighboring cells. The coordinates are sampled at a period of 100 milliseconds. Simultaneously, the MEC retrieves a pre-built knowledge graph subgraph from its local cache or central server. This subgraph covers all relevant nodes and edges within the approximate current area of ​​the UAV (within a 500-meter radius). Real-time observation data is matched against historical patterns in the graph, and the spatiotemporal consistency of the cellular base station topology is verified. Based on the matching results, the system assigns a dynamic confidence score to each set of observation data. The score is based on the chi-square test of the innovation covariance, used to characterize the reliability of the data source in the current environment. An adaptive weighted algorithm based on a Kalman filter framework is employed. The state vector is defined as the UAV's three-dimensional position, velocity, and receiver clock error in the northeast-northeast coordinate system. The vector is composed of the satellite pseudorange equations and the cellular time difference of arrival equations. In each fusion cycle (200 milliseconds), the following steps are performed: Based on the output confidence score, the initial weight of each observation is dynamically calculated. The weight is proportional to the square of the confidence score. Then, iterative reweighted least squares estimation is performed. In each iteration, the weights are further adjusted according to the residual size. Observations with very large residuals (more than 3 times the standard deviation) are reduced in weight or removed. During the fusion process, the final fusion weight of satellite and cellular data is determined by the observation noise covariance matrix calculated in real time. The elements of this matrix are dynamically updated according to the signal quality index (carrier-to-noise ratio). After filtering and fusion, the corrected three-dimensional position coordinates of the UAV are output. The dynamic path replanning unit, based on fused positioning results and combined with environmental changes and mission requirements, optimizes the UAV's flight path in real time, adaptively adjusts the flight path in signal-blocked areas to ensure mission continuity and reliability, dynamically optimizes the flight path according to the real-time environment and mission status to ensure continuous and reliable mission execution in complex terrain, identifies signal-blocked areas, dynamic obstacles and airspace control areas on the current and predicted future flight paths based on fused positioning results and real-time environmental perception data, improves the ability to anticipate environmental threats, and enhances the safety and reliability of the flight path, and generates local path replanning schemes in real time on the MEC side using a rolling optimization method, combined with mission requirements and UAV performance constraints, prioritizing flight corridors with good signal coverage and low collision risk, optimizing flight trajectories, reducing signal interruption and collision risks, ensuring continuous and stable mission execution, dynamically updates the waypoint sequence and flight parameters of each UAV in the next stage according to the path replanning results, and ensures path consistency within the cluster through a collaborative communication mechanism, adapting to continuous and reliable flight in complex terrain, ensuring cluster flight coordination and consistency, and improving the overall mission's adaptability and robustness in complex environments; It should be noted that during the flight of the UAV swarm, based on the high-precision three-dimensional position coordinates output by the multi-source positioning fusion unit (horizontal positioning accuracy better than 1.5 meters, vertical accuracy better than 3.0 meters) and the environmental perception data acquired in real time by the airborne multimodal intelligent sensing system, the flight airspace is dynamically modeled in real time. Using the current position of the UAV as a reference, the airspace within a range of 50 to 200 meters ahead is scanned and analyzed along the current flight path to identify three key areas: signal obstruction areas, dynamic obstacles, and airspace control areas. Specifically, the determination of signal obstruction areas is based on... The satellite navigation signal carrier-to-noise ratio is below 35dB-Hz and the cellular network reference signal reception power is below -110dBm. This is combined with pre-stored terrain elevation and building outline data in the knowledge graph for joint verification. Dynamic obstacles are identified and their trajectories predicted through clustering of LiDAR point clouds over three consecutive frames and visual motion estimation, with the prediction time domain being the next 3 seconds. The coordinate range and altitude restrictions of the airspace control area are dynamically acquired from the air traffic management service interface via the 5G network at a frequency of 1Hz. The identification results are presented in the form of an airspace grid (resolution 0.5m × 0.5m × 0.5m).The data is stored in 5-meter intervals, with each grid cell labeled with its attributes, confidence level, and prediction validity period, forming a dynamic airspace situation map for path planning. Based on the constructed dynamic airspace situation map, at the mobile edge computing node, according to the preset performance constraints of the mission target point, the UAV's maximum turning rate (≤30 degrees / second), maximum climb / descent rate (±5 meters / second), and minimum safety separation (3 meters), a rolling time-domain control method is used for real-time local path optimization. The optimization cycle is synchronized with the dynamic airspace situation map update cycle and is set to 200 milliseconds. Within each optimization cycle, starting from the UAV's current position, the predicted flight space for the next 5 seconds is used as the optimization window, establishing a system with the goal of minimizing total flight risk and constrained by kinematic equations. The optimization model's objective function comprehensively balances three indicators: signal coverage quality (40%), collision risk (40%), and mission progress deviation (20%). Signal coverage quality is quantified by the predicted signal strength (satellite and cellular fusion value) of the grid where the future path point is located. Collision risk is calculated by the minimum distance between the path point and the predicted trajectory of dynamic obstacles (safety threshold 2 meters) and the relative distance with neighboring UAVs (safety threshold 3 meters). The solution uses a sequential quadratic programming algorithm, with the computation time controlled within 50 milliseconds. The final output is an optimal local path composed of a continuous sequence of waypoints (waypoint spacing 1 meter), ensuring that this path preferentially traverses areas with good signal coverage (predicted carrier-to-noise ratio > 40 dB-Hz and RS). Airspace corridors with low collision risk and RP > -105dBm are selected. The output local path waypoint sequence (including coordinates in the northeast-northeast coordinate system, preset arrival timestamp, and expected speed) is encapsulated into standard waypoint commands. Based on the relative relationship between the UAV's current state and the next waypoint, specific attitude adjustment commands (expected roll angle, pitch angle) and throttle are calculated using a PID control law, forming a complete flight control command package. This command package is sent to the corresponding UAV flight control system via a low-latency data link (UDP protocol, end-to-end latency less than 50 milliseconds, transmission frequency 20Hz). To ensure the flight consistency and overall safety of the cluster in complex terrain, a cooperative communication mechanism is introduced: on the one hand, each UAV successfully... After receiving and confirming the execution of a new waypoint sequence, a simplified status confirmation message is broadcast via the cluster's self-organizing network. Meanwhile, the MEC node, acting as a coordinator, continuously monitors the deviation between the command execution status of each drone and its actual trajectory (allowing a deviation radius of 1 meter). If a drone fails to update its status or exceeds the trajectory deviation limit for two consecutive control cycles (400 milliseconds), it is determined to be abnormal and immediately broadcasts the abnormal position information to other drones in the cluster, triggering its local collision avoidance algorithm for emergency avoidance. Simultaneously, the MEC node reassesses the global task allocation and activates contingency plans to ensure continuous and reliable task execution. Through this closed-loop process, continuous, reliable, and coordinated flight of the drone cluster is achieved in complex airspace with fluctuating signals and numerous obstacles. The spatiotemporal state analysis module specifically includes: collecting time-series state data from historical UAV missions, including power consumption curves, motor vibration spectra, positioning drift sequences, and environmental interference records; constructing a spatiotemporal state dataset to form a long-term traceable health record, providing a data foundation for trend analysis and anomaly tracing; using a spatiotemporal graph neural network to model the state data, where nodes represent the UAV's state vector at each moment and edges represent the spatiotemporal dependencies between states; extracting deep spatiotemporal features through graph convolution and temporal convolution to capture the spatiotemporal correlation of state evolution; extracting discriminative high-level features to improve representation capabilities; and based on the extracted deep spatiotemporal features, identifying the UAV's performance degradation trends, abnormal behavior patterns, and their correlation with external environmental factors, providing structured state representations for fault prediction, achieving early risk identification and root cause correlation, and supporting accurate prediction and proactive maintenance decisions. It should be noted that in actual deployment, multi-dimensional time-series status data of each UAV during mission execution are periodically collected and stored to construct a spatiotemporal dataset for UAV status monitoring. Specific data collection items include: recording the remaining battery percentage at a frequency of at least 1Hz to generate a battery consumption curve; acquiring the triaxial vibration acceleration of the motor using a vibration sensor at a sampling frequency of 200Hz and calculating its frequency domain characteristics to form a vibration spectrum sequence; recording the horizontal and vertical positioning deviations output by the fusion positioning module at a frequency of at least 10Hz to construct a positioning drift time-series sequence; and simultaneously collecting environmental interference data, including real-time satellite carrier-to-noise ratio, cellular network reference signal received power, and ambient temperature recorded at a frequency of 1Hz. All data are collected during the process. Each data point is precisely timestamped and includes a unique UAV identifier. The data is then uploaded to cloud storage via 5G network or low-altitude satellite link, forming a structured spatiotemporal state dataset. The data storage period is no less than 90 days, and the daily data volume per UAV is controlled within 500MB. Based on this dataset, a spatiotemporal graph neural network is used to model the UAV state evolution. During modeling, the state of each UAV at a specific moment is defined as a graph node. The node feature vector consists of 12 dimensions, including: battery percentage, vibration frequency amplitude, horizontal positioning deviation, vertical positioning deviation, ambient temperature, satellite carrier-to-noise ratio, RSRP value, flight altitude, airspeed, motor temperature, and two reserved extension dimensions. The edges are divided into two categories: temporal edges are established between adjacent time steps of the same UAV, with weights decaying exponentially according to the time interval; spatial edges are established between different UAVs at the same time, with weights calculated according to their relative distance using a Gaussian kernel function, and the weights are set to zero when the distance exceeds 100 meters. The network structure adopts two layers of spatiotemporal graph convolutional modules and one layer of temporal convolutional modules. The graph convolutional layer outputs a dimension of 64, and the temporal convolutional layer uses dilated convolution with an expansion coefficient of 2. Finally, a 32-dimensional deep spatiotemporal feature representation of each node is output. Using the extracted deep spatiotemporal features, the performance degradation trend of UAVs is identified and abnormal behavior patterns are detected. The performance degradation trend is realized by constructing a long-term and short-term prediction model based on an attention mechanism. This model uses 30 consecutive time steps as the time interval. The system takes the feature sequence of the step as input to predict key state changes within the next 10 time steps. When the predicted vibration frequency amplitude continuously exceeds the baseline value by 15% or the positioning deviation increases by more than 5% per day, it is determined that there is a risk of performance degradation. Abnormal behavior pattern detection adopts a density-based clustering algorithm to identify abnormal outliers in the feature space. When the Mahalanobis distance between the node feature and the normal cluster exceeds 3 standard deviations for a continuous period, an abnormal alarm is triggered. At the same time, a correlation analysis model between state features and environmental parameters is established. By calculating the correlation coefficient matrix between features and environmental variables such as temperature and signal strength, the influence of external environmental factors on the state of the UAV is identified. All analysis results are output to the fault prediction and early warning module in the form of structured state representation. The fault prediction and early warning module specifically includes: using a federated learning framework, training fault prediction sub-models based on spatiotemporal features locally on each UAV, uploading only the model parameters to the cloud for aggregation, achieving collaborative learning under data privacy protection, effectively avoiding the risk of original data leakage, enhancing data security and privacy protection, aggregating each fault prediction sub-model in the cloud to generate a global fault prediction model, receiving spatiotemporal state feature inputs from each UAV in real time, outputting fault probability and anomaly type labels for a future period of time, significantly improving the model's generalization ability and overall accuracy of fault prediction, automatically triggering an early warning signal when the predicted fault probability exceeds a set threshold, and pre-generating a backup UAV scheduling plan, realizing forward-looking autonomous health management and scheduling switch preparation, achieving proactive early warning and rapid response before faults, ensuring mission continuity and system stability; It should be noted that each drone independently trains a fault prediction sub-model on its local computing resources based on its own collected time-series state data. This sub-model takes the extracted 32-dimensional deep spatiotemporal feature vector as input and outputs the fault probability and anomaly type for the next 10 minutes. Local training employs a lightweight temporal neural network structure, with each training cycle lasting 6 hours. Each training session uses the latest local data from the past 24 hours. After training, only the model weight parameters are uploaded to the cloud-based federated learning aggregation server via an encrypted channel. The original state data remains locally on the drone to protect data privacy. The upload frequency is set to once every 12 hours, or... When the loss function converges to a set threshold (loss change rate less than 0.1% for three consecutive training rounds), an immediate upload is triggered, and the drone continuously performs incremental learning of the local model. The cloud-based federated learning server receives encrypted model parameters from each drone and generates a global fault prediction model using a weighted average-based federated aggregation algorithm. The aggregation weights are dynamically calculated based on the timeliness of the parameters uploaded by each drone, the amount of local training data, and the model's performance on the validation set. Data volume accounts for 60% of the weight, and model performance accounts for 40%. The generated global fault prediction model also has the ability to receive 32-dimensional feature vector inputs and can output the fault probability and anomaly type for the next 10 minutes. The global fault prediction model is distributed to each UAV and mobile edge computing node via a secure link. It is updated every 24 hours or immediately when the aggregated model's prediction accuracy improves by more than 2% compared to the previous version. The global model deployed on MEC nodes can receive real-time feature vectors reported by each UAV and return fault prediction results with a latency of less than 50 milliseconds, achieving near real-time cluster health monitoring. A tiered early warning threshold is set: when the fault probability output by the global fault prediction model exceeds 0.7, a level 1 early warning is triggered, notifying the ground station operator; when it exceeds 0.85, a level 2 automatic response early warning is triggered. After the early warning signal is generated, it is immediately combined with real-time data... Based on the mission situation, the health status of each UAV, and remaining resources, a backup scheduling plan is pre-generated. This backup scheduling plan includes: designating a backup UAV identifier (selected from currently idle or low-priority task UAVs), planning the takeover path from the backup UAV to the location of the predicted fault UAV, calculating the estimated task takeover time (not exceeding 120 seconds), and reallocating the remaining workload of the original task. The plan is simulated and verified through a digital twin system to ensure its feasibility. After that, it is put into standby mode. Once the actual condition of the UAV deteriorates or a fault occurs, the plan can be executed automatically or after manual approval, achieving seamless task switching and self-healing scheduling of the cluster, ensuring task continuity and overall reliability. The autonomous scheduling and recovery module specifically includes: real-time monitoring of the health status and task execution progress of each UAV; dynamic assessment of the overall task completion risk and resource availability of the cluster based on fault warning results; significantly improving the early identification and response capability of cluster task risks; when a UAV is determined to be about to fail or has insufficient performance, automatically selecting a healthy UAV from the reserve queue; and triggering a task migration or equipment takeover mechanism based on task priority and resource matching degree to ensure seamless and rapid task switching in case of failure; during the task migration process, dynamically adjusting the task allocation and path planning of the remaining UAVs in the cluster to ensure continuous task execution and optimized utilization of airspace resources; achieving seamless succession and cluster self-healing scheduling in case of failure; and ensuring cluster-level task continuity and overall operational safety. It should be noted that during the execution of the drone swarm mission, a health monitoring service deployed on mobile edge computing nodes and in the cloud collects real-time health status data and mission progress information for each drone at a frequency of no less than 1Hz. Health status data includes remaining battery power, motor vibration spectrum characteristics, positioning deviation, and sensor data integrity. Mission progress information includes current waypoint completion rate, estimated remaining flight time, and achievement status of key mission nodes. The monitoring service's comprehensive fault early warning module outputs predicted fault probability (threshold set at 0.7) and anomaly type labels, combined with the overall resource availability of the swarm (including the number of available drones, average remaining battery power, and airspace channel capacity), dynamically... The risk assessment for cluster task completion employs a weighted scoring method, with task priority accounting for 60% and resource matching degree for 40%. When the overall risk score exceeds a preset threshold of 0.8, contingency plans must be initiated. All assessment results are stored in structured JSON format. If a drone's failure probability is determined by the fault prediction model to continuously exceed 0.85 within the next 10 minutes, or if its key performance indicators exceed safe operating ranges, an automatic equipment takeover mechanism is triggered. This mechanism first selects a suitable backup drone from the reserve queue. Selection criteria include: a health score of at least 85 (out of 100), currently in an idle or low-priority task state, and remaining battery power. If the number of drones exceeds 50% and the straight-line distance to the target drone does not exceed 500 meters, a backup drone is selected. Based on the task priority mapping table and resource matching algorithm, a task migration instruction set is generated within 100 milliseconds. The instruction set includes the target drone's task identifier, remaining waypoint sequence, communication key, and sensor parameter configuration. During the migration, feasibility is verified using a digital twin system to ensure the takeover path meets airspace safety interval requirements (minimum distance between drones 3 meters). After successful verification, the instructions are sent via an encrypted link. During the execution of the task migration instructions, the task allocation and real-time path planning of the remaining drones in the cluster are dynamically adjusted. Task reallocation is performed based on an improved Hungarian algorithm, prioritizing the most efficient and efficient drones. With the goal of minimizing overall mission downtime, and considering the payload capacity and endurance constraints of each drone, the path planning module updates the airspace situation map every 200 milliseconds, resolves conflicts in affected flight corridors in real time, and ensures that the distance between any two drones is always no less than the safety threshold of 3 meters. The entire migration and replanning process requires the total mission downtime to not exceed 120 seconds, and the mission completion rate to be no less than 95% of the original plan. After the self-healing scheduling is completed, the cluster resource status table and mission log are automatically updated, and a recovery report is generated, including the actual takeover time, changes in resource utilization, and possible residual risks, thus completing the seamless handover and autonomous recovery loop of the cluster in the event of a failure.

[0022] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0023] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A UAV control system based on the fusion of low-altitude satellite and cellular network, characterized in that, Includes the following modules: The attitude cooperative navigation module is used to employ an intelligent sensing system and ant colony algorithm that combines edge and cloud. By deploying multimodal sensors on the UAV, it collects multimodal attitude and environmental data of the UAV in real time, and dynamically optimizes the swarm flight attitude and path in the cloud based on the ant colony algorithm. The fusion positioning decision module is used to build a knowledge graph-based fusion positioning model. Combined with the 5GMEC collaborative algorithm, it uses the knowledge graph to associate satellite signal strength, cellular base station topology and historical positioning data, and calculates the optimal fusion positioning weight in real time through mobile edge computing to perform 3D position correction and dynamic path replanning for UAVs. The spatiotemporal state analysis module is used to introduce a spatiotemporal graph neural network to analyze the spatiotemporal correlation between UAV state, environmental data and historical mission data, in order to identify UAV performance degradation trends and abnormal patterns. The fault prediction and early warning module trains a fault prediction model based on federated learning while protecting data privacy, and performs early warning and autonomous scheduling switching before faults occur. The autonomous scheduling and recovery module combines fault warning results with real-time task status to automatically trigger task migration or equipment switching mechanisms when equipment fails or performance degrades. It also dynamically adjusts the allocation of drone tasks within the cluster based on real-time health status and task priority.

2. The UAV control system based on the fusion of low-altitude satellite and cellular network according to claim 1, characterized in that: The attitude cooperative navigation module includes an edge-cloud attitude perception unit and a dynamic collision avoidance unit. The edge-cloud attitude perception unit is used to deploy multimodal sensors on the UAV to form an intelligent sensing system, collect attitude and environmental data in real time, and transmit the data to the cloud to form a real-time attitude state matrix of the UAV. The dynamic collision avoidance unit dynamically optimizes the swarm cooperative path based on the ant colony algorithm, generates the optimal flight attitude and path of the swarm in real time, and dynamically adjusts the spacing and flight direction of the UAVs by combining terrain and obstacle information.

3. The UAV control system based on the fusion of low-altitude satellite and cellular network according to claim 2, characterized in that: The edge-cloud attitude sensing unit specifically includes: A multimodal sensor system is deployed on the drone to form an intelligent sensing system, including an inertial measurement unit, a vision sensor and a lidar, to collect the attitude angle, angular velocity, acceleration and distance information of surrounding obstacles of each drone in real time. The raw data collected by the multimodal sensors is transmitted in real time to the attitude processing platform in the cloud via a low-altitude satellite communication link. At the same time, timestamps and UAV identification are added to construct a time-series attitude data stream. The received time-series attitude data stream is time-aligned and noise-filtered in the cloud, and then fused to form a high-confidence real-time attitude state matrix for the UAV.

4. The UAV control system based on the fusion of low-altitude satellite and cellular network according to claim 2, characterized in that: The dynamic collision avoidance unit specifically includes: Based on the received real-time attitude state matrix of the UAV and environmental obstacle information, a dynamic obstacle map of the current airspace is constructed in the cloud, and the pheromone distribution and path nodes of the ant colony algorithm are initialized in combination with the mission target points of each UAV. The ant colony algorithm is applied for iterative optimization. Based on the relative distance between drones, attitude consistency index and obstacle proximity, the pheromone concentration is dynamically updated, and the attitude adjustment vector and path point sequence of each drone in the next moment are generated in real time. Based on the optimized output path and attitude commands, the commands are sent to each UAV execution unit through a low-latency control link to dynamically adjust the flight direction, speed, and relative distance between UAVs.

5. A UAV control system based on the fusion of low-altitude satellite and cellular network according to claim 2, characterized in that: The fusion positioning decision module includes a multi-source positioning fusion unit and a dynamic path replanning unit; The multi-source positioning fusion unit is used to establish a knowledge graph, associate multi-source information such as satellite signal strength, cellular base station topology and historical positioning data, construct a fusion positioning model, and calculate the optimal fusion positioning weight in real time through the 5GMEC collaborative algorithm to perform dynamic correction of the UAV's three-dimensional position. The dynamic path replanning unit optimizes the UAV's flight path in real time based on the fused positioning results and combined with environmental changes and mission requirements, and adaptively adjusts the flight path in areas with signal obstruction.

6. The UAV control system based on the fusion of low-altitude satellite and cellular network according to claim 5, characterized in that: The multi-source localization fusion unit specifically includes: A knowledge graph-based fusion positioning model is constructed, where nodes include satellite signal strength, cellular base station identifiers, historical positioning trajectories and environmental features, and edges represent signal propagation relationships and spatiotemporal association rules. The satellite navigation signal quality, cellular base station reception strength and topology of the current area of ​​the UAV are obtained in real time through the 5GMEC node, and the historical positioning knowledge graph is combined for matching and confidence assessment. An adaptive weighted fusion algorithm is adopted to calculate the optimal fusion weight of satellite and cellular positioning data in real time at the MEC edge, dynamically correct the three-dimensional position of the UAV, and output high-precision fusion positioning results.

7. A UAV control system based on the fusion of low-altitude satellite and cellular network according to claim 5, characterized in that: The dynamic path replanning unit specifically includes: Based on the fusion of positioning results and real-time environmental perception data, signal blockage areas, dynamic obstacles, and airspace control areas are identified on the current and predicted future flight paths. Combining mission requirements and UAV performance constraints, a rolling optimization method is used to generate local path replanning schemes in real time on the MEC side, prioritizing flight corridors with good signal coverage and low conflict risk. The next-stage waypoint sequence and flight parameters of each UAV are dynamically updated based on the path replanning results, and path consistency within the cluster is ensured through a collaborative communication mechanism.

8. A UAV control system based on the fusion of low-altitude satellite and cellular network according to claim 5, characterized in that: The spatiotemporal state analysis module specifically includes: Collect time-series data of the state of the UAV in historical missions, including power consumption curves, motor vibration spectrum, positioning drift sequence and environmental interference records, and construct a spatiotemporal state dataset; The state data is modeled using a spatiotemporal graph neural network. Nodes represent the state vector of the UAV at each time step, and edges represent the spatiotemporal dependencies between states. Deep spatiotemporal features are extracted through graph convolution and temporal convolution. Based on the extracted deep spatiotemporal features, we can identify the performance degradation trend, abnormal behavior patterns, and their correlation with external environmental factors of UAVs.

9. A UAV control system based on the fusion of low-altitude satellite and cellular network according to claim 8, characterized in that: The fault prediction and early warning module specifically includes: By adopting a federated learning framework, fault prediction sub-models based on spatiotemporal features are trained locally on each drone, and only the model parameters are uploaded to the cloud for aggregation, thus achieving collaborative learning under the protection of data privacy. The cloud aggregates various fault prediction sub-models to generate a global fault prediction model, receives spatiotemporal state feature inputs from each UAV in real time, and outputs the fault probability and anomaly type label for a future period of time. When the predicted failure probability exceeds the set threshold, an early warning signal is automatically triggered, and a backup drone scheduling plan is pre-generated.

10. A UAV control system based on the fusion of low-altitude satellite and cellular network according to claim 9, characterized in that: The autonomous scheduling and recovery module specifically includes: Real-time monitoring of the health status and mission progress of each drone, combined with fault warning results, dynamically assesses the overall mission completion risk and resource availability of the cluster; When a drone is determined to be about to fail or has insufficient performance, a healthy drone is automatically selected from the reserve queue, and a task migration or equipment takeover mechanism is triggered based on task priority and resource matching. During mission migration, the remaining UAV mission allocation and path planning within the cluster are dynamically adjusted to ensure continuous mission execution and optimized utilization of airspace resources.

Images