Autonomous control method and system for unmanned aerial vehicle disconnection

Abstract

The application provides an autonomous control method and system for unmanned aerial vehicle disconnection, and relates to the technical field of unmanned aerial vehicle autonomous control. The method comprises the following steps: acquiring unmanned aerial vehicle data containing multi-source heterogeneous data and task semantic information, and generating a first situation graph by situation analysis; processing the graph and the semantic information by using a lightweight two-stage prediction method to generate a spatial probability distribution graph; quantizing the task semantic information and the probability distribution graph into multiple targets and constraints, solving by using a multi-target particle swarm optimization algorithm guided by a fusion probability gradient, and generating a control instruction; quantitatively evaluating the instruction based on a preset rule, and executing the control instruction. The application significantly improves the autonomous decision-making ability, flight safety and task completion reliability of the unmanned aerial vehicle in complex environments under communication disconnection by dynamically perceiving the communication situation, predicting the disconnection risk and fusing the task semantics for multi-target optimization.

Description

Technical Field

[0001] This application relates to the field of autonomous control technology for unmanned aerial vehicles (UAVs), and more specifically, to autonomous control methods and systems for UAVs that are disconnected from their operating systems. Background Technology

[0002] Currently, with the widespread application of drones in reconnaissance, inspection, logistics, and other fields, the risk of unstable or even interrupted communication links in complex and dynamic environments is becoming increasingly prominent. Traditional drone disconnection response methods mainly rely on preset emergency behavior logic, such as "return to home," "hovering," or flying along a fixed path. While these methods are simple and reliable, they have low levels of intelligence and lack the ability to perceive and adapt to the real-time environment. They cannot make optimal decisions based on remaining tasks, their own status, and the dynamically changing environment during communication interruptions, often leading to mission failure or the drone being in danger.

[0003] Current mainstream research focuses on enhancing the autonomy of unmanned aerial vehicles (UAVs), such as using local path planning algorithms for obstacle avoidance or offline reinforcement learning to train strategies for specific scenarios. While these methods have improved the survivability of UAVs after disconnection to some extent, they still have shortcomings: First, they typically lack dynamic and forward-looking modeling of the core constraint of communication link quality, mostly treating communication as a binary state—connected or interrupted—and failing to quantitatively assess and utilize weak signal areas, let alone predict future communication trends. Second, existing autonomous control schemes are often disconnected from higher-level "task semantics," such as prioritizing the reconnaissance of a certain area or the need to sequentially visit certain key points. Their optimization objectives are singular, making it difficult to efficiently advance complex tasks while ensuring basic safety. Finally, when solving complex high-dimensional, multi-objective optimization problems, traditional optimization algorithms are inefficient and struggle to quickly generate optimal or near-optimal solutions that balance multiple conflicting objectives such as communication quality, mission progress, and flight smoothness under limited onboard computing resources.

[0004] In summary, there is an urgent need for a technical approach that can achieve efficient fusion of multi-source information, lightweight prediction of spatiotemporal situation, and real-time multi-objective planning of fusion task semantics under strict resource constraints. Summary of the Invention

[0005] To address the shortcomings of existing technologies, embodiments of this application provide an autonomous control method and system for unmanned aerial vehicle (UAV) disconnection.

[0006] In a first aspect, embodiments of this application provide an autonomous control method for unmanned aerial vehicle (UAV) disconnection, including:

[0007] Acquire drone data, which includes multi-source heterogeneous data and mission semantic information, and perform situational analysis on the multi-source heterogeneous data to obtain a first situational map.

[0008] A spatial probability distribution map is generated by predicting the first situation map and the task semantic information using a preset two-stage prediction method.

[0009] Based on the task semantic information and the spatial probability distribution map, a multi-objective optimization problem is constructed, which includes at least the task coverage target and the communication maintenance target. A first objective function and a second objective function are constructed based on the task semantic information and the spatial probability distribution map, respectively, and flight safety constraints are generated according to the spatial probability distribution map.

[0010] A multi-objective optimization algorithm is used to solve the flight safety constraints to obtain a set of candidate flight trajectories.

[0011] The spatial probability distribution map is evenly divided into grids according to a preset grid interval. Each grid node is associated with the spatial probability value of the corresponding spatial coordinates, and the center coordinates of the grid node are used as potential waypoints.

[0012] Initialize the particle swarm, where the position vector of each particle represents a candidate flight trajectory, which is composed of consecutive potential waypoints;

[0013] Query the spatial probability value of the potential waypoint corresponding to the particle's current position vector, and calculate the negative gradient of the spatial coordinates of the potential waypoint corresponding to the spatial probability value to obtain gradient information;

[0014] Based on the gradient information, the waypoints of the candidate flight trajectories corresponding to each particle are adjusted and non-dominated sorted to obtain the Pareto front solution set. Each solution in the Pareto front solution set corresponds to a candidate flight trajectory. The cost value of each solution in the Pareto front solution set is calculated.

[0015] The optimal trajectory is selected from the set of candidate flight trajectories to generate control commands, and the control commands are quantitatively evaluated.

[0016] Furthermore, the situational analysis of the multi-source heterogeneous data includes:

[0017] Extract a first feature set from the multi-source heterogeneous data. The first feature set includes feature vectors of signal strength, UAV pose, relay device pose, and relay device trajectory.

[0018] Based on the semantic information of the task, the first feature set is filtered by dynamic pruning technology to obtain the second feature set;

[0019] Dynamic clustering is performed on the feature vectors in the second feature set, the clustering results are mapped to situation element nodes, and weighted directed edges are established on the situation element nodes, the situation element nodes have a state history sequence.

[0020] The weights are calculated by fusing feature vector similarity with task semantic relevance.

[0021] The graph structure formed by the situation element nodes and the weighted directed edges is aggregated through a preset graph attention network to generate a first situation map.

[0022] Furthermore, the filtering of the first feature set using dynamic pruning technology includes:

[0023] The task semantic information is encoded into a task semantic vector, and the cosine similarity between each feature vector and the task semantic vector is calculated as the task relevance score of the feature vector.

[0024] The task relevance score of the feature vector is compared with a preset relevance threshold, and feature vectors with task relevance scores lower than the preset relevance threshold are filtered out to obtain a second feature set.

[0025] Furthermore, the step of processing the first situation map and the task semantic information using a preset two-stage prediction method to predict and generate a spatial probability distribution map includes:

[0026] The preset two-stage prediction method includes a first stage and a second stage;

[0027] In the first stage, the spatiotemporal feature vector of the first situation map is extracted by a spatiotemporal graph neural network;

[0028] The second stage uses a latent variable conditional neural process model to decode the spatiotemporal feature vector and the task semantic vector to generate a first probability distribution parameter.

[0029] Based on the first probability distribution parameters, the spatial probability distribution map is generated using a preset spatial interpolation method.

[0030] Furthermore, the extraction of the spatiotemporal feature vector of the first situation map through the spatiotemporal graph neural network includes:

[0031] The spatiotemporal graph neural network calculates the attention weight between any two nodes in the first situational graph through a preset edge weight-temporal difference collaborative attention mechanism, and sums the attention weights of all nodes by weight, outputting the feature representation of the node as the spatiotemporal feature vector.

[0032] The inputs for calculating the attention weights include: the weights of the directed edges connecting any two nodes, and the covariance of the state sequences of any two nodes within a preset sliding time window;

[0033] The spatiotemporal feature vector is mapped to the first latent variable distribution parameter through a preset encoder network, and the task semantic vector is input into the super network as the first modulation parameter of the generation network.

[0034] Latent variable samples are obtained from the first modulation parameters through latent variable distribution sampling, and the latent variable samples and a preset spatial coordinate grid are input into the generation network for forward propagation to predict and output the first probability distribution parameters.

[0035] Furthermore, the generation of flight safety constraints based on the spatial probability distribution map includes:

[0036] The flight safety constraint is obtained by comparing the spatial probability value of each spatial coordinate with a preset safety threshold, and extracting the set of spatial coordinates whose spatial probability value is lower than the preset safety threshold.

[0037] Furthermore, calculating the cost value of each solution in the Pareto front solution set includes:

[0038] Calculate the total path length of each candidate flight trajectory corresponding to the Pareto front solution set, and use it as the first cost factor;

[0039] The number of spatial probability values ​​corresponding to all waypoints traversed by the candidate flight trajectory that are lower than a preset communication threshold is counted and used as a second cost factor;

[0040] The sum of the absolute values ​​of the turning angles between consecutive segments of the candidate flight trajectory is calculated as the third cost factor;

[0041] Based on the preset first weight coefficient, second weight coefficient, and third weight coefficient, the first cost factor, second cost factor, and third cost factor are weighted and summed to obtain the cost value of the candidate flight trajectory.

[0042] Furthermore, the step of selecting the optimal trajectory from the candidate flight trajectory set to generate the control command includes:

[0043] The optimal trajectory is the minimum solution of the Pareto front solution set.

[0044] Furthermore, the quantitative evaluation of the control command includes:

[0045] Verify whether the dynamic parameters required by the control command are within the preset execution threshold range of the UAV. The dynamic parameters include: flight speed, heading angle and acceleration, as the first evaluation index.

[0046] Calculate the overlap area between the estimated flight path corresponding to the control command and the flight safety constraints, and determine whether it exceeds the preset safety tolerance, as the second evaluation indicator;

[0047] The path coverage of the estimated flight path and the corresponding waypoint sequence in the mission semantic information is compared to see if it exceeds a preset coverage threshold, which is used as the third evaluation indicator.

[0048] If the first evaluation metric is not exceeded, the second evaluation metric is not exceeded, and the third evaluation metric is reached or exceeded, the evaluation is deemed successful and the control command is executed.

[0049] Secondly, embodiments of this application also provide an autonomous control system for UAV disconnection, including: an analysis module, a prediction module, a solution module, and an evaluation module;

[0050] Analysis module: used to acquire UAV data, which includes multi-source heterogeneous data and mission semantic information, and to perform situational analysis on the multi-source heterogeneous data to obtain a first situational map;

[0051] Prediction module: used to process the first situation map and the task semantic information using a preset two-stage prediction method to predict and generate a spatial probability distribution map;

[0052] Solution module: Based on the task semantic information and the spatial probability distribution map, construct a multi-objective optimization problem that includes at least the task coverage target and the communication maintenance target. Construct a first objective function and a second objective function based on the task semantic information and the spatial probability distribution map, respectively, and generate flight safety constraints based on the spatial probability distribution map.

[0053] A multi-objective optimization algorithm is used to solve the flight safety constraints to obtain a set of candidate flight trajectories.

[0054] The spatial probability distribution map is evenly divided into grids according to a preset grid interval. Each grid node is associated with the spatial probability value of the corresponding spatial coordinates, and the center coordinates of the grid node are used as potential waypoints.

[0055] Initialize the particle swarm, where the position vector of each particle represents a candidate flight trajectory, which is composed of consecutive potential waypoints;

[0056] Query the spatial probability value of the potential waypoint corresponding to the particle's current position vector, and calculate the negative gradient of the spatial coordinates of the potential waypoint corresponding to the spatial probability value to obtain gradient information;

[0057] Based on the gradient information, the waypoints of the candidate flight trajectories corresponding to each particle are adjusted and non-dominated sorted to obtain the Pareto front solution set. Each solution in the Pareto front solution set corresponds to a candidate flight trajectory. The cost value of each solution in the Pareto front solution set is calculated.

[0058] Evaluation module: used to select the optimal trajectory from the candidate flight trajectory set to generate control commands, and to perform quantitative evaluation of the control commands.

[0059] Compared with existing technologies, this invention calculates feature correlation in real time through task semantic vectors, uses dynamic pruning technology to filter irrelevant features in advance, and then uses a graph attention network with high parameter sharing efficiency to aggregate the simplified feature relationships to generate a first situation map. This significantly reduces computational complexity while ensuring the depth of situation understanding. In the first stage, this invention extracts compact spatiotemporal feature vectors of situation evolution only through a spatiotemporal graph neural network. In the second stage, a lightweight probabilistic generation model, the latent variable conditional neural process model, decodes the probability distribution parameters based on task semantics and finally generates a spatial probability distribution map, reducing the computational overhead of directly modeling complex distributions and enabling UAVs to obtain forward-looking probabilistic environment maps in real time. The planning space is discretized into a grid, and the negative gradient information of the spatial probability distribution map is innovatively introduced into the multi-objective particle swarm optimization algorithm. This probabilistic gradient guidance mechanism enables the optimization algorithm to quickly converge to a safe and efficient solution region. A lightweight evaluation closed loop oriented towards execution is established to ensure the reliability of control commands and system security. Attached Figure Description

[0060] Figure 1 A flowchart of an autonomous control method for UAV disconnection provided in an embodiment of this application;

[0061] Figure 2 A flowchart for generating a spatial probability distribution map of UAV disconnection provided in an embodiment of this application;

[0062] Figure 3 A flowchart for optimizing multi-target control of UAVs in the event of communication disconnection is provided in an embodiment of this application;

[0063] Figure 4 This is a schematic diagram of an autonomous control system for unmanned aerial vehicle (UAV) disconnection provided in an embodiment of this application. Detailed Implementation

[0064] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0065] See Figure 1 This is a flowchart of an autonomous control method for unmanned aerial vehicle (UAV) disconnection provided in an embodiment of this application. The method includes steps S101 to S104, wherein:

[0066] S101: Acquire UAV data, which includes multi-source heterogeneous data and mission semantic information, and perform situational analysis on the multi-source heterogeneous data to obtain a first situational map.

[0067] S102: The first situation map and the task semantic information are predicted by a preset two-stage prediction method to generate a spatial probability distribution map;

[0068] S103: Based on the task semantic information and the spatial probability distribution map, construct a multi-objective optimization problem that includes at least the task coverage target and the communication maintenance target, construct a first objective function and a second objective function based on the task semantic information and the spatial probability distribution map respectively, and generate flight safety constraints according to the spatial probability distribution map;

[0069] A multi-objective optimization algorithm is used to solve the flight safety constraints to obtain a set of candidate flight trajectories.

[0070] The spatial probability distribution map is evenly divided into grids according to a preset grid interval. Each grid node is associated with the spatial probability value of the corresponding spatial coordinates, and the center coordinates of the grid node are used as potential waypoints.

[0071] Initialize the particle swarm, where the position vector of each particle represents a candidate flight trajectory, which is composed of consecutive potential waypoints;

[0072] Query the spatial probability value of the potential waypoint corresponding to the particle's current position vector, and calculate the negative gradient of the spatial coordinates of the potential waypoint corresponding to the spatial probability value to obtain gradient information;

[0073] Based on the gradient information, the waypoints of the candidate flight trajectories corresponding to each particle are adjusted and non-dominated sorted to obtain the Pareto front solution set. Each solution in the Pareto front solution set corresponds to a candidate flight trajectory. The cost value of each solution in the Pareto front solution set is calculated.

[0074] S104: Select the optimal trajectory from the candidate flight trajectory set to generate control commands, and perform quantitative evaluation on the control commands.

[0075] In specific implementation, the present invention is executed by the onboard computing unit of the autonomous mobile device. In this embodiment, taking a drone as an example, the drone needs to cooperate with one or more relay devices.

[0076] The relay device refers to any fixed or mobile platform that can provide key relay services such as communication, navigation enhancement, and mission information relay for the autonomous mobile device, and its specific form may vary depending on the mission scenario.

[0077] For example, in the airspace, the platform and the UAV automatically establish and maintain multi-hop communication links through self-organizing network protocols, sharing each other's real-time state vectors, including position, speed, heading, and perception data, such as wide-area surveillance images carried by airships.

[0078] Specifically, the airship can periodically broadcast its predicted trajectory packets and communication beacons, and the UAV can integrate this information into its own situational awareness map. It can also upload its own compressed sensing data to the airship and entrust it to use the higher bandwidth satellite-to-ground link for long-distance backhaul.

[0079] In water, the buoy array forms a sparse communication and beacon network. Each buoy receives mission instructions via underwater acoustic communication or satellite link and broadcasts its identifier and location via VHF data link.

[0080] While in flight, the drone's communication module continuously scans these beacons. When critical data transmission or link restoration is required, the drone plans a temporary deviation from its flight path based on a probability distribution map, actively flies close to the target buoy, establishes point-to-point directional high-speed transmission, and then returns to its original mission route. Unmanned surface vessels (USVs) can act as mobile gateways, their flight paths forming convoys with drones to provide accompanying relay services.

[0081] In land or fixed areas, mobile communication vehicles, acting as front-end edge computing nodes, not only provide high-power directional antennas to maintain robust links, but also receive raw perception data uploaded by UAVs. Lightweight models run on the vehicle's server for real-time target recognition and fusion processing, and the processed structured intelligence is then sent downlink to the UAVs as a dynamic update of their mission semantic information. Fixed ground stations, serving as mission planning and monitoring centers, synchronously update global mission instructions and no-fly zone information to both UAVs and mobile communication vehicles via security protocols such as VPNs.

[0082] In this embodiment, an airship is selected as a relay device to serve the unmanned aerial vehicle (UAV).

[0083] In practical implementation, airborne sensor data from UAVs and airships are acquired in real time via embedded system buses or communication protocol interfaces, and the raw data stream is read at a fixed frequency. Multiple parallel data processing threads are initiated to construct features. A first feature set is extracted from the multi-source heterogeneous data, which includes feature vectors of signal strength, UAV pose, airship pose, and airship trajectory.

[0084] For example, a fixed-length circular buffer, such as a queue of length 10, is created to sequentially store the sequence of raw sampled signal strength values ​​over the most recent consecutive time period, and the arithmetic mean and variance of all sampled values ​​within the circular buffer are calculated.

[0085] For the same set of sampled values ​​arranged in chronological order within the circular buffer, a first-order linear fit is performed using the least squares method with the time step as the independent variable and the sampled values ​​as the dependent variable, to obtain the slope of the fitted line.

[0086] The calculated average value, variance, and fitted slope are arranged in order to form a three-dimensional signal intensity feature vector.

[0087] As an optional implementation, two independent processing threads are started, including a first thread and a second thread. The first thread takes snapshots of pose data periodically output by the UAV navigation system, including three-dimensional position coordinates, three-dimensional velocity components and three-axis attitude angles, and performs minimum-maximum normalization on each.

[0088] The nine normalized dimensional values ​​are concatenated in a preset order, such as: X coordinate, Y coordinate, Z coordinate, X-direction velocity, Y-direction velocity, Z-direction velocity, roll angle, pitch angle, and yaw angle, to form the UAV pose feature vector.

[0089] For example, similarly, the second thread performs the same normalization and splicing operations on the received pose data snapshots of the airship's periodic output to form the airship's pose feature vector;

[0090] Simultaneously, a derivative calculation thread is initiated to calculate the relative three-dimensional Euclidean distance and relative azimuth angle between the UAV and the airship based on the two sets of normalized position and velocity data mentioned above, and the derivative values ​​are concatenated into a relative pose feature vector.

[0091] As an optional implementation, a fixed-capacity trajectory buffer queue is established to store the airship's historical three-dimensional position coordinate sequence over a recent period in chronological order.

[0092] In practice, the three-point moving average method is used to smooth the three-dimensional coordinate sequence. On the smoothed trajectory, resampling is performed at fixed time intervals, such as 0.2 seconds, to obtain a regular resampling point sequence.

[0093] The direction vectors of the lines connecting adjacent resampled points are calculated sequentially. The angle between adjacent direction vectors and the length of the line are calculated. The ratio of the angle to the corresponding line length is used as the average curvature of the regular resampled point sequence. The arithmetic mean of the curvature of all points is then calculated.

[0094] The average velocity is obtained by calculating the ratio of the displacement from the start point to the end point of the resampled point sequence to the time taken.

[0095] As an optional implementation, the three-dimensional velocity vector and three-dimensional coordinates of the airship at the current moment are obtained, and a prediction time step is set by using a uniform linear motion extrapolation method. The prediction time step is consistent with the scheduling cycle of the data processing thread.

[0096] In a specific implementation, the predicted time step is used as a scalar and multiplied by the X, Y, and Z components of the current three-dimensional velocity vector to obtain the displacement increments along the three coordinate axes during the predicted time period, including the X-axis displacement increment, the Y-axis displacement increment, and the Z-axis displacement increment.

[0097] The X-axis displacement increment is added to the X-coordinate value of the airship's three-dimensional coordinates to obtain the X-coordinate of the predicted position. Similarly, the Y-coordinate and Z-coordinate of the predicted position are obtained respectively, forming the three-dimensional position coordinates of the predicted next moment.

[0098] The calculated average curvature, average velocity, and predicted three-dimensional position coordinates for the next moment are used to construct the airship trajectory feature vector.

[0099] The signal strength, UAV pose, airship pose, relative pose, and airship trajectory feature vector constitute a first feature set. As an optional implementation, the UAV receives a mission text command via a communication link, such as "While ensuring the wireless communication link strength with the relay airship remains above a safety threshold (e.g., -90 dBm), use onboard visible light and infrared sensors to perform a gridded search on a square region centered at coordinates (X, Y) with side length L, to identify and label potential targets," and inputs the mission text command into a pre-trained lightweight model based on the Transformer architecture, all-MiniLM-L6-v2.

[0100] The model contains six encoding layers and outputs a 384-dimensional task semantic vector by performing forward propagation computation.

[0101] In practice, the string is first preprocessed, such as converting all uppercase English letters to lowercase letters, for example, rewriting "dBm" as "dbm".

[0102] Subsequently, the character filter traverses the entire string, removing all symbols that are not Chinese characters, basic English numbers, spaces, or standard punctuation marks, such as commas, parentheses, and hyphens. At the same time, any consecutive spaces or tabs appearing in the text are uniformly replaced with a single standard space character, resulting in a uniformly formatted text string: "While ensuring that the wireless communication link strength with the relay airship remains above the safety threshold of -90dBm, airborne visible light and infrared sensors are used to perform a gridded search to identify and mark potential targets in a square area centered at coordinates (1000, 2000) with a side length of 500 meters."

[0103] The word segmenter based on the lightweight model breaks down sentences into a series of sub-words or word fragments according to its internal dictionary of approximately 30,000 vocabulary units.

[0104] For example, the sentence is segmented into a series of tokens such as "at", "ensure", "with", "relay", "airship", "of", "wireless", "communication", "link", and "strength". Each token is converted into a corresponding numerical code according to its position in the dictionary, ultimately forming a 52-digit sequence.

[0105] The digital sequence is input into six processing layers, where self-attention calculations and feedforward neural network calculations are performed sequentially to conduct nonlinear transformations of the information. The final output is a feature list containing 384 values.

[0106] As an optional implementation, the first feature set is filtered by dynamic pruning technology, and the cosine similarity between each feature vector and the task semantic vector is calculated as the task relevance score of the feature vector.

[0107] In practice, the cosine similarity is obtained by calculating the dot product of the feature vector and the task semantic vector, and then dividing by the product of their respective moduli (L2 norm).

[0108] For example, the first task relevance score of the signal strength feature vector is calculated by quantifying the task semantic information "wireless communication link strength";

[0109] The UAV pose feature vector represents the UAV's own position, velocity, and attitude. It is used to determine whether the UAV is located above the target area centered at coordinates (X,Y) and whether it can stably perform "grid search". The second task relevance score of the UAV pose feature vector is calculated.

[0110] The airship pose feature vector represents the spatial state of the key node in the communication relay, and the relative pose feature vector represents the geometric relationship between the UAV and the airship. The third and fourth task correlation scores of the airship pose feature vector and the relative pose feature vector can be calculated.

[0111] The direct correlation between the airship's own trajectory and the "grid search" is weak, and the fifth task correlation score of the airship trajectory feature vector is calculated.

[0112] The task relevance score of the feature vector is compared with a preset relevance threshold, and feature vectors with task relevance scores lower than the preset relevance threshold are filtered out to obtain a second feature set; wherein the preset relevance threshold can be determined according to the actual engineering accuracy requirements.

[0113] For example, a historical flight mission dataset is constructed, including full feature data of the airship under different weather conditions and missions. As an optional implementation, domain experts annotate the importance of features in each mission;

[0114] For example, in the search task, relative pose features are labeled as key features, and airship trajectory features are labeled as redundant features; and the relevance score of all samples labeled as key features is calculated.

[0115] For example, the mean correlation score of the samples with the key features is distributed in the range of 0.70 to 0.98; the correlation score of the samples with redundant features is distributed in the range of 0.10 to 0.45.

[0116] As an optional implementation, the middle value of the two distribution intervals or a value slightly higher than the upper limit of the redundancy feature is taken as the initial threshold, for example, 0.60 is taken as the initial preset correlation threshold.

[0117] In practical implementation, when flight missions require the UAV to maintain a highly stable connection with the airship, such as when conducting dense grid searches in complex terrain, it is necessary to retain as many features as possible that may affect link judgment to prevent missed judgments. In this case, it is necessary to lower the threshold; for example, the threshold can be lowered from 0.60 to 0.40.

[0118] In practice, when the drone's battery is low and computing resources are limited, it is necessary to increase the threshold to retain only core information; for example, the threshold can be increased from 0.60 to 0.80.

[0119] Dynamic clustering is performed on the feature vectors in the second feature set, the clustering results are mapped to situation element nodes, and weighted directed edges are established on the situation element nodes, the situation element nodes having a state history sequence.

[0120] As an optional implementation, the feature vectors in the second feature set are dynamically grouped using the online K-Means clustering algorithm;

[0121] For example, in the initialization phase, initial cluster centers are set based on the first batch of feature vectors, and clusters are created based on the cluster centers. The cluster is a set of all feature vectors belonging to the same cluster center.

[0122] For a real-time input feature vector, calculate the Euclidean distance from the feature vector to the current cluster center, and find the cluster center with the smallest distance;

[0123] If the minimum distance is less than a preset fixed distance threshold, the feature vector is assigned to the cluster, and the cluster center coordinates of the cluster are calculated using the feature vector.

[0124] If the minimum distance is greater than or equal to the fixed distance threshold, a new cluster is created based on the feature vector as the starting point. The fixed distance threshold can be obtained by calculating the Euclidean distances between all pairwise historical feature vectors and taking a percentile of the Euclidean distance distribution, such as the 30th percentile.

[0125] In specific implementation, if the cluster does not absorb new feature vectors within a consecutive preset number of processing cycles, such as five consecutive cycles, the cluster is determined to be invalid and deleted; the specific number of the consecutive preset number of cycles can be set according to the requirements of the UAV system for the response speed to situational changes.

[0126] For a cluster that continuously absorbs new feature vectors, if the Euclidean distance between the old and new centers is less than a preset small displacement threshold during each update of the cluster center coordinates in a consecutive preset number of processing cycles, then the cluster is considered to have reached a stable state, and the current center coordinate vector is defined as a situation element node; the small displacement threshold can be determined based on the sensor noise level.

[0127] Each of the aforementioned situation element nodes is associated with a fixed-length first-in-first-out (FIFO) buffer, which stores all feature vectors that have recently been assigned to that cluster in chronological order, forming the state history sequence of the situation element node; wherein, the fixed length is the upper limit of the number of feature vectors that the buffer can store, and is set according to the design requirements of the UAV system.

[0128] In this embodiment, the second feature set is dynamically grouped using an online K-Means clustering algorithm.

[0129] During initialization, standard K-Means clustering is performed using the first 10 sets of data, with a cluster size of 3, generating 3 initial cluster centers and corresponding clusters. For each new input feature vector, its Euclidean distance to each existing cluster center is calculated.

[0130] For example, a fixed distance threshold of 4.5 is set. If the minimum distance of a newly input feature vector is less than the fixed distance threshold, the vector is assigned to the corresponding cluster, and the cluster center is updated; if the minimum distance is greater than or equal to the fixed distance threshold, a new cluster centered on the vector is created.

[0131] Cluster lifecycle management assumes a processing cycle of 100 milliseconds. If a cluster does not absorb new vectors within 5 consecutive processing cycles, it is deemed invalid and deleted.

[0132] If the Euclidean distance between the old and new centers is less than the micro-displacement threshold of 0.1 in each of the eight consecutive updates of the center coordinates of an active cluster, then the cluster is considered stable, and its center vector is defined as a situation element node.

[0133] Each situational feature node is associated with a FIFO buffer of length 50, which is used to store the 50 most recently absorbed feature vectors in the cluster in sequence, forming the node's state history sequence.

[0134] In practice, the weights are calculated by fusing feature vector similarity with task semantic relevance.

[0135] As an optional implementation, two directed edges with opposite directions are created for any two different situation element nodes. The cosine similarity between the current feature vectors (i.e., cluster center vectors) of the two situation element nodes connected by the directed edges is calculated, and the task relevance scores corresponding to the two feature vectors are obtained. The two scores are then multiplied together.

[0136] The product of the calculated cosine similarity and the task relevance score is multiplied and then summed according to a preset weight ratio to obtain an intermediate value. The preset weight ratio is determined according to the actual engineering requirements.

[0137] As an optional implementation, collect all node pairs whose edge weights are to be calculated at the current time. Each node pair includes: the product of cosine similarity and task relevance. Normalize the results by dividing the cosine similarity of each node pair by the sum of the cosine similarities of all node pairs to obtain the proportion of the node pair on the index. Similarly, perform the same processing on the task relevance product index.

[0138] Calculate the information entropy based on the proportion of all node pairs under each indicator. Multiply each proportion by the natural logarithm of that proportion, sum the results, divide by the natural logarithm of the number of node pairs, and take the negative value to obtain the information entropy.

[0139] Subtract the information entropy of the indicator from 1 to obtain the difference coefficient. Add the difference coefficients of the two indicators together, and divide the difference coefficient of each indicator by the sum to obtain the weight of the indicator. The weight of cosine similarity is the proportion of the difference coefficient, and the weight of the task relevance product is calculated similarly.

[0140] The intermediate values ​​are input into the hyperbolic tangent function for nonlinear transformation and scaling, and the final output value is used as the weight of the directed edge, with the weight ranging from [0, 1].

[0141] In practice, it is assumed that there are five node pairs that need to be calculated at the current time, with cosine similarities of 0.75, 0.80, 0.60, 0.70, and 0.65, and task association products of 0.5576, 0.60, 0.50, 0.55, and 0.58, respectively.

[0142] Summing the five values ​​gives 3.5. Dividing each value by the sum yields the percentage sequence. Taking the natural logarithm of each percentage, multiplying it by the percentage, summing the results, taking the negative value, and then dividing by the natural logarithm of the number of node pairs (1.609) gives an information entropy of approximately 0.9966. Similarly, summing the products of task relevance gives 2.7876, and calculating the information entropy gives approximately 0.9980.

[0143] The cosine similarity coefficient is 1 - 0.9966 = 0.0034, and the product similarity coefficient is 1 - 0.9980 = 0.0020. Adding them together gives 0.0054. The cosine similarity weight is 0.0034 divided by 0.0054, approximately 0.63, and the product weight is 0.0020 divided by 0.0054, approximately 0.37. For ease of engineering application, these are rounded to 0.6 and 0.4 respectively.

[0144] In a specific embodiment, the cosine similarity of the current feature vectors of node A (derived from the UAV pose cluster) and node B (derived from the signal strength cluster) is 0.75; their task relevance scores are 0.68 and 0.82 respectively, and their product is 0.5576; the weighted sum gives an intermediate value of 0.67304; after hyperbolic tangent function transformation, the final edge weight is 0.72.

[0145] In practice, the graph structure formed by the situation element nodes and the weighted directed edges is aggregated through a preset graph attention network to generate a first situation map.

[0146] As an optional implementation, the graph attention network adopts a two-layer multi-head graph attention network model with GATv2 architecture. The first layer of the model is configured with 4 independent attention heads, and the second layer is configured with 6 independent attention heads.

[0147] In a specific implementation, the first layer projects the feature vectors of the input situation element nodes into a 64-dimensional space through four independent linear transformation matrices; the second layer projects the 256-dimensional feature vectors output by the first layer into a 64-dimensional space through six independent linear transformation matrices.

[0148] For example, the model can use ReLU as the inter-layer activation function, and the LeakyReLU activation function can be used when calculating the attention coefficients.

[0149] In its implementation, the pre-training steps of the two-layer multi-head graph attention network model include: constructing a historical graph structure sample set based on dynamic clustering and edge weights, and generating node-level target feature vectors as supervision labels for each situation element node in the sample graph. Using the historical graph structure as input and the node target features as supervision signals, a stochastic gradient descent optimizer is employed to minimize the mean square error between the predicted features and the target features. The network parameters are iteratively updated until convergence, and the trained model parameters are then solidified and deployed.

[0150] In the specific implementation, the graph structure constructed in real time is input into the two-layer multi-head graph attention network model. For each situation element node, in each attention head of the first layer, the features of the node and all first-order neighbor nodes are linearly projected. The normalized attention coefficient of each neighbor node is calculated by combining the weight of the directed edge connecting the node and each neighbor node, and the projected features of the neighbor nodes are weighted and summed.

[0151] The outputs of the four attention heads are concatenated along the feature dimension to form the first feature vector output by the situation element node in the first layer.

[0152] The second layer processes the features output by the first layer using the same structure, and performs average pooling on the outputs of the six attention heads to obtain the second feature vector of each situation element node. The second feature vectors of all nodes are then combined to form the first situation map.

[0153] See Figure 2 The flowchart for generating a spatial probability distribution map of UAV disconnection provided in this application embodiment includes:

[0154] S201: Extract features from the first situation map using a spatiotemporal graph neural network to generate a spatiotemporal feature vector;

[0155] S202: The encoder network is used to map the spatiotemporal feature vectors into probability distribution parameters and random sampling is performed to obtain latent variable samples;

[0156] S203: Based on the task semantic information, modulate and generate network parameters, and combine the latent variable samples to predict and calculate the spatial coordinate grid, generating the state probability distribution parameters of each location point;

[0157] S204: Perform spatial interpolation on the probability distribution parameters of all location points to generate a spatial probability distribution map.

[0158] The first situation map and the task semantic information are processed by a preset two-stage prediction method to predict and generate a spatial probability distribution map. The preset two-stage prediction method includes a first stage and a second stage.

[0159] In the first stage, the spatiotemporal feature vector of the first situation map is extracted by a spatiotemporal graph neural network.

[0160] Regarding step S201:

[0161] As an optional implementation, the spatiotemporal graph neural network adopts a network architecture based on graph attention network and gated recurrent unit, including a spatial graph attention layer and a temporal gated recurrent unit layer.

[0162] In a specific implementation, the spatial graph attention layer is configured with four independent attention heads. The spatial graph attention layer projects the feature vectors of the input situation element nodes into a 64-dimensional space through four independent linear transformation matrices to obtain projected features.

[0163] As an optional implementation, the spatial graph attention layer calculates the attention weight between any two situation element nodes in the first situation graph through a preset edge weight-temporal difference collaborative attention mechanism, and sums the attention weights of all situation element nodes by weight, and outputs the feature representation of the situation element node as the spatiotemporal feature vector.

[0164] The inputs for calculating the attention weights include: the weights of the directed edges connecting any two situational element nodes, and the covariance of the state sequences of any two situational element nodes within a preset sliding time window.

[0165] The length of the preset sliding time window is determined based on the data sampling frequency of the UAV system's sensors and the minimum period of the situational change to be analyzed.

[0166] Based on the attention weights, the spatial graph attention layer performs a weighted summation of the projected features of all neighboring nodes of each situation element node, and outputs the spatial features of each situation element node, which are 64-dimensional feature vectors.

[0167] In specific implementation, the time-gated loop unit layer adopts a bidirectional structure, including a forward layer and a reverse layer, with 128 hidden units set in each direction.

[0168] For example, the time-gated cyclic unit layer inputs the spatial features and corresponding state history sequence of each situation element node in the order of the sequence time steps, and the tanh function is used as the activation function inside the time-gated cyclic unit layer; wherein, the time step is directly determined by the length of the state history sequence of the situation element node.

[0169] In practice, the forward layer processes the state history sequence from front to back according to time steps and outputs the forward hidden state; the reverse layer processes the state history sequence from back to front according to time steps and outputs the reverse hidden state.

[0170] The forward hidden state and the reverse hidden state of each situation element node at the final time step are concatenated to obtain a 256-dimensional vector, which serves as the spatiotemporal fusion feature of that node.

[0171] A global average pooling operation is performed on the 256-dimensional spatiotemporal fusion features of all the aforementioned situation element nodes, and a 512-dimensional vector is finally output as the spatiotemporal feature vector.

[0172] Regarding step S202:

[0173] The second stage uses a latent variable conditional neural process model to decode the spatiotemporal feature vector and the task semantic vector to generate a first probability distribution parameter.

[0174] Based on the first probability distribution parameters, the spatial probability distribution map is generated using a preset spatial interpolation method.

[0175] As an optional implementation, the latent variable conditional neural process model includes an encoder network, a supernetwork, and a generator network; the encoder network includes three layers, the first layer receives a 512-dimensional spatiotemporal feature vector and has 512 neurons; the second layer has 256 neurons; and the third layer outputs a 128-dimensional vector, with the ReLU activation function used between layers.

[0176] The output (128-dimensional vector) of the encoder network is the parameter of the latent variable probability distribution. The first 64 values ​​are used as the mean vector of the multivariate Gaussian distribution, and the last 64 values ​​are used as the logarithm of the variance vector of the latent variable probability distribution, thus forming the first latent variable distribution parameter.

[0177] The supernetwork consists of two layers. The first layer receives a 384-dimensional task semantic vector and has 256 neurons. The dimension of the output of the second layer is determined according to the number of weights that need to be dynamically generated in the generator network. The output vector of the supernetwork is used as the first modulation parameter.

[0178] The generative network consists of five layers. The weight matrices of the first three fully connected layers are generated in real time by the first modulation parameters through a linear mapping subnetwork. The input of the generative network is the concatenation of spatial coordinates and latent variable samples, and the output layer has two neurons.

[0179] In a specific implementation, a 512-dimensional spatiotemporal feature vector is input into the encoder network. Through forward propagation calculation, a 128-dimensional vector is output. The first 64 values ​​of the vector are used as the mean vector, and the last 64 values ​​are used as the logarithmic variance vector. An exponential operation is performed on each value in the logarithmic variance vector to obtain the variance vector.

[0180] As an optional implementation, the standard class constructor for creating a multivariate Gaussian distribution in the math library, such as torch.distributions in PyTorch, is invoked, and the mean vector and the variance vector are passed as parameters to the constructor, which returns a 64-dimensional multivariate Gaussian distribution object.

[0181] Based on the Gaussian distribution object, a random sampling operation is performed to obtain a 64-dimensional latent variable sample.

[0182] Regarding step S203:

[0183] In a specific implementation, a 384-dimensional task semantic vector is input into the super network, and through forward propagation calculation, a vector of a specific dimension, namely the first modulation parameter, is output, which is used to dynamically set the weights of the first three layers of the generation network.

[0184] The specific dimension of the first modulation parameter is determined by the total number of parameters required for the weight matrix of the first three fully connected layers of the generator network. For example, if the structure of the first three layers of the generator network requires a certain number of weight parameters, the output dimension of the supernetwork is set to be equal to that number to ensure that its output can be reshaped into the corresponding weight matrix.

[0185] The latent variable samples are concatenated with each coordinate point in a preset spatial coordinate grid to form batch data.

[0186] For example, the geographical boundary description of the task area can be parsed from the task semantic information. For instance, a square area with coordinates (X,Y) as the center and side length L can be extracted from the instruction text to generate a set of two-dimensional point coordinates with equal spacing that completely covers the square area, i.e., the spatial coordinate grid.

[0187] The batch data is input into the generator network, and forward propagation calculation is performed. For each coordinate point, two scalars are output: the first scalar is the mean of the predicted distribution, and the second scalar is the logarithm of the standard deviation of the predicted distribution.

[0188] The mean and the logarithm constitute the first probability distribution parameters of the coordinate point.

[0189] Regarding step S204:

[0190] Based on the first probability distribution parameters of all discrete coordinate points, the bicubic interpolation algorithm is called to calculate the parameters of each point in the continuous space, and finally the spatial probability distribution map is generated.

[0191] In the specific implementation, the mean and logarithmic standard deviation of each coordinate point are used as two independent scalar fields on the discrete grid points, and a bicubic interpolation algorithm is used to perform surface fitting and densification calculations on the two scalar fields respectively.

[0192] For example, the bicubic interpolation algorithm calculates the interpolation result for any point in the regular grid based on the parameter values ​​in the 4×4 neighborhood around each discrete point.

[0193] By performing this calculation on all continuous locations within the entire domain, two continuous parameter fields are finally obtained: a continuous mean field and a continuous log-standard deviation field.

[0194] Combining these two continuous parameters at any point in space creates a complete and continuous probability distribution description of the location, thereby generating the spatial probability distribution map.

[0195] The task semantic information and the spatial probability distribution map are quantified into optimization objectives and constraints, which are then solved using a preset multi-objective optimization algorithm to generate control commands.

[0196] See Figure 3 The flowchart provided in this application embodiment for optimizing multi-target control of UAVs in the event of communication disconnection includes:

[0197] A first objective function and a second objective function are constructed based on the task semantic information and the spatial probability distribution map, respectively, and flight safety constraints are generated based on the spatial probability values ​​of the spatial probability distribution map.

[0198] As an optional implementation, the first objective function is a task coverage objective function, which extracts key areas or waypoint sequences that need to be covered by the UAV flight trajectory by parsing the task semantic information.

[0199] The coverage of a candidate flight path over the key area or waypoint sequence is quantified into a calculable value.

[0200] For example, if the task semantic information is "search area A1", then the first objective function can be calculated as the ratio of the number of grid points in area A1 through which the candidate trajectory passes to the total number of grid points in area A1, with the objective being to maximize the ratio.

[0201] The second objective function is the communication maintenance objective function. Based on the spatial probability distribution map, the spatial probability values ​​corresponding to all the location points passed by a candidate flight trajectory are aggregated, and its comprehensive evaluation value is calculated.

[0202] As an optional implementation, the second objective function can be calculated as the arithmetic mean of the probability values ​​corresponding to all waypoints traversed by the candidate flight trajectory, with the objective being to maximize the arithmetic mean.

[0203] The flight safety constraint is obtained by comparing the spatial probability value of each spatial coordinate with a preset safety threshold, and extracting the set of spatial coordinates whose spatial probability value is lower than the preset safety threshold.

[0204] As an optional implementation, the preset security threshold is determined based on the lowest acceptable probability of success for the communication link maintenance task.

[0205] For example, the success probability requirement for end-to-end data packet transmission can be determined based on mission requirements. For instance, if the mission specifies that the success rate of data packet transmission between the drone and the airship must not be less than 99%, then the lower limit of the end-to-end success probability is 99%.

[0206] In practical implementation, assuming the communication protocol supports a maximum of three automatic retransmission requests, then after a transmission failure, two retransmissions are allowed, meaning a maximum of three attempts in total. Under this mechanism, the conversion relationship between the probability of a single successful transmission and the end-to-end success probability is: the end-to-end success probability equals one minus the cube of the probability of a single transmission failure;

[0207] For example, substituting 99% into the equation 0.99 equals one minus the cube of the probability of a single transmission failure, the cube of the probability of a single transmission failure is 0.01, and the cube root of the probability of a single transmission success is approximately 0.215, or about 78.5%.

[0208] A minimum acceptable success rate of 78.5% is set as the preset security threshold for a single transmission.

[0209] In specific implementation, each spatial coordinate point in the spatial probability distribution map is traversed, and its corresponding spatial probability value is compared with the safety threshold. All coordinate points with spatial probability values ​​lower than the safety threshold are extracted to form a set of no-fly zones. The flight safety constraint requires that the generated candidate flight trajectory must not cross this set of no-fly zones.

[0210] The spatial probability distribution map is gridded, and the Pareto front solution set is obtained by solving a pre-set multi-objective particle swarm optimization algorithm based on probability gradient guidance; wherein, each solution in the Pareto front solution set corresponds to a candidate flight trajectory.

[0211] The spatial probability distribution map is evenly divided into grids according to a preset grid interval. Each grid node is associated with the spatial probability value of the corresponding spatial coordinates, and the center coordinates of the grid node are used as potential waypoints.

[0212] In practice, a fixed grid interval parameter is preset, and based on the grid interval parameter, the horizontal area covered by the spatial probability distribution map is divided at equal intervals along the east-west (X-axis) and north-south (Y-axis) directions to generate a two-dimensional grid.

[0213] Each cell is a grid unit, and the center point of the grid unit is a grid node. The center coordinates of all grid nodes form a discrete set of potential waypoints, where each grid node is associated with two specific data points:

[0214] First, the two-dimensional spatial coordinates of the location of the grid node;

[0215] Second, the spatial probability values ​​at the coordinate points are obtained from the spatial probability distribution map using the bilinear interpolation method.

[0216] Initialize the particle swarm, where the position vector of each particle represents a candidate flight trajectory, which is composed of consecutive potential waypoints.

[0217] In specific implementation, the parameters of the multi-objective particle swarm optimization algorithm based on probability gradient guidance include: particle swarm size, maximum number of iterations, inertia weight coefficient, individual learning factor, social learning factor, and key gradient guidance weight coefficient.

[0218] For example, the parameter settings of the algorithm are determined through offline simulation debugging based on typical scenarios of UAV mission planning. The particle swarm size is usually set between 50 and 100; the maximum number of iterations is set between 100 and 200 depending on the problem complexity; the inertia weight coefficient, individual learning factor, and social learning factor are set using empirical value ranges from the field of particle swarm optimization.

[0219] The key gradient-guided weight coefficients are specific parameters of this algorithm, used to adjust the influence of probabilistic gradient information on particle velocity updates. The algorithm's convergence and solution set distribution settings are observed through simulations of multiple typical task scenarios.

[0220] In the specific implementation, the parameters of the algorithm are first initialized, and the position vector of each particle is instantiated as an integer index sequence. The length of the integer index sequence is fixed, equal to the preset number of waypoints plus two, including the start point and the end point. The preset number of waypoints is determined based on the straight-line distance from the start point to the end point of the mission.

[0221] For example, based on historical mission data, a waypoint can be planned at an average interval of a specific distance, such as 20 meters.

[0222] By using a random number generator to select from the index range of the potential waypoint set (excluding the start and end indexes), each index value in the sequence is obtained, thereby generating an initial candidate flight trajectory for each particle, forming a population containing multiple possible solutions, i.e., trajectories.

[0223] The spatial probability value of the potential waypoint corresponding to the current position vector of the particle is queried, and the negative gradient of the spatial coordinates of the potential waypoint corresponding to the spatial probability value is calculated to obtain the gradient information.

[0224] For example, the potential waypoints are located in a two-dimensional grid based on their spatial coordinates;

[0225] Read the spatial probability values ​​recorded at the four adjacent grid nodes in the east, west, south and north of the grid node to obtain the spatial probability values ​​of the east side, west side, south side and north side;

[0226] Subtract the probability value on the west side from the probability value on the east side, and then divide by twice the grid spacing distance to calculate the rate of change of the spatial probability value in the east-west direction.

[0227] Subtract the probability value on the north side from the probability value on the south side, and then divide by twice the grid spacing distance to calculate the rate of change of the spatial probability value in the north-south direction.

[0228] Using the two rates of change mentioned above as components, a two-dimensional gradient vector is formed. The two-dimensional gradient vector points in the direction of the fastest increase in probability. By inverting the gradient vector, the probability negative gradient vector of the potential waypoint is obtained.

[0229] Based on the gradient information, the waypoints of the candidate flight trajectory corresponding to each particle are adjusted.

[0230] In a specific implementation, the algorithm iterates through each particle in the population to perform an update operation. The new velocity vector of the particle is obtained through vector synthesis, including inertial components, cognitive components, social components, and gradient-guided components.

[0231] The inertial component is the current velocity multiplied by the inertial weight coefficient; the cognitive component is the difference between the individual's historical best position and the current position multiplied by the individual's learning factor and a random number uniformly distributed between zero and one; the social component is the difference between the global historical best position and the current position multiplied by the social learning factor and a random number uniformly distributed between zero and one.

[0232] The gradient guidance component averages the probability negative gradient vectors of all waypoints on the particle's current trajectory, and then multiplies this average vector by the gradient guidance weight coefficient; the probability negative gradient vector of a single waypoint is obtained by inverting the partial derivative of the spatial probability value at the waypoint with respect to its coordinates calculated by the central difference method.

[0233] As an optional implementation, the inertia weight coefficient can be calculated using a linear decreasing strategy, obtained by linear interpolation between the initial value and the final value based on the current iteration number.

[0234] The gradient guidance weight coefficient can be dynamically determined based on the consistency of the direction of the probability negative gradient vector of all waypoints on the current trajectory of the particle. The consistency is evaluated by calculating the angle between the negative gradient vector of each waypoint and their average vector. The smaller the angle, the more consistent the direction, and the larger the value of the gradient guidance weight coefficient.

[0235] For example, suppose in a three-dimensional search space, the algorithm has a total of 100 iterations, and it is currently on the 50th iteration, and needs to update the 3rd particle in the population.

[0236] With an initial inertia weight of 0.9, a final inertia weight of 0.4, and a total of 100 iterations, the weight needs to decrease by 0.5 from the first to the 100th iteration, which means a decrease of 0.005 in each iteration.

[0237] The current iteration is the 50th iteration. From the 1st iteration to the present, there have been 49 reductions. The inertia weight in the 50th iteration is equal to 0.9 minus 49 multiplied by 0.005, which is 0.655. Therefore, the inertia weight coefficient of this particle in this iteration is 0.655.

[0238] For example, suppose the current position has an X coordinate of 10, a Y coordinate of 20, and a Z coordinate of 15; and the current velocity has an X component of 2, a Y component of -1, and a Z component of 1.

[0239] The X-coordinate of the individual's historical best position is 12, the Y-coordinate is 18, and the Z-coordinate is 16; the X-coordinate of the global historical best position is 5, the Y-coordinate is 25, and the Z-coordinate is 10; the individual learning factor is 1.5, and the social learning factor is 1.5. In this iteration, the random number used for the cognitive component is 0.6, and the random number used for the social component is 0.8.

[0240] The particle's trajectory records the five most recently visited waypoints. The negative gradient vector of probability at each waypoint is calculated using the central difference method. In the negative gradient vector of waypoint 1, the X direction is 0.1, the Y direction is 0.2, and the Z direction is 0.1; in waypoint 2, the X direction is 0.15, the Y direction is 0.18, and the Z direction is 0.12; in waypoint 3, the X direction is 0.08, the Y direction is 0.22, and the Z direction is 0.09; in waypoint 4, the X direction is 0.12, the Y direction is 0.15, and the Z direction is 0.15; and in waypoint 5, the X direction is 0.1, the Y direction is 0.25, and the Z direction is 0.05.

[0241] In the specific implementation, the average vector of these five negative gradient vectors is calculated, resulting in an average gradient vector of 0.11 in the X direction, 0.20 in the Y direction, and 0.102 in the Z direction. The consistency between the directions of the five negative gradient vectors and the average vector is evaluated by calculating the cosine of the angle between each vector and the average vector, yielding five cosine values ​​of 0.98, 0.99, 0.97, 0.96, and 0.95. The average of these five cosine values ​​yields a consistency index of 0.97, which is used as the gradient guidance weight coefficient in this iteration.

[0242] As an alternative implementation, a perturbation operation is performed on the waypoint index in the particle position vector based on the new velocity vector.

[0243] In specific implementation, the spatial coordinates of waypoints in the candidate flight trajectory are determined, and a grid neighborhood is defined with the grid cell where the spatial coordinates are located as the center for searching. For example, all potential waypoints within a grid interval in the east, west, south, and north of the grid cell are searched as a set of physically adjacent candidate points.

[0244] Calculate the vector pointing from the current waypoint coordinates to the coordinates of each candidate point in the set, and calculate the dot product between each vector and the probability negative gradient vector of the current waypoint.

[0245] For example, candidate points whose dot product is greater than zero are selected as candidate points located in the gradient guiding direction.

[0246] In practice, the magnitude of the new velocity vector is normalized to obtain a value between 0 and 1 as the replacement probability.

[0247] For the current waypoint, generate a random number between 0 and 1; if the random number is less than the replacement probability, then trigger the replacement operation.

[0248] As an optional implementation, based on the triggered replacement operation, a uniform random sampling method is used to randomly select one of the candidate points as a replacement candidate point, and the index of the replacement candidate point in the potential waypoint set is used to directly replace the original index value representing the current waypoint in the particle position vector.

[0249] By iterating through all intermediate waypoints and performing judgments and operations, a trajectory adjustment is completed, resulting in the first flight trajectory.

[0250] In practice, after the trajectory adjustment is completed, the flight safety constraints of the adjusted first flight trajectory are verified.

[0251] As an optional implementation, the spatial coordinates of each waypoint of the first flight trajectory are traversed to query the set of no-fly zones;

[0252] For example, by performing a hash table lookup operation, it is determined whether the spatial coordinates exist as key values ​​in the no-fly zone set; if any waypoint coordinate is found in the set, the first flight trajectory is determined to violate the constraints and is marked as invalid; if none of the coordinates are found in the set, the first flight trajectory is determined to satisfy the constraints and is marked as valid, and invalid particles and valid particles are obtained respectively.

[0253] For all first flight trajectories marked as valid, calculate two objective function values.

[0254] As an optional implementation, the gridded mission area data is accessed, the first flight trajectory is discretized into line segments, and a scanline algorithm is used to calculate all the grid cells that these line segments pass through.

[0255] Perform an intersection operation between the set of grid cells traversed and the set of grid cells in the task area, and count the number of grid cells within the intersection; divide the number of grid cells by the total number of grid cells in the task area to obtain the task coverage ratio.

[0256] Read the spatial probability value associated with the coordinates of each waypoint on the first flight trajectory and stored in the corresponding grid node data structure; sum the spatial probability values ​​and divide by the total number of waypoints to obtain the arithmetic mean.

[0257] The effective particles are sorted non-dominated to obtain the Pareto front solution set.

[0258] For example, the task coverage ratio of each effective particle is combined with the arithmetic mean to form a two-dimensional vector, and all particles are compared pairwise.

[0259] For example, for particles C and D, if the coverage ratio of C is greater than or equal to the coverage ratio of D, and the average probability value of C is greater than or equal to the average probability value of D, and at least one of these two comparisons is strictly greater than the other, then it is determined that C dominates D.

[0260] In practice, the multi-objective particle swarm optimization algorithm guided by probability gradients counts the number of other particles dominated by each effective particle and the number of particles that dominate it.

[0261] Based on the statistical results, all effective particles that dominate other effective particles with zero number of particles are classified into the first non-dominated layer. The set of all particles in the first non-dominated layer is the Pareto front solution set of the current iteration.

[0262] Calculate the substitution value of each solution in the Pareto front solution set, and select the solution set with the smallest substitution value for decoding, which serves as the control command for the UAV.

[0263] As an optional implementation, the total path length of each candidate flight trajectory corresponding to the Pareto front solution set is calculated as a first cost factor;

[0264] For example, the waypoint coordinate sequence of the candidate flight trajectory is read sequentially, and for each pair of adjacent waypoints in the sequence, their Euclidean distance is calculated;

[0265] The total distance of all consecutive segments of the candidate flight path is summed to obtain the total path length of the path.

[0266] The number of spatial probability values ​​corresponding to all waypoints traversed by the candidate flight trajectory that are lower than a preset communication threshold is counted and used as a second cost factor;

[0267] For example, for each waypoint, a corresponding spatial probability value is obtained based on its spatial coordinates. This spatial probability value is compared with a preset communication quality threshold. If the probability value of that point is lower than the threshold, a counter is incremented. After the traversal is complete, the value of the counter is the number of waypoints in the low communication quality region that the trajectory has passed through, which serves as the second cost factor.

[0268] The communication quality threshold is determined based on the accuracy requirements of the project.

[0269] For example, if the mission specifies that the bit error rate of the images transmitted by the UAV must not exceed 10 to the power of negative five, then this shall be used as an indicator of communication quality.

[0270] In practical implementation, QPSK modulation can be used, and based on the theoretical bit error rate curve of this method, the received signal-to-noise ratio required to achieve a bit error rate of 10 to the power of negative five is approximately 9.5 dB.

[0271] The received signal-to-noise ratio (SNR) is the minimum threshold value required for the receiver to correctly demodulate the signal, and it serves as the physical layer indicator corresponding to the communication quality threshold. QPSK, or Quadrature Phase Shift Keying, is a digital modulation method that carries information by changing the phase of the carrier wave.

[0272] The sum of the absolute values ​​of the turning angles between consecutive segments of the candidate flight trajectory is calculated as the third cost factor;

[0273] For example, the turning angle between flight segments is calculated by sequentially processing three consecutive waypoints on the candidate flight trajectory.

[0274] For example, for consecutive points P1, P2, P3, vector P1P2 represents the direction of the previous flight segment, and vector P2P3 represents the direction of the current flight segment. Calculate the angle between these two direction vectors and take the absolute value of the angle to ensure that the cost is included regardless of whether it is a left turn or a right turn.

[0275] Perform the above operation on all three consecutive points on the trajectory, and sum up the absolute values ​​of all calculated steering angles. The sum is the cumulative steering angle.

[0276] Based on the preset first weight coefficient, second weight coefficient, and third weight coefficient, the first cost factor, second cost factor, and third cost factor are weighted and summed to obtain the cost value of the candidate flight trajectory.

[0277] As an optional implementation, the weighting coefficients are determined by a method based on objective function sensitivity analysis.

[0278] In practical implementation, the ultimate goal of the flight mission is represented as a function of three cost factors. Taking a gridded search mission as an example, the mission objective can be set as minimizing the search completion time. Through theoretical analysis or offline experimental data fitting, an approximate relationship between minimizing the search completion time and path length, the number of waypoints in low-communication areas, and the cumulative turning angle is obtained.

[0279] For example, the minimum search completion time is 0.02 path lengths, 0.5 low-communication area waypoints, and 0.01 cumulative turning angles. The coefficients 0.02, 0.5, and 0.01 reflect the unit impact of each cost factor on the task objective. That is, for every unit increase in path length, the task completion time increases by 0.02 units; for every unit increase in low-communication area waypoints, the task completion time increases by 0.5 units; and for every unit increase in cumulative turning angle, the task completion time increases by 0.01 units.

[0280] Extract the coefficients corresponding to the three cost factors as initial weights, and add them together to obtain the sum.

[0281] Divide each coefficient by the sum to obtain the normalized weight coefficients. The first weight coefficient is approximately 0.038, the second weight coefficient is approximately 0.943, and the third weight coefficient is approximately 0.019. The sum of these three coefficients is 1.

[0282] In specific implementation, the task priority is determined based on the task type in the current task semantic information and a preset weight decision logic, wherein the weight decision logic includes:

[0283] Emergency return-to-home mission: When semantic information parsing reveals keywords such as "signal loss" and "emergency return-to-home," the highest priority is determined to be a safe return as quickly as possible. Therefore, flight safety becomes the primary objective, followed by path length, while flight smoothness has the lowest priority.

[0284] Regional patrol missions: When semantic information is parsed to reveal keywords such as "continuous monitoring" and "regional coverage," the mission requirements are determined to be stable operation and complete coverage. Therefore, communication assurance becomes the primary objective, followed by flight smoothness, with path length having the lowest priority. Flight smoothness represents energy saving and stable observation.

[0285] The control commands are quantitatively evaluated based on preset evaluation rules. The dynamic parameters required by the control commands are verified to be within the preset execution threshold range of the UAV. These dynamic parameters include flight speed, heading angle, and acceleration, which serve as the first evaluation indicators.

[0286] In practice, the hardware parameter table in the UAV system memory is accessed. The hardware parameter table includes the upper and lower limits of flight speed, the upper limit of the rate of change of heading angle, and the upper limits of normal and tangential acceleration.

[0287] As an optional implementation, the list of waypoints is obtained, and for every two consecutive waypoints, the required average speed for this segment is calculated based on the straight-line distance between the two points and the expected time difference.

[0288] For every three consecutive waypoints, calculate the required heading angle change rate at the turning point based on the vector direction changes and time between the two segments.

[0289] Calculate the required acceleration based on the difference in required speed and time between adjacent flight segments.

[0290] In practice, the required average speed for each flight segment is compared with the upper and lower speed limits in the hardware parameter table to check whether it exceeds the limits; the required heading angle change rate at each turning point is compared with the upper limit value; and the required acceleration is compared with the upper limit value of acceleration.

[0291] An evaluation report is generated based on the comparison results, including both Boolean and string types.

[0292] The boolean type is "pass", and the boolean value is true if and only if all comparison results are within bounds; the string type "out-of-bounds details" records the first out-of-bounds parameter found and its value.

[0293] The overlap area between the estimated flight path corresponding to the control command and the flight safety constraint is calculated, and it is determined whether it exceeds the preset safety tolerance, which serves as the second evaluation indicator.

[0294] In specific implementation, the list of waypoints is received, and the path discretizer is invoked to interpolate between every two waypoints at fixed intervals to generate a set of spatial points for the estimated flight path, thereby obtaining a no-fly zone location map of the no-fly zone set.

[0295] As an optional implementation, each point in the spatial point set of the estimated flight path is traversed, and the row and column index in the no-fly zone location map is calculated using a coordinate mapping function based on the latitude and longitude coordinates of the point, and the Boolean value of the index position is read directly.

[0296] Count the number of points whose read values ​​are true, and calculate the total length of the broken line segment formed by connecting these points sequentially in the waypoint sequence.

[0297] The thread compares the calculated total length of the overlapping path with the safety tolerance distance. If the total length is less than the tolerance distance, it outputs "safety flag" as true; otherwise, it outputs "false" and the calculated total length value.

[0298] For example, the root mean square value of the positioning error provided by the positioning system is obtained, the maximum statistical value of the tracking error of the control loop is added, and then multiplied by an empirical safety factor, usually 2-3, to calculate the safety tolerance distance.

[0299] The path coverage of the estimated flight path and the corresponding waypoint sequence in the mission semantic information is compared to see if it exceeds a preset coverage threshold, which is used as the third evaluation indicator.

[0300] In specific implementation, a list of geographic features for obtaining the semantic information of the current task is obtained. The list is either a list of key points accessed in sequence or a polygonal region with defined boundaries.

[0301] Iterate through each key point in the list, find the waypoint with the smallest Euclidean distance to that key point in the decoded waypoint list, and record this minimum distance.

[0302] Compare all recorded minimum distances with a fixed arrival determination radius, such as five meters, count the number of key points smaller than this radius, and calculate the coverage ratio.

[0303] For polygonal region tasks, all decoded waypoints are used as sampling points. Ray-casting or polygon rasterization is used to determine whether each sampling point is located inside the task polygon. The number of internal points is counted and the coverage ratio of the total sampling points is calculated.

[0304] If the first evaluation metric is not exceeded, the second evaluation metric is not exceeded, and the third evaluation metric is reached or exceeded, then the evaluation is deemed successful and the control command is executed.

[0305] Based on the same inventive concept, this application also provides an autonomous control system for UAV disconnection corresponding to the autonomous control method for UAV disconnection.

[0306] See Figure 4 This is a schematic diagram of an autonomous control system for unmanned aerial vehicle (UAV) disconnection provided in an embodiment of this application. The system includes: an analysis module 10, a prediction module 20, a solution module 30, and an evaluation module 40, wherein:

[0307] Analysis module 10: used to acquire UAV data, which includes multi-source heterogeneous data and mission semantic information, and to perform situational analysis on the multi-source heterogeneous data to obtain a first situational map;

[0308] Prediction module 20: is used to process the first situation map and the task semantic information through a preset two-stage prediction method to predict and generate a spatial probability distribution map;

[0309] Solving module 30: Based on the task semantic information and the spatial probability distribution map, construct a multi-objective optimization problem that includes at least the task coverage target and the communication maintenance target, construct a first objective function and a second objective function based on the task semantic information and the spatial probability distribution map respectively, and generate flight safety constraints according to the spatial probability distribution map;

[0310] A multi-objective optimization algorithm is used to solve the flight safety constraints to obtain a set of candidate flight trajectories.

[0311] The spatial probability distribution map is evenly divided into grids according to a preset grid interval. Each grid node is associated with the spatial probability value of the corresponding spatial coordinates, and the center coordinates of the grid node are used as potential waypoints.

[0312] Initialize the particle swarm, where the position vector of each particle represents a candidate flight trajectory, which is composed of consecutive potential waypoints;

[0313] Query the spatial probability value of the potential waypoint corresponding to the particle's current position vector, and calculate the negative gradient of the spatial coordinates of the potential waypoint corresponding to the spatial probability value to obtain gradient information;

[0314] Based on the gradient information, the waypoints of the candidate flight trajectories corresponding to each particle are adjusted and non-dominated sorted to obtain the Pareto front solution set. Each solution in the Pareto front solution set corresponds to a candidate flight trajectory. The cost value of each solution in the Pareto front solution set is calculated.

[0315] Evaluation module 40: used to select the optimal trajectory from the candidate flight trajectory set to generate control commands, and to perform quantitative evaluation of the control commands.

[0316] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0317] In the description of this specification, the terms "exemplary," "for example," "specifically," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

Claims

1. An autonomous control method for unmanned aerial vehicles (UAVs) experiencing disconnection, characterized in that: The method includes: Acquire drone data, which includes multi-source heterogeneous data and mission semantic information, and perform situational analysis on the multi-source heterogeneous data to obtain a first situational map. A spatial probability distribution map is generated by predicting the first situation map and the task semantic information using a preset two-stage prediction method. Based on the task semantic information and the spatial probability distribution map, a multi-objective optimization problem is constructed, which includes at least the task coverage target and the communication maintenance target. A first objective function and a second objective function are constructed based on the task semantic information and the spatial probability distribution map, respectively, and flight safety constraints are generated according to the spatial probability distribution map. A multi-objective optimization algorithm is used to solve the flight safety constraints to obtain a set of candidate flight trajectories; the spatial probability distribution map is uniformly divided into grids according to a preset grid interval, each grid node is associated with the spatial probability value of the corresponding spatial coordinates, and the center coordinates of the grid node are used as potential waypoints; Initialize the particle swarm, where the position vector of each particle represents a candidate flight trajectory, which is composed of consecutive potential waypoints; Query the spatial probability value of the potential waypoint corresponding to the particle's current position vector, and calculate the negative gradient of the spatial coordinates of the potential waypoint corresponding to the spatial probability value to obtain gradient information; Based on the gradient information, the waypoints of the candidate flight trajectories corresponding to each particle are adjusted and non-dominated sorted to obtain the Pareto front solution set. Each solution in the Pareto front solution set corresponds to a candidate flight trajectory. The cost value of each solution in the Pareto front solution set is calculated. The optimal trajectory is selected from the set of candidate flight trajectories to generate control commands, and the control commands are quantitatively evaluated.

2. The autonomous control method for UAV disconnection according to claim 1, characterized in that, The situational analysis of the multi-source heterogeneous data includes: Extract a first feature set from the multi-source heterogeneous data. The first feature set includes feature vectors of signal strength, UAV pose, relay device pose, and relay device trajectory. Based on the semantic information of the task, the first feature set is filtered by dynamic pruning technology to obtain the second feature set; Dynamic clustering is performed on the feature vectors in the second feature set, the clustering results are mapped to situation element nodes, and weighted directed edges are established on the situation element nodes, the situation element nodes have a state history sequence. The weights are calculated by fusing feature vector similarity with task semantic relevance. The graph structure formed by the situation element nodes and the weighted directed edges is aggregated through a preset graph attention network to generate a first situation map.

3. The autonomous control method for UAV disconnection according to claim 2, characterized in that, The filtering of the first feature set using dynamic pruning technology includes: The task semantic information is encoded into a task semantic vector, and the cosine similarity between each feature vector and the task semantic vector is calculated as the task relevance score of the feature vector. The task relevance score of the feature vector is compared with a preset relevance threshold, and feature vectors with task relevance scores lower than the preset relevance threshold are filtered out to obtain a second feature set.

4. The autonomous control method for UAV disconnection according to claim 3, characterized in that, The step of predicting the first situation map and the task semantic information using a preset two-stage prediction method to generate a spatial probability distribution map includes: The preset two-stage prediction method includes a first stage and a second stage; In the first stage, the spatiotemporal feature vector of the first situation map is extracted by a spatiotemporal graph neural network; The second stage uses a latent variable conditional neural process model to decode the spatiotemporal feature vector and the task semantic vector to generate a first probability distribution parameter. Based on the first probability distribution parameters, the spatial probability distribution map is generated using a preset spatial interpolation method.

5. The autonomous control method for UAV disconnection according to claim 4, characterized in that, The spatiotemporal feature vector of the first situation map obtained by extracting it through a spatiotemporal graph neural network includes: The spatiotemporal graph neural network calculates the attention weight between any two nodes in the first situational graph through a preset edge weight-temporal difference collaborative attention mechanism, and sums the attention weights of all nodes by weight, outputting the feature representation of the node as the spatiotemporal feature vector. The inputs for calculating the attention weights include: the weights of the directed edges connecting any two nodes, and the covariance of the state sequences of any two nodes within a preset sliding time window; The spatiotemporal feature vector is mapped to the first latent variable distribution parameter through a preset encoder network, and the task semantic vector is input into the super network as the first modulation parameter of the generation network. Latent variable samples are obtained from the first modulation parameters through latent variable distribution sampling, and the latent variable samples and a preset spatial coordinate grid are input into the generation network for forward propagation to predict and output the first probability distribution parameters.

6. The autonomous control method for UAV disconnection according to claim 1, characterized in that, The generation of flight safety constraints based on the spatial probability distribution map includes: The flight safety constraint is obtained by comparing the spatial probability value of each spatial coordinate with a preset safety threshold, and extracting the set of spatial coordinates whose spatial probability value is lower than the preset safety threshold.

7. The autonomous control method for UAV disconnection according to claim 1, characterized in that, The calculation of the cost value of each solution in the Pareto front solution set includes: Calculate the total path length of each candidate flight trajectory corresponding to the Pareto front solution set, and use it as the first cost factor; The number of spatial probability values ​​corresponding to all waypoints traversed by the candidate flight trajectory that are lower than a preset communication threshold is counted and used as a second cost factor; The sum of the absolute values ​​of the turning angles between consecutive segments of the candidate flight trajectory is calculated as the third cost factor; Based on the preset first weight coefficient, second weight coefficient, and third weight coefficient, the first cost factor, second cost factor, and third cost factor are weighted and summed to obtain the cost value of the candidate flight trajectory.

8. The autonomous control method for UAV disconnection according to claim 1, characterized in that, The control command for selecting the optimal trajectory from the candidate flight trajectory set includes: The optimal trajectory is the minimum solution of the Pareto front solution set.

9. The autonomous control method for UAV disconnection according to claim 1, characterized in that, The quantitative evaluation of the control command includes: Verify whether the dynamic parameters required by the control command are within the preset execution threshold range of the UAV. The dynamic parameters include: flight speed, heading angle and acceleration, as the first evaluation index. Calculate the overlap area between the estimated flight path corresponding to the control command and the flight safety constraints, and determine whether it exceeds the preset safety tolerance, as the second evaluation indicator; The path coverage of the estimated flight path and the corresponding waypoint sequence in the mission semantic information is compared to see if it exceeds a preset coverage threshold, which is used as the third evaluation indicator. If the first evaluation metric is not exceeded, the second evaluation metric is not exceeded, and the third evaluation metric is reached or exceeded, the evaluation is deemed successful and the control command is executed.

10. An autonomous control system for unmanned aerial vehicle (UAV) disconnection, used to implement the autonomous control method for UAV disconnection as described in any one of claims 1-9, characterized in that, The system includes: Analysis module: used to acquire UAV data, which includes multi-source heterogeneous data and mission semantic information, and to perform situational analysis on the multi-source heterogeneous data to obtain a first situational map; Prediction module: used to process the first situation map and the task semantic information using a preset two-stage prediction method to predict and generate a spatial probability distribution map; Solution module: Based on the task semantic information and the spatial probability distribution map, construct a multi-objective optimization problem that includes at least the task coverage target and the communication maintenance target. Construct a first objective function and a second objective function based on the task semantic information and the spatial probability distribution map, respectively, and generate flight safety constraints based on the spatial probability distribution map. A multi-objective optimization algorithm is used to solve the flight safety constraints to obtain a set of candidate flight trajectories. The spatial probability distribution map is evenly divided into grids according to a preset grid interval. Each grid node is associated with the spatial probability value of the corresponding spatial coordinates, and the center coordinates of the grid node are used as potential waypoints. Initialize the particle swarm, where the position vector of each particle represents a candidate flight trajectory, which is composed of consecutive potential waypoints; Query the spatial probability value of the potential waypoint corresponding to the particle's current position vector, and calculate the negative gradient of the spatial coordinates of the potential waypoint corresponding to the spatial probability value to obtain gradient information; Based on the gradient information, the waypoints of the candidate flight trajectories corresponding to each particle are adjusted and non-dominated sorted to obtain the Pareto front solution set. Each solution in the Pareto front solution set corresponds to a candidate flight trajectory. The cost value of each solution in the Pareto front solution set is calculated. Evaluation module: used to select the optimal trajectory from the candidate flight trajectory set to generate control commands, and to perform quantitative evaluation of the control commands.

Images