A compound wing unmanned aerial vehicle swarm intelligence obstacle avoidance control method and system

Abstract

The application relates to a compound-wing unmanned aerial vehicle group intelligence obstacle avoidance control method and system, which comprises the following steps: constructing a state prediction model of each compound-wing unmanned aerial vehicle; acquiring multi-source heterogeneous sensing data, performing joint detection and feature extraction, and generating a local environment map containing obstacle semantic labels and motion states; generating a compressed state summary through feature extraction; broadcasting to adjacent compound-wing unmanned aerial vehicles; receiving compressed state summaries of adjacent compound-wing unmanned aerial vehicles, combining the local environment map and the multi-source heterogeneous sensing data, and generating an obstacle avoidance control increment meeting the dynamics constraint of the state prediction model through a distributed collaborative strategy network; and dynamically reconstructing the formation topology of the unmanned aerial vehicle group according to the current flight mode and the emergency degree of obstacle avoidance. Through decentralized collaborative decision-making, communication anti-interference and dynamic adjustment of the formation topology, the application improves the obstacle avoidance accuracy, robustness, reliability and formation efficiency, and reduces the communication overhead.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) control technology, specifically relating to a method and system for intelligent obstacle avoidance control of a swarm of compound-wing UAVs. Background Technology

[0002] With the gradual liberalization of low-altitude airspace management policies and the accelerated construction of urban air mobility (UAM) systems, compound-wing UAVs, with their unique advantages of combining the efficient cruise capabilities of fixed-wing aircraft with the vertical takeoff and landing capabilities of multi-rotor aircraft, are becoming core carriers for critical scenarios such as urban logistics delivery, emergency medical transport, and infrastructure inspection. When these aircraft operate in the three-dimensional low-altitude urban space, they frequently need to traverse areas with towering buildings, complex electromagnetic environments, and highly dense and dynamically changing obstacles, posing unprecedented challenges to the real-time performance, safety, and robustness of group collaborative obstacle avoidance.

[0003] Among them, the intelligent obstacle avoidance control of compound-wing UAV swarms for urban low-altitude environments aims to achieve safe and autonomous flight of high-density formations while ensuring individual dynamic constraints through distributed perception, collaborative decision-making, and adaptive execution mechanisms. This technical direction not only needs to solve the communication synchronization and path conflict problems in traditional swarm obstacle avoidance, but also needs to deeply couple the unique dynamic constraints of compound-wing platforms during mode switching (such as fixed-wing cruise and multi-rotor hovering), such as nonlinear aerodynamic characteristics, control input saturation, and state response delay.

[0004] Existing technologies have significant limitations in addressing these needs. On the one hand, some solutions rely on centralized architectures, such as using a single virtual leader to dominate global obstacle avoidance decisions. Once its perception is obstructed or the communication link is interrupted, the entire cluster will lose its collaborative obstacle avoidance capability. On the other hand, although some studies have introduced multi-agent reinforcement learning to achieve decentralized collaboration, its modeling objects are limited to isomorphic quadcopter platforms and do not consider the nonlinear dynamics of compound-wing UAVs under velocity-attitude-thrust coupling. This leads to a significant lag in control commands during the mode transition phase. Furthermore, existing methods generally lack multi-source heterogeneous perception fusion mechanisms for typical weakly textured obstacles in cities (such as glass curtain walls, high-voltage cables, and moving vehicles), and still require high-frequency, full-state broadcasting in communication-constrained environments, making them prone to collaborative failures due to packet loss or latency. These shortcomings collectively limit the achievement of highly reliable, efficient, and robust autonomous obstacle avoidance capabilities for composite-wing UAV swarms in real-world low-altitude urban scenarios, necessitating a novel control paradigm that deeply integrates platform dynamics, environmental semantic understanding, and anti-interference collaborative decision-making. Summary of the Invention

[0005] To achieve highly reliable, efficient, and robust swarm autonomous obstacle avoidance capabilities, a first aspect of this invention provides an intelligent obstacle avoidance control method for a swarm of hybrid-wing UAVs, comprising: constructing a state prediction model for each hybrid-wing UAV; acquiring multi-source heterogeneous sensing data; performing joint detection and feature extraction based on the multi-source heterogeneous sensing data to generate a local environment map containing obstacle semantic labels and motion states; generating a compressed state summary based on the multi-source heterogeneous sensing data and the current state characteristics of the hybrid-wing UAVs through feature extraction; broadcasting the compressed state summaries to neighboring hybrid-wing UAVs through an anti-interference communication mechanism; receiving the compressed state summaries from neighboring hybrid-wing UAVs, combining the local environment map and the multi-source heterogeneous sensing data, and generating an obstacle avoidance control increment that satisfies the dynamic constraints of the state prediction model via a distributed cooperative strategy network; dynamically allocating the weights of the trajectory tracking control quantity and the obstacle avoidance control increment according to the current flight mode and the urgency of obstacle avoidance, generating and executing a fused control command, and simultaneously dynamically reconstructing the formation topology of the UAV swarm based on communication quality and the spacing between neighboring hybrid-wing UAVs.

[0006] In some embodiments of the present invention, the state prediction model includes: a state vector, including body axis velocity, attitude, angular velocity, and propulsion unit thrust state; a control input vector, including control input vectors for control surface deflection commands and propeller commands; the control input vector is constrained within a physically feasible range by a saturation function; modal weights, used to characterize the continuous switching between fixed-wing cruise mode, multi-rotor hovering mode, and transition phase; the modal weights are generated logically based on airspeed thresholds; and a thrust coupling model, used to characterize the contribution of propeller thrust to the resultant force or resultant moment of the airframe system and aerodynamic interference effects; the thrust coupling model is constructed based on a thrust mapping matrix and coupling coefficient terms.

[0007] In some embodiments of the present invention, the joint detection and feature extraction includes: acquiring three-dimensional point cloud data through lidar, acquiring penetration distance data through millimeter-wave radar, and acquiring texture semantic data through a visual sensor; jointly detecting weakly textured obstacles and dynamic obstacles, and generating the semantic labels through a lightweight semantic segmentation network; tracking the motion state of the dynamic obstacles using Kalman filtering, predicting their motion trajectory within a predetermined time period in the future, and using the motion trajectory as a constraint condition for generating the obstacle avoidance control increment.

[0008] In some embodiments of the present invention, generating the obstacle avoidance control increment that satisfies the dynamic constraints of the state prediction model via the distributed cooperative strategy network includes: constructing a local communication graph containing its own node and neighboring nodes within a predetermined neighborhood; constructing node feature vectors using the embedding vectors of its own dynamic state and the semantic labels of the obstacles, and constructing edge feature vectors using relative position and relative velocity; aggregating neighborhood information through a graph neural network containing a multi-layer message passing module, and allocating the contribution of neighboring machines in conjunction with an attention mechanism, and outputting the normalized control increment; embedding the safety barrier function as a differentiable penalty term into the loss function of the distributed cooperative strategy network to ensure that the distance between any two machines is always greater than the safety distance dynamically adjusted with flight speed.

[0009] In some embodiments of the present invention, the step of generating a compressed state summary based on the multi-source heterogeneous sensing data and the current state characteristics of the compound-wing UAV by feature extraction includes: extracting position, velocity, acceleration, and distance to the nearest obstacle as the original state summary vector; reducing the dimensionality of the standardized preprocessed original state summary vector by principal component analysis, retaining the principal components that reach the cumulative variance contribution rate threshold, and generating the compressed state summary.

[0010] In some embodiments of the present invention, the step of dynamically allocating the weights of the trajectory tracking control quantity and the obstacle avoidance control increment according to the current flight mode and the obstacle avoidance urgency includes: the obstacle avoidance urgency is determined by the predicted collision time and predicted minimum distance of the nearest obstacle or neighboring aircraft; when the flight speed is in the high-speed cruise range, the trajectory smoothness weight is allocated first; when the flight speed is in the low-speed hovering range, the obstacle avoidance response weight is allocated first; when in the intermediate speed range, linear interpolation is performed, and the gain of the obstacle avoidance urgency is superimposed, and the weights are smoothly switched by first-order filtering.

[0011] A second aspect of the present invention provides a swarm intelligent obstacle avoidance control system for compound-wing unmanned aerial vehicles (UAVs), comprising: a construction module for constructing a state prediction model for each compound-wing UAV, wherein the state prediction model integrates the mode switching and thrust coupling constraints of the compound-wing UAVs and introduces an external disturbance term to characterize environmental interference; a generation module for acquiring multi-source heterogeneous sensing data; performing joint detection and feature extraction based on the multi-source heterogeneous sensing data to generate a local environment map containing obstacle semantic labels and motion states; and an extraction module for extracting the state features of the current compound-wing UAV based on the multi-source heterogeneous sensing data and reducing... The system generates a compressed state summary and broadcasts it to neighboring compound-wing UAVs via an anti-interference communication mechanism. A receiving module receives the compressed state summaries from neighboring compound-wing UAVs, combines them with the local environment map and the multi-source heterogeneous sensing data, and generates obstacle avoidance control increments that satisfy the dynamic constraints of the state prediction model via a distributed cooperative strategy network. An execution module dynamically allocates the weights of the trajectory tracking control quantity and the obstacle avoidance control increments based on the current flight mode and the urgency of obstacle avoidance, generates and executes fused control commands, and simultaneously reconstructs the UAV swarm formation topology based on communication quality and the distance between neighboring compound-wing UAVs.

[0012] A third aspect of the present invention provides an electronic device comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the swarm intelligent obstacle avoidance control method for compound-wing unmanned aerial vehicles provided in the first aspect of the present invention.

[0013] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the intelligent obstacle avoidance control method for swarms of compound-wing unmanned aerial vehicles provided in the first aspect of the present invention.

[0014] The beneficial effects of this invention are: By constructing a state prediction model that explicitly includes mode switching, thrust coupling, and control saturation constraints, the control response capability and obstacle avoidance stability of the compound wing UAV during the high-speed cruise and low-speed hovering switching phase were improved. Through a multi-source heterogeneous perception fusion mechanism, the ability to detect and identify weak texture obstacles or dynamic obstacles such as glass curtain walls, high-voltage cables, and moving vehicles has been enhanced. By adopting a decentralized collaborative decision-making mechanism based on graph neural networks, the obstacle avoidance robustness and collaborative reliability of the group under complex conditions such as single-machine failure and communication limitations are improved. By using state summary compressed broadcast and anti-interference communication mechanisms, the communication load in group collaboration is reduced and the information transmission efficiency in bandwidth-constrained environments is improved. By controlling the dynamic allocation of weights and real-time reconfiguration of formation topology, flight safety, control feasibility, and mission execution efficiency are all taken into account. Attached Figure Description

[0015] Figure 1 This is a basic flowchart illustrating the intelligent obstacle avoidance control method for a swarm of compound-wing UAVs in some embodiments of the present invention. Figure 2 This is a schematic diagram illustrating the specific process of the intelligent obstacle avoidance control method for a swarm of compound-wing UAVs in some embodiments of the present invention. Figure 3 This is a schematic diagram illustrating the core principle framework of decentralized swarm intelligence collaborative decision-making in some embodiments of the present invention; Figure 4 This is a logical flowchart of multi-source heterogeneous sensing fusion and state summary compression transmission in some embodiments of the present invention. Figure 5 This is a schematic diagram of the structure of a composite-wing UAV swarm intelligent obstacle avoidance control device in some embodiments of the present invention; Figure 6 This is a schematic diagram of the structure of an electronic device in some embodiments of the present invention. Detailed Implementation

[0016] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0017] Example 1 refer to Figure 1 and Figure 2 In a first aspect, the present invention provides an intelligent obstacle avoidance control method for a swarm of compound-wing UAVs, comprising: S100. constructing a state prediction model for each compound-wing UAV; S200. acquiring multi-source heterogeneous sensing data; performing joint detection and feature extraction based on the multi-source heterogeneous sensing data to generate a local environment map containing obstacle semantic labels and motion states; S300. generating a compressed state summary based on the multi-source heterogeneous sensing data and the current state characteristics of the compound-wing UAVs through feature extraction; broadcasting the compressed state summary to neighboring compound-wing UAVs through an anti-interference communication mechanism; S400. receiving the compressed state summaries of neighboring compound-wing UAVs, combining the local environment map and the multi-source heterogeneous sensing data, and generating an obstacle avoidance control increment that satisfies the dynamic constraints of the state prediction model through a distributed cooperative strategy network; S500. dynamically allocating the weights of the trajectory tracking control quantity and the obstacle avoidance control increment according to the current flight mode and the urgency of obstacle avoidance, generating and executing a fused control command, and simultaneously reconstructing the formation topology of the UAV swarm based on communication quality and the spacing between neighboring compound-wing UAVs.

[0018] In step S100 of some embodiments of the present invention, a state prediction model is constructed for each compound-wing UAV. The state prediction model includes: a state vector, comprising body axis velocity, attitude, angular velocity, and thrust state of the propulsion unit; a control input vector, comprising control input vectors for control surface deflection commands and thruster commands; the control input vectors are constrained within a physically feasible range by a saturation function; modal weights, used to characterize the continuous switching between fixed-wing cruise mode, multi-rotor hovering mode, and transition phases; the modal weights are generated logically based on airspeed thresholds; and a thrust coupling model, used to characterize the contribution of thruster thrust to the resultant force or moment of the airframe system and aerodynamic interference effects; the thrust coupling model is constructed based on a thrust mapping matrix and coupling coefficient terms. The state prediction model integrates the mode switching and thrust coupling constraints of the compound-wing UAV and introduces external disturbance terms to characterize environmental interference.

[0019] Specifically, an individual state prediction model integrating the dynamic characteristics of the compound wing is constructed: based on the nonlinear aerodynamic coupling characteristics of the compound wing UAV in the mode switching process between fixed-wing cruise and multi-rotor hovering, a state-space model including velocity, attitude, thrust and control input saturation constraints is established, and external disturbance terms are introduced to characterize the impact of urban low-altitude wind field and electromagnetic interference on flight state. The state-space model adopts a nonlinear state-space form that integrates compound wing mode switching and thrust coupling, explicitly defining the state vector, control input vector, control saturation function, mode switching variables, and disturbance terms. The state vector... It should at least include body axis velocity, attitude and angular velocity, and propulsion unit thrust status, control input. It must include at least the control surface deflection command and the thruster speed / duty cycle command; the control input is saturated with a function. Constraints are within physical feasibility limits. Modal weights are set to characterize the continuous switching between fixed-wing cruise, multi-rotor hovering, and transition phases. ,in For multi-rotor mode, 1 represents the high-speed fixed-wing mode; the weights are generated by airspeed threshold logic and can employ hysteresis to suppress jitter. (Nonlinear function) It consists of "aerodynamic terms + thrust mapping terms + thrust dynamic terms + rigid body kinematic terms," ​​where thrust coupling is explicitly modeled through the mapping matrix from thrust to the resultant force / moment of the engine system and the coupling coefficient terms: the mapping matrix is ​​used to characterize the contribution of the tail-thrust propeller and multi-rotor thrust to the resultant force / moment, and the coupling coefficient is used to characterize the influence of rotor downwash on the aerodynamics of the fixed wing and the suppression effect of cruise airflow on the effective thrust of the rotor. External disturbance terms. This is used to characterize the equivalent acceleration / torque disturbances caused by urban low-altitude wind fields and electromagnetic environment, and is updated in real time by an online disturbance estimator. The specific implementation steps are as follows: , , , , , , Where x(t) represents the state vector, f( represents the time derivative of the state vector) ) represents the nonlinear system dynamics function, which consists of aerodynamic terms, thrust mapping terms, thrust dynamic terms, and rigid body kinematic terms; sat(u(t)) represents the saturation treatment of the control input; σ(t) represents the mode switching variable: 0 represents the low-speed multi-rotor mode, and 1 represents the high-speed fixed-wing mode; u(t) represents the control input; Indicates air speed, , Indicates the mode switching threshold; m represents the quality. Indicates aerodynamic force, , The mapping matrix representing the thrust to the resultant force / moment of the engine system; g b Δd(x(t),t) represents gravity; J represents the inertia tensor; Δd(x(t),t) represents the external disturbance term, which represents the equivalent acceleration / torque disturbance caused by wind, electromagnetic interference, etc. η represents the velocity vector (3D) in the body coordinate system; η represents the attitude vector, preferably the Euler angle attitude vector, typically... ω represents the angular velocity vector of the airframe (around the airframe axis); T represents the thrust state vector of the propulsion unit; δ represents one of the control inputs, the control surface deflection angle or the propeller pitch angle; n represents the other control input, the motor speed or the rotor pitch. refer to Figure 4 In step S200 of some embodiments of the present invention, multi-source heterogeneous sensing data is acquired; based on the multi-source heterogeneous sensing data, joint detection and feature extraction are performed to generate a local environment map containing obstacle semantic labels and motion states; wherein, the joint detection and feature extraction includes: S201. Acquire 3D point cloud data using LiDAR, acquire penetration distance data using millimeter-wave radar, and acquire texture semantic data using a visual sensor; S202. Jointly detect weak texture obstacles and dynamic obstacles, and generate the semantic labels using a lightweight semantic segmentation network. Specifically, the lidar operates at a specific wavelength, with a preset scanning frequency and a point cloud density greater than a predetermined density threshold; the millimeter-wave radar operates at a specific frequency band, with a predetermined detection range and a sensitivity to the reflective cross-section of obstacles such as metal cables that is better than a preset sensitivity threshold; the vision sensor uses a global shutter CMOS image sensor with a predetermined resolution and a preset frame rate, and performs pixel-level classification of glass curtain walls and moving vehicles through a lightweight semantic segmentation network, with the classification confidence threshold set to a preset confidence threshold.

[0020] S203. Perform Kalman filtering to track the motion state of the dynamic obstacle, predict its motion trajectory within a predetermined time period in the future, and use the motion trajectory as a constraint condition for generating the obstacle avoidance control increment.

[0021] It is understandable that anti-interference communication and state summary compression transmission are implemented: under communication-limited conditions, only the state summary vector after dimensionality reduction by principal component analysis is broadcast, and its dimension is compressed to below a predetermined proportion of the original state. Sliding window retransmission and timestamp verification mechanisms are used to suppress information loss and delay.

[0022] Therefore, in step S300 of some embodiments of the present invention, based on the multi-source heterogeneous sensing data and the current state characteristics of the compound-wing UAV, a compressed state summary is generated through feature extraction; and broadcast to adjacent compound-wing UAVs through an anti-interference communication mechanism; wherein, the step of generating a compressed state summary based on the multi-source heterogeneous sensing data and the current state characteristics of the compound-wing UAV through feature extraction includes: S301. Extract position, velocity, acceleration, and distance to the nearest obstacle as the original state summary vector; Specifically, state summary generation is triggered at a predetermined period (e.g., initially once every 100ms): five core components are extracted from the local state: two-dimensional position component, velocity magnitude, acceleration magnitude, and distance to the nearest obstacle, and concatenated to form the original summary vector. .

[0023] S302. The original state summary vector after standardization and preprocessing is reduced in dimensionality by principal component analysis, and the principal components that reach the cumulative variance contribution rate threshold are retained to generate the compressed state summary.

[0024] Specifically, the Before PCA calculation, standardization preprocessing is performed, which involves subtracting the mean from the samples within the sliding window and dividing by the standard deviation to avoid principal component bias caused by dimensional differences. The covariance matrix is ​​calculated online with a sliding window length of a preset number of frames (e.g., 100 frames), and eigenvectors and eigenvalues ​​are obtained through singular value decomposition (SVD) or eigenvalue decomposition. The number of principal components is selected based on a cumulative variance contribution rate threshold, preferably retaining the first two principal components to achieve a preset cumulative variance contribution rate (e.g., 96.2%), thus forming the dimensionality-reduced summary vector. To broadcast.

[0025] Dimensionality reduction calculations are preferably performed by the FPGA's internal PCA coprocessor, and the output y and its timestamp are broadcast via a Wi-Fi 6 link. If the ACK is not returned within a predetermined time limit (e.g., 200ms), a retransmission is initiated, with a maximum retransmission count of a preset limit (e.g., 3 times). The receiving end verifies the timestamp against a window (e.g., "current time"). "Summary timestamp ≤ 200ms" filters out expired data to ensure decisions are based on the latest information. To guarantee repeatability and consistency, when using online updates to the covariance matrix, the sending end sets version numbers for the PCA parameters (mean, standard deviation, projection matrix) and broadcasts them synchronously after the update, enabling the receiving end to perform consistent decoding or consistency verification under the same version parameters.

[0026] , , , in, X raw Represents the original input features: location ( px , py ), velocity v, acceleration a, distance to the obstacle ; This represents the standardized feature vector; μ Σ represents the mean vector of the eigenvectors; Σ represents the diagonal matrix of the standard deviations of the eigenvectors. C represents the inverse of the standard deviation diagonal matrix; C represents the covariance matrix. W Represents an orthogonal matrix. denoted as a diagonal matrix; K represents the number of principal components retained (i.e., the dimension after dimensionality reduction). λ Let λi represent the i-th eigenvalue, and ∑λi represent the total variance.

[0027] refer to Figure 3In step S400 of some embodiments of the present invention, generating the obstacle avoidance control increment that satisfies the dynamic constraints of the state prediction model via the distributed cooperative strategy network includes: constructing a local communication graph containing its own node and neighboring nodes within a predetermined neighborhood; constructing node feature vectors using its own dynamic state and the embedding vectors of the obstacle semantic labels, and constructing edge feature vectors using relative position and relative velocity; aggregating neighborhood information through a graph neural network containing a multi-layer message passing module, and allocating the contribution of neighboring machines in conjunction with an attention mechanism, and outputting the normalized control increment; embedding the safety barrier function as a differentiable penalty term into the loss function of the distributed cooperative strategy network to ensure that the distance between any two machines is always greater than the safe distance dynamically adjusted with flight speed.

[0028] A distributed cooperative strategy network based on graph neural networks is used to compute local obstacle avoidance actions. After receiving its own perception data and the state summary broadcast by neighboring drones, each drone constructs a local communication graph centered on itself within a predetermined neighborhood (e.g., a radius of 50m). , where the set of nodes The set of edges includes the node itself and its neighboring nodes. This indicates that the communication link exists and the data is valid as verified by timestamps. Node feature vector. It is obtained by encoding "self-dynamic state summary + semantically aware summary": where the semantically aware summary contains at least the embedding vector of the obstacle semantic label c. The dynamic state summary includes at least the components of position, velocity, acceleration, and distance to the nearest obstacle; edge feature vectors At least include relative position With relative velocity Furthermore, it can include link quality indicators to suppress the impact of low-reliability links. The graph neural network contains three message passing modules, with each layer having a hidden feature dimension of 64 and an activation function of ReLU: the first layer mainly aggregates the relative position and relative velocity information of neighboring machines; the second layer introduces obstacle semantic label embeddings for fusion based on the aggregation results of the first layer; the third layer fuses the aggregated neighborhood information with its own dynamic state to output a control increment that satisfies the constraints of dynamics and control input range. To ensure output stability when the network topology changes, neighborhood aggregation preferably uses weighted summation or mean aggregation; further preferably, attention weights are introduced. The contribution of neighboring machines is automatically allocated based on link quality and interaction strength. The network output layer uses a hyperbolic tangent function to normalize the control increment to […]. The interval [1,1] is scaled and mapped to physical control quantities (such as heading angle / speed / rate of climb, etc.), and then after saturation constraints, it is input into the flight control interface module to be converted into control surface and electronic speed controller executable commands.

[0029] The policy network training phase employs a centralized training, decentralized execution (CTDE) architecture: a centralized experience replay pool records simulation interaction samples exceeding a preset size, with samples formatted as quintuples (...). e) storage format, where Includes local graph structure and node / edge features. To control the increment for each node, For joint rewards, policy updates are preferably performed using the Proximal Policy Optimization (PPO) algorithm or its Multi-Agent Extension (MAPPO) for iterative updates. The reward function includes at least three categories of terms: obstacle avoidance safety, control smoothness, and task efficiency. The violation of the safety barrier function is embedded in the reward / loss as a penalty term, achieving end-to-end differentiability optimization, enabling those skilled in the art to reproduce the training and deployment process.

[0030] , , , , , Among them, the current state As a vector, it contains the local graph structure and node / edge features; action a Represents the control increment vector for each agent; The next state vector after the environment transition The joint reward is in scalar form. The initial hidden state of the graph neural network. Summation based on node features Node category embedding Relative position change Relative velocity change With direction / attitude vector Concatenation and assembly. Message generation function during the message passing phase. Message generation based on node status Through attention weights Weighted aggregation of neighbor messages Combined with linear transformation parameters W (0) b (0) Update the hidden state with the ReLU activation function. Second layer attention weights By scoring function g (0) Calculate and normalize using Softmax to dynamically allocate neighbor importance. In the control output stage, the policy network generates control increments. The hyperbolic tangent function tanh and the upper control limit After scaling, overlay reference control instructions. Then, by limiting the amplitude using the saturation function sat, a safe and feasible control quantity is finally output. Loss function L For multi-objective weighted combinations: including PPO / MAPPO policy loss, LPPO / MAPPO, and value function loss. (weight) λv Strategy Entropy H (weight) λH (Promoting exploration of diversity) and safety barrier penalties (weight) λb (Imposing differentiable penalties on constraint violations) jointly drives the strategy to achieve end-to-end collaborative optimization among task efficiency, control smoothness, system stability and physical security.

[0031] It should be noted that the decentralized swarm intelligence collaborative decision-making is performed as follows: Each UAV calculates local obstacle avoidance actions based on its own perception data and compressed state summaries broadcast by neighboring nodes using a distributed collaborative strategy network based on graph neural networks. This network takes relative position, velocity vector, and obstacle semantic labels as inputs and outputs safety control increments that satisfy dynamic constraints. The decentralized swarm intelligence collaborative decision-making also includes the introduction of a safety barrier function to ensure that the distance between any two UAVs is always greater than a preset safety distance. This safety distance is dynamically adjusted according to the current flight speed, and its value is between a preset minimum safety distance and a preset maximum safety distance. The gradient of the barrier function is directly embedded into the loss function of the strategy network for end-to-end optimization.

[0032] In step S500 of some embodiments of the present invention, the weights of the trajectory tracking control quantity and the obstacle avoidance control increment are dynamically allocated according to the current flight mode and obstacle avoidance urgency, a fused control command is generated and executed, and the formation topology of the UAV group is dynamically reconstructed based on communication quality and the spacing between adjacent compound-wing UAVs. The dynamic allocation of weights between the trajectory tracking control quantity and the obstacle avoidance control increment according to the current flight mode and obstacle avoidance urgency includes: The obstacle avoidance urgency level is determined by the predicted collision time and predicted minimum distance of the nearest obstacle or neighboring aircraft: when the flight speed is in the high-speed cruise range, the flight path smoothness weight is given priority; when the flight speed is in the low-speed hovering range, the obstacle avoidance response weight is given priority; when in the intermediate speed range, linear interpolation is performed, and the gain of the obstacle avoidance urgency level is superimposed, and the weight is smoothly switched by first-order filtering.

[0033] Specifically, the control weight allocation is determined based on both the flight modality recognition results and the urgency of obstacle avoidance, ensuring the stability and implementability of the control law during modality transitions and emergency obstacle avoidance scenarios. In addition to velocity, flight modality recognition further incorporates attitude and propulsion allocation information: the main control module is based on ground speed... Pitch angle and rotor thrust ratio Calculate modal weights Hysteresis and first-order filtering are used to suppress frequent flips near the threshold. The urgency of obstacle avoidance is quantified by indicators. This indicates that the preferred method is to use the predicted collision time (TTC) of the nearest obstacle / neighboring vehicle and the predicted minimum distance. Together, it is determined that when TTC is less than a preset threshold or When the distance is less than the dynamic safety distance, E increases.

[0034] Track smoothness weight Obstacle avoidance response weights Employ a smooth transition instead of a hard switch, and meet the requirements. In one embodiment, when When entering high-speed cruise mode, set , ;when When entering low-speed hovering mode, the weights are reversed; when the speed is in the middle range, press... Linear interpolation and superposition of urgency gain Increase the obstacle avoidance response weights, and apply a time constant to the weights. First-order filtering achieves smooth switching and avoids switching jitter. The aforementioned weights are used to fuse the "basic trajectory tracking command" and the "GNN output control increment," and are sent to the FCU for superposition and execution via the flight control interface module and CAN bus, thereby ensuring control feasibility and stability.

[0035] , , , , , in, A comprehensive indicator representing collision risk, quantifying the urgency of obstacle avoidance. v At the current flight speed, σ This is a continuous variable based on speed normalization, ranging from [0,1], representing the proportion of operating states from low speed to high speed. This value is used to adjust the participation of obstacle avoidance response. Simultaneously, the system utilizes an urgency level index... E Adjust the obstacle avoidance response weights to increase the priority of obstacle avoidance responses in high-risk scenarios: The time to collision threshold is the preset time, and TTC is the currently predicted TTC value. For safe distance threshold, This represents the predicted minimum obstacle distance.E The risk assessment results were combined from both temporal and spatial dimensions, and the results were analyzed using the max and clip functions and filtering coefficients. , The time constant is restricted to the range [0,1] to reflect the urgency of the current obstacle avoidance task. Based on this, the system uses a time constant. A first-order filtering smoothing of obstacle avoidance weights is implemented to prevent drastic fluctuations caused by environmental disturbances; β This is the emergency gain factor, used for amplification. E The impact on obstacle avoidance weights is then smoothed using an exponential decay filter. The "basic trajectory tracking command" typically comes from a traditional controller (such as PID or LQR). The control increment output by the Graph Neural Network (GNN) represents the correction amount provided by the intelligent perception and decision-making modules. Through the aforementioned weighted fusion, the system organically combines "deterministic tracking" and "intelligent obstacle avoidance" to form a unified control command. This command, processed by the flight control module, is sent to the Flight Control Unit (FCU) via the CAN bus, enabling coordinated execution of multi-level control. This ensures flight stability while achieving adaptive response to complex dynamic environments. The entire process achieves closed-loop optimization from perception and decision-making to execution, improving the system's robustness and feasibility.

[0036] Optionally, the multi-source heterogeneous perception fusion mechanism further includes Kalman filtering to track the velocity and acceleration of dynamic obstacles, predicting their trajectory within a predetermined time period in the future, with the prediction error being less than a preset error limit in both the lateral and longitudinal directions, and the predicted trajectory serving as an input constraint for obstacle avoidance decision-making.

[0037] The decentralized swarm intelligence collaborative decision-making system introduces a security barrier function to constrain the distance between any two drones to always be greater than the dynamic safe distance, and embeds the barrier violation amount as a differentiable penalty term into the policy network loss function for end-to-end optimization. For any two drones... and Define relative distance Define security constraint functions ,in A dynamic safety distance is set, adjusted according to speed, and limited to a preset minimum and maximum safety distance (e.g., 3m to 10m). This dynamic safety distance is preferably implemented using a linear-saturation mapping or a piecewise function: the greater the speed, the larger the safety distance; and speed thresholds (e.g., 5m / s and 15m / s) can be used to ensure the safety distance aligns with the mode decision. In reinforcement learning training, the barrier constraint is embedded as a penalty term in the total loss: when... The penalty is 0 when When the safe distance is violated, the penalty increases quadratically with the degree of violation. This penalty term forms a differentiable link between the policy output action and the distance function through system state prediction, so that the gradient can be obtained by automatic differentiation and participate in policy update, thus satisfying the feasibility requirement of "gradient embedding loss function".

[0038] Optionally, a ground monitoring station is also included, which is used to receive cluster status summaries and issue emergency instructions in the event of extreme communication interruption. Emergency instructions have higher priority than local decisions, but are only activated when no signals from neighboring machines are received for a continuous predetermined period of time, so as to avoid frequent intervention that may disrupt the decentralized mechanism.

[0039] Example 2 refer to Figure 5 In a second aspect, the present invention provides a swarm intelligent obstacle avoidance control system 1 for composite-wing unmanned aerial vehicles (UAVs), comprising: a construction module 11 for constructing a state prediction model for each composite-wing UAV, wherein the state prediction model integrates the mode switching and thrust coupling constraints of the composite-wing UAVs and introduces an external disturbance term to characterize environmental interference; a generation module 12 for acquiring multi-source heterogeneous sensing data; performing joint detection and feature extraction based on the multi-source heterogeneous sensing data to generate a local environment map containing obstacle semantic labels and motion states; and an extraction module 13 for extracting the state features of the current composite-wing UAV based on the multi-source heterogeneous sensing data. The system performs dimensionality reduction to generate a compressed state summary; it then broadcasts this summary to adjacent hybrid-wing UAVs via an anti-interference communication mechanism; a receiving module 14 receives the compressed state summary from adjacent hybrid-wing UAVs, combines it with the local environment map and the multi-source heterogeneous perception data, and generates an obstacle avoidance control increment that satisfies the dynamic constraints of the state prediction model via a distributed cooperative strategy network; an execution module 15 dynamically allocates the weights of the trajectory tracking control quantity and the obstacle avoidance control increment according to the current flight mode and the urgency of obstacle avoidance, generates and executes fused control commands, and simultaneously reconstructs the formation topology of the UAV group based on communication quality and the distance between adjacent hybrid-wing UAVs.

[0040] Furthermore, the receiving module 14 includes: a first construction unit, used to construct a local communication graph containing its own node and neighboring nodes within a predetermined neighborhood; a second construction unit, used to construct node feature vectors using its own dynamic state and the embedding vectors of the obstacle semantic labels, and to construct edge feature vectors using relative position and relative velocity; an output unit, used to aggregate neighborhood information through a graph neural network containing a multi-layer message passing module, and to allocate the contribution of neighboring machines in combination with an attention mechanism, and output the normalized control increment; and an embedding unit, used to embed the safety barrier function as a differentiable penalty term into the loss function of the distributed cooperative strategy network to ensure that the distance between any two machines is always greater than the safety distance dynamically adjusted with flight speed.

[0041] In one specific embodiment of the invention, it is applied to an urban low-altitude logistics delivery scenario, where a swarm of 12 compound-wing UAVs needs to perform point-to-point material delivery tasks in densely populated areas of high-rise buildings. The flight altitude ranges from 30 to 80 meters, and the environment includes typical weakly textured or dynamic obstacles such as glass curtain wall buildings, high-voltage power lines, ground mobile vehicles, and temporary construction cranes. To ensure the safe and coordinated obstacle avoidance capabilities of the high-density formation under complex electromagnetic and wind field interference, this embodiment deploys a hardware and software integrated system that deeply integrates platform dynamics characteristics, multi-source perception fusion, and decentralized intelligent decision-making.

[0042] First, from the perspective of system architecture construction, each compound-wing UAV in the cluster is equipped with an integrated airborne intelligent obstacle avoidance control unit, which consists of the following hardware modules: main control computing module, multi-source perception fusion module, communication and status summary processing module, flight control execution interface module, and power management and redundancy monitoring subsystem.

[0043] The main control computing module adopts a heterogeneous computing architecture, including a 1.8GHz quad-core ARM Cortex-A72 processor (for running the operating system and high-level decision-making logic) and a dedicated neural network accelerator (NPU) with a computing power of 4 TOPS (INT8) for real-time inference of graph neural network policy models. This module is interconnected at high speed with the perception fusion module via a PCIe 3.0 x4 interface and connected to the communication module and flight control interface module via dual gigabit Ethernet interfaces, ensuring data throughput latency of less than 5 milliseconds. The operating system is based on a customized Linux kernel and enables a real-time scheduling strategy (SCHED_FIFO) to ensure deterministic response for critical tasks.

[0044] The multi-source sensing fusion module consists of three types of sensors: (1) a mechanically rotating lidar with a wavelength of 905 nm, a horizontal field of view of 360°, a vertical field of view of 30°, a scanning frequency of 20 Hz, and a point cloud output density of 1200 points per square meter, which is connected to the main control via a USB 3.0 interface; (2) a 77 GHz millimeter-wave radar with a MIMO antenna array, a maximum detection distance of 150 meters, a distance resolution of 0.25 meters, a velocity resolution of 0.1 m / s, and a radar cross section (RCS) sensitivity of metal cable obstacles better than -30 dBsm, which transmits raw intermediate frequency data at a clock rate of 10 MHz via an SPI bus; (3) a global shutter CMOS image sensor with a resolution of 1280×720, a frame rate of 30 fps, and a dynamic range of 72 dB, which is transmitted via MIPI. The CSI-2 four-channel interface connects to the main controller, which deploys a lightweight semantic segmentation network (MobileNetV3+DeepLabV3+) at its front end. The model has only 1.2M parameters, an inference latency of 8 milliseconds on the NPU, and a classification confidence threshold of 0.85. It can effectively identify targets such as glass curtain walls, moving vehicles, and construction machinery.

[0045] The communication and state summary processing module is based on a dual-mode wireless communication architecture: the main link uses a 5.8GHz OFDM Wi-Fi6 chipset (supporting 802.11ax), with a theoretical bandwidth of 1.2Gbps and an actual effective throughput of approximately 300Mbps; the backup link is a Sub-1GHz LoRa module (center frequency 868MHz) for low-speed emergency communication under extreme obstruction. This module has a built-in FPGA coprocessor for executing Principal Component Analysis (PCA) dimensionality reduction and sliding window retransmission protocol. In the current configuration of this embodiment, the original state summary vector has 5 dimensions, including two-dimensional position components, velocity magnitude, acceleration magnitude, and the distance to the nearest obstacle; after PCA compression, the first two principal components are retained, resulting in a compressed dimension of 2, with a cumulative variance contribution rate of approximately 96.2%. The broadcast period is adaptively adjusted according to communication quality: when the neighbor signal strength (RSSI) is >-75dBm and the packet loss rate is <5%, the period is set to 100ms; when RSSI is <-85dBm or there are ≥2 consecutive packet losses, the period is extended to 500ms and the retransmission mechanism is started, with a maximum of 3 retransmissions and a timestamp verification tolerance window of ±200ms.

[0046] The flight control execution interface module communicates with the underlying flight controller (FCU) via a CAN 2.0B bus (1Mbps baud rate). The FCU uses an STM32H743 microcontroller and runs the PX4 open-source flight control firmware. This module receives control increment commands (Δu) from the master controller, including elevator deflection angle, aileron deflection angle, rudder deflection angle, and the PWM duty cycle increments of the four rotor motors, and converts them into legal control quantities that comply with the physical actuator saturation constraints. For example, the control surface deflection angle is limited to ±25°, and the PWM signal duty cycle corresponding to the motor speed is limited to 10%-90%, with any excess being hard-limited to prevent actuator overload.

[0047] In addition, the system includes a Ground Control Station (GCS), deployed in the city command center, which establishes a low-priority telemetry link with the cluster via the 4G / 5G public network. The GCS continuously listens to the status summaries broadcast by each drone, and only issues an emergency return-to-home or hovering command if it does not receive a signal from any drone for 5 consecutive seconds and local decision-making fails. This command is broadcast with the highest priority via the LoRa link, with a coverage radius of up to 3 kilometers.

[0048] Based on the above system architecture, the workflow of this embodiment is as follows: Step (1): Construct an individual state prediction model that integrates the dynamic characteristics of the composite wing.

[0049] The main control module of each UAV loads the nonlinear state-space model of the compound wing during the initialization phase and performs sampling according to the sampling period. Discrete operation at 100Hz is used for rolling prediction of the state trajectory over the next 2 seconds. The state vector and control input of the model are defined as follows: State vector: in For the body axis velocity, For Euler angle orientation, Angular velocity, The thrust status of the tail thruster and each rotor; control input ,in For control surface deflection command, This refers to the thruster speed / duty cycle command. To ensure feasibility, all control commands are passed through a saturation function. Limited to physical boundaries.

[0050] Mode switching uses airspeed threshold logic to generate mode weights. It is used for continuous fusion between fixed-wing and multi-rotor dynamics; thrust coupling is achieved through a mapping matrix. Map T to the resultant force / moment of the system and introduce a coupling coefficient. Characterizing rotor downwash and incoming flow suppression effects; propeller thrust dynamics employ a first-order process. Explicitly characterize the thrust response hysteresis during the mode transition phase. External disturbance term. It is estimated and updated in real time by an online disturbance observer to compensate for the effects of urban low-altitude wind fields and electromagnetic interference.

[0051]

[0052] .

[0053] Step (2): Deploy a multi-source heterogeneous sensing fusion mechanism. The lidar outputs a point cloud frame every 50ms, which, after voxel mesh filtering (0.1m voxel size), is input to the obstacle clustering module using the DBSCAN algorithm. , The raw millimeter-wave radar data was processed using 2D-FFT to generate a range-Doppler map. Candidate points for metal cables were extracted using constant false alarm rate (CFAR) detection. The visual sensor outputs a semantic segmentation result every 33ms, labeling categories such as "glass," "vehicle," and "sky." The three data sources were spatiotemporally aligned in a unified world coordinate system (time synchronization accuracy ±2ms, spatial extrinsic calibration error <0.05m), and then fused using weighted voting: for glass curtain walls, visual semantics were prioritized (weight 0.6), supplemented by laser point cloud (weight 0.3), with millimeter-wave radar receiving a weight of 0.1 due to its poor penetration; for high-voltage cables, millimeter-wave radar was prioritized (weight 0.7), supplemented by laser point cloud (weight 0.3); for moving vehicles, all three sources were weighted equally. The fused obstacle list includes location, size, type label, and motion status (static / dynamic). For dynamic obstacles, the system activates a Kalman filter (the state vector includes position, velocity, and acceleration) to update the trajectory at a frequency of 10Hz and predict the position within the next 3 seconds. The lateral and longitudinal prediction errors are both controlled within 1.5 meters.

[0054] Step (3): Implement anti-interference communication and state summary compressed transmission. The main control module triggers state summary generation every 100ms (initial value): extracts five core components: two-dimensional position component, velocity magnitude, acceleration magnitude, and distance to the nearest obstacle, concatenates them into a 5-dimensional vector, and performs dimensionality reduction through the PCA coprocessor in the FPGA (covariance matrix updated online, sliding window length 100 frames), retaining the first two principal components (cumulative variance 96.2%) to form a 2-dimensional summary vector. This vector is broadcast through the Wi-Fi 6 link. If the ACK is not returned within 200ms, retransmission is initiated, up to 3 times. The receiving end filters out expired data through timestamp verification (current time - summary timestamp ≤ 200ms) to ensure that decisions are based on the latest information.

[0055] Step (4): Execute decentralized swarm intelligence collaborative decision-making. After receiving its own perception data and the state summary broadcast by neighboring drones (neighboring drones are defined as those within a 50-meter radius), each drone constructs a local communication graph. , where nodes For itself and its neighboring machines, the edge This indicates that a communication link exists. The Graph Neural Network (GNN) policy model is deployed on the NPU and contains three layers of message passing modules: the first layer aggregates relative positions. With relative velocity The second layer introduces obstacle semantic label embeddings (mapped to 16-dimensional vectors via a learnable lookup table); the third layer integrates its own dynamic states. Each layer's node features have a dimension of 64 and use ReLU activation. The output layer uses the tanh function to control the increment. Mapped to the [-1,1] interval, and then scaled to convert into physical control variables. This policy network employs centralized experience replay during the training phase, collecting over 1 million simulated interaction samples. After convergence, the policy success rate in real-world testing is >98%. Simultaneously, a security barrier function... Embedded loss function, where Meters, ensuring that the distance between any two machines is always greater than the dynamic safety distance (3~10 meters).

[0056] Step (5): Adaptively execute obstacle avoidance control commands and dynamically adjust formation topology. The main control module determines the flight mode based on the current ground speed: if the speed is >15m / s, it enters high-speed cruise mode, the track smoothness weight is set to 0.7, the obstacle avoidance response weight is 0.3, and the control increment mainly corrects the heading angle to bypass obstacles; if the speed is <5m / s, it enters low-speed hovering mode, the weights are reversed, and emergency avoidance actions are executed first. At the same time, the system evaluates the formation topology every 500ms: calculates the Laplace matrix L of the current communication map, and if its second smallest eigenvalue (algebraic connectivity) is <0.1, it triggers topology reconstruction—disconnects from the neighboring aircraft with the worst communication quality (lowest RSSI) and attempts to establish connections with new nodes within a range of 50~80 meters until the algebraic connectivity is >0.1 to ensure group connectivity. All control increments are converted by the flight control interface module and sent to the FCU via the CAN bus. The FCU superimposes them onto the basic track tracking commands, ultimately driving the servos and ESCs to execute.

[0057] Through the coordinated operation of the above systems and methods, this embodiment successfully traversed a 5-kilometer urban canyon route containing 12 100-meter-tall buildings, 3 high-voltage lines, and dense traffic flow in actual tests, with no collisions throughout the entire journey. The average mode switching delay was 0.28 seconds, the effective communication arrival rate was 92.3%, and the task completion time was shortened by 22.7% compared to the fixed formation scheme.

[0058] In another embodiment, in an urban emergency medical supplies delivery scenario, the cluster size is expanded to 25 composite-wing UAVs. The mission area includes temporarily constructed large exhibition venues, temporary traffic control zones, and accident sites, where obstacles are highly dynamic and unpredictable (such as temporarily raised crane booms or suddenly appearing emergency vehicles). To cope with higher density and greater uncertainty, this embodiment introduces enhanced perception redundancy and a hierarchical collaborative decision-making mechanism based on Embodiment 1.

[0059] In terms of system architecture, each UAV has been upgraded with a multi-source perception fusion module: the lidar has been replaced with a 1550nm solid-state lidar (Eye-safe level), increasing the point cloud density to 2000 points per square meter and enhancing its resistance to strong light interference; a new ultrasonic array (operating frequency 40kHz, detection range 0~5 meters) has been added as the last line of defense for close-range obstacle avoidance (<3 meters); the visual sensor has been upgraded to a binocular global shutter camera with a baseline distance of 12cm, capable of generating dense depth maps in real time to assist in the reconstruction of transparent surfaces of glass curtain walls. The communication module has added a UWB (Ultra Wide band) precise positioning subsystem (center frequency 6.5GHz, bandwidth 500MHz), providing centimeter-level relative positioning in GNSS-denied environments, with a positioning update rate of 100Hz, and connecting to the main control via a dedicated UART interface (baud rate 921600).

[0060] The main control computing module has been upgraded to a dual NPU architecture (total computing power 8 TOPS). One NPU is dedicated to running an LSTM network for dynamic obstacle trajectory prediction (128 hidden units, input is the obstacle state of the past 5 frames, output is the trajectory for the next 3 seconds), and the other runs a GNN policy network. The flight control interface module has added a hardware watchdog circuit, which automatically switches to a safe mode (hover or return to home) if the main control malfunctions.

[0061] In terms of methodology, the perception fusion mechanism in step (2) adds a sub-process of "transparent obstacle reconstruction": the binocular visual depth map and the laser point cloud are registered by ICP within a range of 0.5 to 3 meters to complete the missing point cloud on the glass surface. The completion algorithm adopts Poisson reconstruction based on planar prior, and the reconstruction error is <0.1m. Ultrasonic data is only activated when the distance is <3 meters, serving as the final collision warning and triggering a hard braking command.

[0062] The collaborative decision-making in step (3) adopts a hierarchical GNN architecture: the bottom layer is a local GNN with a radius of 50 meters, which is responsible for immediate obstacle avoidance; the upper layer is a sparse GNN with a radius of 100 meters, which is activated only when a large-scale obstacle (such as an entire construction area) is detected, and is used to coordinate the overall detour direction of the formation. The two layers are fused through an attention mechanism, and the decision weight of the upper layer is dynamically adjusted (0~0.4) according to the obstacle size.

[0063] Step (5) introduces a "virtual navigator" auxiliary mechanism for formation topology reconstruction: when the cluster enters an unknown high-risk area (such as an accident site), the ground station designates one UAV as a temporary virtual navigator (not a control center, but only providing a reference trajectory), and the other UAVs adjust their relative positions using it as a reference point, but obstacle avoidance decisions remain decentralized. This mechanism is only activated within a specific geofence and is automatically deactivated after leaving.

[0064] In a test simulating 25 aircraft traversing a sudden fire scene, this embodiment successfully avoided seven suddenly rising fire ladders (height change rate 2m / s), with the closest avoidance distance being 4.2 meters and a cluster task completion rate of 91.6%, verifying the robustness of the system in extreme dynamic environments.

[0065] In another embodiment, for ultra-large-scale urban inspection tasks, the cluster size reaches 50 aircraft, and it needs to complete the inspection of power lines and building facades in a 30-square-kilometer urban area within 2 hours. To balance communication load and obstacle avoidance performance, this embodiment introduces a spatiotemporal clustering and event-driven communication mechanism.

[0066] At the system level, each drone is equipped with a dual-band communication module: 5.8GHz for intra-cluster communication and 900MHz Sub-GHz for inter-cluster backbone. The main control module adds a cluster management coprocessor (based on a RISC-V core) to run clustering algorithms. The perception module adds an infrared thermal imager (640×512 resolution, 25Hz frame rate) for obstacle detection at night or in smoke environments.

[0067] In terms of method, the communication mechanism in step (3) is changed to event-driven: broadcasting is only done when the change in state summary exceeds a threshold (e.g., position change > 2 meters, or a new obstacle is detected); otherwise, silence is maintained. Simultaneously, the cluster is dynamically divided into several sub-clusters (8-12 machines per cluster) based on geographical location. Within each cluster, a fully connected GNN as described in Example 1 is used. Between clusters, compressed summaries are exchanged through an elected cluster head. The cluster head broadcasts a cluster state summary (including cluster center location, average speed, and main obstacle direction) every 2 seconds, and other cluster members adjust their cross-cluster obstacle avoidance strategies accordingly.

[0068] The GNN input in step (4) is enhanced with a "cluster context" feature, which includes the density of the current cluster, the number of neighboring clusters, and the backbone network load status, giving obstacle avoidance decisions a more global perspective. The safety barrier function is expanded to a cluster-level barrier to ensure that the minimum spacing between clusters is >15 meters.

[0069] In actual testing, the 50-machine cluster completed the inspection within 2 hours, with the total communication load being only 28% of that of the full broadcast solution. There were no collisions or communication congestion events, verifying the scalability and efficiency of the invention in ultra-large-scale scenarios.

[0070] Example 3 refer to Figure 6 In a third aspect, the present invention provides an electronic device comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the composite-wing unmanned aerial vehicle swarm intelligent obstacle avoidance control method of the first aspect of the present invention.

[0071] Electronic device 500 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 501, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 502 or a program loaded from storage device 508 into random access memory (RAM) 503. The RAM 503 also stores various programs and data required for the operation of electronic device 500. The processing unit 501, ROM 502, and RAM 503 are interconnected via bus 504. An input / output (I / O) interface 505 is also connected to bus 504.

[0072] Typically, the following devices can be connected to I / O interface 505: input devices 506 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 507 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 508 including, for example, hard disks; and communication devices 509. Communication device 509 allows electronic device 500 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 An electronic device 500 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 6 Each box shown can represent a device or multiple devices as needed.

[0073] Specifically, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented by computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the above-described method steps. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 509, or loaded from a storage device 508 or ROM 502. When the computer program is executed by a processing device 501, it implements the corresponding functions in the methods described in the embodiments of the present invention. Embodiments of the present invention also include a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is used to implement the steps of the above-described intelligent obstacle avoidance control method for a swarm of compound-wing UAVs. The computer-readable storage medium can be RAM, ROM, EPROM, Flash, hard disk, optical disk, or other tangible storage media capable of storing program code. The computer-readable storage medium can be included in the above-described electronic device, or it can exist independently without being assembled into the electronic device. The computer-readable storage medium carries one or more computer programs, which, when executed by the electronic device, cause the electronic device to implement the method steps described in the embodiments of the present invention.

[0074] Computer program code for performing the operations of embodiments of the present invention can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, C++, and Python—as well as conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0075] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It is important to note that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0076] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A swarm intelligent obstacle avoidance control method for compound-wing unmanned aerial vehicles (UAVs), characterized in that, include: Construct a state prediction model for each compound-wing UAV; Acquire multi-source heterogeneous sensing data; Based on the aforementioned multi-source heterogeneous sensing data, joint detection and feature extraction are performed to generate a local environment map containing semantic labels of obstacles and motion states. Based on the multi-source heterogeneous sensing data and the current state characteristics of the hybrid wing UAV, a compressed state summary is generated through feature extraction; and broadcast to adjacent hybrid wing UAVs through an anti-interference communication mechanism. The system receives compressed state summaries from adjacent compound-wing UAVs, combines them with the local environment map and the multi-source heterogeneous sensing data, and generates obstacle avoidance control increments that satisfy the dynamic constraints of the state prediction model through a distributed cooperative strategy network. Based on the current flight mode and obstacle avoidance urgency, the weights of the trajectory tracking control quantity and the obstacle avoidance control increment are dynamically allocated, fused control commands are generated and executed, and the formation topology of the UAV group is dynamically reconstructed based on communication quality and the spacing between adjacent compound-wing UAVs.

2. The intelligent obstacle avoidance control method for a swarm of compound-wing UAVs according to claim 1, characterized in that, The state prediction model includes: The state vector includes the body axis velocity, attitude, angular velocity, and thrust state of the propulsion unit; The control input vector includes control input vectors for rudder deflection commands and thruster commands; the control input vector is constrained within a physically feasible range by a saturation function. Modal weights are used to characterize the continuous switching between fixed-wing cruise mode, multi-rotor hovering mode, and transition phases; the modal weights are generated based on airspeed threshold logic. The thrust coupling model is used to characterize the contribution of the thruster thrust to the resultant force or resultant moment of the engine system and the aerodynamic interference effect; the thrust coupling model is constructed based on the thrust mapping matrix and coupling coefficient terms.

3. The intelligent obstacle avoidance control method for a swarm of compound-wing UAVs according to claim 1, characterized in that, The joint detection and feature extraction process includes: 3D point cloud data is acquired using LiDAR, and penetration distance data is acquired using millimeter-wave radar. Texture semantic data is acquired through a visual sensor; weak texture obstacles and dynamic obstacles are jointly detected, and the semantic labels are generated through a lightweight semantic segmentation network; The motion state of the dynamic obstacle is tracked by Kalman filtering to predict its trajectory within a predetermined time period in the future, and the trajectory is used as a constraint condition for generating the obstacle avoidance control increment.

4. The intelligent obstacle avoidance control method for a swarm of compound-wing UAVs according to claim 1, characterized in that, The obstacle avoidance control increment generated through the distributed cooperative policy network, which satisfies the dynamic constraints of the state prediction model, includes: Construct a local communication graph containing its own node and neighboring nodes within a predetermined neighborhood; Node feature vectors are constructed using their own dynamic state and the embedded vectors of the semantic labels of the obstacles, and edge feature vectors are constructed using their relative positions and relative velocities. The graph neural network, which includes a multi-layer message passing module, aggregates neighborhood information and combines an attention mechanism to allocate the contribution of neighboring machines, outputting a normalized control increment. The safety barrier function is embedded as a differentiable penalty term into the loss function of the distributed cooperative strategy network to ensure that the distance between any two aircraft is always greater than the safe distance that is dynamically adjusted with flight speed.

5. The intelligent obstacle avoidance control method for a swarm of compound-wing UAVs according to claim 1, characterized in that, The process of generating a compressed state summary based on the multi-source heterogeneous sensing data and the current state characteristics of the compound-wing UAV, through feature extraction, includes: Extract position, velocity, acceleration, and distance to the nearest obstacle as the original state summary vector; Principal component analysis is used to reduce the dimensionality of the original state summary vector after standardization and preprocessing, and principal components that reach the cumulative variance contribution rate threshold are retained to generate the compressed state summary.

6. The intelligent obstacle avoidance control method for a swarm of compound-wing UAVs according to claim 1, characterized in that, The dynamic allocation of weights between the trajectory tracking control quantity and the obstacle avoidance control increment based on the current flight mode and obstacle avoidance urgency includes: The urgency level of obstacle avoidance is determined by the predicted collision time of the nearest obstacle or adjacent aircraft and the predicted minimum distance: When the flight speed is in the high-speed cruise range, priority is given to allocating track smoothness weights; When the flight speed is in the low-speed hovering range, obstacle avoidance response weights are allocated first; when it is in the middle speed range, linear interpolation is performed, and the gain of the obstacle avoidance urgency is superimposed, and the weights are smoothly switched through first-order filtering.

7. A swarm intelligent obstacle avoidance control system for composite-wing unmanned aerial vehicles (UAVs), characterized in that, include: A construction module is used to build a state prediction model for each compound wing UAV. The state prediction model integrates the mode switching and thrust coupling constraints of the compound wing UAV and introduces an external disturbance term to characterize environmental disturbances. The generation module is used to acquire multi-source heterogeneous sensing data; based on the multi-source heterogeneous sensing data, joint detection and feature extraction are performed to generate a local environment map containing obstacle semantic labels and motion states; The extraction module is used to extract the current state features of the hybrid wing UAV based on the multi-source heterogeneous sensing data, reduce the dimensionality, and generate a compressed state summary; and broadcast it to neighboring hybrid wing UAVs through an anti-interference communication mechanism. The receiving module is used to receive compressed state summaries of adjacent compound-wing UAVs, combine them with the local environment map and the multi-source heterogeneous perception data, and generate obstacle avoidance control increments that satisfy the dynamic constraints of the state prediction model through a distributed cooperative strategy network. The execution module is used to dynamically allocate the weights of the trajectory tracking control quantity and the obstacle avoidance control increment according to the current flight mode and the urgency of obstacle avoidance, generate fused control commands and execute them, and at the same time dynamically reconstruct the formation topology of the UAV group based on communication quality and the spacing between adjacent compound-wing UAVs.

8. The intelligent obstacle avoidance control system for a swarm of compound-wing unmanned aerial vehicles according to claim 7, characterized in that, The receiving module includes: The first construction unit is used to construct a local communication graph containing its own node and neighboring nodes within a predetermined neighborhood. The second construction unit is used to construct node feature vectors based on its own dynamic state and the embedding vectors of the semantic labels of the obstacles, and to construct edge feature vectors based on relative position and relative velocity. The output unit is used to aggregate neighborhood information through a graph neural network containing a multi-layer message passing module, and to allocate the contribution of neighboring machines by combining an attention mechanism, and output the normalized control increment. The embedding unit is used to embed the safety barrier function as a differentiable penalty term into the loss function of the distributed cooperative strategy network to ensure that the distance between any two aircraft is always greater than the safe distance that is dynamically adjusted with flight speed.

9. An electronic device, comprising: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the intelligent obstacle avoidance control method for a swarm of compound-wing unmanned aerial vehicles as described in any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, wherein, When the computer program is executed by the processor, it implements the intelligent obstacle avoidance control method for swarms of compound-wing unmanned aerial vehicles as described in any one of claims 1 to 6.

Images