Manned unmanned aerial vehicle autonomous obstacle avoidance control method and system based on flight environment perception

Abstract

The invention discloses a manned unmanned aerial vehicle autonomous obstacle avoidance control method and system based on flight environment perception, and relates to the technical field of intelligent control. The method comprises the following steps: acquiring a multi-physical field tensor of a navigation environment through a complementary detector, and converting the multi-physical field tensor into a sparse asynchronous pulse stream through event-driven coding; inputting the sparse asynchronous pulse flow into a first decision maker, executing multi-layer recursive analysis based on spatial voxel-physical feature combination, and determining a reference control instruction; constructing a manifold keeping controller, executing quadratic programming based on a first constraint condition and a second constraint condition on the reference control instruction, and solving an optimal control quantity as a target control sequence; and according to the target control sequence, unmanned aerial vehicle obstacle avoidance navigation control is executed. The technical problem of poor obstacle avoidance reliability caused by insufficient environment perception, lagging obstacle avoidance decision and insufficient control constraint coupling of the manned unmanned aerial vehicle in a complex flight environment in the prior art is solved, and the technical effect of improving the autonomous obstacle avoidance reliability of the manned unmanned aerial vehicle is achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, specifically to a method and system for autonomous obstacle avoidance control of manned unmanned aerial vehicles based on flight environment perception. Background Technology

[0002] With the development of the low-altitude economy and the gradual advancement of urban air transportation, manned drones, as a new type of low-altitude transportation tool, are increasingly being applied to scenarios such as short-distance commuting, emergency rescue, and special operations. Because manned drones typically face complex flight environments during actual flight, such as dense buildings, frequent dynamic obstacles, complex electromagnetic environments, and significant airflow disturbances, higher demands are placed on their environmental perception accuracy, real-time decision-making capabilities, and control safety. In existing technologies, manned drones mostly employ single or simple fusion perception methods based on lidar, visual sensors, or millimeter-wave radar, generating obstacle avoidance commands through traditional path planning algorithms or rule-driven strategies. However, in the process of fusing multi-source heterogeneous data, there are problems such as large data redundancy, asynchronous timing, and insufficient real-time processing efficiency, making it difficult to make stable decisions in a timely manner under rapidly changing dynamic obstacle scenarios. At the same time, existing control methods often separate obstacle avoidance planning from flight control, failing to fully consider the coupling relationship between drone dynamic constraints, attitude stability constraints, and safety boundary constraints. This can easily lead to path oscillations, sudden changes in control variables, or insufficient safety margins, thereby reducing the obstacle avoidance reliability and flight safety of manned drones in complex environments. Summary of the Invention

[0003] This application provides a method and system for autonomous obstacle avoidance control of manned unmanned aerial vehicles (UAVs) based on flight environment perception. It solves the technical problems in the prior art where manned UAVs have insufficient environmental perception, delayed obstacle avoidance decision-making, and insufficient control constraint coupling in complex flight environments, resulting in poor obstacle avoidance reliability.

[0004] The first aspect of this application provides an autonomous obstacle avoidance control method for a manned unmanned aerial vehicle based on flight environment perception, the method comprising: The multiphysics tensor of the navigation environment is acquired by complementary detectors at the airborne front end and converted into a sparse asynchronous pulse stream through event-driven encoding. The sparse asynchronous pulse stream is input into the first decision unit to perform multi-level recursive analysis based on the combination of spatial voxels and physical features to determine the reference control command. A high-dimensional feature space is constructed by embedding tensor kernels and a manifold-maintaining controller is constructed by combining it with safety boundaries. The reference control command is subjected to quadratic programming based on the first and second constraints to solve for the optimal control quantity as the target control sequence. According to the target control sequence, obstacle avoidance navigation control of the UAV is performed.

[0005] A second aspect of this application provides an autonomous obstacle avoidance control system for a manned unmanned aerial vehicle based on flight environment perception, the system comprising: Data acquisition module: Acquires multiphysics tensors of the flight environment through complementary detectors at the airborne front end, and converts them into sparse asynchronous pulse streams via event-driven encoding; Analysis module: Inputs the sparse asynchronous pulse streams into the first decision unit, performs multi-level recursive analysis based on spatial voxel-physical feature combination, and determines the reference control command; Optimization module: Constructs a high-dimensional feature space using tensor kernel embedding, constructs a manifold-maintaining controller in combination with safety boundaries, performs quadratic programming based on the first and second constraints on the reference control command, and solves for the optimal control quantity as the target control sequence; Control module: Performs obstacle avoidance flight control of the UAV according to the target control sequence.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, the multiphysics tensor of the flight environment is acquired through complementary detectors at the airborne front end, and then converted into a sparse asynchronous pulse stream through event-driven encoding. Next, the sparse asynchronous pulse stream is input into the first decision-maker, which performs a multi-layer recursive analysis based on spatial voxel-physical feature combinations to determine the reference control command. Then, a high-dimensional feature space is constructed by embedding tensor kernels, and a manifold-maintaining controller is built by combining it with safety boundaries. A quadratic programming problem based on the first and second constraints is performed on the reference control command to solve for the optimal control quantity as the target control sequence. Finally, obstacle avoidance navigation control is performed on the UAV according to the target control sequence. This solves the technical problems of insufficient environmental perception, delayed obstacle avoidance decision-making, and insufficient coupling of control constraints in existing technologies for manned UAVs in complex flight environments, resulting in poor obstacle avoidance reliability. It achieves the technical effect of improving the reliability of autonomous obstacle avoidance for manned UAVs. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 A schematic flowchart of the autonomous obstacle avoidance control method for manned unmanned aerial vehicles based on flight environment perception provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of an autonomous obstacle avoidance control system for a manned unmanned aerial vehicle based on flight environment perception, provided in an embodiment of this application.

[0009] Figure labeling: Data acquisition module 11, analysis module 12, optimization module 13, control module 14. Detailed Implementation

[0010] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0011] Example 1, as Figure 1 As shown, this application provides an autonomous obstacle avoidance control method for manned unmanned aerial vehicles based on flight environment perception, wherein the method includes: The multiphysics tensor of the flight environment is acquired by complementary detectors at the airborne front end and then converted into a sparse asynchronous pulse stream through event-driven encoding.

[0012] A complementary detection array consisting of visual sensors, millimeter-wave radar, and acoustic sensors is deployed at the front of the UAV. The visual sensors acquire visible light images and depth map data, the millimeter-wave radar acquires target distance, radial velocity, and penetrating environmental echo information, and the acoustic sensors acquire echoes from nearby obstacles and sound pressure signals from airflow disturbances. All sensors are calibrated using a unified clock through a time synchronization module and mapped to a unified UAV coordinate system via coordinate extrinsic calibration. Subsequently, for each sampling period, the raw observation data from different modalities are divided into three-dimensional voxel units according to a spatial grid. Within each voxel, physical quantities such as brightness change rate, radar reflection intensity, radial velocity component, sound pressure amplitude, and spectral energy are fused to construct a multiphysics tensor containing spatial coordinate indices and corresponding physical feature vectors. This multiphysics tensor is continuously updated in a time-series format. Furthermore, event-driven encoding is performed on the multiphysics tensor: using the difference in physical features between two adjacent time steps as the triggering basis, when the change in physical features within any voxel exceeds a preset trigger threshold, a corresponding pulse event data packet is generated. This pulse event data packet includes voxel spatial coordinates, the change in physical features, the polarity of the change direction, and a timestamp. When the change is below the trigger threshold, no event is generated, thereby suppressing redundant information output. Finally, all triggered events are organized into a sparse asynchronous pulse stream output in chronological order, achieving a sparse representation of the multiphysics information of the navigation environment.

[0013] Furthermore, the multiphysics tensor of the navigation environment is obtained and converted into a sparse asynchronous pulse stream through event-driven encoding, including: For airborne detection arrays, lightweight complementary detectors are deployed at the front end. Complementary detection rules are defined based on visual high-confidence regions, millimeter-wave high-confidence regions, and acoustic wave dominant regions. The airborne detection array is driven by the complementary detectors to acquire multiphysics tensors as the UAV flies. The multiphysics tensors are event-driven encoded and converted into sparse asynchronous pulse streams based on pulse events. When the coefficient of the asynchronous pulse stream is less than a preset threshold, no response is made.

[0014] Preferably, a lightweight complementary detector assembly is installed at the front end of the UAV for the airborne detection array. The complementary detector includes a visible light / near-infrared visual module, a millimeter-wave radar module, and an array-type acoustic detection module. Time and space alignment is achieved through a unified clock synchronization module and a body coordinate system calibration module. Complementary detection rules are defined according to the perception reliability of different environmental areas. In areas with sufficient light and rich texture, a high-confidence visual zone is set, with visual ranging and image gradient changes as the dominant perception sources. In areas with smoke, rain, fog, or obstruction, a high-confidence millimeter-wave zone is set, with radar echo intensity, distance, and radial velocity information as the dominant factors. In areas with dense complex obstacles at close range or significant airflow disturbances, an acoustic dominant zone is set, with echo delay, sound pressure amplitude, and spectral energy changes as the dominant features. Multimodal complementarity is achieved through dynamic allocation of confidence weights. Subsequently, based on complementary detection rules, the airborne detection array is driven to collect environmental data in real time as the UAV flies. The raw data of each modality is mapped to a unified spatial index according to a three-dimensional voxel grid. Within each voxel, the brightness change rate, echo intensity, radial velocity component, sound pressure change, and their frequency domain features are fused to form a multiphysics tensor containing spatial coordinates, physical feature vectors, and time stamps. Further, event-driven encoding processing is performed on the multiphysics tensor. The change in voxel physical features within adjacent sampling periods is used as the trigger. When the change exceeds a preset trigger threshold, a corresponding pulse event data packet is generated. The pulse event contains voxel spatial coordinates, physical feature change vectors, change polarity, and timestamp, and is output in chronological order to form a sparse asynchronous pulse stream. When the event coefficient of the asynchronous pulse stream within a unit time window is lower than a preset threshold, it is determined that there is no significant environmental change, and no subsequent response is triggered, thereby achieving sparse representation and real-time efficient transmission of navigation environment information.

[0015] Furthermore, each pulse event includes spatial coordinates, physical feature vectors, timestamps, and polarity. The spatial coordinates represent the voxel position where the change occurs, the physical feature vectors represent the amount of change in physical parameters, the polarity represents the direction of change, the timestamp represents a time marker, and the direction of change includes enhancement or weakening.

[0016] Each pulse event generated during event-driven coding is encapsulated in a structured data format. Each pulse event contains at least four types of fields: spatial coordinates, physical feature vector, timestamp, and polarity. The spatial coordinates are represented using a three-dimensional voxel index, including the voxel's discrete raster number (x, y, z) in the body coordinate system or the corresponding spatial center point coordinates, used to accurately identify the spatial location of the physical change. The physical feature vector is the difference result of the changes in physical parameters of each voxel within adjacent sampling periods, including but not limited to brightness change rate, depth change, millimeter-wave echo intensity difference, radial velocity change, sound pressure amplitude change, and spectral energy change, expressed uniformly in vector form. The timestamp is marked using a high-precision system clock or hardware synchronous clock to record the specific time point when the pulse event is triggered, thus supporting asynchronous timing reconstruction and causal analysis. The polarity characterizes the direction of physical parameter change; a positive physical feature change is marked as enhanced polarity, and a negative physical feature change is marked as weakened polarity, thus enabling directional encoding of environmental change trends through the polarity field. The above-described structured impulse event definition enables the multiphysics information of the environment to be expressed in the form of discrete, traceable event units with directional attributes, providing a standardized input data foundation for subsequent spatiotemporal recursive analysis.

[0017] The sparse asynchronous pulse stream is input into the first decision unit to perform a multi-level recursive analysis based on the combination of spatial voxels and physical features to determine the reference control command.

[0018] Specifically, sparse asynchronous pulse streams are input into a pre-constructed first decision-maker in timestamp order. The first decision-maker employs a multi-layer recursive structure to perform spatiotemporal coupling analysis on the joint input units: In the first recursive unit, spatial neighborhood aggregation is performed on voxel events within the same time window, calculating local obstacle density, change gradient, and risk weight; in the second recursive unit, dynamic prediction of event evolution trends within a continuous window is performed based on the time series, extracting the velocity field change rate, potential collision probability, and traversable area evolution trend; in the third recursive unit, the spatial risk characteristics are fused with the UAV's current pose, velocity, and acceleration state to form a state-environment coupled feature representation. Subsequently, the coupled features are converted into control decision vectors through an output mapping function, calculating heading angle correction, velocity adjustment, and attitude compensation, generating reference control commands for driving servo motors, thus providing an initial control benchmark for subsequent constraint optimization and safety control.

[0019] Furthermore, before inputting the sparse asynchronous pulse stream into the first decision-maker, the construction of the first decision-maker includes: A multi-layered liquid machine architecture is adopted to build a first decision engine, which includes an input layer, a hidden layer, and an output layer, and is associated with a symbolic reasoning supervision branch. The input layer takes event-encoded pulses as input, and each neuron corresponds to a spatial voxel-physical feature combination. The hidden layer contains multiple recursively connected liquid layers to perform spatiotemporal feature extraction and prediction. The output layer takes PWM signals that directly drive servo motors as output. The symbolic reasoning supervision branch is constructed by formalizing aviation rules, safety criteria, and physical laws into logical rules.

[0020] Preferably, a multi-layer liquid machine architecture is used to build the overall network structure of the first decision-maker. The first decision-maker includes an input layer, a hidden layer, and an output layer, and a symbolic reasoning supervision branch is associated outside the backbone network. The input layer is used to receive event-encoded pulse streams. Each input neuron corresponds one-to-one with a "spatial voxel-physical feature combination". The voxel index and physical feature vector are mapped to pulse current input through a weight matrix to realize the encoding access of sparse asynchronous events. The hidden layer consists of multiple recursively connected liquid layers. Each liquid layer contains a randomly sparsely connected spiking neuron network. The neurons form dynamic loops through weighted connections. This is used to perform temporal memory and nonlinear mapping of the input pulses, perform spatial neighborhood association analysis, time trend prediction, and risk evolution modeling, thereby extracting high-dimensional spatiotemporal dynamic features. The output layer uses a linear readout unit to map the liquid state vector of the hidden layer into a continuous control quantity, and converts it into a pulse width modulation signal that can directly drive a servo motor through a PWM modulation module. Meanwhile, the symbolic reasoning supervision branch formalizes aviation flight rules, safety boundary criteria, and flight dynamics and physical constraints into a set of logical rules, constructs a rule knowledge base, and performs consistency verification and constraint feedback on the output results of the liquid network during the training phase. It also supervises and optimizes the network parameters through a joint loss function, so that the first decision-maker can meet the requirements of flight safety and physical feasibility while maintaining data-driven characteristics, thereby forming a control decision structure with spatiotemporal recursive decision-making capabilities and rule constraint consistency.

[0021] Furthermore, the sparse asynchronous pulse stream is input into the first decision-maker to perform intention probability distribution decision based on momentum and potential field cross-attention fusion, and to determine the reference control command for the servo motor; wherein, the momentum element includes kinematic inertial features based on position, velocity and acceleration under the UAV trajectory sequence, and the potential field element includes environmental context based on traversable area, typical flight path structure and obstacle distribution.

[0022] After inputting the sparse asynchronous pulse stream into the first decision-maker, a bi-branch feature representation of momentum and potential field elements is constructed based on the spatiotemporal state vector output by the liquid layer, and a fusion decision is made through a cross-attention mechanism. The momentum element is calculated from the UAV's historical trajectory sequence and specifically includes a kinematic inertial feature vector constructed based on position, velocity, and acceleration data. The kinematic inertial feature vector includes at least the velocity modulus, acceleration modulus, heading angle change rate, attitude angle change rate, and an equivalent momentum representation formed by the product of mass and velocity, which is used to reflect the UAV's motion trend and inertial maintenance capability in the current state. The potential field element is constructed based on the environmental multiphysics tensor and voxel risk assessment results, including the probability distribution of passable areas, the constraint field of typical flight path structure, and the spatial distribution density field of obstacles. The obstacle repulsion potential and the target guidance potential are uniformly expressed as an environmental context feature vector by constructing an artificial potential field function or a risk potential function. Subsequently, a cross-attention module is established within the first decision-maker. Momentum elements are used as query vectors, and potential field elements are used as key-value pairs. The matching degree between momentum trends and environmental risks is calculated through attention weights to generate a fusion feature representation. Based on the fusion feature representation, a multi-dimensional control intention set is constructed, and the probability distribution of each control intention is calculated through the Softmax function. The control intentions include at least categories such as heading correction, deceleration avoidance, acceleration crossing, and attitude adjustment. Finally, the control intention with the highest probability is selected and mapped to a continuous control quantity to generate reference control commands for driving the servo motor, thereby realizing intelligent obstacle avoidance decision-making that takes into account both motion inertia and environmental risk constraints.

[0023] A high-dimensional feature space is constructed by embedding tensor kernels, and a manifold-maintaining controller is constructed by combining safety boundaries. The reference control command is subjected to quadratic programming based on the first and second constraints to solve for the optimal control quantity as the target control sequence.

[0024] After receiving the reference control command, the current state variables and the environmental multiphysics tensor are first processed by tensor kernel embedding to map the low-dimensional state-environment coupling features to a high-dimensional feature space. This tensor kernel embedding method defines a multimodal kernel function to combine the UAV's body state tensor (including six degrees of freedom pose, linear velocity, angular velocity, and their first or second derivatives) with the environmental tensor (including obstacle object field values, wind disturbance field, electromagnetic field intensity distribution, etc.) through tensor product, and then maps it to a regenerating kernel Hilbert space to form a linearly separable high-dimensional state representation, thereby enhancing the expressive power of complex nonlinear coupling relationships. Subsequently, an invariant set control framework based on safety boundaries is constructed in the high-dimensional feature space. The intersection of the zero-level sets of multiple safety barrier functions is defined as the feasible safety region, and a manifold-maintaining controller is designed on this safety region to ensure that the system state trajectory always evolves within a safe manifold.

[0025] Based on this, using the reference control command as the nominal control input, a standard quadratic programming optimization model is constructed. Minimizing the control deviation is taken as the objective function, i.e., minimizing the L2 difference between the actual control quantity and the reference control command. Simultaneously, a first constraint and a second constraint are introduced. The first constraint requires that the linear combination of the time derivative of each safety barrier function with respect to the system state and its function value be greater than or equal to zero, ensuring that the system state does not exceed the safety boundary. The second constraint is the physical limit constraint of the actuator, including the maximum thrust of the motor, the thrust rate of change limit, the range of the control surface deflection angle, and the rate of change limit. By solving the quadratic programming problem online, the optimal control quantity that satisfies all safety and physical constraints is obtained. This optimal control quantity is then organized into a target control sequence in time series form for subsequent closed-loop control execution of the UAV obstacle avoidance navigation.

[0026] Furthermore, a high-dimensional feature space is constructed using tensor kernel embedding, including: The multiphysics tensor and the extended tensor are embedded into the high-dimensional regenerative kernel Hilbert space through a kernel approach to form a high-dimensional feature space. The extended tensor contains the six degrees of freedom pose, linear velocity, angular velocity and their higher-order derivatives of the UAV body, and the physical parameter field values ​​corresponding to each obstacle object in the environment. Based on the high-dimensional feature space, an invariant set control framework based on safety boundaries is adopted to construct a manifold-maintaining controller.

[0027] Preferably, a unified state representation is used for the multiphysics tensor and the extended tensor. The multiphysics tensor characterizes the distribution field of environmental physical parameters after voxelization, including field values ​​such as visual intensity variation, radar echo intensity, radial velocity component, sound pressure variation, electromagnetic field intensity, and wind speed curl. The extended tensor characterizes the dynamic state of the UAV, including six-degree-of-freedom pose (position and attitude angle), linear velocity vector, angular velocity vector, and their first or second-order time derivatives, and further includes the index information of the physical parameter field values ​​corresponding to each obstacle object in the environment. Subsequently, a tensor kernel function is constructed, combining the multiphysics tensor and the extended tensor through tensor product. A nonlinear mapping is performed using a polynomial kernel function or a Gaussian radial basis kernel function to map the original coupled state to a regenerating kernel Hilbert space, forming a high-dimensional eigenvector representation. This makes the originally nonlinearly coupled state-environment relationship linearly separable in the high-dimensional space, thereby enhancing the system's ability to express and discriminate complex dynamic environments.

[0028] After obtaining the high-dimensional feature space representation, an invariant set control framework is constructed based on the safety boundary. The safety domain, which is formed by the intersection of the zero-level sets of multiple safety barrier functions, is defined as the set of feasible states of the system. The embedded manifold of this safety domain in the high-dimensional space is used as the basis for control constraints. On this basis, a manifold-preserving controller is designed. By constraining the direction of system state evolution, the high-dimensional state vector always evolves along the tangent space direction of the safety manifold. When approaching the boundary, the state is pulled back to the interior of the safety domain through the control correction term, thereby realizing the safe accessibility maintenance and stable obstacle avoidance control of UAV in complex multiphysics environments.

[0029] Furthermore, a safety boundary is defined, wherein the safety boundary is the intersection of the zero-level sets of multiple safety barrier functions. Each safety barrier function has a one-to-one correspondence with a failure mode or risk dimension, and all safety barrier functions are uniformly expressed as scalar functions in the state space, with their positive values ​​corresponding to a safe state. Specifically, a first safety barrier function is defined based on geometric collision, characterized as the difference between the Euclidean distance between the UAV and the nearest obstacle and a safety threshold. A second safety barrier function is defined based on turbulent disturbance, characterized as the difference between the local spatial wind speed curl and the aircraft's anti-turbulence capability. A third safety barrier function is defined based on electromagnetic radiation hazards, characterized as the difference between the local field strength and the equipment's anti-interference threshold.

[0030] Preferably, the system security boundary is formally defined; the security boundary is composed of the intersection of the zero-level sets of multiple security barrier functions, each security barrier function corresponds one-to-one with a failure mode or risk dimension, and is uniformly expressed as a continuously differentiable scalar function in the system state space. Where x represents an extended state vector containing UAV state variables and environmental parameters, when The corresponding safe state area, when Corresponding to the risk or failure area, through all The intersection of the zero-level sets is used to form the overall safe and feasible domain.

[0031] Specifically, the first safety barrier function is defined based on the geometric collision risk. It is expressed as the Euclidean distance between the center of mass of the UAV and the center of mass of the nearest obstacle. and the preset safe distance threshold The difference, i.e. When the difference is positive, it indicates that the drone maintains a safe distance from the obstacle; a second safety barrier function is defined based on the risk of turbulence disturbance. It is defined as the local spatial wind speed curl or turbulence intensity index. Upper limit of aircraft's anti-turbulence capability The opposite of the difference, i.e. This is used to characterize whether the current airflow disturbance exceeds the aircraft's stability control capability; the third safety barrier function is defined based on electromagnetic radiation hazards. It is defined as the local electromagnetic field intensity. Interference threshold with airborne equipment The opposite of the difference, i.e. This is used to determine whether the electromagnetic environment poses an interference risk to navigation, communication, or control systems. Through unified modeling and intersection constraints of the aforementioned multiple safety barrier functions, a comprehensive safety boundary characterization of multi-dimensional risk factors such as geometric collisions, aerodynamic disturbances, and electromagnetic interference is achieved, providing a rigorous safety constraint foundation for subsequent control optimization.

[0032] Furthermore, solving for the optimal control quantity as the objective control sequence includes: For the reference control command, the optimal control quantity that satisfies the first constraint and the second constraint is solved in the manifold holding controller as the target control sequence; wherein, the first constraint is that the linear combination of the time derivative of each safety barrier function and the safety barrier function itself is non-negative, and the second constraint is the physical limit of the actuator, including the upper limit of motor thrust, the range of control surface deflection angle and its rate of change limit.

[0033] In solving for the optimal control quantity as the target control sequence, a constrained quadratic programming optimization model is constructed within the framework of a manifold-maintaining controller, using the reference control command as the nominal control input. Specifically, an optimization objective function is first established to minimize the control correction, minimizing the sum of squared deviations between the actual control quantity and the reference control command, thereby ensuring control smoothness while preserving the original decision intent as much as possible. Subsequently, a first constraint and a second constraint are introduced to form a constraint set. The first constraint constructs a control barrier function constraint for each safety barrier function, requiring that its time derivative and the linear combination of the function itself satisfy a value greater than or equal to 0, thus ensuring that the system state remains within the safety domain or converges towards the safety domain under control. The second constraint is a physical limit constraint on the actuator, including that the maximum thrust of the motor does not exceed the preset thrust upper limit, the thrust change rate does not exceed the set slope threshold, the control surface deflection angle is within the allowable angle range, and the control surface angular velocity meets the change rate limit, etc., to ensure that the control command can be realized within the actual hardware capabilities. By solving the constrained quadratic programming problem in real time online, the optimal control quantity that simultaneously satisfies the safety barrier constraint and the physical limit constraint is obtained, and output as the target control sequence in the form of a time series, which is used to drive the UAV to perform safe and stable obstacle avoidance navigation control.

[0034] Based on the target control sequence, perform obstacle avoidance navigation control for the UAV.

[0035] After obtaining the target control sequence, the sequence is divided into discrete control command frames according to the time step and sent to the flight controller via the flight control bus. The flight controller analyzes the thrust commands, attitude angle corrections, and angular velocity adjustments in the target control sequence. The ESC module converts the thrust control quantity into the corresponding motor drive current signal, and the servo drive module converts the attitude correction quantity into differential control quantities for control surface deflection angles or rotor speeds, thereby driving the coordinated action of each actuator. During execution, the flight controller collects feedback status data in real time based on the inertial measurement unit, barometer, and GNSS module, performs a closed-loop comparison with the target control sequence, calculates attitude and position errors, and performs error compensation correction through inner and outer loop control structures. When a new environmental pulse event is detected or the safety barrier function approaches its boundary, the control input is updated in real time and the optimization solution is re-executed, thus forming a perception-decision-control closed-loop linkage mechanism. Through the above execution process, the UAV can complete heading adjustment, speed control, and attitude stabilization while meeting safety and physical constraints, achieving continuous, smooth, and reliable autonomous obstacle avoidance navigation control.

[0036] Furthermore, the UAV obstacle avoidance navigation control is performed according to the target control sequence; the navigation data link is recorded and abnormal obstacle avoidance features are mined, wherein the abnormal obstacle avoidance features satisfy a preset generalization degree; and the flow shape maintaining controller is updated and learned according to the abnormal obstacle avoidance features.

[0037] While executing UAV obstacle avoidance navigation control according to the target control sequence, a complete navigation data chain recording mechanism is constructed to synchronously store the perceived input data, reference control commands, optimal control quantities, actuator feedback status, and real-time values ​​of the safety barrier function, forming a continuous navigation data chain according to timestamps. After the flight mission is completed or within a preset time window, the navigation data chain is analyzed offline or semi-online. An anomaly detection model is constructed to statistically analyze the number of control oscillations, control correction amplitude, barrier function approximation frequency, trajectory deviation, and energy consumption fluctuations during the obstacle avoidance process, extracting feature vectors characterizing abnormal obstacle avoidance behavior. When the feature vector exhibits a recurring pattern that is stable across different environmental scenarios, it is determined to meet a preset generalization threshold, thus identifying it as an abnormal obstacle avoidance feature. Subsequently, the abnormal obstacle avoidance features are input into the parameter update module of the manifold maintaining controller. By adaptively adjusting the weight coefficients of the safety barrier function, the convergence parameters of the control barrier function, and the weights of the optimization target, or by using incremental learning to update the kernel embedding parameters and constraint adjustment factors, the controller can improve its adaptability to complex environmental changes while maintaining the original safety structure, thereby realizing the continuous optimization and autonomous learning update of the manifold maintaining controller.

[0038] In summary, the embodiments of this application have at least the following technical effects: First, the multiphysics tensor of the flight environment is acquired through complementary detectors at the airborne front end, and then converted into a sparse asynchronous pulse stream through event-driven encoding. Next, the sparse asynchronous pulse stream is input into the first decision-maker, which performs a multi-layer recursive analysis based on spatial voxel-physical feature combinations to determine the reference control command. Then, a high-dimensional feature space is constructed by embedding tensor kernels, and a manifold-maintaining controller is built by combining it with safety boundaries. A quadratic programming problem based on the first and second constraints is performed on the reference control command to solve for the optimal control quantity as the target control sequence. Finally, obstacle avoidance navigation control is performed on the UAV according to the target control sequence. This solves the technical problems of insufficient environmental perception, delayed obstacle avoidance decision-making, and insufficient coupling of control constraints in existing technologies for manned UAVs in complex flight environments, resulting in poor obstacle avoidance reliability. It achieves the technical effect of improving the reliability of autonomous obstacle avoidance for manned UAVs.

[0039] Example 2 is based on the same inventive concept as the manned UAV autonomous obstacle avoidance control method based on flight environment perception in the previous examples, such as... Figure 2 As shown, this application provides an autonomous obstacle avoidance control system for manned unmanned aerial vehicles based on flight environment perception, wherein the system includes: Data acquisition module 11: Acquires the multiphysics tensor of the flight environment through complementary detectors at the airborne front end, and converts it into a sparse asynchronous pulse stream through event-driven encoding; Analysis module 12: Inputs the sparse asynchronous pulse stream into the first decision unit, performs multi-level recursive analysis based on spatial voxel-physical feature combination, and determines the reference control command; Optimization module 13: Constructs a high-dimensional feature space by embedding tensor kernels, constructs a manifold-maintaining controller by combining safety boundaries, performs quadratic programming based on the first and second constraints on the reference control command, and solves for the optimal control quantity as the target control sequence; Control module 14: Performs obstacle avoidance flight control of the UAV according to the target control sequence.

[0040] Furthermore, the data acquisition module 11 is used to perform the following methods: For airborne detection arrays, lightweight complementary detectors are deployed at the front end. Complementary detection rules are defined based on visual high-confidence regions, millimeter-wave high-confidence regions, and acoustic wave dominant regions. The airborne detection array is driven by the complementary detectors to acquire multiphysics tensors as the UAV flies. The multiphysics tensors are event-driven encoded and converted into sparse asynchronous pulse streams based on pulse events. When the coefficient of the asynchronous pulse stream is less than a preset threshold, no response is made.

[0041] Furthermore, the data acquisition module 11 is used to perform the following methods: Each pulse event includes spatial coordinates, physical feature vectors, timestamps, and polarity. The spatial coordinates represent the voxel position where the change occurs, the physical feature vectors represent the amount of change in physical parameters, the polarity represents the direction of change, and the timestamp represents a time marker. The direction of change includes either enhancement or weakening.

[0042] Furthermore, the analysis module 12 is used to perform the following methods: A multi-layered liquid machine architecture is adopted to build a first decision engine, which includes an input layer, a hidden layer, and an output layer, and is associated with a symbolic reasoning supervision branch. The input layer takes event-encoded pulses as input, and each neuron corresponds to a spatial voxel-physical feature combination. The hidden layer contains multiple recursively connected liquid layers to perform spatiotemporal feature extraction and prediction. The output layer takes PWM signals that directly drive servo motors as output. The symbolic reasoning supervision branch is constructed by formalizing aviation rules, safety criteria, and physical laws into logical rules.

[0043] Furthermore, the analysis module 12 is used to perform the following methods: The sparse asynchronous pulse stream is input into the first decision-maker to perform intention probability distribution decision based on the cross-attention fusion of momentum and potential field, and to determine the reference control command for the servo motor; wherein, the momentum element includes the kinematic inertial characteristics based on position, velocity and acceleration under the UAV trajectory sequence, and the potential field element includes the environmental context based on traversable area, typical flight path structure and obstacle distribution.

[0044] Furthermore, the optimization module 13 is used to perform the following method: The multiphysics tensor and the extended tensor are embedded into the high-dimensional regenerative kernel Hilbert space through a kernel approach to form a high-dimensional feature space. The extended tensor contains the six degrees of freedom pose, linear velocity, angular velocity and their higher-order derivatives of the UAV body, and the physical parameter field values ​​corresponding to each obstacle object in the environment. Based on the high-dimensional feature space, an invariant set control framework based on safety boundaries is adopted to construct a manifold-maintaining controller.

[0045] Furthermore, the optimization module 13 is used to perform the following method: A safety boundary is defined, wherein the safety boundary is the intersection of the zero-level sets of multiple safety barrier functions. Each safety barrier function has a one-to-one correspondence with a failure mode or risk dimension, and all safety barrier functions are uniformly expressed as scalar functions in the state space, with their positive values ​​corresponding to a safe state. Specifically, a first safety barrier function is defined based on geometric collision, characterized as the difference between the Euclidean distance between the UAV and the nearest obstacle and a safety threshold. A second safety barrier function is defined based on turbulent disturbance, characterized as the difference between the local spatial wind speed curl and the aircraft's anti-turbulence capability. A third safety barrier function is defined based on electromagnetic radiation hazards, characterized as the difference between the local field strength and the equipment's anti-interference threshold.

[0046] Furthermore, the optimization module 13 is used to perform the following method: For the reference control command, the optimal control quantity that satisfies the first constraint and the second constraint is solved in the manifold holding controller as the target control sequence; wherein, the first constraint is that the linear combination of the time derivative of each safety barrier function and the safety barrier function itself is non-negative, and the second constraint is the physical limit of the actuator, including the upper limit of motor thrust, the range of control surface deflection angle and its rate of change limit.

[0047] Furthermore, the control module 14 is used to perform the following methods: The UAV performs obstacle avoidance navigation control according to the target control sequence; the navigation data link is recorded and abnormal obstacle avoidance features are mined, wherein the abnormal obstacle avoidance features satisfy a preset generalization degree; the flow shape maintaining controller is updated and learned according to the abnormal obstacle avoidance features.

[0048] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A manned unmanned aerial vehicle (UAV) autonomous obstacle avoidance control method based on flight environment perception, characterized in that, The method includes: The multiphysics tensor of the flight environment is obtained by complementary detectors at the airborne front end, and then converted into a sparse asynchronous pulse stream through event-driven encoding. The sparse asynchronous pulse stream is input into the first decision unit to perform a multi-level recursive analysis based on the combination of spatial voxels and physical features to determine the reference control command; A high-dimensional feature space is constructed by embedding tensor kernels, and a manifold-maintaining controller is constructed by combining safety boundaries. The reference control command is subjected to quadratic programming based on the first and second constraints to solve for the optimal control quantity as the target control sequence. Based on the target control sequence, perform obstacle avoidance navigation control for the UAV.

2. The autonomous obstacle avoidance control method for manned unmanned aerial vehicles based on flight environment perception as described in claim 1, characterized in that, The multiphysics tensor of the navigation environment is obtained and converted into a sparse asynchronous pulse stream through event-driven encoding, including: For airborne detection arrays, lightweight complementary detectors are deployed at the front end, with complementary detection rules defined based on visual high-confidence region, millimeter-wave high-confidence region and acoustic wave dominant region; The airborne detection array is driven by the complementary detector and travels with the UAV to detect and acquire multiphysics tensors. The multiphysics tensor is event-driven encoded and converted into a sparse asynchronous pulse stream based on pulse events. When the coefficient asynchronous pulse stream is less than a preset threshold, no response is made.

3. The autonomous obstacle avoidance control method for manned unmanned aerial vehicles based on flight environment perception as described in claim 2, characterized in that, Each pulse event includes spatial coordinates, physical feature vectors, timestamps, and polarity. The spatial coordinates represent the voxel position where the change occurs, the physical feature vectors represent the amount of change in physical parameters, the polarity represents the direction of change, and the timestamp represents a time marker. The direction of change includes either enhancement or weakening.

4. The autonomous obstacle avoidance control method for manned unmanned aerial vehicles based on flight environment perception as described in claim 1, characterized in that, Before inputting the sparse asynchronous pulse stream into the first decision-maker, the construction of the first decision-maker includes: A multi-layer liquid machine architecture is adopted to build the first decision engine, which includes an input layer, a hidden layer and an output layer, and an associated symbolic reasoning supervision branch. The input layer takes event-encoded pulses as input, and each neuron corresponds to a spatial voxel-physical feature combination. The hidden layer contains multiple recursively connected liquid layers that perform spatiotemporal feature extraction and prediction. The output layer uses the PWM signal that directly drives the servo motor as the output quantity; The symbolic reasoning supervision branch is constructed by formalizing aviation rules, safety criteria, and physical laws into logical rules.

5. The autonomous obstacle avoidance control method for manned unmanned aerial vehicles based on flight environment perception as described in claim 4, characterized in that, The sparse asynchronous pulse stream is input into the first decision-maker to perform intention probability distribution decision based on the cross-attention fusion of momentum and potential field, and to determine the reference control command for the servo motor. Among them, the momentum element includes the kinematic inertial characteristics based on position, velocity and acceleration under the UAV trajectory sequence, and the potential field element includes the environmental context based on traversable areas, typical flight path structures and obstacle distribution.

6. The autonomous obstacle avoidance control method for manned unmanned aerial vehicles based on flight environment perception as described in claim 1, characterized in that, A high-dimensional feature space is constructed using tensor kernel embedding, including: The multiphysics tensor and the extended tensor are embedded into the high-dimensional regenerative kernel Hilbert space through a kernel method as a high-dimensional feature space. The extended tensor contains the six degrees of freedom pose, linear velocity, angular velocity and their higher-order derivatives of the UAV body, and the physical parameter field values ​​corresponding to each obstacle object in the environment. For the high-dimensional feature space, a manifold-preserving controller is constructed using an invariant set control framework based on safety boundaries.

7. The autonomous obstacle avoidance control method for manned unmanned aerial vehicles based on flight environment perception as described in claim 6, characterized in that, Define a safety boundary, wherein the safety boundary is the intersection of the zero-level sets of multiple safety barrier functions, wherein there is a one-to-one correspondence between the safety barrier functions and the failure modes or risk dimensions, and the multiple safety barrier functions are uniformly expressed as scalar functions in the state space, with their positive value regions corresponding to safe states; The first safety barrier function is defined according to geometric collision, and the first safety barrier function is characterized by the difference between the Euclidean distance between the UAV and the nearest obstacle and the safety threshold. The second safety barrier function is defined based on turbulent disturbances. The second safety barrier function is defined as the difference between the local space wind speed curl and the aircraft's anti-turbulence capability. The third safety barrier function is defined according to the definition of electromagnetic radiation hazards. The third safety barrier function is defined as the difference between the local field strength and the equipment anti-interference threshold.

8. The autonomous obstacle avoidance control method for manned unmanned aerial vehicles based on flight environment perception as described in claim 7, characterized in that, Solving for the optimal control quantity as the target control sequence includes: For the reference control command, the optimal control quantity that satisfies the first constraint and the second constraint is solved in the manifold holding controller, and is used as the target control sequence; The first constraint is that the linear combination of the time derivative of each safety barrier function and the safety barrier function itself is non-negative. The second constraint is the physical limits of the actuator, including the upper limit of motor thrust, the range of control surface deflection angle and its rate of change.

9. The autonomous obstacle avoidance control method for manned unmanned aerial vehicles based on flight environment perception as described in claim 1, characterized in that, Perform obstacle avoidance navigation control for the UAV based on the target control sequence; Record navigation data links and mine abnormal obstacle avoidance features, wherein the abnormal obstacle avoidance features satisfy a preset generalization degree; The flow shape preservation controller is updated and learned based on the abnormal obstacle avoidance characteristics.

10. An autonomous obstacle avoidance control system for manned unmanned aerial vehicles based on flight environment perception, characterized in that: The system is used to implement the autonomous obstacle avoidance control method for manned unmanned aerial vehicles based on flight environment perception as described in any one of claims 1-9, the system comprising: Data acquisition module: Acquires multiphysics tensors of the flight environment through complementary detectors at the airborne front end, and converts them into sparse asynchronous pulse streams through event-driven encoding; Analysis module: Inputs the sparse asynchronous pulse stream into the first decision unit, performs multi-level recursive analysis based on the combination of spatial voxels and physical features, and determines the reference control command; Optimization module: A high-dimensional feature space is constructed by embedding tensor kernels, and a manifold-maintaining controller is constructed by combining safety boundaries. The reference control command is subjected to quadratic programming based on the first and second constraints to solve for the optimal control quantity as the target control sequence. Control module: Performs obstacle avoidance navigation control for the UAV according to the target control sequence.

Images