Heterogeneous multi-vehicle system collision-free cooperative control method based on predictor-based neural network observer and path guidance

Abstract

The application discloses a collision-free cooperative control method of a heterogeneous multi-vehicle system based on a predictor-based neural network observer and path guidance, comprising the following steps: establishing a dynamic model of the heterogeneous multi-vehicle system to provide a model support for layered control; constructing an artificial potential field function and integrating a safety constraint to design a collision avoidance and communication connectivity related potential field function; defining a virtual root leader motion rule according to the constraint to provide a reference path for the formation; designing a predictor-based neural network observer (PNO) to provide uncertainty compensation for the controller; and combining the reference path, the potential field constraint and the predictor-based neural network observer compensation to build a layered control architecture and realize the collision-free cooperative control of the heterogeneous multi-vehicle system. The method can accurately estimate the system uncertainty, significantly reduce the calculation cost and effectively guarantee the safe and stable operation of the system by relying on the path guidance and the PNO design.

Description

Technical Field

[0001] This invention belongs to the field of heterogeneous multi-vehicle system cooperative control technology, and relates to a collision-free cooperative control method for heterogeneous multi-vehicle systems based on a predictor-based neural network observer and path guidance. Background Technology

[0002] Cooperative control of heterogeneous multi-vehicle systems is a core research direction in fields such as intelligent transportation and unmanned systems, and is widely used in critical mission scenarios such as logistics distribution, emergency rescue, and environmental monitoring. In complex real-world environments with dense unknown obstacles and limited communication links, heterogeneous multi-vehicle systems (such as aerial drones and ground vehicles) must simultaneously meet multiple constraints, including path following, formation maintenance, collision-free operation, and communication connectivity maintenance. Their safe and coordinated scheduling has become a technical challenge that urgently needs to be overcome.

[0003] Existing multi-vehicle cooperative control schemes are mostly based on traditional trajectory tracking frameworks, which require pre-setting reference trajectories that change over time and demanding that all vehicles arrive at designated locations at specific times. This results in strong spatiotemporal coupling, significantly increasing the control difficulty and system fault tolerance risks in complex environments. Although path-following frameworks have gradually gained attention due to their characteristic of only constraining spatial paths and not forcing time synchronization, existing research still has significant limitations: First, it focuses on multi-focused two-dimensional planar scenarios and lacks cooperative schemes for heterogeneous systems that address the three-dimensional translation of UAVs and the two-dimensional motion of ground vehicles, making it difficult to adapt to the needs of integrated air-to-ground operations; second, it does not fully consider the dynamic differences of heterogeneous carriers and does not effectively integrate collision-free obstacle avoidance and communication connectivity assurance mechanisms, making it unsuitable for direct application in real-world scenarios with dense obstacles and limited communication.

[0004] Meanwhile, heterogeneous multi-vehicle systems inevitably face uncertainties such as nonlinear dynamics and unknown external disturbances during operation. Traditional observers (such as extended state observers and fuzzy observers) and conventional neural approximation methods have significant shortcomings: the former is prone to initial oscillations and has poor transient estimation performance; the latter requires updating all weight parameters based on the number of neural network nodes, resulting in a heavy online computational burden, and the control-estimation loop is strongly coupled, making it difficult to adapt to high-frequency disturbance scenarios. In addition, existing heterogeneous system control architectures mostly adopt a unified design approach, failing to specifically match the dynamic characteristics of different types of vehicles, further restricting the accuracy of cooperative control and system stability.

[0005] In summary, traditional control methods generally suffer from insufficient control accuracy, poor robustness, and high computational complexity when addressing the integrated requirements of spatiotemporal decoupling, heterogeneous adaptation, uncertainty compensation, collision-free operation, and connectivity assurance in heterogeneous multi-vehicle systems. To ensure that heterogeneous multi-vehicle systems can safely and efficiently complete collaborative tasks in complex and demanding environments, it is urgent to propose a novel collaborative control method that balances heterogeneous adaptability, uncertainty handling capabilities, and safety constraint satisfaction. This would improve the theoretical framework of multi-agent safe collaborative control and promote the engineering application of heterogeneous multi-vehicle systems. Summary of the Invention

[0006] The purpose of this invention is to provide a collision-free cooperative control method for heterogeneous multi-vehicle systems based on a predictor-based neural network observer and path guidance, in order to solve the problems of low system control accuracy, poor robustness, and computational complexity in existing heterogeneous multi-vehicle systems.

[0007] This invention provides a collision-free cooperative control method for heterogeneous multi-vehicle systems based on a predictor-driven neural network observer and path guidance, comprising:

[0008] Step 1: Establish a dynamic model of the heterogeneous multi-vehicle system to provide model support for hierarchical control;

[0009] Step 2: Construct an artificial potential field function, incorporating potential field functions related to collision avoidance and communication connectivity in the design of safety constraints;

[0010] Step 3: Define the virtual root guide's motion rules based on the constraints to provide a reference path for formation;

[0011] Step 4: Design a predictor-based neural network observer PNO to provide uncertainty compensation for the controller;

[0012] Step 5: Combine reference path, potential field constraints and predictor-based neural network observer compensation to build a hierarchical control architecture to achieve collision-free cooperative control of heterogeneous multi-vehicle systems.

[0013] This invention presents a collision-free cooperative control method for heterogeneous multi-vehicle systems based on a predictor-based neural network observer and path guidance. It proposes establishing a time-independent implicit spatial path framework to reduce the control difficulty of heterogeneous systems and improve cooperative flexibility. A hierarchical control architecture is proposed to meet the requirements of three-dimensional spatial cooperation and accurately adapt to the motion characteristics of heterogeneous vehicles, filling a gap in related technologies. The innovative predictor-based neural network observer (PNO) updates only the neural weight norm, significantly reducing the online computational burden and avoiding initial oscillation problems. It can efficiently estimate and compensate for system nonlinearity and unknown disturbances, improving the real-time performance, accuracy, and transient performance of the control. An artificial potential field mechanism is incorporated, achieving collision-free operation through a repulsive potential field and maintaining communication connectivity through a gravitational potential field. Combined with the disturbance resistance compensation of the PNO, it achieves integrated protection of path following, formation maintenance, collision-free operation, and communication connectivity, making the system more stable and reliable in environments with dense obstacles and limited communication. Attached Figure Description

[0014] Figure 1 This is a flowchart of the collision-free cooperative control method for heterogeneous multi-vehicle systems based on predictor neural network observers and path guidance according to the present invention.

[0015] Figure 2 This is a platooning control scenario where a heterogeneous multi-vehicle system follows a path in an environment with unknown obstacles. Detailed Implementation

[0016] like Figure 1 As shown, the present invention provides a collision-free cooperative control method for heterogeneous multi-vehicle systems based on a predictor-driven neural network observer and path guidance, comprising:

[0017] Step 1: Establish a dynamic model of the heterogeneous multi-vehicle system to provide model support for hierarchical control.

[0018] To address the dynamic differences in heterogeneous multi-vehicle systems, a "leader-follower" architecture is constructed:

[0019] like Figure 2 As shown, the heterogeneous multi-vehicle system consists of four quadcopter drones and three ground vehicles. The quadcopter drones fly along a pre-defined route, maintaining formation coordination through the information exchange link shown by the dotted line. The ground vehicles do not need to plan their own paths; they only need to move collaboratively within the convex hull (safe zone) formed by the drone formation. When faced with obstacles such as buildings, trees, and rocks in the scene, the drones and ground vehicles will autonomously avoid them, while all intelligent agents maintain communication connectivity through the information exchange link.

[0020] This architecture leverages the advantages of drones—wide field of view and high maneuverability—while also utilizing the characteristics of ground vehicles—high payload and strong execution capabilities—to complete tasks more efficiently and safely in complex environments.

[0021] Guide layer: Forms and maintains the desired formation along a preset three-dimensional geometric path, while avoiding obstacles and collisions with each other.

[0022] Follower layer: Through distributed neighbor interactions, it converges into the convex hull formed by the four rotors of the leader, achieving a "containment" effect without collisions. Step 1 is as follows:

[0023] Step 1.1: The facilitator layer consists of... The system consists of N aerial quadrotor UAVs, each possessing dynamic characteristics of coupled translation and rotation. Assuming the attitude controller can support precise attitude tracking of the quadrotors, a three-dimensional dynamic model of aerial quadrotor i is constructed:

[0024]

[0025] in, Let be the three-dimensional position and state vector of quadrotor i; Let i be the three-dimensional velocity vector of the quadrotor i; For the control input of the quadcopter i; The model represents the lumped uncertainty of the quadrotor i; it ignores attitude control details and focuses only on position loop control.

[0026] Step 1.2: Follower layer from A two-dimensional dynamic model of the follower layer is constructed, consisting of M ground vehicles:

[0027]

[0028] in, It is the planar position vector of the ground vehicle k; Its linear velocity, Its angular velocity; Its heading angle; Its inertial parameters; Its wheel radius; These are its linear velocity control input and angular velocity control input, respectively; External disturbances; This represents the unmodeled nonlinear term.

[0029] Step 2: Construct an artificial potential field function, incorporating the potential field function related to collision avoidance and communication connectivity in the safety constraint design, specifically as follows:

[0030] Step 2.1: Design of the mutual collision avoidance potential field function: The potential field function takes effect when the distance is between the detection radius and the collision critical distance. The expression is:

[0031] Between quadrotor i and quadrotor j:

[0032]

[0033] in, Let be the collision avoidance potential field function between quadrotor i and quadrotor j; , This is the distance between the two quadcopters; This is the extreme collision avoidance distance between quadcopters; if it is less than this value, there is a risk of collision. The collision detection radius for the quadcopter; the function in The internal monotonous decrease ensures no collisions between the four rotors.

[0034] Between ground vehicle k and ground vehicle m:

[0035]

[0036] in, Let be the collision avoidance potential field function between ground vehicle k and ground vehicle m; , This is the distance between two vehicles on the ground. This represents the extreme collision avoidance distance between vehicles on the ground; if the distance is less than this value, there is a risk of collision. The radius of the ground vehicle collision detection; the function is in The internal structure decreases monotonically to ensure no collisions between vehicles on the ground.

[0037] Step 2.2: Design of the potential field function for collision avoidance with static obstacles: The potential field function takes effect when the distance is between the detection radius and the collision critical distance. The expression is:

[0038] Quadrotor i and static obstacles between:

[0039]

[0040] in, Quadrotor i and static obstacles The collision avoidance potential field function between them; , Obstacles Location, The distance between the quadcopter and the obstacle; This is the minimum collision avoidance distance between the quadcopter and an obstacle; if the distance is less than this value, there is a risk of collision. The function represents the obstacle detection range of the quadcopter. The internal monotonous decrease ensures that the quadcopter avoids static obstacles in the environment.

[0041] Ground vehicles and static obstacles :

[0042]

[0043] in, For ground vehicles and static obstacles The collision avoidance potential field function between them; , Obstacles Location, The distance between ground vehicles and obstacles; This is the minimum avoidance distance between a vehicle and an obstacle on the ground; if the distance is less than this value, there is a risk of collision. The function is the detection range of ground vehicles for obstacles. The internal monotonically decreasing pattern ensures that ground vehicles avoid static obstacles in the environment.

[0044] Step 2.3: Design of the communication connectivity preservation potential field function: The potential field function takes effect when the distance is between the maximum acceptable communication distance and the minimum distance at which the communication mechanism is activated. The expression is:

[0045] Between quadrotor i and quadrotor j:

[0046]

[0047] in, To maintain the potential field function for communication connectivity between quadrotor i and quadrotor j; This is the maximum acceptable communication distance between quadcopters; communication will be interrupted if this value is exceeded. The minimum distance required to activate the communication mechanism; the function is in Internally, the attraction force ensures communication connectivity between the four rotors.

[0048] Between ground vehicle k and ground vehicle m:

[0049]

[0050] in, To maintain the potential field function for the communication connectivity between ground vehicle k and ground vehicle m; This is the maximum acceptable communication distance between ground vehicles; communication will be interrupted if this value is exceeded. The minimum distance required to activate the communication mechanism; the function is in Internally, the attraction force ensures the communication connectivity of ground vehicles.

[0051] Step 3: Define the virtual root guide's motion rules based on the constraints to provide a reference path for the formation, specifically:

[0052] Step 3.1: Definition of the Virtual Root Leader Dynamics Model:

[0053]

[0054] in, The three-dimensional position vector of the virtual root guide; The three-dimensional velocity vector of the virtual root leader.

[0055] Step 3.2: Definition of the spatial constraint function for the reference path:

[0056] The reference path of the virtual root guide is determined by the intersection of two smooth implicit surfaces, and the path function is defined as follows:

[0057]

[0058] in, and for Smooth functions; and It satisfies the conditions of being bounded and not parallel.

[0059] Step 3.3: Define the virtual root bootstrap constraint rules:

[0060] Path preservation constraints: 0 and .

[0061] Uniform motion constraint: ,in The preset constant tangential velocity.

[0062] Ensure the virtual root bootloader always travels along the reference path at a constant speed The velocity vector of the virtual root leader is always related to the motion. and The intersecting curves of the two surfaces are tangent.

[0063] Step 3.4: Define the velocity vector of the virtual root bootstrapper:

[0064]

[0065] in," " represents the cross product of vectors, , Path functions , The gradient; the direction of the cross product result is the tangent of the reference path.

[0066] Step 4: Design a predictor-based neural network observer PNO to provide uncertainty compensation for the controller, specifically:

[0067] Step 4.1: Clarify the composition of system uncertainty: Define the types of uncertainty that need to be compensated to provide a clear basis for the observation target of PNO.

[0068] Quadrotor layer uncertainty definition: The lumped uncertainty of a quadrotor originates from external disturbances and state-dependent damping. Combined with its normalized dynamic model, it is defined as follows:

[0069]

[0070] in, For the lumped uncertainty of quadrotor i The original perturbation of the quadrotor i in the r direction; Let i be the mass of the quadrotor.

[0071] Uncertainty definition for ground vehicle layer: The uncertainty of ground vehicles includes external disturbances and unmodeled nonlinear terms. Based on its dynamic model, two key types of uncertainty are defined:

[0072]

[0073] in, For the linear velocity-related uncertainties of ground vehicle k, The uncertainty is related to the angular velocity of the ground vehicle k.

[0074] Step 4.2: System uncertainties based on neural network approximation:

[0075] Quadrotor layer uncertainty approximation:

[0076]

[0077] in, The ideal neural weight vector; For the basis function vector, Input for the observer; This represents the bounded approximation error.

[0078] Approximate uncertainty for the ground vehicle layer:

[0079]

[0080] in, , The ideal neural weight vector; , For the basis function vector, , Input for the observer; This represents the bounded approximation error.

[0081] Step 4.3: Construct the PNO observation model.

[0082] Construction of the quadrotor layer PNO model: through weight vector norm Simplify learning:

[0083]

[0084] in, for The estimated value; for The estimated value; Let i be the actual velocity of the quadrotor in the r direction. The predicted velocity of quadrotor i in the r direction The velocity prediction error of the quadrotor i in the r direction is denoted as .

[0085] Ground vehicle layer PNO model construction: through weight vector norm Simplify learning:

[0086]

[0087] in, , They are respectively , The estimated value; , They are respectively , The estimated value; Let k be the actual linear velocity of the ground vehicle. Let k be the predicted linear velocity of the ground vehicle. The linear velocity prediction error of ground vehicle k; Let k be the actual angular velocity of the ground vehicle. Let k be the actual angular velocity of the ground vehicle. The angular velocity prediction error for ground vehicle k is given.

[0088] Step 4.4: Design of the State Predictor and Adaptive Learning Law.

[0089] Design of Quadrotor Layer State Predictor and Learning Law:

[0090] State predictor:

[0091] Weight update law:

[0092] in, for The predicted value, For the control input of the quadcopter i, This is the sedation coefficient; For learning gain; This is the forgetting factor; simultaneously, the initial prediction error is set to 0, i.e. This significantly improves transient estimation performance.

[0093] Ground vehicle layer state predictor and learning law design:

[0094] State predictor:

[0095]

[0096] Adaptive learning law:

[0097]

[0098] in, , for , The predicted value, These are the linear velocity control input and angular velocity control input for the ground vehicle k, respectively. This is the sedation coefficient; For learning gain; It is a forgetting factor.

[0099] The uncertainty estimate for quadrotor i is output by PNO; This is the uncertainty estimate of the k-line velocity of ground vehicles output by PNO. This is the uncertainty estimate of the k-angular velocity of the ground vehicle output by PNO.

[0100] Step 5: Combine reference path, potential field constraints, and predictor-based neural network observer compensation to build a hierarchical control architecture, achieving collision-free cooperative control of heterogeneous multi-vehicle systems, specifically:

[0101] Artificial potential field function (step 2): Collision avoidance potential field including quadcopter and ground vehicle ( ), connectivity potential field ( ) and obstacle collision avoidance potential field ( ), providing security constraints for the two-layer system.

[0102] Reference path module (step 3): Implicit path functions of the virtual root bootstrap Based on this, the velocity vector of the virtual root guide is generated through gradient cross product to provide a spatial motion reference for the guide layer.

[0103] PNO Uncertainty Compensation Module (Step 4): Provides uncertainty estimates for the two-layer controller through weight norm updates driven by prediction error. (quadcopter) (Ground vehicles) ensure robustness. The three elements achieve "safety-tracking-robustness" synergy through controller fusion.

[0104] Step 5.1: Design of the Path Tracking Formation Controller for the Guide Layer:

[0105] Step 5.1.1: Definition of fusion error:

[0106]

[0107] in, This refers to the formation error of the quadcopter. Let i be the set of neighbors of the quadrotor i; Let be the relative position vector between quadrotor i and quadrotor j; Let i be the relative position vector between the quadrotor i and the virtual root leader; Let be the desired relative position vector between quadrotor i and quadrotor j. Let i be the expected relative position vector between the quadrotor i and the virtual root leader; The communication weights between quadrotor i and quadrotor j; For virtual root bootstrap communication weights; For safety variables; The number of obstacles detected by the quadcopter i.

[0108] Step 5.1.2: Target velocity generation guided by the reference path:

[0109]

[0110] in, Let be the target velocity vector of quadrotor i; For communication in-degree; This indicates that quadrotor i can receive information from quadrotor j, and vice versa. 0; For virtual root bootstrap communication weights; This indicates that the quadcopter i directly obtains information from the virtual guide, and vice versa. 0; This is the sedation coefficient; Let be the velocity vector of quadrotor j.

[0111] The communication topology among N quadcopters is a directed graph. express; It is a set of nodes for a quadcopter; It is the set of edges; define the Laplace matrix. ; Describe the scale of a communication network consisting of N quadcopters; define the set This represents the communication network between the virtual root leader and the quadcopter. It is assumed that at least one quadcopter can directly obtain information from the virtual root leader, meaning the communication graph of the leader layer contains a spanning tree rooted at the virtual root leader.

[0112] Step 5.1.3: Construction of the complete path tracking formation controller:

[0113]

[0114] in, It is the control input for the quadcopter i; These are preset parameters; This is the path tracking constraint matrix.

[0115] in, , , , ; This is a nonlinear term derived from reference path tracing.

[0116]

[0117]

[0118] ;

[0119] The projection of the target velocity along the path tangent; This is the error stabilization term. Where:

[0120]

[0121]

[0122]

[0123] They are quadcopters i pairs Path tracking error; These are the path tracking errors of the virtual root bootloader; They are respectively The damping constant of the path; The velocity error stabilization coefficient, For speed tracking error; It is a saturation function. Its expression is:

[0124] in, This is the upper limit of the repulsive / attractive force, used to ensure collision avoidance and maintain connectivity.

[0125] Step 5.2: Design of a distributed containment controller in the follower layer:

[0126] Step 5.2.1: Definition of fusion error:

[0127]

[0128]

[0129]

[0130] in, This includes the error for ground vehicles; Let K be the set of follower neighbors of ground vehicle K, which includes other ground vehicles that communicate with K. The leader neighbor set of ground vehicle k contains the quadcopter that communicates with k; The communication weights for ground vehicle k and followers m and quadrotor i are respectively, and the method of value selection is the same as... resemblance; Let m be the position vector of the follower. This is the planar position projection of quadcopter i, that is, ignoring the z-direction position of the quadcopter and only retaining the planar coordinates of the motion along the reference path, which is used to provide a convex hull reference for ground vehicles; Let k be the number of obstacles detected by ground vehicle k. This is a safety variable. Assume that at least one ground vehicle can receive information from at least one quadcopter.

[0131] Step 5.2.2: Virtual Control Law Design:

[0132]

[0133] in, It is a virtual control input; Let k be the communication in-degree of the ground vehicle. The set of all neighbors of ground vehicle k; This includes the error stabilization coefficient; Let m be the linear velocity of the ground vehicle. Let m be the heading angle of the ground vehicle.

[0134] Step 5.2.3: Robust dynamic control law design:

[0135]

[0136] in, , These are the linear velocity control input and angular velocity control input for the ground vehicle k, respectively. This is the dynamic error stabilization coefficient; For linear velocity tracking error, The target linear velocity; For angular velocity tracking error, For the target angular velocity, The heading angle error stabilization coefficient, For heading angle error, The target heading angle.

[0137] Step 6: Prove the stability of the heterogeneous multi-vehicle system using Lyapunov theory, specifically as follows:

[0138] Step 6.1: Construct the Lyapunov function of the mentor layer.

[0139]

[0140] in, For the Lyapunov function of the leader layer; The energy term for the fusion error; It represents the sum of the potential field energy.

[0141] right Taking the time derivative and simplifying, we get:

[0142]

[0143] in, The deviation between the actual speed and the target speed of the quadrotor can be determined from the PNO characteristics. The exponential convergence to a compact set near zero satisfies , It is a bounded constant.

[0144] Step 6.2: Scale derivatives for different scenarios and analyze stability.

[0145] Scenario 1: The quadcopter is outside the safety constraint range: At this time Substitute And by using Young's inequality to expand and contract, we get:

[0146]

[0147] when hour, Therefore, the formation error convergence is proposed.

[0148] Scenario 2: The quadcopter is within the safety constraints: At this time ,right Further scaling yields:

[0149]

[0150] in, ; From the velocity constraint, we know that , There are boundaries, when hour, ,and Since the number of collisions is not increasing, it can be concluded that the quadcopter is non-collision-free and interconnected.

[0151] Step 6.3: Construct the Lyapunov function for the follower layer.

[0152]

[0153] in, For the Lyapunov function of the follower layer; This is the energy term that includes the fusion error; The in-degree coefficient for follower communication; It represents the sum of the potential field energy.

[0154] right Taking the time derivative and simplifying, we get:

[0155]

[0156] in, This refers to virtual control error.

[0157] Step 6.4: Scale derivatives for different scenarios and analyze stability.

[0158] Scenario 1: Ground vehicles are outside the safety constraint area: At this time Substitute And by using Young's inequality to expand and contract, we get:

[0159]

[0160] in, for The upper bound; ;when hour, It can be seen that convergence includes error.

[0161] Scenario 2: Ground vehicles are within the safety constraint range: At this time ,right Further scaling yields:

[0162]

[0163] when hour, It can be seen that, It is bounded, and ground vehicles do not collide and are connected.

[0164] Step 6.5: Convergence analysis of velocity tracking error in the facilitator layer.

[0165] Combining the facilitator layer error dynamics, path tracking error The state space form is as follows:

[0166]

[0167] in, Given a bounded perturbation, the characteristic equation of the system is:

[0168]

[0169] in, for eigenvalues, When the damping constant Time characteristic roots Since all real parts are negative, , Converging to a compact set near zero; then combining this with the definition of velocity tracking error. ( );like Then either (and (Convergence contradictions), or Therefore, the quadcopter speed converges to the target speed, and the path tracking of the target is achieved.

[0170] Step 6.6: Convergence Analysis of Follower Layer Velocity and Heading Angle Errors: Substitute the velocity tracking error into the ground vehicle dynamics controller. Angular velocity tracking error The error dynamics are obtained as follows:

[0171]

[0172] because and Bounded, as can be seen from the input-state stability (ISS) theory, It converges to a compact set near zero, i.e. , Furthermore, combining the dynamics of heading angle error ,because There is a boundary, so Similarly, the heading of ground vehicles converges to the target heading.

[0173] Step 6.7: Conclusion on global system stability: All signals in the closed-loop system, including path tracking error, inclusion error, and uncertainty estimation error, are UUB, which meets the stability requirements.

[0174] Heterogeneous multi-vehicle systems, as core execution units in intelligent transportation and collaborative operations, require high-precision and high-safety collaborative control in complex environments. Heterogeneous formations of aerial quadcopter drones and ground vehicles are a key application of this system, demanding stringent requirements for collaborative response speed and safety assurance capabilities. The proposed method for collision-free collaborative control of heterogeneous multi-vehicle systems, based on a predictor-driven neural network observer and path guidance, accurately estimates system uncertainties, rapidly responds to path tracking and formation inclusion requirements, and effectively ensures collision-free operation and reliable connectivity between drones and ground vehicles in obstacle-filled and communication-constrained environments. This helps heterogeneous multi-vehicle systems safely and efficiently complete critical tasks such as collaborative monitoring, escort, and operations in hazardous areas.

[0175] The above description is only a preferred embodiment of the present invention and is not intended to limit the ideas of 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 collision-free cooperative control method for a heterogeneous multi-vehicle system based on a predictor-based neural network observer and path guidance, characterized in that, include: Step 1: Establish a dynamic model of the heterogeneous multi-vehicle system to provide model support for hierarchical control; Step 2: Construct an artificial potential field function, incorporating potential field functions related to collision avoidance and communication connectivity in the design of safety constraints; Step 3: Define the virtual root guide's motion rules based on the constraints to provide a reference path for formation; Step 4: Design a predictor-based neural network observer PNO to provide uncertainty compensation for the controller; Step 5: Combine reference path, potential field constraints and predictor-based neural network observer compensation to build a hierarchical control architecture to achieve collision-free cooperative control of heterogeneous multi-vehicle systems; Step 3 specifically involves: Step 3.1: Definition of the Virtual Root Leader Dynamics Model: in, The three-dimensional position vector of the virtual root leader; The three-dimensional velocity vector of the virtual root leader; Step 3.2: Definition of Spatial Constraint Function for Reference Path: The reference path of the virtual root guide is determined by the intersection of two smooth implicit surfaces. The path function is defined as follows: in, and for Smooth functions; and They satisfy the conditions of being bounded and not parallel; Step 3.3: Define the virtual root bootstrap constraint rules: Path preservation constraints: 0 and ; Uniform motion constraint: ,in The preset constant tangential velocity; Ensure the virtual root bootloader always travels along the reference path at a constant speed The velocity vector of the virtual root leader is always related to the motion. and The intersecting curves of the two surfaces are tangent; Step 3.4: Define the velocity vector of the virtual root bootstrapper: in," " represents the cross product of vectors, , Path functions , The gradient; the direction of the cross product result is the tangent of the reference path.

2. The method for collision-free cooperative control of heterogeneous multi-vehicle systems based on predictor-driven neural network observers and path guidance according to claim 1, characterized in that, To address the dynamic differences in heterogeneous multi-vehicle systems, a "leader-follower" architecture is constructed. Step 1 is as follows: Step 1.1: The facilitator layer consists of... The system consists of N aerial quadrotor drones. A three-dimensional dynamic model of aerial quadrotor i is constructed: in, Let be the three-dimensional position and state vector of quadrotor i; Let i be the three-dimensional velocity vector of the quadrotor i; For the control input of the quadcopter i; The model represents the lumped uncertainty of the quadrotor i; it ignores attitude control details and focuses only on position loop control. Step 1.2: Follower layer from A two-dimensional dynamic model of the follower layer is constructed, consisting of M ground vehicles: in, It is the planar position vector of the ground vehicle k; Its linear velocity, Its angular velocity; Its heading angle; Its inertial parameters; Its wheel radius; These are its linear velocity control input and angular velocity control input, respectively; External disturbances; This represents the unmodeled nonlinear term.

3. The method for collision-free cooperative control of heterogeneous multi-vehicle systems based on predictor-driven neural network observers and path guidance according to claim 1, characterized in that, Step 2 specifically involves: Step 2.1: Design of the mutual collision avoidance potential field function: The potential field function takes effect when the distance is between the detection radius and the collision critical distance. The expression is: Between quadrotor i and quadrotor j: in, Let be the collision avoidance potential field function between quadrotor i and quadrotor j; , The distance between the two quadcopters; This is the extreme collision avoidance distance between quadcopters; if it is less than this value, there is a risk of collision. The collision detection radius for the quadcopter; the function in The internal monotonous decrease ensures no collisions between the four rotors; Between ground vehicle k and ground vehicle m: in, Let be the collision avoidance potential field function between ground vehicle k and ground vehicle m; , The distance between two vehicles on the ground; This represents the extreme collision avoidance distance between vehicles on the ground; if the distance is less than this value, there is a risk of collision. The radius of the ground vehicle collision detection; the function is in The internal monotonous decrease ensures no collisions between vehicles on the ground; Step 2.2: Design of the potential field function for collision avoidance with static obstacles: The potential field function takes effect when the distance is between the detection radius and the collision critical distance. The expression is: Quadrotor i and static obstacles between: in, Quadrotor i and static obstacles The collision avoidance potential field function between them; , For obstacles Location, The distance between the quadcopter and the obstacle; This is the minimum collision avoidance distance between the quadcopter and an obstacle; if the distance is less than this value, there is a risk of collision. The function represents the obstacle detection range of the quadcopter. The internal monotonic decrease ensures that the quadcopter avoids static obstacles in the environment; Ground vehicles and static obstacles : in, For ground vehicles and static obstacles The collision avoidance potential field function between them; , For obstacles Location, The distance between ground vehicles and obstacles; This is the minimum avoidance distance between a vehicle and an obstacle on the ground; if the distance is less than this value, there is a risk of collision. The function is the detection range of ground vehicles for obstacles. The internal monotonically decreasing pattern ensures that ground vehicles avoid static obstacles in the environment. Step 2.3: Design of the communication connectivity preservation potential field function: The potential field function takes effect when the distance is between the maximum acceptable communication distance and the minimum distance at which the communication mechanism is activated. The expression is: Between quadrotor i and quadrotor j: in, To maintain the potential field function for communication connectivity between quadrotor i and quadrotor j; This is the maximum acceptable communication distance between quadcopters; communication will be interrupted if this value is exceeded. The minimum distance required to activate the communication mechanism; the function is in Internally, the attraction force ensures communication connectivity between the four rotors; Between ground vehicle k and ground vehicle m: in, To maintain the potential field function for the communication connectivity between ground vehicle k and ground vehicle m; This is the maximum acceptable communication distance between ground vehicles; communication will be interrupted if this value is exceeded. The minimum distance required to activate the communication mechanism; the function is in Internally, the attraction force ensures the communication connectivity of ground vehicles.

4. The method for collision-free cooperative control of heterogeneous multi-vehicle systems based on predictor-driven neural network observers and path guidance according to claim 2, characterized in that, Step 4 specifically involves: Step 4.1: Define the types of uncertainties that need to be compensated, providing a clear basis for the observation targets of PNO; Quadrotor layer uncertainty definition: The lumped uncertainty of a quadrotor originates from external disturbances and state-dependent damping. Combined with its normalized dynamic model, it is defined as follows: in, For the lumped uncertainty of quadrotor i The original perturbation of the quadrotor i in the r direction; Let i be the mass of the quadcopter. Uncertainty definition for ground vehicle layer: The uncertainty of ground vehicles includes external disturbances and unmodeled nonlinear terms. Based on its dynamic model, two key types of uncertainty are defined: in, For the linear velocity-related uncertainties of ground vehicle k, Uncertainties related to the angular velocity of ground vehicle k; Step 4.2: System uncertainties based on neural network approximation: Quadrotor layer uncertainty approximation: in, The ideal neural weight vector; For the basis function vector, Input for the observer; This is a bounded approximation error; Approximate uncertainty for the ground vehicle layer: in, , The ideal neural weight vector; , For the basis function vector, , Input for the observer; This is a bounded approximation error; Step 4.3: Construct the PNO observation model: Construction of the quadrotor layer PNO model: through weight vector norm Simplify learning: in, for The estimated value; for The estimated value; Let i be the actual velocity of the quadrotor in the r direction. The predicted velocity of quadrotor i in the r direction The velocity prediction error of quadrotor i in the r direction; Ground vehicle layer PNO model construction: through weight vector norm Simplify learning: in, , They are respectively , The estimated value; , They are respectively , The estimated value; Let k be the actual linear velocity of the ground vehicle. Let k be the predicted linear velocity of the ground vehicle. The linear velocity prediction error of ground vehicle k; Let k be the actual angular velocity of the ground vehicle. Let k be the actual angular velocity of the ground vehicle. The angular velocity prediction error for ground vehicle k; Step 4.4: Design of State Predictor and Adaptive Learning Law: Design of Quadrotor Layer State Predictor and Learning Law: State predictor: Weight update law: in, for The predicted value, For the control input of the quadcopter i, This is the sedation coefficient; For learning gain; This is the forgetting factor; simultaneously, the initial prediction error is set to 0, i.e. This significantly improves transient estimation performance; Ground vehicle layer state predictor and learning law design: State predictor: Adaptive learning law: in, , for , The predicted value, These are the linear velocity control input and angular velocity control input for the ground vehicle k, respectively. This is the sedation coefficient; For learning gain; Forgetting factor; The uncertainty estimate for quadrotor i is output by PNO; This is the uncertainty estimate of the k-line velocity of ground vehicles output by PNO. This is the uncertainty estimate of the k-angular velocity of the ground vehicle output by PNO.

5. The method for collision-free cooperative control of heterogeneous multi-vehicle systems based on predictor-driven neural network observers and path guidance according to claim 1, characterized in that, Step 5 specifically involves: Step 5.1: Design of the Path Tracking Formation Controller for the Guide Layer: Step 5.1.1: Definition of fusion error: in, For the formation error of quadcopters; Let i be the set of neighbors of the quadrotor i; Let be the relative position vector between quadrotor i and quadrotor j; Let i be the relative position vector between the quadrotor i and the virtual root leader; Let be the desired relative position vector between quadrotor i and quadrotor j. Let i be the expected relative position vector between the quadrotor i and the virtual root leader; The communication weights between quadrotor i and quadrotor j; For virtual root bootstrap communication weights; For safety variables; The number of obstacles detected by the quadcopter i; Step 5.1.2: Target velocity generation guided by the reference path: in, Let be the target velocity vector of quadrotor i; For communication in-degree; This indicates that quadrotor i can receive information from quadrotor j, and vice versa. 0; For virtual root bootstrap communication weights; This indicates that the quadcopter i directly obtains information from the virtual guide, and vice versa. 0; This is the sedation coefficient; Let be the velocity vector of quadcopter j; assume that there exists at least one quadcopter that can directly obtain information from the virtual root leader, that is, the communication graph of the leader layer contains a spanning tree rooted at the virtual root leader; Step 5.1.3: Construction of the complete path tracking formation controller: in, It is the control input for the quadcopter i; These are preset parameters; This is the path tracking constraint matrix; in, , , , ; Nonlinear terms derived from reference path tracking; The projection of the target velocity along the tangent of the path; For error stabilization; where: They are quadcopters i pairs Path tracking error; These are the path tracking errors of the virtual root bootloader; They are respectively The damping constant of the path; The velocity error stabilization coefficient, For speed tracking error; The saturation function is expressed as: in, This represents the upper limit of the repulsive / attractive force. Step 5.2: Design of a distributed containment controller in the follower layer: Step 5.2.1: Definition of fusion error: in, This includes the error for ground vehicles; Let K be the set of follower neighbors of ground vehicle K, which includes other ground vehicles that communicate with K. The leader neighbor set of ground vehicle k contains the quadcopter that communicates with k; The communication weights for ground vehicle k and followers m and quadrotor i are respectively, and the method of value selection is the same as... resemblance; Let m be the position vector of the follower. This is the planar position projection of quadcopter i, that is, ignoring the z-direction position of the quadcopter and only retaining the planar coordinates of the motion along the reference path, which is used to provide a convex hull reference for ground vehicles; Let k be the number of obstacles detected by ground vehicle k. For safety variables; assume that there exists at least one ground vehicle capable of receiving information from at least one quadcopter; Step 5.2.2: Virtual Control Law Design: in, It is a virtual control input; Let k be the communication in-degree of the ground vehicle. The set of all neighbors of ground vehicle k; This includes the error stabilization coefficient; Let m be the linear velocity of the ground vehicle. Let m be the heading angle of the ground vehicle; Step 5.2.3: Robust dynamic control law design: in, , These are the linear velocity control input and angular velocity control input for the ground vehicle k, respectively. This is the dynamic error stabilization coefficient; For linear velocity tracking error, The target linear velocity; For angular velocity tracking error, For the target angular velocity, The heading angle error stabilization coefficient, For heading angle error, The target heading angle.

Images