Unmanned vehicle non-convex optimization control method based on collaborative neurodynamics

Abstract

The invention provides an unmanned vehicle non-convex optimization control method based on collaborative neurodynamics, and aims to improve the control stability and accuracy of an unmanned vehicle under a complex non-convex condition. A discrete linear error model of the unmanned vehicle is obtained through Taylor expansion and a forward Euler method; s2, according to the discrete linearization error model of the unmanned vehicle, constructing a linear secondary regulator strategy to obtain an unmanned vehicle path tracking reference control signal; s3, designing optimization performance indexes, obstacle avoidance non-convex constraints and control input constraints according to the reference control input signal, and establishing an unmanned vehicle non-convex optimization control scheme; s4, aiming at the unmanned vehicle non-convex optimization control scheme, constructing a single neurodynamic model to solve and obtain local optimal control signal input; and S5, for a non-convex condition, expanding a single neurodynamics model into a collaborative neurodynamics model to solve and obtain global optimal control signal input, and driving the unmanned vehicle to complete path tracking and obstacle avoidance tasks. According to the invention, the stability and accuracy of path tracking and obstacle avoidance control tasks executed by the unmanned vehicle under the non-convex condition can be improved.

Description

Technical Field

[0001] The present invention relates to the field of unmanned vehicle control technology, and in particular to a non-convex optimization control method for an unmanned vehicle based on collaborative neural dynamics. Background Art

[0002] With the rapid development of artificial intelligence and autonomous driving technologies, unmanned vehicles are gradually becoming a core component of future intelligent transportation systems. The autonomous driving capabilities of unmanned vehicles rely on their real-time perception, decision-making and control in complex environments, and the realization of these capabilities depends largely on the efficiency and stability of optimization algorithms. In practical applications, unmanned vehicles usually face a large number of non-convex optimization problems, such as path planning, obstacle avoidance, speed control, etc. Such problems usually have multiple local optimal solutions, which greatly increases the complexity and difficulty of the solution process. The operating scenarios of unmanned vehicles are complex and changeable, which requires the control system to respond quickly and accurately to dynamic traffic conditions and emergencies. Therefore, how to develop efficient non-convex optimization control technology to improve the performance of unmanned vehicles in complex environments has become the focus of current research.

[0003] Traditional control methods for unmanned vehicles mainly include model-based control, feedback control, and predictive control. Model-based control methods rely on accurate mathematical modeling and achieve control by describing the kinematic characteristics of unmanned vehicles and their environment in detail. However, such methods often perform poorly in the face of complex and changeable actual environments. Feedback control methods such as proportional-integral-differential control adjust the control input in real time to correct errors. Although they work well in simple scenarios, they are difficult to cope with complex scenarios. Predictive control methods optimize current decisions by predicting the behavior of the system in the future. Although they can theoretically handle complex dynamic environments, their computational overhead is large and it is difficult to meet real-time requirements. Given the limitations of traditional control methods, it is particularly important to study a control technology that can effectively deal with non-convex optimization problems when unmanned vehicles perform complex control tasks. This can not only improve the control accuracy and reliability of unmanned vehicles in complex environments, but also lay a solid foundation for the widespread application of unmanned driving technology.

[0004] The present invention proposes a non-convex optimization control method for unmanned vehicles based on collaborative neural dynamics. According to the dynamic information of obstacles and vehicles, a zeroing control obstacle function is constructed, which is described as a dynamic inequality constraint form and incorporated into the path tracking optimization control problem. The method proposed in the present invention can alleviate the challenge of existing control methods falling into local optimality, while meeting the real-time and high-precision control requirements. This advantage is of great significance to the field of unmanned vehicle control technology. In summary, the patent of the present invention is novel and practical. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a non-convex optimization control method for an unmanned vehicle based on collaborative neural dynamics, aiming to improve the control stability and accuracy of the unmanned vehicle under complex non-convex conditions.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A non-convex optimization control method for an unmanned vehicle based on collaborative neural dynamics can achieve obstacle avoidance while completing the unmanned vehicle path tracking task, including the following steps:

[0008] S1: Based on the continuous nonlinear kinematic model of the unmanned vehicle, the discrete linearized error model of the unmanned vehicle is obtained through Taylor expansion and forward Euler method;

[0009] S2: Based on the discrete linearized error model of the unmanned vehicle, a linear quadratic regulator strategy is constructed to obtain the unmanned vehicle path tracking reference control signal;

[0010] S3: Based on the reference control input signal, design optimization performance indicators, obstacle avoidance non-convex constraints and control input constraints, and establish a non-convex optimization control scheme for the unmanned vehicle;

[0011] S4: Construct a single neural dynamics model for the non-convex optimization control scheme of the unmanned vehicle to solve the local optimal control signal input;

[0012] S5: For non-convex situations, the single neural dynamics model is expanded into a collaborative neural dynamics model to solve the global optimal control signal input and drive the unmanned vehicle to complete path tracking and obstacle avoidance tasks.

[0013] Further, the continuous nonlinear kinematic model of the unmanned vehicle described in step S1 is established as

[0014]

[0015] Where v represents the linear velocity; δ represents the front wheel steering angle; L is the distance between the front and rear wheels of the vehicle; (x, y) is the center position of the rear axle of the vehicle; and θ is the vehicle heading angle.

[0016] Furthermore, let the state vector s = [x, y, θ] and the control vector u = [υ, δ]. The relationship between the state and the control vector is expressed by the nonlinear mapping function f(·, ·): In addition, for a path tracking task, there is where s d =[x d ,y d ,θ d ]T and u d =[υ d ,θ d ] Trepresents the state of the reference, and the superscript T is the transposition operation. By performing Taylor series expansion on the reference term and ignoring the higher-order terms, we obtain

[0017]

[0018] Further, obtain

[0019]

[0020] in and Therefore, the control system is transformed into a linearized error model to achieve the purpose of path tracking. Further, discretization of the above model can be obtained

[0021]

[0022] in

[0023]

[0024] Where T s is the sampling time.

[0025] Furthermore, the linear quadratic regulator strategy constructed in step S2 according to the discrete linear error model of the unmanned vehicle is specifically as follows:

[0026]

[0027] in and is a symmetric positive definite weight matrix. Assume but

[0028]

[0029] To obtain the optimal feedback gain matrix Assume a constant symmetric positive definite matrix satisfy Optimal solution This can be further obtained by solving the following Riccati equation:

[0030]

[0031] Therefore, the reference control signal input generated by the linear quadratic regulator is

[0032]

[0033] Further, the step S3 is specifically as follows:

[0034] S301: constructing a zeroing control obstacle function according to dynamic information of obstacles and vehicles, and describing it as a dynamic nonlinear and non-convex inequality constraint;

[0035] S302: Based on the inequality constraints obtained in step S301, the reference control input signal obtained in step S2 is considered to design an optimization performance index, and a non-convex optimization control scheme for the unmanned vehicle is further established.

[0036] Furthermore, the step S301 is

[0037] Consider a closed set Its definition is as follows:

[0038]

[0039] in is a safe collection, is a security function, and Security Set If the function h is in is differentiable when When h(s,u)→0, and for satisfy , where α h belong class and h(s,u) is Lipschitz continuous, then the function h(s,u) for the safe set is a zeroing control barrier function.

[0040] Therefore, the inequality constraint is designed according to the zeroing control barrier function h(s,u) Further construct the nonlinear and non-convex inequality constraints on the unmanned vehicle and obstacles as follows

[0041]

[0042] x p =υ k cos(θ k )nT s ,y p =υ k sin(θ k )nT s

[0043] i∈{1, 2, …, m},

[0044] in is the position of the ith obstacle; (x p ,y p ) is in the time interval [k, k+nT s]s predicted future position of the vehicle; n is the prediction step length; m is the total number of obstacles; b safe ≥L / 2 indicates the safe distance between the vehicle and the obstacle.

[0045] Furthermore, the optimization performance index described in step S302 is

[0046]

[0047] in, represents a positive definite diagonal matrix; is the reference control input signal obtained in step S2. Therefore, the non-convex optimization control scheme for the unmanned vehicle path tracking and obstacle avoidance tasks is established as

[0048]

[0049] u min ≤u k ≤u max

[0050]

[0051] x p =υ k cos(θ k )nT s ,y p =υ k sin(θ k )nT s

[0052] i∈{1, 2,…, m}.

[0053] Furthermore, the single neural dynamics constructed in step S4 are as follows:

[0054] To facilitate the subsequent derivation of neural dynamics, the non-convex optimization control scheme of the unmanned vehicle obtained in step S4 needs to be appropriately transformed. First, an exact penalty function is introduced to handle inequality constraints and control input constraints, and

[0055]

[0056] in and b=[u max ; -u min ]. According to the first-order necessary conditions of the KKT condition, by solving the corresponding KKT equation, the local optimal solution of the non-convex optimization control scheme of the unmanned vehicle can be found:

[0057]

[0058] in and represent the gradient and Clarke's generalized gradient respectively. Specifically,

[0059]

[0060] A similar transformation can be performed. In this sense, the evolutionary model can be described as Further obtain a single neural dynamics model for unmanned vehicle control:

[0061]

[0062] in is a Hessian matrix; is a design parameter. It is worth noting that the neural dynamics model is essentially constructed based on the first-order condition in the KKT condition, which is a necessary condition for global optimality but not a sufficient condition. Therefore, when it is necessary to deal with a non-convex optimization control scheme with multiple local optimal solutions, the neural dynamics model can only provide a local optimal solution under a specific initial state.

[0063] Further, the step S5 is specifically as follows:

[0064] Design the particle swarm optimization algorithm as

[0065]

[0066] Where v j is the velocity vector of the jth particle; u j is the position vector of the jth particle; p best，ι is the individual optimal position of the jth particle at the ιth iteration; g best，ι is the global optimal position of the particle swarm at the ιth iteration; σ is the inertia weight that balances the contribution of the particle's previous velocity, individual optimal position, and global optimal position; κ1 is the learning factor that adjusts the influence of the individual optimal position of the particle on its velocity; κ2 is the learning factor that adjusts the influence of the global optimal position of the particle swarm on the particle velocity.

[0067] Based on the above particle swarm optimization algorithm, the single neural dynamics model obtained in step S4 is expanded into a group of search particles u j , collaborative search for the global optimum of non-convex optimization control schemes for unmanned vehicles To obtain the optimal control input Drive the unmanned vehicle to avoid obstacles. Based on this method, multiple neural dynamics models are initialized under different initial conditions. Subsequently, each neural dynamics model performs a fine local search under its own initial conditions. When all neural dynamics models converge to the equilibrium point, they are collectively evaluated by evaluating the fitness function φ(g best，ι) and participate in information exchange. Subsequently, each neural dynamics model resets its state when receiving information from other members. The process continues to iterate until the termination condition is met Finally, the global optimal control input is output Drive the unmanned vehicle to complete path tracking and obstacle avoidance tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a flow chart of a non-convex optimization control method for unmanned vehicles based on collaborative neural dynamics;

[0069] Figure 2 This is a kinematic analysis diagram of the unmanned vehicle centered on the rear axle;

[0070] Figure 3 The actual driving path of the vehicle and the given reference path diagram when the control method of the present invention is used in the embodiment of the present invention to perform the unmanned vehicle path tracking and obstacle avoidance tasks;

[0071] Figure 4 A diagram showing position state changes in the X and Y directions when the control method of the present invention is used in an embodiment of the present invention to perform path tracking and obstacle avoidance tasks for an unmanned vehicle;

[0072] Figure 5 A diagram showing a change in vehicle heading angle when the control method of the present invention is used in an embodiment of the present invention to perform unmanned vehicle path tracking and obstacle avoidance tasks;

[0073] Figure 6 A diagram showing a change in the minimum distance between a vehicle and an obstacle when the control method of the present invention is used in an embodiment of the present invention to perform path tracking and obstacle avoidance tasks for an unmanned vehicle;

[0074] Figure 7 A speed variation diagram of an unmanned vehicle when performing path tracking and obstacle avoidance tasks using the control method of the present invention in an embodiment of the present invention;

[0075] Figure 8 A diagram showing changes in the front wheel steering angle when the control method of the present invention is used in an embodiment of the present invention to perform path tracking and obstacle avoidance tasks for an unmanned vehicle;

[0076] Fig. 9 It is a convexity change diagram of the inequality constraint when the control method of the present invention is used in an embodiment of the present invention to perform path tracking and obstacle avoidance tasks for an unmanned vehicle. DETAILED DESCRIPTION

[0077] In order to make the purpose and technical solution of the present invention more clearly understood, the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0078] In this embodiment, the reference path is set as a circular track containing different bends, where the center line of the road is set as the reference path, and the ten obstacles are respectively distributed at O1 = (120, 0) meters, O2 = (200, -90) meters, O3 = (200, -120) meters, O4 = (200, -150) meters, O5 = (125, -50) meters, O6 = (45, -60) meters, O7 = (112.5, -190) meters, O8 = (-50, -130) meters, O9 = (-50, -124) meters, and O10 = (-50, -118) meters. The length of the vehicle chassis is L = 2.866 meters. For the linear quadratic regulator control strategy, the weight matrix is ​​set to and In addition, the control input limits of vehicle speed and front wheel steering angle are set as [υ min , max ] = [-12, 12] m / s and [δ min , δ max ] = [-π / 6, π / 6] radians. The sampling time is set to T s = 0.1 seconds. The prediction time step is set to n = 3. The coefficient of the safety-critical constraint is set to α = 1.8. The weight matrix in the objective function is set to In addition, the parameters of the collaborative neural dynamics method of the present invention are set as η=10,N=200,σ=0.9,κ1=0.9,κ2=0.9,ιmax=200. Based on the above parameter settings, the unmanned vehicle is required to move at the desired speed υ d = 10 m / s Follow the given reference path as closely as possible while satisfying control input constraints and giving priority to avoiding obstacles.

[0079] In this embodiment, a 146-second online experiment was performed on the joint simulation platform of Simulink and CarSim. The experimental results are as follows: Figures 3 to 9 In the experiment, the unmanned vehicle needs to follow the circular path while avoiding 10 obstacles O1-O10 distributed on the path. Specifically, Figure 3 The actual path and predefined path map for controlling the unmanned vehicle to perform tasks using the collaborative neural dynamics method of the present invention. Intuitively, the unmanned vehicle starts from the starting point and drives along the reference path. When it drives in front of an obstacle, it automatically avoids the obstacle and returns to the driving path, and finally returns to the starting point. Figure 4 The figure is a diagram of the position change of the unmanned vehicle in the X and Y directions when the unmanned vehicle is controlled to perform a task using the collaborative neural dynamics method of the present invention, wherein Xd and Yd represent the reference trajectories in the X and Y directions, respectively. Figure 5 The diagram of the heading angle change when the unmanned vehicle is controlled to perform a task using the collaborative neural dynamics method of the present invention is consistent with the expected heading angle trajectory. Figure 6 This is a graph showing the minimum distance variation between the unmanned vehicle and 10 obstacles. During the entire mission, it is always greater than the set safety distance. Figure 7 The speed change diagram of the unmanned vehicle is always less than the given maximum speed limit of 12 m / s. The unmanned vehicle will only actively slow down when encountering an obstacle, and maintain the given speed of 10 m / s after safely bypassing the obstacle. Figure 8 is the front wheel steering angle change diagram of the unmanned vehicle, which is always kept within the given upper and lower limits [δ min , δ max ]=[-π / 6,π / 6] radians. Fig. 9 This is the convexity change diagram of the 10 inequality constraints of the unmanned vehicle regarding 10 obstacles in the non-convex optimization control scheme of the unmanned vehicle, that is, h1-h10, where 1 represents non-convexity and 0 represents convexity. It can be seen from the figure that in the entire unmanned vehicle mission execution process, the above 10 inequality constraints show non-convexity in most cases.

[0080] In summary, when the collaborative neurodynamic method of the present invention is used to control an unmanned vehicle to perform a given task, the unmanned vehicle can track a given path and, when encountering an obstacle, can respond accurately to avoid the obstacle.

[0081] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. A non-convex optimization control method for unmanned vehicles based on collaborative neural dynamics, characterized in that: The method comprises the following steps: S1: Based on the continuous nonlinear kinematic model of the unmanned vehicle, the discrete linearized error model of the unmanned vehicle is obtained through Taylor expansion and forward Euler method; S2: Based on the discrete linearized error model of the unmanned vehicle, a linear quadratic regulator strategy is constructed to obtain the unmanned vehicle path tracking reference control signal; S3: Based on the reference control input signal, design optimization performance indicators, obstacle avoidance non-convex constraints and control input constraints, and establish a non-convex optimization control scheme for the unmanned vehicle; S4: Construct a single neural dynamics model for the non-convex optimization control scheme of the unmanned vehicle to solve the local optimal control signal input; S5: For non-convex situations, the single neural dynamics model is expanded into a collaborative neural dynamics model to solve the global optimal control signal input and drive the unmanned vehicle to complete path tracking and obstacle avoidance tasks.

2. According to the non-convex optimization control method of unmanned vehicle based on collaborative neural dynamics as described in claim 1, it is characterized in that: The continuous nonlinear kinematic model of the unmanned vehicle described in step S1 is established as Where v represents the linear velocity; δ represents the front wheel steering angle; L is the distance between the front and rear wheels of the vehicle; (x, y) is the center position of the rear axle of the vehicle; θ is the heading angle of the vehicle; Furthermore, let the state vector s = [x, y, θ] and the control vector u = [υ, δ]; the relationship between the state and the control vector is expressed by the nonlinear mapping function f(·, ·): In addition, for a path tracking task, there is where s d =[x d ,y d ,θ d ] T and u d =[υ d ,θ d ] T represents the state of the reference, and the superscript T is the transposition operation; by performing Taylor series expansion on the reference term and ignoring the higher-order terms, we get Further, obtain in and Therefore, the control system is transformed into a linearized error model to achieve the purpose of path tracking; further, the above model is discretized to obtain in Where T s is the sampling time.

3. According to the non-convex optimization control method of unmanned vehicle based on collaborative neural dynamics as described in claim 1, it is characterized in that: The step S2 is specifically as follows: The linear quadratic regulator strategy constructed based on the discrete linear error model of the unmanned vehicle is as follows: in and is a symmetric positive definite weight matrix; assuming but To obtain the optimal feedback gain matrix Assume a constant symmetric positive definite matrix satisfy Optimal solution This can be further obtained by solving the following Riccati equation: Therefore, the reference control signal input generated by the linear quadratic regulator is 4. According to the non-convex optimization control method of unmanned vehicle based on collaborative neural dynamics as described in claim 1, it is characterized in that: The step S3 is specifically as follows: S301: constructing a zeroing control obstacle function according to dynamic information of obstacles and vehicles, and describing it as a dynamic nonlinear and non-convex inequality constraint; S302: According to the inequality constraint obtained in step S301, the reference control input signal obtained in step 2S2 is considered to design an optimization performance index, and a non-convex optimization control scheme for the unmanned vehicle is further established; Furthermore, the step S301 is Consider a closed set Its definition is as follows: in is a safe collection, is a security function, and Security Set The boundary and interior of is differentiable when When h(s,u)→0, and for Satisfies h(s,u)-αhh(s,u)≥0, where αh belongs to class and h(s,u) is Lipschitz continuous, then the function h(s,u) for the safe set is a zeroing control barrier function; Therefore, the inequality constraint is designed according to the zeroing control barrier function h(s,u) Further construct the nonlinear and non-convex inequality constraints on the unmanned vehicle and obstacles as follows in is the position of the ith obstacle; (x p ,y p ) is in the time interval [k, k+nT s ]s predicted future position of the vehicle; n is the prediction step length; m is the total number of obstacles; b safe ≥L / 2 indicates the safe distance between the vehicle and the obstacle; Furthermore, the optimization performance index described in step S302 is in, represents a positive definite diagonal matrix; is the reference control input signal obtained in step S2; therefore, the non-convex optimization control scheme for the unmanned vehicle path tracking and obstacle avoidance tasks is established as in min in k in max 5. According to the non-convex optimization control method of unmanned vehicle based on collaborative neural dynamics as described in claim 1, it is characterized in that: The step S4 is specifically as follows: To facilitate the subsequent derivation of neural dynamics, the non-convex optimization control scheme of the unmanned vehicle obtained in step S4 needs to be appropriately transformed. First, an exact penalty function is introduced to handle the inequality constraints and control input constraints, and the result is in and b=[u max ; -u min ]; According to the first-order necessary conditions of the KKT condition, by solving the corresponding KKT equation, the local optimal solution of the non-convex optimization control scheme of the unmanned vehicle can be found: in and represent the gradient and Clarke's generalized gradient, respectively; specifically, A similar transformation can be made; in this sense, the evolutionary model can be described as Further obtain a single neural dynamics model for unmanned vehicle control: in is a Hessian matrix; is a design parameter; it is worth noting that the neural dynamics model is essentially constructed based on the first-order condition in the KKT condition, which is a necessary condition for global optimality but not a sufficient condition; Therefore, when it is necessary to deal with non-convex optimization control schemes with multiple local optimal solutions, the neural dynamics model can only provide the local optimal solution under a specific initial state.

6. According to the non-convex optimization control method of unmanned vehicle based on collaborative neural dynamics as described in claim 1, it is characterized in that: The step S5 is specifically as follows: Design the particle swarm optimization algorithm as Where v j is the velocity vector of the jth particle; u j is the position vector of the jth particle; p best,ι is the individual optimal position of the jth particle at the ιth iteration; gbest,l is the global optimal position of the particle swarm at the ιth iteration; σ is the inertia weight that balances the contribution of the particle's previous velocity, individual optimal position, and global optimal position; κ1 is the learning factor that adjusts the influence of the individual optimal position of the particle on its velocity; κ2 is the learning factor that adjusts the influence of the global optimal position of the particle swarm on the particle velocity; Based on the above particle swarm optimization algorithm, the single neural dynamics model obtained in step S4 is expanded into a group of search particles u j , collaborative search for the global optimum of non-convex optimization control schemes for unmanned vehicles To obtain the optimal control input Drive the unmanned vehicle to avoid obstacles; Based on this method, multiple neural dynamic models are initialized under different initial conditions; then, each neural dynamic model performs a refined local search under its own initial conditions; when all neural dynamic models converge to the equilibrium point, they collectively evaluate the fitness function φ(gbest,i) and participate in information exchange; then, each neural dynamic model resets its state when receiving information from other members; This process continues to iterate until the termination condition is met. Until; finally, output the global optimal control input Drive the unmanned vehicle to complete path tracking and obstacle avoidance tasks.

Images