A new energy vehicle obstacle avoidance trajectory tracking control method

Abstract

The application discloses a new energy automobile obstacle avoidance trajectory tracking control method, comprising the following steps: establishing a trajectory tracking error model based on path curvature; constructing a data-driven error compensation model based on a radial basis function network, performing real-time feedforward compensation, and obtaining a corrected vehicle state prediction equation; constructing a high-level adaptive adjustment strategy model based on a deep deterministic policy gradient, and outputting state weight adjustment coefficients and control weight adjustment coefficients of a bottom layer MPC online; reconstructing a dynamic Hessian matrix and a gradient vector of a bottom layer MPC quadratic programming standard type, and then solving an optimal control increment sequence; and extracting optimal control instructions of a current control period and applying the optimal control instructions to vehicle actuators, so as to realize lateral trajectory tracking control of the new energy automobile. The application improves the control precision and stability of the new energy automobile in obstacle avoidance trajectory tracking.

Description

Technical Field

[0001] This invention relates to the field of intelligent vehicle technology, and more specifically to a method for obstacle avoidance trajectory tracking control of new energy vehicles. Background Technology

[0002] With the large-scale development of the new energy vehicle industry and the deep integration of intelligent connected vehicle technology, advanced driver assistance and autonomous driving technologies have become key directions for upgrading new energy passenger vehicle products and building core competitiveness. As the mainstream means of transportation for urban commuting and intercity travel, the lateral trajectory tracking control accuracy and robust stability of the intelligent driving system of new energy passenger vehicles directly determine vehicle driving safety, ride comfort, and road traffic efficiency, and are the core research content in the field of intelligent driving lateral control.

[0003] However, existing trajectory tracking control methods still have significant technical shortcomings when applied to new energy passenger vehicles. Most current mainstream trajectory tracking algorithms are developed for small passenger vehicles and fail to adequately adapt to the structural characteristics and dynamics of medium and large new energy passenger vehicles. These vehicles generally have longer wheelbases and larger body dimensions, resulting in significantly different lateral dynamic response characteristics compared to small passenger vehicles. Directly using trajectory tracking algorithms designed for small passenger vehicles cannot accurately match the real dynamic response patterns of large-sized new energy passenger vehicles, easily leading to control parameter mismatch and control performance degradation. Ultimately, this results in insufficient control accuracy and stability of existing obstacle avoidance trajectory tracking methods for new energy passenger vehicles. Summary of the Invention

[0004] In view of this, the present invention provides a method for obstacle avoidance trajectory tracking control of new energy vehicles, so as to improve the control accuracy and stability of new energy vehicles in obstacle avoidance trajectory tracking.

[0005] A method for tracking and controlling obstacle avoidance trajectory of a new energy vehicle includes: Step S1: Collect the vehicle position coordinates, heading angle and yaw rate of the new energy vehicle, establish a trajectory tracking error model based on path curvature, and map the vehicle body size and minimum turning radius into a set of steering hard constraints for the front wheel steering angle and its control increment. Step S2: Construct a data-driven error compensation model based on radial basis function network, and perform real-time feedforward compensation on the trajectory tracking error model established in step S1 to obtain the corrected vehicle state prediction equation. Step S3: Based on the modified vehicle state prediction equation, construct a high-level adaptive adjustment strategy model based on deep deterministic strategy gradient, and output the state weight adjustment coefficient and control weight adjustment coefficient of the underlying MPC online. Step S4: Combining the adjustment coefficient output from step S3 with the corrected vehicle state prediction equation from step S2, reconstruct the dynamic Hessian matrix and gradient vector of the standard form of the underlying MPC quadratic programming, and use the set of steering hard constraints in step S1 as the solution boundary. Then, use the numerical optimization solver to perform rolling optimization and obtain the optimal control increment sequence. Step S5: Based on the optimal control increment sequence in step S4, extract the optimal control command for the current control cycle and apply it to the vehicle actuator. At the same time, update the system state to step S1 to form a rolling time domain closed-loop feedback control, thereby realizing the lateral trajectory tracking control of the new energy vehicle.

[0006] The obstacle avoidance trajectory tracking control method for new energy vehicles provided by the present invention has the following beneficial effects: 1. This invention constructs a data-driven error compensation model based on radial basis function (RBF) networks. By extracting the residual vector between the actual system and the nominal model, it approximates and outputs the estimated value of the unmodeled dynamic deviation online. This estimated value is then superimposed as a feedforward term into the corrected prediction equation. This enables the prediction model to perceive and actively offset the uncertainties caused by tire nonlinearity and environmental disturbances in real time, thereby significantly reducing the peak values ​​of lateral position error and heading angle error. This allows the vehicle to closely follow the target trajectory even under transient conditions such as continuous obstacle avoidance and lane changing, improving the tracking accuracy of the vehicle's obstacle avoidance trajectory.

[0007] 2. This invention directly utilizes the vehicle's wheelbase and minimum turning radius to map them into a set of hard steering constraints, specifically the front wheel steering angle and its increment. This transforms the physical limit into a feasible domain constraint space for the optimization solver. Simultaneously, it innovatively introduces a logarithmic obstacle penalty function into the high-level policy reward function. When the actual turning radius approaches the physical limit, the reward value drops sharply in a logarithmic fashion. Through the dual protection of hard boundary limits and soft penalty warnings, the invention effectively prevents control commands from exceeding the actuator's physical upper limit, effectively avoiding vehicle loss of control issues caused by planned commands exceeding chassis limits under high-curvature obstacle avoidance paths, thus improving the safety of the vehicle's obstacle avoidance process.

[0008] 3. This invention constructs a high-level adaptive adjustment strategy model based on Deep Deterministic Policy Gradient (DDPG). Taking the corrected state features after feedforward compensation as input, it infers online and outputs dynamic state weight adjustment coefficients and control weight adjustment coefficients. This strategy guides the network to evolve towards "strong correction when there is a large error and weak suppression when there is a small error" through a multi-objective reward function. This mechanism enables the Hessian matrix and gradient vector of the cost function of the underlying MPC to be dynamically reconstructed with the real-time state of the vehicle and the curvature of the environment. This gives the system high sensitivity in the initial stage of lane changing to quickly avoid obstacles, and automatically switches to high smoothness in the end stage of lane changing to suppress overshoot. Ultimately, it improves the control accuracy and stability of new energy vehicles in obstacle avoidance trajectory tracking. Attached Figure Description

[0009] Figure 1 A flowchart illustrating the obstacle avoidance trajectory tracking control method for new energy vehicles; Figure 2 This is a comparison chart of the lateral error between the method of this invention and the traditional MPC tracking control method; Figure 3 This is a comparison chart of the heading error between the method of this invention and the traditional MPC tracking control method. Detailed Implementation

[0010] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain embodiments of the present invention, and should not be construed as limiting the present invention.

[0011] Please see Figure 1 The present invention provides a method for tracking and controlling the obstacle avoidance trajectory of a new energy vehicle, the method comprising steps S1 to S5: Step S1: Collect the vehicle position coordinates, heading angle and yaw rate of the new energy vehicle, establish a trajectory tracking error model based on path curvature, and map the vehicle body size and minimum turning radius into a set of steering hard constraints for the front wheel steering angle and its control increment.

[0012] Specifically, define the error state vector. for:

[0013] in, This represents the lateral error between the vehicle trajectory and the reference path. This represents the rate of change of the lateral error. For heading angle error, This represents the rate of change of heading angle error.

[0014] Previous wheel angle To control the input, the reference path curvature is used. As a disturbance input, a continuous-time tracking error model is established, whose state-space equation is:

[0015] in, for The first derivative, The output vector of a continuous-time system. It is a continuous-time state matrix. For continuous-time control input matrix, The input matrix is ​​a continuous-time perturbation. It is a continuous-time output matrix.

[0016]

[0017]

[0018]

[0019]

[0020] in, For the total mass of the vehicle. For the longitudinal speed of the vehicle, For the front wheel lateral stiffness, For rear wheel lateral stiffness, This is the distance from the center of gravity to the front axle. This is the distance from the center of mass to the rear axle. Let Z be the moment of inertia of the vehicle about the Z-axis.

[0021] Utilizing sampling time Discretizing the equation yields the discrete-time trajectory tracking error model output in step S1, which is the trajectory tracking error model based on path curvature described in step S1. The expression is:

[0022] in, for Error state vector at time step [time]. for Error state vector at time step [time]. The discrete-time state matrix, For discrete-time control matrix, The discrete-time perturbation matrix is... For discrete-time output matrix, for The output vector of a discrete-time system at a given time.

[0023] Turning to hard constraint set for:

[0024] in, This refers to the vehicle's wheelbase. This is the vehicle's minimum turning radius.

[0025] Step S2: Construct a data-driven error compensation model based on radial basis function network to perform real-time feedforward compensation on the trajectory tracking error model established in step S1, so as to obtain the corrected vehicle state prediction equation.

[0026] Specifically, step S2 includes: Step S201: Extract the lateral position deviation, heading angle deviation and their rate of change as system residuals, and use them as the input vector of the radial basis function network; Step S202: The radial basis function network in step S201 is used to approximate the tire side slip nonlinearity characteristics and external environmental lateral disturbances of new energy vehicles under complex working conditions, and the estimated value of the unmodeled lateral dynamics deviation is calculated and output online. In step S203, the unmodeled lateral dynamics deviation estimate output in step S202 is used as a feedforward compensation term and superimposed on the trajectory tracking error model in step S1 to construct the corrected vehicle state prediction equation.

[0027] Specifically, the system residual vector in step S201 is:

[0028] in, for The system residual vector at time t, This represents the vehicle's true lateral error state vector fed back from the sensors. To predict the state vector for an uncompensated ideal base model. for The lateral position error of the vehicle at any given time. This represents the actual lateral speed error of the vehicle. This represents the actual heading angle error of the vehicle. This represents the actual yaw rate error. for The ideal lateral position error of the vehicle at any given time. For the ideal lateral velocity error of the vehicle, This represents the vehicle's ideal heading angle error. This represents the error in the ideal yaw rate.

[0029] The unmodeled deviation estimate of the lateral dynamics in step S202 satisfies:

[0030]

[0031]

[0032] in, This is an estimate of the unmodeled bias in the lateral dynamics. This is the adaptive weight matrix from the hidden layer to the output layer of the radial basis function network. Indicates transpose. This represents the column quantity composed of the nonlinear activation outputs of all nodes in the hidden layer of a radial basis function network. for The first in The nth element (specifically representing the nth element of the radial basis function network) (Gausky nonlinear activation output value of each hidden layer node). ; This represents the total number of hidden layer nodes in a radial basis function network. Indicates the first The center vector of the Gaussian function, Indicates the first The expansion constants of the Gaussian function, This represents the Euclidean norm.

[0033] In step S203, the corrected vehicle state prediction equation is:

[0034]

[0035]

[0036]

[0037]

[0038] in, for The augmented state vector at time step 1. for The augmented state vector at time step 1. The matrix of the discretized augmented system is... The discretized augmented control input matrix. This is the discretized augmented path curvature perturbation matrix. The discretized augmented compensation mapping matrix is... for Incremental front wheel steering angle control at any given moment For reference path curvature, , , These represent the basic state matrix, basic control matrix, and basic curvature perturbation matrix in the discretized trajectory tracking error model, respectively. The dimension is A matrix of all zeros The dimension is The identity matrix.

[0039] Step S3: Based on the modified vehicle state prediction equation, construct a high-level adaptive adjustment strategy model based on deep deterministic strategy gradient, and output the state weight adjustment coefficient and control weight adjustment coefficient of the underlying MPC online.

[0040] Specifically, step S3 includes: Step S301: Extract hysteresis-free lateral features based on the modified vehicle state prediction equation in step S2 to construct the system state space; at the same time, construct a deep deterministic policy network based on a multilayer perceptron to nonlinearly map the system state space into a continuous action space composed of the underlying MPC weight adjustment coefficients. Step S302: To guide the policy network described in step S301 to output the optimal adjustment action, a multi-objective reward function of the turning radius logarithmic obstacle penalty is constructed for policy evaluation. The optimal action is then converted into the dynamic weight matrix required by the underlying MPC in real time through a mapping mechanism, thereby realizing the output of the state weight adjustment coefficient and the control weight adjustment coefficient.

[0041] Specifically, in step S301, the system state space vector for:

[0042] in, Indicates the time step in a continuous interaction process; This is a corrected lateral position deviation to eliminate the hysteresis effect after feedforward compensation; To correct lateral velocity deviation, To correct the heading angle deviation, To correct for the rate of change of heading angle deviation; To output deterministic weight adjustment actions, a definition is defined that includes the network parameter set. A deep policy network. The policy network will define the state space. As input layer features, high-dimensional abstract features are extracted through the non-linear activation function of the hidden layer. .

[0043] Abstract features for:

[0044] in, The learnable weight matrix of the hidden layer of the policy network. For bias vectors, The hidden layer is a nonlinear activation function; The continuous action space vector is:

[0045] in, This is the continuous action space vector output after computation by the policy network. and These represent the state weight adjustment coefficients and control weight adjustment coefficients mapped to the underlying MPC, respectively. For Sigmoid normalized activation function, The learnable weight matrix for the output layer of the policy network. For bias vectors, Random noise is used to enhance the action space exploration ability of the policy network in the early stages of training.

[0046] In step S302, the logarithmic obstacle penalty for the turning radius satisfies:

[0047] in, A logarithmic obstacle penalty function designed for the turning characteristics of large light trucks. The gain coefficient of the logarithmic penalty term. for The actual turning radius of the vehicle at any given time. The set safety warning turning radius threshold, This represents the vehicle's minimum turning radius. This is a constant penalty value when physical limits are exceeded.

[0048] In step S302, the multi-objective reward function is:

[0049] in, A multi-objective reward function optimized for high-level policy networks. and These represent the penalty weighting factors for tracking accuracy, specifically for lateral position deviation and heading angle deviation; Dynamic positive definite state weight matrix With control weight matrix They are respectively:

[0050]

[0051] in, and The basic weight matrix initially set for the underlying MPC.

[0052] Step S4: Combining the adjustment coefficients output in step S3 with the corrected vehicle state prediction equations in step S2, the dynamic Hessian matrix and gradient vector of the underlying MPC quadratic programming standard form are reconstructed. The set of steering hard constraints in step S1 is used as the solution boundary. A rolling optimization is performed using a numerical optimization solver to obtain the optimal control increment sequence.

[0053] Among them, the reconstructed dynamic Hessian matrix With gradient vector The expression is:

[0054]

[0055] in, This represents the block diagonal state weight matrix, which is the result of expanding the dynamic positive definite state weight matrix output in step S3 over the entire prediction time domain. This represents the block diagonal state weight matrix, which is the control weight matrix output in step S3 expanded over the entire prediction time domain. This represents the control increment transfer matrix generated based on the modified prediction equation. This is the state prediction evolution matrix generated based on the modified prediction equation. To generate the path curvature perturbation mapping matrix, This is a prediction sequence of the target path curvature in the future time domain. The mapping matrix for the generated feedforward compensation term is used in the derivation. This is the predicted sequence vector of the unmodeled lateral dynamic bias (feedforward compensation) calculated by the radial basis function network. yes The sequence in the prediction time domain.

[0056] Optimal control increment sequence for:

[0057]

[0058] in, Let be the control increment sequence variable to be solved. To satisfy the feasible region constraint space of minimum turning radius, This involves transforming the control increment sequence into an inequality constraint mapping matrix for the absolute control quantity. These are the constraint boundary vectors.

[0059] Step S5: Based on the optimal control increment sequence in step S4, extract the optimal control command for the current control cycle and apply it to the vehicle actuator. At the same time, update the system state to step S1 to form a rolling time domain closed-loop feedback control, thereby realizing the lateral trajectory tracking control of the new energy vehicle.

[0060] Specifically, step S5 includes: Step S501: Call the optimal control increment sequence output by the numerical optimization solver in step S4, extract the front wheel angle increment corresponding to the first prediction step, and add it to the front wheel angle at the previous moment to calculate the optimal control command at the current moment. In step S502, the optimal control command is sent to the steering actuator of the new energy vehicle for lateral motion control; at the same time, the vehicle sensors collect the latest vehicle position coordinates, heading angle and yaw rate after the control command is executed in real time. Step S503: The latest vehicle status, including the latest vehicle position coordinates, heading angle and yaw rate, collected in step S502, is fed back to step S1. The initial error of trajectory tracking at the next moment is recalculated, and the prediction time domain is rolled forward one step over time to enter the prediction and optimization process of the next control cycle, thereby realizing vehicle trajectory tracking control.

[0061] Specifically, in step S501, the first prediction step corresponds to the front wheel steering angle increment. for:

[0062] in, To extract row vectors from the control sequence, only the first element is extracted; The optimal front wheel steering angle at the current moment As the optimal control command, the expression is:

[0063] in, The front wheel steering angle actually executed at the previous moment. This is the optimal front wheel steering angle control increment extracted from the optimized sequence for the first prediction step at the current moment.

[0064] In step S503, the latest vehicle status satisfy:

[0065] in, For real vehicle function mapping, This provides feedback on the actual state of the sensor at the current moment. This refers to unmodeled environmental disturbances that exist in the real environment, including crosswinds or changes in road surface adhesion coefficient.

[0066] The following simulation experiments were conducted using Matlab R2024a. To make the simulation results closer to the actual driving conditions of new energy vehicles, random dynamic disturbances were introduced to simulate road interference, and steering angle and steering rate constraints were set to simulate the physical characteristics of the actuators. The same operating parameters were used for both the traditional MPC tracking control method and the present invention to ensure the rationality and objectivity of the comparative experiments. To verify the trajectory tracking accuracy of the control method of the present invention, a continuous obstacle avoidance and lane changing scenario at a vehicle speed of 72 km / h was conducted, and the lateral and heading errors of the present invention and the traditional MPC tracking control method were compared.

[0067] The lateral error comparison between the present invention and the traditional MPC tracking control method is as follows: Figure 2 As shown, by Figure 2 It can be seen that, during continuous lane-changing obstacle avoidance, the peak value of the lateral error of the method of the present invention is significantly lower than that of the traditional method, which effectively suppresses the trajectory deviation, makes the vehicle drive closer to the target trajectory, and significantly improves the trajectory tracking accuracy. At the same time, the error fluctuation is smaller and the convergence is faster, the lateral attitude of the vehicle body is more stable, which enhances the driving stability of the vehicle in high-speed continuous maneuvering scenarios and reduces the collision risk caused by excessive trajectory deviation.

[0068] Comparison of heading error between the present invention and the traditional MPC tracking control method Figure 3 As shown, by Figure 3 It can be seen that during lane change and obstacle avoidance, the peak value of the heading error of the method of the present invention is significantly lower than that of the traditional method. It effectively suppresses the deviation between the vehicle's orientation and the tangent of the target trajectory, avoids the violent swaying of the vehicle's front end and the excessive side slip angle of the vehicle during lane change, and makes the vehicle's posture more in line with the trajectory direction. It significantly enhances the vehicle's posture stability and handling smoothness during high-speed maneuvering and reduces the risk of trajectory loss of control caused by heading deviation.

[0069] In summary, the obstacle avoidance trajectory tracking control method for new energy vehicles according to the above embodiments has the following beneficial effects: 1. This invention constructs a data-driven error compensation model based on radial basis function (RBF) networks. By extracting the residual vector between the actual system and the nominal model, it approximates and outputs the estimated value of the unmodeled dynamic deviation online. This estimated value is then superimposed as a feedforward term into the corrected prediction equation. This enables the prediction model to perceive and actively offset the uncertainties caused by tire nonlinearity and environmental disturbances in real time, thereby significantly reducing the peak values ​​of lateral position error and heading angle error. This allows the vehicle to closely follow the target trajectory even under transient conditions such as continuous obstacle avoidance and lane changing, improving the tracking accuracy of the vehicle's obstacle avoidance trajectory.

[0070] 2. This invention directly utilizes the vehicle's wheelbase and minimum turning radius to map them into a set of hard steering constraints, specifically the front wheel steering angle and its increment. This transforms the physical limit into a feasible domain constraint space for the optimization solver. Simultaneously, it innovatively introduces a logarithmic obstacle penalty function into the high-level policy reward function. When the actual turning radius approaches the physical limit, the reward value drops sharply in a logarithmic fashion. Through the dual protection of hard boundary limits and soft penalty warnings, the invention effectively prevents control commands from exceeding the actuator's physical upper limit, effectively avoiding vehicle loss of control issues caused by planned commands exceeding chassis limits under high-curvature obstacle avoidance paths, thus improving the safety of the vehicle's obstacle avoidance process.

[0071] 3. This invention constructs a high-level adaptive adjustment strategy model based on Deep Deterministic Policy Gradient (DDPG). Taking the corrected state features after feedforward compensation as input, it infers online and outputs dynamic state weight adjustment coefficients and control weight adjustment coefficients. This strategy guides the network to evolve towards "strong correction when there is a large error and weak suppression when there is a small error" through a multi-objective reward function. This mechanism enables the Hessian matrix and gradient vector of the cost function of the underlying MPC to be dynamically reconstructed with the real-time state of the vehicle and the curvature of the environment. This gives the system high sensitivity in the initial stage of lane changing to quickly avoid obstacles, and automatically switches to high smoothness in the end stage of lane changing to suppress overshoot. Ultimately, it improves the control accuracy and stability of new energy vehicles in obstacle avoidance trajectory tracking.

[0072] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for tracking and controlling the obstacle avoidance trajectory of a new energy vehicle, characterized in that, include: Step S1: Collect the vehicle position coordinates, heading angle and yaw rate of the new energy vehicle, establish a trajectory tracking error model based on path curvature, and map the vehicle body size and minimum turning radius into a set of steering hard constraints for the front wheel steering angle and its control increment. Step S2: Construct a data-driven error compensation model based on radial basis function network, and perform real-time feedforward compensation on the trajectory tracking error model established in step S1 to obtain the corrected vehicle state prediction equation. Step S3: Based on the modified vehicle state prediction equation, construct a high-level adaptive adjustment strategy model based on deep deterministic strategy gradient, and output the state weight adjustment coefficient and control weight adjustment coefficient of the underlying MPC online. Step S4: Combining the adjustment coefficient output from step S3 with the corrected vehicle state prediction equation from step S2, reconstruct the dynamic Hessian matrix and gradient vector of the standard form of the underlying MPC quadratic programming, and use the set of steering hard constraints in step S1 as the solution boundary. Then, use the numerical optimization solver to perform rolling optimization and obtain the optimal control increment sequence. Step S5: Based on the optimal control increment sequence in step S4, extract the optimal control command for the current control cycle and apply it to the vehicle actuator. At the same time, update the system state to step S1 to form a rolling time domain closed-loop feedback control, thereby realizing the lateral trajectory tracking control of the new energy vehicle.

2. The obstacle avoidance trajectory tracking control method for new energy vehicles according to claim 1, characterized in that, Step S2 specifically includes: Step S201: Extract the lateral position deviation, heading angle deviation and their rate of change as system residuals, and use them as the input vector of the radial basis function network; Step S202: The radial basis function network in step S201 is used to approximate the tire side slip nonlinearity characteristics and external environmental lateral disturbances of new energy vehicles under complex working conditions, and the estimated value of the unmodeled lateral dynamics deviation is calculated and output online. In step S203, the unmodeled lateral dynamics deviation estimate output in step S202 is used as a feedforward compensation term and superimposed on the trajectory tracking error model in step S1 to construct the corrected vehicle state prediction equation.

3. The obstacle avoidance trajectory tracking control method for new energy vehicles according to claim 2, characterized in that, The system residual vector in step S201 is: in, for The system residual vector at time t, This represents the vehicle's true lateral error state vector fed back from the sensors. To predict the state vector for an uncompensated ideal base model. for The lateral position error of the vehicle at any given time. This represents the actual lateral speed error of the vehicle. This represents the actual heading angle error of the vehicle. This represents the actual yaw rate error. for The ideal lateral position error of the vehicle at any given time. For the ideal lateral velocity error of the vehicle, This represents the vehicle's ideal heading angle error. This represents the ideal yaw rate error. The unmodeled deviation estimate of the lateral dynamics in step S202 satisfies: in, This is an estimate of the unmodeled bias in the lateral dynamics. This is the adaptive weight matrix from the hidden layer to the output layer of the radial basis function network. Indicates transpose. This represents the column quantity composed of the nonlinear activation outputs of all nodes in the hidden layer of a radial basis function network. for The first in The nth element (specifically representing the nth element of the radial basis function network) (Gausky nonlinear activation output value of each hidden layer node). ; This represents the total number of hidden layer nodes in a radial basis function network. Indicates the first The center vector of the Gaussian function, Indicates the first The expansion constants of the Gaussian function, Denotes the Euclidean norm; In step S203, the corrected vehicle state prediction equation is: in, for The augmented state vector at time step 1. for The augmented state vector at time step 1. The matrix of the discretized augmented system is... The discretized augmented control input matrix. This is the discretized augmented path curvature perturbation matrix. The discretized augmented compensation mapping matrix is... for Incremental front wheel steering angle control at any given moment For reference path curvature, , , These represent the basic state matrix, basic control matrix, and basic curvature perturbation matrix in the discretized trajectory tracking error model, respectively. The dimension is A matrix of all zeros The dimension is The identity matrix.

4. The obstacle avoidance trajectory tracking control method for new energy vehicles according to claim 3, characterized in that, Step S3 specifically includes: Step S301: Extract hysteresis-free lateral features based on the modified vehicle state prediction equation in step S2 to construct the system state space; at the same time, construct a deep deterministic policy network based on a multilayer perceptron to nonlinearly map the system state space into a continuous action space composed of the underlying MPC weight adjustment coefficients. Step S302: Construct a multi-objective reward function with a turning radius logarithmic obstacle penalty for strategy evaluation, and convert the optimal action into a dynamic weight matrix required by the underlying MPC in real time through a mapping mechanism, thereby realizing the output of state weight adjustment coefficient and control weight adjustment coefficient.

5. The obstacle avoidance trajectory tracking control method for new energy vehicles according to claim 4, characterized in that, In step S301, the system state space vector for: in, Indicates the time step in a continuous interaction process; This is a corrected lateral position deviation to eliminate the hysteresis effect after feedforward compensation; To correct lateral velocity deviation, To correct the heading angle deviation, To correct for the rate of change of heading angle deviation; Abstract features for: in, The learnable weight matrix of the hidden layer of the policy network. For bias vectors, The hidden layer is a nonlinear activation function; The continuous action space vector is: in, This is the continuous action space vector output after computation by the policy network. and These represent the state weight adjustment coefficients and control weight adjustment coefficients mapped to the underlying MPC, respectively. For Sigmoid normalized activation function, The learnable weight matrix for the output layer of the policy network. For bias vectors, It is random noise; In step S302, the logarithmic obstacle penalty for the turning radius satisfies: in, Let logarithmic barrier penalty function, The gain coefficient of the logarithmic penalty term. for The actual turning radius of the vehicle at any given time. The set safety warning turning radius threshold, This represents the vehicle's minimum turning radius. This is a constant penalty value when physical limits are exceeded; In step S302, the multi-objective reward function is: in, A multi-objective reward function optimized for high-level policy networks. and These represent the penalty weighting factors for tracking accuracy, specifically for lateral position deviation and heading angle deviation; Dynamic positive definite state weight matrix With control weight matrix They are respectively: in, and The basic weight matrix initially set for the underlying MPC.

6. The obstacle avoidance trajectory tracking control method for new energy vehicles according to claim 5, characterized in that, In step S4, the reconstructed dynamic Hessian matrix With gradient vector The expression is: in, This represents the block diagonal state weight matrix, which is the result of expanding the dynamic positive definite state weight matrix output in step S3 over the entire prediction time domain. This represents the block diagonal state weight matrix, which is the control weight matrix output in step S3 expanded over the entire prediction time domain. This represents the control increment transfer matrix generated based on the modified prediction equation. This is the state prediction evolution matrix generated based on the modified prediction equation. To generate the path curvature perturbation mapping matrix, This is a prediction sequence of the target path curvature in the future time domain. The mapping matrix for the generated feedforward compensation term is used in the derivation. This is the predicted sequence vector of unmodeled transverse dynamics bias, calculated from the radial basis function network. Optimal control increment sequence for: in, Let be the control increment sequence variable to be solved. To satisfy the feasible region constraint space of minimum turning radius, This involves transforming the control increment sequence into an inequality constraint mapping matrix for the absolute control quantity. These are the constraint boundary vectors.

7. The obstacle avoidance trajectory tracking control method for new energy vehicles according to claim 6, characterized in that, Step S5 specifically includes: Step S501: Call the optimal control increment sequence output by the numerical optimization solver in step S4, extract the front wheel angle increment corresponding to the first prediction step, and add it to the front wheel angle at the previous moment to calculate the optimal control command at the current moment. In step S502, the optimal control command is sent to the steering actuator of the new energy vehicle for lateral motion control; at the same time, the vehicle sensors collect the latest vehicle position coordinates, heading angle and yaw rate after the control command is executed in real time. Step S503: The latest vehicle status, including the latest vehicle position coordinates, heading angle and yaw rate, collected in step S502, is fed back to step S1. The initial error of trajectory tracking at the next moment is recalculated, and the prediction time domain is rolled forward one step over time to enter the prediction and optimization process of the next control cycle, thereby realizing vehicle trajectory tracking control.

8. The obstacle avoidance trajectory tracking control method for new energy vehicles according to claim 7, characterized in that, In step S501, the front wheel steering angle increment corresponding to the first prediction step size for: in, To extract row vectors from the control sequence, only the first element is extracted; The optimal front wheel steering angle at the current moment As the optimal control command, the expression is: in, The front wheel steering angle actually executed at the previous moment. This is the optimal front wheel steering angle control increment extracted from the optimization sequence for the first prediction step at the current moment; In step S503, the latest vehicle status satisfy: in, For real vehicle function mapping, This provides feedback on the actual state of the sensor at the current moment. This represents unmodeled environmental disturbances.

9. The obstacle avoidance trajectory tracking control method for new energy vehicles according to claim 8, characterized in that, In step S1, the set of hard constraints is turned. for: in, This refers to the vehicle's wheelbase.

Images