An autonomous vehicle road adaptive drift control system and method based on neural network dynamics

Abstract

This invention discloses a road adaptive drift control system and method for autonomous vehicles based on neural network dynamics. The control system proposed in this invention comprises five parts: a vehicle dynamics model, a feedforward neural network model, a model predictive control, a path following control, and a road classifier. The designed neural network replaces the traditional vehicle dynamics physical model, effectively overcoming the problem of insufficient accuracy of the traditional physical model in the tire nonlinear region, and significantly improving the trajectory tracking capability of autonomous vehicles under extreme conditions such as drift. The designed road classifier identifies the adhesion characteristics of different roads when the autonomous vehicle brakes, effectively improving the road adaptability of the autonomous vehicle and enhancing its ability to identify the adhesion characteristics of different road surfaces.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving for intelligent vehicles, and in particular to a control system and method for achieving drifting in autonomous vehicles using neural networks. Background Technology

[0002] With the application of artificial intelligence, the development of advanced sensors, and the improvement of computing power in vehicle computers, autonomous vehicles have become a new development direction for the global automotive industry. High-level autonomous vehicles will greatly enhance driving safety and comfort, while also significantly improving road efficiency and reducing emissions.

[0003] Currently, autonomous vehicles mainly consist of several parts: perception, planning, decision-making, and control. Vehicle control primarily focuses on trajectory tracking under low and medium speed conditions, with limited research on control under extreme conditions. Current trajectory tracking control is mainly based on kinematic and dynamic modeling. Kinematic modeling performs well at low speeds and on roads with high curvature, but its effectiveness deteriorates at medium and high speeds due to neglecting tire lateral slippage. Dynamic vehicle control also suffers significant performance degradation under extreme conditions as tires enter nonlinear regions. With the rapid development and application of neural networks, their excellent nonlinear fitting ability, generalization ability, and nonparametric modeling capabilities make neural network modeling a promising research direction. Furthermore, in autonomous vehicle control, compared to other control methods, the good control characteristics of model predictive control algorithms highly depend on accurate models. Therefore, using neural networks to establish accurate dynamic models will significantly improve the trajectory tracking ability of autonomous vehicles under extreme conditions.

[0004] To achieve good generalization ability of neural networks, especially on roads with complex road adhesion characteristics, a large number of training samples covering different road adhesion characteristics are needed for different road characteristics. At the same time, the perception module of autonomous vehicles also has difficulty in identifying road adhesion characteristics. All of these factors will reduce the application potential of neural networks. Therefore, it is necessary to further improve the road adaptability of autonomous vehicles and improve the vehicle's ability to identify road adhesion characteristics. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes an adaptive road drift control system for autonomous vehicles based on neural network dynamics, comprising five parts: a vehicle dynamics model, a feedforward neural network model, a model predictive control, a path following component, and a road classifier.

[0006] The control model in the model predictive controller proposed in this invention is based on a vehicle dynamics model. The model predictive controller acquires the actual state variables of the autonomous vehicle (actual longitudinal and lateral velocities, yaw rates, and angular velocities during vehicle driving) and the reference steady-state variable X of the autonomous vehicle. ss and reference steady-state control quantity U ss and steady-state lateral stiffness C λ,ss With steady-state longitudinal stiffness C α,ss The model predictive controller (MMC) outputs the control increment Δu by solving the problem based on the feedforward neural network output. The control increment Δu output by the MMC is compared with the reference steady-state control quantity u output by the feedforward neural network. ss The sums are used to obtain the actual control quantity.

[0007] The feedforward neural network model uses a driving simulator, IPG-Carmaker, and Simulink to construct a virtual environment. The driver acquires data by virtually drifting on two road surfaces with different radii: high adhesion coefficient (road surface A) and low adhesion coefficient (road surface B). The input to the feedforward neural network is the corrected road curvature output from the controller of the path-following section. The feedforward neural network outputs the steady-state lateral stiffness C, which consists of the desired body slip angle β during vehicle drift. λ,ss Steady-state longitudinal stiffness C α,ss Steady-state matrix X ss and steady-state control matrix U ss .

[0008] The path-following component takes as input the lateral deviation between the autonomous vehicle's position and the reference trajectory, and outputs the corrected road curvature. The corrected road curvature is used as one of the inputs to the feedforward neural network.

[0009] The road classifier proposed in this invention utilizes a neural network. A virtual environment is constructed using a driving simulator, IPG-Carmaker, and Simulink. A driver operates the simulator to perform braking operations on roads A and B to obtain data for network training. The road classifier selects either road A or road B that best matches the current road characteristics based on the road surface friction characteristics acquired during vehicle braking. Determining the road type determines the output of the feedforward neural network. When the road is road A, the output of the feedforward neural network is the reference steady-state state variable X. ssA Reference steady-state control quantity U ssA Steady-state lateral stiffness C α,ssA Steady-state longitudinal stiffness C λ,ssA When the path is road B, the output of the feedforward neural network is the reference steady-state state variable X. ssB Reference steady-state control quantity UssB Steady-state lateral stiffness C α,ssB Steady-state longitudinal stiffness C λ,ssB .

[0010] Based on the above-mentioned vehicle drift control system, this invention also proposes a vehicle drift control method based on a neural network, comprising:

[0011] S1: Establish the vehicle dynamics model, including:

[0012] Establish a dual-track dynamics model for the vehicle, specifically:

[0013] The dual-track dynamics model of the vehicle does not consider the vehicle's roll and pitch motions; the vehicle only moves within the xoy plane. The vehicle is front-wheel steering, and the vehicle's coordinate system lies within the vehicle's left-right symmetry plane. The origin, O, is the vehicle's center of mass. The x-axis is the vehicle's longitudinal axis, with its positive direction pointing towards the front. The positive z-axis is perpendicular to oxy and points upwards, while the y-axis points laterally, its positive direction following the right-hand rule. Based on Newton's laws, rotational equilibrium and force equilibrium equations are established at the vehicle's center of mass, yielding the following expressions:

[0014]

[0015] In the formula, m represents the vehicle mass, and v x ,v y These represent the vehicle's longitudinal and lateral speeds, respectively. Indicates yaw rate. I represents the angular velocity of the yaw. Ψ The moment of inertia of the vehicle about the z-axis, l f ,l r t represents the distance from the center of mass to the front and rear axes, respectively. wf ,t wr These represent the front and rear track widths, respectively. x,i ,F y,i These represent the longitudinal and lateral forces of the tire, respectively. The subscript i represents {front-left, front-right, rear-left, rear-right}, and δ represents the steering angle of the front wheels.

[0016] This invention employs a linear tire force formula. After performing a first-order Taylor expansion on the nonlinear tire force and neglecting the cross stiffness term, the following linear tire force formula is obtained:

[0017] F y,i ≈F yss,i +C α,i △α i

[0018] F x,i ≈F xss,i +C λ,i △λi

[0019] In the formula, C α,i C λ,i F represents the lateral and longitudinal stiffness of the tire, respectively. yss,i F xss,i Δα represents the steady-state tire forces in the lateral and longitudinal directions, respectively. i ,Δλ i These represent the disturbance changes at the horizontal and vertical equilibrium points, respectively.

[0020] Longitudinal slip ratio (λ) i In this invention, it is defined as follows:

[0021]

[0022] In the formula, w i The angular velocity of the wheel, r e V represents the radius of the wheel on the ground. xc,i This indicates the longitudinal velocity at the center of the wheel.

[0023] The wheel angular acceleration is expressed by the following formula:

[0024]

[0025] In the formula, I w This represents the moment of inertia of the wheel. The value of T represents the wheel's angular acceleration. i This indicates the wheel drive torque.

[0026] The velocity at the center of the wheel can be expressed by the following formula:

[0027]

[0028] In the formula, β represents the vehicle body slip angle, v x v y Representing the longitudinal and lateral velocities of the autonomous vehicle, respectively, the tire slip angle is expressed by the following formula:

[0029]

[0030] The control quantity (U), state quantity (X), and parameter (P) in this invention are as follows:

[0031] U={T i ,δ}

[0032]

[0033] P={m,I ψ ,l f ,l r ,t wf ,twr ,r e ,I w C λ,i C α,i}

[0034] S2: Build and train a feedforward neural network, which consists of an input layer, hidden layers, and an output layer. The calculation method of the neural network is shown in the following formula:

[0035]

[0036] a j =f(S) j )

[0037] In the formula, the output of the j-th neuron is determined by S. j It means, w ij b represents the weight from the i-th neuron in the previous layer to the j-th neuron in the current layer. j Let f represent the bias value of the j-th neuron, f represent the activation function, and a j This represents the output value of the j-th node, and the activation function of the hidden layer is chosen as sigmoid.

[0038] This invention utilizes a driving simulator, IPG-Carmaker, and Simulink to construct a virtual environment. The driver operates the simulator to perform virtual drifts on road surfaces A and B with different radii to acquire data. Neural networks are trained separately for two road surface types: high-friction coefficient (road surface A) and low-friction coefficient (road surface B). Each road surface type's trained neural network is further divided into four neural networks (N...). X N U N Cλ N Cα The input to each neural network is the desired vehicle body slip angle β and the corrected road curvature. The neural network outputs are respectively and , and the outputs of the neural network are used to replace the control quantity (U), state quantity (X), and vehicle lateral stiffness C in the aforementioned vehicle dynamics model. α,i and longitudinal stiffness C λ,i The four outputs of the neural network will be used in the model predictive controller to control the vehicle's front wheel steering angle and torque output;

[0039]

[0040] N X This indicates the output of the steady-state state variable X. ss Neural network, N U The neural network U used to output steady-state control quantities ss N Cλ This indicates the output steady-state lateral stiffness C.α,ss Neural network, N Cα This indicates the output of steady-state longitudinal stiffness C. λ,ss Neural networks.

[0041] The acquired data was divided into a 70% training set, a 15% validation set, and a 15% test set. X The neural network has 2 neurons in the input layer, 6 neurons in the hidden layer, and 4 neurons in the output layer. U The neural network has 2 neurons in the input layer, 6 neurons in the hidden layer, and 2 neurons in the output layer. Cλ N Cα The neural network has 2 neurons in the input layer, 4 neurons in the hidden layer, and 1 neuron in the output layer.

[0042] Neural networks (N) in road classifiers f Using a driving simulator, IPG-Carmaker, and Simulink to create a virtual environment, the driver operates the simulator to perform braking operations on roads A and B to obtain the vertical force F of the tires. y Longitudinal force F x And the longitudinal slip ratio λ, which will affect the longitudinal force F of the tire. x With vertical force F y The ratio μ and the slip ratio form a friction curve. The road classifier obtains 100 uniformly distributed μ values ​​from the friction curve, where μ = {μ1μ2…μ}. 100 The actual obtained μ value will be used as the neural network N in the road classifier. f The input data was divided into a 70% training set, a 15% validation set, and a 15% test set.

[0043] N f The neural network has 100 input neurons, 2 hidden neurons, and 1 output neuron.

[0044] S3: Design a path-following PID control algorithm, the control method of which is as follows:

[0045]

[0046] In the formula, k p k represents the proportionality coefficient. i k represents the integral coefficient. d The coefficients are represented by denoted by , k represents the reference path curvature, and when the lateral error e lat When Δk is positive, Δk is greater than 0. The value decreases, meaning the curvature of the current vehicle's trajectory is reduced to approach the reference trajectory, when the lateral error e latWhen it is negative, Δk is less than 0. The value of is increased, that is, the curvature of the current vehicle trajectory is increased to get closer to the reference trajectory. It is defined that when the vehicle's center of mass is on one side of the center of the reference path, it is positive, and otherwise it is negative.

[0047] S4: Design a road classifier that, by acquiring the current road adhesion characteristics, selects the feedforward neural network that best matches those characteristics to output the corresponding X. ss U ss C λ，ss C α,ss .

[0048] The road classifier first detects braking events by monitoring the pressure in the master cylinder to determine if braking has occurred. When braking occurs, it records the vertical force F of the tires. y Longitudinal force F x And the longitudinal slip ratio λ. After braking, a friction curve consisting of the normalized longitudinal force μ and the slip ratio is plotted. The normalized longitudinal force is expressed as:

[0049] μ = F x / F y

[0050] The road classifier obtains 100 uniformly distributed μ values ​​from the friction curve, where μ = {μ1μ2…μ}. 100 The actual obtained μ value will be used as the neural network N in the road classifier. f Given the input, the neural network will output the road type (road A or road B) that best matches the road attachment characteristics of the current input. When the road classifier determines that the current road attachment characteristics best match the characteristics of road A, the neural network (N... X N U N Cλ N Cα The output of ) is

[0051] X ssA U ssA C λ,ssA C λα,ssA If it conforms to the characteristics of road B, then the neural network (N) X N U N Cλ N Cα ) output

[0052] For X ssB U ssB C λ,ssB C α,ssB .

[0053] S5: Design a model predictive control algorithm to obtain X from a feedforward neural network. ss U ssC λ,ss C α,ss The front wheel steering angle and driving torque are obtained by online rolling optimization.

[0054] The vehicle dynamics model adopts a first-order approximation. Discretization is performed using the method described above, resulting in the following linear state-space model:

[0055] △x(k+1)=A ss △x(k)+B ss △u(k)

[0056] A in the formula ss and B ss Both are derived from the output of the neural network, with parameter C in parameter P. α and C λ It also comes from the output of the neural network, where the other values ​​in parameter P are all known quantities, and k represents the k-th time step. The state space passes through N... P Duration prediction, and future N c The state space of the control sequence input of duration can be represented by the following formula:

[0057] △X=F ss △x(k)+Φ ss △U

[0058] In the formula, ΔX, ΔU, F ss , Φ ss As shown in the following formula:

[0059] △X=[△x(k+1|k) T ,△x(k+2|k) T ,…,△x(k+N P |k) T ] T

[0060] △U=[△u(k) T ,△u(k) T ,…,△u(k+N c -1) T ] T

[0061]

[0062] The cost function J(ΔU) is defined as follows:

[0063] J(△U)=△U T H△U+2△x(k) T M T △U

[0064]

[0065] In the formula, and These are the weighting matrices used to determine the tracking error and input energy consumption, respectively.

[0066] Model predictive control calculates the N that minimizes the cost function J. c The control input in the time domain is shown in the following equation:

[0067]

[0068] In the formula, U min U max A represents the minimum and maximum values ​​of the control quantity, respectively. I The value is a lower triangular matrix with a value of 1. Therefore, the optimization problem can be solved by quadratic programming (QP) to obtain ΔU.

[0069] The sampling period T of the model prediction controller in this invention MPC Set to 0.02s, controlling the time domain N. c With a sampling period of 1, the prediction time domain is N. p It is 50 times the sampling period.

[0070] Beneficial effects of this invention:

[0071] 1. The neural network dynamics modeling method proposed in this invention effectively improves the accuracy of autonomous vehicle modeling compared with traditional physical models. The accurate model combined with the model predictive controller significantly improves the vehicle control capability, effectively reduces the lateral error of trajectory tracking, and especially improves the vehicle's control capability when the tires enter the nonlinear region under extreme conditions such as drift.

[0072] 2. The road classifier proposed in this invention effectively improves the road adaptability of autonomous vehicles under different road surfaces, significantly enhances the trajectory tracking ability of autonomous vehicles drifting under various road surface adhesion characteristics, and effectively overcomes the problem of insufficient generalization ability of a single neural network when facing complex road characteristics, reduces the amount of data for training the neural network, and also improves the ability of autonomous vehicles to identify road adhesion characteristics.

[0073] 3. The designed model predictive controller optimizes the output control increment ΔU and the control quantity U output by the feedforward neural network. ss By combining the results, the final control quantity is obtained. The model predictive controller can effectively compensate for the control quantity of the feedforward neural network, thereby making up for the deviation between the output control quantity of the feedforward neural network and the actual expected control quantity of the reference trajectory, so that the autonomous vehicle can drive along the reference trajectory.

[0074] 4. The designed path following control section can correct the driving curvature of the autonomous vehicle based on the deviation between the current autonomous vehicle and the reference path, thereby effectively improving the path tracking capability of the autonomous vehicle and reducing the deviation between the autonomous vehicle trajectory and the reference trajectory. Attached Figure Description

[0075] Figure 1 A dual-track nonlinear model for autonomous vehicles;

[0076] Figure 2 For road classifier modules;

[0077] Figure 3 For PID path following module;

[0078] Figure 4 It is a feedforward neural network structure;

[0079] Figure 5 For data acquisition and training modules;

[0080] Figure 6 For drift control systems of autonomous vehicles. Detailed Implementation

[0081] The invention will now be further described with reference to the accompanying drawings.

[0082] Figure 1 For the dual-track nonlinear model of autonomous vehicles, the following assumptions were made when establishing the vehicle dynamics model:

[0083] (1) Assuming the vehicle is traveling on a flat road, we only consider the lateral and longitudinal motion of the vehicle and ignore the vertical motion of the vehicle.

[0084] (2) Assume that the vehicle's suspension system is a rigid body and ignore the motion of the suspension and its influence on the coupling relationship.

[0085] (3) Ignore the influence of the lateral and longitudinal coupling relationship of the vehicle's tires on the vehicle's steady-state drift.

[0086] (4) Ignore the lateral load transfer and longitudinal load transfer of the vehicle.

[0087] (5) The effect of air resistance on the steady-state characteristics of the vehicle during drift is not considered.

[0088] Based on the above assumptions, the vehicle only moves within the xoy plane. The vehicle is front-wheel steering, and the vehicle's coordinate system lies within the plane of symmetry. The origin of the vehicle's center of mass is O, the x-axis is the vehicle's longitudinal axis with its positive direction pointing towards the front, the positive z-axis is perpendicular to oxy and upwards, and the y-axis points laterally towards the vehicle, its positive direction following the right-hand rule. According to Newton's laws, the rotational equilibrium and force equilibrium equations are established at the vehicle's center of mass, yielding the following expression:

[0089]

[0090] In the formula, m represents the vehicle mass, and v x ,v y These represent the vehicle's longitudinal and lateral speeds, respectively. Indicates yaw rate. I represents the yaw acceleration. Ψ The moment of inertia of the vehicle about the z-axis, l f ,l r t represents the distance from the center of mass to the front and rear axes, respectively. wf ,t wr These represent the front and rear track widths, respectively. x,i ,F y,i These represent the longitudinal and lateral forces of the tire, respectively. The subscript i takes the values ​​fl, fr, rl, and rr, respectively, representing the front left, front right, rear left, and rear right. δ represents the steering angle of the front wheels.

[0091] After performing a first-order Taylor expansion on the nonlinear tire force and neglecting the cross stiffness term, the following linear tire force formula is obtained:

[0092] F y,i ≈F yss,i +C α,i △α i

[0093] F x,i ≈F xss,i +C λ,i △λ i

[0094] In the formula, C α,i C λ,i F represents the lateral and longitudinal stiffness of the tire, respectively. yss,i ,F xss,i Δα represents the steady-state tire forces in the lateral and longitudinal directions, respectively. i ,Δλ i These represent the disturbance changes at the horizontal and vertical equilibrium points, respectively. Where C... α,i C λ,i The value will be directly estimated by the neural network.

[0095] Longitudinal slip ratio (λ) i In this invention, it is defined as follows:

[0096]

[0097] In the formula, w i The angular velocity of the wheel, r e V represents the radius of the wheel on the ground. xc,i This indicates the longitudinal velocity at the center of the wheel.

[0098] The wheel angular acceleration is expressed by the following formula:

[0099]

[0100] In the formula, I w This represents the moment of inertia of the wheel. The value of T represents the wheel's angular acceleration. i This indicates the front wheel drive torque of the wheel.

[0101] The velocity at the center of the wheel can be expressed by the following formula:

[0102]

[0103] In the formula, β represents the vehicle body slip angle, v x v y Representing the longitudinal and lateral velocities of the autonomous vehicle, respectively, the tire slip angle is expressed by the following formula:

[0104]

[0105] Figure 2 This is the road classifier module. This module selects the neural network that best matches the current road adhesion characteristics to output the corresponding X. ss U ss C λ，ss C α,ss .

[0106] The road classifier first detects braking events by monitoring the pressure in the master cylinder to determine if braking has occurred. When braking occurs, it records the vertical force F of the tires. y Longitudinal force F x And the longitudinal slip ratio λ. After braking, a friction curve is plotted with the normalized longitudinal force μ as the abscissa and the longitudinal slip ratio λ as the ordinate. The normalized longitudinal force is expressed as:

[0107] μ = F x / F y

[0108] The road classifier obtains 100 uniformly distributed μ values ​​from the friction curve. The actual obtained μ = {μ1, μ2…μ} 100 The value will be used as the neural network (N) in the road classifier. f The input is the friction curve, from which the corresponding longitudinal slip ratio λ is obtained. Based on the longitudinal slip ratio λ, the neural network will output the road type (road A or road B) that best matches the road adhesion characteristics of the current input. When the road classifier determines that the current road characteristics best match the characteristics of road A, the neural network (N)... X N U NCλ N Cα The output of ) is X ssA U ssA C λ,ssA C α,ssA If it conforms to the characteristics of road B, then the neural network (N) X N U N Cλ N Cα The output of ) is X ssB U SSB C λ,ssB C α,ssB .

[0109] Figure 3 This is a PID path following module. Path tracking uses PID control, and its control method is as follows:

[0110]

[0111] In the formula, k p k represents the proportionality coefficient. i k represents the integral coefficient. d The coefficients are represented by denoted by , k represents the reference path curvature, and when the lateral error e lat When Δk is positive, Δk is greater than 0. The value decreases, meaning the curvature of the current vehicle's trajectory is reduced to approach the reference trajectory, when the lateral error e lat When it is negative, Δk is less than 0. The value of is increased, that is, the curvature of the current vehicle trajectory is increased to get closer to the reference trajectory. It is defined that when the vehicle's center of mass is on one side of the center of the reference path, it is positive, and otherwise it is negative.

[0112] Figure 4 This is a feedforward neural network structure, which consists of an input layer, hidden layers, and an output layer. The computation method of the neural network is shown in the following formula:

[0113]

[0114] a j =f(S) j )

[0115] In the formula, the output of the j-th neuron is determined by S. j It means, w ij b represents the weight from the i-th neuron in the previous layer to the j-th neuron in the current layer. j Let f represent the bias value of the j-th neuron, f represent the activation function, and a j This represents the output value of the j-th node, and the activation function of the hidden layer is chosen as sigmoid.

[0116] Figure 5 For the data acquisition and training module, a virtual environment is constructed using a driving simulator, IPG-Carmaker, and Simulink. The driver performs virtual drifts on road surfaces A and B with different radii by operating the driving simulator to acquire data. Neural networks are trained separately for two road surface types: high-friction coefficient (road surface A) and low-friction coefficient (road surface B). Each road surface type's trained neural network is further divided into four neural networks (N...). X N U N Cλ N Cα The input to each neural network is the desired vehicle body slip angle β and the corrected road curvature. The neural network outputs are X ssA U ssA C λ,ssA C α,ssA and X ssB U ssB C λssB C αssB The output of the neural network is used to replace the control variables (U), state variables (X), and vehicle lateral stiffness C in the aforementioned vehicle dynamics model. α,i and longitudinal stiffness C λ,i The four outputs of the neural network will be used in the model predictive controller to control the vehicle's front wheel steering angle and torque output.

[0117]

[0118] N X This indicates the output of the steady-state state variable X. ss Neural network, N U Indicates the steady-state control quantity U used for output. ss Neural network, N Cλ This indicates the output steady-state lateral stiffness C. α,ss Neural network, N Cα This indicates the output of steady-state longitudinal stiffness C. λ,ss Neural networks.

[0119] The acquired data was divided into a 70% training set, a 15% validation set, and a 15% test set. X The neural network has 2 neurons in the input layer, 6 neurons in the hidden layer, and 4 neurons in the output layer. U The neural network has 2 neurons in the input layer, 6 neurons in the hidden layer, and 2 neurons in the output layer. Cλ N Cα The neural network has 2 neurons in the input layer, 4 neurons in the hidden layer, and 1 neuron in the output layer.

[0120] Neural networks (N) in road classifiers f Using a driving simulator, IPG-Carmaker, and Simulink, a virtual environment was built where drivers operated the simulator to perform braking maneuvers on roads A and B to acquire data. The acquired data was divided into a 70% training set, a 15% validation set, and a 15% test set.

[0121] N f The input to the neural network is the μ value collected during vehicle braking, μ = {μ1, μ2, ..., μ3}. n The output value is either road A or road B. The input layer has 100 neurons, the hidden layer has 2 neurons, and the output layer has 1 neuron. The normalized longitudinal force μ is shown in the following formula:

[0122] μ = F x / F y

[0123] In the formula F x F represents the longitudinal force on the tire during braking. y This indicates the vertical force exerted on the tire during braking.

[0124] Figure 6 For the drift control system of autonomous vehicles, the path tracking controller outputs the corrected reference path curvature combined with the vehicle body sideslip angle expected for drift as the input of the neural network. The road classifier detects braking and obtains the current road adhesion characteristics to select the output of the neural network that best matches the current road characteristics and uses it as the input of the model prediction controller.

[0125] The vehicle dynamics model adopts a first-order approximation. Discretization is performed using the method described above, resulting in the following linear state-space model:

[0126] △x(k+1)=A ss △x(k)+B ss △u(k)

[0127] In the formula, the system matrix Ass and the control input matrix Bss are both derived from the output of the neural network. Ass and Bss are matrices obtained by discretizing the vehicle dynamics model. The parameter C in parameter P... α and C λ It also comes from the output of the neural network, where the other values ​​in parameter P are all known quantities, and k represents the k-th time step. The state space passes through N... P Duration prediction, and future N c The state space of the control sequence input of duration can be represented by the following formula:

[0128] △X=F ss △x(k)+Φss △U

[0129] In the formula, ΔX, ΔU, F ss , Φ ss As shown in the following formula:

[0130] △X=[△x(k+1|k) T ,△x(k+2|k) T ,…,△x(k+N P |k) T ] T

[0131] △U=[△u(k) T ,△u(k) T ,…,△u(k+N c -1) T ] T

[0132]

[0133] The cost function J(ΔU) is defined as follows:

[0134] J(△U)=△U T H△U+2△x(k) T M T △U

[0135]

[0136] In the formula, and These are weighted matrices used to determine the relative importance of tracking error and input energy consumption, respectively.

[0137] Model predictive control calculates the N that minimizes the cost function J. c The control input in the time domain is shown in the following equation:

[0138]

[0139] In the formula, A I U represents a lower triangular matrix with a value of 1. min U max Let represent the minimum and maximum values ​​of the control variable, respectively. Therefore, this optimization problem can be solved using Quadratic Programming (QP) to obtain ΔU. The sampling period T of the model predictive controller... MPC Set to 0.02s, controlling the time domain N. c With a sampling period of 1, the prediction time domain is N. p It is 50 times the sampling period.

[0140] The calculated ΔU represents the front wheel steering angle and driving torque applied to the autonomous vehicle to maintain the desired body slip angle and the curvature of the reference trajectory. The PID controller then outputs a curvature correction Δk based on the actual curvature of the autonomous vehicle's trajectory to obtain the corrected curvature.

[0141] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent methods or modifications that do not depart from the technology of the present invention should be included within the scope of protection of the present invention.

Claims

1. A road adaptive drift control system for autonomous vehicles based on neural network dynamics, characterized in that, It includes a path following component, a road classifier component, a feedforward neural network model component, and a model prediction and control component. The path-following component takes as input the lateral deviation between the autonomous vehicle's position and the reference trajectory, and outputs the corrected road curvature. The resulting corrected road curvature As input to the feedforward neural network; The road classifier selects either road A or road B that best matches the current road characteristics based on the road surface friction characteristics obtained during vehicle braking. Road A is a high-adhesion coefficient road surface, and road B is a low-adhesion coefficient road surface. When road A is selected, the output of the feedforward neural network is the reference steady-state state variable X. ssA Reference steady-state control quantity U ssA Steady-state lateral stiffness C α,ssA Steady-state longitudinal stiffness C λ,ssA When the path is road B, the output of the feedforward neural network is the reference steady-state state variable X. ssB Reference steady-state control quantity U ssB Steady-state lateral stiffness C α,ssB Steady-state longitudinal stiffness C λ,ssB ; The input to the feedforward neural network model includes the corrected road curvature output from the path-following component. And the desired body slip angle β during vehicle drift, the output of the feedforward neural network is the steady-state slip stiffness C. λ,ss Steady-state longitudinal stiffness C α,ss Steady-state matrix X ss and steady-state control matrix U ss When the road classifier selects road A, C λ,ss =C λ,ssA C α,ss =C α,ssA X ss =X ssA U ss =U ssA When the road classifier selects road B, C λ,ss =C λ,ssB C α,ss =C α,ssB X ss =X ssB U ss =U ssB ; The model predictive controller takes as input the actual state variables of the autonomous vehicle and the reference steady-state variable X of the autonomous vehicle as output from the feedforward neural network. ss Reference steady-state control quantity U ss and steady-state lateral stiffness C λ,ss With steady-state longitudinal stiffness C α,ss The model predicts the controller output control increment Δu, which is compared with the reference steady-state control quantity u output by the feedforward neural network. ss The sum of these values ​​yields the actual front wheel steering angle and torque output control value. The model predictive controller is established based on the vehicle's dual-track dynamics model, as detailed below: With the vehicle's center of mass as the origin o, the x-axis as the vehicle's longitudinal axis with its positive direction pointing towards the front of the vehicle, the positive z-axis perpendicular to oxy and pointing upwards, and the y-axis pointing laterally towards the vehicle, its positive direction satisfying the right-hand rule, and according to Newton's laws, establishing rotational equilibrium and force equilibrium equations at the vehicle's center of mass, we obtain the following expression: In the formula, m represents the vehicle mass, and v x ,v y These represent the vehicle's longitudinal and lateral speeds, respectively. Indicates yaw rate. I represents the angular velocity of the yaw. Ψ The moment of inertia of the vehicle about the z-axis, l f ,l r t represents the distance from the center of mass to the front and rear axes, respectively. wf ,t wr These represent the front and rear track widths, respectively. x,i ,F y,i These represent the longitudinal and lateral forces of the tire, respectively, with the subscript i taking the values ​​fl, fr, rl, and rr, respectively, and δ representing the steering angle of the front wheels; By applying the linear tire force formula and performing a first-order Taylor expansion on the nonlinear tire force while neglecting the cross stiffness term, the following linear tire force formula is obtained: F y,i ≈F yss,i +C α,i △α i F x,i ≈F xss,i +C λ,i △λ i In the formula, C α,i C λ,i F represents the lateral and longitudinal stiffness of the tire, respectively. yss,i F xss,i Δα represents the steady-state tire forces in the lateral and longitudinal directions, respectively. i ,Δλ i These represent the disturbance changes at the horizontal and vertical equilibrium points, respectively; Define longitudinal slip ratio (λ) i ): In the formula, w i The angular velocity of the wheel, r e V represents the radius of the wheel on the ground. xc,i Indicates the longitudinal velocity at the center of the wheel; The wheel angular acceleration is expressed by the following formula: In the formula, I w This represents the moment of inertia of the wheel. The value of T represents the wheel's angular acceleration. i The wheel drive torque is indicated; the speed at the wheel center is expressed by the following formula: In the formula, the vehicle body side slip angle The four tire slip angles are represented by the following formula:

2. The road adaptive drift control system for autonomous vehicles based on neural network dynamics according to claim 1, characterized in that, The path following part adopts a PID control algorithm, and its control method is as follows: In the formula, k represents the curvature of the reference path, when the lateral error e lat When Δk is positive, Δk is greater than 0. The value decreases, meaning the curvature of the current vehicle's trajectory is reduced to approach the reference trajectory, when the lateral error e lat When it is negative, Δk is less than 0. The value of is increased, that is, the curvature of the current vehicle trajectory is increased to get closer to the reference trajectory. It is defined that when the vehicle's center of mass is on one side of the center of the reference path, it is positive, and otherwise it is negative.

3. The road adaptive drift control system for autonomous vehicles based on neural network dynamics according to claim 1, characterized in that, The road classifier comprises a neural network with 100 input neurons, 2 hidden neurons, and 1 output neuron. A virtual environment is constructed using a driving simulator, IPG-Carmaker, and Simulink, and the driver performs braking operations on roads A and B by operating the driving simulator to obtain training data.

4. The road adaptive drift control system for autonomous vehicles based on neural network dynamics according to claim 3, characterized in that, The road classifier detects braking events by monitoring the pressure in the master cylinder to determine whether braking has occurred, and records the vertical force F of the tires when braking occurs. y Longitudinal force F x Given the longitudinal slip ratio λ, after braking, a friction curve consisting of the normalized longitudinal force μ and the slip ratio λ is plotted. The normalized longitudinal force is expressed as: μ=F x / F y The road classifier obtains the actual μ value, which is used as the neural network (N) in the road classifier. f The neural network takes the friction curve as input and outputs the corresponding slip ratio. Based on the slip ratio, it obtains the road type that best matches the road adhesion characteristics of the current input: road A or road B.

5. The road adaptive drift control system for autonomous vehicles based on neural network dynamics according to claim 1, characterized in that, The feedforward neural network model includes an input layer, a hidden layer, and an output layer, and its calculation method is shown in the following formula: a j =f(S j ) In the formula, the output of the j-th neuron is determined by S. j It means, w ij b represents the weight from the i-th neuron in the previous layer to the j-th neuron in the current layer. j Let f represent the bias value of the j-th neuron, f represent the activation function, and a j This represents the output value of the j-th node, and the activation function of the hidden layer is chosen as sigmoid; The feedforward neural network uses a driving simulator, IPG-Carmaker, and Simulink to construct a virtual environment. The driver obtains data by virtually drifting on road surfaces A and B with different radii through the driving simulator. Neural networks are trained separately for road surface A (high adhesion coefficient) and road surface B (low adhesion coefficient). Each type of road surface's trained neural network is further divided into four neural networks N. X N U N Cλ N Cα The input to each neural network is the desired vehicle body slip angle β and the corrected road curvature. The neural network outputs are X ssA U ssA C λ,ssA C α,ssA and X ssB U ssB C λssB C αssB The four outputs of the neural network will be used in the model predictive controller to control the front wheel steering angle and torque output of the vehicle; Where, N X The neural network has 2 neurons in the input layer, 6 neurons in the hidden layer, and 4 neurons in the output layer; N U The neural network has 2 neurons in the input layer, 6 neurons in the hidden layer, and 2 neurons in the output layer; N Cλ , and N Cα The neural network has 2 neurons in the input layer, 4 neurons in the hidden layer, and 1 neuron in the output layer.

6. The road adaptive drift control system for autonomous vehicles based on neural network dynamics according to claim 1, characterized in that, It also includes defining the control variable U, the state variable X, and the parameter P, as shown below: U={T i ,δ} P={m,I ψ ,l f ,l r ,t wf ,t wr ,r e ,I w ,C λ,i ,C α,i }。 7. The road adaptive drift control system for autonomous vehicles based on neural network dynamics according to claim 6, characterized in that, The prediction algorithm of the model prediction controller includes the following: The vehicle dynamics model adopts a first-order approximation. Discretization is performed using the method described above, resulting in the following linear state-space model: △x(k+1)=A ss △x(k)+B ss △u(k) The state matrix A in the formula ss and control matrix B ss Both are derived from the output of the neural network, with parameter C in parameter P. λ,i C α,i It also comes from the output of the neural network. The other values ​​in the parameter P are all known quantities. In the formula, k represents the state space after N times. P Duration prediction, and future N c The duration of the control sequence input, the state space is represented by the following formula: △X=F ss △x(k)+Φ ss △U In the formula, ΔX, ΔU, F ss , Φ ss As shown in the following formula: △X=[△x(k+1|k) T ,△x(k+2|k) T ,…,△x(k+N P |k) T ] T △U=[△u(k) T ,△u(k) T ,…,△u(k+N c -1) T ] T The cost function J(ΔU) is defined as follows: J(△U)=△U T H△U+2△x(k) T M T △U In the formula, and These are the weighting matrices used to determine the tracking error and input energy consumption, respectively. Model predictive control calculates the N that minimizes the cost function J. c The control input in the time domain is shown in the following equation: In the formula, U min U max Let represent the minimum and maximum values ​​of the control quantity, respectively. Therefore, the optimization problem can be solved by quadratic programming to obtain ΔU.

8. The road adaptive drift control system for autonomous vehicles based on neural network dynamics according to claim 7, characterized in that, The sampling period T of the model predictive controller MPC Set to 0.02s, controlling the time domain N. c The value is 1, and the prediction time domain is N. p It is 50.

9. A road adaptive drift control method for autonomous vehicles based on neural network dynamics, characterized in that, S1: Establish a dual-track dynamics model for the vehicle, including: The dual-track dynamics model of the vehicle does not consider the vehicle's roll and pitch motions; the vehicle only moves in the xoy plane. The vehicle is front-wheel steering, and the vehicle coordinate system lies in the vehicle's left-right symmetry plane. The origin of the vehicle's center of mass is o, the x-axis is the vehicle's longitudinal axis with its positive direction pointing towards the front, the z-axis is perpendicular to oxy and pointing upwards, and the y-axis points laterally towards the vehicle, its positive direction satisfying the right-hand rule. According to Newton's laws, rotational equilibrium and force equilibrium equations are established at the vehicle's center of mass, resulting in the following expression: In the formula, m represents the vehicle mass, and v x ,v y These represent the vehicle's longitudinal and lateral speeds, respectively. Indicates yaw rate. I represents the angular velocity of the yaw. Ψ The moment of inertia of the vehicle about the z-axis, l f ,l r t represents the distance from the center of mass to the front and rear axes, respectively. wf ,t wr These represent the front and rear track widths, respectively. x,i ,F y,i These represent the longitudinal and lateral forces of the tire, respectively, with the subscript i taking the values ​​fl, fr, rl, and rr, respectively, and δ representing the steering angle of the front wheels; By applying the linear tire force formula and performing a first-order Taylor expansion on the nonlinear tire force while neglecting the cross stiffness term, the following linear tire force formula is obtained: F y,i ≈F yss,i +C α,i △α i F x,i ≈F xss,i +C λ,i △λ i In the formula, C α,i C λ,i F represents the lateral and longitudinal stiffness of the tire, respectively. yss,i F xss,i Δα represents the steady-state tire forces in the lateral and longitudinal directions, respectively. i ,Δλ i These represent the disturbance changes at the horizontal and vertical equilibrium points, respectively; The longitudinal slip ratio λ is defined as follows: In the formula, w i The angular velocity of the wheel, r e V represents the radius of the wheel on the ground. xc,i This indicates the longitudinal velocity at the center of the wheel. The wheel angular acceleration is expressed by the following formula: In the formula, I w This represents the moment of inertia of the wheel. The value of T represents the wheel's angular acceleration. i Indicates the wheel drive torque; The velocity at the center of the wheel is expressed by the following formula: In the formula, the vehicle body side slip angle The four tire slip angles are represented by the following formula: The control variable U, the state variable X, and the intermediate parameter P are designed as follows: U={T i ,δ} P={m,I ψ ,l f ,l r ,t wf ,t wr ,r e ,I w ,C λ,i ,C α,i } S2: Establish a feedforward neural network, which consists of an input layer, hidden layers, and an output layer. The calculation method of the neural network is shown in the following formula: a j =f(S j ) In the formula, the output of the j-th neuron is determined by S. j It means, w ij b represents the weight from the i-th neuron in the previous layer to the j-th neuron in the current layer. j Let f represent the bias value of the j-th neuron, f represent the activation function, and a j This represents the output value of the j-th node, and the activation function of the hidden layer is chosen as sigmoid; Using a driving simulator, IPG-Carmaker, and Simulink construct a virtual environment. Drivers manipulate the simulator to perform virtual drifts on road surfaces A and B with different radii to acquire data. Neural networks are trained separately for road surface A (high adhesion coefficient) and road surface B (low adhesion coefficient). Each road surface type's trained neural network is further divided into four neural networks N. X N U N Cλ N Cα The input to each neural network is the desired vehicle body slip angle β and the corrected road curvature. The neural network outputs are X ssA U ssA C λ,ssA C α,ssA and X ssB U ssB C λssB C αssB The four outputs of the neural network will be used in the model predictive controller to control the front wheel steering angle and torque output of the vehicle; N X The neural network has 2 neurons in the input layer, 6 neurons in the hidden layer, and 4 neurons in the output layer. N U The neural network has 2 neurons in the input layer, 6 neurons in the hidden layer, and 2 neurons in the output layer. N Cλ N Cα The neural network has 2 neurons in the input layer, 4 neurons in the hidden layer, and 1 neuron in the output layer; Neural network N in road classifier f By using a driving simulator, IPG-Carmaker, and Simulink to build a virtual environment, drivers operate the driving simulator to perform braking operations on roads A and B to obtain training data. f The neural network has 100 input neurons, 2 hidden neurons, and 1 output neuron. S3: Design a path-following PID control algorithm: In the formula, k represents the curvature of the reference path, when the lateral error e lat When Δk is positive, Δk is greater than 0. The value decreases, meaning the curvature of the current vehicle's trajectory is reduced to approach the reference trajectory, when the lateral error e lat When it is negative, Δk is less than 0. The value of is increased, that is, the curvature of the current vehicle trajectory is increased to get closer to the reference trajectory. It is defined that when the vehicle's center of mass is on one side of the center of the reference path, it is positive, and vice versa. S4: Design a road classifier that, by acquiring the current road adhesion characteristics, selects the feedforward neural network that best matches those characteristics to output the corresponding X. ss U ss C λ，ss C α,ss ; The road classifier detects braking events by monitoring the pressure in the master cylinder to determine if braking has occurred, and records the vertical force F of the tires when braking occurs. y Longitudinal force F x Given the longitudinal slip ratio λ, after braking, a friction curve consisting of the normalized longitudinal force μ and the slip ratio is plotted. The normalized longitudinal force is expressed as: μ=F x / F y The road classifier obtains a series of uniformly distributed μ values ​​from the friction curve, μ = {μ1, μ2, ..., μ3}. n The actual μ value obtained will be used as the neural network N in the road classifier, along with the corresponding slip ratio value. f Given the input, the neural network will output either road A or road B that best matches the current input's road attachment characteristics. When the road classifier determines that the current road attachment characteristics best match those of road A, the neural network N... X N U N Cλ N Cα The output is X ssA U ssA C λ,ssA C λα,ssA If it conforms to the characteristics of road B, then the neural network N X N U N Cλ N Cα of The output is X ssB U ssB C λ,ssB C α,ssB ; S5: Design Model Predictive Control Algorithm: Use a first-order approximation of the vehicle dynamics model. Discretization is performed using the method described above, resulting in the following linear state-space model: △x(k+1)=A ss △x(k)+B ss △u(k) A in the formula ss and B ss Both are derived from the output of the neural network, with parameter C in parameter P. α and C λ It also comes from the output of the neural network. The other values ​​in the parameter P are all known quantities. In the formula, k represents the state space after N times. P Duration prediction, and future N c The duration of the control sequence input, the state space is represented by the following formula: △X=F ss △x(k)+Φ ss △U In the formula, ΔX, ΔU, F ss Φ ss As shown in the following formula: △X=[△x(k+1|k) T ,△x(k+2|k) T ,…,△x(k+N P |k) T ] T △U=[△u(k) T ,△u(k) T ,…,△u(k+N c -1) T 】 T The cost function J(ΔU) is defined as follows: J(△U)=△U T H△U+2△x(k) T M T △U In the formula, and These are the weighting matrices used to determine the tracking error and input energy consumption, respectively. Model predictive control calculates the N that minimizes the cost function J. c The control input in the time domain is shown in the following equation: In the formula, U min U max Let represent the minimum and maximum values ​​of the control quantity, respectively. Therefore, the optimization problem is solved using quadratic programming to obtain ΔU, which is then compared with the reference steady-state control quantity u output by the feedforward neural network. ss The sums are used to obtain the actual control quantity.

Images