A mechanism-data hybrid-driven autonomous driving control method, system, and medium

Abstract

The present invention relates to a mechanism-data hybrid-driven autonomous driving control method, system, and medium. The method comprises: controlling a vehicle to output a preset trajectory and speed using a model predictive controller driven by a sampling frequency and a pure mechanism model, which serves as an offline dataset; constructing a residual model based on a Gaussian process regression algorithm, training a residual model of the mechanism model based on the offline dataset, establishing a prediction model for the model predictive controller based on the mechanism model and the residual model, and outputting a total driving force requirement; utilizing a low-order PID controller to obtain additional vertical force, additional roll moment, and additional pitch moment; constructing a torque allocation matrix based on the total driving force requirement, additional vertical force, additional roll moment, and additional pitch moment using vehicle structural parameters, and calculating the motor torque provided by each of the four wheel hub motors. Compared with the prior art, the present invention effectively suppresses the impact of external disturbances on tracking performance, achieving more intelligent and precise autonomous driving control.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a mechanism-data hybrid driven autonomous driving control method, system, and medium. Background Art

[0002] Autonomous driving has been a major research hotspot in academia and industry in recent years and is considered one of the key development directions for future vehicles. As the underlying layer of autonomous vehicles, the control system is responsible for receiving data from sensors, analyzing environmental information, and determining the vehicle's direction, speed, and other maneuvers. It directly determines the vehicle's driving performance. Currently, public acceptance of autonomous vehicles driven by four-wheel hub motors is limited, largely due to the fact that the control variables calculated by the autonomous driving control system cannot balance the six key performance indicators of the vehicle and cannot adapt to time-varying conditions such as unknown random road disturbances. Therefore, enabling adaptive and self-learning autonomous driving control systems driven by four-wheel hub motors can further expand the autonomous vehicle market and promote their development.

[0003] With the advancement of computer hardware and improved computational efficiency, model predictive control (MPC) algorithms have gradually become the mainstream of intelligent control algorithms for autonomous vehicles. Based on the form of the predictive model, MPC algorithms can be broadly categorized into three types: purely mechanism-driven, purely data-driven, and hybrid data-mechanism-driven. With the development of artificial intelligence (AI), the latter two have become research hotspots. The control performance of purely mechanism-driven algorithms depends entirely on the accuracy of the simplified mechanism model. Due to the conflict between modeling accuracy and model complexity, purely mechanism-driven models can only guarantee basic control performance and are unable to adapt to time-varying conditions such as random road disturbances. When the mismatch between the mechanism model and the real system is significant, the purely mechanism-driven model may even lose its ability to control the autonomous vehicle system, resulting in inestimable losses. Purely data-driven algorithms possess the ability to adapt to operating conditions, but their control performance depends entirely on the completeness of the training dataset's coverage of the vehicle's feasible state space. This requires a high-quality human driving dataset, which is difficult and expensive to obtain. Furthermore, the predictive models of purely data-driven algorithms are black-box models, making them difficult for human drivers to trust. Therefore, there is currently a lack of a model predictive control algorithm that can adapt to time-varying conditions such as random road disturbances and does not rely entirely on the quality of the data set to achieve autonomous driving control. Summary of the Invention

[0004] The purpose of the present invention is to provide an autonomous driving control method, system and medium driven by a hybrid mechanism data, which uses human driving data to guide the self-learning of the residual model to achieve autonomous driving control with adaptive speed, trajectory tracking control and vehicle body posture adaptive adjustment under complex time-varying working conditions.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A mechanism-data hybrid driven autonomous driving control method comprises the following steps:

[0007] Offline dataset construction: Based on the sampling frequency and the model predictive controller driven by the pure mechanism model, the vehicle is controlled to output the preset trajectory and speed, which is used as the offline dataset;

[0008] Offline training of the residual model: A residual model is constructed based on the Gaussian process regression algorithm. Based on the offline data set, a residual model of the mechanism model is trained. Based on the mechanism model and the residual model, a prediction model of the model predictive controller is established to output the total driving force demand.

[0009] Low-level driving torque distribution: A low-order PID controller is used to obtain the additional vertical force, additional roll moment, and additional pitch moment. Based on the total driving force requirement, additional vertical force, additional roll moment, and additional pitch moment, the torque distribution matrix is ​​constructed using vehicle structural parameters to calculate the motor torque provided by each of the four wheel hub motors.

[0010] In the offline dataset construction step, a sampling time equal to the single-step prediction duration of the model predictive controller is used to collect measurement data of the four-wheel hub motor driven autonomous driving vehicle under test conditions, thereby obtaining a sequence dataset containing the state and control quantities of the autonomous driving vehicle system.

[0011] The offline dataset is constructed as follows: a sampling frequency t is preset, and under test conditions, a purely mechanism-driven model predictive controller is used to control the vehicle to track the preset trajectory and speed, and to collect the measured values ​​of the vehicle state and control variables:

[0012]

[0013] in represents the vehicle longitudinal velocity, represents the lateral speed of the vehicle, represents the vehicle's yaw rate, represents the front wheel angle, Indicates the total driving torque;

[0014] The mechanism model is used to calculate the predicted value of the vehicle state at the next sampling moment based on the current state and control input:

[0015]

[0016] in, represents the predicted value of the vehicle longitudinal velocity, represents the predicted value of the vehicle's lateral velocity, represents the predicted value of the vehicle's yaw rate, represents the predicted value of the front wheel angle, Indicates the predicted value of total driving torque;

[0017] Define the prediction error of the mechanism model at the current moment

[0018] The offline dataset is constructed as follows:

[0019]

[0020] in, is the feature data, e is the label data, its subscript represents the data number, and m is the total number of samples in the data set.

[0021] Based on the Gaussian process regression algorithm, the residual model learning mathematical model of the mechanism model is established:

[0022]

[0023]

[0024] Among them, [Y].a represents the a-th column of matrix Y, which consists of all output vectors y k The ath dimension of ; I represents an m×m identity matrix, represents the variance of the measurement noise in the ath dimension; Represents a real number, obtained by the kernel function: in, and L a is the hyperparameter in the kernel function; Represents an m-dimensional row vector, through calculate; Represents an m×m symmetric matrix, through Calculation; m is the total number of samples in the data set, z, All are samples from the dataset.

[0025] The mathematical model is learned through the residual model of the mechanism model to estimate each dimension of the output vector. When the input is z, the estimated value of the residual of the mechanism model is:

[0026] in, n d is the dimension of the output vector.

[0027] The maximum likelihood estimation method is used to calculate the hyperparameters in the residual model learning mathematical model to obtain the residual model of the mechanism model. The mechanism model and the residual model are linearly superimposed to obtain the prediction model of the model predictive controller.

[0028] The objective function and constraints of the prediction model of the model predictive controller are expressed as:

[0029]

[0030] in,

[0031] e c,k =sin(Φ ref )(X k -X ref )-cos(Φ ref )(Y k -Y ref )

[0032] e l,k =-cos(Φ ref )(X k -X ref )-sin(Φ ref )(Y k -Y ref )

[0033] Among them, e l,k is the lateral tracking error, e c,k is the longitudinal tracking error, V x is the longitudinal speed, V ref is the reference speed, δ is the front wheel angle, T is the total required torque, q l is the lateral tracking error weight, q c is the longitudinal tracking error weight, q v is the speed tracking error weight, q δ To suppress the front wheel angle weight, q T To suppress the total demand moment weight, x t is the current state of the vehicle, f(x k ,u k ) is the system mechanism model, is the compensation model for the system state prediction error, is the current sampling sample required to compensate the model, θ is the model parameter, Φ ref is the reference yaw angle, X k is the longitudinal position of the vehicle in the earth coordinate system, Y k is the horizontal position in the geodetic coordinate system, X ref is the reference longitudinal position, Y ref is the reference horizontal position.

[0034] For the underlying driving torque distribution part of the adaptive adjustment of the vehicle body posture, due to the random unknown uneven road disturbances under the test conditions, its suspension system will be subjected to random vertical excitation. Since the various systems of the vehicle are integrated with each other and the lateral, longitudinal and vertical forces are coupled with each other, this random vertical excitation not only affects the smoothness and comfort of the vehicle, but also affects the vehicle's speed, trajectory tracking and other performances. The research object of the present invention is a four-wheel hub motor driven automatic driving vehicle. Due to its anti-pitch geometric characteristics, the driving torque of the four motors can be distributed in the lower-level controller to achieve adaptive adjustment of the vehicle posture. Considering the real-time and versatility of the control system, the lower-level controller adopts a low-order PID controller to obtain the additional vertical force, roll torque and pitch torque, and realizes torque distribution through the driving torque distribution matrix. The underlying driving torque distribution steps are as follows:

[0035] Step 1: Design three low-order PID controllers with the following control objectives: minimize the vertical acceleration of the vehicle body, minimize the roll angle of the vehicle body, and minimize the pitch angle of the vehicle body, and output additional vertical force, additional roll moment, and additional pitch moment;

[0036] Step 2: According to the force transmission mechanism of the suspension when the four-wheel hub motor is driven, the parameter values ​​in the torque distribution matrix are determined based on the vehicle structural parameters;

[0037] Step 3: Obtain the total driving force requirement given by the prediction model and the additional vertical force, additional roll moment, and additional pitch moment given by the low-order PID controller. Based on the torque distribution matrix, calculate the motor torque that each of the four wheel hub motors should provide for adaptively adjusting the vehicle body posture.

[0038] The formula for calculating the motor torque that each of the four wheel hub motors should provide for adaptively adjusting the vehicle body posture based on the torque distribution matrix is ​​as follows:

[0039]

[0040] Among them, θ f is the angle between the line connecting the front suspension pitch center and the front wheel contact point and the horizontal line, θ r is the angle between the line connecting the rear suspension pitch center and the rear tire contact point and the horizontal line, t f is the front axle track, t r is the rear axle track; F x is the total driving force demand, F z is the additional vertical force, M x is the additional rolling moment, M z is the additional pitching moment; F xfl is the longitudinal force of the left front wheel, F xfr Right front wheel longitudinal force, F xrl Left rear wheel longitudinal force, Fxrr Longitudinal force on the right rear wheel.

[0041] An automatic driving control system driven by a hybrid mechanism and data, the automatic driving control system includes an upper controller and a lower controller, for implementing the method as described above, wherein:

[0042] The upper-level controller is implemented through a data-mechanism hybrid drive model predictive control algorithm, receives measurement information about the vehicle state from the sensor, calculates the control quantity of the vehicle based on the prediction model of the model predictive controller, and sends it to the lower-level controller, wherein the control quantity of the vehicle includes the total required driving torque and the front wheel angle, and the prediction model compensates for the prediction error through the Gaussian process regression algorithm; the lower-level controller obtains the additional vertical force, roll moment and pitch moment through a low-order PID controller, and combines the total required driving torque sent by the upper-level controller, and calculates the torque that each of the four hub motors needs to provide through the torque distribution matrix, so that the vehicle can achieve automatic driving.

[0043] A storage medium stores a program thereon, and when the program is executed, the method described above is implemented.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] This invention integrates a data-driven hybrid model predictive control algorithm that utilizes driving data to guide predictive model self-learning, and a low-level drive torque distribution scheme tailored to the anti-pitch geometry characteristics of four-wheel hub motor drive, resulting in an adaptive autonomous driving control system and method for driving conditions. This invention enables the control system of a four-wheel hub motor-driven autonomous vehicle to adapt to time-varying operating conditions and vehicle body posture, alleviating the problem of the predictive model's inability to adapt to operating conditions in mechanism-driven model predictive control algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of the method of the present invention;

[0047] Figure 2 It is a system structure block diagram of the present invention;

[0048] Figure 3 This is a schematic diagram of the bottom driving torque distribution of the present invention;

[0049] Figure 4 Schematic diagram of a three-degree-of-freedom mechanism model in one embodiment

[0050] Figure 5 A schematic diagram of suspension force transmission for a vehicle driven by four wheel hub motors in one embodiment;

[0051] Figure 66a is a comparison diagram of the results of the automatic driving vehicle control system under low-adhesion straight-line acceleration and deceleration conditions obtained according to different prediction models in an embodiment, wherein (6a) is a torque comparison diagram, (6b) is a speed and yaw rate comparison diagram, (6c) is an acceleration, roll angle and pitch angle comparison diagram, and (6d) is an energy consumption and speed tracking error comparison diagram;

[0052] Figure 7 7 is a comparison diagram of the results of the automatic driving vehicle control system under high adhesion double lane change conditions obtained according to different prediction models in an embodiment, wherein (7a) is a comparison diagram of the front wheel angle and torque, (7b) is a comparison diagram of the vehicle speed and yaw angular velocity, (7c) is a comparison diagram of the acceleration, roll angle and pitch angle, and (7d) is a comparison diagram of the energy consumption and trajectory tracking error. DETAILED DESCRIPTION

[0053] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0054] The data-mechanism hybrid drive algorithm combines the advantages of both pure mechanism-driven and pure data-driven algorithms. The mechanism model describes the modelable portion of the physical system, while the learning model approximates the non-modelable portion. This algorithm achieves high modeling accuracy for autonomous vehicle system models and relatively low dataset construction costs. Because the control performance of autonomous driving control systems designed based on model predictive control algorithms is strongly correlated with the accuracy of the predictive model, the data-mechanism hybrid drive algorithm has become the preferred algorithm for upper-level control systems, balancing modeling accuracy, model complexity, and the economic efficiency of dataset construction.

[0055] For autonomous vehicles traveling on unknown, random road surfaces, their suspension systems are subject to random vertical excitations. Because the various vehicle systems are integrated and the lateral, longitudinal, and vertical forces are coupled, this random vertical excitation not only affects the vehicle's ride comfort and ride quality, but also its speed, trajectory tracking, and other performance. This research focuses on autonomous vehicles driven by four in-wheel motors. Due to their anti-pitch geometry, adaptive adjustment of the vehicle's attitude can be achieved by distributing the drive torque of the four motors in a lower-level controller.

[0056] To achieve adaptive control of a four-wheel-hub motor-driven autonomous vehicle control system to time-varying operating conditions and adaptive adjustment of vehicle body posture, and to alleviate the inability of the predictive model in mechanism-driven model predictive control algorithms to adapt to operating conditions, the present invention proposes a data-mechanism hybrid-driven model predictive control algorithm that uses driving data to guide the self-learning of the predictive model. This algorithm also develops a low-level drive torque distribution scheme tailored to the anti-pitch geometric characteristics of the four-wheel-hub motor drive, integrating it into an autonomous driving control system and method that is adaptive to driving conditions. Due to the diverse driving scenarios encountered by vehicles, this embodiment utilizes two classic test conditions: a straight-line acceleration and deceleration condition on a low-adhesion road surface, and a high-adhesion double lane change condition, both of which involve unknown road surface disturbances.

[0057] This embodiment first provides a mechanism data hybrid driven automatic driving control method, such as Figure 1 As shown, the following steps are included:

[0058] Offline dataset construction: Based on the sampling frequency and the model predictive controller driven by the pure mechanism model, the vehicle is controlled to output the preset trajectory and speed, which is used as the offline dataset;

[0059] Offline training of the residual model: A residual model is constructed based on the Gaussian process regression algorithm. Based on the offline data set, a residual model of the mechanism model is trained. Based on the mechanism model and the residual model, a prediction model of the model predictive controller is established to output the total driving force demand.

[0060] Low-level driving torque distribution: A low-order PID controller is used to obtain the additional vertical force, additional roll moment, and additional pitch moment. Based on the total driving force requirement, additional vertical force, additional roll moment, and additional pitch moment, the torque distribution matrix is ​​constructed using vehicle structural parameters to calculate the motor torque provided by each of the four wheel hub motors.

[0061] This embodiment also provides a mechanism data hybrid driven automatic driving control system, including an upper controller and a lower controller, whose structure is as follows: Figure 2 As shown, it is used to implement the method as described above, wherein,

[0062] The upper-level controller is implemented through a data-mechanism hybrid drive model predictive control algorithm. It receives measurement information about the vehicle status from sensors, calculates the control quantity of the vehicle based on the prediction model of the model predictive controller, and sends it to the lower-level controller. The control quantity of the vehicle includes the total required driving torque and the front wheel angle. The prediction model compensates for the prediction error through the Gaussian process regression algorithm; the lower-level controller obtains the additional vertical force, roll moment and pitch moment through a low-order PID controller, and combines it with the total required driving torque sent by the upper-level controller. The torque distribution matrix calculates the torque that each of the four hub motors needs to provide, enabling the vehicle to achieve autonomous driving.

[0063] The following describes in detail the implementation process of the above method and system:

[0064] 1. Offline dataset construction

[0065] Using a sampling time that is the same as the single-step prediction duration of the model predictive controller, measurement data of a four-wheel hub motor-driven autonomous driving vehicle under test conditions is collected to obtain a sequence dataset containing the state and control variables of the autonomous driving vehicle system.

[0066] Determine a certain sampling frequency t, and under the test conditions, design a purely mechanism-driven model predictive controller to control the vehicle to track the preset trajectory and speed. In this embodiment, the mechanism model adopts a three-degree-of-freedom vehicle dynamics model. The mechanism model is as follows Figure 4 The specific form of the mechanism model is as follows:

[0067]

[0068] F R,y =Dsin(Catan(Bα R -E(Bα R -atan(Bα R ))))

[0069] F F,y =Dsin(Catan(Bα F -E(Bα F -atan(Bα F ))))

[0070]

[0071] Wherein, X is the horizontal coordinate of the vehicle in the geodetic coordinate system; Y is the vertical coordinate of the vehicle in the geodetic coordinate system; is the vehicle yaw angle; V x V is the component of the vehicle's center of mass velocity along the x-axis of the vehicle coordinate system; y is the component of the vehicle's center of mass velocity along the y-axis of the vehicle coordinate system; r is the vehicle's yaw angular velocity; δ is the front wheel angle; the total driving torque T acting on the driving wheel; m is the vehicle mass, I z is the vehicle’s moment of inertia about the z axis, F x is the longitudinal force of the vehicle, F F,y is the total lateral force on the front axle of the vehicle, F R,y is the total lateral force on the rear axle of the vehicle, l F is the distance from the front axle to the center of mass of the vehicle, l R is the distance from the rear axle to the center of mass of the vehicle, D, C, B, E are the coefficients in the magic formula, α F is the front axle slip angle, α R is the rear axle slip angle, R eis the effective rolling radius of the wheel, g is the acceleration of gravity, f is the rolling resistance coefficient, C D is the drag coefficient, ρ is the air density, A is the vehicle's frontal area, and α is the road slope angle. The state variables are as follows:

[0072]

[0073] Select the control variable u = [δ; T], and after discretization by Euler method, we can get: x k+1 =f(x k ,u k ).

[0074] The control objective of the purely mechanism-driven model predictive controller of the present invention is to track the preset trajectory and speed while making the control input as small as possible. Therefore, the mathematical description of the control algorithm is as follows:

[0075]

[0076] in,

[0077] e c,k =sin(Φ ref )(X k -X ref )-cos(Φ ref )(Y k -Y ref )

[0078] e l,k =-cos(Φ ref )(X k -X ref )-sin(Φ ref )(Y k -Y ref )

[0079] Among them, e l,k is the lateral tracking error, e c,k is the longitudinal tracking error, V x is the longitudinal speed, V ref is the reference speed, δ is the front wheel angle, T is the total required torque, q l is the lateral tracking error weight, q c is the longitudinal tracking error weight, q v is the speed tracking error weight, q δ To suppress the front wheel angle weight, q T To suppress the total demand moment weight, x t is the current state of the vehicle, f(x k ,u k ) is the system mechanism model, Φ ref is the reference yaw angle, X kis the longitudinal position of the vehicle in the earth coordinate system, Y k is the horizontal position in the geodetic coordinate system, X ref is the reference longitudinal position, Y ref is the reference horizontal position.

[0080] The upper-level controller of the control system receives sensor measurement information about the vehicle's state. By solving the above optimization problem, the first set of elements of the solution is the control quantity of the vehicle (total required driving torque and front wheel angle). After the control quantity acts on the autonomous vehicle system, the upper-level controller receives new sensor measurement information about the vehicle's state, updates the new state information to the above optimization problem, and solves the problem in a loop, so that the autonomous vehicle can track the preset speed and trajectory under the test conditions. After the purely mechanism-driven model predictive controller completes the above tracking control task, it extracts the sensor measurement information about the vehicle's state under the test conditions from the autonomous vehicle storage module:

[0081]

[0082] in represents the vehicle longitudinal velocity, represents the lateral speed of the vehicle, represents the vehicle's yaw rate, represents the front wheel angle, Indicates the total driving torque.

[0083] At the same time, using the mechanism model x k+1 =f(x k ,u k ) Calculate the predicted value of the vehicle state at the next sampling moment based on the current state and control input:

[0084]

[0085] in, represents the predicted value of the vehicle longitudinal velocity, represents the predicted value of the vehicle's lateral velocity, represents the predicted value of the vehicle's yaw rate, represents the predicted value of the front wheel angle, Indicates the predicted value of the total driving torque.

[0086] Define the prediction error of the mechanism model at the current moment

[0087] The offline dataset is constructed as follows:

[0088]

[0089] in, is the feature data, e is the label data, its subscript represents the data number, and m is the total number of samples in the data set.

[0090] 2. Residual model learning of mechanism model

[0091] Based on the Gaussian process regression algorithm, the residual model learning mathematical model of the mechanism model is established:

[0092]

[0093]

[0094] Definition: dV x Indicates the longitudinal velocity V x The difference between the sensor measurement value and the predicted value given by the mechanism model; dV y Indicates the lateral velocity V y The difference between the sensor measurement value and the predicted value given by the mechanism model; dr represents the difference between the yaw angular velocity r sensor measurement value and the predicted value given by the mechanism model; let d k =[dV x,k ,dV y,k ,dr k ], measured value y k =d k +w k ,y k It can be calculated by the following formula: k =x k -f(x k-1 ,u k-1 )+w k , where x k represents the measured value of the system state at time k, f(x k-1 ,u k-1 ) represents the predicted value of the system state at time k-1. As the eigenvalue of the Gaussian process regression algorithm, the residual model learning process of the mechanism model can be realized by using the data set D containing m samples obtained by sampling. The specific steps are as follows:

[0095] Assume that the residual model of the mechanism model is an unknown function d, which represents the input Mapping to output e, that is: d:R nz →R nd In addition, the output vector e k It is obtained by measurement and is contaminated by independent and identically distributed Gaussian noise, that is: Among them, w k In the following form:

[0096] w k ~N(0,Σ w )

[0097]

[0098] Where diag represents a diagonal matrix.

[0099] Assume that the output vector e k Any dimension a∈(1,2,...,n d ) are independent of each other, then for a new input Its mean and variance at the output dimension a can be expressed as follows:

[0100]

[0101] Among them, [Y].a represents the a-th column of matrix Y, which consists of all output vectors y k The ath dimension of ; I represents an m×m identity matrix, represents the variance of the measurement noise in the ath dimension; represents a real number, obtained by the kernel function. In this embodiment, the square exponential kernel is selected: in, and L a is the hyperparameter in the kernel function; Represents an m-dimensional row vector, through calculate; Represents an m×m symmetric matrix, through Calculation: m is the total number of samples in the dataset, and z and z are the samples in the dataset.

[0102] By e k The calculation formula of the variance and mean can be used to calculate the value of e k Make an estimate for each dimension of the vector, then, when the input is When the residual model e k The estimated value of the vector is:

[0103]

[0104]

[0105] The present invention uses the mean estimate of the residual of the mechanism model as the residual model of the mechanism model, that is, By using the maximum likelihood estimation method, the following optimization problem is constructed:

[0106]

[0107] You can calculate The unknown parameter θ in .

[0108] When the residual model of the mechanism model After the undetermined parameters in the Gaussian process regression algorithm are determined during the training process, they will remain unchanged in the subsequent verification phase. Therefore, the upper-level model predictive control-based data mechanism hybrid drive speed trajectory tracking control algorithm has the following description form:

[0109]

[0110] in,

[0111] e c,k =sin(Φ ref )(X k -X ref )-cos(Φ ref )(Y k -Y ref )

[0112] e l,k =-cos(Φ ref )(X k -X ref )-sin(Φ ref )(Y k -Y ref )

[0113] Among them, e l,k is the lateral tracking error, e c,k is the longitudinal tracking error, V x is the longitudinal speed, V ref is the reference speed, δ is the front wheel angle, T is the total required torque, q l is the lateral tracking error weight, q c is the longitudinal tracking error weight, q v is the speed tracking error weight, q δ To suppress the front wheel angle weight, q T To suppress the total demand moment weight, x t is the current state of the vehicle, f(x k ,u k ) is the system mechanism model, is the compensation model for the system state prediction error, is the current sampling sample required to compensate the model, θ is the model parameter, Φ ref is the reference yaw angle, X k is the longitudinal position of the vehicle in the earth coordinate system, Y k is the horizontal position in the geodetic coordinate system, X ref is the reference longitudinal position, Y ref is the reference lateral position. In the description of the above data mechanism hybrid drive algorithm, it is obvious that the prediction model x k+1 f(x k ,u k) is the mechanism driving part, which is derived from the three-degree-of-freedom horizontal and vertical coupled dynamic model; It is the data-driven part, which is based on the Gaussian process regression algorithm and uses the sampling data set D to determine the model parameters through the maximum likelihood estimation algorithm.

[0114] 3. Underlying driving torque distribution for adaptive adjustment of vehicle posture

[0115] For the underlying driving torque distribution part of the adaptive adjustment of the vehicle body posture, due to the existence of random unknown uneven road disturbances under the test conditions, its suspension system will be subjected to random vertical excitation. Since the various systems of the vehicle are integrated with each other, the lateral, longitudinal and vertical forces are coupled with each other. This random vertical excitation not only affects the smoothness and comfort of the vehicle, but also affects the vehicle's speed, trajectory tracking and other performances. The research object of the present invention is a four-wheel hub motor driven autonomous driving vehicle. Due to its anti-pitch geometric characteristics, the driving torque of the four motors can be distributed in the lower-level controller to achieve adaptive adjustment of the vehicle posture and weaken the influence of uneven road excitation on the vehicle body posture. This force transmission mechanism, such as Figure 5 As shown. Considering the real-time and universality of the control system, the lower-level controller uses a low-order PID controller to obtain the additional vertical force, roll moment, and pitch moment, and realizes the torque distribution through the driving torque distribution matrix. The steps of the bottom-level driving torque distribution are as follows:

[0116] Step 1: Design three low-order PID controllers with the following control objectives: minimize the vertical acceleration of the vehicle body, minimize the roll angle of the vehicle body, and minimize the pitch angle of the vehicle body, and output additional vertical force, additional roll moment, and additional pitch moment;

[0117] Step 2: According to the force transmission mechanism of the suspension when the four-wheel hub motor is driven, the parameter values ​​in the torque distribution matrix are determined based on the vehicle structural parameters;

[0118] Step 3, such as Figure 3 As shown in the figure, the total driving force requirement given by the prediction model and the additional vertical force, additional roll moment, and additional pitch moment given by the low-order PID controller are obtained. Based on the torque distribution matrix, the motor torque that each of the four wheel hub motors should provide for adaptively adjusting the vehicle body posture is calculated:

[0119]

[0120] Among them, θ f is the angle between the line connecting the front suspension pitch center and the front wheel contact point and the horizontal line, θ r is the angle between the line connecting the rear suspension pitch center and the rear tire contact point and the horizontal line, t f is the front axle track, t r is the rear axle track; F xis the total driving force demand, F z is the additional vertical force, M x is the additional rolling moment, M z is the additional pitching moment; F xfl is the longitudinal force of the left front wheel, F xfr Right front wheel longitudinal force, F xrl Left rear wheel longitudinal force, F xrr Longitudinal force on the right rear wheel.

[0121] The proposed system was simulated and verified under low-adhesion straight-line acceleration and deceleration conditions and high-adhesion double lane change conditions in the presence of random road roughness. The results were compared with those from a purely mechanism-driven model predictive control approach. The differences between these approaches and the proposed system lie solely in the control system; the perception and execution systems remain the same.

[0122] To better simulate the dynamic characteristics of a four-wheel hub motor-driven autonomous vehicle, the simulation environment of the present invention was built in Modelon. The integrated data mechanism hybrid drive autonomous driving control system of the present invention was implemented in Simulink.

[0123] At the same time, the present invention designs three indicators to evaluate the performance of the method:

[0124] 1. Track and control performance indicators.

[0125]

[0126] Among them, t e is the driving time under the test conditions, v Target is the preset reference speed under test conditions, x Target and y Target are the preset longitudinal reference trajectory and lateral reference trajectory respectively. v1 Measuring the control performance of the control system to control the autonomous vehicle to follow the preset speed, J v2 Measures the control performance of the control system in controlling the autonomous vehicle to follow the preset trajectory

[0127] 2. Energy consumption index.

[0128]

[0129] Among them, t e is the driving time under the test conditions, V Bi I is the voltage of each of the four hub motors when they output torque. i is the current of each of the four hub motors when they output torque, J e1 Measures the energy consumption of the control system while controlling the autonomous vehicle under test conditions.

[0130] 3. Constraint violation indicators.

[0131] The test conditions selected by the present invention involve random uneven road disturbances. In order to verify the performance of the control system in adaptively adjusting the body posture of the four-wheel hub motor-driven autonomous driving vehicle, restricted ranges are designed for the pitch angle, roll angle and vertical acceleration of the autonomous driving vehicle.

[0132] In low-adhesion straight-line acceleration and deceleration conditions, the vehicle posture constraints are as follows:

[0133] |a z (t)|<a z_Target1 (0.4[m / s 2 ])

[0134] |φ(t)|<φ _Target1 (0.004[rad])

[0135] |ψ(t)|<ψ _Target1 (0.0015[rad])

[0136] In the high-adhesion double-lane-shifting condition, the vehicle posture constraints are as follows:

[0137] |a z (t)|<a z_Target1 (1.4[m / s 2 ])

[0138] |φ(t)|<φ _Target1 (0.7[rad])

[0139] |ψ(t)|<ψ _Target1 (0.14[rad])

[0140] Among them, a z is the vertical acceleration of the vehicle body, φ is the roll angle of the vehicle body, and ψ is the pitch angle of the vehicle body. The design vehicle body posture constraint violation index function is as follows:

[0141]

[0142]

[0143] Among them, J s Measures the number of violations of the vehicle posture constraints when the control system controls the autonomous vehicle driving under test conditions.

[0144] In order to illustrate the impact of autonomous driving control systems based on different types of model predictive control algorithms on the method proposed in this invention, in addition to the prediction model based on the data-mechanism hybrid drive algorithm, the present invention also established a pure mechanism prediction model based on the three-freedom vehicle dynamics model, and integrated the obtained prediction model into the model predictive control framework to obtain an autonomous driving control system. The obtained autonomous driving control system was verified in the test conditions, and the verification results are shown in Figure 2. Figure 6 and Figure 7 As shown in the figure, the autonomous driving control system based on the data-mechanism hybrid drive algorithm performs best. This is attributed to the advantages of the Gaussian process regression algorithm in modeling the residual model of the mechanism model. The prediction model obtained through learning improves the prediction accuracy of the future operating dynamics of the autonomous driving vehicle system, enabling the original purely mechanism-driven autonomous driving vehicle control system to adapt to complex time-varying working conditions and effectively suppress the impact of uneven road surface on tracking performance. It also verifies the rationality of the present invention's selection of the Gaussian process regression algorithm to establish a residual model to compensate for the three-degree-of-freedom mechanism model as the basis of the prediction model.

[0145] From the above simulation results, it can be seen that the present invention realizes the adaptive time-varying working conditions and adaptive adjustment of the vehicle body posture of the four-wheel hub motor driven automatic driving vehicle control system by utilizing the data mechanism hybrid drive model predictive control algorithm that uses driving data to guide the self-learning of the predictive model and the underlying driving torque distribution scheme for the anti-pitch geometric characteristics of the four-wheel hub motor drive, thereby alleviating the problem that the predictive model in the model predictive control algorithm driven by the mechanism model cannot adapt to the working conditions, and is of great significance in the development of adaptive working condition automatic driving vehicle control systems.

[0146] The present invention discloses an automatic driving control system and method of mechanism-data hybrid-driven model predictive control, which is suitable for automatic driving vehicles driven by four-wheel hub motors. The system covers a hierarchical control scheme for speed tracking, trajectory tracking, and posture adjustment. By comparing the actual measurement data of the vehicle state with the vehicle state data calculated by the simplified mechanism model, the residual of the mechanism model is modeled using the Gaussian process regression algorithm, and a prediction model driven by data-mechanism hybrid is learned; the prediction model is then integrated into the model predictive control framework to improve the prediction accuracy of the future operating dynamics of the system, so that the original pure mechanism model-driven control system has the ability to adapt to complex time-varying working conditions and effectively suppress the influence of external disturbances on tracking performance; finally, an automatic driving control system and adaptive method driven by mechanism data hybrid for all driving conditions are obtained, realizing more intelligent and more accurate dynamic control of automatic driving vehicles. In addition, the present invention utilizes the anti-pitch geometric characteristics of the automatic driving vehicle driven by four-wheel hub motors, adopts a low-order PID controller to obtain additional vertical force, roll moment and pitch moment, effectively weakening the excitation of external disturbances on the vehicle body posture, and the control scheme has high real-time and versatility. The proposed method consists of three main parts: dataset construction, residual modeling of the mechanism model, and underlying driving torque allocation for adaptive vehicle posture adjustment. The dataset is constructed by sampling feasible state points; residual model learning exploits the Gaussian process regression algorithm's ability to model datasets contaminated by Gaussian noise; and the underlying driving torque allocation for adaptive vehicle posture adjustment exploits the anti-pitch geometry of a four-wheel hub motor-driven autonomous vehicle.

[0147] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0148] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A mechanism-data hybrid driven automatic driving control method, characterized in that: The following steps are involved: Offline dataset construction: Based on the sampling frequency and a model predictive controller driven by a pure mechanism model, the vehicle is controlled to output a preset trajectory and speed, which is used as the offline dataset. The mechanism model adopts a three-degree-of-freedom vehicle dynamics model. Offline training of the residual model: A residual model is constructed based on the Gaussian process regression algorithm. Based on the offline data set, a residual model of the mechanism model is trained. Based on the mechanism model and the residual model, a prediction model of the model predictive controller is established to output the total driving force demand. Low-level driving torque distribution: A low-order PID controller is used to obtain the additional vertical force, additional roll moment, and additional pitch moment. Based on the total driving force requirement, additional vertical force, additional roll moment, and additional pitch moment, a torque distribution matrix is ​​constructed using vehicle structural parameters to calculate the motor torque provided by each of the four wheel hub motors. The objective function and constraints of the prediction model of the model predictive controller are expressed as: in, and c,k =sin(Φ ref )(X k -X ref )-cos(Φ ref )(AND k -AND ref ) and l,k =-cos(Φ ref )(X k -X ref )-sin(Φ ref )(AND k -AND ref ) Among them, e l,k is the lateral tracking error, e c,k is the longitudinal tracking error, V x is the longitudinal speed, V ref is the reference speed, δ is the front wheel angle, T is the total required torque, q l is the lateral tracking error weight, q c is the longitudinal tracking error weight, q v is the speed tracking error weight, q δ To suppress the front wheel angle weight, q T To suppress the total demand moment weight, x t is the current state of the vehicle, f(x k ,u k ) is the system mechanism model, is the compensation model for the system state prediction error, is the current sampling sample required to compensate the model, θ is the model parameter, Φ ref is the reference yaw angle, X k is the longitudinal position of the vehicle in the earth coordinate system, Y k is the horizontal position in the geodetic coordinate system, X ref is the reference longitudinal position, Y ref is the reference horizontal position.

2. The automatic driving control method driven by a hybrid mechanism and data according to claim 1, characterized in that: In the offline dataset construction step, a sampling time equal to the single-step prediction duration of the model predictive controller is used to collect measurement data of the four-wheel hub motor driven autonomous driving vehicle under test conditions, thereby obtaining a sequence dataset containing the state and control quantities of the autonomous driving vehicle system.

3. The automatic driving control method driven by a hybrid mechanism and data according to claim 1, characterized in that: The offline dataset is constructed as follows: a sampling frequency t is preset, and under test conditions, a purely mechanism-driven model predictive controller is used to control the vehicle to track the preset trajectory and speed, and to collect the measured values ​​of the vehicle state and control variables: in represents the vehicle longitudinal velocity, represents the lateral speed of the vehicle, represents the vehicle's yaw rate, represents the front wheel angle, Indicates the total driving torque; The mechanism model is used to calculate the predicted value of the vehicle state at the next sampling moment based on the current state and control input: in, represents the predicted value of the vehicle longitudinal velocity, represents the predicted value of the vehicle's lateral velocity, represents the predicted value of the vehicle's yaw rate, represents the predicted value of the front wheel angle, Indicates the predicted value of total driving torque; Define the prediction error of the mechanism model at the current moment The offline dataset is constructed as follows: in, is the feature data, e is the label data, its subscript represents the data number, and m is the total number of samples in the data set.

4. The automatic driving control method based on a hybrid mechanism and data drive according to claim 3, characterized in that: Based on the Gaussian process regression algorithm, the residual model learning mathematical model of the mechanism model is established: Among them, [Y].a represents the a-th column of matrix Y, which consists of all output vectors y k The ath dimension of ; I represents an m×m identity matrix, represents the variance of the measurement noise in the ath dimension; Represents a real number, obtained by the kernel function: in, and L a is the hyperparameter in the kernel function; Represents an m-dimensional row vector, through calculate; Represents an m×m symmetric matrix, through Calculation; m is the total number of samples in the data set, z, All are samples from the dataset.

5. The automatic driving control method driven by a hybrid mechanism and data according to claim 4, characterized in that: The mathematical model is learned through the residual model of the mechanism model to estimate each dimension of the output vector. When the input is z, the estimated value of the residual of the mechanism model is: in, n d is the dimension of the output vector; The maximum likelihood estimation method is used to calculate the hyperparameters in the residual model learning mathematical model to obtain the residual model of the mechanism model. The mechanism model and the residual model are linearly superimposed to obtain the prediction model of the model predictive controller.

6. The automatic driving control method driven by a hybrid mechanism and data according to claim 1, characterized in that: The bottom driving torque distribution steps are as follows: Step 1: Design three low-order PID controllers with the following control objectives: minimize the vertical acceleration of the vehicle body, minimize the roll angle of the vehicle body, and minimize the pitch angle of the vehicle body, and output additional vertical force, additional roll moment, and additional pitch moment; Step 2: According to the force transmission mechanism of the suspension when the four-wheel hub motor is driven, the parameter values ​​in the torque distribution matrix are determined based on the vehicle structural parameters; Step 3: Obtain the total driving force requirement given by the prediction model and the additional vertical force, additional roll moment, and additional pitch moment given by the low-order PID controller. Based on the torque distribution matrix, calculate the motor torque that each of the four wheel hub motors should provide for adaptively adjusting the vehicle body posture.

7. The automatic driving control method driven by a hybrid mechanism and data according to claim 6, characterized in that: The formula for calculating the motor torque that each of the four wheel hub motors should provide for adaptively adjusting the vehicle body posture based on the torque distribution matrix is ​​as follows: Among them, θ f is the angle between the line connecting the front suspension pitch center and the front wheel contact point and the horizontal line, θ r is the angle between the line connecting the rear suspension pitch center and the rear tire contact point and the horizontal line, t f is the front axle track, t r is the rear axle track; F x is the total driving force demand, F z is the additional vertical force, M x is the additional rolling moment, M z is the additional pitching moment; F xfl is the longitudinal force of the left front wheel, F xfr Right front wheel longitudinal force, F xrl Left rear wheel longitudinal force, F xrr Longitudinal force on the right rear wheel.

8. An automatic driving control system driven by a hybrid mechanism and data, characterized in that: The automatic driving control system includes an upper controller and a lower controller, configured to implement the method according to any one of claims 1 to 7, wherein: The upper-level controller is implemented through a data-mechanism hybrid drive model predictive control algorithm, receives measurement information about the vehicle state from the sensor, calculates the control quantity of the vehicle based on the prediction model of the model predictive controller, and sends it to the lower-level controller, wherein the control quantity of the vehicle includes the total required driving torque and the front wheel angle, and the prediction model compensates for the prediction error through the Gaussian process regression algorithm; the lower-level controller obtains the additional vertical force, roll moment and pitch moment through a low-order PID controller, and combines the total required driving torque sent by the upper-level controller, and calculates the torque that each of the four hub motors needs to provide through the torque distribution matrix, so that the vehicle can achieve automatic driving.

9. A storage medium having a program stored thereon, characterized in that: When the program is executed, the method according to any one of claims 1 to 7 is implemented.

Images