CN-121560086-B - Vehicle path tracking and longitudinal speed control method based on Gaussian process compensation

CN121560086BCN 121560086 BCN121560086 BCN 121560086BCN-121560086-B

Abstract

The invention discloses a vehicle path tracking and longitudinal speed control method based on Gaussian process compensation, which belongs to the technical field of dynamics control and comprises the following steps of S1, collecting vehicle data on a calibration road, constructing a Gaussian process residual model, performing offline training, S2, performing online updating on the Gaussian process residual model after offline training, S3, calculating three-channel compensation quantity, S4, constructing a residual correction prediction model, determining output prediction, S5, constructing a quadratic programming problem, solving, obtaining control increment, steering input and longitudinal acceleration input at the current moment, and applying the control increment, steering input and longitudinal acceleration input to a vehicle actuator. The sparse residual Gaussian process structure and the guidance point maintenance strategy provided by the invention give consideration to online learning capability and operation complexity, and are suitable for being deployed on the existing automatic driving controller hardware platform.

Inventors

XIE YONGQI
SUN CHAO
Ning Changjiu
YANG XIONGJI
LENG JIANGHAO
WEN DA

Assignees

北京理工大学深圳汽车研究院(电动车辆国家工程实验室深圳研究院)
北京理工大学

Dates

Publication Date: 20260508
Application Date: 20260123

Claims (9)

1. A vehicle path tracking and longitudinal speed control method based on gaussian process compensation, comprising the steps of: S1, collecting data of a vehicle on a calibration road, constructing a Gaussian process residual model, and performing offline training; s2, collecting data of a vehicle on a target road, determining an uncertainty index and a normalized distance threshold, and updating a Gaussian process residual model after offline training on line; S3, calculating three-channel compensation quantity by using the Gaussian process residual error model after online updating; s4, constructing a residual error correction prediction model according to the three-channel compensation quantity, and determining output prediction; S5, constructing a quadratic programming problem according to output prediction, solving the quadratic programming problem to obtain a control increment, a steering input and a longitudinal acceleration input at the current moment, and applying the control increment, the steering input and the longitudinal acceleration input to a vehicle actuator; the step S1 comprises the following substeps: s11, collecting a vehicle state vector and a control input vector on a calibration road; S12, constructing a vehicle nominal dynamics model according to the vehicle state vector and the control input vector, and performing discretization processing to obtain a discrete time nominal model; S13, collecting time sequence data on a typical road, and calculating a nominal prediction state corresponding to the time sequence data by using a discrete time nominal model; s14, constructing residual output and input features of a Gaussian process according to the nominal prediction state; S15, stacking residual output and input features of a Gaussian process to obtain a training data set as a discrete training sample; S16, calculating a sample mean value and a sample variance of input features in the discrete training samples, and outputting the sample mean value and the sample variance of residual errors; S17, calculating an input standard deviation vector according to the sample mean and sample variance of the input features; s18, calculating a first standardized variable and a second standardized variable according to the input standard deviation vector and the output standard deviation vector; S19, determining standardized samples corresponding to the first standardized variable and the second standardized variable; s110, constructing a Gaussian process residual model; and S111, carrying out parameter identification on the standardized sample to obtain a super parameter, taking the super parameter and the induction point set as an initial state of a Gaussian process residual model, and carrying out offline training.
2. The gaussian process compensation based vehicle path tracking and longitudinal speed control method according to claim 1, wherein in S11, a vehicle state vector is used The expression of (2) is: ; Wherein, the To be an absolute position of the vehicle centroid along the X-axis in the world coordinate system, To be an absolute position of the vehicle's centroid along the Y-axis in the world coordinate system, In order to be the heading angle, For the longitudinal speed in the vehicle body coordinate system, For the lateral velocity in the vehicle body coordinate system, In order to be the yaw rate, Is a transposition; In S11, the input vector is controlled The expression of (2) is: ; Wherein, the Is the rotation angle of the front wheel, Is the first Longitudinal acceleration at each sampling instant; in the S12, a vehicle nominal dynamics model The expression of (2) is: ; Wherein, the To map the current state and inputs to the vehicle dynamics equations of state derivatives for a given vehicle parameter, Is a state vector that is made up of vehicle state variables, For the vehicle to control the input vector, Is a set of vehicle parameters; in the S12, a discrete time nominal model The expression of (2) is: ; Wherein, the Is a discrete kinetic function; in the S14, residual error output The expression of (2) is: ; Wherein, the Is the first The residuals of the longitudinal speed at each sampling instant, Is the first The transverse velocity residual at each sampling instant, Is the first The yaw rate residual at each sample time, Is the first The longitudinal speed of the vehicle measured at each sampling instant, Is the first The lateral speed of the vehicle measured at each sampling instant, Is nominal model in the first The predicted longitudinal speed of the vehicle at each sampling instant, Is the nominal model The predicted lateral speed of the vehicle at each sampling instant, Is the first The yaw rate of the vehicle measured at each sampling instant, Is nominal model in the first Predicting the yaw rate of the vehicle at each sampling moment; in S14, input features of the Gaussian process The expression of (2) is: ; Wherein, the A matrix is selected for the states and, For the input of the selection matrix, Is the first The vehicle at each sampling instant measures a state vector, To at the first The vehicle control input vector for each sampling instant, Is the first The front wheel rotation angle at each sampling instant, Is the first Longitudinal acceleration at each sampling instant, Is a 5-dimensional real vector space; In the S15, training data set The expression of (2) is: ; Wherein, the The total number of samples in the offline training data set; In the S18, a first standardized variable The expression of (2) is: ; Wherein, the A sample mean vector representing the input features, Representing an input standard deviation vector; In the S18, a second normalized variable The expression of (2) is: ; Wherein, the A sample mean vector representing the residual output, Representing the output standard deviation vector.
3. The gaussian process compensation based vehicle path tracking and longitudinal speed control method according to claim 1, wherein said S2 comprises the sub-steps of: s21, collecting a state vector and a control input vector of a vehicle at each sampling moment on a target road; S22, predicting a state vector and a control input vector of a vehicle on a target road at each sampling moment by using a discrete time nominal model to obtain a nominal next moment state; s23, constructing an online residual error measurement vector according to the nominal next moment state and the actual measurement state; s24, determining standardized input according to the online residual error measurement vector; S25, taking standardized input as input of a Gaussian process residual model after offline training to obtain prior distribution of residual prediction, and entering S26; S26, calculating an uncertainty index according to the prediction covariance; S27, calculating a normalized distance according to the current induction point set; s28, when the uncertainty index meets the gating condition constructed by the normalized distance And updating the Gaussian process residual model on line, wherein, In order to be an uncertainty index, In order for the uncertainty threshold to be a high, In order to normalize the distance, Is a normalized distance threshold.
4. The gaussian process compensation-based vehicle path tracking and longitudinal speed control method according to claim 3, wherein in S26, an uncertainty index is used The expression of (2) is: ; Wherein, the Is the first The residual covariance matrix of the sample instants, Performing matrix trace operation; in the S27, the distance is normalized The expression of (2) is: ; Wherein, the Is the current induction point set Index of the points of induction, As a first standardized variable, a second standardized variable, For the current set of induction points, A length scale vector for each input dimension.
5. The gaussian process compensation based vehicle path tracking and longitudinal speed control method according to claim 1, wherein said S3 comprises the sub-steps of: s31, constructing Gaussian process input according to the current actual measurement state and the current time control quantity or the last time control quantity of the control quantity to be optimized; s32, normalizing Gaussian process input; S33, inputting the standardized Gaussian process as an input of an updated Gaussian process residual model to obtain prediction distribution of three-channel residual; S34, mapping the prediction distribution of the three-channel residual error from the standardized domain back to the physical quantity domain, and calculating three-channel compensation quantity.
6. The gaussian process compensation based vehicle path tracking and longitudinal speed control method according to claim 5, wherein in S31, a gaussian process input is performed The expression of (2) is: ; Wherein, the For the lateral velocity in the vehicle body coordinate system, For the longitudinal speed in the vehicle body coordinate system, In order to be the yaw rate, Is the rotation angle of the front wheel, In the case of a longitudinal acceleration command or a braking force command, Is a transposition; in S33, the expression of the prediction distribution of the three-channel residual error is: ; Wherein, the For the purpose of normalizing the mean value of the residual errors, In order to normalize the residual covariance matrix, Is a multi-element gaussian distribution, A normalized residual prediction vector that obeys the distribution; in the S34, three-channel compensation amount The expression of (2) is: ; Wherein, the As the amount of compensation for the lateral velocity, For the amount of compensation of the longitudinal speed, Is the compensation amount of the yaw rate.
7. The gaussian process compensation based vehicle path tracking and longitudinal speed control method according to claim 1, wherein said S4 comprises the sub-steps of: S41, constructing a residual injection matrix; S42, calculating a residual correction prediction model according to the residual injection matrix and the three-channel compensation quantity; S43, performing first-order linearization on the residual error correction prediction model to obtain a linear time-varying state space model; S44, constructing an overlapped state prediction equation on the linear time-varying state space model based on the predicted time domain length and the control time domain length; S45, constructing an output superposition equation in a prediction time domain based on the superposition state prediction equation; s46, determining output prediction by using an output superposition party in the prediction time domain.
8. The gaussian process compensation based vehicle path tracking and longitudinal speed control method according to claim 7, wherein in S41, a residual injection matrix is used The expression of (2) is: ; Wherein, the In the form of a 3 x 3 zero matrix, For a3 x 3 matrix of units, Is a6 x 3 dimensional real matrix space; In S42, a residual correction prediction model The expression of (2) is: ; Wherein, the Is a function of the nominal discrete dynamics of the vehicle, As a function of the current state vector, For the current control input vector to be used, As a set of parameters of the vehicle, Weights are injected for the residual errors and, The compensation quantity is three channels; In S43, the expression of the linear time-varying state space model is: ; Wherein, the For the next time state vector after linearization, As a matrix of states, In order to input the matrix of the data, Is a constant term; In S44, the expression of the superimposed state prediction equation is: ; Wherein, the For predicting the superimposed vector of the states of each step in the time domain, In order to control the incremental superposition vector, In order to predict the matrix of time domain state transition coefficients, To predict the time domain control delta coefficient matrix, For predicting a time domain constant term vector; in S45, the expression of the output superposition equation in the prediction domain is: ; Wherein, the In order to predict the output of the superimposed vector, To output the state mapping matrix of the superimposed vector, To output a matrix of coefficients for the control increment, Is a constant term vector in the output equation.
9. The gaussian process compensation based vehicle path tracking and longitudinal speed control method according to claim 1, wherein said S5 comprises the sub-steps of: s51, determining reference output of each step in the prediction time domain according to a reference path of a target road; S52, determining an output error according to the reference output and the output prediction of each step in the prediction time domain; S53, constructing a first diagonal weighting matrix and a second diagonal weighting matrix; S54, calculating a weighted performance index in a prediction time domain according to the output error, the first diagonal weight matrix and the second diagonal weight matrix; s55, constructing a state constraint condition; S56, integrating the weighted performance index and the state constraint condition in the prediction time domain into a quadratic programming problem according to the superposition state prediction equation and the output superposition equation in the prediction time domain; S57, carrying out online solving on the secondary optimization problem by using a quadratic programming solver to obtain an optimal control increment sequence, taking a first item of the optimal control increment sequence as a control increment at the current moment, and applying steering input and longitudinal acceleration input to a vehicle executor; In S51, the reference output of each step in the prediction time domain The expression of (2) is: ; Wherein, the In order to achieve the desired longitudinal velocity, In order to predict the length of the time domain, Is a transposition; in S52, an error is outputted The expression of (2) is: ; Wherein, the Predicting for output; In the S53, a first diagonal weight matrix The expression of (2) is: ; Wherein, the As a result of the longitudinal velocity error, In the event of a lateral error, As the heading angle error weight, Is a diagonal matrix operator; In the S53, a second diagonal weight matrix The expression of (2) is: ; Wherein, the For the incremental weight of the steering acceleration, The longitudinal acceleration increment weight; in S54, the weighted performance index in the prediction time domain The expression of (2) is: ; Wherein, the In order to control the length of the time domain, To control the increment; In S55, the expression of the state constraint condition is: ; ; ; Wherein, the For the lower boundary of the front wheel steering angle, Is the first The rotating angle of the front wheel of the walking machine, Is the upper boundary of the front wheel corner, For the lower bound of the longitudinal acceleration, Is the first The longitudinal acceleration of the step is determined, For the upper bound of the longitudinal acceleration, For the lower bound of the increment of the rotation angle, Is the first The increment of the step angle, For the upper bound of the increment of the rotation angle, For the lower bound of the acceleration increment, Is the first The step acceleration is increased by a step acceleration, For the upper bound of the acceleration, For the lower bound of the lateral deviation, To be at the moment Predicted first The lateral deviation of the steps is calculated, For the upper bound of the lateral deviation, For the lower boundary of the longitudinal velocity, To be at the moment S prediction of the first The longitudinal speed of the step, Is the upper boundary of the longitudinal speed; In S56, the expression of the quadratic programming problem is: ; Wherein, the In order to control the incremental stacking direction, For a Hessian matrix of quadratic costs, Is a vector of coefficients of the linear term, In the form of a matrix of inequalities, Constraint vectors for inequality; in S57, the control increment of the current time The expression of (2) is: ; Wherein, the For the increment of the steering angle of the front wheels, Is the longitudinal acceleration increment.

Description

Vehicle path tracking and longitudinal speed control method based on Gaussian process compensation Technical Field The invention belongs to the technical field of dynamics control, and particularly relates to a vehicle path tracking and longitudinal speed control method based on Gaussian process compensation. Background Along with the acceleration application of the automatic driving and intelligent networking technology in the scenes of highways, urban roads, mining transportation, closed test fields and the like, the ground unmanned vehicle realizes the high-precision path tracking and longitudinal speed joint control under the medium-high speed working condition, and becomes an important basic capability for guaranteeing the driving safety and riding comfort. Particularly, under complex working conditions of large curve radius change, obvious road adhesion coefficient fluctuation, frequent vehicle load change and the like, the vehicle dynamics presents strong nonlinearity and strong coupling characteristics, and higher requirements are provided for modeling accuracy, self-adaption capability and instantaneity of the controller. At present, model predictive control methods such as LQR, MPC and the like based on a simplified monorail model and a linear tire model are widely adopted in engineering, and fixed parameters are usually identified offline under a calibration test or a single road condition and kept unchanged for a long time in actual operation. The method is difficult to accurately describe the influences of factors such as nonlinear cornering characteristics, longitudinal and transverse coupling effects, road adhesion coefficients, load transfer and the like of the tire on vehicle dynamics. When the vehicle switches from the calibration condition to different roads, different speeds or different loading conditions, the residual error between the nominal model and the real dynamics can be significantly increased, resulting in increased path tracking error, softer or looser steering response and reduced safety margin. The traditional self-adaptive control or parameter estimation method depends on a low-dimensional parameterized model, is difficult to capture complex and time-varying high-dimensional residual error dynamics on the premise of ensuring stability, and is difficult to achieve both control precision and implementation complexity. In recent years, a residual modeling method based on Gaussian Process (GP) is introduced into vehicle dynamics and trajectory tracking control for learning a mapping relationship between "model error-state/control" on a nominal model basis and performing feedforward compensation in MPC, thereby improving tracking accuracy. However, existing GP-MPC methods rely on off-line centralized training in which a large number of samples are acquired on a target road or similar conditions, whereby hyper-parameters are estimated and regression models are trained, and in which once the road geometry, attachment conditions or desired speed profile change, the model generalization capability is limited, often requiring re-acquisition of data and off-line training. In addition, the calculation complexity of standard Gaussian process reasoning and updating increases in a cubic manner along with the number of samples, and online updating and real-time control coupling are difficult to achieve under the conditions of high sampling frequency and limited vehicle-mounted computing power. The existing sparse GP or online GP works reduce the calculated amount through mechanisms such as induction points, but the problems that the number of the induction points is increased unconstrained, the addition/replacement strategy is lack of correlation with closed-loop control performance, instantaneity and safety constraint are not fully considered, and even when residual error learning is improper, new unstable factors such as overcompensation and oscillation are introduced. Therefore, a sparse online Gaussian process residual modeling method for joint control of unmanned vehicle path tracking and longitudinal speed is needed, online learning and updating of model residual are achieved under the condition of limited induction point scale and limited computing resources, and track tracking precision and speed control performance of the unmanned vehicle under different roads and working conditions are improved on the premise of guaranteeing closed loop stability and constraint safety through tight coupling with a model prediction control framework. Disclosure of Invention The invention provides a vehicle path tracking and longitudinal speed control method based on Gaussian process compensation. The technical scheme of the invention is that the vehicle path tracking and longitudinal speed control method based on Gaussian process compensation comprises the following steps: S1, collecting data of a vehicle on a calibration road, constructing a Gaussian process residual model, and performing offlin