CN-121995917-A - Unmanned vehicle track tracking model predictive control method based on hybrid offline-online Gaussian process regression

CN121995917ACN 121995917 ACN121995917 ACN 121995917ACN-121995917-A

Abstract

The invention provides an unmanned vehicle track tracking model prediction control method based on mixed offline-online Gaussian process regression, which comprises the steps of constructing a nominal vehicle dynamics model to serve as a state prediction standard, training an offline Gaussian process regression model, learning unmodeled dynamics in vehicle dynamics based on historical driving data, constructing an online Gaussian process regression model in the real-time running process of a vehicle, dynamically maintaining a capacity-limited online data set based on real-time acquired vehicle states and control input data increment training, evaluating a prediction variance of new data, respectively weighting and fusing a prediction mean value and a prediction variance output by the two models to obtain a mixed Gaussian process regression model, combining the mixed Gaussian process regression model and the nominal vehicle dynamics model to construct an enhanced prediction model, solving a limited time domain optimization problem with constraint in each control period, and outputting a control instruction through rolling optimization to realize track tracking.

Inventors

CHEN YUTAO
HE JIE
WU CHANGHUA
ZHOU YU
ZHENG FENG

Assignees

福州大学

Dates

Publication Date: 20260508
Application Date: 20260202

Claims (10)

1. An unmanned vehicle trajectory tracking control method, comprising: Constructing a nominal vehicle dynamics model serving as a state prediction reference; training an offline Gaussian process regression model, and learning unmodeled dynamics in vehicle dynamics based on historical driving data; In the real-time running process of the vehicle, an online Gaussian process regression model is built, based on the vehicle state acquired in real time and the incremental training of control input data, a capacity-limited online data set is dynamically maintained by evaluating the prediction variance of new data, an induction point set is dynamically configured along a prediction state track generated by model prediction control, and Gaussian process reasoning is performed by adopting a sparse approximation method; Based on prediction error statistics of the offline and online Gaussian process regression models in the sliding time window, dynamically determining fusion weights, and respectively weighting and fusing prediction means and prediction variances output by the two models to obtain a Gaussian mixture process regression model; and combining the Gaussian mixture process regression model with a nominal vehicle dynamics model to construct an enhanced prediction model, solving a constrained finite time domain optimization problem in each control period, and outputting a control instruction through rolling optimization to realize track tracking.
2. The method of claim 1, wherein the nominal vehicle dynamics model is a bicycle model structure, and the state space comprises a longitudinal position, a transverse position, a heading angle, a longitudinal speed and a yaw rate, and the control input comprises a longitudinal acceleration and a front wheel rotation angle.
3. The unmanned vehicle trajectory tracking control method of claim 1, wherein the offline Gaussian process regression model is trained based on a historical experimental data set of a nominal vehicle dynamics model, ultra-parameter post-curing is optimized, a joint vector of a vehicle state and a control quantity is used as input, and a residual error of a system actual state and a predicted output of the nominal vehicle dynamics model is used as a learning target.
4. The method for tracking and controlling the track of the unmanned vehicle according to claim 1, wherein the dynamic maintenance capacity-limited online data set is characterized in that the data set is added when new data meets the prediction variance related screening condition, and low-value samples are eliminated preferentially in an aging weighting mode when the data set reaches the upper limit of the capacity.
5. The unmanned vehicle track following control method of claim 1, wherein the sparse approximation method approximates the statistical characteristics of the original dataset by a function distribution of induction points, and the induction points are dynamically configured along a state track generated by model predictive control.
6. The unmanned vehicle track tracking control method of claim 1, wherein the parameters of the online Gaussian process regression model are updated according to a preset period, the updating process is executed in a background thread, a new collected sample is temporarily stored in a buffer queue during updating, and the new collected sample is uniformly added into an online data set after updating is completed.
7. The method for tracking and controlling the track of the unmanned vehicle according to claim 1, wherein the step of dynamically determining the fusion weight is to calculate the average absolute error of two models in a sliding time window and adaptively allocate the fusion proportion of the two models based on the average absolute error.
8. The method for tracking and controlling the trajectory of an unmanned vehicle according to claim 1, wherein the enhanced predictive model maps the state residuals to an unmodeled dynamic space through a preset mapping matrix, and the unmodeled dynamic and process noise mainly affects the longitudinal speed and the yaw rate of the vehicle.
9. The method of claim 1, wherein the constraints in the constrained finite time domain optimization problem include control input upper and lower limit constraints and opportunity constraints, and the opportunity constraints are converted into deterministic constraints by a standard normal distribution quantile function to ensure that the system meets state constraints at a preset confidence level.
10. An unmanned vehicle comprising a controller configured to perform the method of any of claims 1-9.

Description

Unmanned vehicle track tracking model predictive control method based on hybrid offline-online Gaussian process regression Technical Field The invention belongs to the technical field of intelligent control and automatic driving, and particularly relates to an unmanned vehicle track tracking model predictive control method based on mixed offline-online Gaussian process regression. Background With the deep development of automatic driving technology, the requirements of model-based control methods for model accuracy in vehicle motion control are continuously improved. The MPC has the advantages of being capable of explicitly processing multidimensional constraint, simultaneously considering multiple physical limitations such as actuator saturation, stability boundary and the like, being capable of predicting future dynamics of a system based on an optimization mechanism of forward prediction and providing prospective for control decision, and being naturally suitable for a multi-input multi-output system and capable of processing coupling relation of longitudinal control and transverse control. However, the control performance of MPC is highly dependent on the accuracy of the predictive model. The actual vehicle system has complex nonlinear dynamics characteristics, including complex phenomena such as tire force saturation, load transfer and the like, and the traditional modeling method based on mechanism analysis is difficult to completely describe all dynamic characteristics of the system, so that a significant mismatch phenomenon exists between the model and the actual system. GPR is used as a non-parameterization method based on a Bayesian framework, firstly, the GPR can provide complete prediction distribution, not only output a prediction mean value, but also explicitly quantize prediction uncertainty, which provides important basis for risk assessment and robust decision making of a control system, secondly, the GPR has strong nonlinear fitting capability, complex system dynamic characteristics can be accurately learned without presetting a model structure, and furthermore, the method has lower prior knowledge requirement, and can realize flexible modeling by defining similarity relations among data through a kernel function. The existing GPR application research is mainly developed around two technical routes, namely a GPR method based on offline learning utilizes sufficient historical data to establish a static compensation model, higher fitting precision can be ensured in a training working condition range, but adaptability to environment dynamic changes is limited, and a GPR method based on online learning tracks system changes by updating model parameters in real time, so that the GPR method has good environment adaptability, and the calculation complexity of the GPR method is increased in a cubic level along with accumulation of data quantity, so that the real-time requirement of a control system is difficult to meet. Although the fusion research of the GPR and the MPC has been advanced to a certain extent, a plurality of challenges still face in engineering application, namely firstly, a single offline or online learning mode is difficult to balance between model stability and environmental adaptability, secondly, the existing method lacks an effective dynamic fusion mechanism and cannot intelligently adjust a learning strategy according to working condition changes, and in addition, the application of the method on a resource-limited platform is restricted by the bottleneck of calculation efficiency in the online learning process. Particularly, when dealing with highly nonlinear vehicle dynamics systems, how to ensure the real-time performance of the controller while maintaining higher tracking accuracy is still a technical problem to be solved. Disclosure of Invention Aiming at the defects and shortcomings in the prior art, the invention provides an unmanned vehicle track tracking model prediction control method based on mixed offline-online Gaussian process regression, which aims to solve the technical problems of insufficient control precision caused by model mismatch in the traditional Model Prediction Control (MPC), weak offline adaptability of a single Gaussian Process Regression (GPR) model, high online calculation complexity and lack of an effective fusion mechanism. The method comprises the steps of firstly constructing a vehicle nominal dynamics model as a state prediction standard, simultaneously constructing a residual error learning framework of the cooperation of an offline GPR model and an online GPR model, wherein the offline GPR model is trained based on a large-scale historical data set and solidifies super parameters to provide priori knowledge of unmodeled dynamics of a system, the online GPR model is trained by vehicle states acquired in real time and control input data increment, the information value of new data is evaluated based on prediction variance to dynamically ma