CN-122021240-A - Unscented Kalman filter friction coefficient estimation method, unscented Kalman filter friction coefficient estimation system and unscented Kalman filter friction coefficient estimation application based on time convolution residual network driving

CN122021240ACN 122021240 ACN122021240 ACN 122021240ACN-122021240-A

Abstract

The invention relates to a unscented Kalman filter friction coefficient estimation method, a unscented Kalman filter friction coefficient estimation system and application based on a time convolution residual error network drive, which are used for establishing a vehicle dynamics model, constructing a time convolution residual error integrated network and a double-channel unscented Kalman filter structure, inputting time sequence data of a vehicle state into the time convolution residual error integrated network, sequentially inputting vehicle observation information and network output obtained based on a Dugoff model into the double-channel unscented Kalman filter structure as pseudo-measurement input, carrying out state correction, and taking an updated state as an optimal friction coefficient estimation value at the current moment. The method and the device remarkably improve stability and precision of the estimation result, can effectively correct systematic deviation in preliminary friction coefficient estimation, provide a more robust estimation result under complex working conditions, and realize smooth, accurate and stable dynamic estimation of the friction coefficient.

Inventors

SONG XIULAN
TONG ZHENGYANG
HE DEFENG
WU YANHONG
MU JIANBIN
LU WEIDANG

Assignees

浙江工业大学

Dates

Publication Date: 20260512
Application Date: 20251217

Claims (8)

1. A unscented Kalman filter friction coefficient estimation method based on time convolution residual error network driving is characterized by establishing a vehicle dynamics model; And inputting the time-series data of the vehicle state into the time convolution residual error integrated network, sequentially inputting the vehicle observation information obtained based on the Dugoff model and the output of the network into the dual-channel unscented Kalman filtering structure as pseudo-measurement input, and carrying out state correction, wherein the updated state is the friction coefficient optimal estimated value at the current moment.
2. The unscented Kalman filter friction coefficient estimation method based on the time convolution residual network driving of claim 1, wherein the time convolution residual integrated network comprises a plurality of sub-networks, and any one of the sub-networks comprises a time convolution module and a residual regression module which are cooperatively arranged.
3. The method for estimating a unscented Kalman filter friction coefficient based on a time convolution residual network drive of claim 2, characterized by characterizing the vehicle state Output of time convolution module And estimation value of prediction residual And training the residual regression module, and taking the difference value between the time convolution module and the residual regression module as friction coefficient correction estimation.
4. The unscented Kalman filter friction coefficient estimation method based on the time convolution residual network driving of claim 3, wherein all the sub-networks have the same structure and different initial parameters; And calculating the mean value and variance of the integrated prediction based on the friction coefficient correction estimates output by all the sub-networks.
5. The method for estimating the unscented Kalman filter friction coefficient based on the time convolution residual network driving of claim 1, wherein the state correction by the dual-channel unscented Kalman filter structure comprises the following steps: s3.1, modeling a system state, and establishing a double-observation mechanism; S3.2, setting an initial state estimation value and covariance, and generating a preset Sigma point through unscented transformation; S3.3, performing state prediction and covariance prediction based on a nonlinear state transfer function of the system; S3.4, updating the state and the covariance for the first time based on a first observation equation of a double observation mechanism; s3.5, generating a new Sigma point based on the state mean value and covariance after the first updating, and executing the second updating based on a second observation equation of the double observation mechanism.
6. The method for estimating a unscented Kalman filter friction coefficient based on a time convolution residual network drive of claim 5 wherein the first observation equation is associated with a state vector of the vehicle, a normalized tire force, and Gaussian white noise; And the second observation equation is associated with the mean value and the variance of the output of the time convolution residual error integrated network and Gaussian white noise.
7. A unscented Kalman filter friction coefficient estimation system based on time convolution residual error network driving is characterized by comprising At least one processor, and A memory communicatively coupled to at least one of the processors; wherein the memory stores instructions executable by the processor for execution by the processor to implement the unscented kalman filter friction coefficient estimation method based on a time convolution residual network drive of one of claims 1 to 6.
8. An application of the unscented Kalman filter friction coefficient estimation method based on the time convolution residual network driving of one of the claims 1-6 is characterized in that the estimation method is applied to the estimation of the friction coefficient in the vehicle driving process.

Description

Unscented Kalman filter friction coefficient estimation method, unscented Kalman filter friction coefficient estimation system and unscented Kalman filter friction coefficient estimation application based on time convolution residual network driving Technical Field The invention relates to the technical field of judgment or calculation of driving parameters of a road vehicle driving control system which is not related to control of a specific subsystem, in particular to a unscented Kalman filter friction coefficient estimation method, a unscented Kalman filter friction coefficient estimation system and application based on time convolution residual error network driving. Background With the development of automatic driving technology, the control safety and stability of intelligent vehicles become particularly important. Tire-road friction coefficient (TRFC) is a key parameter affecting vehicle dynamic performance and stability, and its accurate estimation is critical to the control accuracy of intelligent driving systems. However, TRFCs present indirect testability and operating condition dependencies that make accurate estimation thereof a significant challenge. At present, the mainstream TRFC estimation method mostly depends on a traditional physical model or an empirical formula, and the method has a certain effect under the conventional working condition, but has larger estimation error when facing complex driving scenes and severe road conditions. Moreover, such methods typically have high computational resource requirements that limit their real-time application in real-vehicle systems. In recent years, with breakthrough of deep learning technology, a TRFC estimation method based on data driving has great potential. Among other things, the time-sequential convolutional network (TCN) offers a new solution for TRFC estimation by virtue of its unique advantage in processing time-series data. However, the deep learning method has a limitation in coping with the uncertainty of the system, which may cause a decrease in estimation accuracy. To solve this problem, unscented Kalman Filtering (UKF) was introduced as an efficient nonlinear filtering technique to improve estimation accuracy and system robustness. Therefore, how to realize accurate and real-time estimation of the tire-road friction coefficient in complex and changeable driving environments has become a key technical problem for improving the performance of the intelligent driving system, and development of an innovative solution integrating the advantages of multiple methods is needed. Disclosure of Invention The invention solves the problems in the prior art, and provides a unscented Kalman filtering friction coefficient estimation method, a unscented Kalman filtering friction coefficient estimation system and application based on a time convolution residual error network drive. The technical scheme adopted by the invention is that the unscented Kalman filter friction coefficient estimation method based on the time convolution residual error network drive establishes a vehicle dynamics model; And inputting the time-series data of the vehicle state into the time convolution residual error integrated network, sequentially inputting the vehicle observation information obtained based on the Dugoff model and the output of the network into the dual-channel unscented Kalman filtering structure as pseudo-measurement input, and carrying out state correction, wherein the updated state is the friction coefficient optimal estimated value at the current moment. Preferably, the time convolution residual integrated network comprises a plurality of sub-networks, and any sub-network comprises a time convolution module and a residual regression module which are cooperatively arranged. Preferably, characterised by the state of the vehicleOutput of time convolution moduleAnd estimation value of prediction residualAnd training the residual regression module, and taking the difference value between the time convolution module and the residual regression module as friction coefficient correction estimation. Preferably, all the sub-networks have the same structure and different initial parameters; And calculating the mean value and variance of the integrated prediction based on the friction coefficient correction estimates output by all the sub-networks. Preferably, the state correction with the dual-channel unscented kalman filter structure includes the steps of: s3.1, modeling a system state, and establishing a double-observation mechanism; S3.2, setting an initial state estimation value and covariance, and generating a preset Sigma point through unscented transformation; S3.3, performing state prediction and covariance prediction based on a nonlinear state transfer function of the system; S3.4, updating the state and the covariance for the first time based on a first observation equation of a double observation mechanism; s3.5, generating a new Sigma point based on