A data-driven based real-time estimation method for tire cornering stiffness

Abstract

The present application is suitable for the technical field of online estimation of vehicle dynamics parameters and intelligent chassis control, and provides a real-time estimation method for tire cornering stiffness based on data driving, which comprises the following steps: obtaining a tire mechanics anchor point data set and performing data intensive processing; using a parameterized tire model for global fitting and expansion across load; equivalent road adhesion difference to load scale change, constructing load-cornering angle domain enhanced training data set; based on the load-cornering angle domain enhanced training data set, the key feature subset is screened through feature importance evaluation, and an explicit closed cornering stiffness proxy model is constructed. Real-time acquisition of cornering angle and vertical load, calculation of equivalent load, input of the cornering angle and vertical load into the proxy model, and direct output of the tire cornering stiffness estimation value. The method realizes low complexity, high real-time and numerical stability of the cornering stiffness estimation, is convenient for on-board deployment, and provides reliable parameter support for vehicle state estimation and motion control.

Description

Technical Field

[0001] This invention belongs to the field of online estimation of vehicle dynamic parameters and intelligent chassis control technology, and particularly relates to a data-driven real-time estimation method for tire lateral stiffness. Background Technology

[0002] With the development of vehicle electrification and intelligence, vehicle chassis control is gradually shifting from traditional axle-level control to wheel-level coordinated control. Tire lateral stiffness, as a key parameter reflecting the variation of tire lateral force with slip angle, directly affects the accuracy and robustness of algorithms for vehicle state estimation, yaw stability control, trajectory tracking control, and wheel-end force distribution. In actual driving, tire lateral stiffness is significantly affected by factors such as vertical load, road surface adhesion level, temperature, tire pressure, wear, and tire operating range, exhibiting strong nonlinearity and time-varying characteristics.

[0003] In existing technologies, one type of method relies on parametric tire models for parameter identification or online updates. However, under multiple loads, multiple adhesions, and rapid dynamic conditions, it is prone to problems such as numerical sensitivity, unstable derivatives in the saturation region, and limited convergence speed. Furthermore, embedded implementations typically involve high computational complexity, making it difficult to balance real-time performance and generalization. Another type of method uses machine learning models to directly establish a mapping from load parameters to side slip stiffness. While this can improve fitting ability, it generally suffers from strong dependence on training data coverage, insufficient cross-load condition transfer capability, weak interpretability, and high computational burden for online deployment. In addition, the high cost of tire mechanical testing and calibration leads to sparse and uneven distribution of training samples in the load-side slip angle domain, further exacerbating the risk of overfitting and the difficulty of engineering implementation.

[0004] Therefore, there is an urgent need for a real-time tire lateral stiffness estimation method that can take into account cross-load / cross-adhesion coverage, interpretability, and low-complexity deployment based on engineering-accessible data. Summary of the Invention

[0005] The purpose of this invention is to provide a data-driven real-time estimation method for tire lateral stiffness, which aims to solve the problems mentioned in the background art.

[0006] The present invention is implemented as follows: a data-driven real-time estimation method for tire lateral stiffness includes the following steps:

[0007] Step 1: Obtain the tire mechanical anchor point dataset, which includes at least the slip angle, vertical load, and lateral force or the slip stiffness derived from the lateral force.

[0008] Step 2: Based on the tire mechanical anchor point dataset, perform data densification processing on the sideslip angle dimension to improve the sample coverage of small sideslip angles and transition areas, and obtain denser samples;

[0009] Step 3: Based on the dense samples, a parametric tire model is used to perform global fitting and expansion across loads to generate cross-load expanded samples, thereby improving the consistency and usability of the data under multi-load conditions.

[0010] Step 4: Equivalent the road surface adhesion difference to the load scale change, normalize the road surface adhesion coefficient, and perform scale transformation on the vertical load based on the normalized road surface adhesion coefficient to obtain the equivalent load, thereby constructing a load-side slip angle domain enhancement training dataset to cover more adhesion levels with less data.

[0011] Step 5: Establish a machine learning prediction model based on the load-side deviation angle domain enhanced training dataset, and determine the key feature subset through feature importance evaluation, so as to reduce feature redundancy and overfitting risk while ensuring the expression of the main physical relationships;

[0012] Step 6: Construct an explicit closed lateral stiffness surrogate model on the key feature subset using sparse regression and regularized regression, so that the model outputs the lateral stiffness estimate in the form of a fixed-order mapping.

[0013] Step 7: During online operation, the sideslip angle and vertical load are acquired in real time. After calculating the equivalent load, the equivalent load and sideslip angle are input into the explicit closed sideslip stiffness surrogate model, and the estimated tire sideslip stiffness is output. Online inference does not require iterative solution, which meets the requirements of real-time vehicle deployment.

[0014] In a further technical solution, in step 1, the anchor point data in the tire mechanical anchor point dataset can come from tire test benches, real vehicle data, or simulation data, or a combination of the above sources;

[0015] The tire mechanics anchor point dataset consists of single-wheel measured data or single-wheel equivalent data obtained by converting vehicle state estimation results with tire mechanics relationships.

[0016] In a further technical solution, in step 2, the data intensive processing includes one or more of the following methods: reducing the side slip angle sampling interval for encrypted sampling; interpolating and resampling the side slip angle-lateral force curve; and performing weighted encrypted sampling in the small side slip angle region and the transition region.

[0017] In a further technical solution, in step 3, the parametric tire model is preferably a semi-empirical tire model, a magic formula tire model, or an equivalent form thereof;

[0018] The global fitting across loads is extended to nonlinear optimization fitting, preferably nonlinear least squares fitting; the global fitting across loads uses joint optimization across loads to obtain model parameters, so as to avoid inconsistencies across loads caused by fitting each load separately.

[0019] In a further technical solution, in step 4, the normalized road surface adhesion coefficient is obtained by scaling, interval normalization or segmented mapping of the road surface adhesion coefficient, and is limited to a preset range; when the road surface adhesion coefficient is missing, the default value or the most recent valid value is used as a substitute.

[0020] A further technical solution is to introduce symmetric data processing when constructing the load-side offset angle domain enhanced training dataset in step 4, so as to reduce zero-point jitter and improve sample coverage.

[0021] Among them, symmetry data processing aims to reduce jitter caused by sign switching near the zero point and improve the utilization rate of effective samples. Absolute value transformation is performed on negative side-slip angle samples to concentrate the samples in the positive side-slip angle range. If necessary, the direction label can be retained to support the sign recovery of subsequent control links.

[0022] In a further technical solution, in step 5, the machine learning prediction model is a tree ensemble regression model, a lightweight neural network regression model, a kernel method regression model, or any combination thereof; preferably, the tree ensemble regression model includes a random forest model and / or a gradient boosting tree model; the above models are only used as carriers for offline modeling and feature selection, and are not limited to specific network structures or tree model forms.

[0023] In a further technical solution, in step 5, the feature importance assessment includes permutation importance assessment: using the validation set error or out-of-bag error as the baseline error, the feature is randomly permuted and the error is recalculated, and the error increment is used as the feature importance, thereby determining the minimum but sufficient subset of key features.

[0024] A further technical solution is that, in step 6, the explicit closed lateral stiffness surrogate model is obtained through "sparse regression + regularized regression"; first, LASSO is used for sparse selection, and then ridge regression is used for parameter fine-tuning; the parameter fine-tuning introduces physical constraints, which include at least one of the following: non-negative lateral stiffness constraint; lateral stiffness monotonically does not decrease with respect to equivalent load within a preset small lateral angle range; output continuity or smoothness constraint with respect to lateral angle; saturation or boundary constraints are applied in the high lateral angle region.

[0025] A further technical solution is that the explicit closed lateral stiffness surrogate model is a low-order polynomial surrogate model, whose coefficients are constants obtained through offline training. In the online stage, the equivalent load and lateral angle are used as inputs to perform order determination calculations and output the estimated lateral stiffness value.

[0026] A further technical solution, in step 7, includes out-of-domain processing and confidence output: when the online input exceeds the training domain threshold, the input is subjected to at least one of the following: amplitude limiting, segmented extrapolation, rollback to the parameterized tire model output, or rollback to the previous time-series estimate; and based on at least one of the following: the distance of the input from the training domain boundary, the model residual statistic, or the normalized squared innovation (NIS), the confidence or health index of the lateral stiffness estimate is output; the lateral stiffness estimate and / or confidence or health index are used for lateral force estimation, vehicle state estimation update, trajectory tracking controller parameter adaptation, or wheel-end force distribution control.

[0027] The present invention provides a data-driven real-time estimation method for tire lateral stiffness, the advantages of which are as follows:

[0028] (1) By increasing the data density of anchor points and expanding the global fitting across loads, the sample coverage is improved and the consistency across loads is enhanced, thus alleviating the problem of sample sparsity and unevenness.

[0029] (2) By constructing an equivalent load that “equivalent to load scale change”, more adhesion levels can be approximated with a small amount of data, significantly reducing the cost of cross-road calibration;

[0030] (3) By “selecting key features by permutation importance + constructing an explicit closed surrogate model by sparse / regular regression”, the risk of overfitting and the burden of online computation are reduced while ensuring the expression of the main physical relationships;

[0031] (4) The online stage uses a fixed-order mapping form to output the estimated lateral stiffness value, which does not require iterative solution, the delay is predictable, the value is more stable, and it is suitable for real-time deployment on the vehicle. Attached Figure Description

[0032] Figure 1 A flowchart of a data-driven real-time estimation method for tire lateral stiffness provided in an embodiment of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0034] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0035] like Figure 1 As shown, an embodiment of the present invention provides a data-driven real-time estimation method for tire lateral stiffness, comprising the following steps:

[0036] Step 1: Obtain the tire mechanical anchor point dataset, which includes at least the slip angle, vertical load, and lateral force or the slip stiffness derived from the lateral force.

[0037] Step 2: Based on the tire mechanical anchor point dataset, perform data densification processing on the sideslip angle dimension to improve the sample coverage of small sideslip angles and transition areas, and obtain denser samples;

[0038] Step 3: Based on the dense samples, a parametric tire model is used to perform global fitting and expansion across loads to generate cross-load expanded samples, thereby improving the consistency and usability of the data under multi-load conditions.

[0039] Step 4: Equivalent the road surface adhesion difference to the load scale change, normalize the road surface adhesion coefficient, and perform scale transformation on the vertical load based on the normalized road surface adhesion coefficient to obtain the equivalent load, thereby constructing a load-side slip angle domain enhancement training dataset to cover more adhesion levels with less data.

[0040] Step 5: Establish a machine learning prediction model based on the load-side deviation angle domain enhanced training dataset, and determine the key feature subset through feature importance evaluation, so as to reduce feature redundancy and overfitting risk while ensuring the expression of the main physical relationships;

[0041] Step 6: Construct an explicit closed lateral stiffness surrogate model on the key feature subset using sparse regression and regularized regression, so that the model outputs the lateral stiffness estimate in the form of a fixed-order mapping.

[0042] Step 7: During online operation, the sideslip angle and vertical load are acquired in real time. After calculating the equivalent load, the equivalent load and sideslip angle are input into the explicit closed sideslip stiffness surrogate model, and the estimated tire sideslip stiffness is output. Online inference does not require iterative solution, which meets the requirements of real-time vehicle deployment.

[0043] In a preferred embodiment of the present invention, in step 1, the anchor point data in the tire mechanics anchor point dataset can come from tire test benches, real vehicle data, or simulation data, or a combination of the above sources. The tire mechanics anchor point dataset is single-wheel measured data or single-wheel equivalent data obtained by converting vehicle state estimation results with tire mechanics relationships.

[0044] In a preferred embodiment of the present invention, in step 2, the data intensive processing includes one or more of the following methods: reducing the side slip angle sampling interval for encrypted sampling; interpolating and resampling the side slip angle-lateral force curve; and performing weighted encrypted sampling in the small side slip angle region and the transition region.

[0045] As a preferred embodiment of the present invention, in step 3, the parametric tire model is preferably a semi-empirical tire model, a magic formula tire model, or an equivalent form thereof.

[0046] The global fitting across loads is extended to nonlinear optimization fitting, preferably nonlinear least squares fitting; the global fitting across loads uses joint optimization across loads to obtain model parameters, so as to avoid inconsistencies across loads caused by fitting each load separately.

[0047] In a preferred embodiment of the present invention, in step 4, the normalized road surface adhesion coefficient is obtained by scaling, interval normalization or segmented mapping of the road surface adhesion coefficient, and is limited to a preset range; when the road surface adhesion coefficient is missing, a default value or the most recent valid value is used as a substitute.

[0048] In a preferred embodiment of the present invention, in step 4, when constructing the load-side slip angle domain augmentation training dataset, symmetric data processing is introduced to reduce zero-point jitter and improve sample coverage. Specifically, symmetric data processing aims to reduce jitter caused by sign switching near the zero point and improve the utilization rate of effective samples. Absolute value transformation is performed on negative slip angle samples to concentrate samples within the positive slip angle range; if necessary, direction labels can be retained to support sign recovery in subsequent control steps.

[0049] In a preferred embodiment of the present invention, in step 5, the machine learning prediction model is a tree ensemble regression model, a lightweight neural network regression model, a kernel method regression model, or any combination thereof; preferably, the tree ensemble regression model includes a random forest model and / or a gradient boosting tree model; the above models are only used as carriers for offline modeling and feature selection, and are not limited to specific network structures or tree model forms.

[0050] In a preferred embodiment of the present invention, in step 5, the feature importance assessment includes permutation importance assessment: using the validation set error or out-of-bag error as the baseline error, the feature is randomly permuted and the error is recalculated, and the error increment is used as the feature importance, thereby determining the minimum but sufficient subset of key features.

[0051] In a preferred embodiment of the present invention, in step 6, the explicit closed lateral stiffness surrogate model is obtained through "sparse regression + regularized regression"; first, LASSO is used for sparse selection, and then ridge regression is used for parameter fine-tuning; the parameter fine-tuning introduces physical constraints, which include at least one of the following: non-negative lateral stiffness constraint; lateral stiffness monotonically does not decrease with respect to equivalent load within a preset small lateral angle range; output continuity or smoothness constraint with respect to lateral angle; saturation or boundary constraints are applied in the high lateral angle region.

[0052] In a preferred embodiment of the present invention, the explicit closed lateral stiffness surrogate model is a low-order polynomial surrogate model, whose coefficients are constants obtained through offline training. In the online stage, the equivalent load and lateral angle are used as inputs to perform order determination calculations and output the estimated lateral stiffness value.

[0053] In a preferred embodiment of the present invention, step 7 further includes out-of-domain processing and confidence output: when the online input exceeds the training domain threshold, the input is subjected to at least one of the following: amplitude limiting, segmented extrapolation, rollback to the parameterized tire model output, or rollback to the previous time-series estimate; and based on at least one of the following: the distance of the input from the training domain boundary, the model residual statistic, or the normalized squared innovation (NIS), the confidence or health index of the lateral stiffness estimate is output; the lateral stiffness estimate and / or confidence or health index are used for lateral force estimation, vehicle state estimation update, trajectory tracking controller parameter adaptation, or wheel-end force distribution control.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A data-driven real-time estimation method for tire lateral stiffness, characterized in that, Includes the following steps: Step 1: Obtain the tire mechanical anchor point dataset, which includes at least the slip angle, vertical load, and lateral force or the slip stiffness derived from the lateral force. Step 2: Based on the tire mechanical anchor point dataset, perform data densification processing on the sideslip angle dimension to improve the sample coverage of small sideslip angles and transition areas, and obtain denser samples; Step 3: Based on the dense samples, a parametric tire model is used to perform global fitting and expansion across loads to generate expanded samples across loads; Step 4: Equivalently transform the road surface adhesion difference into a load scale change, normalize the road surface adhesion coefficient, and perform a scale transformation on the vertical load based on the normalized road surface adhesion coefficient to obtain the equivalent load, thereby constructing a load-side slip angle domain augmentation training dataset. Step 5: Build a machine learning prediction model based on the load-side deviation angle domain augmented training dataset, and determine the key feature subset through feature importance evaluation; Step 6: Construct an explicit closed lateral stiffness surrogate model on the key feature subset using sparse regression and regularized regression, so that the model outputs the lateral stiffness estimate in the form of a fixed-order mapping. Step 7: During online operation, the slip angle and vertical load are acquired in real time. After calculating the equivalent load, the equivalent load and slip angle are input into the explicit closed slip stiffness surrogate model, and the estimated tire slip stiffness value is output.

2. The data-driven real-time tire lateral stiffness estimation method according to claim 1, characterized in that, In step 1, the anchor point data in the tire mechanical anchor point dataset comes from one or more of the following types: tire test bench, real vehicle data, or simulation data. The tire mechanics anchor point dataset consists of single-wheel measured data or single-wheel equivalent data obtained by converting vehicle state estimation results with tire mechanics relationships.

3. The data-driven real-time estimation method for tire lateral stiffness according to claim 1, characterized in that, In step 2, the data-intensive processing includes one or more of the following methods: reducing the side slip angle sampling interval for encrypted sampling; interpolating and resampling the side slip angle-lateral force curve; and performing weighted encrypted sampling in the small side slip angle region and the transition region.

4. The data-driven real-time estimation method for tire lateral stiffness according to claim 1, characterized in that, In step 3, the parameterized tire model is a semi-empirical tire model, a magic formula tire model, or an equivalent form thereof; The global fitting and extension across loads employ nonlinear least squares fitting; the global fitting across loads uses joint optimization across loads to obtain model parameters.

5. The data-driven real-time estimation method for tire lateral stiffness according to claim 1, characterized in that, In step 4, the normalized road surface adhesion coefficient is obtained by scaling, interval normalization or piecewise mapping of the road surface adhesion coefficient, and is limited to a preset range; when the road surface adhesion coefficient is missing, the default value or the most recent valid value is used as a substitute.

6. The data-driven real-time estimation method for tire lateral stiffness according to claim 1, characterized in that, In step 4, when constructing the load-side slip angle domain augmentation training dataset, symmetry data processing is introduced; Symmetrical data processing involves performing an absolute value transformation on the negative side deflection angle samples to concentrate the samples within the positive side deflection angle range.

7. The data-driven real-time estimation method for tire lateral stiffness according to claim 1, characterized in that, In step 5, the machine learning prediction model is a tree ensemble regression model, a lightweight neural network regression model, a kernel method regression model, or any combination thereof.

8. The data-driven real-time estimation method for tire lateral stiffness according to claim 1, characterized in that, In step 5, the feature importance assessment includes permutation importance assessment: using the validation set error or out-of-bag error as the baseline error, the feature is randomly permuted and the error is recalculated, and the error increment is used as the feature importance, thereby determining the minimum but sufficient subset of key features.

9. The data-driven real-time estimation method for tire lateral stiffness according to claim 1, characterized in that, In step 6, the explicit closed lateral stiffness surrogate model is obtained through "sparse regression + regularized regression"; first, LASSO is used for sparsity selection, and then ridge regression is used for parameter fine-tuning; the parameter fine-tuning introduces physical constraints, which include at least one of the following: non-negativity constraint of lateral stiffness; constraint that lateral stiffness is monotonically non-decreasing with respect to equivalent load within a preset small lateral angle range; constraint that the output is continuous or smooth with respect to the lateral angle; and constraint that saturation or boundary constraints are applied in the high lateral angle region.

10. The data-driven real-time estimation method for tire lateral stiffness according to claim 1, characterized in that, Step 7 further includes out-of-domain processing and confidence output: when the online input exceeds the training domain threshold, the input is subjected to at least one of the following: amplitude limiting, segmented extrapolation, backtracking to the parameterized tire model output, or backtracking to the previous time-series estimate; and based on at least one of the following: the distance of the input from the training domain boundary, the model residual statistic, or the normalized squared innovation (NIS), the confidence or health index of the lateral stiffness estimate is output; the lateral stiffness estimate and / or confidence or health index are used for lateral force estimation, vehicle state estimation update, trajectory tracking controller parameter adaptation, or wheel-end force distribution control.

Images