A neural network-based tire six-force prediction method and system

Abstract

The present application relates to the technical field of tire performance prediction based on artificial intelligence, and particularly relates to a tire six-component force prediction method and system based on a neural network. The method obtains structural parameters, material parameters, working condition parameters and corresponding six-component force test data of a plurality of tire samples, and constructs a structure representation vector, a material representation vector and a working condition coupling state vector after preprocessing. A prediction network including a three-branch encoding module, a feature fusion module, an intermediate physical quantity prediction module, a basic mechanics model module and a residual correction module is established. The network is trained in layers using pure side slip, pure longitudinal slip and complex working conditions, and a consistent constraint on the return torque is introduced. Finally, the target tire parameters and target working condition parameters are input, and the longitudinal force, lateral force, normal force, roll torque, rolling resistance torque and return torque are output. The method has the advantages of high prediction accuracy, strong physical rationality, wide working condition coverage and high engineering application value.

Description

Technical Field

[0001] This invention relates to the field of tire performance prediction technology based on artificial intelligence, and in particular to a method and system for predicting the six components of tire force based on neural networks. Background Technology

[0002] As the only component in direct contact between a vehicle and the road surface, the tire's stress state directly determines the vehicle's driving, braking, steering, and driving stability. In vehicle dynamics analysis, the six components of tire forces typically refer to the forces and moments generated in three directions during tire-road contact, specifically including longitudinal force, lateral force, normal force, rollover torque, rolling resistance torque, and self-aligning torque. These six components are not only crucial foundational data for tire structure design, tread compound development, and vehicle matching optimization, but also key input parameters for vehicle stability control, path tracking control, active safety control, and digital prototype simulation.

[0003] In existing technologies, the main methods for obtaining the six-component forces of a tire include experimental testing, theoretical modeling, and numerical simulation. Experimental testing typically relies on a six-component force test bench, a vehicle road testing system, or dedicated tire dynamics testing equipment to directly measure the tire's mechanical response under predetermined inflation pressure, load, slip angle, slip ratio, camber angle, and road surface conditions. The advantage of this method is the high realism and reliability of the test results, but its disadvantages are equally apparent: firstly, the testing equipment is expensive, the testing cycle is long, and the coverage of sampling conditions is limited; secondly, repeated testing is required for different tire specifications, different material systems, and different road surface conditions, resulting in high R&D costs and making it difficult to meet the needs of rapid tire design iteration and batch analysis under complex operating conditions.

[0004] Theoretical modeling methods are typically based on empirical, semi-empirical, or simplified mechanical models. They predict tire mechanical properties by establishing mathematical relationships between tire structural parameters, operating parameters, and force output. While this method can reduce reliance on experiments to some extent, it usually depends on extensive experimental data for parameter identification, and the model parameters are strongly coupled with tire specifications, material properties, and operating condition ranges. When tire structural design, material systems, or operating condition boundaries change significantly, the original model often needs to be re-identified, thus limiting its adaptability to complex operating conditions, tires of different specifications, or highly nonlinear regions. Especially for torque quantities such as self-aligning torque, rollover torque, and rolling resistance torque, existing simplified models often struggle to simultaneously achieve accuracy, stability, and computational efficiency.

[0005] Numerical simulation methods primarily rely on finite element analysis, multibody dynamics coupling analysis, or fine-grained ground contact modeling. By performing high-precision discretization of the tire structure, material layers, and contact boundaries, they obtain the tire's force response under different loads and motion states. While this method can reflect the tire's mechanical behavior under complex nonlinear conditions in detail, its modeling is complex, the solution time is long, and computational resources are high. It is generally more suitable for offline analysis and local mechanism studies, and is difficult to directly apply to large-scale parameter scanning, rapid design optimization, or real-time online estimation scenarios.

[0006] In recent years, with the development of machine learning and neural network technologies, researchers have begun to introduce data-driven methods into the field of tire force estimation, aiming to improve the speed and accuracy of tire force prediction by learning the complex nonlinear mapping relationships inherent in sample data. Compared with traditional empirical models, neural networks have certain advantages in handling strongly coupled, multivariable, and highly nonlinear problems, and therefore are gradually being applied in areas such as tire longitudinal force, lateral force, and vehicle state estimation.

[0007] Chinese patent document CN111708977B discloses an online tire force estimation method based on neural networks. This document uses the vertical load, sideslip angle, and longitudinal slip ratio of the wheel as inputs, and vehicle state information such as longitudinal acceleration, lateral acceleration, and yaw acceleration as labels. It employs a parallel structure of four neural networks to achieve online estimation of the tire force of each wheel, with the output mainly consisting of the longitudinal and lateral forces of each wheel. This method can, to some extent, avoid the use of expensive tire force sensors and reduce the workload of complex tire model identification, thus possessing certain application value in vehicle active safety and online condition monitoring.

[0008] However, further analysis reveals that existing technologies still have at least the following shortcomings: First, CN111708977B focuses on online estimation of tire forces during vehicle operation. Its input parameters are mainly concentrated at the level of operating condition variables, with insufficient consideration given to the structural and material parameters of the tire itself. Therefore, it is difficult to directly reflect the influence of different tire specifications, different tire carcass configurations, and different compound systems on the six-component force characteristics. Second, this literature mainly focuses on estimating longitudinal and lateral forces, and has not formed a complete six-component force prediction system covering longitudinal force, lateral force, normal force, as well as rollover torque, rolling resistance torque, and self-aligning torque. Therefore, it is difficult to fully meet the needs of tire design analysis, vehicle handling and stability simulation, and comprehensive tire performance evaluation. Third, existing technologies usually adopt a black-box mapping method from input variables to output mechanical quantities, which fails to adequately explore intermediate physical characteristics such as tire contact patch state, pressure center offset, and force line offset. As a result, although the model can obtain certain prediction results, it is prone to problems such as insufficient physical interpretability, decreased generalization stability, or local outputs that do not conform to the laws of tire mechanics under complex composite operating conditions and operating conditions that exceed the training distribution.

[0009] Furthermore, there are significant coupling relationships among the six components of tire force. For example, the self-aligning torque is closely related to the lateral force and its line of action offset; the rollover torque is closely related to the normal load distribution and the position of the ground pressure center; and the rolling resistance torque is affected by the combined effects of rolling conditions, material hysteresis loss, and ground contact mark deformation. Existing technologies, if based solely on single-output or simple multi-output regression methods to establish prediction models, often fail to fully reflect the inherent relationships between the six components, and are also not conducive to maintaining consistent prediction accuracy and physical consistency across pure sideslip, pure longitudinal slip, and combined conditions.

[0010] Therefore, existing technologies still require a new method for predicting the six components of tire force, which can establish a more complete prediction framework for pure sideslip, pure longitudinal slip, and combined operating conditions, based on a comprehensive consideration of tire structural parameters, material parameters, and operating condition parameters. At the same time, this method should also take into account prediction accuracy, computational efficiency, operating condition coverage, and physical rationality, so as to provide a more comprehensive and reliable six-component force input for rapid tire design, vehicle dynamics simulation, online state estimation, and active safety control. Summary of the Invention

[0011] The technical objective of this invention is to address the problems of existing methods for obtaining the six components of tire force, such as high testing costs, long modeling cycles, insufficient coverage of complex working conditions, weak physical consistency of prediction results, and difficulty in taking into account differences in tire structure and materials. This invention proposes a neural network-based method for predicting the six components of tire force. By comprehensively incorporating tire structural parameters, material parameters, and working condition parameters, a high-precision prediction model for the six components of force is established for pure sideslip, pure longitudinal slip, and combined working conditions. This enables rapid, accurate, and physically plausible prediction of longitudinal force, lateral force, normal force, and each torque component, providing reliable data support for tire design optimization, vehicle dynamics analysis, and vehicle control.

[0012] Firstly, in order to achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0013] A method for predicting the six components of tire force based on a neural network includes the following steps:

[0014] S1. Obtain the original sample data of multiple tire samples under different working conditions. The original sample data includes at least the tire structure parameter set, tire material parameter set, tire working condition parameter set, and the corresponding tire six-component force measured label set.

[0015] S2. Preprocess the original sample data and construct the structural representation vector, material representation vector and working condition coupling state vector respectively. At the same time, according to the working condition type, divide the samples into pure lateral slip working condition sample subset, pure longitudinal slip working condition sample subset and composite working condition sample subset to form a hierarchical training dataset.

[0016] S3. Construct a neural network for predicting the six components of tire force;

[0017] S4. The hierarchical training dataset is used to train the tire six-component force prediction neural network. Training loss weights are set for the pure sideslip condition sample subset, the pure longitudinal slip condition sample subset, and the composite condition sample subset, respectively. A consistency constraint on the offset of the lateral force action line is introduced for the prediction of the return torque, so as to obtain the trained tire six-component force prediction model.

[0018] S5. Obtain the target structure parameter set, target material parameter set, and target working condition parameter set of the target tire. After processing by the trained tire six-component force prediction model, output the longitudinal force prediction value, lateral force prediction value, normal force prediction value, rollover torque prediction value, rolling resistance torque prediction value, and self-aligning torque prediction value of the target tire under the target working condition.

[0019] As a further improvement, in step S1, the tire structure parameter set includes at least one or more of the following parameters: tread width, section height-to-width ratio, rim diameter, carcass cord angle, belt layer angle, belt layer width, sidewall thickness, and tread pattern type code.

[0020] And / or,

[0021] In step S1, the tire material parameter set includes at least one or more of the following parameters: tread rubber equivalent elastic modulus, carcass rubber equivalent elastic modulus, belt layer rubber equivalent elastic modulus, sidewall rubber equivalent elastic modulus, loss factor, cord layer equivalent elastic modulus, and material equivalent density.

[0022] And / or,

[0023] In step S1, the tire operating condition parameter set includes at least one or more of the following parameters: inflation pressure, vertical load, slip angle, longitudinal slip ratio, camber angle, driving speed, road adhesion coefficient, and road roughness factor;

[0024] And / or,

[0025] In step S1, the tire six-component force measurement label set includes longitudinal force, lateral force, normal force, rollover torque, rolling resistance torque, and self-aligning torque.

[0026] As a further improvement, in step S2, the preprocessing includes: outlier removal or correction, missing value completion, discrete variable encoding, continuous variable normalization, time series signal filtering, and steady-state interval truncation.

[0027] For continuous variables, standardization is performed by subtracting the mean of the original variable from the training sample and then dividing by the standard deviation of the original variable from the training sample to obtain the standardized variable.

[0028] And / or,

[0029] In step S2, the working condition type classification rule is as follows: when the absolute value of the longitudinal slip ratio is less than the preset slip threshold and the absolute value of the side slip angle is greater than the preset side slip threshold, the sample is classified into a pure side slip working condition sample subset; when the absolute value of the side slip angle is less than the preset side slip threshold and the absolute value of the longitudinal slip ratio is greater than the preset slip threshold, the sample is classified into a pure longitudinal slip working condition sample subset; when the absolute value of the side slip angle is greater than the preset side slip threshold and the absolute value of the longitudinal slip ratio is greater than the preset slip threshold, the sample is classified into a composite working condition sample subset.

[0030] As a further improvement, in step S3, the tire six-component force prediction neural network includes:

[0031] The structural encoding branch is used to receive structural representation vectors and output structural latent features.

[0032] The material encoding branch is used to receive the material characterization vector and output the material's implicit features;

[0033] The operating condition coding branch is used to receive the operating condition coupled state vector and output the operating condition implicit features;

[0034] The feature fusion module is used to fuse the implicit structural features, implicit material features, and implicit working condition features, and output shared features;

[0035] Intermediate physical quantity prediction module, used to predict the set of intermediate physical characterization quantities of grounding imprints based on the shared features;

[0036] The basic mechanical model module is used to calculate the initial value set of the basic six-component forces based on the tire structure parameter set, tire material parameter set, tire operating condition parameter set, and intermediate physical characterization quantity set of the ground imprint.

[0037] The residual correction module is used to output the residual correction set corresponding to the initial value set of the basic six component forces;

[0038] The six-component force generation module is used to output the final tire six-component force prediction result based on the basic six-component force initial value set and the residual correction amount set.

[0039] As a further improvement, the intermediate physical characterization set of the grounding imprint includes at least one or more of the following parameters: grounding imprint length, grounding imprint width, longitudinal offset of the normal pressure center, lateral offset of the normal pressure center, equivalent offset distance of the lateral force action line, and average tangential shear stress of the grounding area.

[0040] As a further improvement, the training loss function in step S4 includes a six-component force error term, a layered working condition weighting term, and a self-correcting torque consistency constraint term;

[0041] The total loss function is obtained by weighted summation of a six-component force error term, a working condition weighting term, and a aligning torque consistency constraint term. The six-component force error term characterizes the error between the predicted six-component force and the measured label. The layered working condition weighting term characterizes the weighted loss of different working condition subsets. The aligning torque consistency constraint term characterizes the consistency between the predicted aligning torque and the offset of the lateral force line of action. The coefficients of the working condition weighting term and the aligning torque consistency constraint term are used to adjust the weights of the corresponding terms in the total loss function.

[0042] As a further improvement, in step S5, the final tire six-component force prediction result is obtained by adding the basic six-component force initial value set and the residual correction amount set;

[0043] And / or,

[0044] In step S5, the predicted value of the return torque is obtained by multiplying the predicted value of the lateral force by the equivalent offset distance of the line of action of the lateral force relative to the center of the tire contact area, and then adding it to the residual correction amount of the return torque.

[0045] Secondly, the present invention also provides a tire six-component force prediction system based on a neural network, comprising:

[0046] The data acquisition module is used to execute step S1;

[0047] The data preprocessing and hierarchical construction module is used to perform step S2;

[0048] A neural network building module is used to execute step S3;

[0049] The model training module is used to perform step S4;

[0050] The six-component force prediction module is used to execute step S5;

[0051] The system is configured to implement the tire six-component force prediction method.

[0052] Thirdly, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the method.

[0053] Fourthly, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the method.

[0054] This invention constructs a joint input prediction framework for tire structural parameters, material parameters, and operating condition parameters. By combining intermediate physical characterization quantities of the ground imprint, initial values ​​of the basic mechanical model, and a neural network residual correction mechanism, it achieves coordinated prediction of tire longitudinal force, lateral force, normal force, rollover moment, rolling resistance moment, and self-aligning moment. This allows for a more comprehensive characterization of the complex nonlinear mechanical response of the tire under pure sideslip, pure longitudinal slip, and combined operating conditions. Compared to traditional methods that rely on numerous bench tests, empirical formula fitting, or high-cost finite element calculations, this invention significantly shortens the six-component force acquisition cycle, improves prediction efficiency and operating condition coverage, especially… It maintains good prediction accuracy and generalization ability under different tire specifications, different material systems, and variable operating conditions. At the same time, since this invention does not simply use black-box regression to output the six-component force, but incorporates physical correlations such as the aligning torque, ground pressure distribution characteristics, and force line offset into the prediction process, it effectively enhances the consistency between the prediction results and the actual force mechanism of the tire, reduces the risk of non-physical output, and thus provides faster, more accurate, and engineering-application-value six-component force data support for tire structure design optimization, material formulation screening, vehicle handling stability analysis, virtual prototype development, and vehicle active safety control. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the system structure of the present invention.

[0056] Figure 2 This is a flowchart of the method of the present invention.

[0057] Figure 3 This is a schematic diagram of the three-branch coding and feature fusion network structure of the present invention.

[0058] Figure 4 This is a schematic diagram showing the relationship between the intermediate physical characterization quantity of the grounding imprint and the generation of the six components of force in this invention.

[0059] Figure 5 This is a flowchart of the hierarchical training and loss function construction process of this invention.

[0060] Figure 6 This is a schematic diagram of the predicted output of the six components of the target tire in this invention.

[0061] Figure 7 This is a schematic diagram of the sample distribution under different working conditions according to the present invention.

[0062] Figure 8 This is a comparison chart of the test errors of the various methods of this invention.

[0063] Figure 9 This is a fitting curve of lateral force and self-aligning torque under pure lateral deviation conditions according to the present invention.

[0064] Figure 10 This is a comparison diagram of the combined prediction of longitudinal force and lateral force under the combined working conditions of this invention. Detailed Implementation

[0065] This invention relates to a neural network-based method for predicting the six-component forces of tires. This method addresses the problem of rapid prediction of the six-component forces of radial tires under different structural, material, and operational parameters. It proposes a technical route consisting of: data acquisition—data preprocessing and operational condition layering—three-branch encoding and feature fusion—prediction of intermediate physical representations of the ground contact imprint—initial value generation of the basic mechanical model—residual correction—layered training—target tire six-component force output. Unlike traditional direct black-box regression, this invention uses intermediate physical representations of the ground contact imprint as a bridging layer between input parameters and the six-component force output, establishing a connection between the network learning process and the tire ground contact mechanics mechanism. Simultaneously, it introduces a layered training strategy for pure sideslip, pure longitudinal slip, and combined operating conditions, enabling the model to maintain higher prediction accuracy and better physical plausibility under multiple operating conditions.

[0066] like Figure 1As shown, the system of the present invention mainly includes: a data acquisition module, a data preprocessing and hierarchical construction module, a prediction network, a model training module, and a six-component force prediction output module; wherein, the prediction network further includes a structure encoding branch, a material encoding branch, a working condition encoding branch, a feature fusion module, an intermediate physical quantity prediction module, a basic mechanical model module, and a residual correction module. Figure 2 The overall flow of the method of the present invention is shown. Figure 3 The three-branch coding and feature fusion network structure is shown. Figure 4 The relationship between the intermediate physical quantities of the grounding imprint and the generation of the six components of force is shown. Figure 5 This illustrates the hierarchical training and loss function construction process. Figure 6 The process of predicting the output of the six components of the target tire force is shown. Figure 7 The distribution of samples under different working conditions is shown. Figure 8 The differences in testing error between the method of the present invention and the comparative method are shown. Figure 9 The fitting results are shown under the pure lateral deviation condition. Figure 10 The combined prediction effect under combined operating conditions is shown.

[0067] To enable those skilled in the art to implement this invention without inventive effort, the following will provide a structured disclosure of the terminology, system structure, specific technical routes for each step, model structure, training methods, parameter selection, and dataset usage.

[0068] I. Terminology Explanation

[0069] To enable those skilled in the art to accurately understand the technical solution of the present invention, some of the terms used herein will be explained first.

[0070] The six components of tire force refer to the forces and moments generated in three directions when the tire is in contact with the road surface and is under load, rolling, steering, or braking conditions, including longitudinal force. Lateral force Normal force Overturning torque Rolling resistance torque and the restoring torque .

[0071] Structural parameter set : Refers to the set of parameters used to describe the tire's geometry and skeleton layout, including but not limited to tread width, section height-to-width ratio, rim diameter, tire carcass cord angle, belt layer angle, belt layer width, sidewall thickness, and tread pattern type code.

[0072] Material parameter set This refers to a set of parameters used to describe the mechanical and wear characteristics of the various components of a tire, including but not limited to the equivalent elastic modulus of the tread compound, the equivalent elastic modulus of the carcass compound, the equivalent elastic modulus of the belt layer compound, the equivalent elastic modulus of the sidewall compound, the wear factor, the equivalent elastic modulus of the cord layer, and the equivalent density of the material.

[0073] Operating Parameter Set : Refers to the set of parameters used to describe the external operating state and contact conditions of a tire, including but not limited to inflation pressure, vertical load, slip angle, longitudinal slip ratio, camber angle, driving speed, road adhesion coefficient, and road roughness factor.

[0074] Structural representation vector : refers to the set of structural parameters The vectors generated through normalization, encoding, scaling reconstruction, and combination are used to represent comprehensive information about the tire skeleton and geometric configuration.

[0075] Material characterization vector : refers to the set of material parameters The vector formed after normalization, hierarchical combination and feature compression is used to characterize the equivalent stiffness, energy dissipation and mass distribution characteristics of materials in different parts of the tire.

[0076] Operating condition coupling state vector : refers to the set of operating parameters The vector formed by the tire and its coupling derivatives is used to characterize the stress state of the tire under specific external conditions and its nonlinear coupling relationship.

[0077] Grounding Imprint Intermediate Physical Characterization Set : Refers to the set of intermediate quantities used to connect the physical mechanism between the input characteristics and the output six components of force, including the length of the ground imprint, the width of the ground imprint, the longitudinal offset of the normal pressure center, the lateral offset of the normal pressure center, the equivalent offset distance of the lateral force action line, and the average tangential shear stress of the grounding area, etc.

[0078] Basic Six-Component Force Initial Value Set : refers to the set of preliminary predicted values ​​of the six components of force calculated by the basic mechanical model based on tire parameters and intermediate physical characteristics.

[0079] Residual correction set : Refers to the set of correction quantities learned by the neural network based on the difference between the initial value of the basic six-component force and the actual six-component force, which is used to compensate for nonlinear errors that are difficult to cover by simplified mechanical models.

[0080] II. System Structure

[0081] like Figure 1As shown, the tire six-component force prediction system based on neural networks of the present invention includes at least a data acquisition module, a data preprocessing and hierarchical construction module, a three-branch encoding module, a feature fusion module, an intermediate physical quantity prediction module, a basic mechanical model module, a residual correction module, a model training module, and a six-component force prediction output module.

[0082] The data acquisition module is used to collect raw sample data from multiple tire samples under different working conditions. The raw sample data can originate from six-component bench tests, high-confidence finite element simulation data calibrated with the tests, or mixed test and simulation data. Preferably, test data serves as the primary data source, and simulation data serves as an enhanced data source for boundary conditions.

[0083] The data preprocessing and hierarchical construction module is used to clean, normalize, encode, extract steady-state data, remove anomalies, and classify operating conditions into different types. This module outputs a structural representation vector. Material characterization vector Operating condition coupling state vector And the training dataset stratified by working condition category.

[0084] The three-branch coding module includes a structural coding branch, a material coding branch, and a working condition coding branch. The structural coding branch extracts tire skeleton and geometric configuration features, the material coding branch extracts material nonlinearity and layered stiffness features, and the working condition coding branch extracts working condition variables such as load, slip, yaw, and camber, as well as their coupling effects.

[0085] The feature fusion module fuses the implicit features from the three branches to form shared features. Shared features It can be implemented using fully connected layers, residual layers, gating units, or attention fusion structures. This implementation preferably adopts a fusion approach of fully connected layers + residual connections + gating weighting to balance stability and expressive power.

[0086] The intermediate physical quantity prediction module is used to predict based on shared features Output ground imprint intermediate physical characterization set Instead of directly outputting the six components of force, this module first predicts physical quantities that reflect the tire contact state and the position of the line of action, thereby improving the interpretability and physical consistency of the model.

[0087] The basic mechanical model module is based on tire parameters and intermediate physical quantity sets. Calculate the initial value set of the six fundamental forces. This module can be implemented using simplified empirical formulas, linearized ground stress models, parameter-reduced tire mechanics models, or semi-empirical models.

[0088] The residual correction module is based on shared features and the initial value set of the basic six forces Output residual correction set It is used to compensate for complex material nonlinearity, coupled deformation, and extreme condition deviations that are difficult to describe by basic mechanical models.

[0089] The model training module trains the entire network according to the predetermined loss function and hierarchical working condition strategy, and completes parameter updates, early stopping control, regularization and model selection.

[0090] The six-component force prediction output module is used to output the six-component force prediction results of the target tire under the input conditions of the target tire.

[0091] III. Overall Technical Route for Implementing the Method of the Invention

[0092] like Figure 2 As shown, the method of this invention is implemented according to steps S1 to S5. The overall technical approach is not simply to input tire parameters into a black-box neural network and directly output the six-component force, but rather to employ a multi-source parameter representation—intermediate physical quantity bridging—basic mechanical initial value generation—residual compensation—working condition-layered training approach. This approach has two core characteristics:

[0093] Firstly, by connecting the input and output sides through intermediate physical characterization quantities of grounding imprints, the black box nature is reduced;

[0094] Secondly, by training under tiered working conditions, the model can master the tire mechanics laws under pure sideslip, pure longitudinal slip, and combined working conditions, and then achieve collaborative output within a unified network framework.

[0095] The following combination Figures 3 to 6 The implementation method of each step will be explained in further detail.

[0096] (I) Step S1: Sample data acquisition and raw data construction

[0097] Step S1 mainly aims to provide a real, stable, and learnable sample base for subsequent modeling.

[0098] 1. Selection of sample tires

[0099] Preferably, no fewer than 20 tire samples are selected, covering different specifications, aspect ratios, belt ply structures, tread patterns, and compound systems. For example, these may include passenger car tires, SUV tires, and light commercial vehicle tires. To enable the model to generalize across tire types, the tire samples should have sufficient dispersion in structural and material parameters.

[0100] 2. Input parameter acquisition

[0101] For each tire sample, its structural parameter set is collected. Material parameter set and operating condition parameter set .

[0102] Structural parameter set This may include: tread width Units are mm; cross-sectional height-to-width ratio , is a dimensionless proportional value; rim diameter The unit is inch or mm; tire cord angle The unit is °; belt layer angle Unit is °; belt layer width The unit is mm; tire sidewall thickness Unit is mm; tread pattern category code , where is a discrete variable.

[0103] Material parameter set This may include: the equivalent elastic modulus of the tread compound. The unit is MPa; the equivalent elastic modulus of the tire carcass rubber. The unit is MPa; the equivalent elastic modulus of the belt layer adhesive. The unit is MPa; the equivalent elastic modulus of the sidewall rubber. Unit is MPa; loss factor , is a dimensionless variable; equivalent elastic modulus of the cord layer The unit is MPa; the material's equivalent density. The unit is g / cm³ or kg / m³.

[0104] Operating Parameter Set This may include: inflation pressure The unit is kPa; vertical load The unit is N; sideslip angle The unit is °; longitudinal slip ratio , is a dimensionless variable; outward inclination angle The unit is °; driving speed The unit is km / h or m / s; road surface adhesion coefficient Dimensionless; road surface roughness factor , dimensionless or defined according to calibration.

[0105] 3. Obtaining Six-Point Strength Tags

[0106] For each working condition sample, the measured label set was collected using a six-component force test bench. ,include , , , , , Preferably, each sample point is labeled with the average value of the steady-state period before being entered into the model, in order to reduce the impact of instantaneous noise.

[0107] 4. Sample working condition design

[0108] To improve the model's adaptability to multiple operating conditions, the following operating condition coverage strategy is recommended:

[0109] Pure lateral slip condition: Approaching 0 It changes gradually within a preset range;

[0110] Pure longitudinal slip condition: Approaching 0 It changes gradually within a preset range;

[0111] Combined working conditions: and Simultaneously, non-zero values ​​are taken, and variables such as load, air pressure, and speed are combined.

[0112] For example, the preferred sampling interval is as follows:

[0113] It varies within the range of 180 kPa to 320 kPa;

[0114] It varies within the range of 2000N to 7000N;

[0115] It varies within the range of -12° to 12°;

[0116] It varies within the range of -0.25 to 0.25;

[0117] It varies within the range of -4° to 4°;

[0118] It varies within the range of 20km / h to 120km / h.

[0119] The above method forms the original sample database, laying the foundation for the hierarchical modeling in the subsequent step S2.

[0120] (ii) Step S2: Preprocessing, representation vector construction and working condition stratification

[0121] Step S2 differs from the conventional machine learning approach of simple normalization and training set partitioning. In this step, the present invention completes three key tasks: First, reconstructing the original parameters into a dedicated representation vector suitable for the tire object; second, introducing working condition coupling derivatives to enhance the model's perception of nonlinear boundaries; and third, stratifying the samples into pure sideslip, pure longitudinal slip, and combined working conditions, thereby making the training process guided by tire working condition mechanisms.

[0122] 1. Data cleaning and anomaly handling

[0123] The original sample data is checked item by item. For missing values, the following rules are applied: if the missing percentage is less than a preset threshold, such as 3%, then continuous variables are imputed using interpolation between adjacent operating conditions, the mean of similar samples, or regression. If the missing percentage exceeds the preset threshold, the sample is discarded or retested. For outliers, a combination of statistical thresholds and physical boundaries is used. For example, when the six-component force label exceeds the equipment's range, shows obvious sign errors, or deviates from adjacent operating points by more than three standard deviations, it is marked as an outlier and reviewed. Samples confirmed as outliers after review are removed or corrected.

[0124] 2. Steady-state extraction of time-series signals

[0125] When the raw output of the test bench is time-series data, it is necessary to extract the steady-state segment for each operating condition. This can be done by first performing a low-pass filter on the raw time-series data, then calculating the standard deviation of the sliding window. When the standard deviation of several consecutive windows is below a threshold, the segment is considered to have entered the steady-state interval. Then, the average value within the steady-state interval is taken as the six-component force label for that operating condition.

[0126] 3. Standardization of continuous variables

[0127] For continuous variables, standardization is preferred. Standardize the fractions to satisfy:

[0128] ;

[0129] in, For original variables, For standardized variables, This represents the mean of the variable in the training set. This represents the standard deviation of the variable in the training set.

[0130] For categorical variables such as tire tread pattern category codes One-hot encoding or embedding encoding is preferred for processing. In this embodiment, if the number of pattern categories is no more than 8, one-hot encoding can be used; if there are more pattern categories, trainable embedding encoding is preferred.

[0131] 4. Structural representation vector Construction

[0132] This invention does not directly use the structural parameter set Instead of inputting the data into the network as is, a structural representation vector is constructed. Preferably, This includes the original geometric parameters, normalized geometric parameters, and several derived variables reflecting the skeleton layout. For example, the following derived variables can be constructed:

[0133] relative width ratio of belt layer ,satisfy:

[0134] ;

[0135] in, The relative width ratio of the belt layer. For the width of the belt layer, This refers to the tread width.

[0136] It can also construct cross-sectional shape factors :

[0137] ;

[0138] in, For cross-sectional shape factor, The aspect ratio of the cross section. This refers to the tread width. The diameter is the rim diameter.

[0139] Through the above construction, the structural input not only retains information in a single dimension but also reflects the relative relationships between structures. Ultimately, the structural representation vector... It is formed by splicing together the original structural parameters and derived structural factors.

[0140] 5. Material Characterization Vector Construction

[0141] Material characterization vector It is mainly used to reflect the equivalent stiffness and hysteresis characteristics of different parts of the tire. Preferably, the material parameters are grouped according to the tread layer, carcass layer, belt layer, and sidewall layer, and then the following derived parameters are extracted. For example, an overall equivalent stiffness factor can be constructed. :

[0142] ;

[0143] in, The overall equivalent stiffness factor. , , , These are the weighting coefficients for the tread layer, carcass layer, belt layer, and sidewall layer, respectively. , , , These are the equivalent elastic moduli for the corresponding parts.

[0144] It can also construct material hysteresis characteristic factors. :

[0145] ;

[0146] in, For material hysteresis characteristic factors, The overall equivalent stiffness factor. This is the material loss factor.

[0147] By using the above method, material parameters are transformed from simple parameter tables into comprehensive characterization vectors with engineering significance, which helps in subsequent learning about the influence of material nonlinearity on rolling resistance torque, self-correcting torque, and normal load distribution.

[0148] 6. Operating condition coupling state vector Construction

[0149] This is one of the key differences between this invention and conventional input methods. Operating condition coupling state vector. In addition to the original operating condition parameters, it also includes several derived variables that reflect the coupling relationship between operating conditions. For example:

[0150] Sliding coupling term :

[0151] ;

[0152] in, For the sliding coupling term, Side slip angle, It represents the longitudinal slip ratio.

[0153] Normalized load ratio :

[0154] ;

[0155] in, For normalized load ratio, For the current vertical load, This is the reference load.

[0156] Adhesion Demand Index :

[0157] ;

[0158] in, For the attached demand index, Side slip angle, For longitudinal slip ratio, The road surface adhesion coefficient, To prevent positive decimals with a denominator of zero.

[0159] Through these derived variables, the network can not only perceive individual operating parameters, but also whether the tire is close to the adhesion limit or is in a strongly coupled operating condition, thereby improving the prediction capability at complex boundaries.

[0160] 7. Working Condition Stratification Rules

[0161] In this invention, the samples are not randomly mixed for training, but are first divided into three subsets of operating conditions according to the mechanism. The following rules are preferably adopted:

[0162] when and At that time, it was determined to be a pure sideslip condition;

[0163] when and At that time, it was determined to be a pure longitudinal sliding condition;

[0164] when and When this condition is met, it is determined to be a composite working condition.

[0165] in, To preset the sliding threshold, This is a preset lateral deviation threshold. Preferably, Take a value between 0.01 and 0.02. Take a value between 0.5° and 1.0°.

[0166] 8. Dataset partitioning

[0167] The various operating condition samples are further divided into training, validation, and test sets. Ideally, stratified sampling should be performed simultaneously along both the tire sample dimension and the operating condition dimension to avoid information leakage between training and testing for the same tire and the same operating condition point. For example, a 70%:15%:15% ratio could be used.

[0168] Through the above step S2, three types of input vectors and hierarchical training datasets suitable for the network structure of this invention can be obtained.

[0169] (III) Step S3: Neural network structure, intermediate physical quantity prediction and six-component force generation mechanism

[0170] Step S3 does not construct a typical multilayer perceptron, but rather a hierarchical network architecture consisting of structural branches, material branches, operating condition branches, intermediate physical quantities, fundamental mechanical models, residual correction, and six-component force outputs. This step is crucial to the interpretability, accuracy, and feasibility of this invention, and will be described in detail below.

[0171] 1. Three-branch coding structure

[0172] like Figure 3 As shown, the tire six-component force prediction neural network includes three input encoding branches. The structure encoding branch receives the structure representation vector. Preferably, it consists of three fully connected layers, with each layer having, for example, 64, 64, or 32 nodes. The ReLU activation function is preferred. The material encoding branch receives the material characterization vector. It can be configured as a three-layer structure with 64, 32, or 32 layers. The operating condition coding branch receives the operating condition coupling state vector. Since operating condition variables have the most direct impact on the output, a higher capacity is preferred, such as a three-layer structure with 128, 64, and 32 layers. Each branch can have batch normalization layers and Dropout layers added between layers. A Dropout ratio of 0.1 to 0.3 is preferred to prevent overfitting.

[0173] 2. Feature Fusion Module

[0174] The implicit features of the three branch outputs are denoted as follows: , and Then concatenate them into joint features:

[0175] ;

[0176] in, For the combined features after splicing, These are implicit structural features. As a latent feature of the material, These are implicit characteristics of the operating conditions.

[0177] Subsequently, joint features Shared features are output via the fusion layer The preferred fusion layer consists of two fully connected layers and one residual connection to maintain stability during deep training.

[0178] 3. Intermediate physical quantity prediction module

[0179] Traditional models often directly map from the input to... , , , , , The black box characteristics are obvious. This invention first predicts the intermediate physical characterization set of grounding imprints. In this embodiment, At least including: grounding imprint length ; width of grounding imprint Longitudinal offset of the normal pressure center Lateral offset of the normal pressure center ; Equivalent offset distance of the line of action of lateral force ; average tangential shear stress in the receiving area .in, and Reflecting the geometric shape of the grounding imprint, and Reflects the location of the pressure center. Reflects the offset of the line of action of the lateral force. These quantities reflect the average tangential force level in the contact area. They can all serve as intermediate mechanism quantities for the generation of the six component forces.

[0180] The intermediate physical quantity prediction module can be implemented through a single multi-output full-connection header, or it can be split into geometric sub-headers, pressure center sub-headers, and shear stress sub-headers. This implementation prefers splitting the outputs to improve the learning stability of various physical quantities.

[0181] 4. Basic Mechanical Model Module

[0182] Obtaining the intermediate physical quantity set Then, the basic mechanics model module calculates the initial value set of the six component forces of the foundation. This module does not require the use of complex finite element models; instead, simplified models that are feasible in engineering can be used.

[0183] Initial values ​​of longitudinal force foundation It can be estimated using the following formula:

[0184] ;

[0185] in, For the initial values ​​of the longitudinal force foundation, For longitudinal calibration coefficients, The average tangential shear stress in the receiving area. The length of the grounding imprint. For the width of the grounding imprint, This is the sign function for the longitudinal slip ratio.

[0186] Lateral force foundation initial values It can be estimated using the following formula:

[0187] ;

[0188] in, The initial values ​​for the lateral force foundation are as follows: These are the lateral calibration coefficients. Side slip angle, For vertical loads, This is a correction function that varies with the road surface adhesion coefficient and the limit state.

[0189] Initial values ​​of normal force It can be directly determined from the relationship between load and ground pressure distribution, with the following preferred value:

[0190] ;

[0191] in, The initial value for the normal force is given. This refers to the vertical load in the input operating condition.

[0192] Initial value of the basic overturning torque It can be estimated using the following formula:

[0193] ;

[0194] in, This is the initial value of the basic overturning torque. The initial value for the normal force is given. This represents the lateral offset of the normal pressure center.

[0195] Initial values ​​of rolling resistance torque It can be estimated using the following formula:

[0196] ;

[0197] in, This is the initial value of the rolling resistance torque. This is the rolling resistance calibration coefficient. For material hysteresis characteristic factors, The initial value for the normal force is given. This is the length of the grounding imprint.

[0198] Initial value of the restoring torque foundation It can be estimated using the following formula:

[0199] ;

[0200] in, The initial value of the restoring torque is the foundation value. The initial values ​​for the lateral force foundation are as follows: This is the equivalent offset distance of the line of action of the lateral force.

[0201] The significance of the aforementioned basic mechanics model module lies in the fact that it first uses intermediate physical quantities to express the main mechanical relationships of the six component forces, and then uses neural networks to compensate for the remaining complex nonlinearities, thereby reducing the model's degrees of freedom and enhancing its interpretability.

[0202] 5. Residual Correction Module

[0203] The residual correction module receives shared features. and the initial value set of the basic six forces Output residual correction set ,Right now:

[0204] ;

[0205] in, , , These are the residual corrections for the triaxial forces, respectively. , , These are the residual corrections for the three-dimensional torques, respectively.

[0206] The residual correction module preferably consists of a shared layer and six independent output heads. The shared layer is used to extract high-level features that contribute to all outputs, and the independent output heads fit the specific bias of each output.

[0207] 6. Six-component force generation module

[0208] The six-component force generation module generates the final prediction result using the following formula:

[0209] ;

[0210] in, This is the final vector of the six-component tire force prediction results. The initial value vector of the basic six-component force. This is the residual correction vector.

[0211] For the restoring torque, a coupling expression can be further used:

[0212] ;

[0213] in, This is the predicted value of the aligning torque. This is the predicted value of the lateral force. The equivalent offset distance of the line of action of the lateral force. This is the correction amount for the remaining residuals.

[0214] This is particularly important because the restoring torque is not obtained through complete free regression, but rather establishes a clear relationship with lateral forces and the offset of the line of action, thereby significantly reducing the risk of non-physical output.

[0215] (iv) Step S4: Layered training strategy, loss function and parameter selection

[0216] Step S4 does not involve training the model using a typical training set. Instead, it employs a hierarchical training strategy with multiple constraints, designed around the characteristics of tire mechanics. This will be explained in detail below.

[0217] 1. Overall Idea of ​​Differentiated Training

[0218] Because the formation mechanisms of tire lateral force and self-centering torque are prominent under pure sideslip conditions, longitudinal force and rolling resistance torque are more sensitive under pure longitudinal slip conditions, and multivariate strong coupling is observed under combined conditions, it is easy to cause some conditions to dominate and some outputs to be weakened if all samples are mixed and trained uniformly. This invention uses a hierarchical training dataset and sets weights for each subset of conditions in the loss function to ensure that the model can simultaneously grasp the laws of the three types of conditions.

[0219] 2. Training Batch Organization

[0220] In each training epoch, it is preferable to draw small batches of samples from the pure lateral slip condition sample subset, the pure longitudinal slip condition sample subset, and the combined condition sample subset in a fixed ratio. For example, the sampling can be balanced at a ratio of 1:1:1; or it can be adjusted to 2:2:3 according to engineering needs, so that the combined condition receives more training resources.

[0221] 3. Design of the total loss function

[0222] Total loss function Including six component force error terms Layered working condition weighted items and the consistency constraint term for the restoring torque ,satisfy: ;in, For the total loss function, This is a six-component force error term. For weighted loss terms of different operating condition subsets, This is a consistency constraint term for the normalizing torque. The weighting coefficients for the operating conditions. This is the coefficient of the consistency constraint term for the aligning torque.

[0223] 4. Six-component force error term

[0224] The weighted mean square error of the six force components is preferred. It is expressed as:

[0225] ;

[0226] in, This is a six-component force error term; , , , , , These are the error weights for longitudinal force, lateral force, normal force, overturning torque, rolling resistance torque, and self-aligning torque, respectively. , , , , , These represent the errors between the predicted and measured values, respectively.

[0227] Preferably, different weights can be assigned based on the dimensions and engineering importance of the six output quantities. For example, those with a significant impact on handling and stability performance... and Give it higher weight.

[0228] 5. Layered working condition weighting items

[0229] To prevent the model from being biased towards a particular work condition due to a dominant number of samples, this invention weights the three work conditions separately. This can be expressed as: ;in, Weighting terms for operating conditions, For pure sideslip conditions, For losses under pure longitudinal sliding conditions, For losses under combined working conditions, , , These are the weighting coefficients for the three types of working conditions.

[0230] In a preferred embodiment, set , , This allows for stronger constraints on complex operating conditions during training.

[0231] 6. Consistency constraint for return torque

[0232] To enhance the physical consistency of the aligning torque, this invention introduces a aligning torque consistency constraint. This constraint is not simply... Instead of fitting labels, we make them consistent with lateral forces and line-of-action offsets. This can be expressed as:

[0233] ;

[0234] in, This is a consistency constraint term for the normalizing torque. This is the predicted value of the aligning torque. This is the predicted value of the lateral force. This is the equivalent offset distance of the line of action of the lateral force.

[0235] This means that even when there are local fluctuations in the data labels, the model will tend to output a aligning torque result that is more consistent with the tire mechanics mechanism.

[0236] 7. Optimize the algorithm and training parameters

[0237] The Adam optimizer is preferred during training. The initial learning rate can be set to... The batch size can be 64 or 128, and the number of training epochs can be set to 200 to 500. If the validation set loss no longer decreases within 20 consecutive epochs, the early stopping mechanism is enabled.

[0238] To avoid overfitting, the following measures can be used in conjunction:

[0239] Firstly, Dropout, with a ratio of 0.1 to 0.3;

[0240] Secondly, the coefficient of the L2 regularization term can be taken as... to ;

[0241] Third, a learning rate decay strategy, such as multiplying the learning rate by 0.5 if the validation loss does not improve after 10 consecutive epochs.

[0242] 8. Training Process

[0243] During training, samples within a batch are input into the three-branch encoding module, and shared features are output after feature fusion. Then, the intermediate physical quantity prediction module outputs the result. And generated by the basic mechanics model module. Then output by the residual correction module. Finally obtained The model is based on the total loss function. Backpropagation updates parameters.

[0244] 9. Model Selection Criteria

[0245] After training, in addition to examining the average error on the validation set, we also prioritize examining the mean absolute error, root mean square error, and coefficient of determination under three different operating conditions. Only when all the main outputs under the three operating conditions reach the preset thresholds is the current model determined as the final model.

[0246] (V) Step S5: Target Tire Prediction and Engineering Deployment

[0247] Step S5 determines whether the present invention can be truly implemented in an engineering scenario.

[0248] 1. Target tire input

[0249] For the target tire to be predicted, first collect its target structural parameter set, target material parameter set, and target working condition parameter set. The collection method should be consistent with step S1 to ensure a uniform input format.

[0250] 2. Unified preprocessing

[0251] The target tire input is processed using the same standardized parameters and encoding method as in the training phase, yielding the target structure representation vector, target material representation vector, and target operating condition coupled state vector. It is important to note that the mean and standard deviation must not be recalculated during the inference phase; instead, the values ​​saved during the training phase should be used. and .

[0252] 3. Model Reasoning

[0253] The target input is fed into the trained tire six-component force prediction model, which then undergoes three-branch encoding, feature fusion, intermediate physical quantity prediction, basic mechanical initial value generation, and residual correction to obtain the target tire six-component force prediction result:

[0254] ;

[0255] in, This is the predicted value of the longitudinal force. This is the predicted value of the lateral force. This is the predicted value of the normal force. This is the predicted value of the overturning torque. This is the predicted value of the rolling resistance torque. This is the predicted value of the aligning torque.

[0256] 4. Result Output Format

[0257] The output results can be used in the following scenarios:

[0258] Firstly, offline parameter analysis during the tire design phase;

[0259] Secondly, the tire six-component force input for the multibody dynamics simulation of the whole vehicle;

[0260] Third, the online estimation module in the vehicle real-time control system;

[0261] Fourth, tire status updates in digital twin or virtual prototype systems.

[0262] 5. Reasoning efficiency

[0263] Because this invention involves only a small number of matrix multiplications and simple physical calculations during the inference phase, its inference speed is much faster than high-precision finite element calculations. For a single operating point, prediction can be completed within milliseconds to seconds, making it suitable for batch scanning and real-time applications.

[0264] IV. Specific Application Examples

[0265] (I) Experimental conditions and experimental platform

[0266] This application example was completed on a tire six-component force indoor test bench. The test bench is equipped with tire vertical loading, inflation pressure control, rolling speed control, sideslip angle adjustment, longitudinal slip ratio adjustment, and simultaneous acquisition of six-component forces. The six-component force sampling frequency was set to 200Hz, the steady-state sampling time window was set to 5s, and the average value of the steady-state interval was taken as the label value for each working point.

[0267] 1. Tire Sample

[0268] Twenty-four radial tires were selected as sample tires, including:

[0269] Sample group A: 16 passenger car tires;

[0270] Sample group B: 8 SUV tires.

[0271] Tires come in various sizes, structures, and material systems. Preferably, the tread width... The cross-sectional height-to-width ratio ranges from 195mm to 255mm. The rim diameter is 40 to 65. For tires ranging from 16 inches to 19 inches, the tire cord angle... The angle of the belt layer is 78° to 90°. The angle ranges from 18° to 28°. Among the material parameters, the equivalent elastic modulus of the tread rubber is... The equivalent elastic modulus of the sidewall rubber ranges from 6.5 MPa to 11.8 MPa. The loss factor ranges from 3.2 MPa to 6.4 MPa. The values ​​range from 0.072 to 0.146.

[0272] 2. Working condition design

[0273] To cover the main usage scenarios where tires exert six-tenths of their force, the following operating condition ranges are defined:

[0274] inflation pressure :200kPa, 220kPa, 240kPa, 260kPa;

[0275] Vertical load : 2500N, 3500N, 4500N, 5500N;

[0276] Side slip angle -10°, -8°, -6°, -4°, -2°, 0°, 2°, 4°, 6°, 8°, 10°;

[0277] Longitudinal slip ratio : -0.20, -0.15, -0.10, -0.05, 0, 0.05, 0.10, 0.15, 0.20;

[0278] outward tilt -2°, 0°, 2°;

[0279] driving speed : 40km / h, 60km / h, 80km / h.

[0280] The operating conditions are divided into three categories: pure sideslip, pure longitudinal slip, and combined operating conditions. Among them, the pure sideslip condition is... Or close to 0; pure longitudinal slip condition is Or close to 0; composite operating conditions are and At the same time, it is not zero.

[0281] 3. Dataset Composition

[0282] A total of 48,960 valid sample points were obtained. Among them:

[0283] 16,320 samples of pure sideslip conditions;

[0284] 12,240 samples of pure longitudinal sliding conditions;

[0285] 20,400 sets of composite working condition samples.

[0286] The samples were stratified using both tire sample dimensions and working condition dimensions, and divided into training set, validation set, and test set, with a ratio of 70%:15%:15%.

[0287] (ii) Proportional Setting

[0288] To demonstrate the technical effects of this invention, the following comparative examples are provided.

[0289] Comparative Example 1: Traditional End-to-End Multilayer Feedforward Neural Network Method

[0290] The structural parameters, material parameters, and working condition parameters are directly concatenated and input into a conventional multilayer feedforward neural network. No intermediate physical characterization prediction module is set up, no basic mechanical model module is set up, and no layered working condition training strategy is adopted. The six component forces are directly output end-to-end.

[0291] Comparative Example 2: Direct Prediction Method Based on Fundamental Mechanical Models

[0292] A simplified tire mechanics model is adopted, and the tire six-component force is directly calculated based on the working condition parameters and some tire parameters, without using neural network residual compensation.

[0293] Comparative Example 3: Initial physical values ​​+ residual network, but without intermediate physical quantity bridging and layered training.

[0294] The initial values ​​of the six components of force are first calculated by the basic mechanical model, and then corrected by the residual network. However, the input does not construct dedicated structural characterization vectors, material characterization vectors, and working condition coupling state vectors, and does not distinguish between pure lateral slip, pure longitudinal slip, and composite working conditions during the training process.

[0295] Example 1: The Method of the Invention

[0296] It adopts a three-branch coding structure, bridging of intermediate physical representation quantities of ground imprint, initial value of six component forces of the foundation plus residual correction, layered working condition training, and consistency constraint of return torque.

[0297] III. Model Training Settings

[0298] 1. Network settings of the method of the present invention

[0299] The structure coding branch is a 64-64-32 three-layer fully connected structure;

[0300] The material coding branch has a 64-32-32 three-layer fully connected structure;

[0301] The operating condition coding branch has a 128-64-32 three-layer fully connected structure;

[0302] The feature fusion module is a 128-128 two-layer fully connected structure with a residual connection;

[0303] The intermediate physical quantity prediction module outputs the ground imprint length. grounding imprint width Longitudinal offset of normal pressure center Lateral offset of the normal pressure center Equivalent offset distance of the line of action of lateral force Average tangential shear stress in the junction area Residual correction module output , , , , and .

[0304] 2. Training parameters

[0305] The optimizer uses Adam;

[0306] The initial learning rate is ;

[0307] Batch size is 128;

[0308] The maximum number of training rounds is 300;

[0309] The dropout ratio is 0.2;

[0310] The L2 regularization coefficient is ;

[0311] Early stopping is triggered when the validation set loss does not decrease for 20 consecutive epochs.

[0312] 3. Loss Function

[0313] Total loss function for:

[0314] ;

[0315] in, This is a six-component force error term. For weighted terms of stratified operating conditions, This is a consistency constraint term for the normalizing torque. The weighting factor for the operating conditions. This is the consistency constraint coefficient for the restoring torque. In this embodiment, , .

[0316] The evaluation indicators used are mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination. .

[0317] The mean absolute error (MAE) is:

[0318] ;

[0319] in, For the sample size, For the first Measured values ​​of a sample For the first The predicted value for each sample.

[0320] The root mean square error (RMSE) is:

[0321] ;

[0322] The meanings of each parameter are the same as above.

[0323] Coefficient of determination for:

[0324] ;

[0325] in, This is the average of the measured values.

[0326] (iv) Experimental Results and Data Analysis

[0327] 1. Overall test set error comparison

[0328] Table 1 presents the main evaluation results of Example 1 and each comparative example on the overall test set.

[0329] Table 1 Comparison of prediction errors for the six components of the overall test set.

[0330]

[0331] As shown in Table 1, the RMSE of Embodiment 1 of the present invention across all six output quantities is significantly lower than that of the comparative embodiments, with an average... The value reached 0.978. Compared to Comparative Example 1, Example 1 showed better performance in longitudinal force. Lateral force , backing torque The RMSE decreased by approximately 40.5%, 40.9%, and 44.4%, respectively; compared to Comparative Example 3, the rolling resistance torque... and the restoring torque The accuracy was reduced by approximately 28.2% and 32.4% respectively. This indicates that even with the existing physical initial value + residual network, this invention can further improve prediction accuracy by introducing intermediate physical representations of grounding imprints and hierarchical working condition training.

[0332] 2. Test results under different working conditions

[0333] Table 2 shows the main error results of Example 1 under three types of working conditions.

[0334] Table 2 Error results of Embodiment 1 of the present invention under different working conditions

[0335]

[0336] As shown in Table 2, the present invention exhibits high prediction accuracy under both pure sideslip and pure longitudinal slip conditions. While the error increases under more complex combined conditions, it still maintains a high coefficient of determination and a low RMSE. This indicates that the present invention, through layered training based on different conditions, enables the model to possess strong learning capabilities for all three types of conditions, avoiding the significant error amplification issue seen in traditional hybrid training methods under combined conditions.

[0337] 3. Technical effect analysis under pure lateral deflection condition

[0338] Select a 225 / 55R17 passenger car tire, and inflate it to the following pressure: kPa, vertical load N, Inclination angle Longitudinal slip ratio Under the conditions, test the sideslip angle Lateral force and restoring torque when varying from -10° to 10°.

[0339] Table 3 Comparison of lateral force and restoring moment under pure sideslip condition

[0340]

[0341] From Table 3 and in combination Figure 9 It can be seen that the lateral force curve predicted by this invention is highly consistent with the measured curve, maintaining a good fit even when the absolute value of the sideslip angle is large. Simultaneously, the predicted and measured values ​​of the aligning torque show good agreement in terms of variation trend and peak position, without exhibiting the excessive oscillation phenomenon common in end-to-end black-box models at high sideslip angles. This is related to the offsetting distance between the aligning torque and the lateral force action line in this invention. The establishment of a coupling relationship is directly related, indicating that the present invention can effectively improve the physical consistency of the return torque output.

[0342] 4. Technical effect analysis under pure longitudinal sliding conditions

[0343] Select the same tire, at the same inflation pressure kPa, vertical load N, Side slip angle outward tilt angle Under these conditions, test the longitudinal slip ratio Longitudinal force and rolling resistance torque as it varies from -0.20 to 0.20.

[0344] Table 4 Comparison of longitudinal force and rolling resistance torque under pure longitudinal sliding condition

[0345]

[0346] As shown in Table 4, the present invention can accurately characterize the change law of longitudinal force from the linear region to the saturation region as the slip ratio increases under pure longitudinal slip conditions. At the same time, since the material hysteresis characteristic factor and the ground imprint length are used to construct the initial value of the rolling resistance torque, the prediction of the rolling resistance torque is more stable and has good consistency with the changing trend of the positive and negative slip ratio ranges.

[0347] 5. Analysis of technical effects under combined working conditions

[0348] Select the same tire, at the same inflation pressure kPa, vertical load N, Inclination angle Nine composite working conditions were set up to compare the combined prediction capabilities of longitudinal force and lateral force.

[0349] Table 5. Joint prediction results under combined working conditions

[0350]

[0351] From Table 5 and in combination Figure 10It can be seen that under combined working conditions, the lateral force decreases to a certain extent as the longitudinal slip ratio increases, and this invention can capture this coupling phenomenon well. Compared with Comparative Example 1, the joint prediction scatter points of this invention under combined working conditions are more concentrated near the ideal diagonal, indicating that the working condition coupling state vector... With its layered training mechanism, the model can learn the complex coupling relationship between slip, bias, and load more accurately.

[0352] (v) Ablation test analysis

[0353] To further illustrate the contribution of the key technical features of this invention, ablation experiments were conducted. The ablation targets included:

[0354] A: Remove the intermediate physical characterization prediction module for grounding imprints;

[0355] B: Remove the stratified training based on work conditions and only use hybrid training;

[0356] C: Remove the consistency constraint term for the aligning torque.

[0357] Table 6 Ablation test results of key modules of the present invention

[0358]

[0359] As shown in Table 6, removing the intermediate physical quantity module significantly increases the errors of lateral force and restoring torque, indicating that bridging the intermediate physical quantity has a significant effect on improving the interpretability and accuracy of the model; removing the hierarchical training significantly reduces the prediction ability of composite working conditions, indicating that the working condition hierarchical strategy is an important means to improve the generalization ability of multiple working conditions; removing the consistency constraint of the restoring torque, although the overall... The decrease was limited, but the RMSE of the restoring torque increased significantly, indicating that the constraint was effective for stability. Forecasting is especially crucial.

[0360] (vi) Conclusion on Technical Effectiveness

[0361] Based on the above experimental results, the following conclusions can be drawn:

[0362] 1. This invention can uniformly predict longitudinal force, lateral force, normal force, rollover torque, rolling resistance torque and self-aligning torque under the combined effect of tire structural parameters, material parameters and working condition parameters. The overall prediction accuracy is significantly better than traditional end-to-end neural network methods, simple basic mechanical model methods and hybrid model methods that do not use intermediate physical quantities for bridging and hierarchical training.

[0363] 2. The present invention exhibits high prediction accuracy under pure sideslip, pure longitudinal slip, and combined conditions. In particular, it maintains a good coefficient of determination and low error under combined conditions, indicating that the present invention has strong condition coverage and generalization ability.

[0364] 3. This invention establishes a physical bridging relationship from input parameters to six-component force output by using intermediate physical characterization quantities of grounding imprints, making the prediction of torque quantities such as aligning torque, overturning torque and rolling resistance torque more stable, and reducing the risk of non-physical results generated by pure black box models under extreme working conditions.

[0365] 4. This invention employs layered training based on working conditions and consistency constraints on the return torque, which significantly enhances the ability to express the intrinsic correlation between the six components of tire force, thereby improving the model's adaptability to complex coupled working conditions.

[0366] Therefore, compared with the prior art, the present invention can significantly improve the accuracy, physical rationality and engineering applicability of tire six-component force prediction while maintaining high prediction efficiency. It can be widely used in tire structure design optimization, material formulation screening, vehicle dynamics analysis, digital prototype development and vehicle active safety control.

[0367] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

[0368] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0369] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0370] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0371] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0372] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0373] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0374] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

Claims

1. A method for predicting the six components of tire force based on neural networks, characterized in that, The steps include the following: S1. Obtain the original sample data of multiple tire samples under different working conditions. The original sample data includes at least the tire structure parameter set, tire material parameter set, tire working condition parameter set, and the corresponding tire six-component force measured label set. S2. Preprocess the original sample data and construct the structural representation vector, material representation vector and working condition coupling state vector respectively. At the same time, according to the working condition type, divide the samples into pure lateral slip working condition sample subset, pure longitudinal slip working condition sample subset and composite working condition sample subset to form a hierarchical training dataset. S3. Construct a neural network for predicting the six components of tire force; S4. The hierarchical training dataset is used to train the tire six-component force prediction neural network. Training loss weights are set for the pure sideslip condition sample subset, the pure longitudinal slip condition sample subset, and the composite condition sample subset, respectively. A consistency constraint on the offset of the lateral force action line is introduced for the prediction of the return torque, so as to obtain the trained tire six-component force prediction model. S5. Obtain the target structure parameter set, target material parameter set, and target working condition parameter set of the target tire. After processing by the trained tire six-component force prediction model, output the longitudinal force prediction value, lateral force prediction value, normal force prediction value, rollover torque prediction value, rolling resistance torque prediction value, and self-aligning torque prediction value of the target tire under the target working condition.

2. The tire six-component force prediction method based on neural networks according to claim 1, characterized in that: In step S1, the tire structure parameter set includes at least one or more of the following parameters: tread width, section height-to-width ratio, rim diameter, carcass cord angle, belt layer angle, belt layer width, sidewall thickness, and tread pattern type code. And / or, In step S1, the tire material parameter set includes at least one or more of the following parameters: tread rubber equivalent elastic modulus, carcass rubber equivalent elastic modulus, belt layer rubber equivalent elastic modulus, sidewall rubber equivalent elastic modulus, loss factor, cord layer equivalent elastic modulus, and material equivalent density. And / or, In step S1, the tire operating condition parameter set includes at least one or more of the following parameters: inflation pressure, vertical load, slip angle, longitudinal slip ratio, camber angle, driving speed, road adhesion coefficient, and road roughness factor; And / or, In step S1, the tire six-component force measurement label set includes longitudinal force, lateral force, normal force, rollover torque, rolling resistance torque, and self-aligning torque.

3. The tire six-component force prediction method based on neural networks according to claim 1, characterized in that: In step S2, the preprocessing includes: outlier removal or correction, missing value completion, discrete variable encoding, continuous variable normalization, time series signal filtering, and steady-state interval truncation. For continuous variables, standardization is performed by subtracting the mean of the original variable from the training sample and then dividing by the standard deviation of the original variable from the training sample to obtain the standardized variable. And / or, In step S2, the working condition type classification rule is as follows: when the absolute value of the longitudinal slip ratio is less than the preset slip threshold and the absolute value of the side slip angle is greater than the preset side slip threshold, the sample is classified into a pure side slip working condition sample subset; when the absolute value of the side slip angle is less than the preset side slip threshold and the absolute value of the longitudinal slip ratio is greater than the preset slip threshold, the sample is classified into a pure longitudinal slip working condition sample subset; when the absolute value of the side slip angle is greater than the preset side slip threshold and the absolute value of the longitudinal slip ratio is greater than the preset slip threshold, the sample is classified into a composite working condition sample subset.

4. The tire six-component force prediction method based on neural networks according to claim 1, characterized in that: In step S3, the tire six-component force prediction neural network includes: The structural encoding branch is used to receive structural representation vectors and output structural latent features. The material encoding branch is used to receive the material characterization vector and output the material's implicit features; The operating condition coding branch is used to receive the operating condition coupled state vector and output the operating condition implicit features; The feature fusion module is used to fuse the implicit structural features, implicit material features, and implicit working condition features, and output shared features; Intermediate physical quantity prediction module, used to predict the set of intermediate physical characterization quantities of grounding imprints based on the shared features; The basic mechanical model module is used to calculate the initial value set of the basic six-component forces based on the tire structure parameter set, tire material parameter set, tire operating condition parameter set, and intermediate physical characterization quantity set of the ground imprint. The residual correction module is used to output the residual correction set corresponding to the initial value set of the basic six component forces; The six-component force generation module is used to output the final tire six-component force prediction result based on the basic six-component force initial value set and the residual correction amount set.

5. The tire six-component force prediction method based on neural networks according to claim 4, characterized in that: The intermediate physical characterization set of the grounding imprint includes at least one or more of the following parameters: grounding imprint length, grounding imprint width, longitudinal offset of the normal pressure center, lateral offset of the normal pressure center, equivalent offset distance of the lateral force action line, and average tangential shear stress of the grounding area.

6. The tire six-component force prediction method based on neural networks according to claim 1, characterized in that: The training loss function in step S4 includes a six-component force error term, a layered working condition weighting term, and a self-correcting torque consistency constraint term; The total loss function is obtained by weighted summation of a six-component force error term, a working condition weighting term, and a aligning torque consistency constraint term. The six-component force error term characterizes the error between the predicted six-component force and the measured label. The layered working condition weighting term characterizes the weighted loss of different working condition subsets. The aligning torque consistency constraint term characterizes the consistency between the predicted aligning torque and the offset of the lateral force line of action. The coefficients of the working condition weighting term and the aligning torque consistency constraint term are used to adjust the weights of the corresponding terms in the total loss function.

7. The tire six-component force prediction method based on neural networks according to claim 1, characterized in that: In step S5, the final tire six-component force prediction result is obtained by adding the basic six-component force initial value set and the residual correction amount set; And / or, In step S5, the predicted value of the return torque is obtained by multiplying the predicted value of the lateral force by the equivalent offset distance of the line of action of the lateral force relative to the center of the tire contact area, and then adding it to the residual correction amount of the return torque.

8. A tire six-component force prediction system based on neural networks, characterized in that, include: The data acquisition module is used to execute step S1; The data preprocessing and hierarchical construction module is used to perform step S2; A neural network building module is used to execute step S3; The model training module is used to perform step S4; The six-component force prediction module is used to execute step S5; The system is configured to implement the method described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the method described in any one of claims 1-7.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the method described in any one of claims 1-7.

Images