A neural network-based tire wear 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 wear prediction method and system based on a neural network. The method acquires tire structure parameters, material parameters, use condition time sequence parameters, finite element wear response data and measured wear calibration data; maps the finite element wear response to a normalized tread coordinate grid and corrects it with measured data to construct a supervised wear field label; establishes a training sample accordingly and constructs a neural network model containing a static parameter branch, a time sequence condition branch, a wear field decoding branch and a physical consistency constraint branch; and outputs the wear distribution field, total wear amount and partition wear amount of the target tire after training. The present application can improve the accuracy of tire wear prediction, accurately represent uneven wear characteristics, reduce test costs and be suitable for tire design optimization, uneven wear warning and intelligent maintenance.

Description

Technical Field

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

[0002] As the only component of a vehicle in contact with the road surface, the wear condition of tires directly affects not only their grip, braking performance, and handling stability, but also vehicle safety, energy consumption, and tire lifespan. Especially during long-term service, tire wear is not a simple, uniform thinning process, but rather is influenced by a combination of factors, including tire structural parameters, tread rubber material properties, inflation pressure, load, vehicle speed, slippage, road conditions, and ambient temperature, exhibiting significant nonlinearity, time-varying characteristics, and spatial non-uniformity. Therefore, accurately and efficiently predicting tire wear under different operating conditions has always been a crucial technical challenge in tire design, testing and evaluation, and intelligent maintenance.

[0003] Existing tire wear assessment methods mainly include road vehicle testing, bench acceleration testing, and analytical methods based on empirical formulas or physical mechanisms. While road vehicle testing can realistically reflect tire wear under actual operating conditions, it suffers from long testing cycles, high costs, significant environmental influences, and relatively poor repeatability. Bench testing, although it can shorten the evaluation cycle to some extent, typically struggles to fully reproduce the real-world processes of tire stress, heat generation, slippage, and ground pressure evolution under complex road conditions. On the other hand, wear analysis methods based on physical mechanisms or finite element simulation can explain tire wear processes from the perspectives of contact mechanics, tribology, and material response. However, these methods often require complex material constitutive models, contact models, and wear evolution models, placing high demands on parameter acquisition and computational resources. Therefore, they still have limitations in rapid engineering prediction and online applications.

[0004] With the development of sensor technology, data processing technology, and artificial intelligence methods, predicting tire wear using data-driven models has become an important development direction in recent years. In existing technology, document CN113239599A discloses a "Smart Tire Wear Life Prediction Method and Device Based on BP Neural Network." This scheme acquires a dataset containing tire pressure, vehicle speed, load, and the radial 2nd to 6th order elevation modal frequencies of the tire. The data is divided into training, validation, and test sets to establish a BP neural network model, using tire wear as the output to predict tire wear life. This existing technology demonstrates that using neural network methods to replace traditional manual detection or purely empirical judgment can, to a certain extent, reduce the cost of tire wear prediction and improve the efficiency of wear life prediction, providing a feasible technical route for intelligent tire condition prediction.

[0005] However, for those skilled in the art, the aforementioned prior art still has at least the following shortcomings: First, the input features are mainly concentrated on limited parameters such as air pressure, vehicle speed, load, and modal frequency, which do not adequately consider the comprehensive coupling effect of tire structural parameters, tread rubber material parameters, and complex operating condition parameters, making it difficult to fully reflect the multi-factor coupling characteristics in the tire wear formation mechanism; Second, this type of method is mainly aimed at predicting overall wear or life, and is more focused on single scalar results, which are insufficient for characterizing the wear differences between different areas such as the tire shoulder, tire crown, and central area, as well as the non-uniformity of tread wear distribution, making it difficult to meet the needs of tire wear analysis and structural analysis. The existing solutions typically train networks directly based on existing measurement data, which fails to adequately address the issues of consistent representation and comparability of wear samples under different tire specifications, tread patterns, and operating conditions, thus affecting the model's generalization ability and cross-sample applicability. Furthermore, the current technology lacks a more systematic constraint mechanism for the physical laws governing the gradual evolution of wear with increasing mileage, as well as the inherent consistency between wear and the evolution of friction work, slip work, and ground pressure distribution. This results in room for improvement in the physical interpretability and engineering reliability of the model output.

[0006] Therefore, it is still necessary to propose a new tire wear prediction method based on existing technologies, so as to further improve the ability to characterize the coupling relationship of multi-source parameters, the spatial distribution characteristics of wear, and the physical laws of wear evolution while retaining the advantages of neural network prediction efficiency, so as to obtain more accurate, stable and engineering application-valued tire wear prediction results. Summary of the Invention

[0007] The technical objective of this invention is to address the problems of existing tire wear prediction methods, such as high testing costs, long testing cycles, insufficient characterization of the coupling effects of multiple factors, and difficulty in accurately reflecting the uneven wear distribution and wear evolution of the tire tread. This invention provides a tire wear prediction method based on neural networks. By integrating tire structural parameters, material parameters, operating condition parameters, and wear field label information, a tire wear prediction model is established that balances prediction accuracy, physical consistency, and engineering applicability. This enables efficient and accurate prediction of total tire wear, zoned wear, and wear distribution.

[0008] Firstly, in order to achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0009] A method for predicting tire wear based on neural networks includes the following steps:

[0010] S1. Construct a tire sample set and obtain the structural parameter vector, material parameter vector, operating condition time sequence vector, finite element wear response data, and measured wear calibration data for each tire sample.

[0011] S2. Map the finite element wear response data to a unified grid in the normalized tread coordinate system to form an initial wear field, and correct the deviation of the initial wear field based on the measured wear calibration data to form a supervised wear field label.

[0012] S3. Calculate the total wear amount, multiple tread zone wear amount vectors, and physical consistency feature vectors based on the supervised wear field labels, and construct training samples;

[0013] S4. Construct a neural network model and train it. The neural network model includes a static parameter feature extraction branch, a time-series operating condition feature extraction branch, a wear field decoding branch, and a physical consistency constraint branch.

[0014] S5. Obtain the structural parameter vector, material parameter vector, and preset operating condition time sequence vector of the target tire, input them into the trained neural network model, first output the predicted wear field of the target tire, and then calculate the total wear amount and wear amount of each zone of the target tire from the predicted wear field to obtain the wear prediction result of the target tire under the preset operating conditions.

[0015] As a further improvement, in step S1, the structural parameter vector includes at least one of the following: tread width, cross-sectional aspect ratio, carcass cord angle, belt layer angle, tread groove depth, tread block pitch distribution parameters, and shoulder transition radius.

[0016] And / or,

[0017] In step S1, the operating condition parameters in the operating condition time sequence vector include at least one of the following parameters: inflation pressure, vertical load, sideslip angle, longitudinal slip ratio, driving speed, road surface adhesion coefficient, ambient temperature, cumulative driving mileage, and road surface roughness.

[0018] As a further improvement, in step S2, the normalized tread coordinate system consists of lateral normalized coordinates and circumferential normalized coordinates. The lateral normalized coordinates are obtained by the ratio of the actual coordinates in the tread width direction to the tread width, and the circumferential normalized coordinates are obtained by the ratio of the tread circumferential arc length coordinates to the effective circumferential length of the tread. The actual coordinates in the tread width direction are the actual position coordinates in the tire tread width direction, the tread width is the effective width of the tire tread, the tread circumferential arc length coordinates are the arc length position coordinates of the tire tread after circumferential unfolding, and the effective circumferential length of the tread is the effective unfolded length of the tire tread in the circumferential direction.

[0019] And / or,

[0020] In step S2, the deviation correction is achieved using a calibration function between the finite element wear field and the measured wear data. The corrected monitoring wear field label is obtained by inputting the initial wear field and the measured wear calibration data into the calibration function. The calibration function is used to correct the deviation between the finite element results and the measured results.

[0021] As a further improvement, in step S3, the physical consistency feature vector includes at least one of the following: cumulative friction work, cumulative slip work, and ground pressure distribution evolution features.

[0022] As a further improvement, in step S4, the static parameter feature extraction branch is used to extract the static features of the structural parameter vector and the material parameter vector; the time-series working condition feature extraction branch is used to extract the evolution features of the working condition time-series vector; the wear field decoding branch is used to output the predicted wear field and derive the predicted total wear amount and predicted partition wear amount vector based on the predicted wear field; and the physical consistency constraint branch is used to constrain the correspondence between the prediction result and the physical consistency feature vector.

[0023] And / or,

[0024] In step S4, a comprehensive loss function is used when training the neural network model. The comprehensive loss function is composed of a weighted sum of the wear field error term between the predicted wear field and the supervised wear field label, the total amount error term between the predicted total wear amount and the total wear amount, the partition error term between the predicted partition wear amount vector and the partition wear amount vector, and the physical consistency constraint error term.

[0025] Wherein, the wear field error term is used to characterize the difference between the predicted wear field and the supervised wear field label, the total error term is used to characterize the difference between the predicted total wear amount and the total wear amount, the partition error term is used to characterize the difference between the predicted partition wear amount vector and the partition wear amount vector, and the physical consistency constraint error term is used to characterize the deviation between the prediction result and the physical consistency constraint. The weight coefficients corresponding to each error term are all real numbers greater than 0.

[0026] The physical consistency constraint error term includes at least one of the following constraints:

[0027] Predicted wear is monotonically non-decreasing with increasing cumulative mileage;

[0028] The increase in tread wear in different zones follows the same trend as the increase in the cumulative friction work or cumulative slip work of the corresponding zones.

[0029] The trend of the tread zone wear increment is consistent with the evolution characteristics of the ground pressure distribution;

[0030] Among them, monotonic non-decreasing means that the predicted wear value of subsequent mileage nodes is not less than the predicted wear value of the previous mileage node.

[0031] As a further improvement, in step S4, the static parameter feature extraction branch is a multi-layer fully connected network, the temporal condition feature extraction branch is any one of a long short-term memory network, a gated recurrent unit, or a temporal convolutional network, and the wear field decoding branch is any one of a deconvolutional network, an upsampling convolutional network, or a fully connected mapping network.

[0032] As a further improvement, step S5 is followed by an online correction step: real-time acquisition of at least one of the tire temperature signal, tire pressure signal, acceleration signal, and wheel speed signal; extraction of incremental features of actual operating conditions; updating of the preset operating condition time-series vector; and inputting the updated time-series vector back into the trained neural network model to obtain dynamically updated wear field, total wear amount, and zone wear amount.

[0033] Secondly, the present invention also provides a tire wear prediction system based on a neural network, comprising:

[0034] The data acquisition module is used to execute step S1;

[0035] The tag building module is used to perform step S2;

[0036] The sample construction module is used to execute step S3;

[0037] The model training module is used to perform step S4;

[0038] The prediction output module is used to execute step S5.

[0039] 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 method.

[0040] Fourthly, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the method.

[0041] Compared with existing technologies, this invention jointly models tire structural parameters, material parameters, and time-series parameters of operating conditions, and constructs a unified wear field label by combining finite element wear response. Then, it uses measured data for calibration to form supervisory information. This allows the model to not only predict the overall wear level of the tire but also output the zonal wear amount and wear distribution of different areas of the tread. Therefore, it can more realistically reflect the non-uniform wear characteristics of tires under complex service conditions. Simultaneously, this invention employs a neural network structure combining static parameter branches and time-series operating condition branches, and introduces physical consistency constraints on wear evolution. This makes the prediction results more consistent with the actual tire wear mechanism in terms of the trend of changes with mileage growth and the correspondence with friction work or ground pressure evolution, thereby effectively improving the model's prediction accuracy, stability, and interpretability. Furthermore, this invention uses a combination of simulation and measurement to construct training samples, reducing reliance on long-cycle road tests and large-scale bench tests. This significantly shortens the tire wear evaluation cycle, reduces testing costs, and provides reliable data support and technical means for tire structure optimization design, uneven wear risk warning, use and maintenance strategy formulation, and intelligent tire online condition assessment. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the structure of the tire wear prediction system based on neural networks according to the present invention.

[0043] Figure 2 This is a flowchart of the tire wear prediction method based on neural networks of the present invention.

[0044] Figure 3 This is a schematic diagram illustrating the construction of the wear field monitoring label of the present invention.

[0045] Figure 4 This is a schematic diagram of the neural network model structure of the present invention.

[0046] Figure 5 This is a schematic diagram of the online correction and prediction output of the present invention.

[0047] Figure 6 This is a convergence curve of the training loss and validation loss of this invention.

[0048] Figure 7 This is a comparison chart of the predicted wear field and the measured wear field of this invention.

[0049] Figure 8 This is a bar chart comparing the prediction errors of the wear amount in different zones according to the present invention.

[0050] Figure 9 This is a comparison chart of the prediction accuracy of total wear amount under different methods of the present invention. Detailed Implementation

[0051] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0052] This invention provides a neural network-based method for predicting tire wear. This approach addresses three long-standing core problems in tire wear prediction: First, the lack of a unified representation of wear data among tires of different specifications, tread patterns, and operating conditions makes direct comparison of samples difficult; second, relying solely on road or bench tests to build training sets is time-consuming and costly, while relying solely on finite element simulation suffers from domain bias compared to measured results; third, while many existing data-driven solutions can output a certain total wear amount or lifespan value, they do not adequately reflect the spatial distribution of wear, regional wear patterns, and wear evolution laws. To address these problems, this invention uses a core technical approach of "unified wear field label construction, simulation and measured coupled calibration, joint modeling of static parameters and time-series operating conditions, physical consistency constraint training, and hierarchical prediction output" to achieve unified prediction of total tire wear, regional wear, and wear distribution.

[0053] I. Terminology Explanation

[0054] To enable those skilled in the art to accurately understand this invention, the main terms used in this specification are explained below.

[0055] Structural parameters: These are the set of parameters used to characterize the tire's geometry, laminated structure, and tread pattern, including but not limited to tread width, section height-to-width ratio, carcass cord angle, belt layer angle, tread groove depth, tread pitch distribution parameters, and shoulder transition radius.

[0056] Material parameters: refers to the set of parameters used to characterize the physical and mechanical properties, viscoelastic properties and wear resistance of tread rubber and related component materials, including but not limited to storage modulus, loss factor, hardness, wear resistance coefficient, filler addition amount, crosslinking density and glass transition temperature.

[0057] Using the operating condition time sequence vector: refers to the set of operating condition parameters arranged in order of cumulative mileage or time. Each time sequence node corresponds to the tire service status within a mileage interval or time interval, including but not limited to inflation pressure, vertical load, vehicle speed, sideslip angle, longitudinal slip ratio, ambient temperature, road adhesion coefficient, and road roughness.

[0058] Wear field: refers to the spatial distribution data set formed by mapping the wear state of each discrete position of the tire tread to a unified coordinate grid. It can be expressed in the form of wear depth, remaining groove depth, wear rate or volume loss.

[0059] Normalized tread coordinate system: refers to a unified coordinate system established after normalizing the tread width and circumferential position of tires of different specifications. It is used to achieve comparability and learnability of wear distribution of different tire samples.

[0060] Supervisory wear field label: refers to the training label formed after the initial wear field obtained by finite element simulation is combined with the measured groove depth or contour scan data for deviation correction.

[0061] Zoned wear: refers to the wear characteristics obtained by dividing the tire tread into several zones in the lateral or lateral and circumferential directions, and then statistically analyzing the wear values ​​in the corresponding zones, such as the wear of the left shoulder zone, the central zone, and the right shoulder zone.

[0062] Physical consistency constraints refer to the introduction of constraints that are consistent with the actual wear patterns of tires during the model training process, so that the network output, in addition to fitting the labels, also satisfies the laws that wear does not decrease with cumulative mileage, is consistent with the growth trend of friction work, and is coordinated with the evolution of ground pressure distribution.

[0063] Static parameter branch: refers to the neural network branch used to process structural and material parameters. Its input does not change with the mileage sequence and mainly extracts the inherent design properties of the tire.

[0064] Timing-based operating condition branch: This refers to the neural network branch used to process the timing vector of operating conditions. Its input changes gradually with the mileage interval and mainly extracts the operating condition evolution characteristics during the tire's service life.

[0065] II. System Structure of the Invention

[0066] like Figure 1 As shown, the tire wear prediction system of the present invention may include a data acquisition module 10, a label construction module 20, a sample construction and preprocessing module 30, a model training module 40, a prediction output module 50, and an optional online correction module 60.

[0067] The data acquisition module 10 is used to collect or retrieve structural parameters, material parameters, operating parameters, finite element analysis results, and measured wear data of tire samples. Structural parameters can be obtained from tire design drawings, CAD models, finite element preprocessing parameter files, or product databases; material parameters can be obtained from laboratory DMA tests, hardness tests, wear tests, and formula databases; operating parameters can be obtained from road tests, bench tests, vehicle CAN bus, smart tire sensors, or preset operating condition libraries; finite element wear response data can be output from the tire finite element analysis platform; and measured wear data can be obtained through laser profile scanning, contact groove depth gauges, three-dimensional surface scanning, or bench testing.

[0068] The label construction module 20 is used to convert the wear response output by the finite element method into an initial wear field under a unified coordinate grid, and to correct for deviations by combining it with measured data, outputting supervised wear field labels that can be used for model training. This module is one of the important modules of this invention, which solves the problems of inconsistent scales and expression methods of wear data from different tire samples, as well as the systematic deviation between simulation and measurement.

[0069] The sample construction and preprocessing module 30 is used to further calculate the total wear amount, partition wear amount and physical consistency features based on the supervised wear field label, and to complete operations such as sample cleaning, outlier removal, discrete variable encoding, continuous variable normalization, time series alignment and training set / validation set / test set partitioning.

[0070] The model training module 40 is used to build and train a neural network model. The neural network model includes at least a static parameter feature extraction branch, a time-series operating condition feature extraction branch, a feature fusion layer, a wear field decoding layer, and a physical consistency constraint layer. After training, it outputs a tire wear prediction model parameter file.

[0071] The prediction output module 50 is used to receive the structural parameters, material parameters and preset operating condition time vectors of the target tire during the deployment phase, call the trained model, first output the wear field of the target tire, and then calculate the total wear amount and the wear amount of each zone to give the wear prediction result of the target tire under the target service conditions.

[0072] The online correction module 60 is an optional module. It is used to receive online monitoring signals such as temperature, tire pressure, acceleration, and wheel speed in real time during actual tire use, dynamically correct the original operating condition sequence, and trigger the model to perform rolling update predictions. Through this module, the present invention can be extended from offline design prediction to online state estimation.

[0073] III. Specific Technical Route for Implementing the Method of the Invention

[0074] like Figure 2As shown, the method of the present invention can be summarized into five steps, S1 to S5. Each step is described in detail below.

[0075] (a) Step S1: Construct a tire sample set and obtain multi-source basic data

[0076] Step S1 is used to form the original sample library and is the starting point of the entire method. This step includes three sub-processes: sample object determination, sample data item definition, and data collection and storage.

[0077] 1. Determining the Sample Object

[0078] In one embodiment, the tire sample can be selected from passenger car radial tires, light truck tires, commercial vehicle tires, or new energy vehicle tires. To ensure the representativeness of the model, the tire sample should cover at least two specifications, two tread patterns, and two tread compound systems. For passenger car tires, specifications such as 205 / 55R16, 215 / 55R17, and 225 / 45R18 can be selected. For each specification, different belt layer angles, different tread pitch arrangements, and different shoulder profile designs can be further covered.

[0079] In a set of recommended embodiments, no fewer than 300 complete sample records can be constructed; in a more preferred embodiment, the number of complete sample records is no fewer than 800; and in industrial deployment, the total number of samples can be increased to more than 2,000. Here, "one complete sample record" can be understood as a set of input features and their wear labels corresponding to a certain tire design object under a certain working condition sequence.

[0080] 2. Data Item Definition

[0081] The data collected by this invention includes at least three types of inputs and two types of label sources.

[0082] The first type of input is a structure parameter vector. It may include, but is not limited to, the following parameters: tread width. The unit can be millimeters; the aspect ratio of the cross-section. tire cord angle The unit can be degrees; belt layer angle The unit can be degrees; the depth of the groove. The unit can be millimeters; pattern pitch distribution parameters It can be expressed as several pitch lengths and their proportions; tire shoulder transition radius The unit can be millimeters.

[0083] The second type of input is a material parameter vector. It may include: energy storage modulus The unit can be MPa; loss factor Shore hardness abrasion resistance coefficient Filler addition amount For example, phr; crosslinking density Glass transition temperature The unit can be Celsius.

[0084] The third type of input is the operating condition time sequence vector. In one embodiment, the target service process can be divided into: There are several mileage intervals, each with a length of 100km, 200km, 500km, or 1000km; this invention does not impose an absolute limitation on these lengths. Operating condition nodes are defined within each mileage interval. It may include at least: inflation pressure Vertical load Side slip angle Longitudinal slip ratio driving speed Road surface adhesion coefficient Ambient temperature Cumulative mileage Road surface roughness parameters .

[0085] Finite element wear response data The first type of label source can be obtained through finite element simulation. Specific content may include tread node wear depth, ground pressure distribution, friction work distribution, temperature rise distribution, slip energy dissipation, etc.

[0086] Actual wear calibration data The second type of label source can be obtained through actual measurement, including but not limited to the remaining groove depth at each measuring point, the contour scan surface shape, the volume loss of the region, and the results of accelerated wear on the test bench.

[0087] 3. Data collection and storage

[0088] Structural parameters can be directly read from the tire design system; material parameters can be imported from material performance test data tables; operating parameters can be collected from test databases or on-board records. Finite element wear response data can be output from commercial software or self-developed platforms, such as Abaqus, ANSYS, LSDYNA, Marc, or dedicated wear solvers based on user subroutines. Measured data can be collected by laser scanning equipment, contact depth gauges, or online detection systems.

[0089] To facilitate subsequent training, a unified data table structure should be used when collecting and storing data. The following fields are recommended: Sample Number, Tire Specification, Tread Pattern Type, Structural Parameter Field Group, Material Parameter Field Group, Operating Condition Sequence Field Group, Finite Element Wear Data File Path, Measured Calibration Data File Path, Data Source Marker, Timestamp, and Validity Marker.

[0090] In one embodiment, the raw data is managed using a combination of relational database and file system, wherein parametric data is stored in database tables and field distribution data is stored in the form of matrix files or tensor files.

[0091] (ii) Step S2: Constructing supervised wear field labels under a unified coordinate grid

[0092] Step S2 does not simply input the finite element simulation results directly as labels into the network. Instead, it uses a two-stage process of unified mesh mapping and measured calibration to create supervised wear field labels that can learn across specifications, patterns, and operating conditions. This step plays a crucial role in resolving sample inconsistency and simulation-measurement discrepancies.

[0093] 1. Establish a normalized tire tread coordinate system

[0094] Because different tire samples differ in tread width, outer circumferential dimensions, and tread pitch distribution, directly inputting the finite element nodal wear results into the model makes it difficult to establish a one-to-one correspondence between different samples in spatial location, and the model struggles to learn the "wear patterns of a certain lateral region and a certain circumferential region." Therefore, this invention first establishes a normalized tread coordinate system.

[0095] In one embodiment, a lateral normalized coordinate is established along the tread width direction. Establish circumferential normalized coordinates based on the circumferential direction of the tire tread. Its expression is:

[0096] ;

[0097] in, These are the actual coordinates along the tread width direction; This refers to the tread width; The coordinates are the circumferential arc length of the tire tread. This refers to the effective circumferential length of the tire tread. For horizontally normalized coordinates; For circumferential normalized coordinates.

[0098] In practice, The range of values ​​can be set to or , The range of values ​​can be set to Those skilled in the art can flexibly choose the specific interval definition based on whether the pattern is symmetrical or whether it is necessary to retain absolute left and right position information.

[0099] 2. Unified grid division

[0100] After establishing a normalized coordinate system, the tire tread is further divided into a uniform grid. In a recommended embodiment, the lateral division can be... Each grid cell is divided circumferentially into [number] grid units. There are [number] grid cells, and the total number of grid cells is [number]. For scenarios with high computational efficiency requirements, it can also be used. Mesh; for scenarios requiring higher spatial resolution, a mesh can be used. Grid.

[0101] For the nodal or element wear response output from the finite element analysis, a spatial interpolation method is used to map it into a unified mesh. Optional interpolation methods include nearest neighbor interpolation, bilinear interpolation, inverse distance weighted interpolation, spline interpolation, or area-weighted averaging based on element coverage relationships. If the finite element output is nodal wear depth, bilinear interpolation is preferred; if the output is element volume loss, area-weighted averaging is preferred.

[0102] After mapping, the initial wear field matrix can be obtained. Among them, matrix elements Indicates the position in the horizontal direction. line, circumferential direction Initial wear characterization value of column grid cells.

[0103] 3. Formation of the finite element wear field

[0104] This invention does not strictly limit the source of the finite element wear field, as long as a spatial response reflecting the tread wear trend can be obtained. In one embodiment, the finite element analysis is divided into two levels.

[0105] The first level is contact condition analysis, which involves solving the tire ground pressure distribution, ground shear force distribution, and local friction work density under given tire pressure, load, speed, sideslip angle, and slip ratio.

[0106] The second level is wear accumulation analysis, which involves calculating the wear increment of each node or unit based on variables such as local friction work, sliding speed, ground pressure, and temperature field, combined with the wear law model, and then gradually accumulating it along the mileage direction to form the end-point wear distribution.

[0107] In a recommended embodiment, the wear increment relationship can be used in the finite element solution in the following form:

[0108] ;

[0109] in, For the first The wear increment at a discrete location within this mileage step; The wear coefficient; For the first The local wear driving force at a discrete location can be represented by local contact pressure, sliding velocity, friction work density, or a combination thereof; This corresponds to the increment in driving mileage.

[0110] It should be noted that this formula is only one possible expression of wear increment, used to facilitate implementation by those skilled in the art, and does not constitute a limitation on the present invention. The core of the present invention does not lie in the finite element wear formula itself, but in the construction method of finite element wear response and unified wear field label.

[0111] 4. Calibration of measured data

[0112] While relying solely on the wear field obtained through finite element simulation as a label offers high sample acquisition efficiency, systematic deviations often exist between simulation results and actual wear due to factors such as simplified material parameters, idealized boundary conditions, and approximate wear mechanism models. Therefore, this invention further incorporates measured wear calibration data to correct the deviations in the initial wear field.

[0113] Actual wear calibration data There can be multiple sources. In one embodiment, after the tire has traveled to a predetermined mileage, the three-dimensional surface profile of the tire tread is obtained by laser contour scanning and then converted into groove depth change values; in another embodiment, multiple measuring points are selected on the left shoulder, middle and right shoulder of the tire, and the remaining groove depth is measured with a groove depth gauge; in yet another embodiment, bench test results can be used as the basis for actual measurement calibration.

[0114] Calibration methods can include linear calibration, zonal calibration, nonlinear regression calibration, or a small calibrator based on neural networks. To balance ease of implementation and calibration effectiveness, a recommended embodiment employs zonal linear calibration. This involves first dividing the tire tread into left shoulder, middle, and right shoulder zones, establishing a linear correction relationship between the finite element mean and the measured mean for each zone, and then correcting each grid point within that zone. In another preferred embodiment, a lightweight calibration network is used, taking local grid point values, average values ​​of adjacent regions, and operating condition variables as inputs, to correct complex deviations.

[0115] The revised monitoring wear field label can be represented as:

[0116] ;

[0117] in, This is the initial wear field; These are actual wear calibration data; For calibration functions; This is the revised label for monitoring wear field.

[0118] In practice, if a training sample lacks complete measured data, semi-supervised calibration can be performed using only local measured data. For example, if only the average remaining groove depth of the left shoulder, middle, and right shoulder regions is measured, the measured average of the three regions can be used to perform a zonal scale correction on the initial wear field, making its overall level closer to the measured value.

[0119] 5. The engineering significance of monitoring wear zone labels

[0120] After S2 processing, all tire samples were transformed to a unified normalized mesh, and simulation bias was reduced through experimental correction. This labeling serves three purposes:

[0121] First, it makes the wear distribution of tires of different specifications and tread patterns comparable;

[0122] Second, it provides a unified target for subsequent convolutional or decoding network outputs;

[0123] Third, it lays a unified data foundation for the analysis of zoned wear, total wear, and online uneven wear.

[0124] Therefore, those skilled in the art will understand that S2 is not simply a data format conversion step, but a key step in this invention that combines tire engineering knowledge with data-driven learning.

[0125] (III) Step S3: Sample construction, hierarchical label generation and preprocessing

[0126] The main task of step S3 is not simply to clean the data, but to further generate total wear, partition wear and physical consistency features based on the supervised wear field labels, and to construct standardized training samples suitable for network learning.

[0127] 1. Calculation of total wear

[0128] Total wear Supervised wear field label Statistical results were obtained. In one embodiment, the total wear was defined as the average wear value of all meshes:

[0129] ;

[0130] in, This represents the number of horizontal grid cells. Circumferential grid number; For the first Line 1 Wear characterization values ​​of column grids; This represents the total wear amount.

[0131] In another embodiment, the total wear can also be defined as the weighted average of all mesh values ​​or the sum of volumetric losses, depending on the type of wear characterization used. For example, if If the value represents the difference in remaining trench depth, then the average can be taken; if it represents the volume loss per unit area, then the area can be integrated.

[0132] 2. Calculation of zoned wear amount

[0133] To reflect uneven or localized wear differences, this invention further generates multiple zoned wear amounts from a uniform wear field. In a recommended embodiment, the tread is laterally divided into a left shoulder zone, a center zone, and a right shoulder zone. For implementations requiring higher resolution, it can be further subdivided into a left shoulder zone, a left center zone, a center zone, a right center zone, and a right shoulder zone.

[0134] If divided into three horizontal regions, then the first... Differential wear amount This can be expressed as the average value of the mesh wear within that region:

[0135] ;

[0136] in, For the first The grid set corresponding to each partition; This represents the number of grid cells within the partition. This corresponds to the wear characterization value of the mesh. For the first Wear and tear of each zone.

[0137] This process yields the partition wear vector:

[0138] ;

[0139] in, This represents the number of partitions.

[0140] 3. Construction of Physical Consistency Feature Vectors

[0141] This invention does not merely fit labels to data, but further constructs physically consistent feature vectors. In a recommended embodiment, At least includes cumulative friction work Cumulative sliding work and the evolution characteristics of ground pressure distribution One or more of them.

[0142] Cumulative frictional work can be expressed as the accumulation of frictional work density within the contact patch across each mileage interval. In one embodiment:

[0143] ;

[0144] in, This represents the total number of mileage intervals. This represents the number of discrete contact areas; For the first Mileage range, the first Friction work density or equivalent friction work index of each contact area; For the first The increase in mileage for each interval; This is the cumulative frictional work.

[0145] Cumulative sliding work A similar calculation method can be used, only... Replace it with the slip energy dissipation index.

[0146] Evolution characteristics of ground pressure distribution It can be a sequence of changes in average ground pressure in each zone with mileage, or it can be a low-dimensional eigenvector obtained after principal component analysis of the pressure field. In one embodiment, the eigenvector can be composed of the mean, variance, and extreme values ​​of the average ground pressure in the left shoulder, middle, and right shoulder regions.

[0147] 4. Categorical variable coding and normalization

[0148] Discrete variables such as tread pattern type, road surface type, and tire series need to be encoded before entering the network. One-hot encoding, target encoding, or embedded encoding can be used. For scenarios with a large sample size and many categories, embedded encoding is preferred; for scenarios with fewer categories, one-hot encoding is sufficient.

[0149] Continuous variables should be normalized. Mini-maximum normalization is recommended.

[0150] ;

[0151] in, These are the parameter values ​​before normalization; These are the normalized parameter values; This represents the minimum value of the corresponding parameter in the training samples; This represents the maximum value of the corresponding parameter in the training samples.

[0152] For parameters that may have abnormally large values ​​or long-tailed distributions, such as the instantaneous slip index under certain operating conditions, logarithmic transformation or truncation can be performed first, followed by normalization.

[0153] 5. Outlier handling and data partitioning

[0154] Abnormal samples may originate from experimental errors, data recording errors, or finite element solution anomalies. This invention recommends a combined approach of "rule-based screening + statistical screening + manual review." Rule-based screening includes, for example, identifying tire pressure outside the set range, negative remaining groove depth, missing speed fields, or corrupted finite element metadata files. Statistical screening includes, for example, marking samples where the Z-score for a certain parameter exceeds a threshold for tires of the same specification as suspicious. For important samples, manual review can be arranged to avoid mistakenly deleting valid extreme condition samples.

[0155] When partitioning the dataset, it is recommended to divide it into training, validation, and test sets based on tire sample objects. This avoids the same tire object falling into both training and testing sets simultaneously at different mileage points, which could lead to data leakage. A recommended ratio is 70% training set, 15% validation set, and 15% test set. For scenarios with a small total sample size, K-fold cross-validation can also be used.

[0156] 6. Technical effects of step S3

[0157] Through step S3, this invention transforms the complex tire wear problem into a set of training samples with clear hierarchy, explicit physical meaning, and direct input into the network: the input side consists of static parameters and time-series operating conditions, the label side consists of the wear field, total wear amount, and zone wear amount, and the constraint side consists of friction work, slip work, and pressure evolution. This structured sample design enables the model to learn statistical laws while also improving its generalization ability by leveraging engineering physical constraints.

[0158] (iv) Step S4: Neural network model construction, training methods and parameter selection

[0159] Step S4 is to meet the requirement that "the model structure, training method, parameter selection and dataset usage must be disclosed in a structured manner". The following is a full explanation of this step.

[0160] 1. Overall Model Structure

[0161] like Figure 4 As shown, the neural network model used in this invention generally includes the following parts: static parameter feature extraction branch 41, time-series operating condition feature extraction branch 42, feature fusion layer 43, wear field decoding branch 44, derived computation layer 45, and physical consistency constraint layer 46.

[0162] Static parameter feature extraction branch 41 receives the structural parameter vector. and material parameter vector Its task is to extract the inherent design features of tires, namely, the prior attributes that do not change with mileage but significantly affect the wear pattern. For example, under the same operating condition sequence, different tread pitch schemes, different belt layer angles, and different tire shoulder profiles will cause different ground pressure migration and local wear trends.

[0163] Timing condition feature extraction branch 42 receives the timing vector of the operating condition. Its task is to extract the time-varying characteristics of tires caused by variables such as load, tire pressure, speed, temperature, and slippage during tire service. Since tire wear essentially accumulates over time, using a branch specifically for processing time series helps to learn the impact of operating condition evolution on wear.

[0164] Feature fusion layer 43 is used to jointly represent static and temporal features. The fusion method can be vector concatenation, gated weighted fusion, attention fusion, or bilinear fusion. For ease of implementation and optimal results, concatenation followed by a fully connected layer is recommended; for higher accuracy, an attention mechanism can be used to dynamically adjust the intensity of structural parameter features based on the current operating conditions.

[0165] Wear field decoding branch 44 is used to map the fused high-dimensional features into a wear field prediction matrix under a unified coordinate grid. The derived computation layer 45 calculates the total wear amount based on the wear field prediction matrix. and partition wear vector In other words, this invention does not adopt the ordinary multi-output mode of "direct parallel output of total quantity and partition quantity", but takes the wear field as the core output, and then derives the total quantity and partition quantity from the wear field, thereby maintaining the internal consistency among the three.

[0166] Physical consistency constraint layer 46 is used to calculate the prediction results and physical consistency features. The deviation between the two is fed back into the total loss function to make the model output conform to the actual law of tire wear.

[0167] 2. Implementation of static parameter branches

[0168] The static parameter branches can employ multi-layer fully connected networks. In a recommended embodiment, the structure is as follows: the input layer dimension equals the total dimension of the static parameters, and it is mapped to the intermediate feature space through two to four fully connected layers. Each fully connected layer can be followed by a batch normalization layer and an activation function. The preferred activation functions are ReLU, LeakyReLU, or GELU.

[0169] For example, if the static input dimension is 24, the following network size can be used:

[0170] First fully connected layer: 24→64;

[0171] Second fully connected layer: 64→128;

[0172] Third fully connected layer: 128→128.

[0173] To prevent overfitting, Dropout can be added after the second or third layer, with the Dropout ratio selectable from 0.1 to 0.3. For cases with a large sample size and abundant parameters, the output dimension of the third layer can also be increased to 256.

[0174] 3. Implementation methods of timing condition branches

[0175] The temporal condition branches preferably use LSTM, GRU, or a temporal convolutional network (TCN). Considering the cumulative dependency of tire wear on condition evolution, a two-layer LSTM is used in a preferred embodiment. Assume that the dimension of each temporal node is... The timing length is The shape of the input tensor is .

[0176] For example, if the working condition node includes eight characteristics such as tire pressure, load, slip angle, speed, slip ratio, temperature, adhesion coefficient, and roughness, then it can be set as follows: If the total mileage is divided into 20 mileage intervals, then .

[0177] The recommended number of hidden units for a two-layer LSTM can be set to 64 for the first layer and 128 for the second layer. The output can use the hidden state at the last moment, or it can be attention-weighted pooling applied to the hidden states at all time points. For scenarios where wear is sensitive to short-term impacts under certain high-load conditions, attention pooling often achieves better results.

[0178] If deployment efficiency is a concern, GRU can be used instead of LSTM. If the data sampling frequency is high and the sequence is long, a Temporal Convolutional Network (TCN) can be used to reduce training overhead.

[0179] 4. Feature fusion and wear field decoding

[0180] The static branch output feature vector is denoted as The output feature vector of the time-series branch is denoted as In one embodiment, the two are concatenated to form a fused vector:

[0181] ;

[0182] in, To fuse feature vectors; This is a static branch characteristic; Features of temporal branches; symbols This indicates vector concatenation.

[0183] After fusion, two fully connected layers can be applied to map it to a low-resolution feature map, and then deconvolution or upsampling convolution can be used to restore it to the target wear field grid size. For example, if the target output is The grid can be mapped first as The feature map is then restored through two levels of upsampling. .

[0184] Another possible implementation is to output directly through a fully connected layer instead of using convolutional decoding. The sample size is then counted, and the matrix is ​​reshaped to form the wear field matrix. This approach is simpler to implement if the sample size is small and the grid resolution is low. However, from the perspective of learning local correlations in the learning space, convolutional decoding is usually superior.

[0185] 5. Implementation methods of the derived computation layer

[0186] In this invention, the total wear amount and the zoned wear amount are preferably determined by the predicted wear field. Instead of learning in isolation, calculations are derived directly. This design ensures that if the predicted wear field in a certain region increases, the wear amount in the corresponding region will naturally increase; the total wear amount will necessarily be consistent with the wear values ​​of all grids, thereby improving the consistency of the results.

[0187] One example of deriving total wear is as follows:

[0188] ;

[0189] in, To predict the predicted values ​​of the corresponding grid in the wear field; To predict the total wear.

[0190] Derivative One example of wear measurement for a single zone is as follows:

[0191] ;

[0192] in, For the first A grid set of partitions; This represents the number of grid cells in the partition. For the first Predicted wear amount for each partition.

[0193] 6. Loss Function Design

[0194] The loss function of this invention adopts a weighted combination of field error, total error, partitioning error, and physical consistency error. In one embodiment, the total loss function is:

[0195] ;

[0196] in, This is the wear field error term; This is the error term for total wear. This refers to the error term for wear in different zones; This is the error term for physical consistency constraints; , , , These are the weighting coefficients for each loss term.

[0197] In one embodiment, Mean square error (MSE) can be used; and Mean absolute error (MAE) or Huber loss can be used; It can consist of multiple sub-constraints.

[0198] For example, a monotonically non-decreasing constraint can be expressed as: for two adjacent mileage nodes, if the predicted wear value of the later node is lower than that of the earlier node, a penalty is applied. For the total wear, the monotonically constrained sub-item can be written as:

[0199] ;

[0200] in, For the first The predicted total wear at each mileage node; This is a monotonic penalty item.

[0201] If the consistency between zoned wear and cumulative friction work is considered, a trend-consistent loss can also be set. For example, the difference between the wear increment and the friction work increment of a certain zone can be constrained to ensure that their trends are consistent.

[0202] 7. Training methods and hyperparameter selection

[0203] In a recommended embodiment, training uses the Adam optimizer with an initial learning rate of 0.001 and batch sizes of 16, 32, or 64. The number of training epochs can be 100 to 300, and a validation set early stopping strategy is employed: training terminates if the validation loss does not decrease within 20 consecutive epochs.

[0204] The recommended weight initialization methods are Xavier initialization or He initialization. To prevent overfitting, one or more of the following measures can be used: Dropout, L2 regularization, data augmentation, K-fold cross-validation, and early stopping.

[0205] In an exemplary combination of parameters, the following can be taken: ; ; ; If the focus in actual projects is more on identifying localized uneven wear, then the accuracy can be improved. If greater emphasis is placed on physical credibility, then the level can be appropriately increased. .

[0206] 8. Dataset Usage and Training Process

[0207] During model training, the training set is used for parameter updates, the validation set is used for hyperparameter tuning and early stopping criteria, and the test set is used for final performance evaluation. To ensure the reliability of the evaluation, the test set must not be used for parameter tuning during the training process.

[0208] A complete training process can be as follows:

[0209] First, read the training set samples and complete the batch loading of static parameters, temporal parameters and labels;

[0210] Secondly, forward propagation is used to obtain the predicted wear field, the predicted total wear amount, and the predicted regional wear amount;

[0211] Then calculate the total loss function and perform backpropagation;

[0212] Then, after each round of training, the validation set is used to calculate the validation loss and evaluation metrics;

[0213] Finally, the model is saved at the weight corresponding to the optimal validation performance.

[0214] Evaluation metrics should include at least: Total Wear Effect (MAE), Total Wear Effect Reduction (RMSE), MAE for each zone, Average Pixel Error of Wear Field, and Coefficient of Determination. For engineering deployments, it is recommended to output the left shoulder area error, middle area error, and right shoulder area error simultaneously to assess the model's ability to identify uneven wear.

[0215] 9. Technical effects of step S4

[0216] Step S4 achieves a systematic expression of the complex causes of tire wear through joint modeling of "static parameters + time-series operating conditions" and hierarchical output of "field-region-total". Furthermore, due to the introduction of physical consistency constraints, the model does not rely solely on statistical correlation but also conforms to the actual evolution of tire wear to a certain extent. Therefore, this step not only improves prediction accuracy but also enhances model interpretability and engineering reliability.

[0217] (v) Step S5: Target tire prediction and online correction

[0218] Step S5 is mainly used for the deployment of the present invention in actual engineering.

[0219] In offline prediction scenarios, the structural and material parameters of the target tire are first obtained, and then a preset operating condition time-series vector is constructed based on the target usage conditions. For example, for a new energy vehicle tire, the average speed distribution, load distribution, tire pressure distribution, and ambient temperature range within the next 20,000 kilometers can be set based on the user's operating condition profile, constructing an operating condition sequence of length 20. This sequence is then input into a trained model, outputting the predicted wear field of the target tire. The total wear and wear of each zone are further calculated from this wear field. If the prediction results show that the wear in a certain tire shoulder area is significantly higher than that in other areas, it can be used for subsequent structural optimization or user maintenance suggestions.

[0220] In online correction scenarios, such as Figure 5 As shown, during actual tire use, the online correction module can receive real-time signals, such as tire pressure from the tire pressure sensor, tire cavity temperature from the temperature sensor, road excitation characteristics from the acceleration sensor, and wheel speed information from the wheel speed sensor. The system can fuse these real-time signals with the original preset operating conditions to form an updated operating condition time-series vector, and then periodically call the model to re-predict the wear field.

[0221] For example, if the preset tire pressure is 240 kPa, but online monitoring shows that it has been consistently below 220 kPa, the tire pressure value in the corresponding interval of the time-series operating condition vector can be corrected, and the wear field output by the model may show a higher wear trend in the tire shoulder area. In this way, the present invention can provide users with dynamic wear prediction results that are closer to the actual service conditions.

[0222] IV. Specific Application Examples

[0223] To verify the effectiveness of the neural network-based tire wear prediction method described in this invention in predicting total wear, zoned wear, and wear distribution, the applicant constructed a tire sample database, completed finite element analysis, bench wear tests, groove depth measurement, three-dimensional contour scanning, and model training, and compared it with existing conventional methods. Experimental results show that this invention can significantly improve the accuracy of tire wear prediction, especially in the characterization of uneven wear and the identification of uneven wear patterns.

[0224] (I) Experimental Objective

[0225] The experimental objectives of this application example mainly include the following three aspects:

[0226] First, it verifies the accuracy of the method of the present invention in predicting the total wear of tires under given structural parameters, material parameters, and operating conditions.

[0227] Secondly, the method of the present invention is verified to predict the wear amount and wear distribution field of different regions of the tire tread.

[0228] Third, this study verifies whether the technical approach of "finite element wear response + measured calibration + physical consistency constraint + hierarchical output" can achieve better technical results compared to the traditional BP neural network single-value prediction method and the ordinary deep network method without physical consistency constraint.

[0229] (II) Experimental conditions and sample setup

[0230] 1. Test subjects

[0231] Passenger car radial tires were selected as the research object, and three tire specifications were set: 205 / 55R16; 215 / 55R17; 225 / 45R18.

[0232] Two tread pattern schemes and two tread compound formulations were designed for each specification, resulting in a total of 12 tire design combinations. Four prototype tires were prepared for each tire design combination, for a total of 48 prototype tires. Among them: tread pattern scheme A is a symmetrical tread pattern with equal pitch; tread pattern scheme B is an asymmetrical tread pattern with variable pitch; formulation scheme M1 is a conventional wear-resistant tread compound; and formulation scheme M2 is a modified tread compound that improves the wear factor while also taking into account wear resistance.

[0233] 2. Structural and material parameters

[0234] The following structural parameters were recorded for all sample tires: tread width Cross-sectional height-to-width ratio tire cord angle ; Belt layer angle ; Pattern groove depth Shoulder transition radius Pattern pitch distribution parameters .

[0235] The following material parameters were recorded for the tread compound: storage modulus Loss factor Shore hardness abrasion resistance coefficient Filler addition amount Glass transition temperature .

[0236] 3. Operating Condition Settings

[0237] Accelerated wear tests were conducted using an indoor tire wear testing bench. Each tire sample was assembled, inflated, and balanced according to standard procedures before testing. The following operating parameters were recorded during the test: inflation pressure. Vertical load driving speed Side slip angle Longitudinal slip ratio Ambient temperature Road surface adhesion coefficient .

[0238] To simulate actual service conditions, three types of operating condition spectra were set up for the test:

[0239] Operating condition spectrum G1: Low-speed start-stop type for urban roads;

[0240] Operating condition spectrum G2: Integrated type of urban expressway and ring road;

[0241] Operating condition spectrum G3: High-speed cruise and medium load type.

[0242] The test mileage was uniformly set at an equivalent cumulative mileage of 20,000 km, with each 2,000 km mark serving as a test node, forming a total of 10 mileage nodes.

[0243] 4. Methods for obtaining measured wear data

[0244] At each inspection point, the sample tire undergoes the following tests:

[0245] Five groove depth measuring points were set up on the left shoulder, the middle, and the right shoulder, for a total of 15 measuring points. The remaining groove depth was measured using a digital groove depth gauge.

[0246] The point cloud data of the tire tread surface was acquired using a laser 3D contour scanner and converted into a tire wear height distribution map.

[0247] Record the tire appearance to determine if there is obvious uneven wear, serrated wear, or localized abnormal wear.

[0248] Tire pressure, temperature, and wheel speed signals are collected simultaneously to construct online operating condition correction data.

[0249] (III) Dataset Construction and Method Implementation

[0250] 1. Number of data samples

[0251] A total of 432 valid wear records were obtained under 48 sample tires, 10 testing nodes, and 3 types of operating conditions. Among them, 12 invalid samples due to sensor frame loss and scanning distortion were removed, and 420 samples were finally included in the modeling samples.

[0252] The 420 samples were divided into:

[0253] Training set: 294 groups;

[0254] Validation set: 63 groups;

[0255] Test set: 63 groups.

[0256] 2. Construction of Finite Element Wear Field

[0257] Combination Figure 3For each tire sample, a finite element model of the tire was established, and contact analysis was performed under inflation, loading, rolling contact, lateral slip, and sliding conditions. Subsequently, based on the relationship between contact pressure distribution, friction work density, and wear increment, the wear response at each discrete location of the tire tread was calculated.

[0258] To facilitate comparison between samples of different specifications, the finite element analysis results are mapped to a unified normalized tread coordinate grid. The grid is divided into 32 meshes laterally and 64 meshes circumferentially, forming... The wear field matrix.

[0259] 3. Actual measurement calibration

[0260] The 3D contour scan results and data from 15 groove depth measurement points were used together as measured wear calibration data to correct the deviation of the initial finite element wear field, thus obtaining the supervised wear field label. This step corresponds to... Figure 3 The technical chain in the process is "finite element wear response - unified mesh mapping - measured calibration - supervised wear field labeling".

[0261] 4. Model Setup

[0262] according to Figure 4 The network structure shown is used to establish the model of this invention. It includes:

[0263] Static parameter branches: 3-layer fully connected network;

[0264] Timing-based branch: 2-layer LSTM;

[0265] Feature fusion layer: splicing and fusion;

[0266] Wear field decoding branch: 2-stage upsampling convolutional decoding;

[0267] Derived calculation layer: Calculates the total wear and the wear of each zone from the predicted wear field;

[0268] Physical consistency constraint layer: including mileage monotonic non-decreasing constraint and zone wear increment and cumulative friction work change trend consistency constraint.

[0269] Simultaneously set up two sets of comparative methods:

[0270] Comparative Example 1: Traditional BP Neural Network Method

[0271] Input structural parameters, material parameters, and average operating conditions; only the total wear amount will be output.

[0272] Comparative Example 2: Ordinary LSTM-FC Deep Network Method

[0273] Input structural parameters, material parameters, and operating condition timing. Output the total wear amount and the wear amount of three zones, but do not output the wear field and do not set physical consistency constraints.

[0274] (iv) Evaluation Indicators

[0275] To demonstrate the effectiveness of the technology, the following evaluation index is used: the average absolute error of total wear, denoted as... The root mean square error of total wear is denoted as . The coefficient of determination for total wear is denoted as . The average absolute error of the wear amount in each zone is recorded as follows: left shoulder Central and right shoulder The average pixel error of the wear field is denoted as... The average spatial error between the predicted and measured wear fields is used to characterize the difference between the predicted and measured wear fields. The accuracy rate for identifying uneven wear is used to evaluate the ability to distinguish between samples with obvious left-sided, right-sided, and central wear anomalies. The accuracy rate for uneven wear identification is defined as follows: when the wear amount in a certain region is significantly higher than in other regions, and this is consistent with the measured judgment result, it is considered correctly identified.

[0276] (V) Experimental Results and Data

[0277] 1. Model convergence status

[0278] During training, both training and validation losses stabilized around round 60, while the validation loss essentially stopped decreasing after round 95. The model reached its optimum at round 102. It is recommended to plot this process as follows: Figure 6 .from Figure 6 As can be seen, the model of this invention is stable during training and has no obvious overfitting, indicating that the supervised wear field label and physical consistency constraint adopted are beneficial to model convergence.

[0279] 2. Total Wear Prediction Results

[0280] The total wear prediction results for the three methods on the test set are shown in Table 1.

[0281] Table 1 Comparison of total wear prediction results using different methods

[0282]

[0283] As shown in Table 1, the method of the present invention is superior to the two comparative examples in predicting total wear. Compared with Comparative Example 1, A decrease of approximately 58.97%. A decrease of approximately 56.86%; compared to Comparative Example 2, A decrease of approximately 40.74%. The accuracy was reduced by approximately 39.30%. This demonstrates that the present invention does not simply replace the conventional network structure, but rather significantly improves the prediction accuracy of total wear by unifying wear field labels, hierarchical output, and physical consistency constraints. Figure 9The results can be plotted as a bar chart or scatter plot comparing the prediction errors of the three methods on the test set. The plot shows more intuitively that the predicted points of the method of this invention are closer to the ideal diagonal than the measured points.

[0284] 3. Prediction results of wear amount by zone

[0285] Table 2 shows the prediction results of wear on the left shoulder, middle and right shoulder areas using the three methods on the test set.

[0286] Table 2 Comparison of prediction errors for zoned wear amount using different methods

[0287]

[0288] Note: Comparative Example 1 only outputs the total wear amount and does not have the ability to predict the wear amount in different zones. Therefore, it cannot be used for uneven wear identification and regional wear diagnosis.

[0289] As shown in Table 2, the prediction errors of the method of the present invention in the left shoulder, middle, and right shoulder regions are significantly lower than those in Comparative Example 2. Specifically, the error is reduced by approximately 45.64% in the left shoulder, approximately 44.10% in the middle, and approximately 45.92% in the right shoulder. This result indicates that the technical approach of the present invention, which first predicts the wear field and then derives the wear amount for each region, can better reflect the wear differences in different areas of the tire. Figure 8 The error comparison of each method in the three partitions can be displayed intuitively in the form of a grouped bar chart. It can be seen from the figure that the present invention has a significant error advantage in all three partitions.

[0290] 4. Wear field prediction results

[0291] Three typical samples were randomly selected from the test set, and their wear fields were predicted and compared with the actual measurements. The results are shown in Table 3.

[0292] Table 3. Prediction accuracy of the wear field using the method of the present invention.

[0293]

[0294] As shown in Table 3, the present invention can correctly identify the area of ​​maximum wear on typical samples, and the average spatial error of the wear field is controlled within 0.123 mm. This indicates that the model can not only output numerical values, but also output spatial distribution results that are highly consistent with the actual wear morphology. Figure 7 The measured and predicted wear field thermograms for samples T-17, T-28, and T-43 are presented respectively. The comparison shows that the predicted results of this invention are highly consistent with the measured results in terms of the location, extent, and lateral offset trend of the high wear zone, especially in the accurate characterization of uneven wear on the tire shoulder.

[0295] 5. Uneven wear identification results

[0296] Twenty-one samples with obvious uneven wear trends were selected from the test set for uneven wear identification and evaluation. The results are shown in Table 4.

[0297] Table 4 Comparison of uneven wear identification results

[0298]

[0299] As shown in Table 4, the present invention has significant advantages in the task of identifying uneven wear. Comparative Example 1, because it only outputs the total wear amount, cannot identify uneven wear; Comparative Example 2, although it can output the partition amount, will show bias in the judgment of uneven wear areas in some complex samples due to the lack of wear field supervision labels and physical consistency constraints; while the method of the present invention significantly improves the accuracy of uneven wear identification through the core output of the wear field and hierarchical derivation.

[0300] (vi) Typical Sample Analysis

[0301] The following is combined Figure 7 Two typical samples are described in detail.

[0302] 1. Sample T-17: Right tire shoulder wear sample

[0303] Sample T-17 is a 215 / 55R17 tire, using tread pattern scheme B and compound scheme M2, with a performance profile of G1. During testing, this sample was subjected to prolonged periods of low tire pressure and high cornering lateral slip. The measured results showed that after 20,000 km, the remaining groove depth in the right shoulder area was significantly less than that in the left shoulder and center areas, indicating significant right-side wear.

[0304] The model of this invention predicts a significant high-wear zone in the right shoulder region of the sample, which is highly consistent with the measured contour scan results. While Comparative Example 2 also indicates greater wear on the right side, the extent of the high-wear zone deviates from the measured values, and the wear in the central region is overestimated. This demonstrates that the unified wear field label and physical consistency constraints introduced in this invention help in learning the shoulder wear mechanism under the coupling of tire pressure reduction and side-clamp conditions.

[0305] 2. Sample T-28: Sample with uneven wear in the middle

[0306] Sample T-28 is a 225 / 45R18 tire, using tread pattern scheme A and compound scheme M1, with a performance spectrum of G2. This sample was tested under relatively high inflation pressure and at medium to high speeds. The test results show that wear is mainly concentrated in the right-center area, classifying it as a typical sample with high wear in the center.

[0307] The predicted wear field of this invention not only shows the high wear trend in the middle, but also accurately reflects the morphological features of the high wear region shifting slightly to the right. In contrast, although Comparative Example 2 can determine that the wear is high in the middle, it cannot accurately give the lateral shift position. This result shows that this invention is more suitable for handling complex wear problems with local morphological differences than ordinary networks that only output partition values.

[0308] The above application examples and experimental data demonstrate that the present invention has at least the following technical effects:

[0309] First, this invention can significantly improve the accuracy of tire total wear prediction. Compared with traditional BP neural networks, the average absolute error of total wear is reduced by approximately 58.97%; compared with ordinary deep time series networks, the average absolute error of total wear is reduced by approximately 40.74%.

[0310] Secondly, the present invention can accurately predict the wear amount of different areas of the tire tread, and achieves a significantly lower error level than the comparative example in the three areas of left shoulder, middle and right shoulder, indicating that the present invention has better technical effect in characterizing uneven wear.

[0311] Third, this invention uses the wear field as the core output, which can directly provide the wear distribution in the tread space and correctly identify the maximum wear area on typical uneven wear samples. The average spatial error of the wear field is low, making it suitable for tire structure optimization, uneven wear warning and maintenance strategy formulation.

[0312] Fourth, by combining physical consistency constraints and online operating condition correction, the prediction results of this invention not only have high accuracy, but also better conform to the evolution law of tire wear with mileage growth, and have good stability and interpretability in engineering applications.

[0313] Therefore, the above experimental results fully demonstrate that the present invention has made significant technological progress compared with the prior art in terms of prediction accuracy, regional wear identification ability, wear distribution expression ability and online correction ability.

[0314] 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.

[0315] 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.

[0316] 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.

[0317] 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.

[0318] 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.

[0319] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0320] 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.

[0321] 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 tire wear based on neural networks, characterized in that, Includes the following steps: S1. Construct a tire sample set and obtain the structural parameter vector, material parameter vector, operating condition time sequence vector, finite element wear response data, and measured wear calibration data for each tire sample. S2. Map the finite element wear response data to a unified grid in the normalized tread coordinate system to form an initial wear field, and correct the deviation of the initial wear field based on the measured wear calibration data to form a supervised wear field label. S3. Calculate the total wear amount, multiple tread zone wear amount vectors, and physical consistency feature vectors based on the supervised wear field labels, and construct training samples; S4. Construct a neural network model and train it. The neural network model includes a static parameter feature extraction branch, a time-series operating condition feature extraction branch, a wear field decoding branch, and a physical consistency constraint branch. S5. Obtain the structural parameter vector, material parameter vector, and preset operating condition time sequence vector of the target tire, input them into the trained neural network model, first output the predicted wear field of the target tire, and then calculate the total wear amount and wear amount of each zone of the target tire from the predicted wear field to obtain the wear prediction result of the target tire under the preset operating conditions.

2. The method according to claim 1, characterized in that: In step S1, the structural parameter vector includes at least one of the following: tread width, cross-sectional height-to-width ratio, carcass cord angle, belt layer angle, tread groove depth, tread block pitch distribution parameters, and shoulder transition radius. And / or, In step S1, the operating condition parameters in the operating condition time sequence vector include at least one of the following parameters: inflation pressure, vertical load, sideslip angle, longitudinal slip ratio, driving speed, road surface adhesion coefficient, ambient temperature, cumulative driving mileage, and road surface roughness.

3. The method according to claim 1, characterized in that: In step S2, the normalized tread coordinate system consists of lateral normalized coordinates and circumferential normalized coordinates. The lateral normalized coordinates are obtained by the ratio of the actual coordinates in the tread width direction to the tread width, and the circumferential normalized coordinates are obtained by the ratio of the circumferential arc length coordinates to the effective circumferential length of the tread. The actual coordinates in the tread width direction are the actual position coordinates in the tread width direction, the tread width is the effective width of the tire tread, the circumferential arc length coordinates are the arc length position coordinates of the tire tread after it is unfolded circumferentially, and the effective circumferential length of the tread is the effective unfolded length of the tire tread circumferentially. And / or, In step S2, the deviation correction is achieved using a calibration function between the finite element wear field and the measured wear data. The corrected monitoring wear field label is obtained by inputting the initial wear field and the measured wear calibration data into the calibration function. The calibration function is used to correct the deviation between the finite element results and the measured results.

4. The tire wear prediction method based on neural networks according to claim 1, characterized in that: In step S3, the physical consistency feature vector includes at least one of the following: cumulative friction work, cumulative slip work, and ground pressure distribution evolution features.

5. The tire wear prediction method based on neural networks according to claim 1, characterized in that: In step S4, the static parameter feature extraction branch is used to extract the static features of the structural parameter vector and the material parameter vector; the time-series working condition feature extraction branch is used to extract the evolution features of the working condition time-series vector; the wear field decoding branch is used to output the predicted wear field and derive the predicted total wear amount and the predicted partition wear amount vector based on the predicted wear field; and the physical consistency constraint branch is used to constrain the correspondence between the prediction result and the physical consistency feature vector. And / or, In step S4, a comprehensive loss function is used when training the neural network model. The comprehensive loss function is composed of a weighted sum of the wear field error term between the predicted wear field and the supervised wear field label, the total amount error term between the predicted total wear amount and the total wear amount, the partition error term between the predicted partition wear amount vector and the partition wear amount vector, and the physical consistency constraint error term. Wherein, the wear field error term is used to characterize the difference between the predicted wear field and the supervised wear field label, the total error term is used to characterize the difference between the predicted total wear amount and the total wear amount, the partition error term is used to characterize the difference between the predicted partition wear amount vector and the partition wear amount vector, and the physical consistency constraint error term is used to characterize the deviation between the prediction result and the physical consistency constraint. The weight coefficients corresponding to each error term are all real numbers greater than 0. The physical consistency constraint error term includes at least one of the following constraints: Predicted wear is monotonically non-decreasing with increasing cumulative mileage; The increase in tread wear in different zones follows the same trend as the increase in the cumulative friction work or cumulative slip work of the corresponding zones. The trend of the tread zone wear increment is consistent with the evolution characteristics of the ground pressure distribution; Among them, monotonic non-decreasing means that the predicted wear value of subsequent mileage nodes is not less than the predicted wear value of the previous mileage node.

6. The tire wear prediction method based on neural networks according to claim 1, characterized in that: In step S4, the static parameter feature extraction branch is a multilayer fully connected network, the temporal condition feature extraction branch is any one of a long short-term memory network, a gated recurrent unit, or a temporal convolutional network, and the wear field decoding branch is any one of a deconvolutional network, an upsampling convolutional network, or a fully connected mapping network.

7. The tire wear prediction method based on neural networks according to claim 1, characterized in that: Step S5 is followed by an online correction step: real-time acquisition of at least one of the tire temperature signal, tire pressure signal, acceleration signal, and wheel speed signal; extraction of incremental features of actual operating conditions; updating of the preset operating condition time-series vector; and inputting the updated time-series vector back into the trained neural network model to obtain dynamically updated wear field, total wear amount, and zone wear amount.

8. A tire wear prediction system based on neural networks, characterized in that, include: The data acquisition module is used to execute step S1; The tag building module is used to perform step S2; The sample construction module is used to execute step S3; The model training module is used to perform step S4; The prediction output module is used to execute step S5.

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 steps of 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 steps of the method described in any one of claims 1-7.

Images