A parameterized model-based wheel hub NVH performance prediction method

Abstract

The application provides a hub NVH performance prediction method based on a parameterized model, and relates to the cross field of computer aided engineering and artificial intelligence. In view of the problems of long time consumption of traditional CAE simulation analysis and low extraction efficiency of complex geometric characteristics of the hub, the method firstly constructs a 3D geometric parameter standardization description system of the hub and defines a standard data dictionary containing continuous variable characteristics and discrete topological characteristics, extracts corresponding characteristics by combining geometric topological analysis and machine vision dimension reduction recognition, and fuses and constructs a joint feature vector; then, a machine learning prediction model is trained relying on historical hub design data and corresponding NVH performance evaluation data, and the joint feature vector of the hub to be tested can output the NVH performance prediction result. The application can realize efficient and rapid evaluation of the hub NVH performance, significantly shorten the design iteration cycle, and is suitable for batch design verification scene of the NVH performance of the automobile hub.

Description

Technical Field

[0001] This invention relates to the field of computer-aided engineering and artificial intelligence, and in particular to a method for predicting the NVH performance of wheel hubs based on a parametric model. Background Technology

[0002] As a core safety and load-bearing component of the automotive chassis, the NVH performance (such as natural frequency and dynamic / static stiffness) directly affects the ride comfort and safety of the entire vehicle. In existing industrial software development and wheel design processes, traditional CAE simulation software such as Abaqus and ANSYS are typically used for finite element analysis. However, existing technologies have the following significant technical shortcomings: First, there is the problem of low analysis efficiency. Traditional CAE simulation requires a tedious process of geometry cleaning, mesh generation, boundary condition setting, and solution calculation. Each analysis is time-consuming and cannot meet the rapid iterative design requirements of the current automotive industry.

[0003] Second, current direct prediction methods suffer from difficulties in feature extraction and computational redundancy. Direct prediction methods based on 3D deep learning, which have emerged in recent years, consume enormous computational resources when processing industrial CAD models with complex topological structures. Furthermore, relying solely on 3D models makes it difficult to efficiently and robustly extract discrete topological features such as the number of wheel spokes and specific types of bolt holes.

[0004] Third, the prediction results lack engineering guidance. Most existing AI prediction models are black-box models, only outputting the final NVH performance values, unable to tell designers which geometric features caused the prediction results, let alone provide reverse parameter optimization suggestions.

[0005] Therefore, it is necessary to propose a wheel hub NVH performance prediction method based on a parametric model to solve or at least alleviate the above-mentioned defects. Summary of the Invention

[0006] The main objective of this invention is to provide a wheel hub NVH performance prediction method based on a parametric model, so as to solve the problems of low analysis efficiency, difficulty in feature extraction and computational redundancy in the existing technology.

[0007] To achieve the above objectives, this invention provides a method for predicting wheel hub NVH performance based on a parametric model, comprising the following steps: S1, acquire the three-dimensional geometric model data of the wheel hub under test and the corresponding operating condition parameters; S2, extract features from the three-dimensional geometric model data to obtain continuous variable features and discrete topological features, and fuse the continuous variable features, discrete topological features and the operating condition parameters to construct a joint feature vector; S3, acquire historical wheel hub data and corresponding historical NVH performance evaluation data, construct a training set based on the historical wheel hub data and the historical NVH performance evaluation data, and train a pre-constructed NVH performance prediction model based on the training set; S4. Input the joint feature vector into the trained NVH performance prediction model and output the NVH performance prediction result of the wheel hub to be tested.

[0008] Preferably, the operating condition parameters in step S1 include material properties and boundary constraints; wherein, when the boundary constraints are set to unconstrained full degrees of freedom, the NVH performance prediction result of the wheel hub under test is the natural frequency of the free mode of the corresponding wheel hub. When the boundary constraint condition is set to a six-degree-of-freedom full constraint on the bolt mounting hole, the NVH performance prediction result of the wheel hub under test is the natural frequency of the constraint mode of the corresponding wheel hub.

[0009] Preferably, step S2 specifically includes the following steps: S21, call the three-dimensional geometry kernel to parse the three-dimensional geometric model data of the wheel hub to be tested, and extract continuous variable features through geometric calculation; S22, the three-dimensional geometric model data of the wheel hub to be tested is projected in two dimensions or sliced ​​along a preset axis to generate a binary image, and discrete topological features are identified and calculated using image morphology and contour extraction algorithms. S23, standardize the continuous variable features, and fuse the standardized continuous variable features and discrete topological features with preset operating condition parameters to generate a joint feature vector.

[0010] Preferably, step S3 specifically includes the following steps: S31, construct a multi-source dataset of historical wheel hubs; wherein, the multi-source dataset includes source domain features composed of CAD simulation features of historical wheel hubs and target domain features composed of real bench test features; S32, map the source domain features and the target domain features to the regenerating kernel Hilbert space, and use the maximum mean difference algorithm to calculate the distribution distance between them; S33, construct a joint loss function that includes prediction task loss and distribution alignment loss by combining the distribution distance, and train a pre-constructed NVH performance prediction model based on the joint loss function.

[0011] Preferably, step S4 specifically includes the following steps: S41, decouple the joint feature vector and route it; wherein, the discrete topological features are input to the gradient boosting tree branch, and the continuous variable features are input to the multilayer perceptron branch; S42, the discrete topological features are input into the pre-trained tree model, the leaf node landing index of the sample in all decision trees is extracted, the leaf node landing index is one-hot encoded and passed through the embedding layer to be transformed into a low-dimensional tree representation vector. S43, Construct a fully connected multilayer perceptron containing multiple hidden layers, process the continuous variable features, and output a deep physical representation vector; S44, the low-dimensional tree representation vector and the deep physical representation vector are spliced ​​and fused across modes. After feature cross-calculation is completed by a fully connected layer, the NVH performance prediction result is mapped and output.

[0012] Preferably, the continuous variable features include rim thickness and spoke tilt angle, and the discrete topological features include the number of bolt holes and the number of spokes.

[0013] Preferably, the extraction of the rim thickness includes: identifying the inner and outer cylindrical surface models of the hub, extracting the outer and inner cylindrical surfaces, and calculating the rim thickness using the following formula: in, For the rim thickness, For sampling points on the inner surface, For point Along the local normal vector on the corresponding outer cylindrical surface The projection point on the surface, The inner cylindrical surface; The extraction of the spoke tilt angle includes: extracting the normal vector of the spoke center plane and the principal axis vector of the hub rotation, and calculating the spoke tilt angle by the vector dot product. : in, The spoke tilt angle, Let be the normal vector of the center plane of the wheel spokes. The principal axis vector of hub rotation. This represents the inverse cosine function.

[0014] Preferably, the distribution alignment loss is calculated using the following formula: in, Represents the distribution alignment loss. Represents the first term of the source domain Individual sample feature data, This represents the total number of samples in the source domain. The first term of the target domain Individual sample feature data, This represents the total number of samples in the target domain. This represents the feature mapping function that maps the original features to a high-dimensional kernel space.

[0015] Preferably, the joint loss function is expressed as: in, Denotes the joint loss function. This indicates the source domain's predicted NVH performance results. Indicates the true label of the source domain. This indicates the predicted NVH performance results for the target domain. Indicates the true label of the target domain. This represents the loss function for the prediction task. This is a dynamically adaptive penalty coefficient.

[0016] Preferably, the NVH performance prediction result is expressed as: in, This indicates the NVH performance prediction results. Represented as the weight matrix of the output layer, Represented as a low-dimensional tree representation vector, Represented as a deep physical representation vector. It is represented as the bias vector of the output layer.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a wheel hub NVH performance prediction method based on a parametric model. By constructing a standardized description system for wheel hub 3D geometric parameters, it unifies the classification rules and data specifications for wheel hub geometric features, providing a unified mapping benchmark for feature extraction from wheel hub models with different structures and sources, ensuring the consistency and universality of feature data. It employs a combination of geometric topology analysis and machine vision dimensionality reduction recognition to extract corresponding features and fuse them to construct a joint feature vector. This efficiently completes the structured feature transformation of complex 3D wheel hub models, balancing the extraction accuracy of continuous physical parameters with the recognition efficiency of discrete topological structures, solving the problems of high difficulty and low efficiency in extracting complex wheel hub geometric features. Based on historical design data and corresponding NVH performance evaluation data from enterprises, a machine learning prediction model is trained to establish an end-to-end mapping relationship between geometric parameters and NVH performance, replacing the cumbersome geometric cleaning, mesh generation, and iterative solution process of traditional CAE simulation. The wheel hub under test only needs to complete feature extraction to quickly output NVH performance prediction results through the model, significantly shortening the evaluation time of wheel hub NVH performance, effectively improving design iteration efficiency, and reducing the consumption of computing resources and time costs in the R&D process. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the overall method in one embodiment of the present invention.

[0020] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 scope of protection of the present invention.

[0023] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0024] Please see the appendix Figure 1 The present invention provides a method for predicting the NVH performance of a wheel hub based on a parametric model, comprising the following steps: S1: Acquire the 3D geometric model data of the wheel hub under test and the corresponding operating condition parameters; read the original 3D geometric model files of the wheel hub in industry standard formats such as STEP, IGES, and x_t through the file interface, and simultaneously collect the operating condition parameters matching the wheel hub under test from the design document or operating condition configuration module to complete the basic data entry and format verification of the prediction process; thus, it fully covers the geometric structural elements and physical operating condition elements that affect the NVH performance of the wheel hub, providing complete and accurate raw data input for subsequent feature extraction and model inference, and ensuring that the final prediction results correspond to the design conditions and actual application scenarios.

[0025] In a preferred embodiment of the present invention, the operating condition parameters in step S1 include material properties and boundary constraints. Specifically, the material properties include core physical parameters such as the elastic modulus, Poisson's ratio, and density of the wheel hub material. The boundary constraints define the force constraint state of the wheel hub. By incorporating these two types of parameters into the prediction input system, the model can switch the corresponding calculation logic based on the input constraint parameters. When the boundary constraints are set to unconstrained full degrees of freedom, the model simulates the test state of the wheel hub's free suspension, and the output NVH performance prediction result for the wheel hub is the natural frequency of the corresponding free mode. When the boundary constraints are set to fully constrained six degrees of freedom for the bolt mounting holes, the model simulates the test state of the wheel hub's fixed mounting, and the output NVH performance prediction result for the wheel hub is the natural frequency of the corresponding constrained mode. This embodiment allows a single prediction model to simultaneously adapt to two typical NVH evaluation scenarios: free mode and constrained mode. It eliminates the need for separate modeling and training for different mode types, improving the model's reusability and applicability, and comprehensively covering the conventional testing and evaluation conditions for wheel hub NVH performance.

[0026] S2, feature extraction is performed on the three-dimensional geometric model data to obtain continuous variable features and discrete topological features. The continuous variable features, discrete topological features, and operating condition parameters are fused to construct a joint feature vector. A dual-path feature extraction mechanism is adopted for the different numerical characteristics of the wheel hub geometric parameters, processing the continuously changing size parameters and the discretely classified topological structure parameters respectively, transforming the unstructured three-dimensional model data into structured numerical feature data. Then, the two types of geometric features and the numericalized operating condition parameters are concatenated in a fixed order along the feature dimension to form a standardized joint feature vector with fixed dimensions, clear semantics, and direct input to the machine learning model. This achieves the digital and structured representation of the complex three-dimensional wheel hub model, takes into account the adaptability of different types of geometric feature extraction, eliminates the input inconsistency problem caused by the difference in the original three-dimensional model format, and provides input data with a unified format and complete information for subsequent prediction models.

[0027] S3. Acquire historical wheel hub data and corresponding historical NVH performance evaluation data, and construct a training set based on the historical wheel hub data and the historical NVH performance evaluation data. Train a pre-constructed NVH performance prediction model based on the training set. By connecting to the enterprise's product data management system, simulation database, and test management system, batch collect wheel hub 3D design model files from past R&D and mass production stages, as well as NVH performance simulation and test result data for each wheel hub. Simultaneously, invalid samples with missing data or abnormal operating conditions are removed to establish a sample dataset with one-to-one correspondence between design parameters and performance labels. This provides sufficient sample support for the training of the machine learning model. Relying on the historical data accumulated by the enterprise, a sample space covering multiple structures, sizes, and operating conditions is constructed to ensure the diversity and engineering representativeness of the training samples, providing a data foundation for the model's generalization ability.

[0028] S4, input the joint feature vector into the trained NVH performance prediction model, and output the NVH performance prediction result of the wheel hub under test; input the joint feature vector corresponding to the wheel hub under test into the trained prediction model, and the model performs forward calculation through the trained parameters, directly outputting the corresponding NVH performance index value; in this way, there is no need to perform the cumbersome process of geometry cleaning, mesh generation, and iterative solution of traditional CAE simulation, and the performance result can be obtained only through the forward calculation of the model, which greatly shortens the evaluation time of wheel hub NVH performance and improves the design iteration efficiency.

[0029] In a preferred embodiment of the present invention, step S2 specifically includes the following steps: S21, the three-dimensional geometry kernel is invoked to parse the three-dimensional geometric model data of the wheel hub under test, and continuous variable features are extracted through geometric calculations; by calling the application interface of the three-dimensional geometry kernel at the underlying level, the boundary representation topology of the three-dimensional model is read and parsed, the geometric elements such as faces and edges in the model are traversed and topological relationships are established, and various dimensional parameters are accurately measured and calculated based on spatial geometric operation rules to extract the precise values ​​of various continuous variable features; thus, physical quantity calculations are performed directly based on the original topology of the three-dimensional model, and the extracted continuous variables have clear physical dimensions and engineering meanings. The numerical accuracy is consistent with the modeling accuracy of the original three-dimensional model, and there is no accuracy loss caused by dimensionality reduction processing, which can provide accurate continuous dimensional parameter inputs for subsequent performance prediction.

[0030] Specifically, the continuous variable features include rim thickness, spoke inclination angle, and offset, which directly determine the structural stiffness and mass distribution characteristics of the wheel hub body. This application uses the extraction of rim thickness and spoke inclination angle as examples for illustration below: The extraction of the rim thickness includes: identifying the inner and outer cylindrical surface models of the hub, extracting the outer and inner cylindrical surfaces, and calculating the rim thickness using the following formula: in, For the rim thickness, For sampling points on the inner surface, For point Along the local normal vector on the corresponding outer cylindrical surface The projection point on the surface, The inner cylindrical surface is used; the minimum Euclidean distance between the sampling points on the inner surface and the corresponding outer cylindrical surface is calculated using the formula. The extraction of the spoke tilt angle includes: extracting the normal vector of the spoke center plane and the principal axis vector of the hub rotation, and calculating the spoke tilt angle by the vector dot product. : in, The spoke tilt angle, Let be the normal vector of the center plane of the wheel spokes. The principal axis vector of hub rotation. This represents the inverse cosine function; it is obtained by calculating the inverse cosine of the angle between two vectors through the dot product of vectors. Through the geometric analysis described above, the system combines the extracted physical features such as rim thickness, spoke inclination angle, and offset into a continuous variable feature vector. ; This represents the number of features of a continuous variable, while each in the vector set... This corresponds to a feature, for example Represents the rim thickness. This represents the spoke tilt angle.

[0031] S22, the three-dimensional geometric model data of the wheel hub to be tested is projected in two dimensions or sliced ​​along a preset axis to generate a binary image. Discrete topological features are identified and calculated using image morphology and contour extraction algorithms. Taking the rotation center axis of the wheel hub as a preset reference axis, the boundary contour of the three-dimensional model is orthogonally projected onto a plane perpendicular to this axis, or the cross-sectional contour at the corresponding position is extracted. After rasterization rendering and threshold segmentation, a binary image is generated that distinguishes between solid areas and hollow areas. Then, the image contour is optimized and projection noise is eliminated through dilation and erosion operations of image morphology. Independent closed contours in the image are extracted through connected component analysis. After the contours are classified and identified, the number is counted to obtain the relevant parameters of discrete topological features. In this way, by using dimensionality reduction projection, the complex topological search problem in three-dimensional space is transformed into a contour recognition problem in two-dimensional images, which greatly reduces the amount of search and statistical calculation for discrete holes and spoke structures. It can efficiently complete the counting and classification of discrete structures such as spokes and bolt holes, while ensuring the extraction efficiency and recognition accuracy of discrete topological features.

[0032] In detail, the discrete topological features include the number of bolt holes, the number of spokes, etc., which determine the overall structural form of the hub and the load transmission path. This application uses the extraction of the number of bolt holes and the number of spokes as an example for illustration below: The number of bolt holes and the number of spokes are obtained through the following steps: The rotation center axis of the wheel hub is set as The axis directs the boundary contour of the three-dimensional geometric model of the wheel hub under test to... The plane is orthogonally projected, and the projected model is orthogonally rasterized and rendered to generate a two-dimensional binary image matrix. The orthogonal projection transformation matrix is ​​used to complete the perspective-free projection of the three-dimensional boundary contour to the two-dimensional plane, ensuring that the projected planar structure corresponds one-to-one with the actual topological layout of the wheel hub. Traversing a 2D binarized image matrix using computer vision libraries (such as OpenCV) Connected components, extract closed contour sets ; For any contour Calculate its zeroth-order spatial moment (i.e., area property) ): And calculate its corresponding normalized central moments to derive seven translation-, scaling-, and rotation-invariant moments. Invariant moments, The feature vectors of each contour are obtained. ; Next, the feature vectors of each contour are matched with the preset bolt hole template and spoke cutout template for similarity. Based on the matching and classification results, the discrete topological features of the number of bolt holes and the number of spokes are counted and aggregated to construct discrete topological feature vectors. ; This represents the number of discrete topological features, while each of the vector sets... This corresponds to a discrete topological feature, for example Represents the number of bolt holes. This represents the number of spokes.

[0033] S23, the continuous variable features are standardized. The standardized continuous variable features and discrete topological features are fused with preset operating condition parameters to generate a joint feature vector. Based on the statistical parameters of each continuous variable in the historical dataset, the currently extracted continuous variable features are numerically standardized to eliminate differences in physical dimensions and numerical magnitudes between different parameters, making each parameter fall within a similar numerical range. The processed continuous variable features, discrete topological features, and numerically encoded operating condition parameters are then concatenated along the feature dimension in a preset fixed order to obtain a joint feature vector with unified dimensions and standardized numerical distribution. The standardization process unifies the numerical scale of different parameters, avoiding adverse interference from differences in numerical magnitudes on the gradient descent process during model training, ensuring the stability of the model training and inference process. At the same time, the feature fusion forms a complete input vector, fully meeting the input format requirements of the machine learning model.

[0034] Among these, the global mean of this type of continuous variable is extracted from the historical database. with standard deviation For the continuous variable vector of the current wheel hub under test conduct standardization: in, Indicates the first Features of continuous variables Represents the standardized first Features of continuous variables; Characteristics of standardized continuous variables Discrete topological features with clear classification and counting significance Direct concatenation along the feature channel dimension generates a high-dimensional joint feature vector that can be directly processed by machine learning models. This approach abandons the computationally expensive pure 3D mesh or point cloud deep learning schemes and achieves a fast, robust feature digitization process with extremely high signal-to-noise ratio for complex industrial wheel hub models by integrating 3D B-Rep calculation with 2D invariant moment visual recognition.

[0035] In a preferred embodiment of the present invention, step S3 specifically includes the following steps: S31, construct a multi-source dataset of historical wheel hubs; wherein, the multi-source dataset includes source domain features composed of CAD simulation features of historical wheel hubs and target domain features composed of real bench test features; the feature set corresponding to the historical wheel hub CAD simulation data is divided into source domain features, and the feature set corresponding to the real bench test data is divided into target domain features, and the two together form a multi-source training dataset containing two types of data; wherein, the source domain data comes from simulation calculations and has the characteristics of large sample size and low acquisition cost, while the target domain data comes from actual physical tests and has the characteristics of high engineering authenticity; in order to integrate data resources from two different sources, make full use of the company's existing simulation data accumulation and test data accumulation, and take into account both data scale and data authenticity, so as to provide a complete multi-source data foundation for subsequent feature distribution alignment and model transfer training.

[0036] Here, a source domain dataset consisting of a large number of historical CAE simulation features is defined. and a target domain dataset consisting of a small number of real bench test features. ,in For joint feature vectors, For true NVH performance labels, and Let be the sample size of each, and satisfy . .

[0037] S32, the source domain features and the target domain features are mapped to the reproducing kernel Hilbert space, and the maximum mean difference algorithm is used to calculate their distribution distance. Through a nonlinear feature mapping function, the source domain features and the target domain features are simultaneously mapped to the high-dimensional reproducing kernel Hilbert space, transforming the nonlinear distribution difference in the original feature space into a measurable distribution difference in the high-dimensional space. Then, the maximum mean difference algorithm is used to calculate the squared distance between the mean vectors of the feature samples of the two domains in the high-dimensional space, thereby quantifying the degree of feature distribution difference between the two types of data and obtaining the corresponding distribution distance value. By using high-dimensional space mapping to achieve accurate quantitative measurement of nonlinear distribution differences, the systematic distribution deviation between simulation data and measured data can be accurately captured, providing a clear quantitative basis for subsequent distribution alignment optimization.

[0038] S33, a joint loss function comprising prediction task loss and distribution alignment loss is constructed based on the distribution distance. A pre-built NVH performance prediction model is trained based on this joint loss function. The prediction task loss of the source domain samples, the prediction task loss of the target domain samples, and the distribution alignment loss representing the distribution difference are combined to form the overall optimization objective of the model training, i.e., the joint loss function. The prediction task loss measures the error between the model's predicted value and the true label value, while the distribution alignment loss measures the difference in feature distribution between the source and target domains. During model training, an optimization algorithm iterates backpropagation with the goal of minimizing the value of the joint loss function, simultaneously optimizing the model's prediction accuracy and feature distribution consistency. This allows the model to learn the NVH performance prediction rules while gradually reducing the feature distribution difference between simulation data and measured data, achieving knowledge transfer from simulation data to measured data, and effectively improving the model's prediction accuracy and generalization ability in actual engineering scenarios.

[0039] The distribution alignment loss is calculated using the following formula: in, Represents the distribution alignment loss. Represents the first term of the source domain Individual sample feature data, This represents the total number of samples in the source domain. The first term of the target domain Individual sample feature data, This represents the total number of samples in the target domain. This represents the feature mapping function that maps the original features to a high-dimensional kernel space. This is represented as the source domain feature dataset. This is represented as the target domain feature dataset. Let denote the 2-norm in Hilbert space.

[0040] The joint loss function is expressed as: in, Denotes the joint loss function. This indicates the source domain's predicted NVH performance results. Indicates the true label of the source domain. This indicates the predicted NVH performance results for the target domain. Indicates the true label of the target domain. This represents the loss function for the prediction task. The dynamic adaptive penalty coefficient enables forced alignment between virtual simulation and real-world physical test data in the latent space. This allows the model to use extremely scarce bench data as an anchor point during inference to correct systematic biases caused by massive amounts of simulation data, greatly improving the model's generalization ability in the real physical world.

[0041] In a preferred embodiment of the present invention, step S4 specifically includes the following steps: S41, the joint feature vector is decoupled and routed; wherein, the discrete topological features are input to the gradient boosting tree branch, and the continuous variable features are input to the multilayer perceptron branch; according to the numerical attributes and types of the features, the joint feature vector is split into two groups: discrete topological features and continuous variable features, which are respectively sent to the two corresponding parallel computing branches in the model; wherein the discrete topological features are input to the gradient boosting tree branch for processing, and the continuous variable features are input to the multilayer perceptron branch for processing; for the two types of features with significantly different attributes, a separate processing mode is adopted to avoid the step change of discrete topology from interfering with the smooth fitting process of continuous physical quantities, so that the two computing branches can perform optimal fitting for the characteristics of the corresponding features, give full play to the adaptation advantages of different algorithm architectures, and improve the overall fitting ability and prediction accuracy of the model.

[0042] S42, the discrete topological features are input into the pre-trained tree model. The leaf node landing indexes of the sample in all decision trees are extracted. Each leaf node landing index is one-hot encoded and passed through an embedding layer to be transformed into a low-dimensional tree representation vector. Gradient boosting trees are used as the basic architecture for the tree model branches. After the discrete topological features are input into the tree model, the sample is judged layer by layer according to the splitting rules of each decision tree, and finally falls on a leaf node of each tree. The final leaf node position index of the sample in all decision trees is recorded. After converting the position indices of all leaf nodes into high-dimensional sparse one-hot encoded vectors, they are input into the embedding layer for linear mapping to be transformed into dense low-dimensional feature vectors, i.e., low-dimensional tree representation vectors. : ;in, Represented as a feature embedding function, This is represented as a tree decision path extraction function.

[0043] S43, Construct a fully connected multilayer perceptron with multiple hidden layers to process the continuous variable features and output a deep physical representation vector; Construct a fully connected multilayer perceptron network with multiple hidden layers, where each hidden layer undergoes a nonlinear transformation through an activation function; After inputting the continuous variable features into the multilayer perceptron, feature abstraction and extraction are performed step by step through multiple nonlinear transformations to fit the complex mapping relationship between continuous geometric parameters and NVH performance, ultimately outputting a high-dimensional deep physical representation vector; Each layer applies an activation function with leakage correction linear units and batch normalization to smoothly fit the continuous differential relationship between physical quantities such as rim thickness and offset and modal frequencies, outputting a deep physical representation vector. : in, Represented as a non-linear activation function, Represented as the first The weight matrix of the hidden layer, Represented as the first Bias vector of the hidden layer.

[0044] S44, the low-dimensional tree representation vector and the deep physical representation vector are cross-modal concatenated and fused. After feature cross-calculation is completed by a fully connected layer, the NVH performance prediction result is mapped and output. The low-dimensional tree representation vector output by the tree model branch and the deep physical representation vector output by the multilayer perceptron branch are concatenated in the feature channel dimension to obtain a fused feature vector. The fused feature vector is input into a fully connected network for feature cross-calculation to learn the coupling relationship between discrete topological features and continuous variable features. Finally, it is mapped to the corresponding NVH performance index value through a linear output layer to form a complete NVH performance prediction result.

[0045] Furthermore, the NVH performance prediction result is expressed as follows: in, This indicates the NVH performance prediction results. Represented as the weight matrix of the output layer, Represented as a low-dimensional tree representation vector, Represented as a deep physical representation vector. It is represented as the bias vector of the output layer.

[0046] The final output is = , This is expressed as the predicted static stiffness value of the wheel hub under test. This is expressed as the predicted value of dynamic stiffness at a specific evaluation frequency. These are expressed as the predicted natural frequencies from the 1st to the 10th order.

[0047] Furthermore, regarding the natural frequencies, the multimodal prediction model can output the free modal frequencies of the wheel hub in an unconstrained state and the constrained modal frequencies in a simulated fixed state (such as bolt hole constraints) by receiving different operating condition parameters.

[0048] It should be noted that the static stiffness in the prediction results is characterized as the normal displacement deformation of the wheel hub under test when a unit static load is applied at a specific force-bearing node such as the wheel flange; the dynamic stiffness is characterized as the frequency domain characteristic value of the ratio of dynamic force to dynamic displacement at the wheel hub response point under a specific excitation frequency (such as the road noise sensitive frequency band).

[0049] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for predicting the NVH performance of wheel hubs based on a parametric model, characterized in that, Includes the following steps: S1, acquire the three-dimensional geometric model data of the wheel hub under test and the corresponding operating condition parameters; S2, extract features from the three-dimensional geometric model data to obtain continuous variable features and discrete topological features, and fuse the continuous variable features, discrete topological features and the operating condition parameters to construct a joint feature vector; S3, acquire historical wheel hub data and corresponding historical NVH performance evaluation data, construct a training set based on the historical wheel hub data and the historical NVH performance evaluation data, and train a pre-constructed NVH performance prediction model based on the training set; S4. Input the joint feature vector into the trained NVH performance prediction model and output the NVH performance prediction result of the wheel hub to be tested.

2. The wheel hub NVH performance prediction method based on a parametric model according to claim 1, characterized in that, The operating parameters in step S1 include material properties and boundary constraints; wherein... When the boundary constraint condition is set to unconstrained full degrees of freedom, the NVH performance prediction result of the wheel hub under test is the natural frequency of the free mode of the corresponding wheel hub. When the boundary constraint condition is set to a six-degree-of-freedom full constraint on the bolt mounting hole, the NVH performance prediction result of the wheel hub under test is the natural frequency of the constraint mode of the corresponding wheel hub.

3. The wheel hub NVH performance prediction method based on a parametric model according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21, call the three-dimensional geometry kernel to parse the three-dimensional geometric model data of the wheel hub to be tested, and extract continuous variable features through geometric calculation; S22, the three-dimensional geometric model data of the wheel hub to be tested is projected in two dimensions or sliced ​​along a preset axis to generate a binary image, and discrete topological features are identified and calculated using image morphology and contour extraction algorithms. S23, standardize the continuous variable features, and fuse the standardized continuous variable features and discrete topological features with preset operating condition parameters to generate a joint feature vector.

4. The wheel hub NVH performance prediction method based on a parametric model according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31, construct a multi-source dataset of historical wheel hubs; wherein, the multi-source dataset includes source domain features composed of CAD simulation features of historical wheel hubs and target domain features composed of real bench test features; S32, map the source domain features and the target domain features to the regenerating kernel Hilbert space, and use the maximum mean difference algorithm to calculate the distribution distance between them; S33, construct a joint loss function that includes prediction task loss and distribution alignment loss by combining the distribution distance, and train a pre-constructed NVH performance prediction model based on the joint loss function.

5. The wheel hub NVH performance prediction method based on a parametric model according to claim 1, characterized in that, Step S4 specifically includes the following steps: S41, decouple the joint feature vector and route it; wherein, the discrete topological features are input to the gradient boosting tree branch, and the continuous variable features are input to the multilayer perceptron branch; S42, the discrete topological features are input into the pre-trained tree model, the leaf node landing index of the sample in all decision trees is extracted, the leaf node landing index is one-hot encoded and passed through the embedding layer to be transformed into a low-dimensional tree representation vector. S43, construct a fully connected multilayer perceptron containing multiple hidden layers, process the continuous variable features, and output a deep physical representation vector; S44, the low-dimensional tree representation vector and the deep physical representation vector are spliced ​​and fused across modes. After feature cross-calculation is completed by a fully connected layer, the NVH performance prediction result is mapped and output.

6. The wheel hub NVH performance prediction method based on a parametric model according to claim 3, characterized in that, The continuous variable features include rim thickness and spoke tilt angle, while the discrete topological features include the number of bolt holes and the number of spokes.

7. The wheel hub NVH performance prediction method based on a parametric model according to claim 6, characterized in that, The extraction of the rim thickness includes: identifying the inner and outer cylindrical surface models of the hub, extracting the outer and inner cylindrical surfaces, and calculating the rim thickness using the following formula: in, For the rim thickness, For sampling points on the inner surface, For point Along the local normal vector on the corresponding outer cylindrical surface Projection points on, The inner cylindrical surface; The extraction of the spoke tilt angle includes: extracting the normal vector of the spoke center plane and the principal axis vector of the hub rotation, and calculating the spoke tilt angle by the vector dot product. : in, The spoke tilt angle, Let be the normal vector of the center plane of the wheel spokes. The principal axis vector of hub rotation. This represents the inverse cosine function.

8. The wheel hub NVH performance prediction method based on a parametric model according to claim 4, characterized in that, The distribution alignment loss is calculated using the following formula: in, Represents the distribution alignment loss. Represents the first term of the source domain Individual sample feature data, This represents the total number of samples in the source domain. The first term of the target domain Individual sample feature data, This represents the total number of samples in the target domain. This represents the feature mapping function that maps the original features to a high-dimensional kernel space.

9. The wheel hub NVH performance prediction method based on a parametric model according to claim 8, characterized in that, The joint loss function is expressed as: in, Denotes the joint loss function. This indicates the source domain prediction of NVH performance results. Indicates the true label of the source domain. This indicates the predicted NVH performance results for the target domain. Indicates the true label of the target domain. This represents the loss function for the prediction task. This is a dynamically adaptive penalty coefficient.

10. The wheel hub NVH performance prediction method based on a parametric model according to claim 5, characterized in that, The NVH performance prediction results are expressed as follows: in, This indicates the NVH performance prediction results. Represented as the weight matrix of the output layer, Represented as a low-dimensional tree representation vector, Represented as a deep physical representation vector. It is represented as the bias vector of the output layer.

Images