1. Introduction
The vehicle development process is subject to strong competitive pressure because of the continuously increasing number of vehicle derivatives and shorter vehicle development cycles [
1]. Aggressive new players with innovative technologies and progressing digitalization in the automotive industry are causing further challenges for vehicle manufacturers [
2].
According to Klostermeier [
3], the use of digital twins is important for the automotive industry as increasingly efficient and faster product development with a simultaneous reduction in real prototypes helps to reduce costs and development time.
The full vehicle specifications defined at the beginning of the vehicle development process from the domains of ride comfort, driving dynamics, acoustics, and so forth often lead to component requirements at the component level with a high demand on design and material selection [
4]. With the use of digital twins in the development process, it is possible to generate design concepts at an early phase and use them to make virtual estimates of the required installation space as well as forecasts of the target value fulfillment.
Specifically, the design of elastomeric bearings is important for achieving the required full-vehicle characteristics. The vibration-isolating behavior based on the viscoelastic characteristics of elastomers helps to reduce vibrations and noise from engine-, transmission-, and road-induced vibrations [
4]. In addition, elastomeric bearings (chassis bushings) in the chassis contribute significantly to the elasto-kinematic characteristics of the wheel control. The design of the compliance of chassis bushings helps to adjust wheel position values under the action of forces [
4].
Due to the nonlinear material behavior of rubber [
4], the design of the component behavior is very demanding and requires robust virtual calculation and simulation methods.
Adkins [
5] and Göbel [
6] have already developed analytical calculation models for the stiffnesses of cylindrical bushings at radial, axial, and torsional deformation. They assumed linear elastic material behavior for quasi-static deformation of the elastomer material. Their calculation models are based on cylindrical bushings with a constant cross-section.
The validity of analytical calculation models based on linear elastic material behavior has been investigated in a previous study [
7] that considered eight simple bushings (cylinders with constant cross-section) with different design parameters. The mean relative deviation of the calculated to the measured static stiffness was 10% for radial deformation.
Bushings in modern vehicles are subject to strong requirements. To fulfill the target values, it is often necessary to use component designs that deviate significantly from simple bushings with constant cross-sections. Complex component designs (e.g., contoured inner sleeves) lead to an inhomogeneous stress distribution in the elastomer material and require numerical methods [
4,
8].
For the design of elastomer components with complex geometries, the finite element method (FEM) is usually used. This method enables the variation of individual model parameters (e.g., geometry parameters) and the calculation of the mechanical properties [
9]. The use of FEM in early vehicle development phases enables a prediction of the packaging as well as forecasts of the target value fulfillment.
For the design of an engine mount, Liu [
10] has presented an optimization method based on FEM. Liu describes a geometric optimization problem using an objective function to minimize the error between calculated and required stiffness in two different loading directions. The design parameters are optimized in a given design space. Kaya [
11] has described another optimization method using FEM for the design of suspension bushings, presenting the geometric optimization problem by minimizing the error between the calculated and measured force signal. The design parameters are also optimized in a given design space. However, the optimization of design parameters to fulfill the target values using FEM leads to high simulation efforts and, therefore, to high component costs.
A more efficient method for predicting component properties in relation to design parameters is the use of artificial neural networks (ANNs). However, this requires a large database containing the design parameters depending on the target values.
Jung’s [
12] method utilizes FEM results to train ANNs. In this study, at given design parameters, the stiffness of bushings was predicted using ANNs. Cernuda [
13] has investigated the usability of ANNs in the vehicle development process. The dataset was generated by FEM as well. This study also shows that ANNs can be used to predict the stiffnesses of bushings depending on their design parameters.
In addition to the design of the static behavior of elastomeric bearings, the structural durability of elastomeric components is important in the vehicle development process. Ernst [
14] has introduced a virtual method to define the requirements for the structural durability of engine mounts. He developed a complex modeling approach based on empirical rheological descriptions, which represents the component property changes due to multi-axial loads. The complex rheological substitute model of an engine mount was parameterized using particle swarm optimization based on test rig experiments.
Uriarte [
15] has shown the use of particle swarm optimization for the determination of material parameters of the hyper-elastic material model following
Mooney-Rivlin [
16]. The study describes particle swarm optimization as a simple and intelligent solution method that requires only a few parameters and leads to a solution quickly.
In the current work, ANNs are used to design the quasi-static component behavior of bushings. The database required for the network training is generated using FEM and includes different design parameter sets and the associated quasi-static stiffness of various loading directions. The difference of this study compared to the mentioned studies is the design optimization under consideration of constraints of the mechanical as well as geometrical properties of the bushing based on artificial neural networks. With the help of the presented method, the design process can be inverted to require target values and to calculate design parameters.
This paper is structured as follows:
Section 2 describes the experimental characterization of a physical bushing, the present component design, and the FEM of a reference bushing. The determination of the material parameters of the reference bushing-model based on material parameter optimization is presented in
Section 3.
Section 4 shows the generation of different design parameter sets using the design of experiments (DoE) method and the generation of the FEM database. The training of ANNs is described in
Section 5. Finally,
Section 6 presents a design parameter optimization under constraints based on the ANN applying different usecases as examples.
7. Conclusions
In this work, the application of ANNs was investigated for the design process of chassis bushings. The nonlinear and strongly geometry-dependent transfer behavior resulting from the elastomer of the bushing was modeled by ANNs.
In the first step, a physical chassis bushing was characterized on a multi-axial elastomer test rig. Subsequently, a reference bushing based on the physical bushing was created as a finite element model in the simulation environment ABAQUS. The geometrical modeling of the inner core and outer sleeve was approximated with basic geometrical elements and relations between them. The free elastomer contour required an optical measurement and was modeled as a spline. To model the material behavior, the material model according to Yeoh was used. The material parameters for the Yeoh model were determined using a subsequent material parameter optimization based on PSO and the measurement results of the physical bushing. For the subsequent design study, a geometrical model simplification was conducted regarding the optically measured elastomer contour. Subsequently, the geometrical model simplification was validated in relation to the reference bushing model. The resulting FEM model represented the simplified bushing model for the design study.
In the next step, six design parameters were defined for the geometrical design of the bushing, which represented the basis for the subsequent design study. Based on the DoE method of Latin hypercube sampling, 20,000 samples were created in a design parameter space to generate the training data. To avoid invalid geometrical models within the design parameter space, geometrical constraints were defined. Subsequently, the design parameter space was cleaned, resulting in 2322 valid samples.
In the following step, the cleaned design parameter space formed the basis for the training of the ANNs. For the stiffness prediction in the radial, axial, and torsional load directions, one ANN was trained each. The network training was based on the Adam optimizer and 80% of the data was used for the training process. The remaining 20% of the data was used for the validation of the trained networks. To increase the prediction accuracy of the ANN, a hyper-parameter optimization based on Bayesian optimization was performed, resulting in three different ANN topologies.
Finally, a DO was presented using the ANN for the design process. The basis for this was PSO which was extended to address the geometrical constraints. To showcase the performance of the DO, two usecase studies were discussed and the accuracy of the DO was evaluated. In the first usecase, target values in the form of quasi-static stiffnesses of the radial, axial, and torsional load directions were required. In the second usecase, in addition to the required target values of the quasi-static stiffnesses, two installation space requirements (width and outer diameter of the elastomer body) were defined. Using DO, the optimal design parameters to fulfill the target values were determined in each case. In both cases, a MAPE < 3% was achieved between the target values and the stiffnesses from the DO. To evaluate the deviation between the ANN prediction and the results of the FEM simulation using the resulting design parameters from the DO, the MAPE was calculated. In both usecases, the MAPE was smaller than 2%.
The small prediction errors represent a very good prediction accuracy of the ANN models as well as good design accuracy of the entire method. In future work, the transferability of the method to other types of chassis bushings (e.g., bushings with an intermediate sleeve) should be tested. Furthermore, the method can also be generalized and used for installation space issues with other components. In addition, the extension to more complex optimization tasks, including multi-objective optimization, should be investigated.