PSI - Issue 58

Kay Büttner et al. / Procedia Structural Integrity 58 (2024) 95–101 Kay Büttner et al. / Structural Integrity Procedia 00 (2019) 000–000

99 5

Fig. 4. Development process of elastomeric bushings.

The calculation of quasi-static stiffnesses in early vehicle development phases shall be based on simplified physical calculation models from Göbel (1969) and Töpel et al. (2019). These simple models were limited to simple bushing concepts with constant cross sections. Due to the steadily increasing requirements at the total vehicle level and thus also at the component level, significantly more complex bushing concepts are used today (for example: Bushing concepts with variable cross-sections, intermediate plates or preloaded bushings). These requires new approaches like the data-driven design for the estimation of target values and geometries in early vehicle development phases. The development of such data-driven models for establishing the parameter-property relationships requires a data base of component characteristics and geometry. With the help of experimentally and simulatively (FEM) determined empirical data, artificial neural networks can be trained and used for the design of complex bushing concepts. In the vehicle development process, target values (TV) are specified and the package must be predicted as accurately as possible. The challenge here is the inversion of the physical as well as data-driven models, because the models are working by prescribing geometric design parameters and estimating quasi-static stiffnesses. Especially for strongly non-linear relationships, an effective optimization method is necessary. In this work, the particle swarm optimization is used. The optimization problem is shown in Fig. 5. The loss function is defined by the mean absolute percentage error of the quasi-static stiffnesses based on the experimental characterization and the predicted stiffnesses of the developed models. Thus, the feed forward neural networks (DP in and TV out) can be used for the design process in early vehicle development phases and the design space estimation can be done on the basis of the target value requirement (TV in and DP out). With the developed models, forecast accuracies of ±15% can be achieved. This range corresponds approximately to the tolerance specifications due to manufacturing process influences. With the help of robust calculation model, precise results and a reduction in development time and cost-intensive prototypes are achieved.

Fig. 5. Optimization problem for inverting the physical and data-driven models.