Growing interest in the use of virtual simulation tools as part of the automotive product development process is driven by the need for automotive manufacturers and parts suppliers to develop better quality products in shorter time at lower cost.

Component and full vehicle durability testing is one aspect of product development for which time savings can be realized. Traditionally, accelerated durability simulations have been performed using full vehicles by driving physical prototypes on specially designed road surfaces, simulating the vehicles’ service life. In the last thirty years, durability testing has been accelerated in the laboratory environment where measured vehicle excitation inputs have been edited to contain only the most damaging portions. The goal of the current research is to advance the process further through the use of high-fidelity virtual prototype durability simulations, which reveal the consequences of design decisions made much earlier in the product development cycle before the first physical prototypes are built.

Virtual durability full vehicle models are computationally complex. Linearizing the individual models of nonlinear components such as shock absorbers and elastomeric bushings has been a typical method used to simplify the vehicle model. The focus of the current research is to develop a methodology to increase the fidelity of these nonlinear component models using computationally economical techniques, thus increasing the precision of the results of the full vehicle model and the speed at which the results are obtained.

Neural networks are mathematical models that possess the flexibility and computational efficiency desired for this application. These models are capable of generalizing component behaviour using training data that represents the full range of component behaviour that is to be modeled.

The current research describes the methodology required to develop and implement neural network models of nonlinear automotive components into simplified and full-vehicle virtual durability models. The data used to train the neural networks includes hysteresis effects that are not modeled with the methods currently available in the multibody dynamics software package. Correlation of the results of the virtual durability simulation with the laboratory test results is performed to show the validity of the methodology that was developed.