Automated evaluation of crash simulations with different stages in vehicle development using machine learning methods
Increased digital vehicle development is a key driver of the sharp rise in the number of crash simulations that are performed using the finite element (FE) method. The simultaneous increase in the complexity of the simulation models and the complex crash behavior of a vehicle mean that manual evaluation of the simulations is time-consuming and incomplete. As a result, there is a risk that conspicuous crash behavior will be overlooked, leading to incorrect design measures being taken on the vehicle. Automated evaluation of the crash behavior is difficult because the FE meshes of the individual simulations differ.
In this dissertation, a discretization method is used to ensure the comparability of different FE meshes, which ensures further analysis of the simulations. Outlier detection is presented as a new method to automatically analyze crash behavior for all components, time steps and evaluation variables. In this process, conspicuous crash behavior is identified by comparison with simulations from the history and clearly presented to the user in the post-processor. With the crash behavior detector (CVD) the engineer is warned in case of recurrence of a given crash behavior in new simulations. This allows an individualized evaluation, according to the user's specifications. The outlier detection and the CVD ensure a more complete evaluation of the crash simulations by analyzing all information automatically and directing the focus specifically to the relevant areas of a crash simulation.
In addition, dimensionality reduction methods for visualizing a bunch of simulations are considered. The extent to which nonlinear algorithms provide advantages over linear methods is investigated. Furthermore, different possibilities are considered how the temporal behavior can be considered. Dimensionality reduction enables the user to clearly visualize the crash behavior for a large number of simulations for selected components in scatter and line diagrams. This allows the identification of simulations from the history that show similar crash behavior. This newly gained information helps the engineer to define new design measures to improve the crash behavior.
