Process for Extraction of Knowledge from Crash Simulations by means of Dimensionality Reduction and Rule Mining
This thesis proposes an efficient process flow for analyzing an ensemble of vehicle crash simulations. The process has two goals. The
first goal is to algorithmically detect the largest and most notable types of deformation. The second goal is to find out how to avoid or trigger
a user-specified deformation behavior.
The first goal of deformation behavior segmentation, is approached with a novel dimensionality reduction technique. This dimensionality
reduction technique not only makes it possible to derive a lightweight, intermediate, and mesh-free representation of the simulation results,
but also makes it possible to compute a normalized simulation similarity.
These simulation similarities can be used to find groups of similar deformation behaviors by using clustering algorithms and lowdimensional
embeddings. An engineer then has to decide which types of deformation are acceptable, and which are not. Having chosen the desired types of deformation, it is then possible to find out how they can be achieved by using rule mining. The rule mining algorithm returns multiple safe design spaces, in which the engineer’s demand is fulfilled. Because the rules might be unsafe at their boundaries, an optimization can limit the probability of a rule being wrong to within a specified limit.