Automatic analysis of crash simulations with dimensionality reduction algorithms such as PCA and t-SNE

The increasing number of crash simulations and the growing complexity of the models require an efficiently designed evaluation of the simulation results. Nowadays a full vehicle model consists of approximately 10 million shell elements. Each of them contains various evaluation variables that describe the physical behavior of the element. Therefore, the simulation models are very high dimensional. During vehicle development, a large number of models is created that differ in geometry, wall thicknesses and other properties. These model changes lead to different physical behavior during a vehicle crash. This behavior is to be analyzed and evaluated automatically. In this article, potentials of several algorithms for dimensionality reduction are investigated. The linear Principal Component Analysis (PCA) is compared to the non-linear t-distributed stochastic neighbor embedding (t-SNE) algorithm. For those algorithms, it is necessary that the input data always has an identical feature space. Geometrical modifications of the model lead to changes of finite element meshes and therefore to different data representations. Therefore several2D and 3D discretization approaches are considered and evaluated (sphere, voxel). In order to assess the quality of the results, a scale-independent quality criterion is used for the discretization and the subsequent dimensionality reduction. The simulations used in this paper are carried out with LS-DYNA. The aim of the presented study is to develop an efficient process for the investigation of different data transformation approaches, dimensionality reduction algorithms, and physical evaluation quantities. The resulting evaluation method should represent physically relevant effects in the existing simulations in a low-dimensional space without human interaction and thus support the engineer in the evaluation of the results.