FEMMINER: Data Mining Techniques for Car Simulation Results

More product variants, the increase in government regulations and the use of stochastic and multidisciplinary optimization cause a drastic increase of the number of numerical simulation performed in the car design process. The availability of this data creates the opportunity to compare thousands of simulation variants, to find optimal design variants (even from previous projects) and to evaluate their impact on the functionality of the car. The huge volume of data (large number of simulations and large data-sizes per simulation) and the diverse content of simulation data are major obstacles that can be overcome by data mining methods based on dimension reduction.

Aspects of the car design process in automotive industry
Our activities in data mining of simulation results focus on two aspects of the virtual product development­ process:

  • In order to find an optimal design while taking specific constraints into account (like safety regulations), many model variants are analyzed in which small modifications in material parameters and geometries are introduced. This is an iterative and time consuming process. Post-processing software tools are able to display the model geometry and the results of the numerical simulation in an intuitive way, but engineering judgment is limited to the analysis of just a few (perhaps at most five) simulations­ at once.
  • The other aspect of the car design process involves the reuse of information from previous projects or development­ phases­. Geometrical shapes and geometrically­ distributed properties like stresses or deformations­, defined on very fine meshes, are the essential information­ to evaluate. Unfortunately, this information is not accessible for direct data base queries­ because of its size and its missing common structure­.

In both application cases, nonlinear dimension reduction methods are used that exploit the high correlation arising from the basic similarities in the models and problems to be solved. The major trends of the differences, which are hidden in the detailed results, are computed and simple parameters provide information about the contribution of each of these trends to the individual geometries or simulation results. A simple illustration of this idea can be given by considering several pictures of a car that has been rotated around one axis (see Figure 1). After application of a nonlinear dimension reduction to adjust the images consisting of pixel values via the so-called Multidimensional Scaling (MDS) method, we are able to extract dependency on the rotation angle and represent all images in a parametric way.

Figure 1: The rotation angle of the images is a parameter that can be identified by a reduction method­ applied to the images. We use a multidimensional scaling technique to assign a location in N-dimensional space to each data, such that neighboring­ data are more related to each other than distant ones. For sufficiently small N, the resulting locations may be displayed in a graph or 3D visualization.

Each circle in the MDS plot (see Figure 1 and its explanation) represents a snapshot of the car and the major parameter resulting from the mathematical­ analysis turns out to be just the rotation angle.

Applying nonlinear dimension reduction methods
Figure 2 shows the application of these methods to vehicle crash simulation. The deformations­ of a set of parts at the front of a car are extracted from 132 simulations and nonlinear reduction methods are applied to these data sets. Here, each point corresponds to one of these 132 simulation results. Its position in the graph is given by the contribution of the two major deformation­ modes to the simulation result for this point. The formation of the points is almost a one-dimensional curve, indicating that the results are determined by one major deformation mode. By comparing two extreme results, it can be seen that the behavior of the deformation mode is determined by lower longitudinal rail. The color of each point indicates the thickness of the frontal part of the longitudinal rail and it is obvious that the major mode is highly correlated with this thickness. The example shows that the essence of 132 simulation­ results can be derived by analyzing the MDS plot and visualizing a very small subset of carefully­ selected simulation results.

Figure 2: Analysis of a set of crash simulation results using­ high-dimensional data reduction­ (testcase calculated­ using a truck model from the US National Crash Analysis Center).

The main objective of Noise, Vibration and Harshness (NVH) simulations of a car is the comfort of the passengers when exposed to car vibrations. Figure 3 shows examples of 32 possible changes (red colored parts) along with the 32 corresponding vibration response curves from such simulations. Our dimension reduction methods generate a MDS plot with a point for each variant that is based on the similarity of the vibration response curves. It shows groups of different curve shapes due to changes in 1) the seat masses, 2) the cooler properties, 3) the lateral structure (at the door) and 4) the part thickness. The MDS plot not only allows for the identification of these groups, it can also be seen that changes in the thickness do not influence­ the curve significantly (big cluster on the right). Having all 32 curves organized in this way allows the engineer to fast and easily evaluate his designs.

Figure 3: Analysis of a set of NVH simulation results using high-dimensional data reduction­ (testcase provided­ by AUDI AG as part of the SIMDAT EU project­).

The described development was initiated by a Fraunhofer Challenge Project running from 2009 to 2010 and is continued within the project SIMDATA-NL – funded by the German Federal Ministry­ of Education and Research (BMBF) – until 2013. The project focuses on the application­ of efficient mathematical methods for dimension reduction of large data sets. In another project, named FEMMINER (funded by the BMBF from 2010 to 2012), SCAI and the company GNS mbH in Braunschweig implement these methods in GNS’s post-processing software environment.

Our expectation is that the new analysis environment will substantially increase the productivity­ of engineers. Consequently, a faster way to car designs will be possible.