Development of an automated simulation environment for Body in White joining techniques

Abstract

Parts in the joining station are usually arranged and clamped to result in an assembly that perfectly resembles the CAD0 construction. The perfect set-up is usually determined with the help of numerical optimization. In reality there are imperfections specific for each part, which are usually adjusted in the workshop manually. To do numerical optimization for each part is computationally expensive. Here the need for a more efficient Data mining analysis method arises. The present thesis investigates the metamodeling techniques in order to approximate the geometry response to the change of the joining station parameters. The aim is to provide the evaluation of the model by interpolating between the geometry and station set-up variations.

The suitable regressions, which are able to deal with non-linear geometry behaviour, are selected with the help of the literature research. The theory behind the numerical simulation, regressions and sampling is studied to ensure the right choice of the metamodel. An automated simulation environment is programmed to assist with the creation of variation in geometry and joining station, numerical solution and analysis. The chosen regressions - Support Vector Regression and Kernel Ridge Regression - are tested on models of different complexity. The errors are evaluated to provide the quantitative measure of the quality of regressions. Possible improvements of the metamodels are studied, such as Latin Hypercube sampling techniques.

The reduced complexity metamodels proved to provide a good approximation, while dealing well with non-linearities. On the other hand, it was shown that models with many design variables require improvement in sampling technique to provide better result within reasonable computational costs. At the same time, Latin Hypercube did not provide visible advancements in the tested cases. The Automated Simulation Environment and the tested metamodels are a base for the future implementation of an Artificial Neural Networks for defining the perfect set-up of Body in White joining stations for imperfect parts.