Design of Automotive Joints Using Response Surface Polynomials and Neural Networks
MetadataShow full item record
In the early design stages of a car body, a simplified model, which represents the constituent components of the car body by their performance characteristics, is used to optimize the overall car. The determined optimum performance characteristics of the components are used as performance targets to design these components. Since designers do not know the relation between the performance characteristics of the components and their dimensions and mass, this may lead to unreasonable performance targets for the components. Moreover, this process is inefficient because design engineers use empirical procedures to design the components that should meet these targets. To design the component more efficiently, design tools are needed to link the performance targets with the physical design variables of the components. General methodologies for developing two design tools for the design of car joints are presented. These tools can be viewed as translators since they translate the performance characteristics of the joint into its dimensions and vice-versa. The first tool, called translator A, quickly predicts the stiffness and the mass of a given joint. The second tool, called translator B, finds the dimensions and mass of the most efficient joint design that meets given stiffness requirements, packaging, manufacturing and styling constraints. Putting bulkheads in the joint structure is an efficient way to increase stiffness. This thesis investigates the effect of transverse bulkheads on the stiffness of an actual B-pillar to rocker joint. It also develops a translator A for the B-pillar to rocker joint with transverse bulkheads. The developed translator A can quickly predict the stiffness of the reinforced joint. Translator B uses optimization to find the most efficient, feasible joint design that meets given targets. Sequential Linear Programming (SLP) and the Modified Feasible Direction (MFD) method are used for optimization. Both Response Surface Polynomial (RSP) translator B and Neural Network (NN) translator B are developed and validated. Translator A is implemented in an MS-Excel program. Translator B is implemented in a MATHEMATICA program. The methodology for developing translator B is demonstrated on the B-pillar to rocker joint of an actual car. The convergence of the optimizer is checked by solving the optimization problem many times starting from different initial designs. The results from translator B are also checked against FEA results to ensure the feasibility of the optimum designs. By observing the optimum designs and by performing parametric studies for the effect of some important design variables on the joint mass we can establish guidelines for design of joints.
- Masters Theses