Development of prediction models
for property prediction in assembly products
In several cases, the final assembly property will closely depend on the property under investigation, but mainly on the manufacturer process. An example can be borrowed from the assembly process of electrical transformers, several properties of which (e.g., power losses) depend closely on the corresponding properties of the component materials but also on the manufacturer process. In such assembly lines, it is necessary to be able to predict accurate enough the properties of the final product based on measurements of previously built transformers. Given the fact that the nature of the phenomena involved (e.g., magnetic-field properties in an electrical transformer) cannot be modeled accurately as a function of the materials used as well as the manufacturer process, we wish to develop prediction models that provide property predictions based on material and manufacturer parameters. Currently developed models are focusing on standard linear regression formulations, however we wish to develop nonlinear prediction models (e.g., based on neural networks) that will be able to capture more accurately the underlying physical phenomena. This thesis will primarily consist of the following steps:
- Advancing an existing testing framework for statistically analyzing the impact of assembly and component parameters in the final assembly properties.
- Development of prediction models that relate the assembly, component and manufacturer parameters with the final assembly properties.
The currently used programming platform is C++.