Matthieu Loustaunau: Industrial process error estimation by machine learning

Abstract:

Performing an set-up on a complex machine may be difficult. This problem arises frequently for the industry, especially when the relation between input and output data cannot be defined precisely. Heavy methods of optimization may be used to perform an set-up. This thesis investigate the possibility to use a machine learning approach on a specific machine. We study the structure of the relation between input and output data. We show the variations are smooth. We define a set of test to evaluate future models. We design and test several models on simulation data, and select the best one. We design a strategy to use data in the best possible way. The selected model is then tested on actual data in order to be optimized.