Machine learning methods for manufacturing

Abstract

Machine learning methods have become increasingly popular with the release of numerous open-source tools and libraries. Nevertheless the adoption of these techniques for use in manufacturing has been limited in practice. Manufacturing is still mostly dependent on traditional statistical methods and tools, even though machine learning methods could be applied to data that is already being collected from measurements done during manufacturing processes. The purpose of this thesis is to introduce four different machine learning methods, that could prove to be useful in a manufacturing setting, and several different methods relating to the preprocessing of data and preliminary data analysis. The machine learning methods introduced are support vector machines, random forests, neural networks and NARX (non-linear autoregressive exogenous) neural networks. The algorithms and the history behind the methods introduced are explained, along with suggestions for some popular implementations of the algorithms, and the performance of each the methods is evaluated using a domain appropriate dataset. Knowledge of the machine learning methods introduced in this thesis are an important addition to the toolkit of anyone doing predictive analytics.