Machine learning for sensor-based manufacturing processes

Abstract

The increasing availability of relevant information, events and constraints in the environment of the modern factories due to deployment of IoT sensor technologies on the production line has led to an “explosion” in contextual big data. At the same time the advancements in the machine learning field from the last years opened new approaches for the analysis of the manufacturing processes datasets that are characterized by noisy data, a large number of features and an imbalanced classification of the samples. In this paper we investigate the applicability and the impact of machine learning techniques for managing production processes considering the data from a semiconductor manufacturing process (SECOM dataset). We have applied algorithms such as Boruta and MARS for the selection of the most relevant features and the Random Forest and the Gradient Boosted Trees for the samples classification. The results show better values for precision when the features are selected using Boruta and MARS rather than PCA and better values for accuracy when the data is unsampled and classified using Random Forest and Logistic Regression rather than Gradient Boosted Trees.