Automated Machine Learning for Failure Detection Model in Predictive Maintenance

Abstract

The advent of new technologies has precipitated a swift proliferation of digitalization within the manufacturing sector, giving rise to novel conceptualizations. The present study is concerned with the topic of predictive maintenance. The advent of the Digital Twin (DT) approach has led to a notable rise in prominence for predictive maintenance. In order to perform predictive maintenance in an effective manner, it is of paramount importance to predict potential failures in advance. In this study, a dataset comprising five distinct failure classes, determined based on factors such as air temperature, process temperature, rotational speed, torque, and tool wear, is considered. While previous studies have primarily focused on machine learning (ML) algorithms, this study makes a distinctive contribution to the field by employing automated machine learning (AutoML) libraries. The objective of AutoML libraries is to autonomize ML algorithms in order to obtain optimal results. Despite their recent use in a number of studies, they have not yet gained widespread acceptance. In this study, three open-source libraries of the Python programming language, namely AutoSklearn, AutoKeras, and PyCaret, will be employed for data analysis and comparison of the resulting outputs. A systematic comparison will be conducted to identify the most suitable algorithm. Additionally, this study aims to utilise a hyperparameter optimisation approach, which will enhance the prevalence and applicability of predictive maintenance studies. This study contributes to the advancement of predictive maintenance applications. © 2025 Elsevier B.V., All rights reserved.