Using supervised and one-class automated machine learning for predictive maintenance
Luís Ferreira, André Pilastri, Filipe Romano, Paulo Cortez
Abstract
Predictive Maintenance (PdM) is a critical area that is benefiting from the Industry 4.0 advent. Recently, several attempts have been made to apply Machine Learning (ML) to PdM, with the majority of the research studies assuming an expert-based ML modeling. In contrast with these works, this paper explores a purely Automated Machine Learning (AutoML) modeling for PdM under two main approaches. Firstly, we adapt and compare ten recent open-source AutoML technologies focused on a Supervised Learning. Secondly, we propose a novel AutoML approach focused on a One-Class (OC) Learning (AutoOneClass) that employs a Grammatical Evolution (GE) to search for the best PdM model using three types of learners (OC Support Vector Machines, Isolation Forests and deep Autoencoders). Using recently collected data from a Portuguese software company client, we performed a benchmark comparison study with the Supervised AutoML tools and the proposed AutoOneClass method to predict the number of days until the next failure of an equipment and also determine if the equipments will fail in a fixed amount of days. Overall, the results were close among the compared AutoML tools, with supervised AutoGluon obtaining the best results for all ML tasks. Moreover, the best supervised AutoML and AutoOneClass predictive results were compared with two manual ML modeling approaches (using a ML expert and a non-ML expert), revealing competitive results.