An Effective Methodology for Imbalanced Data Handling in Predictive Maintenance for Offset Printing

The printing industry is one of the largest manufacturing industries in the world, being characterized by high production volumes, where continuous maintenance of machine performance is key. Predictive Maintenance (PdM) enables the use of a maintenance policy based on novel Machine Learning (ML) algorithms, in or-der to provide valuable insights for both diagnostics and prognostics. However, real-world data used for PdM model training are characterized by great class im-balances, as failure events have a significant lower rate of happening compared to the normal no failure operations. Furthermore, ML models that are subjected to imbalanced datasets, are prone to be highly biased while having misleading accuracy scores. This can prohibit systems to accurately predict machine failure, leading to excessive costs while affecting the safety of the workers. This work proposes a data sampling methodology for predictive maintenance algorithms used mainly in Offset Printing environments, aiming to improve model performance. Based on a historical dataset extracted by an Offset Printing manufacturer, a methodology consisting of multiple classification algorithms utilizing different sampling techniques (SMOTE, ADASYN, and RUS), was trained and evaluated using cross-validation. The evaluation outcomes demonstrated the ability of the proposed methodology to effectively handle data imbalances while significantly enhancing model performance, outperforming other state-of-the-art techniques.