Machine Learning Approaches for Predictive Maintenance in Smart Manufacturing Environments
Abstract
Predictive maintenance (PdM) is a key enabler for Industry 4.0. Predictive maintenance has one goal: to predict equipment failures in the future to minimize downtimes (and resulting costs) in production. Today's intelligent smartmanufacturing environment of sensors, big data analytics and machine learning (ML) has enabled around-the-clock machine health surveillance. This manuscript provides a review and a critical evaluation of state-of-the-art ML methodologies for PdM by categorizing the methodologies in supervised, unsupervised and reinforcement learning. The applicability of these methods in the fault detection, remaining useful life (RUL) estimation, and anomaly detection is investigated. A case study using vibration and temperature datasets obtained from CNC machinery is used to show the comparative performance of Random Forest (RF), Support Vector Machines (SVM), and Long Short Shane Memory (LSTMs) Neural Networks. According to evaluation performances including accuracy, precision, recall and mean absolute error (MAE), LSTM-based models outperformed the conventional ML approaches in terms of modeling the temporal dynamics of sensor data. It also discusses the problems such as data imbalance, model interpretability, and real-time application, and provides recommendations for hybrid architecture and edgecomputing implementation. Accordingly, a framework for ML algorithm selection and implementation in PdM systems in smart manufacturing scenarios is proposed in this study, aiming at increased reliability, safety, and cost efficiency.