Machine Learning for Equipment Failure Prediction in Smart Manufacturing Plants: A Review

INTRODUCTION:Smart manufacturing plants generate continuous streams of vibration, acoustic, thermal, electrical, process-control and quality data. These data make it possible to move from reactive or calendar-based maintenance toward predictive maintenance, but they also introduce problems of missing labels, non-stationary operating regimes, rare failures and unequal costs of false alarms and missed failures. OBJECTIVES:This review analyses how machine learning is used to forecast equipment failures and remaining useful life in intelligent production environments, with emphasis on the suitability of methods for real industrial deployment rather than benchmark accuracy alone. METHODS:A structured narrative review was conducted over established prognostics and health management literature, Industry 4.0 predictive-maintenance surveys, and representative empirical studies using public and industrial datasets such as C-MAPSS, PRONOSTIA, IMS bearings, Bosch Production Line Performance, semiconductor ion-implantation data and IoT-enabled production-line data. RESULTS:The reviewed literature shows that feature-based classifiers, gradient boosting, support vector machines and random forests remain competitive when labels are scarce and process knowledge is available. Deep learning improves representation learning for multivariate sequences, especially for vibration and run-to-failure data, but requires careful validation against temporal leakage and domain shift. Hybrid approaches that combine signal processing, physics, digital twins and cost-aware decision rules are increasingly important for plant-level use. CONCLUSION:Machine learning can substantially improve failure prediction in smart factories, yet its value depends on data governance, uncertainty handling, maintainability and integration with maintenance planning. The most credible systems treat prediction as a socio-technical decision process, not as a stand-alone model.