Machine Learning-Assisted Fault Diagnosis in Smart Manufacturing Systems

Abstract

Rotating machinery fault diagnosis has moved, over roughly the past decade, from hand-engineered vibration and current-signal features paired with comparatively simple classifiers toward end-to-end deep learning architectures that learn diagnostic features directly from raw sensor signals, following the same representation-learning shift that transformed image classification. This article synthesizes that literature, tracing the field from convolutional and recurrent architectures for vibration-based diagnosis, through the transfer learning, few-shot learning, and synthetic data augmentation techniques developed to address the chronic labeled-fault-data scarcity industrial settings present, to the integration of fault diagnosis models into the broader smart manufacturing stack of digital twins, industrial IoT prognostics frameworks, and privacy-preserving federated learning across facilities. We give particular attention to a growing strand of work on lightweight and explainable architectures, which pushes back against the assumption that the largest, least constrained model is automatically the right one to deploy. Read together, the evidence supports machine learning-assisted fault diagnosis as an increasingly mature and convergent methodology, while leaving the gap between strong benchmark performance on controlled laboratory datasets and reliable detection of novel fault modes under genuinely heterogeneous, multi-site field conditions as a substantially open research agenda.