Deep Learning for Predictive Maintenance in Smart Manufacturing — A Review

Abstract

Deep learning (DL) has emerged as a transformative tool for predictive maintenance (PdM) and fault diagnosis across industrial domains such as aerospace, automotive, energy, and process systems. This review synthesizes 35 recent studies employing diverse DL models—including CNN, LSTM, Transformer, Autoencoder, GAN, GNN, and hybrid physics-informed architectures—applied to sensor, acoustic, vibration, and process signals. The findings reveal that DL significantly improves Remaining Useful Life (RUL) estimation, anomaly detection, and fault classification, outperforming traditional machine learning approaches. Despite these advances, challenges persist: large data requirements, limited cross-domain generalization, model interpretability gaps, high computational cost, unstable training in generative methods, and unclear thresholds in unsupervised detection. Moreover, most research is constrained to component-level validation, with limited industrial deployment. This review identifies critical research gaps and provides future directions to guide the development of scalable, explainable, and resource-efficient DL solutions for real-world predictive maintenance.