AI-Driven Digital Twin Technology for Predictive Maintenance in Smart Manufacturing Environments: A Comprehensive Review of Architectures, Algorithms, and Industrial Applications
Abstract
ABSTRACT
Unplanned equipment downtime remains one of the most costly disruptions in discrete and process manufacturing, and conventional reactive and time-based preventive strategies are poorly matched to the stochastic degradation behaviour of modern cyber-physical production systems. Digital twin (DT) technology — a continuously synchronised virtual representation of a physical asset — combined with artificial intelligence (AI) offers a route to condition-aware, prognostic maintenance. This review-based study synthesises peer-reviewed literature on AI-driven digital twins for predictive maintenance (PdM) in smart manufacturing, following a structured protocol of question formulation, source screening, thematic coding, and cross-study comparison. A five-layer reference architecture and a closed-loop PdM workflow are proposed, and the reviewed approaches are compared across modelling paradigm, learning technique, and deployment maturity. The synthesis indicates that hybrid twins, which couple physics-based models with deep-learning prognostics, offer the most favourable balance of early fault detection, adaptability, and interpretability, whereas purely data-driven pipelines are easier to deploy but depend heavily on labelled failure data. Persistent barriers include data scarcity, synchronisation latency, interoperability gaps, and the absence of standardised validation benchmarks. Promising research directions include physics-informed learning, edge-deployed and federated twins, and uncertainty-aware decision support. The paper provides engineers and researchers with a consolidated technical reference for designing and evaluating DT-based maintenance systems.
KEYWORDS: Digital twin, Predictive maintenance, Smart manufacturing, Industry 4.0, Remaining useful life, Deep learning, Cyber-physical systems, Physics-informed machine learning, Condition monitoring, Fault diagnosis
Unplanned equipment downtime remains one of the most costly disruptions in discrete and process manufacturing, and conventional reactive and time-based preventive strategies are poorly matched to the stochastic degradation behaviour of modern cyber-physical production systems. Digital twin (DT) technology — a continuously synchronised virtual representation of a physical asset — combined with artificial intelligence (AI) offers a route to condition-aware, prognostic maintenance. This review-based study synthesises peer-reviewed literature on AI-driven digital twins for predictive maintenance (PdM) in smart manufacturing, following a structured protocol of question formulation, source screening, thematic coding, and cross-study comparison. A five-layer reference architecture and a closed-loop PdM workflow are proposed, and the reviewed approaches are compared across modelling paradigm, learning technique, and deployment maturity. The synthesis indicates that hybrid twins, which couple physics-based models with deep-learning prognostics, offer the most favourable balance of early fault detection, adaptability, and interpretability, whereas purely data-driven pipelines are easier to deploy but depend heavily on labelled failure data. Persistent barriers include data scarcity, synchronisation latency, interoperability gaps, and the absence of standardised validation benchmarks. Promising research directions include physics-informed learning, edge-deployed and federated twins, and uncertainty-aware decision support. The paper provides engineers and researchers with a consolidated technical reference for designing and evaluating DT-based maintenance systems.
KEYWORDS: Digital twin, Predictive maintenance, Smart manufacturing, Industry 4.0, Remaining useful life, Deep learning, Cyber-physical systems, Physics-informed machine learning, Condition monitoring, Fault diagnosis