Digital Twin Frameworks for Predictive Maintenance in Smart Manufacturing

Abstract

Manufacturers have long faced a choice between reactive maintenance, which wastes money on unplanned downtime, and
fixed-schedule preventive maintenance, which wastes good component life, and the digital twin has emerged as the organizing
concept for a third option: a continuously synchronized virtual counterpart of a physical asset that can forecast a specific
failure mode far enough in advance to act on it. This article synthesizes the literature on digital twin frameworks for predictive
maintenance in smart manufacturing, tracing the concept from Grieves and Vickers' original physical-virtual-connection
model through Kritzinger et al.'s influential three-tier taxonomy distinguishing a digital model, a digital shadow, and a true
digital twin, into the reference architectures -- Lee et al.'s cyber-physical 5C pipeline, Tao and Zhang's shop-floor paradigm, and
Lu et al.'s broader reference model -- that operationalize it on the factory floor. We review the physics-based, data-driven, and
hybrid methods used to turn a digital twin into a working prognosis, and we examine the comparatively rare field evaluations,
including a recent CNC-machining framework reporting R² above 0.94 for two prediction tasks and an 18-month steel-industry
field trial with a measured, transparent detection record, that ground the field's architectural claims in evidence. We conclude
that the conceptual and architectural foundations are mature and convergent across independent research groups, while the
empirical base for sustained field performance, the closing of the automated loop that separates a digital twin from a digital
shadow, and the integration of asset-level prognoses into system-level production planning remain open engineering problems.