EdgeTwin as a Deployment-Aware Edge AI Digital Twin for Predictive Maintenance and Visual Anomaly Detection in SME Smart Manufacturing

Abstract

Smart manufacturing increasingly uses AI-driven digital twins, predictive maintenance, and visual inspection, but less is known about how these capabilities can be integrated into lightweight Edge AI architectures usable by resource-constrained SMEs, making deployment-aware evaluation important. This study develops and evaluates EdgeTwin through design science and experimental applied computing. The framework combines machine-condition data, industrial inspection images, edge-oriented inference, digital twin representation, and human-supervised decision support. Machine-failure prediction was tested using AI4I 2020, visual anomaly detection was evaluated on MVTec AD, INT8 optimization was evaluated in a workstation-based edge-oriented simulation, and contextual feasibility was assessed by eight Greek manufacturing experts. XGBoost achieved the strongest predictive-maintenance performance, PatchCore achieved the strongest visual detection performance, and EfficientAD offered the best balance between detection quality and latency. The study contributes a modular, deployment-aware, and human-supervised computing artifact for integrating maintenance and quality inspection intelligence in constrained manufacturing settings.