Application of Graph Neural Networks for Structural Dependency Analysis in Industrial Digital Twins Using the UCI AI4I 2020 Predictive Maintenance Dataset

作者：A. Stupina, S. Kukartseva, M. Stepantsevich · 发表于：International Conference on Industrial Engineering, Applications and Manufacturing · 年份：2026 · DOI：10.1109/ICIEAM69213.2026.11549674

Industrial digital twins represent complex, dynamically evolving systems composed of heterogeneous equipment, technological processes, and information flows. Traditional monitoring and predictive maintenance approaches often treat assets in isolation or rely on linear correlation models, which fails to capture high-order structural dependencies and cascade failure propagation. This paper proposes a graph-based representation of industrial digital twins and introduces a Graph Neural Network (GNN) framework for dependency analysis and early anomaly detection. The manufacturing process is formalized as a directed heterogeneous graph where nodes correspond to physical assets and process stages, while edges encode material, energy, and control dependencies. A Graph Attention Network (GAT) architecture is employed to learn contextual node representations and predict the likelihood of fault propagation across the production line. The model is validated on the real, publicly available UCI AI4I 2020 Predictive Maintenance Dataset, containing 10,000 records of production cycles with equipment failure labels. Comparative experiments demonstrate that the proposed GNN approach outperforms baseline methods (Random Forest, LSTM, and traditional graph metrics) in terms of F1-score ($+12.8 \%$), early detection lead time (+1.9 production cycles), and false positive rate reduction $(-16.3 \%)$. The results confirm that graph-based deep learning significantly enhances the analytical capabilitie...