A Physics-Aware Graph Network for Geometry Online Monitoring in Smart Manufacturing

Industrial Internet of Things (IIoT) empowers smart manufacturing by enabling real-time data collection, analysis, and optimization. Wire Arc Additive Manufacturing (WAAM), as a smart manufacturing method, is attracting considerable attention in industry and academia due to its ability to fabricate large com-ponents, whilst allowing material savings. Monitoring the geome-try of the bead is essential for ensuring the quality and consistency of the manufactured components, and even provide feedback to enhance the control of the manufacturing process. In recent years, deep learning-based geometry monitoring approaches have gained significant attention. However, several challenges remain. Existing methods often overlook the physical relationships among process signals, rely solely on data-driven graph constructions, and suffer from poor interpretability due to the black-box nature of neural networks. Additionally, fixed receptive fields in graph convolu-tional networks (GCNs) limit feature extraction and model perfor-mance. To address these issues, this study proposes a physics-aware graph network (PAGN). A sensor network is designed to construct graph data based on based on the physical relationships among signals. Leveraging the interpretability of wavelet trans-form, PAGN integrates a multi-scale filter (MSF) module and an adaptive multi-scale stacking (AMSS) module, enabling multi-res-olution analysis of graph signals. Lastly, the effectiveness of the proposed PAGN is evaluated on the copper and chrome zirconium copper dataset. Experimental results on both datasets demon-strate the effectiveness of the proposed PAGN framework, achiev-ing RMSE reductions of 10.37% and 12.02% and MAE reductions of 8.9% and 5.15%compared to the other methods.