Predicting and Modeling Perturbation Propagation in Shipbuilding Workshops with Attention-based GNN and Cellular Automata

Abstract

Dynamic and highly resource-coupled shipbuilding workshops are highly susceptible to cascading effects stemming from local perturbations, often leading to global scheduling failures. While Digital Twin (DT) technology offers a solution, most existing research focuses on "reactive" rescheduling after a perturbation occurs, lacking the capability for "proactive" prediction and modeling of the perturbation propagation process. To bridge this gap, this paper proposes a novel hybrid model (AGNN-CA) that integrates an Attention-based Graph Neural Network (GNN) with Cellular Automata (CA), deeply embedded within a DT framework. First, the manufacturing workshop is abstracted as a heterogeneous graph, where nodes (workstations, tasks, materials) and edges (process flows, material flows) possess multi-dimensional attributes. An attention-based GNN is employed to learn the complex spatio-temporal dependencies within the graph, enabling accurate prediction of the initial perturbation state at individual workstations. Subsequently, these predictions serve as the initial conditions for a Cellular Automata model, where each cell represents a manufacturing resource. The state transition rules of the CA are defined by process constraints and resource competition dynamics, simulating the propagation process and impact scope of the perturbation throughout the workshop network. Experiments based on real-world ship block assembly workshop data show that the proposed model achieves an accuracy of 94.2% in predicting perturbation propagation paths. More importantly, it identifies affected bottleneck stations 25% earlier than baseline models (e.g., pure simulation or traditional machine learning models), providing a critical time window for proactive scheduling decisions. This research presents a new paradigm of data-driven and mechanism-model fusion for disturbance management in complex manufacturing systems.