An explainable multi-layer graph attention network for product completion time prediction in aircraft final assembly lines

Abstract

Predicting product completion time (PCT) is a critical challenge in aircraft manufacturing systems, especially for make-to-order production. This necessitates manufacturers to comprehensively analyze operational state features, including task completion, resource allocation, and material supply, to estimate delivery dates effectively. With the increasing availability of production site perception, data-driven methods for PCT prediction have gained significant attention. However, the coupled interactions among various manufacturing elements, combined with the demand for real-time scheduling in digital twin scenarios, have limited the accuracy and explainability of traditional black-box predictive models. To address these challenges, this paper proposes an explainable multi-layer heterogeneous graph attention network (M-HGAT) customized for predicting PCT in the aircraft final assembly line (AFAL). First, a heterogeneous graph representation method is introduced to model the aircraft assembly status, focusing on the interactions among assembly tasks, materials, and workers. Then, a two-layer state feature aggregation neural network is designed to learn the mapping relationship between the target PCT and input features, incorporating logical and demand constraints among various elements inherent in the aircraft assembly process. Finally, the accuracy and explainability of the proposed model have been validated through an industrial case study focused on PCT prediction. Compared to four benchmark predictive models, the proposed model achieves superior predicted results, reducing the root mean square error by 48 % compared to the best benchmark. Furthermore, the explainability of the M-HGAT is demonstrated through its ability to identify key manufacturing elements and bottleneck assembly stations by analyzing attention weights within the neural network, which provides valuable insights for production managers to optimize AFAL operations and enhance production efficiency.