Graph transfer learning approach for aircraft assembly line productivity prediction amidst evolving process network

Data-driven machine learning models are widely used to predict productivity in complex assembly lines with notable accuracy and efficiency. However, the evolving nature of assembly processes results in modifications to localized segments of the process network, leading to new scenarios with varying input dimensions and distributions. This makes direct application of existing models impractical, while small-batch production characteristics make retraining for each new scenario unfeasible. To address these challenges, this study proposes a graph transfer learning approach for predicting productivity in aircraft assembly lines with evolving process networks. First, a baseline model is established using Graph Neural Networks, where parameters of manufacturing elements serve as inputs and productivity indicators as outputs. The model is based on a graph representation of the assembly line, with nodes denoting processes, resources and materials. Next, a transfer learning framework is developed to adapt the prediction model structure to changes in the process network and select representative samples for new scenarios. Evaluations on a real industrial case demonstrate that the proposed model excels not only in predicting productivity for the original scenario, but also in rapidly adapting to new process networks after transfer, effectively handling changes in input dimensions and limited data samples.