Automated Fiber Placement (AFP) is a critical process in composite manufacturing, where precise fiber tow placement is essential for achieving high-quality and high-performance engineering components. However, deviations in process variables frequently lead to defects such as gaps and overlaps, which can compromise structural integrity. While various monitoring techniques exist, accurately predicting and understanding the formation of these defects from complex sensor data remains challenging. This work introduces a novel application of a Transformer-based deep learning architecture to enhance the estimation of gap widths in AFP. Leveraging a publicly available industrial AFP dataset, our methodology incorporates a customized positional encoding scheme to effectively integrate the critical spatial context of the tow layup process. The model’s predictive performance was evaluated, achieving a Mean Absolute Percentage Error (MAPE) of 1.04% and an R-squared (R2) value of 0.9143, demonstrating its capability for accurate gap width estimation. Furthermore, SHapley Additive exPlanations (SHAP) analysis was employed to assess the complex interplay between sources of manufacturing process variation. This study establishes the Transformer architecture as a promising and interpretable data-driven tool for AFP process monitoring. The results serve as a proof of concept for attention-based virtual metrology, offering a pathway towards deeper process understanding and defect mitigation.
Cardoso et al. (2026) studied this question.