A lightweight vision-driven control framework for enhancing geometric accuracy in WAAM

• A control framework to enhance geometric accuracy in wire arc additive manufacturing. • A lightweight and efficient geometry measurement method using molten pool images. • Real-time, accurate measurement enabled by integrating multiple vision algorithms. • An effective predictive control strategy enhanced by machine learning techniques. • Two real-world aircraft wing rib parts are fabricated to validate the framework. Wire arc additive manufacturing (WAAM) has emerged as a promising metal additive manufacturing technique, characterized by a high deposition rate, relatively low equipment and operating cost, and strong compatibility with standard arc welding systems. These characteristics make WAAM particularly suitable for fabricating medium to large-sized metal parts, especially in high-end manufacturing sectors such as aerospace and maritime. Enhancing geometric accuracy in these applications is critical, as deviations can lead to low yield rates and compromised product reliability. Current WAAM systems primarily rely on offline process optimization to achieve high geometric accuracy; however, such offline strategies are inherently limited in their ability to respond to dynamic process variations such as sporadic defects, distortions, and material inconsistencies. To overcome these challenges, this study proposes a lightweight and efficient online control framework, which incorporates real-time molten pool image analysis and data-driven intelligent feedback loop control strategies. A single-camera vision module is developed to extract geometric features from molten pool images, enabling accurate estimation of weld bead width and layer height without relying on time-consuming laser scanning. Vector quantized variational autoencoder (VQVAE) and deep learning techniques are employed to enhance feature representation. The control module combines AutoRegressive with eXogenous input (ARX)-based model predictive control (MPC) with support vector regression (SVR), enabling precise adjustment of deposition parameters while compensating for cumulative layer height deviations. The effectiveness of the proposed framework is validated by manufacturing two real-world metal parts from an identical CAD model, where one part is fabricated without control and the other is fabricated with the proposed control framework, resulting in an improved geometric accuracy. The process efficiency is also improved due to the reliance on in-process molten pool images, thereby avoiding the additional measurement time associated with conventional inter-layer profile scanning. Specifically, the geometric accuracy is improved by around 50.8%, and the overall production efficiency is enhanced by 31.5%. This work offers a practical solution for high-precision WAAM fabrication, promoting its broader adoption in diverse manufacturing scenarios.