Iterative Vision-Based Model to Measure the Contact-Tip-Working-Distance for WAAM Interlayer Control

Abstract

Dimensional accuracy in Wire Arc Additive Manufacturing is frequently compromised by stochastic geometric variations, primarily layer height undulations and humping, which cause uncontrollable fluctuations between the build-up component and the welding wire, commonly referred to as Contact Tip-to-Work Distance (CTWD). This missdistance leads to arc instability and insufficient melt pool shielding, degrading final component quality. Therefore, this paper aims to develop an iterative vision-based model to detect and measure the CTWD losses by monitoring the model decay that triggers a continuous learning loop with a High-Performance Computing system, enabling the model to be retrained and updated to adapt to environmental changes, such as reflections, spatter, or new robot trajectories. This model has been quantized to run on an industrial PC for low-latency inference, while challenging frames from the welding camera are forwarded to an edge device for operator data annotation. This vision-based approach significantly improves control system efficacy by providing a proactive, measurable feedback signal for inter-layer adjustment decisions (repeat, skip, or proceed), thereby maintaining layer geometry and ensuring the long-term reliability of the WAAM process in dynamic manufacturing environments. As a result, a single-shot detector is selected as the object detection model, which weighs 8MB and runs at 60 frames per second.

Keywords: Additive manufacturing, Object detection model, Contact tip-to-work distance, ML models, WAAM

DOI: 10.54941/ahfe1007189