Deep learning prediction of weld bead geometry during multi-pass multi-layer welding process

초록

Accurate prediction of weld bead geometry is essential for optimizing welding procedure specifications in multi-layer, multi-pass welding. Most existing studies predict only limited geometric parameters such as bead width or height and usually assume a fixed initial shape, without considering variation in the input geometry or the influence of welding position. This work addresses these limitations by proposing a convolutional-neural-network-based method that predicts the full cross-sectional bead geometry for arbitrary initial geometries. A tip-referenced region-of-interest extraction technique is introduced to reduce geometric fluctuations. A comprehensive dataset was acquired using a line-scanner sensor through a series of welding experiments on three common materials to ensure dataset diversity. The model is evaluated against linear regression, random forest, and multilayer perceptron baselines, yielding a determination coefficient of 0.94 and a geometric overlap of 82.4% on the validation set. Analysis of the welding parameters indicates that bead-geometry dimensions generally exhibit a significant inverse relationship with travel speed, while showing positive correlations with wire-feed speed and welding currents. The model is successfully applied to predict welding geometry for multi-layer, multi-pass welding processes, demonstrating its potential as an effective tool for data-driven welding-procedure development. © 2026 Elsevier Ltd.

키워드