Vision-based penetration monitoring method for climbing helium arc welding using deep learning with multivariate time series forecasting

Abstract

Climbing helium arc welding is one of the key technologies in aerospace, military and other high-end equipment industries. Due to the time-varying position and pose of molten pool, the heat and mass transfer process during such welding are complicated and pose a challenge to ensure molten pool stability and weld penetration consistency. Recently, the weld penetration prediction based on molten pool images has shown great application potential in online weld monitoring. However, to accurately predict weld penetration is still a difficult task because of the high similarity of molten pool surface morphology under different weld penetration, especially during non-flat welding process of medium-thickness aluminum alloy plates. This paper proposes a novel vision-based two-phase framework for weld penetration prediction and applies it to climbing helium arc welding. The framework consists of multi-level characteristics extraction phase and deep time series forecasting phase. Firstly, the Deeplabv3+ is used to segment multi-region from molten pool image sequences, extracting the Molten Pool Region (MPR) and the Exposed Aluminum Liquid Region (EALR) within each image. Then, a time series comprising the extracted multi-level characteristics is constructed, and a Kansformer model is subsequently utilized to predict the weld back width based on this characteristic time series. The validity of the proposed method was verified by using data retained from an actual production platform of 2219 aluminum alloy rocket propellant canisters, and the experimental results showed that it was superior to existing methods. Moreover, its generalization ability is verified under varying process parameters, and the average inference speed of a single frame can reach 17.12 fps. In particular, SHAP analysis is utilized to explain the key role of the extracted dynamic characteristics in weld back width prediction.