Process Optimization of Robotic Laser Weld Brazing for AI-friendly Dataset Generation to Enable Smart Ultrasonic Inspection of Braze Joint Quality in Industry 4.0
Loading...
Date
Journal Title
Journal ISSN
Volume Title
Publisher
Elsevier
Abstract
This study establishes a custom-designed process window for a robotic laser weld brazing (LWB) process to control defects in braze seams used in automotive applications. This window is essential for producing braze seams of varying quality that meet custom specifications, enabling the development of an industrial-level ultrasonic testing (UT) defect detection system, with artificial intelligence (AI) automating UT data processing and measurement extraction. To obtain the required custom samples for training the AI model, a novel wire-adjusted heat input approach was defined, expounding a clear correlation between heat input and defect formation, unlike the traditional nominal heat input, which lacked consistency under varied wire feed speeds. This novel approach enabled precise delineation of heat input-related boundaries to distinguish defect-free conditions from those with lack of adhesion (LoA), pores, base metal melting (BMM), and non-conforming braze geometry. The process window was validated using micro-computed tomography (μCT) to scan and analyze defect distribution along the entire braze seam length. A semi-automated segmentation method was developed to facilitate statistical analysis of defect evolution, ensuring reliable classification of the defects. UT signatures for both defects and seam geometry were evaluated, and reliable signal feedback for defect detection was substantiated. A deep learning model, trained using μCT-derived ground truth, predicted seam geometry directly from UT data, achieving high alignment with μCT measurements across varied geometries and defects. The process window establishes the groundwork for both UT detection and AI-based geometry prediction, while extending AI-based detection to other defects warrants future investigation.
Description
The final publication is available at Elsevier via https://doi.org/10.1016/j.optlastec.2026.117033. © 2027. This manuscript version is made available under the CC-BY-NC-ND 4.0 license http://creativecommons.org/licenses/by-nc-nd/4.0/