Thesis

Towards in-process robotic ultrasonic weld geometry estimation and lack of sidewall fusion monitoring

Downloadable Content

Download PDF
Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2026
Thesis identifier
  • T18125
Person Identifier (Local)
  • 202291562
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • Robotic arc welding is central to the automation of modern manufacturing, yet the integration of volumetric quality assurance directly into the manufacturing workflow remains an open challenge. This thesis presents four distinct contributions for robotic phased array ultrasonic inspection that enables lack of sidewall fusion (LOSWF) monitoring and continuous weld bevel geometry estimation during and after deposition. Four interconnected contributions are established. First, it is demonstrated for the first time that LOSWF can be detected in real time using ultrasonics during active Gas Tungsten Arc Welding as B-Scan signal dropout: proper sidewall fusion causes longitudinal wave absorption into the melt, while LOSWF maintains signal stability, providing a binary, physically interpretable proxy for joint integrity, as shown in Chapter 4. Second, a comprehensive experimental framework is presented to characterise and reduce robot-induced measurement error across a range of robotic parameters; roll, lateral offset, and coupling force; identifying a critical transition zone that governs the optimal selection between linear and sector scan modalities, as described in Chapter 5. Third, a low-latency signal and image processing pipeline is introduced in Chapter 6, combining Area Under Curve (AUC)-based sparsification with Graphics Processing Unit (GPU) accelerated Histogram of Oriented Gradients (HOG) estimation to achieve Full Matrix Capture data compression exceeding 88% while extracting bevel orientation to within 0.4° in under 250 ms. Finally, a machine learning pipeline is presented in Chapter 7, employing data augmentation; using Gaussian Process Regression (GPR) to augment scarce training data and a stacked ensemble of Support Vector Machine and polynomial regressor for inference; that achieves bevel length estimation within 1.2 mm, consistent with ISO 13920 weld preparation tolerances. Together, these contributions demonstrate that high-resolution volumetric ultrasonic data can be compressed, processed, and interpreted within the latency constraints of industrial robotic path planning. This work establishes a foundation for closed-loop, self-correcting robotic welding systems capable of providing real-time quality assurance in safety-critical applications such as shipbuilding and nuclear fabrication.
Advisor / supervisor
  • MacLeod, Charles Norman
  • Sweeney, Nina
  • Mohseni, Ehsan
  • Loukas, Charalampos
Resource Type
DOI

Relations

Items

Thumbnail Title Date Uploaded Visibility Actions
file details PDF of thesis T18125 2026-09-04 Public Download