High-fidelity data reduction and feature extraction for real-time robotic ultrasonic geometry estimation in nuclear manufacturing

Tools

Dimakos, Angelos and Nicolson, Ewan and Thompson, Christopher and Serjeant, Sam and Tunukovic, Vedran and MacLeod, Charles and Sibson, James and Lines, David and Sweeney, Nina and Loukas, Charalampos and Mohseni, Ehsan (2026) High-fidelity data reduction and feature extraction for real-time robotic ultrasonic geometry estimation in nuclear manufacturing. In: Postdocs in Nuclear Energy 2026, 2026-04-14 - 2026-04-15, University of Strathclyde.

[thumbnail of Dimakos-etal-PiNE-2026-High-fidelity-data-reduction-and-feature-extraction-for-real-time-robotic-ultrasonic-geometry-estimation]
Preview
 Text. Filename: Dimakos-etal-PiNE-2026-High-fidelity-data-reduction-and-feature-extraction-for-real-time-robotic-ultrasonic-geometry-estimation.pdf
Accepted Author Manuscript
License: Strathprints license 1.0
Download (543kB)| Preview

Abstract

Nuclear manufacturing for SMRs requires right-first-time production to minimize repairs. While real-time ultrasonic sidewall fusion monitoring shows promise, with amplitude signal drop from 60% to 0% indicating proper fusion, robotic inspection introduces measurement challenges. Advanced ultrasonic data acquisition methods improve robustness but generate data volumes incompatible with real-time processing (up to 20 million data points). This Babcock-supported research proposes the combination of automated ultrasonic inspection with machine learning to enable rapid weld geometry estimation. It extracts features from compressed data through a three-stage process: (1) identifying the most predictive hand crafted features, (2) augmenting sparse data using oversampling and gaussian process regression, and (3) creating lightweight models (stacked support vector machine and ridge regression), suitable for real-time deployment. This achieves 90% data reduction, maintaining 1.2 mm compression fidelity and computation time below 200 ms using hybrid CPU-GPU architecture. Validated on production-relevant weld geometries (120 mm narrow gap and 15.8 mm V-groove), the approach characterises robotic positioning uncertainties by propagating positioning through linear regression for ultrasonic measurement and could enable autonomous quality control during welding. With in-process defect detection, this framework represents a step toward closed-loop quality control in nuclear manufacturing and could significantly reduce rework cost in next-generation SMRs.

ORCID iDs

Dimakos, Angelos ORCID logoORCID: https://orcid.org/0009-0007-9531-2216, Nicolson, Ewan ORCID logoORCID: https://orcid.org/0000-0002-8174-1665, Thompson, Christopher, Serjeant, Sam, Tunukovic, Vedran ORCID logoORCID: https://orcid.org/0000-0002-3102-9098, MacLeod, Charles ORCID logoORCID: https://orcid.org/0000-0003-4364-9769, Sibson, James, Lines, David ORCID logoORCID: https://orcid.org/0000-0001-8538-2914, Sweeney, Nina ORCID logoORCID: https://orcid.org/0000-0002-4495-4688, Loukas, Charalampos ORCID logoORCID: https://orcid.org/0000-0002-3465-8076 and Mohseni, Ehsan ORCID logoORCID: https://orcid.org/0000-0002-0819-6592;
CORE (COnnecting REpositories)