Thesis

Unsupervised data-driven analysis of ultrasonic inspection data

Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2022
Thesis identifier
  • T16549
Person Identifier (Local)
  • 201552659
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • Ultrasonic inspection is a key part of condition-based maintenance in the nuclear power industry, and it is widely used for flaw detection and characterisation in critical components. The analysis of ultrasonic inspection datasets is a time-consuming and complex task commonly carried by expert analysts. The need for robust and efficient interpretation of inspection data is especially important now that advancements in ultrasonic hardware enable the capture of high-resolution data in far greater rates and volumes than in the recent past, driving research efforts to create automated procedures for signal classification and flaw detection. This thesis provides new data-driven approaches for analysing large volumes of ultrasonic inspection data in an unsupervised manner, without requiring individually labelled ultrasonic signals. The first method utilises the DBSCAN clustering algorithm at its core, and along with the proposed subsampling method, twostage clustering procedure, and automated parameter estimation procedure, it provides efficient flaw detection without a pre-defined state of normality. The analysis is then extended to large-scale ultrasonic inspection datasets that offer challenges both in terms of size, but also heterogeneity across the different neighbourhoods of the inspected surface. The proposed method utilises a neighbourhood-based transformation of the signals’ variability measures as a means of creating a more homogenous feature space that allows for distance metric that is relevant across the surface, offering significant computational benefits. The application of the proposed methods is focused on pressure tubes, a critical reactor asset of CANDU reactors, and is tested across a large number of real-world datasets showing satisfactory detection rates and efficient performance.
Advisor / supervisor
  • West, Graeme
Resource Type
DOI
Funder

Relations

Items