University of Twente Student Theses


Towards deep learning frameworks for the analysis of magnetic flux leakage captures

Stoian, N.A. (2022) Towards deep learning frameworks for the analysis of magnetic flux leakage captures.

This is the latest version of this item.

[img] PDF
Abstract:In the petrochemical industry, one of the most important NDT methods is magnetic flux leakage (MFL), a method which uses the magnetic properties of pipes and storage tanks to detect defects in them. Machine learning frameworks are usually split into two parts in this context: a detection algorithm and a classification algorithm. ROSEN already has a detection algorithm and a deep learning model deployed as part of its MFL scanning tool, the TBIT, however the results obtained so far tend to differ by a large margin from scan to scan. In this case, the project has focused towards improving the second component, the classification model. For the purposes of this project, ROSEN has provided 135000 entries split over 20 datasets captured from 10 different storage tanks scanned over a number of years. This thesis has a two-fold contribution: the first is to try to identify the potential causes behind the discrepancy in the results of the baseline deep learning classification model between the different datasets, and the second is to test multiple generalization techniques in order to decrease the generalization error of the model.
Item Type:Essay (Master)
ROSEN Technology and Research Center GmbH, Enschede, Netherlands
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science MSc (60300)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page