Improving Thermal Anomaly Detection Using Robot Pose Conditioning

Master Thesis (2025)
Author(s)

M.J.H. Owusu (TU Delft - Mechanical Engineering)

Contributor(s)

J. Kober – Mentor (TU Delft - Learning & Autonomous Control)

J.F.P. Kooij – Graduation committee member (TU Delft - Intelligent Vehicles)

Y. van Warmerdam – Graduation committee member (Alliander)

Faculty
Mechanical Engineering
More Info
expand_more
Publication Year
2025
Language
English
Graduation Date
07-08-2025
Awarding Institution
Delft University of Technology
Programme
['Mechanical Engineering | Vehicle Engineering | Cognitive Robotics']
Faculty
Mechanical Engineering
Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Abstract

Conditional variational auto-encoders showed improved anomaly detection abilities over standard variational auto-encoders in literature. This paper explores the effectiveness of robot pose conditioning on thermal anomaly detection in the context of electrical substations. We introduce a multi-modal conditional variational auto-encoder framework, capable of reconstructing thermal images and robot poses. It utilises a multi-objective loss function consisting of mean squared error image and pose reconstruction loss and Kullback-Leibler divergence. Orientation showed to be the most effective conditioning pose, in the context of anomaly detection. A well performing network effectively reconstructs the original assets based on the latent space representation, contains only slightly blurred reconstructions in cases of uncertainty, has a structured latent space as principal component analysis reveals and shows high separability between the distributions of the image reconstruction errors for normal and anomalous samples.

Files

License info not available