Abstract
Nuclear fusion, the process that takes place in stars is a potentially game-changing source of low-carbon energy to meet future demands. It is, however, considerably challenging to sustain fusion reactions on earth. One of the critical barriers to the commercialisation of fusion energy being the ability to efficiently maintain and inspect fusion reactor systems.
Detection of issues in plant is critical for both productivity, in terms of maximising up-time of plant, as well as for regulatory and safety related functions. Detection of general or unspecified anomalies in industrial environments using camera images is a challenging task, and one which is the focus of much current research.
Many studies related to anomaly and defect detection are application focused, limiting wider applicability, and demonstrate results on domain-specific datasets which are often not widely available, and hence it is difficult to directly re-apply, or compare performance of new techniques.
This work introduces the concept of Regenerative Anomaly Detection, whereby images taken in an environment are regenerated using a learned model of what the environment normally looks like, and the comparison of the two images is used as a metric of anomalousness. It explores the use of Generative Adversarial Networks (GAN)s, a deep learning-based architecture in performing generalised visual anomaly detection within the context of fusion remote maintenance. A framework for performing these operations and characterising performance is established.
The work then focuses on three areas, where novel techniques are proposed and evaluated in order to improve capability in terms of performance through establishment of optimal image comparison metrics, combining techniques for selective in-component anomaly detection, and optimising image regeneration techniques by combining predictive and numerical approaches.
Results and conclusions across all of these areas are presented, and then finally summarised, with a description of possible future directions of the work, and a discussion around links to other work including that in the worlds of perceptual and cognitive psychology.