Anomaly detection and fault diagnostics for underwater gliders using deep learning

Wu, Peng; Harris, Catherine A.; Salavasidis, Georgios; Kamarudzaman, Izzat; Phillips, Alexander B.; Thomas, Giles; Anderlini, Enrico. 2022 Anomaly detection and fault diagnostics for underwater gliders using deep learning. In: Oceans 2021: San Diego - Porto, San Diego, 20 - 23 September 2021. 1-6.

Before downloading, please read NORA policies.

Preview

Text Anomaly_Detection_and_Fault_Diagnostics_for_Underwater_Gliders_Using_Deep_Learning.pdf Download (1MB) | Preview

Abstract/Summary

Underwater Gliders (UGs) (Fig. 1) are a type of Autonomous Underwater Vehicle (AUV) that are being used extensively for long-term observation of key physical oceanographic parameters [1]. They operate remotely at a low surge speed of approximately 0.3ms−1, with deployments of several months [2]. However, developing Near Real-Time (NRT) anomaly detection and fault diagnostics systems for such vehicles remains challenging as decimated sensor data can only be transmitted off-board periodically during operations when the UG is on the surface. As part of an ongoing collaboration, the authors have previously developed anomaly detection systems for UGs via different approaches. In [3], a simple but effective system was developed to detect the wing loss using the roll angle. In [4], system identification techniques were employed to detect changes in model parameters which further successfully deduced simulated and natural marine growth. Anderlini, et al. [5] further conducted a field test to validate a marine growth detection system for UGs using ensembles of regression trees. In [6], the use of a range of deep learning techniques was investigated to achieve over-the-horizon anomaly detection for UGs. In [7], an anomaly detection system based on an improved Bi-directional Generative Adversarial Network (BiGAN) was prototyped to enable generic anomaly detection for different types of anomalies. For UGs operated over the horizon, some faults can only be revealed when the faulty UGs are recovered. Also, it is not clear when the faults developed. Some undetected faults can lead to critical failures and the loss of vehicle and/or data cargo. Therefore, it is essential to understand the actual causes of high anomaly scores during remote monitoring to allow operators to take appropriate mitigations to minimise subsequent risks and maximise the successful delivery of the remainder of the deployment. This paper further compares the results acquired in [7] with other baseline approaches. In addition, a new supervised fault diagnostics method for UGs is proposed. The BiGAN-based anomaly detection system is applied to estimate when the faults are developed, such that the training dataset for the supervised fault diagnostics model can be accurately annotated. The results suggest that the BiGAN-based anomaly detection system has successfully detected different types of anomalies, in good agreement with model-based and rule-based approaches. The supervised fault diagnostics system has achieved high fault diagnostics accuracy on the available test dataset.

Item Type:	Publication - Conference Item (Paper)
Digital Object Identifier (DOI):	https://doi.org/10.23919/OCEANS44145.2021.9705774
Date made live:	17 Jan 2023 13:48 +0 (UTC)
URI:	https://nora.nerc.ac.uk/id/eprint/533874

Actions (login required)

View Item

Document Downloads

More statistics for this item...