Saving lives, improving ocean data collection, and supporting green energy - how we can all benefit from safe unmanned marine systems?
At present, in the absence of current regulatory guidelines, unmanned marine systems (UMS) are piloted and monitored by experts, keeping operational costs high and limiting the scale of fleets that can be deployed simultaneously. This project will increase the safety of UMS by helping the vehicles identify the cause of their adverse behaviour.
The team is developing a generalised automated detection and diagnosis protocol for adverse vehicle behaviour to inform and improve the management of UMS. They will:
- develop a monitoring tool that detects adverse UMS behaviour and alerts the user
- use data from the operation of actual devices to develop a classification tool that correctly detects and diagnoses unexpected behaviour as due to system malfunctions or environmental disturbances
- validate the tools using autonomous underwater gliders
- work with Lloyd’s Register to inform regulations for UMS
In the absence of accepted international terminology for marine autonomous systems (MAS), new vocabulary for condition monitoring of MAS has been introduced by the team. This is designed to be general and transferable to any RAS technology, although failure modes are specific to each design. A hierarchical procedure has been designed to autonomously label RAS time-series data as either standard or anomalous. Human feedback is used to correct erroneous tagging.
Furthermore, specific datasets of deployments of a range of MAS have been selected to be used in the remainder of the project, enabling the design of transferable solutions. Remote-sensing data will complement the datasets to capture the environmental disturbances during the deployments.
The team has been studying the definition of sensing requirements and has introduced guidance for the definition of sensing requirements. They have also undertaken work on the identification of sensing deviations for robotic and autonomous systems, with a focus on marine autonomous systems. In particular, they have designed a novel method using a bidirectional generative adversarial network (Bi-GAN) with added hints for smart anomaly detection for RAS.
After being trained with data from healthy baseline RAS, the system recognises anomalous behaviour by tracking sensing deviations. The autonomous detection method has been developed with deployment datasets of real underwater gliders, showing excellent performance on the unseen test set. Hence, the system will enable real-time, over-the-horizon condition monitoring of marine autonomous systems (MAS) to achieve round-the-clock autonomous operations.
Fault-diagnostics methods for marine autonomous systems have been introduced and developed using supervised and semi-supervised learning. The systems were tested with available data from glider deployments, which is being enriched with additional publicly available datasets as well as future demonstration field tests. The previously developed anomaly detection algorithms are being implemented on the over-the-horizon control and command (C2) infrastructure of the National Oceanography Centre for future testing.
Presentations and papers
- Wu, P., Harris, C., Salavasidis, G., Lorenzo-Lopez, A., Kamarudzaman, I., Phillips, A., Thomas, G., and Anderlini, E. "Unsupervised anomaly detection for underwater gliders using generative adversarial networks" in Engineering Applications of Artificial Intelligence, Volume 104, September 2021. Github repository linked in the paper.
- E Anderlini, “Machine learning: friend or foe?”, Royal Institution of Naval Architects (RINA) London Branch, 10th December
- 13/01/2021, “Machine Learning: friend or Foe?”, National Oceanography Centre: Innovation Centre, audience: 25 people, representatives from the oceanographic and hydrographic companies within the Innovation Centre and marine roboticists of the National Oceanography Centre,
- 24/02/2021, “Identification of the dynamics of biofouled underwater gliders”, EUMarineRobots: “Coffee with EUMR”, audience: 50 people, marine roboticists,
- 04/03/2021, “Artificial Intelligence ABCs: Automation”, Science Innovation Union: “Artificial Intelligence ABCs: Automation, Blockchain, Communication”, audience: 250 people, doctoral students and early career researchers.