Author: Chris Nelson, Data Scientist
There is an ever-growing acknowledgement that risks to subsea cable infrastructure are increasingly due to malicious and targeted acts from international adversaries. A cursory glance at the news highlights numerous examples of targeted activities that increase risks to subsea cables, from foreign vessels surveying cables in UK waters, to potentially deliberate acts of sabotage through anchor dragging over long distances in the Baltic Sea.
In the UK, Defence Secretary John Healey has vocally called out these activities and the current government has pledged significant funds to developing new methods to track both surface and sub-surface vessels around subsea cables. A high proportion of these funds are targeted at hardware development, including unmanned underwater drones which can monitor along cables. Of course, as the number of cables increases, the number of drones required will also increase leading to ever-expanding costs meaning more passive methods that can be more easily scaled are also needed.
Another method for monitoring cables that may be more scalable, as highlighted by a recent Parliamentary Joint Committee report, is through the use of Distributed Acoustic Sensing (DAS). With DAS, it is possible to turn each subsea cable into thousands of individual hydrophones, each of which are sensitive to engine and propeller noise from nearby ships. This, in theory, allows the cables to monitor themselves without the need for external hardware in the water. To make this work in near real-time, though, requires a high degree of knowledge about DAS and advanced data science capabilities. Our expertise in both of these areas places Indeximate at this new frontier of subsea cable defense.
Vessel detection from DAS data has challenges in data handling (for both development and implementation) and performance limitations in rough weather.
Historical Approaches – Classic Acoustic Techniques
Traditional methods for vessel detection using DAS data typically rely on at-the-edge processing using simple, single-band, acoustic energy detectors. In this scenario, a large increase in acoustic energy at a specific location along a cable in the right acoustic band can be used to identify an ‘event’ which, if it matches specific criteria, can be classified as a vessel, or trawling, or an anchor strike. This can be reasonably effective, with numerous examples in the academic literature and elsewhere of specific vessels being identified and tracked.
However, this method is also susceptible to failure. Increases in environmental noise along the cable, potentially due to increased wave heights during adverse weather, can quickly begin to drown out the acoustic signatures of vessels. In our experience, deterioration in performance can occur at significant wave heights of just 1 m. Above this value, algorithms acting at-the-edge can either return a plethora of false positives, or must be switched off leading to the missed detection of numerous true positives.
At-the-edge processing is also a limiting factor for both historical forensic analysis and the development of advanced analytical techniques, such as machine learning approaches. DAS is extremely data intensive with TBs of data being collected in a single day along a single cable. This data volume also poses a challenge to automated analysis techniques with most studies being lab based, carried out on “ideal” sections of cable and on short recordings that can practically be handled. Without real-time compression, this data must be over-written within several days. In terms of forensic analysis, operators must, therefore, be very quick in identifying vessel related issues in order to extract the data before it is deleted. In terms of advanced analytics, these massive data volumes mean it is difficult to collect a dataset large enough to develop effective machine learning models. These combined effects are clearly limiting with regards the overall success of at-the-edge approaches, particularly in the realm of critical national asset monitoring. As such, new approaches are required.
Historical Approaches – Machine Learning
Recent academic studies have also highlighted that Machine Learning techniques that exploit visualisations of acoustics or machine based analysis of spectral content offer an alternative to the classic acoustic approach, however again studies have struggled to gather sufficient volume of training data (due to the same data intensity issue) and have to date not been fielded in an automated manner on practical computing hardware.
The Indeximate Approach
The foundation of the Indeximate approach to vessel detection in DAS data lies in our proprietary data compression techniques. Using these techniques, we can upload data to the cloud where it can all be stored. This storage builds up a repository allowing forensic analysis to take place weeks or months after any event, as well as, importantly, a database of vessel signatures that can be used for the development of machine learning models. Working with the Science and Technologies Funding Council Hartree Centre (UKRI), Indeximate successfully developed and deployed such a model that can be applied to our near real-time cloud data stream, working in a fully automated deployment with a a web based GUI.
A key benefit of machine learning over classical approaches is that, with sufficient training, the returned models can identify real vessel signatures in even the most adverse weather conditions. On datasets where historical approaches started to struggle at 1m significant wave heights, our machine learning models continued to perform well at significant wave heights up to and above 4m. This is a remarkable leap in performance. Essentially, our machine learning models can pick out the vessel signals from even the noisiest data, independent of whether the vessel’s AIS transponder is on or off. Not only this, but these models can return valuable meta-data, such as the distance of the vessel to each 10 m block along the cable which, in turn, allows us to triangulate the vessel positions through time returning tracking with a higher cadence than AIS.
The successful development of these Indeximate models moves DAS vessel detection from a technique that works in a laboratory ondata with favourable environmental conditions are available, to something that can work in (near) real-time in an operational environment. These models can be used across a range of areas, from identifying on-going careless fishing or shipping practices close to subsea cables, to tracking vessels that deliberately have their AIS transponders switched off for potentially nefarious purposes. Whether you are worried about the risks to a single cable, or more broadly across critical subsea national infrastructure, Indeximate are here to help. Indeximate are currently fielding this capability operationally as a proof of concept TRL-6 service.
If you want to learn more, feel free to reach out at: https://indeximate.com/contact-us/
Figure 1: Cable infrastructure in northwest Europe (source KIS-ORCA). Red denotes telecommunications cables, green denotes renewable energy infrastructure cables, and orange denotes both national and international inter-connectors. The scale of monitoring required is clearly enormous.
Figure 2: A hyper-spectral time (x-axis) distance (y-axis) image displaying a vessel transit in DAS data. Predominantly blue areas are dominated by waves on the sea surface, whilst green and red areas are dominated by engine and propeller noise from vessels, respectively. All three are used in our algorithms.
Figure 3: Difference in detections between a machine learning model (top panel) and a classical acoustic method (bottom panel) at a single time along a cable in the Irish Sea plotted in our web-based Scattersphere for Vessel Detection GUI. The dark green line plots the cable route whilst the purple (top) and green (bottom) circles denote three minutes of machine learning and classical acoustic detections, respectively at a given point in time. The grey dots denote the most recent AIS reported locations of the two vessels in question and the grey lines plot their paths within the last hour. The machine learning model readily detects both vessels, whilst the classical acoustic method can only detect the right most vessel.
