Paper deep dive
Fully Unsupervised Detection of Physical Contacts on Subsea Cables via State-of-Polarization Monitoring
Agastya Raj, Alvaro Doval, Tian Tian, Steinar Bjørnstad, Marco Ruffini
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 97%
Last extracted: 7/5/2026, 10:38:28 AM
Summary
The paper presents the Fast-Slow DSVDD, a fully unsupervised deep learning detector designed for continuous State-of-Polarization (SoP) monitoring on subsea cables. The model uses a dual-head architecture (fast and slow) to capture both impulsive transients (2â100 Hz) and slower environmental dynamics (below 2 Hz). In a 92-day field trial on the LowestoftâLista subsea cable, the model successfully ranked all five confirmed trawler contact events within the top 13 of 122,174 recordings, outperforming vanilla DSVDD and classical STA/LTA methods. The framework also discovered additional candidate events corroborated by DAS and AIS data.
Entities (7)
Relation Signals (4)
Fast-Slow DSVDD â detects â Trawler Contact
confidence 100% ¡ The model ranks all five confirmed trawler contacts within the top 13 of 122,174 recordings
State-of-Polarization (SoP) â monitoredon â LowestoftâLista subsea cable
confidence 100% ¡ We use 92 days of continuous SoP recordings from the LowestoftâLista cable
Fast-Slow DSVDD â monitors â State-of-Polarization (SoP)
confidence 100% ¡ We present a fully unsupervised Fast-Slow DSVDD detector for continuous State-of-Polarization monitoring
Trawler Contact â causesdamageto â LowestoftâLista subsea cable
confidence 90% ¡ trawler fishing and anchor dragging are leading causes of cable damage
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:We present a fully unsupervised Fast-Slow DSVDD detector for continuous State-of-Polarization monitoring on a deployed subsea cable. Trained without event labels, it ranks all five confirmed trawler contacts within the top 13 of 122,174 recordings and surfaces additional corroborated cable-contact events.
Tags
Links
- Source: https://arxiv.org/abs/2607.01484v1
- Canonical: https://arxiv.org/abs/2607.01484v1
Trouble viewing inline? Open PDF directly â
Full Text
18,221 characters extracted from source content.
Expand or collapse full text
Fully Unsupervised Detection of Physical Contacts on Subsea Cables via State-of-Polarization Monitoring Agastya Raj (1) , Alvaro Doval (2) , Tian Tian (1) , Steinar Bjørnstad (2) , Marco Ruffini (1) , (1) School of Computer Science and Statistics, IRIS Research Group, ADAPT Research Centre, Trinity College Dublin,Agastya.Raj@tcd.ie (2) Tampnet AS, Stavanger, Norway. Abstract We present a fully unsupervised Fast-Slow DSVDD detector for continuous State-of-Polarization monitoring on a deployed subsea cable. Trained without event labels, it ranks all five confirmed trawler contacts within the top 13 of 122,174 recordings and surfaces additional corroborated cable-contact events. Š2026 The Author(s) 1 Introduction Subsea fibre-optic cables carry over 97% of inter- continental data traffic, and trawler fishing and an- chor dragging are leading causes of cable damage, with dragged-anchor incidents alone accounting for 30â40% of offshore cable faults [1]. Distributed Acoustic Sensing (DAS) and State of Polarization (SoP) fibre sensing technologies mitigate these risks by monitoring vibrations and physical dis- turbances along the cable [2].DASlocalises ap- proaching trawlers [3], but suffers from saturation effects and limited dynamic range for strong sig- nals such as direct impacts [4].SoPmonitoring provides a complementary approach [5, 6]: it can be extracted directly from existing coherent re- ceivers at no marginal hardware cost, does not saturate during strong motions, is compatible with inline amplifiers [7], and produces unique signa- tures for distinct physical impacts [8]. A recent long-term field trial on the Lowest- oftâLista subsea cable established thatSoPre- sponds to real trawler contacts, anchor drags, and environmental forces over multi-month observation periods [9]. However, this was performed manu- ally by visual correlation ofSoPwaveforms, and does not scale to network-wide real-time moni- toring. Furthermore,SoPmonitoring through live transmission systems is impacted by environmen- tal and equipment noise, within which physical contacts are rare and subtle. This motivates a Machine Learning (ML) approach over classical thresholding. However, the absence of labelled event data from real field trials has largely confined existingML-basedSoPdetection to controlled set- tings: Supervised classifiers trained on labelled ex- amples achieve high accuracy on short sequences with predefined event types [10â13], and semi- supervised methods relax this to one-class fits on 1 This paper is a preprint of a paper accepted in ECOC 2026 and is subject to Institution of Engineering and Technology Copyright. A copy of record will be available at IET Digital Library Fig. 1: Experimental Setup labelled baselines [14], but both evaluate on con- trolled testbed scenarios rather than continuous deployed data. Automated detection on deployed cables has been demonstrated in terrestrial set- tings [7, 15, 16] but relies on supervised training. In this work, we transition from manual and su- pervised approaches to fully unsupervised event detection on continuous long-durationSoPdata from a deployed subsea cable. We use 92 days of continuousSoPrecordings from the Lowest- oftâLista cable (Tampnet, North Sea)-the same system used in our previous field trial [9], compris- ing 122,174 one-minute recordings at 44.1 kHz. We evaluate a deep one-class detection model under unsupervised conditions: no trawler labels are used at any stage of training, model selection, or hyperparameter tuning. The model ranks all ground-truth confirmed trawler contacts within the top 13 of the full 122,174 archive, where lower ranks indicate recordings judged more anoma- lous by the detector. Beyond the 5 logged events, the framework produces additional recordings not previously identified during manual reviews of SoPdata. Post-hoc review against theDASand Automatic Identification System (AIS) records con- firmed crossings in which the vessel disabled its AIS transponder during cable crossing. These results demonstrate fully unsupervised detection on continuous deployed cable as a low- cost practical approach, while uncovering signif- icant events missed by manual monitoring even under diverse event characteristics. Field Trial and Data Pipeline Data was acquired from the Tampnet subsea com- munication system connecting Lowestoft, U.K., to Lista, Norway (Fig. 1) using the same configuration as [9]. A continuous-wave (CW) laser was injected arXiv:2607.01484v1 [cs.NI] 1 Jul 2026 Fig. 2: Detection framework of the Fast-Slow DSVDD model. The input SoP recording is processed by fast and slow heads, with four dilated convolutional branches. Branch-wise DSVDD distances are fused into a single recording-level anomaly rank. Fig. 3: Anomaly ranks of the five confirmed trawler contacts across the 122,174-recording archive. The model scores each recording by how unusual it is relative to normal cable activity; rank 1 corresponds to the most anomalous recording. as an alien wavelength at Lowestoft, multiplexed with live traffic through an amplified Wavelength Division Multiplexing (WDM) transmission system spanning passive submarine cable segments. At Lista, theCWwavelength was demultiplexed to a Polarising Beam Splitter (PBS) unit measuring the relativeS 1 Stokes parameter at 44.1 kHz with 16-bit resolution.DASwas operated on a paral- lel dark fibre from Lowestoft, covering 120 km of the cable without passing through inline amplifiers; DASdata is not used in the detection framework but provides independent corroboration for events discussed in Section 4. We analyse 92 days of continuous recordings from this system (JuneâAugust 2025), compris- ing 122,174 one-minute stereo FLAC files. The sensing observable isS 1 =V 1 âV 2 , the differential output of the twoPBSchannels. During this pe- riod, five trawler physical contacts were confirmed by manual cross-referencing ofSoPtransients, DASwaterfall signatures, andAISvessel-tracking records. These five events, spanning from sub- second (E1, 0.5 s) to sustained (E5, 10.6 s), are summarised in Fig. 3 and its accompanying table. These five events constitute the sole ground-truth for all detection results; no labels are used to train, select, or tune any model. Detection Framework Physical contact events on subsea cables are rare, brief, and unlabelled in operational settings. In our 92-day archive, five confirmed trawler con- tacts occupy a combinedâ25 seconds of sig- nal across 122,174 one-minute recordings, cor- responding to a class prevalence below 0.004%. Supervised classification is infeasible at this la- bel scarcity, while semi-supervised approaches assume a stationary reference distribution that months of deployed-cable data cannot provide. Standard unsupervised anomaly detectors such as one-class SVMs, Isolation Forests, autoen- coders, and Deep Support Vector Data Descrip- tion (DSVDD) [17] operate at a single temporal scale. TheSoPsignal in this setting, however, contains two physically different anomaly regimes: trawler contacts produce impulsive transients in the 2â100 Hz band lasting 0.5â10 seconds [9], while environmental and equipment dynamics pro- duce slower variations below 2 Hz. A single-scale detector must therefore prioritise one regime at the expense of the other. Reliable detection requires a model that learns normal behaviour separately at the characteristic timescales of both regimes. We extendDSVDDto a dual-head architec- ture, with each head designed for a distinct regime (Fig. 2). The fast head bandpass filters the input to 2â50 Hz and operates on 1-second windows to capture impulsive transients, while the slow head filters to 0.1â2 Hz and operates on 10- second windows for slower dynamics. Within each head, four parallel dilated convolutional branches provide multi-resolution coverage. The dilation schedules are matched to the regime: the fast head uses1, 3, 9, 27and the slow head1, 9, 81, 729, spanning receptive fields from milliseconds to the full 10-second window. Each branch is associated with a 32-dimensional hypersphere centre, giving eight detection channels. The fast head applies average pooling across 119 overlapping windows, whereas the slow head applies max pooling across 11 windows so that a single anomalous segment can dominate the output. These responses are aggregated into recording-level anomaly scores and fused into a single archive-wide ranking. Each head is trained on 200,000 uniformly sam- pled windows from the archive using an unsuper- visedDSVDDobjective that encourages embed- Fig. 4: Logged and newly surfaced cable-contact events. Panel (a) shows a logged trawler contact, while panels (b)â(d) show three additional high-ranked candidates from the Fast-Slow DSVDD model. In each panel, the left subfigure shows the 30 s SoP waveform and the right shows the croppedDASwaterfall. ClearDASsignatures appear for the 3 June and 17 August events, while none is observed for the 10 June event within the whole 120 km interrogated section. dings of training windows to lie close to a fixed hypersphere centre. For branchb, with encoder Ď b and centrec b , the anomaly score assigned to a windowxis the squared embedding distance d b (x) =âĽĎ b (x)â c b ⼠2 , which is minimised dur- ing training over unlabelled archive windows using Adam (lr= 10 â4 , 8 epochs). At inference, each branch assigns a recording-level score given by max w d b (w) across that recordingâs windows. The final anomaly rank of a recording is then defined as the minimum across the eight branch-specific ranks. A recording is therefore flagged if it appears anomalous in any one detection channel. Results We compare Fast-SlowDSVDDagainst two base- lines: an Short-Term Average (STA)/Long-Term Average (LTA) trigger (STA = 50 ms, LTA = 5 s, bandpass 0.5â20 Hz onS 1 ), representative of classical transient-detection approaches [18]; and vanillaDSVDD, a single-head, single-scale variant of our architecture operating on 1-second windows without the fast-slow band split. We evaluate each method by sorting the full 122,174-recording archive by anomaly score and reporting the ranks of the five confirmed trawler events. Rank 1 corresponds to the recording the method considers most anomalous, so lower ranks indicate better detection performance. A method that places all five events within the topKrequires an operator to review at mostKrecordings to achieve complete detection. Each ground-truth event is matched to its containing recording using aÂą10 s containment interval around the logged timestamp. We summarise performance using the worst-of-5 metric, defined as the highest anomaly rank among the five confirmed events. Fig. 3 and the accompanying table summarise the per-event anomaly ranks together with the cor- responding event durations. Fast-SlowDSVDD places all five confirmed trawler contacts within the top 13 of the archive, with the sustained 10.6- second event E05 ranked 1st and the sub-second, hardest-to-detect event E03 ranked 13th.STA/LTA achieves worst-of-5 = 91, with its failure concen- trated on the shortest events, including E03 at rank 91. VanillaDSVDDachieves worst-of-5 = 1,219, failing on E05 (rank 1,219), the sustained contact that its fixed 1-second window scale cannot re- solve. The remaining four events are ranked within the top 30 by both baselines, indicating that sim- pler methods can recover short impulsive contacts, but not with the same consistency across event durations. What distinguishes Fast-SlowDSVDD is consistent performance across the full range of event morphologies, from sub-second impulses to multi-second sustained contacts. At the operating point where all five trawler con- tacts are recovered, Fast-SlowDSVDDproduces 13 alarms over 92 days, approximately one per week. This represents a review burden that is op- erationally tractable for monitoring. The proposed fast-slow design therefore provides consistent un- supervised detection across both short impulsive and sustained trawler-contact events. Discovery: The modelâs highest-ranked record- ings included three additional candidate events, on 3 rd June, 10 th June, and 17 th August 2025, that were absent from the event log and had not been identified during manualSoPreview. Figure 4 compares these detections with a logged trawler- contact event. Post-hoc review against concurrent DASandAISrecords confirmed physical cable- interaction signatures for the 10 th June and 17 th August events. No correspondingDASsignature was observed for the 10 th June event through- out 120 km, suggesting that it occurred outside theDAS-covered cable span in this experiment. These findings show that the proposed framework can surface previously undetected events of physi- cal contact from continuous SoP monitoring. Conclusions We demonstrated fully unsupervised detection of physical contact events on a deployed subsea ca- ble using continuousSoPmonitoring. The pro- posed Fast-SlowDSVDDmodel recovered all five confirmed trawler contacts within the top 13 of a 122,174-recording archive and also surfaced additional corroborated candidate events. These results establishSoP-based monitoring as a prac- tical low-cost basis for automated screening of long-duration subsea cable recordings. Acknowledgements This work was supported by the European Unionâs Horizon Europe research and innova- tion programme (Grant No. 10113933) ICON project, CELTIC-NEXT SUSTAINET-ADVANCED (C2024/3-3), Taighde Ăireann â Research Ireland under Grant No. 18/RI/5721 (OpenIreland Re- search Infrastructure), 13/RC/2106_P2 (ADAPT centre). References [1]International Cable Protection Committee. âDamage to submarine cables from dragged anchorsâ. ICPC Viewpoints, updated 24 February 2025, accessed 21 April 2026, International Cable Protection Com- mittee. [Online]. Available:https : / / w . iscpc . org/publications/icpc- viewpoints/damage- to- submarine-cables-from-dragged-anchors/ [2] S. Bjørnstad, K. S. Yamase Skarvang, D. Roar Hjelme, A. Tunheim, F. Fjermestad, and E. Ăsterli, âFirst Im- pact Movement Characterization of Shallow Buried Live Subsea-Cableâ, in 2024 Optical Fiber Communications Conference and Exhibition (OFC), Mar. 2024, p. 1â3. [3]O. H. Waagaard, J. P. Morten, E. Rønnekleiv, and S. Bjørnstad, âExperience from Long-term Monitoring of Subsea Cables using Distributed Acoustic Sensingâ, en, in 27th International Conference on Optical Fiber Sensors, Alexandria, Virginia: Optica Publishing Group, 2022, Th2.4. DOI: 10.1364/OFS.2022.Th2.4 [4] C.-R. Lin, S. von Specht, K.-F. Ma, M. Ohrnberger, and F. Cotton, âAnalysis of saturation effects of distributed acoustic sensing and detection on signal clipping for strong motionsâ, en, Geophysical Journal International, vol. 241, no. 2, p. 971â985, Mar. 2025. DOI:10.1093/ gji/ggaf089 [5]K. Alexoudis, F. Azendorf, A. Doval, and S. Bjørnstad, âMulti-Modal Fiber Sensing for Offshore Environmental and Infrastructure Monitoringâ, en, 2026. [6]K. S. Vaskinn et al., âSensing earthquakes in aerial single- and multi-core fiber optic networksâ, en, Jour- nal of Optical Communications and Networking, vol. 18, no. 4, B119, Apr. 2026. DOI: 10.1364/JOCN.584845 [7] F. Usmani et al., âA Smart Sensing Grid for Road Traf- fic Detection Using Terrestrial Optical Networks and Attention-Enhanced Bi-LSTMâ, en, Journal of Lightwave Technology, vol. 43, no. 10, p. 4624â4634, May 2025. DOI: 10.1109/JLT.2025.3543180 [8]K. Abdelli, M. Lonardi, J. Gripp, S. Olsson, F. Boitier, and P. Layec, âRisky event classification leveraging transfer learning for very limited datasets in optical networksâ, en, Journal of Optical Communications and Networking, vol. 16, no. 7, p. C51, Jul. 2024. DOI:10.1364/JOCN. 517529 [9]S. Bjørnstad, A. Doval, and A. Tysdal, âFirst observations of subsea cable physical contacts combining State of Polarization (SoP) and Distributed Acoustic Sensing (DAS)â, en, 2026. [10]K. Abdelli et al., âVision Transformers for Anomaly Clas- sification and Localization in Optical Networks Using SOP Spectrogramsâ, en, Journal of Lightwave Tech- nology, vol. 43, no. 4, p. 1902â1914, Feb. 2025. DOI: 10.1109/JLT.2024.3519755 [11] L. Sadighi, S. Karlsson, C. Natalino, L. Wosinska, M. Ruffini, and M. Furdek, âDeep Learning for Detection of Harmful Events in Real-World, Noisy Optical Fiber Deploymentsâ, en, Journal of Lightwave Technology, vol. 43, no. 13, p. 6092â6101, Jul. 2025. DOI:10.1109/ JLT.2025.3557748 [12]B. Yang et al., âLow-complexity SOP-based vibration broadband sensing and efficient recognition for stable IM/D optical interconnects in data centersâ, en, Jour- nal of Optical Communications and Networking, vol. 17, no. 8, p. 692, Aug. 2025. DOI: 10.1364/JOCN.559810 [13] W. Qin, X. Gong, W. Hou, L. Gan, and L. Guo, âEx- perimental Demonstration of Bending Eavesdropping Detection in Optical Communications Using a Physics- Informed Convolutional Networkâ, en, Journal of Light- wave Technology, vol. 44, no. 5, p. 1636â1646, Mar. 2026. DOI: 10.1109/JLT.2025.3647694 [14]L. Sadighi, S. Karlsson, C. Natalino, and M. Furdek, âML- Based State of Polarization Analysis to Detect Emerging Threats to Optical Fiber Securityâ, IEEE Transactions on Network and Service Management, vol. 23, p. 432â 442, 2026. DOI: 10.1109/TNSM.2025.3607022 [15] S. Pellegrini et al., âOverview on the state of polarization sensing: Application scenarios and anomaly detection algorithmsâ, en, Journal of Optical Communications and Networking, vol. 17, no. 2, A196, Feb. 2025. DOI:10. 1364/JOCN.537881 [16] K. Abdelli, âPrompt Once, Manage All: A Unified LLM Framework for Multi-Task Optical Link Managementâ, en, Journal of Lightwave Technology, p. 1â12, 2026. DOI: 10.1109/JLT.2026.3660177 [17] L. Ruff et al., âDeep one-class classificationâ, in Proceed- ings of the 35th International Conference on Machine Learning, J. Dy and A. Krause, Eds., ser. Proceedings of Machine Learning Research, vol. 80, PMLR, Oct. 2018, p. 4393â4402. [18] C. J. Carver and X. Zhou, âPolarization sensing of network health and seismic activity over a live terres- trial fiber-optic cableâ, en, Communications Engineering, vol. 3, no. 1, p. 91, Jul. 2024. DOI:10.1038/s44172- 024-00237-w