DOMAIN MAP 2: Structural Health Monitoring, Scour Detection, and Environmental Compensation¶
Synthesised from 45 batch summaries covering ~1,800 papers (1981--2026)
Domain at a glance¶
mindmap
root((D2 · SHM + scour<br/>1 800 papers))
Vibration-based SHM
Natural frequency primary indicator
OMA SSI FDD standard
Continuous monitoring 15+ turbine-years
Weijtjens 2014-17
McAdam 2023
PolyMAX Peeters 2004
Scour detection
Monopile drop 5-15% per 1D
Tripod drop max 5.3% at 0.6D
Bridge piers up to 40%
Global scour > local
Equivalent cantilever models
EOV compensation
Regression Weijtjens 2014
PCA Fremmelev 2023
GP Avendano-Valencia 2017
Bayesian Drangsfeldt 2023
Cointegration Cross-Worden
Feature engineering
Frequency most validated
Mode shapes self-compensating
Higher modes sensitive
MAC COMAC MSF
Feature ranking Buckingham-Pi
Mode-shape approaches
Kariyawasam 2020
Malekjafarian 2018
Khan 2024
Jawalageri 2022
Algorithms
SSI-COV preferred
Automated pipelines
Deep-learning estimators
State-function decomposition
Relevance to PhD
V1 32-month field
V2 state-function EOV
B Buckingham-Pi features
A Bayesian decision
E encoder for SHMWorkflow of a scour-detection pipeline¶
%%{init: {"theme":"base","themeVariables":{"fontSize":"16px","fontFamily":"Inter, system-ui, -apple-system, sans-serif","primaryTextColor":"#1a1a1a","lineColor":"#666"},"flowchart":{"nodeSpacing":25,"rankSpacing":40,"padding":16,"useMaxWidth":true}}}%%
flowchart TB
Sensors["<b>1 · Sensors</b><br/><span style='font-size:14px'>triaxial accelerometers<br/>+ strain gauges · 1500 + 500 Hz</span>"]:::sens
OMA["<b>2 · Operational modal analysis</b><br/><span style='font-size:14px'>SSI-COV · model-order selection<br/>MAC check</span>"]:::oma
Features["<b>3 · Feature extraction</b><br/><span style='font-size:14px'>f₁ · f₂ · damping<br/>MAC · COMAC · MSF</span>"]:::feat
EOV["<b>4 · EOV compensation</b><br/><span style='font-size:14px'>state-function · PCA · GP</span>"]:::eov
Detection["<b>5 · Detection</b><br/><span style='font-size:14px'>CUSUM · RANSAC<br/>Bayesian fusion</span>"]:::det
Decision["<b>6 · Maintenance decision</b>"]:::dec
Sensors ==> OMA ==> Features ==> EOV ==> Detection ==> Decision
classDef sens fill:#e3f2fd,stroke:#1565c0,stroke-width:2px,color:#0d47a1
classDef oma fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px,color:#1b5e20
classDef feat fill:#fff8e1,stroke:#f57f17,stroke-width:2px,color:#e65100
classDef eov fill:#fce4ec,stroke:#c2185b,stroke-width:2px,color:#880e4f
classDef det fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px,color:#4a148c
classDef dec fill:#ffebee,stroke:#c62828,stroke-width:3px,color:#b71c1c1. ESTABLISHED KNOWLEDGE¶
1.1 Vibration-Based SHM Fundamentals¶
Natural frequency is the most widely validated dynamic health indicator for both offshore wind turbines and bridges. The physical basis is unambiguous: foundation stiffness governs natural frequency, and any process that degrades stiffness (scour, soil softening, bolt loosening) shifts the frequency downward. This is confirmed across monopiles (Arany 2014/2015/2016; Prendergast 2013/2015; Weijtjens 2014/2015/2017), tripod suction buckets (Kim 2025; Ma 2018; Seo 2020; Ryu 2020), bridge piers (Bao 2015/2016/2019; Belmokhtar 2021; Boujia 2019/2021; Kariyawasam 2019/2020), and gravity-base foundations (Futai 2021). The equivalent cantilever concept and Winkler spring-beam models are the standard analytical tools for relating frequency to scour depth.
Operational Modal Analysis (OMA) is the accepted field technique for extracting modal parameters from ambient excitation. Stochastic Subspace Identification (SSI) and Frequency Domain Decomposition (FDD) are the dominant algorithms. Automated OMA pipelines have been demonstrated for continuous OWT monitoring over 15+ turbine-years (Weijtjens 2014--2017; McAdam 2023; Moynihan 2023; Song 2023). PolyMAX provides a complementary frequency-domain estimator for controlled-excitation contexts (Peeters 2004).
1.2 Scour Detection via Frequency Shift¶
Scour is the leading cause of bridge failure in the United States (~22 collapses/year; Yao 2013) and a dominant threat to OWT monopile integrity. Every study examining scour effects on foundations reports capacity reduction and frequency decrease: monopile frequency drops 5--15% per 1D scour depth (Li 2020; Jawalageri 2022; Tseng 2018 reports up to 14% for horizontal bending modes); bridge pier frequencies can drop up to 40% for 30% embedment loss (Kariyawasam 2020). Tripod suction buckets show smaller sensitivity -- max 5.3% at scour depth = 0.6D (Kim 2025) -- because multi-footing configurations redistribute stiffness loss.
Global scour is consistently more detrimental than local scour for the same nominal depth, because it removes overburden stress across a wider area (Li 2020; Kariyawasam 2020; Qi 2016). Local scour, however, preserves overconsolidation effects that make the sub-scour soil stiffer than a simple mudline-lowering assumption predicts (Qi 2016; Chortis 2020).
1.3 Environmental and Operational Variability (EOV) as a Confounding Factor¶
Temperature, wind speed, traffic, and tidal state all produce frequency variations that can mask or mimic damage signals. This is universally acknowledged (Casas 2017; Avendano-Valencia 2017; Cao 2017; Weijtjens 2014; Kariyawasam 2019a; Weil 2025). Regression-based normalisation (Weijtjens 2014), PCA-based dimensionality reduction (Fremmelev 2023; Drangsfeldt 2023), Gaussian process time-series models (Avendano-Valencia 2017), and Bayesian regression (Drangsfeldt 2023) are the established compensation approaches.
1.4 Foundation Stiffness Governs OWT Dynamics¶
Monitored natural frequencies for monopile OWTs are systematically 5--15% higher than design predictions (Stuyts 2022/2023; Kallehave 2015; Arany 2016). This reflects conservative soil stiffness assumptions in API/DNVGL p-y curves. The PISA framework (Jurado 2022; Kheffache 2024; McAdam 2023) and OxCaisson Winkler hierarchy (Suryasentana 2017--2023) provide more accurate alternatives. Soil damping values reported in the literature span 0.17--1.3% of critical (Rezaei 2018), making damping the most uncertain parameter for fatigue life prediction.
1.5 Damage Detection Theory¶
Worden's axioms and the Rytter (1993) four-level hierarchy (detection, localisation, quantification, prognosis) provide the theoretical backbone. Data-driven SHM is well suited for Level 1 (detection) using unsupervised methods trained on healthy-state data only (Svendsen 2021; Cevasco 2020; Abdelhak 2024). Advancing to Levels 2--4 generally requires physics-based model integration (Cross 2021; Kita 2021; Ritto 2021).
2. ACTIVE FRONTIERS (2023--2026)¶
2.1 Physics-Informed SHM and Digital Twins¶
The fusion of physics-based models with data-driven methods is the dominant research frontier. Approaches include: - GPLFM (Gaussian Process Latent Force Model): First in-situ below-mudline strain validation at Westermeerwind Park (Zou 2022); virtual sensing of subsoil strains from above-water accelerations (Nayek 2019; Bilbao 2022). - Differentiable modelling: Embeds physical equations within neural network training loops for end-to-end learning that respects conservation laws (Shen 2023, Nature Reviews). - Hierarchical Bayesian population models: Smith (2023) infers population- and local-level soil stiffness distributions across a wind farm, detecting scour as an anomaly within the population-based SHM (PBSHM) framework. - Digital twin with Kalman filter: Branlard (2024) achieves 10--15% error on fatigue DEL for a full-scale floating OWT (TetraSpar prototype). - Physics-informed RANSAC + CUSUM: Kim (2026 submitted) reduces EOV-induced frequency scatter by 70% and detects scour at 0.39D with 95% probability and zero false alarms using 32 months of field data from a 4.2 MW tripod OWT.
2.2 Unsupervised and Data-Driven Scour Detection¶
Novelty detection using only healthy-condition data is emerging as the practical paradigm for scour monitoring, since labelled scour-state data is virtually nonexistent in operational settings (Abdelhak 2024; Drangsfeldt 2023). Autoencoder-based anomaly detection (LSTM-AE, sparse AE + SVM) has been validated for multivariate industrial time-series (Park 2023; Terbuch 2023; Wei 2023; Roy 2024). Extended Isolation Forest on triaxial accelerometer data detects blade ice accumulation using healthy-condition training only (Gomez ~2024).
2.3 Distributed and Remote Sensing¶
- DAS (Distributed Acoustic Sensing): Xu & Soga (2024) provide the first full-scale OWT validation of phi-OTDR DAS, capturing both local (loose bolt) and global tower strain over km-scale distances.
- DFOS (Distributed Fibre Optic Sensing): Shi (2021) reviews continuous spatial monitoring at mm resolution; Glisic (2022) highlights advances in damage localisation.
- GNSS-based bridge SHM: Wang (2021) reviews displacement monitoring and dynamic characterisation.
- Computer vision for scour: Millar (2024) links structural response to scour conditions via the ArchIMEDES algorithm on masonry arch bridges.
2.4 Surrogate Models for Lifetime Reassessment¶
Kriging meta-models replace ~1M aeroelastic simulations for fatigue assessment (Schmidt 2025/2026). GP regression surrogates trained on small sea-state subsets predict long-term fatigue damage at tower base and mooring fairlead (Liu 2023). ML-based DEL surrogates with transfer learning extend to wake conditions (Haghi 2024).
2.5 Mode-Shape and Higher-Mode Indicators¶
Second mode shape is more sensitive to scour than the first mode (Jawalageri 2022). Mode Shape Ratio (MSR) is more sensitive to scour than natural frequency and its temperature dependence is inversely related, enabling effect separation (Malekjafarian 2019). Mean-Normalised Mode Shape (MNMS) can locate scour and quantify stiffness loss (Malekjafarian 2020). MAC values track scour progression (Jawalageri 2022). The trade-off is that mode-shape-based indicators require more sensors.
3. CONTESTED CLAIMS¶
3.1 Frequency vs. Mode Shape as Primary Scour Indicator¶
Kariyawasam (2019a) found that field frequency variability can exceed scour-induced frequency shifts, questioning the reliability of frequency-only approaches for small-span bridges. Khan (2021) and Malekjafarian (2020) advocate mode-shape-based indicators as more robust. Jawalageri (2022) shows second-mode sensitivity is higher but requires more instrumentation. No consensus exists on which indicator is optimal for operational deployment.
3.2 Local vs. General Scour Treatment in Design¶
Ciancimino (2022) demonstrates that local scour (38% moment capacity reduction) and general scour (48% reduction) produce fundamentally different failure mechanisms, challenging the DNV 1.3D uniform scour assumption. Chortis (2020) shows that p-y curves must be modified for scour-hole geometry (4th vs. 7th order polynomial extraction is itself debated). No standardised code provision accounts for these differences.
3.3 Physics-Based vs. Purely Data-Driven SHM¶
Cross (2021) argues that grey-box models (physics + GP kernels) generalise beyond training data, while purely data-driven CNN/LSTM approaches (Zhang 2020; Ghimire 2021) demonstrate strong within-domain accuracy. The trade-off between interpretability/generalisability and raw predictive power is unresolved. Chen, Cao & Zhu (2023, IEEE IoT Journal) highlight that most data-driven SHM shows "digressive performance" in real-world deployment versus simulation.
3.4 Cyclic Loading: Stiffness Increase or Degradation?¶
LeBlanc (2010) demonstrated that pile stiffness in sand increases with number of cycles, directly contradicting the API approach that degrades static p-y curves. Lombardi (2013) found frequency changes in clay depend on shear strain level. Xu (2019) observed frequency increase after cycling in sand due to compaction. Nielsen (2017) showed two-way loading produces largest rotation under partly drained conditions, contradicting drained-condition findings. The direction of stiffness evolution remains soil-type, strain-level, and drainage-regime dependent.
3.5 10-Minute Analysis Window Adequacy¶
Sadeghi (2023) demonstrates that up to 65% of monopile fatigue damage relates to low-frequency dynamics with periods >1 day, invisible in standard 10-minute analysis windows. This directly challenges the DNV/IEC DLC framework. Whether short-to-long-term correction factors suffice or whether fundamental changes to data segmentation are needed is actively debated.
4. VERIFIED GAPS¶
4.1 No Validated Frequency-Based Scour Monitoring System for Operational OWTs¶
All frequency-based scour detection work on OWTs remains at laboratory, centrifuge, or numerical stage. Crotti (2019) provides rare long-term field data for bridges; Kariyawasam (2019a/b) deployed field accelerometers on bridges. Kim (2026 submitted) is the first to use extended field data from an operational OWT tripod, but the method awaits peer review and fleet-scale validation. No commercial system exists.
4.2 Combined Scour + Cyclic Loading Interaction¶
Cox and Bhattacharya study cyclic loading without scour; Chen (2018) and Chortis (2020) study scour without cycling. No experimental or numerical study systematically investigates the interaction of progressive scour with millions of load cycles on OWT foundations. This is the critical operational scenario.
4.3 Scour Effects on Suction Bucket Foundations¶
Scour research is heavily concentrated on monopiles and bridge piers. Only Chen (2018), Cheng (2024), Yuan (2019), Jin (2025), and Kim (2025) examine scour on suction buckets. Bucket-specific scour morphology, time-dependent scour evolution, and combined scour-VHM-cyclic capacity are largely unexplored.
4.4 Long-Term Field Monitoring Linking Scour Progression to Dynamic Response¶
Song (2023) provides 1-year OWT modal monitoring (Block Island) but does not address scour. Weijtjens (2017) tracks 15 turbine-years but scour depth measurements are not co-located with vibration data. No integrated dataset exists that simultaneously records scour bathymetry, environmental conditions, and structural dynamic response over multi-year timescales.
4.5 Cross-Domain Transferability of SHM Methods¶
Weil (2025) and bridge SHM papers (Bao, Prendergast) develop parallel but disconnected ML-based frequency tracking methods. No framework unifies OWT and bridge SHM, despite shared physics (frequency-scour relationship, EOV compensation, Winkler spring models). Transfer learning across structure types is unexplored.
4.6 Scour in Cohesive and Mixed Soils¶
Harris (2023) explicitly states that scour depth estimation in cohesive marine soils carries large uncertainty. Bao (2019) identifies cohesive soil effects on vibration-based detection as a critical gap. Most scour studies use clean sand. Time-dependent scour in clay/silt is poorly characterised for both bridges and OWTs.
4.7 Quantitative (Not Just Qualitative) Scour Assessment from Vibration Data¶
While detection (binary: scour present or not) is established, quantitative estimation of scour depth from vibration data is nascent. Weil (2023) proposes a digital-twin approach; Kim (2026 submitted) uses a coupled 3D-to-1D framework. Both are single-site demonstrations. Fleet-scale quantification is absent.
5. QUANTITATIVE BENCHMARKS¶
5.1 Scour Sensitivity¶
| Parameter | Value | Source |
|---|---|---|
| Monopile fn drop per 1D scour | 5--15% (varies with soil, L/D) | Li 2020; Jawalageri 2022 |
| Tripod suction bucket fn drop at 0.6D scour | 5.3% max | Kim 2025 |
| Bridge pier fn change for 30% embedment loss | Up to 40% | Kariyawasam 2020 |
| Scour-induced fatigue life reduction (1.3D) | ~24% | Cao 2024 |
| TMD fatigue life increase (1% mass ratio) | ~65% at 1.3D scour | Cao 2024 |
| Moment capacity loss per 1D scour (monopile) | Near-linear, geometry dependent | Li 2021 |
| Local scour capacity reduction | 38% of Mult | Ciancimino 2022 |
| General scour capacity reduction | 48% of Mult | Ciancimino 2022 |
5.2 Foundation Stiffness and Natural Frequency¶
| Parameter | Value | Source |
|---|---|---|
| Design fn underestimation (API p-y) | 5--15% below measured | Stuyts 2022; McAdam 2023 |
| Closed-form fn prediction error (3-spring model) | <3.5% across 10 European OWFs | Arany 2016 |
| Soil damping ratio range (monopile) | 0.17--1.3% critical | Rezaei 2018 |
| Aerodynamic damping (operational) | 2--8% critical | Rezaei 2018 |
| Foundation cost share of OWT | 15--40% of total | Houlsby 2000; Tran 2017 |
| Monopile market share | 77--80% | Abdelhak 2024; Weijtjens 2017 |
5.3 EOV Compensation and Detection Performance¶
| Parameter | Value | Source |
|---|---|---|
| EOV scatter reduction (physics-informed pipeline) | 70% | Kim 2026 (submitted) |
| Scour detection threshold | 0.39D at 95% probability, zero false alarms | Kim 2026 (submitted) |
| Digital twin fatigue DEL error (field) | 10--15% | Branlard 2024 |
| GPLFM virtual sensing accuracy | First in-situ validation below mudline | Zou 2022 |
| Monopile DEL extrapolation (strain-based) | <4% monthly error | Ziegler 2019 |
| Jacket model updating fn error reduction | 30% to 1% via SSI-based updating | Augustyn 2020 |
| Bridge scour ML forecasting | 1-week ahead prediction | Yousefpour 2021 |
5.4 Field Monitoring Campaigns¶
| Site / System | Duration | Sensors | Key Finding | Source |
|---|---|---|---|---|
| Belwind monopile (BE) | 5+ years | Accelerometers, OMA | Stiffening trend detected; scour tracked | Weijtjens 2014--2017 |
| Block Island OWT (US) | 1 year | Accelerometers | FA modes highly variable; SS modes stable | Song 2023 |
| Westermeerwind Park (NL) | -- | GPLFM virtual sensing | First below-mudline strain validation | Zou 2022 |
| Korean 4.2 MW tripod OWT | 32 months | Accelerometers + SCADA | Scour detected at 0.39D, EOV compensated | Kim 2026 |
| Robin Rigg (UK) | Multi-year bathymetry | Sonar surveys | Scour-induced decommissioning of 2 turbines | Garcia 2023; E.ON ~2015 |
| Borkum Riffgrund 1 SBJ (DE) | Continuous since 2015 | Pore pressure, strain | First SBJ field data; drained/undrained boundaries identified | Shonberg 2017 |
| Bothkennar (UK) / Luce Bay (UK) | Field trials | Load cells, displacement | Suction caisson field benchmark (clay/sand) | Houlsby 2005/2006 |
| Baildon Bridge (UK) | Multi-year | Accelerometers, FDD | Frequency variability exceeds scour-induced shift | Kariyawasam 2019a/b |
| Po River bridge (IT) | Long-term | Sedimeter, echo sounder | Real-time vulnerability during floods | Crotti 2019 |
| EuroProteas prototype (GR) | Single campaign | Ambient vibration | Full-scale shallow foundation under progressive scour | Tubaldi 2022 |
| Korean tripod suction pile | Installation stages | Impact/ambient | fn matches analysis within 1.5% per stage | Ryu 2019 |
5.5 Key Reference Datasets and Benchmarks¶
- Robin Rigg failure case: Two monopiles decommissioned after 6 years due to 15 m seabed drop and 120x60x20 m scour hole (Carlos 2023; E.ON).
- DNV 1.3D scour depth rule: Industry standard design scour depth for monopiles under current-only; shown to be context-dependent.
- Butterfield (1994) VHM envelope: Foundational dataset for combined loading failure surfaces on sand.
- Oztoprak & Bolton (2013) G/G0 database: 454 lab tests, 3860 data points; prediction within factor 1.13.
- Ching CLAY-Cc/6/6203: 6203 records from 429 studies; largest public clay compressibility database.
- JCSS PMC Section 3.7: Probabilistic soil property characterisation; scales of fluctuation; CoV tables.
- CWRU bearing dataset: De facto DL benchmark for machinery fault detection.
- OC5 Phase III (Alpha Ventus): Multi-code OWT simulation validation against full-scale measurements.
- Cox centrifuge caisson tests: 12,000 cycles at 1:200 scale; reference dataset for cyclic stiffness evolution.
- PISA Dunkirk/Cowden field tests: Benchmark for new monopile p-y methods.
- TAMU-scour method (Briaud 2015): Scour prediction incorporating soil erodibility; validated against 10 databases.
- Ciancimino hybrid methodology: 1g flume + Ng centrifuge with 3D-printed scour holes; methodological benchmark for combined hydraulic-mechanical scour testing.
Generated 2026-04-17 from 45 batch synthesis files (batch01_agent1 through batch09_agent5), covering approximately 1800 unique papers spanning 1981--2026.