A review on environmental load identification and structural health monitoring of offshore structures

Wind

Wind is a major component of the environmental load. In the design of offshore structures, international marine and offshore engineering standards are the main references tools. In view of the current standards and calculation procedures that treat wind as quasi-stable and ignore the relevant parameters of atmospheric turbulence in nature, Kozmar (Kozmar et al., 2022) took relevant international standards, rule-based procedures, and actual wind data as input conditions to calculate the environmental loads acting on single-pile offshore wind turbines (OWT). Li et al. (Li et al., 2022a) used the observation data of offshore platform to analyze the changes of the average wind field, turbulence and turbulence flux of three typhoons landing in Guangdong, and discussed the influence of the average wind direction on the turbulence fluctuation intensity and friction velocity of the typhoon boundary layer. Field measurement provides the most authentic data but is restricted by the scarcity of measuring points and the high cost of maintenance in harsh environments (Jiang et al., 2021). In contrast, existing standards are convenient but often conservative, failing to capture the complex transient effects of turbulence, which suggests a need for more refined aerodynamic models.

For nearly 50 years, many scholars have applied Computational Fluid Dynamics (CFD) to the evaluation of wind load, which has grown rapidly recently (Guo et al., 2021a). Wnek et al. (Wnek and Soares, 2015) studied the wind load of two kinds of ship models using CFD, and compared the results with the wind tunnel experiments. Janssen et al. (Janssen et al., 2017) conducted a three-dimensional stable RANS CFD simulation of the wind load on container ships, discussed the influence of simplifying the hull and deck container stacking on the results, and studied the blocking effect caused by the domain boundary. Although CFD has become a powerful tool to complement experimental data, a major limitation lies in the Reynolds number effect (Zhao et al., 2023). Most CFD studies and wind tunnel tests are performed at model scale, and extrapolating these results to full-scale offshore structures involves uncertainties. Furthermore, the accuracy of CFD heavily depends on the selection of turbulence models and mesh quality, which requires rigorous validation against field data.

In addition to deterministic descriptions based on standards and CFD, random wind loads are often modelled as stochastic processes when analysing and identifying the dynamic response of offshore structures. In this context, the turbulent wind field is usually characterised by power spectral density (PSD) functions and coherence models, or simulated in the time domain using spectral representation and multivariate ARMA-type models (Ma and Liu, 2025; Wang et al., 2025). Such stochastic wind models provide not only realistic input for fatigue and ultimate limit state assessments, but also a natural framework for integrating wind effects into SHM.

Wave

The significant height and peak time of wave are the main factors affecting the response of offshore structure. The current wave monitoring technology mainly includes the measurement of various sensors and the indirect method of estimating wave height.

The development of wave sensors mainly includes the buoys with accelerometers or Global Positioning System (GPS). In recent years, many scholars have developed wave monitoring sensors and took them to the measurement of wave in the actual sea. While sensors like buoys offer high reliability, they suffer from high deployment costs and vulnerability to extreme weather. Bio-inspired sensors offer a novel solution for power supplies. However, their long-term stability and measurement precision in complex sea states still require further verification compared to mature commercial sensors (Chen et al., 2024; Jensen et al., 2021; Yurovsky and Dulov, 2020).

Compared with the direct measurement method, the wave monitoring technology combined with machine learning methods has the advantages of lower cost. Many scholars have applied machine learning methods to the monitoring of wave characteristics parameters, the recent research achievements are summarized in Table 1. As shown in Table 1, data-driven methods significantly reduce hardware dependency. However, a critical challenge for these intelligent algorithms is their generalization ability. Most models are trained on historical data within specific ranges; thus, their prediction accuracy may degrade significantly under unseen extreme wave conditions. Additionally, the “black box” nature of deep learning models lacks physical interpretability, which is a concern for engineering safety assessment.

OPEN IN VIEWER

Table 1.

Summary and comparison of machine learning methods for wave characteristic prediction.

Reference	Method	Source of data	Monitoring indicator	Accuracy
Kumar et al. (Kumar et al., 2017)	Minimal resource allocation network and the growing/pruning radial basis function network	Gulf of Mexico, Korean region and UK region (Jan 1, 2011-Aug 30, 2015)	Wave height	With the RMSE of 0.2661 m (MRAN) and 0.2808 m (GAP-RBF)
Kumar et al. (Kumar et al., 2018)	Deep belief networks	The same data set with (Kumar et al., 2017)	Significant wave height, wave period, crossing period	–
Yang et al. (Yang et al., 2022)	Bayesian optimization algorithm combined with automated machine learning	Nine buoys from NOAA national data buoy center	Significant wave height	With the RMSE of no more than 0.14 m
Hwang et al. (Hwang et al., 2023)	Convolutional neural networks	Camera mounted on the deck of ship	Wave height	With the MAD of 0.098 m

Ice

Ice load is the main control load for the safe operation of ships and offshore structures in the ice area. Wide structures (such as gravity platform, semi-submersible platform, etc.) are usually adopted to resist the severe polar sea ice conditions. The structure-ice interaction process is complex, it is necessary to pay attention to the overall and local ice load of the structure in the design of structural ice resistance (Wang et al., 2018c). Wang et al. (Wang et al., 2018a) analyzed the dynamic characteristics of the platform under different ice excitation levels based on the long-term measured response of the ice-structure interaction. The results show that the inherent nonlinearity of the structure varies with excitation amplitude and bandwidth. It can be seen that monitoring of ice load is of great significance for the health assessment of offshore structures. Figure 1 shows the timeline of research on ice loads over the past 10 years. It can be seen that research on the monitoring technology of ice load mainly includes numerical simulation and the establishment of ice force prediction model. Moreover, image identification and machine learning are also used in the study of ice loads (Kim and Nam, 2020; Wang et al., 2023b). Current numerical simulations often simplify ice constitutive models, leading to discrepancies with reality. Moreover, the applicability of image identification techniques is severely limited by visibility. In polar regions, frequent fog, snow, and polar night conditions can render visual-based monitoring ineffective. Future research should focus on fusing multi-source data to ensure robustness.OPEN IN VIEWER

For ice loads, the highly intermittent and stochastic nature of ice–structure interaction has also motivated the development of random load models. Instead of prescribing deterministic ice forces, these studies represent ice-induced loads as stochastic processes that account for random breaking patterns(Chu and Lindenschmidt, 2019), intermittent contact (Hossain and Taylor, 2024) and switching between crushing and bending failure modes (Teo et al., 2017). In such models, ice loads are often described by stochastic differential equations, random pulse trains or Markov-type switching processes, and their spectra are used as input for dynamic response and fatigue analyses. Within an SHM context, these stochastic ice load models can be combined with measured structural responses to identify key parameters and to update reliability assessments of structures.

Theory of structural dynamics

Firstly, the relationship between structural motion, strain and stress is established through theoretical analysis. The spatial discrete dynamic system of offshore structure can be represented by the following continuous time differential equation:

M u ¨ ( t ) + C u ˙ ( t ) + Ku ( t ) = f ( t )

(1)

[ K − ω 2 M ] ϕ = 0

(2)

The relationship between strain mode and displacement mode can be used to obtain the strain time history. The two kinds of modes have the same frequency and damping ratio, but the mode expression is different. In the finite element, the strain vector at any point of the element can be expressed as a function of the displacement vector, the formula is as follows (Vettori et al., 2023):

ε = Su

(3)

S = [ ∂ ∂ x ∂ ∂ y ∂ ∂ z ∂ ∂ y ∂ ∂ x ∂ ∂ z ∂ ∂ y ∂ ∂ z ∂ ∂ x ]

(4)

Based on this, the strain time history can be obtained indirectly, the stress can be further obtained from the constitutive relationship of the material, which can be used to evaluate the fatigue damage of offshore structures.

Modal parameter identification

Modal identification of offshore structures has always been a focus of attention. The dynamic characteristics of offshore structures will vary with time because of the time-varying environmental loads such as wave, wind and currents, as well as the change of soil hardness. For large and complex engineering structures, extracting modal parameters from output data is a common method because of the difficulty in measuring input forces.

From a practical implementation point of view, field modal identification for offshore structures is also strongly affected by data quality and uncertainty issues. Long-term response measurements may suffer from sensor drift, intermittent outages, clipping during extreme events and timing inconsistencies between different acquisition systems (Zhang et al., 2025). In addition, the excitation is often weak and broadband, especially for stiff structures, which leads to low signal-to-noise ratios and increases the uncertainty of identified modal parameters. Therefore, it is necessary to develop effective modal parameter identification methods to obtain the accurate operating state. Some method of filtering out noise and improve the accuracy of modal parameter identification are listed in Table 2.

OPEN IN VIEWER

Table 2.

Methods of filtering out the noise effect on modal parameter identification.

Year	Reference	Method
2015	Tang et al. (Tang et al., 2015)	Unilateral moving average and random decrement technique
2018	Dong et al. (Dong et al., 2018)	Harmonic correction
2019	Liu et al. (Liu et al., 2019)	Repeat extraction of the maximum energy component
	Bao et al. (Bao et al., 2019)	Partially described inverse singular value problem
	Bao et al. (Bao and Shi, 2019)	Structured low rank approximation
2020	Liu et al. (Liu et al., 2020)	Repeat extraction of the maximum energy component
2022	Li et al. (Li et al., 2022b)	Hilbert-Huang transform
	Zhang et al. (Zhang et al., 2022a)	Kalman filter
2024	Xiang et al. (Xiang et al., 2024)	Density-based noise application spatial clustering method
2025	Sales et al. (Sales and Chang, 2025)	An improved stochastic subspace identification framework

At present, the modal identification research of offshore structure mainly includes natural excitation technology and modal parameter identification under different operating states, which is also the covered aspects of the chapter.

Motion status monitoring

The motion state of offshore structure is one of the most important measurement indexes to describe the dynamic characteristics in structural engineering, which can be directly transformed into strain and deformation, and provide useful information of the structure. At present, the monitoring of the movement of offshore structures can be obtained by Global positioning system (GPS), inertial navigation system (INS), acoustic positioning system (APS), Lidar, satellite and so on. However, the above techniques mainly focus on the attitude, position, movement trajectory and distribution of offshore structures, which are not enough for the dynamic response analysis. The analysis of the motion state of the structure can be based on the vibration equation, which is described briefly in Section 3.1. The current research mainly includes theory, simulation and experiment. In practice, the computational cost, model transparency and ease of integration into existing SCADA and monitoring infrastructure are important constraints that influence the choice between purely physics-based and data-driven approaches.

Theoretical research

In the aspects of theoretical research, it mainly includes the establishment of the offshore structure vibration equation, the treatment of non-zero initial conditions, the solution of the traditional drift problem and the low efficiency of calculation. The theoretical studies on the motion states of offshore structure in recent years and are summarized in Table 3. The theoretical approaches summarized in Table 3 offer high computational efficiency and clear physical meaning by solving or simplifying the governing equations. However, their main limitation is the model reliance. The accuracy of these methods heavily depends on the precision of the finite element model, which is often uncertain and requires continuous updating in real-world applications.

OPEN IN VIEWER

Table 3.

The theoretical studies on the motion state of offshore structure.

Reference	Achievement	Advantages
Lu et al. (Lu et al., 2022b)	A new Laplacian domain algorithm	High computational efficiency, suitable for steady-state analysis
Lu et al. (Lu et al., 2022a)	The state space equation of the floating structure by introducing the discrete frequency transfer function	High computational efficiency, can avoid the dissipation in numerical integration algorithms
Shi et al. (Shi et al., 2022)	Stable forced vibration equations for OWT systems with multiple DOF	High accuracy
Liu et al. (Liu et al., 2021)	Non-integral displacement reconstruction method using complex exponential series	Suppressed the drift of traditional methods
Liu et al. (Liu et al., 2018)	Time-domain response estimation method with non-zero initial conditions	Estimate the response of non-zero initial conditions effectively
Park et al. (Park et al., 2017)	The motion equation of a floating OWT using the flexible multi-body dynamics method	High accuracy
Lan et al. (Lan et al., 2022)	The floor acceleration response spectrum of the platform under wave and earthquake action	High computational efficiency, high accuracy
Kefal et al. (Kefal, 2017; Li et al., 2020a)	Stablished a four-node inverse shell element (iQS4)	Can reconstruct global displacement field of the structures, high accuracy

In addition, some machine learning methods have been used to the response estimation of offshore structure. Wang et al. (Wang et al., 2022b) proposed an identification method of tower top acceleration and tower foot force based on the deep learning (DL) using the coupling relationship between tower force and floating foundation motion. Guo et al. (Guo et al., 2021b) developed a machine learning model based on long short-term memory (LSTM) to predict the fluctuation and surge motion of semi-submersible vehicles. Subsequently, the DL model was extended using dropout layer, which can be used to predict the motion of the offshore platform in advance and accurately (Guo et al., 2022). Huang et al. (Huang et al., 2024) used the optimized traditional VMD method to decompose the response data of OWT under extreme conditions, eliminated modal components with similar complexity to reduce the calculation cost, and used the PSO-LSTM model to predict the motion of turbines with six degrees of freedom. Data-driven methods are adept at learning the complex, nonlinear mapping between input and response without explicit physics models. Conversely, they suffer from two major drawbacks: lack of generalizability and interpretability. Their performance degrades significantly when encountering data outside the training domain. This compromises their trustworthiness in safety-critical structural health assessments, demanding more physics-informed machine learning solutions.

Experiment method

For the study of experimental methods, Zhang et al. (Zhang et al., 2019) derived hydroelastic similarity based on elastic similarity, Froude similarity and cross-section bending stiffness similarity, which was used to guide the experiment of a 10,000-ton offshore power platform under wind, wave and current. On this basis, Sun et al. (Sun et al., 2019) studied the dynamic characteristics of a million-ton offshore converter station under typical earthquake loads based on the physical model test. The acceleration and displacement of key parts were collected by data acquisition instrument combined with the strain gauge, acceleration sensor and laser displacement meter. The layout position of the sensor was shown in Figure 2. Based on the experiments, the effect of water, air, single and bidirectional seismic excitation on structural dynamic response were analyzed (Zhao et al., 2021a). Hu et al. (Hu et al., 2020) built a test platform (Figure 3) to simulate various loading states such as wind-wave, wind–current, wave-current and wind-wave-current for the first time, and measured the displacement, velocity and acceleration of the OWT model at three points on the tower wall to study the influence of different working conditions on the dynamic response. Xi et al. (Xibing et al., 2025) revealed the fluid-soil-wind turbine interaction mechanism and the influence on dynamic response through shaking table tests.OPEN IN VIEWER

Figure 2.

Placement of sensor ref (Sun et al., 2019).

OPEN IN VIEWER

Figure 3.

Experimental model in ref (Hu et al., 2020).

In addition to monitoring and analysing the motion status of offshore structures, a number of studies have proposed vibration control devices to mitigate structure vibrations. For example, Lei et al. (Lei et al., 2023) proposed a novel inner-platform tuned mass damper (TMD) for steel wind turbine towers and showed numerically that it can effectively reduce vibrations under wind and wave loads. Based on this research, Lei et al. (Lei et al., 2026) further studied vibration control of wind turbine towers using a nonlinear tuned liquid damper (TLD) with aero-elastic coupling. Although vibration control is not the main focus of the review, it is worth noting that the design and performance assessment of such devices strongly rely on reliable monitoring and identification of structural responses, which links closely to SHM techniques.

From a structural health monitoring perspective, motion status monitoring mainly provides global kinematic information of offshore structures, such as rigid-body translations, rotations and overall vibration amplitudes. These motion quantities are directly linked to internal forces and deformations through the structural dynamics relations introduced in Section 3.1. However, global motion is often insufficient to quantify local fatigue damage or to detect early-stage defects at critical details. For this reason, SHM systems typically combine motion measurements with strain/stress monitoring, which offers more localized and damage-sensitive indicators. Therefore, the next section focuses on methods for strain/stress monitoring that complement motion-based monitoring within an integrated SHM framework.

Strain/stress monitoring

Real-time monitoring of stress distribution and changes of offshore structures can detect potential damage and failures, such as stress concentration, crack initiation, etc., so as to take timely maintenance measures to avoid catastrophic accidents and ensure the safe operation of structures. This section addresses “strain/stress monitoring”, which targets structural responses at critical locations and provides more direct input for fatigue assessment and damage identification. At present, the research of stress/strain monitoring includes direct measurement and indirect identification methods.

Direct measurement

As mentioned earlier, using sensors for measurement is the most common method, which can provide the most accurate results for strain/stress. In 2015, Shen et al. (Shen et al., 2015) proposed two types of FBG strain rosette to measure the strain of hull plane, as shown in Figure 4. The type and size of the strain rosette were optimized by analyzing the factors affecting the strain transfer ratio. Subsequently, Brewick et al. (Brewick et al., 2022) proposed the structure of strain rosette as shown in Figure 5, which was laid on the hull to predict the local pressure. While direct measurement provides high-fidelity local information, it is fundamentally limited by sensor placement and durability. For large offshore structures, it is economically and practically impossible to cover all critical areas, especially in the high-stress areas underwater, which has accelerated the development of virtual sensing techniques for structural response identification.OPEN IN VIEWER

Figure 4.

structure of FBG strain rosette proposed in (Shen et al., 2015).

OPEN IN VIEWER

Figure 5.

Typical strain gauge rosette set-up proposed in (Brewick et al., 2022).

Virtual sensing

For offshore structures, the maximum stress/strain position is generally located underwater, which is difficult to be directly measured due to the influence of operating conditions and economy. At present, there are some methods to estimate stress/strain using measured data and structural parameters, which can be called virtual sensing (VS) technology. The VS technology mainly includes inverse finite element method (iFEM), Kalman filter (KF), mode extension (ME) and machine learning (ML). Recent studies on the application of VS to strain/stress identification of offshore structures is summarized briefly in Table 4, which will be described in detail later. Virtual Sensing (VS) techniques overcome the sensor coverage limitation, but they introduce a new set of challenges. Their performance is highly dependent on two factors: the fidelity of the structural model and the placement of the limited sensor network. An improperly calibrated FE model will amplify errors throughout the reconstruction process, and suboptimal sensor locations can lead to unobservable or inaccurate results for specific damage modes. The key difference among iFEM, KF, and ME lies in their error treatment: iFEM is based on a variational principle, KF is based on statistical estimation, and ME relies on modal truncation.

OPEN IN VIEWER

Table 4.

Study on strain/stress identification of offshore structures.

Method	Reference	Research approach and object
iFEM	Kefal et al. (Kefal and Oterkus, 2016)	Numerical simulation of Panama container ship
	Kefal (Kefal, 2019)	Numerical simulation of cylindrical offshore structures
	Zhu et al. (Zhu et al., 2024)	Numerical simulation and model test of hull stiffened plates
KF	Maes et al. (Maes et al., 2015, 2016)	Field experiment of 3 MW single-pile OWT in the North sea of Belgium
	Tatsis et al. (Tatsis et al., 2021)	Numerical simulation of NREL 5 MW OWT
	Terrazas et al. (Terrazas et al., 2022)	Numerical simulation of an OWT
	Palanisamy et al. (Palanisamy et al., 2019)	Model test of offshore structure in flume
	Song et al.(Song et al., 2022)	Field experiment of FINO3 platform located in the North sea
ME	Iliopoulos et al. (Iliopoulos et al., 2016)	Field experiment of a Vestas V90 3 MW OWT in the Belgian North sea.
	Skafte et al. (Skafte et al., 2017)	Model test of a 1:50 geometrical scaled model of a tripod jacket structure
	Tarpo et al. (Tarpo et al., 2020, 2021)
	Augustyn et al. (Augustyn et al., 2021)	Numerical simulation of jacket type substructure of OWT
	Henkel et al. (Henkel et al., 2020)

Inverse finite element method

iFEM is usually used to reconstruct the strain field of the whole structure, which is a response reconstruction method based on the least square variational principle proposed by Tessler et al. (Tessler and Spangler, 2004) of NASA Langley Research Center in 2002. Based on the known structural responses (such as displacement, strain, etc.), the structural parameters, full-field response and the unknown external loads causing these responses can be obtained by iFEM. Subsequently, many scholars have applied iFEM to the field of civil engineering structures, offshore engineering equipment, aerospace structures and biomedical engineering widely.

The study of iFEM on offshore structures is mainly done by Kefal’s team. Based on the research of Tessler, Kefal and Oterkus (Kefal and Oterkus, 2016) established a simple inverse four-node quadrilateral shell element (iQS4) in 2016, which realized the reconstruction of the displacement and stress of Panama container ship using the measured strain. Subsequently, in 2020, iQS4 was used to obtain von Mises distribution, which was used to estimate the damage under different conditions (Li et al., 2020b). In 2019, Kefal (Kefal, 2019) developed a novel eight-node curved inverse-shell element (iCS8) and corresponding iFEM formula, then the element was used to predicted the full-field deformation shape, global strain and stress distribution of cylindrical offshore structures accurately. In 2023, Zhu et al. (Zhu et al., 2024) combined optical fiber sensors with strain function-inverse finite element method (SF-iFEM) to establish the strain mapping function between the inner and outer surfaces of the hull stiffened plates, which was used to reconstruct the strain field accurately.

The inverse finite element method is efficient and real-time in solving problems where loads or stresses are inferred from displacement data. However, it has several limitations. First, the method’s accuracy depends on the quality and distribution of the measurement data. Insufficient measurement points or noisy data may lead to inaccurate results. Second, iFEM is highly sensitive to geometric modeling and boundary conditions, the systematic errors can be caused if the actual model does not match the theoretical one. Additionally, when dealing with strong nonlinearity, large deformations, or contact problems, the inverse problem may become ill-posed, leading to non-unique or unstable solutions. Overall, iFEM is suitable for situations where the structure is linear, with small deformations, sufficient measurement, and low noise. For complex structures and loads, it should be applied with more caution.

Kalman filter

KF is a recursive state estimation method that updates the optimal state estimation through the system model and measured response. Therefore, the response time history of the whole field of the offshore structure can be estimated using KF and the measured acceleration or strain or the combination of responses. The current trend is to use multi-type responses combined with simplified structural models to improve the accuracy and efficiency of response estimation.

Based on KF, Maes et al. (Maes et al., 2015, 2016) used limited measured data and structural models to obtain the strain time histories of 3 MW single-pile OWT in the North Sea of Belgium under different working conditions and locations. Tatsis et al. (Tatsis et al., 2021) decomposed the OWT system into substructures independent of the whole system, and recursively obtained solutions by fusing the reduced order model of the structural subdomain with the available response through Bayesian filter (BF). Terrazas et al. (Terrazas et al., 2022) used modal space formula and simplification model to reconstruct the dynamic response of the whole physical system through KF, and obtained the multi-axial stress time history to determine the critical plane of maximum fatigue damage accumulation, which was used to calculate the cumulative damage of the structure. Palanisamy et al. (Palanisamy et al., 2019) used the measured acceleration and strain to improve the estimation accuracy of the unmeasured strain in the low and high frequency domains respectively, and introduced buffering technology to realize the estimation of the strain of fixed submarine structures under non-stationary random input. Song et al. (Song et al., 2022) proposed a recursive Bayesian inference framework for the prediction of sparse output strain, the estimation of modeling parameters and unknown inputs of offshore platforms FINO3. In order to solve the problem of traditional argument Kalman filter (AKF) requiring offline selection of covariance matrix related to model uncertainty, Vettori et al. (Vettori et al., 2023) proposed an automated process noise covariance adaptation method to estimate the full-field strain of blade, the results showed that the response prediction accuracy is better than traditional AKF.

The KF method is widely used in structural response identification, but it also has some limitations. KF relies on an accurate system models and noise assumptions. If the model is inaccurate or the noise characteristics do not match, the identification accuracy may decrease. Additionally, KF is sensitive to initial conditions, inaccurate initial estimates can lead to significant deviations in the identification results. Furthermore, when dealing with large-scale systems, the computational burden can be heavy. Therefore, in the application process of KF, attention should be paid to the handling of factors such as noise environment, initial conditions, and computational efficiency.

Mode extension

ME technology can also solve the problem of limited information due to the limited measurement data. Accurate finite element models are fundamental to ME, which can be calibrated by modal assurance criteria (MAC). The response time history of the structure at different positions can be predicted based on the measured response and modal vector extracted from the calibrated FE model. ME method was proposed by Hjelm et al. (Hjelm et al., 2005) in 2005. Subsequently, based on the modal decomposition (MD) and ME method, Iliopoulos et al. (Iliopoulos et al., 2016) combined the limited measured acceleration set with the FE model to obtain the response of the unmeasured position, which was verified in OWT with different operating states. However, the acceleration sensor is not sufficient to reconstruct the low-frequency stress, the effect of noise is amplified by the double integral. In order to solve the problem, Henkel et al. (Henkel et al., 2020) added strain gauges to the sensor network, and predicted the stress of K-type and X-type joints of the jacket structure with the measured acceleration and strain through MD and ME, so as to evaluate the fatigue. In order to improve the accuracy of strain estimation, Skafte et al. (Skafte et al., 2017) divided the structural response into the low and high frequency. Then, the strain time histories of low frequency and high frequency were added to get the combined strain. In order to improve the quality of ME results, Augustyn et al. (Augustyn et al., 2021) proposed to deploy seabed sensors and wave radar sensors to extract local subsea brace modes and static wave modes respectively, which were included in the extended modes. Tarpo et al. (Tarpo et al., 2020) reduced modeling errors using numerical mode to extend experimental mode and converted the measured system response into full-field strain/stress using ME. Then, a method of estimating the stress/strain of the subsystem using a set of undamped and linear modal shapes was proposed and verified by numerical analysis and experiments of offshore platforms (Tarpo et al., 2021).

The ME method in structural response identification relies on accurate finite element models and measurement data. If the finite element model is inaccurate, the predictions may have significant errors. Additionally, estimating low-frequency stresses is challenging, as acceleration sensors cannot effectively reconstruct low-frequency stresses, and noise can affect the prediction accuracy. Furthermore, limitations in sensor placement and quantity may result in insufficient coverage of key areas, which in turn affects the accuracy of stress/strain predictions. While separating low-frequency and high-frequency responses helps improve accuracy, the process is complex and prone to inaccurate separation. Overall, the ME method currently faces issues related to model dependence, noise interference, and sensor limitations.

Corrosion

Corrosion is one of the main causes of failure of offshore structures. Currently, preliminary anti-corrosion measures have been taken for key components. However, experience shows that these measures cannot completely prevent long-term corrosion deterioration in highly corrosive marine environments. In fact, corrosion management still mainly relies on regular inspections and conservative assumptions, and there is still a significant gap between this and the long-term performance requirements of offshore structures. That is, the current system cannot provide reliable, continuous and quantifiable corrosion evolution information to meet these requirements.

Corrosion monitoring techniques for offshore structures mainly include ultrasonic (US) methods, dedicated corrosion sensors and structural-response-based identification methods. US techniques are generally sensitive to thickness loss and localized defects, but their performance may degrade in the presence of complex geometries, coatings, marine growth and environmental noise, and they often require sophisticated signal processing. Dedicated sensors (for example electrochemical or mechanical probes) can provide continuous local information such as corrosion rate, but their spatial coverage is limited and their long-term stability in harsh marine environments remains to be fully demonstrated. Structural-response-based methods, which infer corrosion from reconstructed strain or displacement fields, have the potential to capture more global effects but depend strongly on model fidelity and on the density and quality of the measurement network.

For corrosion localization, Kefal’s team (Ghasemzadeh et al., 2023) used the iQS4 element and both full-sensor and GA-optimized reduced-sensor iFEM schemes to reconstruct global strain fields and identify corrosion regions from deviations relative to the undamaged state. This approach converts sparse measurements into a pseudo full-field response and can reveal stiffness loss without direct access to the damaged area, but it requires accurate numerical models, appropriate sensor placement and careful treatment of operational and environmental variability; its robustness in real offshore conditions is still to be established.

Ultrasonic methods have been proposed by several authors. Sharma et al. (Sharma and Mukherjee, 2015) developed a non-contact, in-situ US guided-wave technique for underwater steel plates and explored its capability for estimating corrosion level. Chathurani et al. (Chathurani Thibbotuwa et al., 2022) designed a US-based corrosion sensor and performed field tests on OWT components to assess thickness loss. Thibbotuwa et al. (Thibbotuwa et al., 2022) further proposed an intelligent corrosion monitoring system for OWTs based on US pulse-echo and defined the system requirements listed in Table 5, while Silionis et al. (Silionis and Anyfantis, 2024) combined Monte Carlo simulation and a genetic algorithm to optimize sensor layouts. These studies demonstrate the feasibility of US-based corrosion monitoring, but also reveal practical challenges: non-contact methods face complex wave propagation and noise in real structures, whereas contact sensors need reliable long-term coupling and protection, which increases installation and maintenance efforts. The requirements in Table 5 mainly address hardware, communication and compactness; issues such as long-term power supply, data management, cybersecurity and integration into risk-based inspection and maintenance strategies are less discussed, and fixed low monitoring frequencies may be insufficient for faster degradation scenarios. All the aforementioned functional requirements will directly affect the total cost of the corrosion monitoring system.

OPEN IN VIEWER

Table 5.

Requirements of corrosion monitoring system proposed in (Thibbotuwa et al., 2022).

Aspects	Requirements
Applicable environment	Harsh environment, unattended for a long time
Way of working	Wireless transmission of field measurement
Monitoring frequency	Measure once a month and can increase when necessary
Size and weight	Be up to 6 × 11 cm with a weight of up to 100 g

To capture corrosion rate and related parameters, Ahuir-Torres et al. (Ahuir-Torres et al., 2019) integrated non-destructive and passive electrochemical technologies into a sensor interface for autonomous real-time remote sensing of OWT corrosion. Alonso-Valdesueiro et al. (Alonso-Valdesueiro et al., 2022) proposed a highly sensitive submarine corrosion sensor using thin steel tubes as probes. These developments show the potential of combining electrochemical and mechanical concepts with remote sensing, but most prototypes have only been tested under simplified exposure conditions and over limited time spans. Their performance with respect to drift, fouling, mechanical robustness and calibration stability over the full service life is still insufficiently documented. Moreover, most current solutions adopt a single-physics, single-sensor perspective, whereas corrosion in offshore structures is strongly coupled with mechanical loading, temperature and hydrodynamic action.

In summary, corrosion monitoring for offshore structures is evolving from local, periodic inspections towards more continuous and system-level solutions. Nevertheless, important challenges remain in achieving multi-scale coverage (from local defects to global structural impact), ensuring long-term robustness in harsh environments, providing adequate redundancy at system level and establishing standardized procedures for data interpretation and decision support. Addressing these issues and better integrating heterogeneous data with probabilistic and data-driven models will be crucial for deploying corrosion monitoring within digital-twin-based integrity management schemes.

Crack

Cracks are another major form of damage in offshore structures, typically initiating at stress concentration zones under cyclic loading. Fatigue crack initiation and propagation have been identified as dominant failure mechanisms for many critical offshore details. However, most existing monitoring schemes still target cracks that have already grown to a detectable size, and robust strategies for detecting early-stage crack initiation under realistic marine loading and environmental conditions are still limited. In addition, the combined influence of corrosion, variable-amplitude loading and multiaxial stresses on crack evolution is not yet fully reflected in practical monitoring and assessment frameworks.

For crack detection, both acoustic-based and model-assisted experimental approaches have been studied. Focusing on acoustic detection at the rear edge of OWT blades, Zhang et al. (Zhang et al., 2020) numerically analysed aerodynamic acoustic noise for cracked and uncracked flange sections and compared the results with a semi-empirical model. Ali et al. (Ali et al., 2021) established a finite element model of a T-joint, identified high stress/strain regions and successfully detected fatigue damage by combining FEM with experiments. Acoustic techniques are non-contact and can potentially cover large areas, but are sensitive to background noise, aerodynamic conditions and sensor placement in real offshore environments. FEM-supported methods, in contrast, can accurately identify critical regions and provide clear physical interpretation, but are localized, model-dependent and more difficult to scale up to full structures.

Based on numerical simulations, Hlaing et al. (Hlaing et al., 2022) used heuristics and partially observable Markov decision processes (POMDPs) to analyse crack propagation and proposed several crack growth models and failure criteria for optimizing the management of OWTs under fatigue degradation. POMDPs explicitly represent uncertainties in crack growth and inspection and link monitoring to risk-informed decisions, but their models are generally calibrated from limited data and simplified assumptions, and the computational cost of solving them can hinder direct application to large-scale systems.

Adedipe et al. (Adedipe et al., 2015) performed experiments on fatigue crack growth in S355J2+N steel under corrosive conditions and proposed a crack length monitoring method combining back-face strain (BFS), alternating current potential difference (ACPD) and direct current potential difference (DCPD). This multi-sensor fusion improves robustness in corrosive environments, but remains essentially local and requires sensors near anticipated crack paths, presupposing accurate prior knowledge of critical locations; long-term stability and maintenance needs in harsh marine environments still require further study.

Data-driven approaches have been increasingly used to quantify crack characteristics. Yang et al. (Yang et al., 2020) proposed a multiple back-propagation neural network using a signal difference coefficient (SDC) as input to predict crack length and angle, while Qiu et al. (Qiu et al., 2020) developed a BP neural-network method improved by GRNN and GA to gradually determine crack location and severity in OWT structures. These neural-network-based methods demonstrate that data-driven models can capture complex nonlinear relationships between signals and crack parameters and reduce explicit reliance on detailed physical models, but their performance strongly depends on the representativeness of training data, which are often obtained under simplified laboratory conditions. Dai et al. (Dai et al., 2022) proposed a crack identification method using multidimensional strain data and support vector regression (SVR), showing good agreement between identified and actual crack lengths and growth paths (Figures 6 and 7). SVR is comparatively robust for small datasets, yet still requires careful feature selection and parameter tuning.OPEN IN VIEWER

OPEN IN VIEWER

Figure 7.

Identification results of crack growth path.

In summary, for crack growth monitoring, available approaches can be broadly grouped into three categories: (i) direct measurement techniques that infer crack size or growth rate from local sensors (e.g. strain or potential difference); (ii) model-based methods that use numerical or analytical models to interpret structural response and predict crack evolution; and (iii) data-driven and decision-oriented frameworks that employ machine learning or probabilistic models to link monitoring results with inspection and maintenance planning. Direct methods are typically sensitive but intrinsically local and require prior knowledge of critical locations; model-based methods offer clearer physical insight but depend on model fidelity and boundary-condition assumptions; data-driven and decision frameworks can treat complex dependencies and support life-cycle management, but rely on representative data and careful validation.

Taken together, these studies show a clear shift from simple crack/no-crack detection towards quantitative characterization of crack growth and its integration into integrity management. However, most methods are still validated at specimen or numerical model level, and long-term field demonstrations on full-scale offshore structures are scarce. There is also a lack of unified benchmark datasets and performance metrics for comparing different monitoring approaches in terms of detection threshold, false alarm rate, robustness to environmental and operational variability and cost-effectiveness. Filling these gaps will be essential for turning current crack monitoring techniques into mature, widely applicable tools for the life-cycle management of offshore structures.

Local damage of offshore structures

Most of the research on the damage monitoring of offshore structures has so far been carried out using numerical simulation or model-scale tests, in which local damage is artificially introduced by weakening cross-sections or reducing the stiffness of individual members. These studies have clearly shown that local stiffness reduction or cross-section weakening is an effective way to investigate the sensitivity of different damage identification methods in a controlled setting and to understand their basic capabilities. However, most existing work is still confined to such predefined damage scenarios, robust strategies for detecting and quantifying local damage in full-scale offshore structures, under strong environmental and operational variability, are still not sufficiently established.

A considerable number of contributions fall into the category of model/response-based methods, which exploit changes in modal parameters or frequency response functions (FRFs), often combined with model updating. Shokrgozar et al. (Shokrgozar et al., 2022) employed signal energy decomposition techniques to localize damage by extracting features from the casing structure in both intact and damaged states. Sakaris et al. (Sakaris et al., 2021) simulated damage by reducing the stiffness of the connection points between the tendons and the storage tanks on the upper part of a platform, and proposed to use a single response signal and a random functional model to represent the dynamic characteristics of the structure in different states, thus achieving damage detection, localization and quantification. Zargarzadeh et al. (Zargarzadeh et al., 2020) proposed a new iterative cross-model cross-mode (CMCM) algorithm based on changes in the modal parameters of a platform steel frame to identify and locate damage. Similarly, Zhang et al. (Zhang et al., 2022b) identified damaged regions by comparing the modal characteristics of the structure in undamaged and damaged conditions, and thereby determined both the location and severity of the damage. Fathi et al. (Fathi et al., 2020) proposed a Bayesian model updating and damage detection framework based on incomplete FRF data; by incorporating FRFs at different excitation frequencies, the amount of information in the objective function was increased to alleviate the incomplete measurement problem, and the method was applied to damage identification in offshore structures, with its performance verified using simulated damage via local component stiffness reduction.

Overall, these approaches share the common feature of inferring local damage from global structural responses. Their main advantage lies in the fact that they can work with spatially sparse or incomplete measurements, and they offer an interpretable link between damage and changes in modal or FRF characteristics. However, their performance is highly dependent on the fidelity of the underlying numerical model, the quality of the identified modal parameters and FRFs, and the treatment of measurement noise and environmental effects. In addition, most case studies consider relatively simple damage patterns (e.g., single-location or severe stiffness reduction) and do not fully address multiple simultaneous damage locations or strong parameter uncertainties, which may limit their direct applicability to complex, full-scale offshore systems.

More recently, intelligent and data-driven methods have been introduced for local damage identification. Guo et al. (Guo et al., 2018) proposed a damage identification index that combines sixth-order modal identification results and natural frequency information, and then trained a neural network model for component damage prediction using the damage index and damage degree obtained from finite element simulations. Kuai et al. (Kuai et al., 2024) simulated damage by reducing element stiffness and employed two machine-learning regression models—support vector machine (SVM) and relevance vector machine (RVM)—to define cointegration relationships; damage in fixed steel platforms was then identified by combining anomaly detection with frequency-domain decomposition. To overcome the limitation of conventional artificial neural networks in capturing the temporal correlation between features and damage patterns, Wang et al. (Wang et al., 2024a) proposed an integrated deep learning method in which a squeeze-and-excitation network (SENet) attention mechanism is used to select the most informative features, and a bidirectional long-short term memory (BiLSTM) network is employed to learn the mapping between the extracted features and structural damage patterns. Numerical simulations and experiments on offshore jacket structures were carried out to validate the proposed method.

In summary, these machine learning methods mark a clear shift from purely model-based localized damage identification toward data-driven approaches that leverage high-dimensional response data and capture complex nonlinear relationships. Compared to classical modal analysis and frequency response function methods, they offer greater flexibility in handling diverse features and patterns. However, these approaches often face challenges such as the aforementioned issues with generalization across different operating conditions and the availability of benchmark datasets. Addressing these challenges is crucial for transforming existing localized damage monitoring technologies into mature, widely applicable tools for the full-lifecycle integrity management of offshore structures.

Conclusion

(1) From the perspective of marine environmental loads, existing work on wind, wave and ice monitoring has significantly improved the understanding of offshore load spectra and extreme events. Field measurements, CFD simulations and intelligent data-driven models have been widely used. However, most studies still treat wind, wave and current as separate loads, and full-scale multi-field coupled monitoring and identification remain limited. In addition, Reynolds number effects in wind and CFD studies, the generalization ability of intelligent wave models under extreme conditions, and the robustness of ice-load monitoring under poor visibility are not fully resolved.

Suggestion for future research and applications

Based on the above findings and the author’s understanding, the following suggestions are proposed for future research and engineering application of SHM in offshore structures.

(1) Develop multi-source, coupled load monitoring for wind–wave–ice environments. Future work should move from single load monitoring towards integrated monitoring of coupled environmental fields. For offshore engineering structures, this implies combining field measurements, CFD/CFD–FE coupled models and intelligent algorithms to jointly identify wind, wave and ice loads. Particular attention should be paid to (i) validating CFD and data-driven models against full-scale measurements at realistic Reynolds numbers, and (ii) constructing load descriptors that can be directly used as covariates in response/damage-based SHM indicators.

In summary, the future development of SHM for offshore engineering structures should focus on: (i) coupled and multi-scale monitoring of environmental loads and structural responses; (ii) offshore-specific sensor and virtual sensing system design with quantified long-term performance; (iii) integration of physics-based models, intelligent algorithms and probabilistic decision tools; and (iv) large-scale field validation and digital-twin-based integrity management. These directions are expected to promote the transition of SHM for offshore structures from concept demonstration to robust, widely adopted industrial practice.