Search

Generalization remains a significant challenge in the application of deep learning methods, particularly in predicting mooring line tension for moored floating structures. The dynamic and complex nature of marine environments makes accurate prediction of mooring line tension a formidable task. Existing deep learning models, constrained by the data distribution of their training sets, typically perform well within the scope of these datasets. However, their limited generalization capability becomes apparent when confronted with unknown and changing marine conditions, resulting in a decline in prediction accuracy. To address this issue, this paper proposes a Domain Generalization (DG) model approach, which incorporates a distribution matching regularization term to the model's loss function, aimed at reducing distributional disparities between data from different domains. The proposed method was compared with existing Deep Neural Network (DNN) model method through validation with engineering case studies. The results indicate that the DG model method significantly surpasses the DNN model in terms of generalization capabilities. The mooring line tension prediction model developed herein holds potential for constructing digital twins, offering near real-time data support for moored floating engineering projects. This facilitates timely and accurate decision-making in response to environmental changes. • Develops a Domain Generalization model to predict mooring tension in unknown sea conditions. • Demonstrates improvements in model generalization, surpassing traditional Deep Neural Network (DNN) methods. • The effectiveness of the model was validated through a practical engineering case study of a floating wind turbine. [ABSTRACT FROM AUTHOR]

Irregular ocean waves contribute to random and non-linear hydrodynamic forces on marine and offshore platforms, in particular those made of slender bodies such as jack-ups. Keeping track of the structural fatigue, integrity, and reliability renders benefits to these platforms such as enabling predictive maintenance and risk-based decision support. While computational techniques are useful in providing insights to deepen our understanding of the physics involved, these are resource-intensive and not suited for predictive capabilities. Here we introduce an effective deep learning method based on Long Short Time Memory (LSTM) to predict unsteady hydrodynamic forces on jack-up legs, which are mainly governed by non-linear drag forces, as well as dynamic responses of the jack-up in terms of reaction forces and hull motion. The numerical OF-ABAQUS solver (one-way fluid-structure coupling) designed for jack-up by Li et al. (2022, 2023) is used to generate a database for the deep neural network, which takes in wave elevation time series as the input. A leaky non-linear rectification is used to trace the relationship between wave elevation (as the input) and hydrodynamic forces as well as structural responses (as the outputs). The LSTM model is trained using a stochastic gradient descent method, and the results are compared with the full-order numerical simulations. Once the hyper-parameters of the LSTM algorithm are identified, further tuning for each sea-state is not required. Instead, the same LSTM algorithm can be used for different sea-states. The predictions made with our LSTM model are real-time, 96,000 times faster than the full-order simulations, within a 6% error threshold, and use a negligible amount of central processing unit (CPU) power. This powerful tool can be used to obtain real-time predictions using computers on offshore platforms without high-performance computers inland. Further, the proposed LSTM-based method can be applied to create digital twins for jack-up platforms. [ABSTRACT FROM AUTHOR]

The significance of understanding the flow past a bluff body (BB) lies in its relevance to ocean, structural, and environmental applications. Capturing the transient flow behaviors with fine details requires extensive computational power. To address this, the present study develops an improved method for modeling the complex flow dynamics around a BB under steady and unsteady conditions. It is a deep learning (DL)-enhanced reduced-order model (ROM) that leverages the strengths of proper orthogonal decomposition (POD) for model reduction, convolutional neural network-long short-term memory (CNN-LSTM) for feature extraction and temporal modeling, and Bayesian optimization for hyperparameter tuning. The model starts with dimensionality reduction, followed by DL optimization and forecasting, and terminates with flow field reconstruction by combining dominant POD modes and predicted amplitudes. The goal is to establish a DL-driven ROM for fast and accurate modeling of the flow evolution. Based on the comparison of millions of data samples, the predictions from the ROM and CFD are considerably consistent, with a coefficient of determination of 0.99. Furthermore, the ROM is ∼10 times faster than the CFD and exhibits a robust noise resistance capability. This study contributes a novel modeling approach for complex flows, enabling rapid decision-making and interactive visualization in various applications, e.g., digital twins and predictive maintenance. • Developing a deep learning-enhanced reduced-order model for transient flow modeling. • Establishing a Bayesian-optimized CNN-LSTM for accurate flow forecasting. • Investigating spatiotemporal bluff body flows under steady and unsteady conditions. [ABSTRACT FROM AUTHOR]

Blow-outs occurred on offshore platform and associated fires have been recurrent during the previous few decades, and poses a potential safety hazard to humans, property and the surrounding environment. Although the real-time forecast based on deep learning have shown promise in the fields of fire modelling and hazardous area evaluations, jet fire spatio-temporal modelling has not yet undergone sufficient investigation in complex ocean engineering cases like offshore platforms. This research therefore proposes a deep learning-based framework for jet fire spatio-temporal probabilistic real-time forecast by developing the Hybrid-VB-ConvSTnn model, which integratesConvGRU and variational Bayesian inference. And the significant hyperparameters were locally optimized through sensitivity analysis and finally identified as Monte Carlo (MC) sampling number m = 100 and dropout probability p = 0.1. By performance comparison with different models, the Hybrid-VB-ConvSTnn model shows competitive spatio-temporal forecasting capabilities in terms of both real-time (Inference time = 0.83s) and accuracy (R 2 = 0.982). Moreover, the Hybrid-VB-ConvSTnn model could provide the additional uncertainty inferences based on the probability density of the Bernoulli distribution, which avoids the inherent shortcomings of "overconfidence" for traditional point-estimate models and lends credibility to flame boundary identification. The proposed framework could support the digital twin-based fire emergency management on offshore platforms by more comprehensive and robust risk evaluation. • Deep learning-based advanced approach framework for natural gas jet fire spatiotemporal probabilistic forecast is developed. • Uncertainty inference in flame boundary forecast can be determined with variational Bayesian inference. • Hyperparameters that determine model performance are locally optimized through sensitivity analysis. • Real-time forecast provides an alternative for digital twin model of jet fire accidents management on offshore platform. [ABSTRACT FROM AUTHOR]

Engineering has many necessary fields, and Structural Health Monitoring (SHM) is one of the most important of them. Sometimes in industrial environments, it is difficult and even impossible to collect data containing different real damages. Therefore, the problem of data acquisition represents a primary challenge in designing damage detection systems. The application of digital twin methods based on simulated models and/or Machine Learning (ML) models is a practical way to solve this problem. In this approach, a digital twin is generated for a compromised structure, utilizing a physics-based model to analyze diverse damage scenarios. Subsequently, an ML model is trained using data extracted from the physics-based model, functioning as the digital twin. This research proposes a method based on a digital twin for detecting damages in structures. The data produced from a Floating Wind Turbine (FWT) model was used to evaluate the performance of the proposed digital twin-based method. For this purpose, the FWT structure was simulated using a numerical model to address the data collection problem in the face of various uncertainties, such as changing loading conditions. In line with the concept of digital twin and to reduce the computational time, a Deep Convolution Long Short-Term Memory Neural Network (DCLSTMNN) model was designed and trained only with the frequency data of various scenarios of the simulated FWT model under constant loads (deterministic loads, including constant wind speed and airy wave model) to learn the damage-sensitive features. Then, to demonstrate the robustness of the proposed model under different uncertainties, the DCLSTMNN model was evaluated using the frequency data of the simulated FWT model under variable loading conditions (including Kaimal wind model and JONSWAP wave theory). Some vibration response components unrelated to the nature of the FWT model were removed using the Complete Ensemble Empirical Mode Decomposition (CEEMD) method. Then, the reconstructed vibration responses were used to create the frequency data using the Frequency Domain Decomposition (FDD) technique. The study results show that the proposed digital twin-based method can detect the location and severity of damage more accurately than other comparable methods despite various uncertainties. • A novel digital twin-based method is proposed for learning damage-sensitive features from data of various scenarios of the simulated simple FWT model to detect damage in the complex FWT model. • A Deep Convolution Long Short-Term Memory Neural Network (DCLSTMNN) model is designed to learn features and damage detection of the complex FWT model. • The proposed DCLSTMNN is trained with extracted data from the simulated simple FWT model and then tested with extracted data from the complex FWT model. [ABSTRACT FROM AUTHOR]