| dc.description.abstract | Machine learning (ML), a subset of artificial intelligence (AI), has gained significant traction in engineering due to its capacity to enhance the analysis, design, and operation of complex systems by learning from data, identifying patterns, and making decisions with minimal human intervention. This popularity surge is driven by advancements in computational power, allowing the training of sophisticated ML models on large datasets derived from simulations, sensors, and operational histories.
This thesis uses a ML algorithm which predicts Gumbel parameters using statistical estimation and ML techniques. Initially, it reads data files containing significant wave height (Hs), spectral peak period (Tp), and a third structure-specific value, then selects subsets based on initial training points. Gumbel distributions are fitted to these subsets to estimate extreme values, with parameters validated with a stopping criterion. Using Gaussian Process Regression (GPR) and Kriging techniques, the algorithm predicts location and scale parameters for all sea states, iteratively updating based on convergence criteria to ensure accuracy and quantify prediction uncertainty. The process involves training the GPR model on initial data, making predictions, and evaluating these against a stopping criterion, iterating as necessary until reliable outputs are achieved.
The ML algorithm was tested on three different offshore platforms. The aim of the thesis was to evaluate the accuracy of the ML algorithm for different datasets. It was used to predict the offset for a mobile offshore drilling unit, displacement for a tension leg platform and the stress on a specific joint on a jacket. The platforms operate under various environmental conditions, and the study aims to test how this will affect the ML algorithm’s efficiency and accuracy.
For the mobile offshore drilling unit (MODU), the ML model shows good accuracy for location parameters but less so for scale parameters, with mean absolute percentage error (MAPE) stabilizing around 6% for scale and 2.4% for location after six iterations. For the tension leg platform (TLP)'s sway motion, the model predicts the location parameter accurately, with MAPE dropping to 0.19% after four iterations, while the scale parameter stabilizes at a higher MAPE of 9.38%. For the TLP's surge motion, the location parameter's MAPE drops to 0.49%, whereas the scale parameter remains high at 15.40% after ten iterations. The jacket's results show that scaling data to giga pascal (GPa) significantly improves accuracy, with MAPE for the location parameter at 0.13% and for the scale parameter at 12.80% after four iterations. The findings indicate that the ML algorithm generally predicts location parameters more accurately than scale parameters, and additional iterations beyond a certain point yield minimal improvement. | |
