Nicolás Gorostidik bere tesia defendatuko du irailaren 15ean, astelehenean

Defentsa Adela Moyua aretoan egingo da, UPV/EHUren Leioako Campuseko Zientzia eta Teknologia Fakultatean, 11:00etan.

Nicolás Gorostidi matematikako eta estatistikako doktoregaia da, eta deep learning teknikak aplikatzen ditu itsasoko haize-sorgailu eoliko flotatzaileen amarratze-sistemen egiturazko egoeraren jarraipena egiteko. Matematika aplikatuak, mekanika konputazionala eta modelizazio estatistikoa konbinatzen ditu, itsasoko sistema eolikoetan mantentze prediktiboko estrategiak garatzeko. Era berean, Fluido Konputazionalen Dinamikako Masterra du Cranfield Universityn, eta Ingeniaritza Mekanikoko Gradua Oviedoko Unibertsitatean. Esparru akademikotik kanpo, jazz pianista, maratoi korrikalaria eta gozogintza zalea da.

“Structural Health Monitoring of Floating Offshore Wind Turbine Mooring Lines using Deep Learning” izeneko tesia David Pardo irakasleak (BCAM, Ikerbasque eta UPV/EHU) eta Vincenzo Nava doktoreak (Politecnico di Torino) zuzendu dute. Defentsa 2025eko irailaren 15ean egingo da, 11:00etan, EHUko Zientzia eta Teknologia Fakultateko Adela Moyua aretoan (Leioako campusa).

BCAM osatzen dugun pertsona guztien izenean, zure etorkizunerako onena opa nahi dizugu, bai arlo profesionalean, bai pertsonalean.

Abstract

Floating offshore wind energy has strong potential as a clean energy source, but high operational costs and harsh environmental conditions limit its widespread deployment. Maintenance is often slowed by difficult weather, and conventional monitoring techniques provide unreliable assessments of turbine structural integrity. Structural health monitoring (SHM), an inverse problem from a mathematical perspective, offers a promising solution, using sensor data to infer system properties. This dissertation focuses on applying deep learning to SHM for floating turbines, specifically to detect degradation in mooring systems.

To address this problem, we develop deep learning models to identify two types of mooring damage: biological fouling and anchor displacements. Due to limited real-world data, we simulate turbine behavior using OpenFAST and extract meaningful statistical features from the platform’s responses to various operational conditions. Deterministic models, including multi-layer perceptrons and autoencoders, are trained to detect anomalies and estimate damage severity, providing early warnings for potential mooring failure, and enabling predictive maintenance strategies before performance losses occur.

We explore common sources of underlying uncertainty arising from structural design, measurement error, and long-term operation. We implement methods to alleviate the effects of these uncertainties, and quantify the individual and collective impact of each source of uncertainty in our predictions.

A probabilistic autoencoder is then implemented to quantify damage as a probability distribution, capturing cases with multiple plausible solutions. By comparing deterministic and probabilistic approaches, we demonstrate how uncertainty influences model predictions and quantify the confidence of our model’s estimations.