Bedömning av skrovpåväxt genom datadriven modellering

The marine industry faces increasing pressure to reduce greenhouse gas emissions, as vessel fuel consumption remains a major contributor to environmental impact. One critical factor influencing energy efficiency is hull fouling—the accumulation of marine organisms on the ship’s hull—which leads to increased hydrodynamic resistance and higher propulsion power demand. Despite its practical importance, existing data-driven methods for monitoring fouling progression under real-world conditions remain limited in scope and reliability. Most vessels still rely on fixed-interval cleaning schedules, which can result in both unnecessary maintenance and avoidable fuel costs. This thesis addresses this challenge by developing a machine learningbased framework to estimate fouling-induced performance degradation from operational ship data, without relying on direct fouling measurements. Two complementary approaches were explored: (i) a residual-based supervised model trained on high-confidence clean-condition data, and (ii) an unsupervised representation learning model using autoencoders and clustering. The supervised approach proved highly effective. Using convex hull and quantile-filtering techniques to define clean baselines, the predictive regression models were trained to predict delivered power. The resulting residuals consistently increased during periods of presumed fouling and dropped sharply after major hull cleaning, demonstrating clear operational value. A forward-looking test using only early (clean) data for training showed that the model can detect fouling trends several weeks in advance, indicating real-world applicability for condition-based maintenance. The unsupervised approach, built on autoencoders and clustering, revealed operational patterns but did not consistently capture progressive fouling without auxiliary supervision. Incorporating a weak proxy signal (time since last cleaning) improved temporal ordering in the latent space, though this signal remained too noisy to serve as a standalone fouling indicator. Overall, the thesis presents a practical and interpretable data-driven framework for monitoring hull fouling using standard onboard measurements. The proposed framework enables condition-based maintenance decisions using existing onboard sensors and aligns with broader initiatives toward sustainable and efficient maritime operations.