Abstract
Performance data from ships is subject to distributional shifts, sometimes referred to as concept drift. In this study, synthetic monitoring data is simulated for the KVLCC2, considering publicly available reference data and a semi-empirical simulation framework. Neural networks are trained to predict the required shaft power and to overcome the deterioration in model accuracy due to concept drift, three methods of incremental learning are applied and compared: (1) Layer freezing, (2) regularization, and (3) elastic weight consolidation. Furthermore, an implicit methodology for quantifying the changing hull and propeller performance is presented. In addition, a generic feature engineering framework is used for eliminating insignificant features. In two investigations, sudden and incremental concept drift scenarios are examined, and the effect of different uncertainty categories on model performance is studied in parallel based on three different datasets. As a main finding, it is confirmed that data quality is of great importance for accurate machine learning-driven performance monitoring — even in simulated environments. Furthermore, the study shows that freezing layers during incremental learning proves to be most robust and accurate, but it will be part of future work to examine this on actual sensor data.
Original language | English |
---|---|
Article number | 103619 |
Journal | Applied Ocean Research |
Volume | 138 |
Number of pages | 15 |
ISSN | 0141-1187 |
DOIs | |
Publication status | Published - 2023 |
Keywords
- Concept drift
- Hull biofouling
- Incremental learning
- KVLCC2
- Synthetic data