Performance Analysis of a Gas Carrier using Continual Learning in a Data Stream Context

In the present study, model drift – mostly driven by biofouling – is mitigated by applying continual learning and the hydrodynamic performance of a large LNG tanker trading in worldwide service is assessed in parallel. The available auto-logged sensor data spans five years including two dry docking intervals. An adaptive training methodology of an artificial neural network is established, and the drift score is derived from simulated data under sea trial conditions. In addition, hindcast metocean data from ERA5 is included in the model’s feature space for capturing environmental conditions. Crucially, it was found that the obtained drift estimate is in general accordance to ISO 19030 results, proving the method’s validity. Still, the methodology itself is subject to considerable uncertainty. Finally, limitations and possible extensions of the proposed methodology are presented.