An attention-neural-processes approach to reconstructing unsteady water load for seaplanes.

The distribution of holographic unsteady water load offers important information for evaluating the hydrodynamic performance of seaplanes. However, traditional tank test is limited by the number of sensors that can be deployed on the bottom of hull, thus only providing sparse data for estimating load distribution and leading to inaccurate evaluation of seaplane performance. To achieve accurate and rapid reconstruction of holographic load distribution, a machine learning load-reconstruction model based on Attention Neural Processes is proposed. This model performs spatiotemporal modeling of seaplane water load utilizing sparse sensor data. It directly learns the load patterns across multiple time steps and employs Attention modules to capture the spatial distribution of load. Comparisons with alternative methods demonstrate the model's superior ability to simultaneously capture spatial and temporal dependencies of the unsteady load data. In addition, the model's robust generalization capability is also validated by reducing the number of sensors in the training data. The results indicate that the proposed model exhibits high prediction efficiency, accuracy, and generalization for the reconstruction of unsteady water load distribution, which is of great significance for comprehensively evaluating the hydrodynamic performance of seaplanes. [ABSTRACT FROM AUTHOR]