Data-driven prediction for the hydrodynamic behavior of propellers using a clustering ensemble learning network
Abstract
Aiming to address the time-consuming challenges associated with hydrodynamic behavior assessment of propellers, this work provides an artificial intelligence model of propellers' hydrodynamic behavior called PlatProp. The basis of PlatProp is a proposed clustering ensemble learning network (CELN). The CELN consists of a bootstrap sampling layer, meta-learner layer, cluster layer, hidden layer, and output layer. We generated 2000 propeller CAD models by a parametric model as training data to achieve PlatProp. Specifically, Latin hypercube sampling (LHS) was used to select 2000 sample points across an extensive design space for the model's parameters. These sample propellers include uncommon hydrofoil profiles and blade geometries, thereby facilitating the exploration of innovative design possibilities. In addition, the Reynolds-averaged Navier-Stokes (RANS) method was utilized to calculate the hydrodynamic behavior of the aforementioned sample propellers, ensuring the acquisition of a high-high-fidelity dataset. Finally, the CELN model was constructed and trained using backpropagation, serving as the core of PlatProp. In result part, Extensive numerical investigations have been conducted to analyze the CELN method by constructing 60 CELN variants to evaluate its inherent performance. A two-way ANOVA method was performed to assess the effectiveness of the cluster layer. Furthermore, numerous numerical experiments illustrate that the PlatProp is competitive with the other classical machine learning models. An application of full-scale ship's propeller hydrodynamic behavior prediction was carried out. The results demonstrate that the CELN is competitive with existing classical machine learning methods. PlatProp can offer reliable hydrodynamic behavior for propellers.