| With the increasinged application of Computational Fluid Dynamics(CFD)in Ship Hydrodyanmics,uncertainties are becoming a more important problem and challenging issue.How to improve the CFD techinuqes in Ship Hydrodyanmics is a long-standing problem and an investable demand in order to raise the design level.In this dissertation,we outline an investigation of two important uncertainty sources including turbulence model and hull form geometry around the numerical simulation of hull-propeller interaction.A framework for forward uncertainty propagation and revesed uncertainty quantification is proposed for uncertainties quantification including parametric uncertainty,model form uncertainty and geometric uncertainty.Firstly,a uniform deep Bayesian framework for forward uncertainty propation and reversed uncertainty quantification is proposed based on variational autoencoders(VAE)according the requirement of uncertainty analysis and specific high-dimensional data for the quantities of interests.The framework is designed with extensible ability to incorporate different forms of encoders and decoders for different application scenarios and well verified.Results show satisfactory ability in learning the high-dimensional relationships between inputs and outputs of non-linear system.On this basis,different turbulence closures are compared from the perspective of wake field prediction and the parametric uncertainty analysis of a frequently-used two equation model is performed with the proposed framework.A gaussian process model is adopted as the encoder and2 D convolutional neural network is selected for decoder.Additionally,the iterative ensemble Kalman filter is applied to estimates the posterior distribution of model paramters with highfidelity PIV obersevations.The non-parametric analysis for the model-form uncertainty is perform to overcome the shortcoming of linear eddy-viscosity model.The representation of model-form uncertaint is obtained based on anisotropy tensor decomposition of Reynolds stresses.Realizability analysis of several commen turbulence models are performed and Reynolds stressed tensor perturbation is achieved by perturbating the eigenvalue and turbulence kinetics.The proposed deep Bayesian framework is extended with 3D-CNN for the propagation of turbulence uncertainties.Similarly,the iterative ensemble Kalman filter is applied to estimates the posterior distribution of Reynols stressed with high-fidelity PIV obersevations.For the geometric uncertainty of hull form,a series of deep encoders and decoders based on NURBS representation and modified 3D Fisher vectors are proposed for uncertainty propagation of system inputs.On this basis,different scales of geometric perturbation of the hull form of ship stern region are analyzed to investigate the propagation from geometric perturbation to ship resistance and wake field.Finally,a coupled viscous-potential approach is applied to analyze the propagation of turbulence model uncertainty and geometric uncertainty in hull-propeller interaction.Two additional deep Bayesian model is constructed accounting for the effect of hull to propeller performance and the effect of propeller to ship resistance.On this basis,the application mode of the proposed uncertainty framework is illustrated for different reference ship.The ability of the proposed method is well-verified to have application potential in ship design. |
