Several Sequence Sampling Strategies For Single Surrogate Model And Hybrid Surrogate Model And Their Applications

Structure design of hull is a multi-parameter and highly complex design problem,accompanied by a large number of expensive and time-consuming computational simulation or experiments.Surrogate model technology can make use of high-precision simulation results of a small number of design schemes to predict the response results of any design scheme in global design space.It can provide design assistance to the structural designers.How to improve and guarantee the prediction accuracy of the surrogate model is an important key problem.This thesis studies the single surrogate model based and the hybrid surrogate model based sequence sampling strategies.The main work contents of this thesis are as follows:In the study of updating surrogate model,it is highly appreciated that the sample points are uniformed and the local regions with high-nonlinearity are specially considered.Therefore the concept of curve distance between spatial curves is proposed,which includes the information of both design space and response space.Referring to the minimum distance maximization method,the minimum space curve distance maximization criterion is proposed,which is independent of types of surrogate models and therefore has the generalization.The effectiveness of the proposed method is demonstrated by testing 17 test functions.Results show that the new sampling criterion has the higher overall prediction accuracy of the final surrogate model than the one of the classical MSE criteria.After the basic theory of hybrid surrogate model is concisely introduced,a mixing strategy is described.The strategy can give the variable weighted for different surrogate models in different design space regions by using the generalized cross-validation method.By combining the hybrid surrogate model and the sequence sampling method,two sequence sampling methods are proposed.In the first one,the sub-surrogate model based sampling is first conducted and sampled points within a certain distance are then integrated and the hybrid surrogate model is finally built based on the updated sub-surrogate models.The second one is the hybrid surrogate model based sequence sampling,in which the sampling points are obtained by using the hybrid surrogate model itself rather than the sub-surrogate models.Numerical experiments are conducted on some mathematical test functions.Results show that the final hybrid surrogate model updated by the hybrid surrogate model based sequence sampling has higher overall prediction accuracy than that of the sub-surrogate model based sampling strategy.In some practical problems,it usually pays more attention to the response prediction accuracy in certain regions.Therefore a sampling strategy for local semi-closed region and closed region is proposed.Sigmoid function and normal distribution function are introduced as complementary coefficients to adjust the importance of different regions in order to achieve the local important sampling strategy.The numerical experiments on eight mathematical functions show that the proposed local important sampling strategy makes the higher prediction accuracy than whole space sampling strategies in the concerned local areas.Finally,the presented single surrogate model based sequential sampling strategy,the hybrid surrogate model based sequential sampling strategy and the local important sequential sampling strategy are applied to predict the maximum bending stress of a typical double bottom grillage structure.The crossed beam system model is used to calculate the stress of the double bottom structure,and seven parameters are employed as the input of the surrogate model.The results show that all above methods improve the prediction accuracy of the surrogate model greatly.The proposed sequence sampling strategy based on the minimum space curve distance maximization criterion,hybrid surrogate model based sequence sampling strategy and sequence important sampling strategies can improve the prediction accuracy of surrogate model under different requirements.The works done in the thesis provide useful references for fast prediction of time-consuming simulation in practical engineering design.