| The high-precision CFD modeling technology for the strong nonlinear motion of damaged ships in waves faces the bottleneck challenges of long time-consuming and low efficiency.Focusing on some key problems in the regression prediction framework of machine learning model,combined with CFD modeling technology,by introducing k-fold cross validation method and combinatorial learning concept,an adaptive and sequential parallel sampling method(KAPS)and a pointwise ensemble metamodeling method(Ensemble)are developed,It is used to predict the two degree of freedom motion response of damaged ships in regular beam sea,and the inputoutput function mapping relationship of nonlinear motion of damaged ships is reconstructed by combined agent model.The research results of this paper provide a new idea for the stability evaluation of damaged ships under different wave conditions.It has the advantages of minimizing the modeling times of response amplitude operator(RAO)analysis process of damaged ship motion and minimizing the dependence of CFD modeling process on users’ skills.It has the advantages of simple concept and convenient implementation.The main research work and conclusions of this paper are as follows:(1)In order to improve sampling efficiency and sampling quality,an adaptive and sequential parallel sampling method(KAPS)based on agent model technology is developed.This method can not only adaptively sample according to the response information of the target problem,but also consider the nonlinearity and sparsity of the target problem samples.KAPS uses Thiessen Polygon method to divide the design space into multiple sub regions,and there is only one sample in each sub region.The k-fold cross validation error is used as the sample error criteria,and the sub region space size is used as the sample sparsity criteria.The error criteria and sparsity criteria are linearly weighted to jointly determine the target polygon area.In addition,the number of parallel samples that KAPS can obtain through one iteration can be determined according to the actual computing resources,that is,multiple target polygon regions can be determined in one iteration process,and only one sample can be selected for each sub region.The selection of new samples takes into account both the distance from the samples in the same area and the distance from the samples added in the same iteration process,and avoids the aggregation of new samples at the boundary of each sub area by setting the distance threshold and the maximum and minimum distance criterion.The test shows that compared with the existing typical adaptive and sequential sampling methods,KAPS can improve the sampling quality.In addition,most of the existing adaptive and sequential sampling methods are only suitable for single response systems,while the actual engineering problems are often multi response systems.There is a serious mismatch and inapplicability between the existing technology and the actual needs.Therefore,this paper develops an adaptive and sequential parallel sampling method(MKAPS)for multi response systems.(2)For the research on the modeling method of ensemble metamodel,firstly,the prediction performance of single agent model(such as artificial neural network,support vector regression,radial basis function interpolation and Kriging interpolation model)is studied,and the internal hyperparametric parameter adjustment laws of each single agent model are summarized.Then,based on four base metamodels,a pointwise weight ensemble metamodeling method(Ensemble)is developed: the training sample point weight is determined point by point by using the0-1 strategy,the unknown sample point weight is obtained by inverse distance interpolation,and the ensemble metamodel is obtained by linear weighted combination.The effects of the number of training samples and the number of base metamodels on the prediction accuracy of the ensemble metamodel are further analyzed.The first mock exam of 8 test functions and Taylor series of residual resistance coefficient sets shows that the Ensemble method can give full play to the local precision advantage of each base metamodel and restrain the model with poor local prediction accuracy.Regardless of the size of training samples,the pointwise ensemble metamodel always has incomparable advantages over the existing typical ensemble metamodeling methods and base metamodel,and when the number of base metamodels is maintained at 4,the comprehensive prediction ability of the ensemble metamodel is the best.(3)Based on the adaptive and sequential parallel sampling method and the construction method of pointwise ensemble metamodel proposed in this paper,combined with CFD numerical simulation technology,the motion response prediction of two degrees of freedom(roll and heave)of DTMB 5415 ship model under the condition of zero speed and damaged side double compartments in regular beam sea is carried out.Compared with the CFD calculation results,the prediction error of roll motion response is 1.88%,and the prediction error of heave motion response is3.56%.On the premise of acceptable prediction accuracy,the ensemble metamodeling technology significantly improves the prediction efficiency of damaged ship motion response and realizes the rapid prediction of damaged ship motion response in waves. |
