Reliability-based Design Optimization Of Ship Structures Based On SMOTE And Surrogate Model

The lightweight technology of ship structures,which meets structural performance requirements and achieves the goal of minimum quality,can not only reduce the consumption of construction steel to reduce the total cost,but also improve the ship maneuverability,thereby obtaining a structural form that meets both economics and safety requirements.The ship’s structural composition and loads are complicated,and the finite element method is often used to obtain the relationship between the structural parameters and the response with high accuracy.However,repeated iterations are required during the optimization process,resulting in long time and high calculation costs.Therefore,it is necessary to find a surrogate model technology that improves the optimization efficiency and has a certain accuracy.At present,the optimizations of ships and other engineering structures are mostly based on deterministic structural parameters.But there are uncertain factors in the actual structures in terms of material properties,structural dimensions,and external environment,sometimes making deterministic design solutions fail to meet safety requirements.Thus structural optimization design based on reliability is necessary.In this paper,in order to solve the problem of low calculation efficiency and difficult convergence in conventional reliability-based design optimization（RBDO）,the oversampling method SMOTE algorithm and dynamic neural network are used for RBDO.In the single loop method（SLA）of RBDO,the simulated annealing method is used as the optimization algorithm to get the current optimal solution.By using the SMOTE algorithm during the iteration process,the sample points near the failure surface around the current optimal solution are gradually increased,and then the BP neural network model is updated.Thus the local accuracy of the model is gradually improved and it makes the network model better approximate the limit state function near the global optimal solution.Finally,it is applied to mathematical examples and RBDO of real ship section.The calculation results show that the method can effectively obtain the optimal solution that meets the accuracy and reduce the calculation cost.Aimed at the requirement of the approximate accuracy of the failure surface in the RBDO,an improved point adding strategy based on the SMOTE algorithm is proposed.In the iterative process,a large number of candidate sample points are synthesized using the SMOTE algorithm.The satisfing new sample points are selected by the condition that the evaluation function is less than a fixed value and then delete the points that are too close.Add these points to the training set to update the surrogate model.It can obtain higher approximation accuracy for the failure surface with fewer sample points.The improved point adding strategy and neural network model are integrated into the single loop method of RBDO,and the simulated annealing method is used as the optimization algorithm.The points that meet the point adding criteria and the current optimal solution are added to the training set to update the surrogate model.The proposed strategy is used to a real ship section.The results show that the method based on neural network model and dynamic point adding strategy has better accuracy while reducing the number of finite element calculations,and effectively solves the RBDO of complex structures.