陈松坤，王德禹

上海交通大学 海洋工程国家重点实验室; 高新船舶与深海开发装备协同创新中心， 上海 200240

通讯作者:陈松坤（1994-），男，湖北省十堰市人，硕士生，主要研究方向为船舶与海洋工程结构物设计与制造. 通信作者：王德禹，男，教授，博士生导师，E-mail: dywang@sjtu.edu.cn.

An Improved Monte Carlo Reliability Analysis Method Based on Neural Network

CHEN Songkun,WANG Deyu

State Key Laboratory of Ocean Engineering; Collaborative Innovation Center for Advanced Ship and Deep-Sea Exploration, Shanghai Jiao Tong University, Shanghai 200240, China

摘要/Abstract

摘要： 在结构可靠性分析中，蒙特卡罗（MC）是最准确的方法，但是对大量样本点的精确计算限制了它在工程实际中的应用.为了减少分析次数，以BP(Back Propagation)神经网络技术为基础，提出了一种改进的MC方法（BP-MC）.该方法通过进行实验设计（DOE）构建BP模型，以权重因子和到失效面的距离作为筛选准则，从MC样本点中筛选出失效面附近的点添加至训练集，重新训练BP模型直至满足收敛准则.随后以该BP模型识别样本点是否处于失效域，从而计算结构的失效概率.最后,以数学模型和加筋板极限强度可靠性计算为例，验证了BP-MC算法的准确与高效.
关键词: 可靠性, 蒙特卡罗抽样, 神经网络, 加筋板, 极限强度, 船舶结构
Abstract: Monte Carlo (MC) is a very accurate method in the structure reliability calculation, however, its application is limited due to a large number of computation when it comes to complex engineering structures. It is time-consuming even in a single analysis. To reduce the calculation, the neural network approach is adopted to construct the BP-MC method. The back propagation (BP) neural network is built through design of experiments (DOE), then the weighting factors and the distance to failure surface are used as filters to pick up the design points out of the MC points. Those picked points are prone to cause the structure failure, and transferred into the training set to update the BP model. The filter-update process continues until the convergence of the BP, and then reliability index is calculated with the BP model on the MC points. The efficiency and usability are elucidated with a mathematic model and a stiffened panel model at the end of this paper.
Key words: reliability, Monte Carlo sampling, neural network, stiffened panel, ultimate limit strength, ship structure


PDF全文下载地址:

点我下载PDF