Due to high sensitivity to various imperfections, buckling loads of thin-walled cylindrical shells subjected to axial load vary dramatically. In order to predict lower-bound buckling loads for axially loaded cylindrical shells rationally, a probabilistic analysis approach named Probabilistic Random Perturbation Load Approach (PRPLA) is developed in this study. Firstly, a Back-Propagation Neural Network (BPNN) based method is established to describe measured imperfection patterns. Next, Random Single Perturbation Load Approach (RSPLA) is loaded upon BPNN-based depicted traditional imperfection patterns to construct a stochastic dimple imperfection. The aforementioned scattering traditional imperfections, as well as a variety of scattering non-traditional imperfections, are then sampled using Monte-Carlo simulation to generate cylindrical shell models differentiating from a nominal one. The probabilistic distribution of lower-bound buckling loads is obtained by finite element analysis. A nominal shell's realistic lower-bound buckling load is determined by choosing a specified reliability level lastly. The results show that describing measured imperfection patterns via BPNN is very close to real ones, and PRPLA presented is an improved method to find lower-bound buckling loads efficiently compared with NASA SP-8007 and many commonly used numerical approaches.

由于对各种缺陷的高度敏感性，承受轴向载荷的薄壁圆柱壳的屈曲载荷变化很大。为了合理地预测轴向载荷圆柱壳的下限屈曲载荷，本研究开发了一种称为概率随机扰动载荷法（PRPLA）的概率分析方法。首先，建立了一种基于BP神经网络的方法来描述测量的缺陷模式。接下来，将随机单扰动加载方法（RSPLA）加载到基于BPNN的传统缺陷模式上，以构建随机酒窝缺陷。上述散射的传统缺陷以及各种散射的非传统缺陷，然后使用Monte-Carlo仿真对样本进行采样，以生成与标称值不同的圆柱壳模型。通过有限元分析获得下界屈曲载荷的概率分布。最后通过选择指定的可靠性等级来确定名义壳体的实际下界屈曲载荷。结果表明，通过BPNN描述测量的缺陷模式非常接近真实模式，与NASA SP-8007和许多常用的数值方法相比，提出的PRPLA是一种改进的方法，可以有效地找到下界屈曲载荷。最终通过选择指定的可靠性级别来确定实际的下限屈曲载荷。结果表明，通过BPNN描述测量的缺陷模式非常接近真实模式，与NASA SP-8007和许多常用的数值方法相比，提出的PRPLA是一种改进的方法，可以有效地找到下界屈曲载荷。最终通过选择指定的可靠性级别来确定实际的下限屈曲载荷。结果表明，通过BPNN描述测量的缺陷模式非常接近真实模式，与NASA SP-8007和许多常用的数值方法相比，提出的PRPLA是一种改进的方法，可以有效地找到下界屈曲载荷。