基于改进混沌粒子群的混合神经网络和蒙特卡洛法的结构随机灵敏度分析方法

STOCHASTIC SENSITIVITY ANALYSIS METHOD BASED ON HYBRID NEURAL NETWORK OF IMPROVED CHAOTIC PARTICLE SWARM AND MONTE CARLO

针对不确定性参数对结构力学性能的随机影响,该文利用混合神经网络良好的小样本学习和泛化能力构建结构响应复杂的函数关系,采用改进的混沌粒子群算法优化网络寻址结构。结合蒙特卡洛法对结构进行随机性分析,并根据该文提出的新的灵敏度度量参数计算随机变量的全局灵敏度系数。通过数学算例和工程算例验证了所提方法的可行性,且结构响应的概率分布曲线也可以真实的反应实际情况。同时,利用该文所提出的随机灵敏度计算方法可以更好的反应各随机变量对结构响应的相关性和敏感性。

In order to study the random effects of uncertain parameters on the performance of structure mechanics, the hybrid neural network possessing significant learning capacity and generalization capability at a small amount of information is used to construct the relativities of complex functions between the input and output samples. The networks would be optimized by the improved chaotic particle swarm for high efficiency and accuracy. According to the proposed method, global sensitivity coefficients of random variables can be calculated through random analysis by Monte Carlo simulation. In order to verify the feasibility of the proposed method, several examples are analyzed including mathematical examples and engineering examples. The results of these examples indicate that the proposed algorithm increases the precision and the response distributions of engineering examples could be reflected by the fitting probability density curves, meanwhile the structure response sensitivity and correlation could be better reflected based on the method of this paper.