【作者】 王广帅；

【导师】 郑世杰；

【作者基本信息】 南京航空航天大学 ， 工程力学， 2009， 硕士

【摘要】 在利用神经网络进行复合材料结构健康监测的研究中,一些学者只从纯数值仿真角度验证了神经网络的模型辨识能力,并没有进行实验验证,而缺少实验依据的纯粹计算机仿真没有足够的说服力;也有一些学者通过实验采集各种损伤情况的特征信息,以此构建训练样本,但是单纯依靠实验方法获取大量训练样本是非常困难的。鉴于上述情况,本文提出借助模型修正技术,结合模态分析实验结果自动校正有限元模型,从而建立一个较为符合实际情况的数学模型,由修正后的有限元模型得到的各种损伤情况下结构的特征信息,以此构建神经网络的训练样本。本文制备了两个碳纤维增强复合材料试验件,并进行了模态分析实验,测得两个试验件的前五阶弯曲模态频率;根据复合材料结构力学以及边界条件的近似处理建立了复合材料梁结构的有限元模型;由于材料参数和边界条件是有限元建模中存在的主要误差,将遗传算法作为优化工具,结合模态分析实验测得数据分别对材料参数以及边界参数进行了优化修正,修正后的有限元模型具有很高的精度;利用修正后的有限元模型得到复合材料结构含不同脱层损伤的前五阶弯曲模态频率,频率信息经过规格化处理后作为训练样本对混合递阶遗传RBF神经网络进行训练;将实验测得数据送入训练好的混合递阶遗传RBF神经网络进行预测,实现了对复合材料结构脱层损伤的定位和损伤程度的评估。更多还原

【Abstract】 The research on health monitoring for composite structures based on neural network is comprehensively discussed in this paper. Some researchers have verified the ability of neural network to identify the model from numerical method without any experiments, yet this is not enough. Some other researchers collect features of a variety of structural damages by experiments to obtain samples for training neural network, but the number of samples is not enough. So the model updating method based on modal analysis experiment is used to acquire a more realistic model. Features of various structural damages are derived by the model to form enough samples for training neural network,Firstly, two carbon fiber reinforced composite beams are fabricated, and their first five flexure modal frequencies are measured by an experiment method. Then, the initial model based on composite structural mechanics is updated by genetic algorithm and frequencies obtained by modal analysis experiment. The material parameters of composite structure are optimized and revised, as well as the boundary condition. Then, samples to train the neural network are obtained by the revised FEM model. Finally, the first five flexure modal frequencies obtained by experiment are input to the neural network to predict the location and extent of demalination.更多还原