利用振动模态测量值和神经网络方法的结构损伤识别研究

提出了一种基于模态测量参数和神经网络的结构损伤检测方法，建造了两种输入方式的BP神经网络，即自振频率以及结合自振频率与振型，并讨论了不同数量的输入信息对结构损伤检测精度和计算效率的影响。证明了输入的参数越多，神经网络就越聪明，训练的收敛速度越快；以及在保证一定的测量精度的情况下，基于频率与振型的损伤识别结果要好于基于频率的检测结果。最后，通过对3层框架模型的4种损伤工况下的结构损伤检测结果的分析，认为利用模态测量参数和神经网络方法能够准确地识别结构损伤的位置，而且能较精确地识别结构损伤的大小。

Neural networks-based structural damage detection through modal parameter measurements

The neural networks-based structural damage detection using measured modal data has been proposed. Two neural networks with different input modes (i.e., natural frequencies, the combination of natural frequencies and modal shapes) were built and analyzed in this paper, which show that BP neural network would be more intelligent and more quickly convergent with more input parameters. The combination of measured frequencies and modal shapes can improve the degree of accuracy of damage detection if the measured errors could be eliminated at most. A 3-storey steel frame model with 4 damage cases is used to verify the degree of accuracy and computation efficiency of the proposed approach. Numerical results show that the proposed approach can not only localize correctly the damage, but also identify the magnitude of damage with a relatively high degree of accuracy.