基于神经网络的壳体结构损伤诊断研究

STUDY OF DAMAGE FAULT DIAGNOSIS IN SHELL BASED ON NEURAL NETWORK

神经网络输入参数的选择将直接影响工程结构损伤识别的精度和准确性。本文提出以反映结构损伤位置和程度的固有频率与频率下降率的组合作为神经网络输入的特征参数 ,以增加对损伤程度敏感的参数项 ,克服单独使用某种参数的缺陷。针对使用BP算法的多层感知器中存在的网络学习收敛速度慢 ,容易陷入局部极小点等问题 ,采用一个改进算法。并以门座起重机筒形支柱———圆柱壳结构损伤为例 ,进行计算分析 ,从中可以看出 ,采用此组合特征参数和改进算法提高了诊断的精度 。

The selection of neural network input parameters will influence the precise and accuracy of damage identification in structure. Some papers took inherent frequencies, frequency response functions, displacement and stress modal et al as neural network input parameters. From the point of the dynamic properties, the inherent frequencies were the most direct input parameters. But the ratio of frequency drops was more sensitive than inherent frequencies. In order to improve the sensitive parameters of damage degree, overcome the shortcomings of using single dynamic signatures, it was put forward to take the inherent frequencies and ratio of frequency drops as characteristic parameters of neural network inputs. The characteristic parameters can reflect the location and the degree of damage in structure. A new improved back propagation algorithm (IBP) was adopted aiming at the problems about slow convergence rate and easily to fall into part minimums in network studying of perceptions using back propagation algorithm (BP). Taking the cylindrical supporting shell of crane as examples, the damage fault occurred in it was analyzed adopting the characteristic parameters and IBP algorithms. It can be obtained that this method improves the diagnosis precise of the location and the degree of damage in engineering structure, and quickens the tempo of convergence.