摘要
通过理论推导 ,得到了模型参数误差均匀分布时五种输入向量相对误差的计算公式 ,这些公式为结构损伤识别神经网络的输入向量的选择提供了理论指导。理论分析表明 ,用归一后的频率变化比或按模态点归一的一阶模态损伤信号指标构造神经网络的输入向量 ,能有效地降低甚至消除均匀建模误差的影响。这些公式和结论在梁和框架损伤识别的数值算例中得到了证实。此外 ,本文对非均匀建模误差对神经网络输入向量及识别结果的影响进行了数值模拟和分析。
The influence of modeling errors on input vectors to artificial neural networks (ANN) is studied. Relative error formulas of five input vectors, applicable to the case of uniform modeling errors, are derived for selecting the input vectors to ANN. Theoretical analysis shows that the influence of uniform modeling errors may be minimized or even eliminated when {NFCRi} or {NDSI1} is chosen to be the input vector. All the formulas and conclusions are verified by numerical results for damage detection on a beam and on a frame. The numerical simulation analysis is performed for damage identification of the same frame with nonuniform modeling errors for comparison.
出处
《振动与冲击》
EI
CSCD
北大核心
2004年第2期39-43,共5页
Journal of Vibration and Shock
基金
国家自然科学基金项目 (599780 1 5)
