考虑残差学习的深层卷积神经网络在结构损伤识别中的应用研究

Application Research on Deep Convolutional Neural Network Considering Residual Learning in Structural Damage Identification

提出了一种考虑残差学习的深层卷积神经网络损伤识别方法,并将其应用到框架结构节点损伤识别中。采用试验研究方式对所提方法进行了深入探讨,结果表明该方法可以很好地解决网络深化带来的网络退化或梯度爆炸、弥散导致的收敛困难和识别准确率差等问题,能对结构损伤诊断中的损伤定位这一复杂问题进行有效识别。在对试验框架节点损伤位置识别的对比研究中,考虑残差学习的深层卷积神经网络收敛速度和准确率均高于常规浅层神经网络和深层神经网络,有极高的准确率和稳定性,从而使得对于工程中复杂结构损伤诊断所需要的更深层、更复杂网络的搭建成为可能。此外,为提升网络用训练样本的质量和数量,依据样本划分规律提出了一种新的数据样本扩增方法,该方法在相同条件下可以显著增加用以训练的样本量并能弱化数据截断带来的信息缺失,识别准确率和收敛速度也大幅提高,研究显示了该处理方式的有效性和适用性。

A deep convolutional neural network damage identification method considering residual learning was proposed and applied to the damage identification of the frame structure joints.The proposed method was deeply discussed by means of experimental research,and the results showed that this method could solve the problems of convergence difficulty and poor recognition accuracy caused by the network degradation and gradient explosion,dispersion problems when the network deepening.In the comparative study of joint damage identification of test frame,the convergence speed and accuracy of deep convolutional neural network considering residual learning were higher than those of shallow conventional neural network and deep neural network,and had high accuracy and stability,and increased the possibilities to build a deeper and more complex network for damage diagnosis of complex structures in engineering.In addition,in order to improve the quality and quantity of training samples for network,a new data processing method was proposed according to the law of sample division.This method could significantly increase the sample size for training,weaken the information loss caused by data truncation under the same conditions,and greatly improve the recognition accuracy and convergence speed,and the research showed its effectiveness and applicability.