基于多头自注意力机制和卷积神经网络的结构损伤识别研究  

Research on structural damage identification based on multi-head self-attention and convolutional neural networks

作　　者：张健飞[1] 黄朝东 王子凡 ZHANG Jianfei;HUANG Chaodong;WANG Zifan(College of Mechanics and Materials,Hohai University,Nanjing 210098,China)

机构地区：[1]河海大学力学与材料学院,南京210098

出　　处：《振动与冲击》2022年第24期60-71,共12页Journal of Vibration and Shock

基　　金：国家重点研发计划(2018YFC0406703)。

摘　　要：为了提高卷积神经网络(convolutional neural networks,CNN)的结构损伤识别性能,提出了一种以结构振动加速度信号为输入的基于多头自注意力的CNN模型。模型首先利用一维CNN学习加速度信号中的局部特征,然后利用多头自注意力机制关注输入数据中不同位置和不同表征子空间中的重要信息、学习信号中的全局特征,最后利用学习到的特征进行结构损伤模式识别。悬臂梁数值试验和振动台试验的结果显示出:相比于CNN模型、CNN-长短期记忆网络(long short-term memory,LSTM)联合模型和CNN-双向LSTM(bidirectional LSTM,BiLSTM)联合模型,基于多头自注意力的CNN模型复杂度低、易于训练,且具有更高的损伤识别精度和更强的抗噪性以及对于损伤特征相近的损伤模式具有更好的辨识能力。In order to improve the performance of the convolutional neural networks(CNN)model for structural damage identification,a CNN model based multi-head self-attention was proposed,which takes structural vibration acceleration signal as input.The acceleration signal of structural vibration was firstly fed into the model and processed with one-dimensional CNN to extract local features.The multi-head self-attention was then utilized to attend to important information in different positions and different representation subspaces of the input data to learn global features.All the learned features were finally used for structural damage pattern recognition.The results of numerical and shaking table tests of a cantilever beam show that compared with a CNN model,a CNN-LSTM(long short-term memory)joint model and a CNN-BiLSTM(bidirectional LSTM)joint model,the CNN model based on multi-head self-attention has lower complexity,easier training,higher damage detecting accuracy,stronger anti-noise capacity and better ability to distinguish damage modes with similar characteristics.

关 键 词：深度学习 多头自注意力 卷积神经网络(CNN) 结构损伤识别 

分 类 号：TU317[建筑科学—结构工程]