噪声与振动控制

噪声与振动控制 ›› 2021, Vol. 41 ›› Issue (5): 127-133.DOI: 10.3969/j.issn.1006-1355.2021.05.021

• 信号处理与故障诊断 • 上一篇    下一篇

一种基于LSTM循环神经网络和振动测试的结构损伤检测方法

王子凡，张健飞   

  1. （ 河海大学力学与材料学院，南京211100 ）
  • 收稿日期:2021-01-04 修回日期:2021-02-23 出版日期:2021-10-18 发布日期:2021-10-18
  • 通讯作者: 王子凡

A Structural Damage Detection Method Based on LSTM Recurrent Neural Network and Vibration Testing

  • Received:2021-01-04 Revised:2021-02-23 Online:2021-10-18 Published:2021-10-18
  • Contact: Zi fan Wang

PDF

摘要：

本文提出了一种基于长短期记忆（LSTM）循环神经网络的损伤检测方法，通过直接提取结构动态测试时域数据中的特征实现结构的损伤识别，基于不同损伤情况的重力坝有限元模型生成的加速度数据对LSTM网络进行在不同噪声水平下进行训练和测试，采用网格搜索方法对网络超参数进行优化。数值试验和试验室悬臂梁振动试验结果表明基于LSTM的损伤检测方法具有很高的损伤识别准确率和抗噪能力，其性能相对传统的循环神经网络（RNN）和门控循环单元（GRU）神经网络，不同损伤工况测试准确率均有提升，最高达16.25%。

关键词: 故障诊断, 长短期记忆循环神经网络, 损伤识别, 加速度, 网格搜索, 抗噪性

Abstract:

In this paper, a damage detection method based on long-short term memory (LSTM) recurrent neural network is proposed. The structural damage identification is realized by directly extracting the features in the time-domain data of structural dynamic test. The LSTM network is trained and tested under different noise levels based on the acceleration data generated by the finite element model of gravity dam under different damage conditions The grind search method is used to optimize the network parameters. The results of numerical test and cantilever vibration test show that the damage detection method based on LSTM has high damage identification accuracy and anti noise ability. Compared with the traditional recurrent neural network (RNN) and gated recurrent unit (GRU) neural network, the test accuracies of LSTM on diffferent damage conditions increase and the maximum improvement is up to 16.25%.