Damage Detection Based on Wavelet Transform and Convolution Neural Networks

Abstract

The performance of conventional damage detection systems depends mainly on the physical and geometrical damage characteristics and the choice of damage classifier. Some works directly use Convolutional Neural Networks (CNN) for damage pattern recognition analysis of experimentally measured vibration signals. This work proposes a method that combines wavelet transform and CNN for Structural Health Monitoring (SHM). Firstly, we obtain numerically simulated structures with sensors arranged on them to collect data and perform the cut-off; then, we perform the wavelet transform to the acceleration signals of different simulated damage patterns and use them for training the CNN; finally, the trained CNN can predict the structural damage patterns. A four-level benchmark building introduced by the IASC-ASCE Structural Health Monitoring Working Group is used to validate this damage identification method. The numerical results show that the proposed method can effectively solve the problem of quantifying structural damage.