.png){kind=icon}
.png){kind=icon}
Return
Explainable 1-D convolutional neural network for damage detection using Lamb wave
.png){kind=icon}
PRE
AI
DOI：10.1016/j.ymssp.2021.108220.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Recently researchers are exploring the data-driven machine learning techniques to identify damage from Lamb wave response. This is primarily because the physics based model are often difficult to implement for complicated aerospace structures. In this study, one-dimensional convolutional neural network (1D CNN) is used for automated damage detection using Lamb wave response of a thin aluminium plate. An attempt is made to interpret the 1D CNN model in terms of damage feature contributions using Local Interpretable Model-Agnostic Explanations (LIME) and find that the interpretation corresponds to insightful damage signature. The database consists of finite element (FE) simulated data (180 cases) along with experimental data (48 cases) which is divided into Training, Testing, and Validation datasets. Each of the training and validation datasets contains FE simulated data with the major part used for training, whereas, experimental data are used for testing. Here, FE simulation responses are only used as a surrogate of experimental results. FE model is not used in the damage detection process rendering the technique to be purely data-driven. Finite element simulations of Lamb wave response are obtained using commercial package (ABAQUS) for different frequencies (100 KHz, 125 KHz, 150 KHz). The experimental data was generated for a 1.6 mm thick aluminium plate using piezoelectric wafer transducers. The 1D CNN architecture gave around 100% accuracy. Using the significant features affecting classification which is obtained from LIME, damage localization was achieved with around 7% deviation from actual damage localization.
Keywords:
Structural health monitoring
Deep learning
Lamb wave
1D CNN
Explainable AI
LIME
Journal
IF:
8.9
Papers:
1.3W
Citations:
6.6W