Deep neural networks for crack detection inside structures

Fatahlla Moreh ¹, Hao Lyu ^2,3, Zarghaam Haider Rizvi ^1,4

⁰Geomechanics and Geotechnics, Kiel University, Kiel, 24118, Germany.

Scientific Reports +

|

February 23, 2024

View abstract on PubMed

Summary

This study enhances seismic-wave crack detection for plates using deep learning. Advanced networks like DenseNet and data normalization significantly improve the accuracy of identifying small cracks in structures.

Area Of Science

Structural Health Monitoring
Non-Destructive Testing
Applied Geophysics

Background

Conventional crack detection methods are labor-intensive and costly.
Deep neural networks offer automated solutions for structural damage assessment.
Seismic-wave-based techniques show promise for non-invasive crack detection.

Purpose Of The Study

To improve seismic-wave-based crack detection in plate structures using deep learning.
To investigate the impact of network architecture and data preprocessing on detection accuracy.
To enhance the identification of small and subtle cracks.

Main Methods

Utilized an encoder-decoder deep neural network architecture.
Experimented with various network components, including Densely Connected Convolutional Network (DenseNet) as a backbone.
Implemented a data preprocessing strategy involving reference wave field normalization.
Tested methods on an expanded dataset for crack detection in plate structures.

Main Results

DenseNet effectively extracts crack-indicative features from seismic wave signals.
Reference wave field normalization significantly improves the detection accuracy of small cracks.
The enhanced deep learning approach demonstrates superior performance in crack detection.

Conclusions

Deep neural networks, particularly with robust backbones like DenseNet, are effective for automated seismic-wave crack detection.
Data preprocessing, including normalization, is crucial for improving the sensitivity and accuracy of detecting minor structural damage.
This research advances non-destructive structural health monitoring capabilities for plate structures.

Keywords:

Crack detection Deep learning Neural network Wavefield