Data Augmentation for Deep-Learning-Based Multiclass Structural Damage Detection Using Limited Information | JoVE Visualize

Area of Science:

Civil Engineering
Computer Science
Artificial Intelligence

Background:

Global infrastructure deterioration necessitates advanced Structural Health Monitoring (SHM) methods.
Traditional manual inspections for SHM are inefficient, subjective, and data-limited.
Deep Learning (DL) shows promise for automated SHM, but suffers from data scarcity.

Purpose of the Study:

To investigate the performance of DL-based multiclass damage identification using synthetic data generated by Generative Adversarial Networks (GANs).
To evaluate a convolutional neural network (CNN) architecture for concrete surface damage detection using GAN-generated synthetic images.
To quantify the correlation between classification accuracy and the quantity/diversity of synthetic data for data augmentation in SHM.

Main Methods:

Utilized Generative Adversarial Networks (GANs) to generate synthetic images of concrete surface damage.
Trained a convolutional neural network (CNN) architecture on datasets comprising real images, synthetic images, and hybrid combinations.
Quantified classification performance metrics (accuracy, validation, testing) for various training configurations.

Main Results:

CNN performance on hybrid datasets (real + synthetic) decreased by 10.6% (validation) and 7.4% (testing) compared to models trained solely on real data.
Models trained with both real and synthetic samples showed an average performance decrease of 1.6% compared to models trained only on real samples.
Increased synthetic data quantity led to decreased sample diversity and increased correlation between synthetic data and classification accuracy.

Conclusions:

While GANs can augment data for SHM, their use in multiclass damage identification with CNNs results in a slight decrease in classification performance.
The study highlights the trade-off between data quantity and diversity when using GANs for SHM data augmentation.
Further research is needed to optimize GAN architectures and training strategies for effective synthetic data generation in SHM applications.