Optimizing delamination imaging via full wavefield segmentation using augmented simulated wavefield data.

Area of Science:

• Materials Science
• Structural Health Monitoring
• Artificial Intelligence

Background:

• Catastrophic structural failures in carbon-fiber reinforced polymers can result from hidden delamination.
• Accurate delamination detection is crucial for preventing such failures.
• Deep learning for full wavefield segmentation shows promise but faces challenges with data acquisition costs and complex wavefield superposition.

Purpose of the Study:

• To propose and evaluate a data preprocessing strategy to improve deep learning-based delamination imaging in carbon-fiber reinforced polymers.
• To enhance the performance and generalization of deep learning models for delamination detection.
• To enable effective delamination imaging using simulated data for real-world applications.

Main Methods:

• A data preprocessing strategy combining wavenumber filtering and hybrid noise-flipping augmentation was developed.
• Wavenumber filtering isolates guided wave modes introduced by delamination in the frequency domain.
• Noise and flipping augmentation were used to improve model generalization under varying measurement conditions.

Main Results:

• The proposed preprocessing strategy significantly enhanced deep learning model performance for delamination imaging.
• Models trained on simulated data demonstrated effective application to experimental measurements.
• A highest intersection over union score of 0.8634 was achieved, producing artifact-free delamination images.

Conclusions:

• The data preprocessing strategy effectively guides deep learning models to focus on delamination-relevant features.
• The approach improves model generalization, allowing successful application to experimental data.
• This method offers a cost-effective solution for accurate delamination imaging in structural health monitoring.