A UNet++-Based Approach for Delamination Imaging in CFRP Laminates Using Full Wavefield.

Area of Science:

• Materials Science
• Structural Health Monitoring
• Non-Destructive Testing

Background:

• Timely delamination detection is critical for carbon fiber-reinforced polymer (CFRP) integrity and longevity.
• Full wavefield data in CFRP contains rich information for damage mapping but is challenging to interpret due to guided wave complexities.
• Existing methods struggle with the multimodal and dispersive nature of guided waves, hindering accurate delamination imaging.

Purpose of the Study:

• To develop an end-to-end deep learning approach for accurate delamination imaging in CFRP structures.
• To leverage 2D frequency domain spectra (FDS) derived from full wavefield data for enhanced damage detection.
• To validate the proposed method on simulated, experimental, and public datasets.

Main Methods:

• Implementation of a UNet++ deep learning architecture for image segmentation.
• Utilizing 2D frequency domain spectra (FDS) extracted from full wavefield data as input.
• Validation using a self-constructed simulation dataset, Scanning Laser Doppler Vibrometry experimental data, and a public dataset.

Main Results:

• The UNet++ model accurately predicts delamination location, shape, and size using multi-frequency FDS.
• The approach demonstrates robustness against frequency offsets and noise interference in FDS.
• The model trained on simulated data shows direct applicability to real-world experimental data, producing artifact-free images.

Conclusions:

• The proposed UNet++ based approach provides an effective end-to-end solution for delamination imaging in CFRP.
• The method successfully addresses the challenges posed by multimodal and dispersive guided waves.
• The model's ability to generalize from simulated to experimental data highlights its practical potential for structural health monitoring.