SDFormer: A Novel Transformer Neural Network for Structural Damage Identification by Segmenting the Strain Field Map.

Area of Science:

• Structural engineering
• Artificial intelligence
• Materials science

Background:

• Structural health monitoring (SHM) is crucial for infrastructure safety.
• Traditional damage identification methods require expert knowledge and lack versatility.
• Existing techniques struggle with adaptability across different structural types.

Purpose of the Study:

• To introduce SDFormer, a novel deep learning network for structural strain damage identification.
• To treat damage identification as an image segmentation task using advanced AI.
• To develop a model that directly maps strain fields to damage distributions.

Main Methods:

• Proposed SDFormer (structural damage transformer), a U-shaped network utilizing self-attention mechanisms.
• Input: structural strain field maps.
• Output: predicted damage location and severity.
• Employed skip connections for feature fusion and self-attention for enhanced damage feature extraction.

Main Results:

• SDFormer demonstrated superior accuracy and efficiency in numerical experiments compared to advanced Convolutional Neural Networks (CNNs).
• The model successfully mapped strain fields to damage distributions without complex damage index design.
• Validated robustness and anti-noise capabilities through dedicated experiments, outperforming comparison models.

Conclusions:

• SDFormer offers a highly effective and efficient solution for structural damage identification.
• The self-attention-based network provides robust performance, even in noisy conditions.
• This approach advances SHM by enabling direct, data-driven damage assessment.