Study on damage identification of slender beams based on SE-CNN and frequency domain grayscale image

To address high costs in structural damage detection and shortcomings of traditional deep learning models (low accuracy, weak noise robustness, poor reliability), this paper proposes a structural damage identification method combining SE-CNN (with SE attention mechanism) and frequency domain grayscale image (FDGSI). A slender beam finite element model was first established to collect vibration signals with damage information. These signals were numerically encoded into grayscale images (GSI) via 2D-DFT for frequency domain conversion. Then SE-CNN and test models were built to conduct identification studies on datasets with different models, analyzing noise and heterogeneous data effects. The research results indicate that the model proposed in this paper achieves high recognition accuracy in raw data identification; demonstrates good noise resistance in noisy data identification; and exhibits excellent generalization performance in heterogeneous data identification, proving that the proposed method possesses certain generalization capability, reliability, and noise-robustness.