A hybrid SGMD-CNN-transformer model for noise-robust structural damage detection with time-frequency feature integration

Abstract

Structural damage detection often suffers from noise interference and complex modal behaviors. To address these issues, this study proposes a novel hybrid approach combining Symplectic Geometry Mode Decomposition (SGMD), Wavelet Synchrosqueezed Transform (WSST), Convolutional Neural Network (CNN), and Transformer architectures for robust and accurate damage detection. Experiments on a six-story steel frame structure and the IASC-ASCE Benchmark model dataset demonstrate that the complete SGMD-CNN-Transformer model achieves outstanding performance, significantly outperforming all ablated variants. Notably, the model exhibits exceptional noise immunity, maintaining consistently high detection accuracy even under severe noise conditions. This work provides strong evidence for the effectiveness and practical utility of the proposed method in structural health monitoring.