FFformer: A Transformer Architecture for Structural Natural Frequency Estimation from Video Data

To address frequency feature extraction in vibrating structures, this study developed a computer vision-based Transformer architecture incorporating temporal feature learning, designed to acquire modal information from structural vibration sequences and precisely extract frequency-domain characteristics. The proposed framework initially employed optical flow method to capture vibration displacement sequences from video recordings. Subsequently, the frequency-enhanced Transformer processed these temporal signals through seasonal-trend decomposition, followed by spectral analysis of the decomposed vibration components to identify natural frequencies. The architecture integrates specialized frequency feature extraction modules with an attention mechanism that prioritizes spectral characteristics in vibration signals. Operating directly on raw vibration videos as input, the system successfully predicted modal frequencies with high precision. Furthermore, defect simulation experiments demonstrated the learned representations' strong generalization capability across novel vibrating structures. Experimental validation revealed superior performance in both prediction accuracy (95%) and robustness, significantly outperforming conventional frequency analysis methods.