Deep optical flow to identify structural vibration modal parameters

Abstract

Accurate measurement of vibration signals is an important prerequisite for accurate analysis of structural vibration modes. Optical flow estimation is suitable for non-contact full-field signal measurement. However, optical flow datasets are usually synthesized by computers and lack practical application in engineering structures. In addition, the algorithm has deficiencies in processing sparse textures and edge features, and the sub-pixel estimation accuracy needs to be improved. Therefore, this paper constructs real structural vibration optical flow datasets to provide basic data support for structural health monitoring. For the deep optical flow algorithm, an adaptive Gabor filter component is proposed. By dynamically adjusting the direction and scale, the edge and texture features in the optical flow field are enhanced, and the fusion with the convolutional neural network further models the relationship between global and local features. At the same time, the local similarity matching mechanism is combined to improve the accuracy of optical flow estimation and the ability to identify local details. In addition, the smoothing term and the loss function of the learnable parameters are added to improve the physical consistency and prediction accuracy of the model, and effectively suppress the generation of noise interference and discontinuous optical flow. The results of qualitative and quantitative scene and ablation experiments show that the proposed deep optical flow algorithm is superior to the existing methods in terms of optical flow estimation accuracy, structural edge preservation ability and low light adaptability, and shows excellent performance in sparse texture scenes and low signal-to-noise ratio environments. In addition, by performing modal frequency error evaluation and energy ratio analysis on the identification results, the effectiveness and stability of the proposed method in modal main frequency identification and high-order mode extraction are demonstrated, providing a high-precision solution for non-contact modal analysis.