172 / 2021-03-31 17:39:03

Modal frequency extraction based on 3D-CNN

3D-CNN,Modal analysis,Computer vision,Deep learing

Existing modal analysis methods include contact measurement and non-contact measurement for structural vibration measurement. Contact measurement has its limitations for light and small structures, such as the additional quality impact brought by the sensor. At the same time, the installation and calibration of the sensor is also a time-consuming and laborious process. Non-contact measurement can solve the above problems. In this article, we introduce a method of modal frequency extraction based on 3D-CNN. The deep learning network is used to automatically extract feature information from the video data, finally the modal frequency of the structure is output. The key idea of this paper is to use high-speed cameras to collect vibration information and use 3D convolution kernel to extract the features of spatio-temporal information from video data. The dynamic characteristics of the system will not be changed by non-contact measurement which has the characteristics of low cost and simple operation. Compared with DIC technology, the trained deep learning network has a faster calculation speed. This article also tested the robustness of 3D-CNN. Structures of different sizes show high accuracy. Experimental results show that 3D-CNN has good robustness and accuracy. Finally, this article analyzed the process of 3D-CNN. We hope that the process of 3D-CNN is interpretable. Therefore, we visualized the features extracted by the 3D-CNN convolution kernel.