This is published under the terms of CC BY-NC.

Downloaded: 411 times

Abstract

Vibration measurements play a critical role in troubleshooting mechanical equipment and assessing the structural integrity of buildings. However, conventional vibration measurement methods rely on contact-based approaches, such as the attachment of accelerometers to the target object, leading to complex equipment deployment. Therefore, non-contact vibration measurement has attracted great attention but has yet to be fully addressed. In this paper, we propose DeepVib, a non-contact vibration measurement system that enables accurate micron-level vibration monitoring. First, we introduce a series of signal processing algorithms to extract the vibration object motion from mmWave reflection signals. Then, we design a deep neural network to effectively suppress noise interference and achieve outputting higher signal-to-noise ratio data. Finally, we eliminate static reflections with geometric-based method to recover the vibrations of the target. The experimental results show that our non-contact measurement method can accurately measure the vibration at the micron level with the average error of vibration frequency less than 0.1%. For the amplitude below 100μm, the median error of estimation is 7.23%. In addition, DeepVib reduces the estimation error of 80th-percentile amplitude by 56.60% compared with the conventional method.