Full-Field Visual Vibration Measurement of Rotating Machine Under Complex Conditions via Unsupervised Retinex Model

Abstract

Vision-based measurement, as the most common non-contact vibration measurement method, has received increasing attention and development in recent years. However, it is still highly dependent on the shooting environment and data quality, and is difficult to apply to complex mechanical structures. This article focuses on the full-field extraction of micro-vibration signals of high-speed rotating machinery under complex conditions. A novel visual vibration measurement scheme is proposed, called the texture-perception motion estimation frame (TPME). First, an unsupervised image decomposition model called the texture-perception network (TPNet) is designed to linearly separate the image reflection layer and illumination layer. The high-frequency texture features of the reflection layer can strengthen the data representation ability. Using it alone for phase-based micro-motion extraction (PME) can not only expand the measurement range and accuracy but also ensure measurement robustness under complex luminance changes. Furthermore, we evaluate the computational cost of this approach, which achieves significant measurement improvements with a very low computational overhead. Experiments on three different rotor platforms demonstrate the effectiveness of TPME.