A Pixel Array Framework of Vision-based Remote Vibration Measurements for Modal Analysis and Condition Monitoring

  • Miaoshuo Li

Student thesis: Doctoral Thesis


Vision-based Remote Vibration Measurement (VRVM) has been actively researched in recent years, because its non-contact and full-field property can fulfil non-intrusive monitoring of large infrastructures and measuring vibration of structure that is difficult to install sensors. As data collected from camera are images rather than vibration response, there is a necessary process of extracting vibration signal from video footages, however, existing VRVM methods provide low accuracy in acquiring vibration responses, which is due to its inherent weakness.

Targeting at addressing this key limit, this PhD study dedicates to develop a novel framework based on the principle of array signal processing (ASP). In recent decades, ASP has advanced rapidly on extensive applications including radar, sonar, seismology, acoustic and so forth. In this study, a pixel array signal processing (PASP) framework is established by analogy between the pixel array (selected pixels of image) and other advanced arrays in relevant area. The proposed framework consists of three phases: The first is to constitute a pixel array by selecting characteristic pixels from reference frame of video. The second is to implement array signal processing paradigm to obtain vibration responses with high signalto-noise ratio (SNR). Finally, it is to extract the diagnostic features or modal parameters from the filtered signals. Under this framework, a group of methods are developed to meet requirements of different scenarios, and thus lead to following key findings and novel contributions:

It has been derived in theory that the maximum gain of Signal-to-noise ratio (SNR) is a determined value for a given video dataset, which is equal to the number of array elements (characteristic pixels). It suggests that higher number of the characteristic pixels should be used for achieving higher SNR vibration responses.

Subsequently, a mode-shape-based adaptive spatial filtering (MASF) approach, a method based on singular value decomposition value (SVD) and a data independent spatial filtering (DISF) approach are proposed respectively. In particular, the proposed MASF was mathematically proved to be a least mean square filter (LMSF). Moreover, the gain of MASF was verified through simulation on synthetic video footages, in which MASF enhanced SNR by times of pixel number (about 2000) to identify weak mode from noisy images.

Finally, extensive experiments were carried out for fulfilling two common scenarios, including: (1) an experimental modal analysis for a free-free beam with high frequency (up to 6,389Hz) modes but small responses (at micrometre lever), along with (2) the condition monitoring of a multistage gearbox with mesh frequency (up to 1,145Hz); and (3) a freefixed wind turbine blade with low frequency (up to 250Hz) modes but high responses (over 0.2 meter), along with (4) the condition monitoring of a reciprocating compressor with rotating frequency (about 15Hz). These results validate the outperformance of proposed methods in accuracy, efficiency, and robustness, which is highly conducive to development
of the vision-based modal analysis and condition monitoring.
Date of Award23 May 2022
Original languageEnglish
SupervisorFengshou Gu (Main Supervisor) & Andrew Ball (Co-Supervisor)

Cite this