Bayesian wavelet PCA methodology for turbomachinery damage diagnosis under uncertainty
发表时间：2021-09-13
点击次数：
- 论文类型：
- 期刊论文
- 第一作者：
- 徐胜利
- 通讯作者：
- Jiang, Xiaomo,Huang, Jinzhi,Yang, Shuhua,王晓放
- 发表时间：
- 2021-01-30
- 发表刊物：
- MECHANICAL SYSTEMS AND SIGNAL PROCESSING
- 文献类型：
- J
- 卷号：
- 80
- 页面范围：
- 1-18
- ISSN号：
- 0888-3270
- 关键字：
- Wavelet packet transform; Bayesian hypothesis testing; Probabilistic PCA; Damage diagnostics; Turbomachinery
- 摘要：
- Centrifugal compressor often suffers various defects such as impeller cracking, resulting in forced outage of the total plant. Damage diagnostics and condition monitoring of such a turbomachinery system has become an increasingly important and powerful tool to prevent potential failure in components and reduce unplanned forced outage and further maintenance costs, while improving reliability, availability and maintainability of a turbomachinery system. This paper presents a probabilistic signal processing methodology for damage diagnostics using multiple time history data collected from different locations of a turbomachine, considering data uncertainty and multivariate correlation. The proposed methodology is based on the integration of three advanced state-of-the-art data mining techniques: discrete wavelet packet transform, Bayesian hypothesis testing, and probabilistic principal component analysis. The multiresolution wavelet analysis approach is employed to decompose a time series signal into different levels of wavelet coefficients. These coefficients represent multiple time-frequency resolutions of a signal. Bayesian hypothesis testing is then applied to each level of wavelet coefficient to remove possible imperfections. The ratio of posterior odds Bayesian approach provides a direct means to assess whether there is imperfection in the decomposed coefficients, thus avoiding over-denoising. Power spectral density estimated by the Welch method is utilized to evaluate the effectiveness of Bayesian wavelet cleansing method. Furthermore, the probabilistic principal component analysis approach is developed to reduce dimensionality of multiple time series and to address multivariate correlation and data uncertainty for damage diagnostics. The proposed methodology and generalized framework is demonstrated with a set of sensor data collected from a real-world centrifugal compressor with impeller cracks, through both time series and contour analyses of vibration signal and principal components. (C) 2016 Elsevier Ltd. All rights reserved.
- 是否译文：
- 否
