PCA-based Monitoring of Power Plant Vibration Signal by Discrete Wavelet Decomposition Features

SUMMARY AND CONCLUSIONSThis paper studies the vibration signals obtained from a reactor coolant pump (RCP) of a power plant. A discrete wavelet transform (DWT) is used to decompose the vibration signals generated within a time window into approximation (cA2) and detailed coefficients (cD2 and cD1). Additionally, twelve features consisting of some statistics are generated from each of the DWT coefficients or sub-bands, resulting in a total of 36 features per vibration signal. These features are used as inputs for the principal component analysis (PCA). From the PCA results, only 4 principal components (PCs) explain most variability (about 96%) of the features. The variability was found in features X2, X13, X14, X25, and X26 which correspond to the 75
^th
percentile of cA2, 25
^th
percentile of cD2, 75
^th
percentile of cD2, 25
^th
percentile of cD1, and 75
^th
percentile of cD1 respectively. Therefore, we can monitor these small number of 4 PCs variables instead of the original 36 PCs. These 4` PCs can provide the behavior of normal vibration signals and how to detect abnormal signals in the future.