Unsupervised clustering of vibration signals for identifying anomalous conditions in a nuclear turbine

Abstract

We consider a real industrial case concerning 148 shut-down multidimensional transients of a nuclear power plant NPP turbine. The objective is to identify groups of transients with similar functional behaviors, and distinguish transients with peculiar behaviors which can be representative of anomalous conditions in the turbine. This objective is pursued by analyzing 7 vibration signals referred to the turbine shaft. The novelty of the work consists in transforming the signals into the “turbine speed-domain” for aligning them according to the turbine speed, so as to easily recognize outlier transients and then performing a fuzzy similarity analysis based on pointwise differences. Spectral analysis and Fuzzy C-Means FMC clustering are applied to identify the turbine anomalous conditions.