Acoustic monitoring of rotating machine by advanced signal processing technology

KANEMOTO Shigeru¹

1. School of Computer Science & Eng. The University of Aizu, Tsuruga, Ikki-machi, Aizuwakamatsu city, Fukushima, 965-8580, Japan (kanemoto@u-aizu.ac.jp)

Abstract: The acoustic data remotely measured by hand held type microphones are investigated for monitoring and diagnosing the rotational machine integrity in nuclear power plants. The plant operator’s patrol monitoring is one of the important activities for condition monitoring. However, remotely measured sound has some difficulties to be considered for precise diagnosis or quantitative judgment of rotating machine anomaly, since the measurement sensitivity is different in each measurement, and also, the sensitivity deteriorates in comparison with an attached type sensor. Hence, in the present study, several advanced signal processing methods are examined and compared in order to find optimum anomaly monitoring technology from the viewpoints of both sensitivity and robustness of performance. The dimension of pre-processed signal feature patterns are reduced into two-dimensional space for the visualization by using the standard principal component analysis (PCA) or the kernel based PCA. Then, the normal state is classified by using probabilistic neural network (PNN) or support vector data description (SVDD). By using the mockup test facility of rotating machine, it is shown that the appropriate combination of the above algorithms gives sensitive and robust anomaly monitoring performance.

Keyword: Acoustic Monitoring, PCA, Kernel-based PCA, PNN, SVDD, Cepstrum