Nonnegative factorization of spectrogram for local damage detection of belt conveyor gearboxes

Abstract

The problem of local damage detection for rotating machines is widely studied in the literature. In order to extract information of the damage, the approach based on the vibration signal analysis is applied very often. However, in many cases the analysis of the signal in the time domain is not sufficient. A good choice is the analysis of the signal into the time-frequency domain and extraction of the information on damage from this representation. In this paper, we propose to separate the time-frequency map (spectrogram) of the vibration signal into new partial spectrograms. We assume that one of the new time-frequency representation carries information about the damage, i.e. cyclic impulsive behavior. To perform the separation we apply the idea that is based on a very important topic in numerical algebra, which is matrix factorization. In the presented approach, Convex Non-negative Matrix Factorization (Convex NMF) is used. In the experiments, we analyze a real vibration signal from the heavy duty gearbox used in mining industry. The results obtained for real signal are compared with those obtained using the Spectral Kurtosis approach, one of the classical method used in the problem of local damage detection.