The local maxima method for enhancement of time–frequency map and its application to local damage detection in rotating machines

Abstract

In this paper a new method of fault detection in rotating machinery is presented. It is based on a vibration time series analysis in time–frequency domain. A raw vibration signal is decomposed via the short-time Fourier transform (STFT). The time–frequency map is considered as matrix (M×N) with N sub-signals with length M. Each sub-signal is considered as a time series and might be interpreted as energy variation for narrow frequency bins. Each sub-signal is processed using a novel approach called the local maxima method. Basically, we search for local maxima because they should appear in the signal if local damage in bearings or gearbox exists. Finally, information for all sub-signals is combined in order to validate impulsive behavior of energy. Due to random character of the obtained time series, each maximum occurrence has to be checked for its significance. If there are time points for which the average number of local maxima for all sub-signals is significantly higher than for the other time instances, then location of these maxima is “weighted” as more important (at this time instance local maxima create for a set of Δf a pattern on the time–frequency map). This information, called vector of weights, is used for enhancement of spectrogram. When vector of weights is applied for spectrogram, non-informative energy is suppressed while informative features on spectrogram are enhanced. If the distribution of local maxima on spectrogram creates a pattern of wide-band cyclic energy growth, the machine is suspected of being damaged. For healthy condition, the vector of the average number of maxima for each time point should not have outliers, aggregation of information from all sub-signals is rather random and does not create any pattern. The method is illustrated by analysis of very noisy both real and simulated signals.