Qualitative Analysis of the Time-Frequency Images of Vibrations in Faulty Bearings

Abstract

The paper considers time-frequency transforms of vibrating signal in bearings with faults. The objective is the classification of faults based on time-frequency image processing. Direct classification of such images does not provide satisfactory results. The processing chain, from data acquisition to image processing, is complex and needs a qualitative analysis, highlighting also the importance of the right preprocessing. Two pre-processing steps are considered, at the level of time and time-frequency planes. It reveals a source separation and stationarity analysis, for the first pre-processing chain, and, for the second one, the image scaling, and registration. Content analysis of time-frequency images shows a variance of the content across the observation frame/window of the same fault. It is concluding that images with high complexity, evaluated here with Renyi entropy, must be excluded from the registration process of the images. Finally, an alternative approach, for classification of the time-frequency images, feature selection and extraction must be considered. Computer-based experiments sustain the observation and the comments during such an analysis.