Aspects of Features Selection and Extraction from Time-Frequency Images of Vibration Signals

The research domain of the paper is the time-frequency image processing. Firstly, a comparison among the features selection and extraction methods from time-frequency images of vibration signals in bearings with faults is made. Both time and frequency methods are considered and discussed, from classification performance point of view. A new method of feature selection based on the decomposition of time-frequency images in sub-bands is introduced. For each sub-band, discrete cosine transform is applied. The coefficients of the transform are features which must be processed for classification task. Distance based classifier is considered with vectors of features of various lengths. Computer based experiments are conducted with real data from a benchmark database with vibration signals generated by various type and size of faults. The results are very encouraging and show the feasibility of the method.