| Abstract [eng] |
The aim of this thesis is to investigate signal classification methods suitable for classifying vibration signals of various electromechanical systems and to assess the accuracy of their classification across different datasets. The methods of vibration signal measurement, signal analysis and classification were described. Three different datasets were used in this thesis. One of them consisted of vibration signals of bearings with different fault types recorded in the CWRU bearing database. The MaFaulDa machinery fault simulator database, containing different fault cases of an electric motor system was also used. Another dataset used in the research was obtained in the laboratory by measuring vibration signals of a motor with varying load conditions. For these data sets, time domain, spectrum, cepstrum and autocorrelation function plots of the signals were generated. Based on selected informative features, a classifier was developed, operation of which was based on calculating Euclidean distances between points in the feature space and the centers of their classes. The classification of all data sets using CNN, RF, SVM and KNN methods was also tested. In the CNN method, spectrograms generated from vibration signals were used for classification, while feature vectors extracted from spectrograms were used for the RF, SVM, and KNN methods. The classification accuracy of these methods was compared, and confusion matrices were obtained in order to evaluate their suitability for the condition and state classification of electromechanical systems. |
