Fault diagnosis of rolling bearings using spectral peaks and neural networks

The most important components of machine parts are rolling bearings, the condition of which is necessary to control, since possible defects in their design can lead to incorrect operation or general failure of machines. Modern solutions on fault diagnosis of bearings typically use complex feature extraction processes, such as their Hilbert spectrum imaging and a further powerful neural network to classify them. In this article, we propose a simple, but, nevertheless, an effective algorithm for solving this problem. To extract features from a signal, we divide the signal spectrum into equal subintervals and find the amplitude maximum and the corresponding frequency value in each of them. In the article, based on the t-SNE method, it is shown that the features selected in this way, despite their small size, represent different types of signals well. At the second stage, the selected features are fed to the input of a simple classifier neural network. The proposed method is computationally simple, both at the stage of feature extraction and at the stage of neural network training. Despite this, the method gives 100% accuracy for all types of signals on short data from the IMS dataset.