Scientists from South Ural State University have developed a new method for diagnosing rolling bearing defects, which allows to significantly increase the accuracy and speed of detecting of faults. Deputy Head of the Research Laboratory of Technical Self-Diagnostics and Self-Control of Devices and Systems Vladimir Sinitsyn and his young colleagues Olga Ibryaeva, Viktoriia Eremeeva, and Mohammad Mohammad from the SUSU School of Electronic Engineering and Computer Science have published this result in the Algorithms highly rated international journal early in 2025.
Rolling bearings are used in many different areas, from aviation to CNC machines. That is why a contactless method of their diagnostics is so important – not only to detect existing defects, but also to prevent and warn of breakdowns.
With the advent of neural networks and pattern recognition technologies, these methods have also begun to be applied for bearings monitoring. Acoustic signals of bearing vibration undergo primary processing, the frequency spectrum is studied, features are identified, and then the neural network classifies the features and draws a conclusion about the bearing condition.
In order to improve the accuracy of this diagnostic method, deep learning methods of complex neural networks are usually used, which requires computational and time costs, as well as a large amount of data for training.
Vladimir Sinitsyn and his colleagues from SUSU proposed to process the signal using the LPC (linear predictive coding) algorithm before feeding it to the neural network. This method is known and is mainly used for recognizing human speech and synthesizing it on a computer.
The LPC algorithm is easy to implement. As a result, the data transmitted to the neural network (feature vector) is not very large – only 50 values. A simple neural network, without deep learning, is sufficient for its processing. This significantly speeds up the signal processing.
SUSU scientists tested the new algorithm on test data sets and achieved almost 100% accuracy in detecting defects, as well as an advantage in the speed of calculations compared to the previously used hybrid model Hybrid MLP-CNN, which uses convolutional neural networks and a multilayer perceptron.