The work describes rolling bearings operation data processing, and their use in the problem of constructing a mathematical model of the binary classification of the operating state of bearings by the method of a convolutional neural network with varying factors of dilatation of the kernel of convolutional layers. To classify bearings with defects, we used vibration acceleration data from our own test bench and a publicly available data set. The work also investigated a method for generalizing the classification of bearing signals obtained as a result of fundamentally different experiments and having different standard sizes. To unify signals, the following processing method is proposed: select data areas with displacement, go to the frequency space using fast Fourier transform, cut off frequencies exceeding 10 times the shaft rotation frequency, restore the signal while maintaining 10 shaft rotation periods, scale the received signal by dividing it by its diameter orbits of the rolling body and interpolate the signal at 2048 points. This algorithm also allows to generate a balanced sample for building a mathematical model. This feature is provided by varying the step of splitting the initial signal. The advantage of this algorithm over the classical methods of oversampling or undersampling is the generation of new objects that specify the statistical parameters of the general population. The signal processing algorithm was used both for binary classification problems within one dataset, and for training on one and testing on another. To increase the data set for training and testing the mathematical model, the bootstrapping method is used, based on multiple generation of samples using the Monte Carlo method. The quality of the mathematical model of binary classification was assessed by the proportion of correct answers. The problem is formulated as the problem of minimizing binary cross entropy. The results obtained are presented in the form of graphs demonstrating the neural network training process and graphs of the distribution density of metrics.

An experimental research facility has been developed to receive vibration signals from mechanisms with installed rolling bearings. A control block for all equipment has been created. For the repeatability of the experiment, an external microcontroller with a programmed proportional-integral-derivative regulator was used. Experiments were carried out to obtain initial data for different types of bearings. The processed data were grouped and made publicly available for further research. It is proposed to solve the problem of emergency stop of the generator, arising during operation due to bearings worn, by recognizing the pre-emergency conditions of rotary rig based on the use of advanced machine learning techniques: to highlight the signs of vibration and build clusters according to the degree of worn.

No abstract

The paper investigates the relationship between vibration acceleration of bearings with their operational state. To determine these dependencies, a testbench was built and 112 experiments were carried out with different bearings: 100 bearings that developed an internal defect during operation and 12bearings without a defect. From the obtained records, a dataset was formed, which was used to build classifiers. Dataset is freely available. A methodfor classifying new and used bearings was proposed, which consists in searching for dependencies and regularities of the signal using descriptive functions: statistical, entropy, fractal dimensions and others. In addition to processing the signal itself, the frequency domain of the bearing operationsignal was also used to complement the feature space. The paper considered the possibility of generalizing the classification for its application on thosesignals that were not obtained in the course of laboratory experiments. An extraneous dataset was found in the public domain. This dataset was used todetermine how accurate a classifier was when it was trained and tested on significantly different signals. Training and validation were carried out usingthe bootstrapping method to eradicate the effect of randomness, given the small amount of training data available. To estimate the quality of theclassifiers, the F1-measure was used as the main metric due to the imbalance of the data sets. The following supervised machine learning methodswere chosen as classifier models: logistic regression, support vector machine, random forest, and K nearest neighbors. The results are presented in theform of plots of density distribution and diagrams.