Application of Machine Learning Methods to Predict and Analyze the Condition of Rolling Bearings

Abstract Rising demands for industrial machinery quality and reliability lead to automating the analysis and prediction of failures, shifting from emergency repairs to planned maintenance. The article proposes methods for the analysis of possible rolling bearing failures. Freely accessible damage-type data sources adapted for autoregressive modeling were used. Classification techniques such as logistic regression, XGBoost, and artificial neural networks are compared. The analysis is performed in the time and frequency domains, and an attempt is made to apply the Cepstrum analysis in the field of mechanical vibrations. Some anomaly detection methods are applied, including ARIMA and autoencoder. The results are consolidated in a purpose-built web-based application. In this way, a comprehensive diagnostic system is proposed that uses modern machine learning techniques and methods for detecting and classifying anomalies in rolling bearing failures. The final goal of the proposed research is to create an integrated system with high speed and increased reliability for decision-making, thereby prolonging the life of mechatronic systems and improving the production indicators, which would have a strong economic effect.