Improvements in Prediction of Remaining Useful Life of Roller Bearings Using Deep Learning Models

The purpose of this study is to enhance the evaluation of damage in roller bearings by utilizing a multi-way machine and deep learning algorithms. Regular maintenance and repair are necessary for smooth operation and optimal harvesting of energy. Bearings have a significant impact on the functioning of rotating machinery and are prone to frequent damage and deterioration during operations. Hence, it is crucial to establish a reliable and instantaneous approach to forecast the remaining useful life (RUL). This paper proposes a two-stage prognostic analysis using machine learning and deep learning models. The suggested model is validated using the XJTU-SY dataset. The collected vibration data from the rolling bearings undergo piecewise aggregate approximation and singular spectrum analysis. The central moment approach is used to determine the exact point at which bearing degradation occurs, based on the established relationship between the health index and the RUL. Both temporal and spectral properties are extracted to train the algorithms. Furthermore, the process is subsequently divided into two separate stages: the classification of the bearing condition using the extreme gradient boosting (XGBoost) and the prognosis assessment that employs a mixed convolution neural network and a gated recurrent unit (CNN-GRU). The proposed model demonstrated high efficiency, achieving an F1 score of over 98% when forecasting the health of the bearing using XGBoost. Furthermore, the model demonstrates outcomes similar to those of conventional methods in predicting the RUL.