Remaining Useful Life Prediction Study of Rolling Bearings Based on TCN-Transformer and KCA

Abstract

With the modernization of industry, more and more attention is being paid to predicting the remaining useful life (RUL) of rolling bearings, which are key components of machinery and equipment. Since rolling bearings often operate in very complex environments, this makes prediction difficult. We propose an approach that combines the cross-attention mechanism with a kolmogorov-arnold networks layer (KCA) and fuses it with a temporal convolution network (TCN) and transformer networks to better capture degraded features of raw bearing data. To better capture the friction features of the original bearing data, this article adopts the data preprocessing method of time domain feature extraction and principal component analysis (PCA) feature screening combined with start prediction time (SPT) point division data. The present study has been validated using the XJTU-SY dataset, and a comparison has been made with the classical models long short term memory (LSTM), convolutional neural network (CNN), and gate recurrent unit (GRU). The root mean square error (RMSE) and mean absolute error (MAE) have been reduced by 58.0% and 66.2%, 71.4% and 75.8%, and 52.3% and 62.5%, respectively. The experimental evidence presented in this article demonstrates the superiority of the research method employed in predicting the RUL of rolling bearings.