Rolling Bearing Remaining Useful Life Prediction Based on LSTM-Transformer Algorithm

Abstract

Bearings are the most critical components in modern industrial rotating machinery. If a bearing is damaged, it can lead to serious consequences such as an interruption to a production line and financial losses. It is important to monitor the bearing operation condition and to predict the remaining useful life (RUL) of bearings so that a scheduled maintenance can be planned ahead. In order to improve the accuracy of a bearing RUL prediction, a new data-driven RUL prediction technique based on Long Short-Term Memory (LSTM) network and Transformer network is proposed. Firstly, a total of 8 degradation characteristics in both time and frequency domains are extracted from the bearing data to be used as the input features. After the data preprocessing steps such as normalization and sliding window interception, the degradation characteristic dataset is obtained. Then, the proposed LSTM-Transformer technique is applied to the characteristic dataset for training and prediction. The prediction result shows that the proposed technique can effectively overcomes the information loss of LSTM network caused by the increase distance between the input and output sequences to produce a more accurate RUL prediction. The RUL prediction obtained using the proposed technique is compared with those using existing techniques such as GRU, LSTM and CNN networks for an evaluation of the effectiveness and efficiency of the proposed technique. It is confirmed that the proposed technique can yield a more accurate bearing RUL prediction than the existing techniques.