Prediction of remaining useful life of rolling element bearings based on LSTM and exponential model

Fault in rolling element bearings is a very common fault in mechanical systems. It may lead to abnormal operation of equipment, even to serious accidents or significant losses. Periodical monitoring of bearings plays a vital role in reducing unplanned maintenance and improving the reliability of machines. However, the existing methods for determining faults in rolling element bearings introduce too many artificial factors, and the results are often subjective. In order to solve this problem, the present paper proposes a hybrid real-time method for determining the starting time of a fault in a rolling element bearing. Based on the dynamic 3σ interval and voting mechanism, our method can adaptively predict the starting time. Firstly, the long short-term memory (LSTM) neural network is used to predict the trend of the future operation of the bearing. Then, an exponential model is used to estimate its remaining useful life (RUL). The obtained experimental results show that the proposed approach can significantly reduce artificial interference, adaptively divide the state of rolling element bearings, and accurately predict RUL.