3 / 2024-05-08 15:00:04

Bayesian Long Short Term Memory Network for Remaining Useful Life Prediction of Rolling Bearings

LSTM,Bayesian statistics,Rolling bearing,Uncertainty quantification

The safe development and peaceful application of nuclear energy are important forces in solving the energy crisis, promoting China's energy transformation, and achieving environmental protection goals. However, with the long-term operation of nuclear power plants, equipment such as circulating water pumps are facing safety hazards caused by performance degradation. As the key driving equipment for the final cold source of the secondary circuit system in nuclear power plants, the reliability of circulating water pumps plays an important role in ensuring the continuous and stable operation of the units. During the operation of the circulating water pump, the roller bearing, as a key component of the pump body transmission core, plays a crucial role in bearing equipment loads and absorbing system vibrations. Once a fault occurs, it will directly affect the performance of the circulating water pump and pose a challenge to the safe operation of the entire nuclear power plant. However, currently, the maintenance strategy for bearings and other components in commercial nuclear power plants in China still focuses on preventive maintenance through periodic inspections, which has the drawbacks of "insufficient maintenance" and "excessive maintenance", making it difficult to meet the application needs of equipment predictive maintenance. The effective prediction of Remaining Useful Life (RUL) of circulating water pump roller bearings through the combination of artificial intelligence and information technology has important engineering value. Therefore, this article constructs multiple characteristic indicators based on historical operating data combined with the vibration characteristics during bearing operation, to achieve the characterization of the degradation process of rolling bearings throughout their entire life cycle and complete the detection of bearing health status. This provides a basic condition for the subsequent evaluation of degradation status and RUL prediction of rolling bearings. Secondly, by combining the temporal analysis ability of Long Short Term Memory (LSTM) networks, the problem of traditional neural networks not being able to fully utilize historical data is overcome. At the same time, Bayesian thinking is combined to quantify the uncertainty in neural networks and achieve more accurate predictions. Finally, based on the data collected from the scaled test bench, the predictive performance of the model was verified. The results showed that the model can accurately and effectively predict the life of rolling bearings, and quantify the uncertainty of prediction results caused by parameter uncertainty and other factors. This solves the problem of low accuracy and credibility of prediction results currently faced by RUL technology, and provides effective auxiliary guidance and decision support for operators.