Predictions of Component Remaining Useful Life Using Bayesian LSTM for Reactor Coolant Pump

Abstract With the increase of service life and coupling between components, reactor coolant pumps are faced with bearing damage, mechanical wear, shaft seal leakage and other performance degradation and safety problems. In order to predict the remaining useful life (RUL) of typical components of reactor coolant pump and meet the application requirements of predictive maintenance, a RUL prediction method for reactor coolant pump is proposed by combining Bayesian theory with data-driven method. In this work, we use Bayesian LSTM to predict RUL and its uncertainty, After clarifying the interface relationship of the prediction model, the model introduces arbitrary uncertainty and cognitive uncertainty into the deep learning network, and solves the problem of quantifying the uncertainty of the prediction results of data-driven method by constructing the RUL prediction model based on Bayesian LSTM. The research shows that compared with the neural network model with a single point estimate, the Bayesian LSTM can not only give the point estimate of the predicted result, but also provide the confidence interval of the RUL predicted result according to the data characteristics, thus providing more accurate technical support for the predictive maintenance of subsequent equipment. This has important prospect value for improving the status detection capability and intelligent operation and maintenance level of nuclear power equipment, and finally realizing advanced goals such as “no monitoring, few people on duty”.