Predicting the Temper Embrittlement of Steam Turbine Rotor Steels with Bayesian Neural Network

Abstract

Bayesian neural network was proposed to predict the temper embrittlement of steam turbine rotor in service. The FATT50 (the fracture appearance transition temperature) of the rotors was predicted as a function of ratio of the two peak current densities (I p/Ipr ) tested by electrochemical potentiodynamic reaction method, temperature of electrolyte, J-factor and grain size (N). A database was obtained from the test of electrochemical potentiodynamic reaction and Charpy impact. The Bayesian neural network technique was used for modeling of temper embrittlement. The neural network shows a more precise prediction of temper embrittlement of rotor steels than the prediction using multiple linear regression. The training error and verifying error is with the scatter of plusmn20 degC. The results show that, for the temper embrittlement of rotor steels prediction, the prediction model based on Bayesian neural network is feasible and effective