摘要
为提高电化学动电位再活化法(EPR)检测汽轮机转子钢(30Cr2MoV)热脆性的检测精度,利用Bayesian神经网络建立了预测模型。根据EPR法测定的60组不同苦味酸电解液温度下,30Cr2MoV转子钢的活化峰电流密度与再活化峰电流密度比(Ia/Ir)的数据、电解液温度、转子钢化学成分J参数和晶粒度参数(N),采用Bayesian正则化训练的神经网络,建立了转子钢脆性转变温度(FATT50)与电化学特征值、电解液温度、转子钢化学成分J参数和晶粒度参数(N)之间的映射模型。利用训练好的网络预测了新的转子钢材料的脆性转变温度。结果表明:网络的训练误差和检验误差都在±20℃范围内,小于多元线性回归法得到的误差。因此,Bayesian神经网络能较准确地用来预测转子钢材料的脆性转变温度。
Making use of Bayesian Neural network, a prediction model has been established for improving the accuracy of
checking temper embrittlement of steam turbine rotor steel （30Cr2MoV） by the electrochemical potentiodynamic reaction （EPR） method. A model has been established that reflects the relationship between fracture appearance transition temperature （ FATF50 ） of rotor steel and electrochemical eigenvalues, temperature of the electrolyte, the steel＇s chemical ingredients, J parameter and crystal granularity N. The model＇s regularly trained Bayesian Neural Network was constructed by making use of data concerning the ratio of activating to reactivating peak current densities （ Io/Ir ） of rotor steel （30Cr2MoV）, temperature of the electrolyte, the rotor＇s steel chemical ingredient, J parameter and the crystal＇s granularity parameter N, obtained by the EPR method under sixty different temperature conditions of the picric electrolyte. The trained neural network was then used to predict the fracture appearance transition temperature of some new rotor steel material. Results showed that training errors of the neural network and verifying test errors were all within a scatter band of ± 20℃, which is smaller than that obtainable by the multiple linear regression method. Thus the fracture appearance transition temperature of rotor steel can be predicted more accurately by means of Bayesian neural networks. Figs 3, tables 3 and refs 13.
出处
《动力工程》
EI
CSCD
北大核心
2005年第6期860-864,共5页
Power Engineering
基金
国家电力公司科技项目(SP11-2001-02-29)
