基于锆基合金退火参数与硬度的BP神经网络模型

BP ANN Model for Annealing Parameter and Hardness of Zirconium Base Alloy

摘要:

基于Trainrp算法建立锆基合金HANA-4(Zr-1.5Nb-0.4Sn-0.2Fe-0.1Cr）和HANA-6(Zr-1.1Nb-0.05Cu）退火参数与硬度的BP神经网络预测模型。模型输入单元为合金成分、退火温度和退火时间，输出单元为硬度。神经网络为3-7-1结构，动量因子和学习速率均为0.6。以实验结果验证网络的可靠性。预测结果表明，相对误差为7%，相对拟合率R值为0.98534。该模型可为锆基合金退火参数的制定提供参考。网络敏感性分析表明：退火温度和退火时间对网络的精度影响很大，而合金成分则影响很小

Abstract:

Back－Propagation artificial neural network (BP ANN) module with Trainrp algorithm was used to predict annealing parameter and hardness of HANA-4(Zr-1.5Nb-0.4Sn-0.2Fe-0.1Cr)and HANA-6(Zr-1.1Nb-0.05Cu). In this module, chemical compositions, annealing temperature and annealing time were employed as input units while hardness was employed as output unit. The optimal network architecture was considered to be 3-7-1 with momentum factor and learning rate all chosen to be 0.6, respectively. The result of ANN module shows that the biggest relative error (RE) is 7%, and the correlation coefficients (R-value) is 0.98534. The result strongly indicates that the ANN model is an efficient tool to provide annealing parameter. Moreover, Sensitivity analysis (SA) shows that annealing temperature and annealing time are the most important parameters to the ANN model accuracy, the effect of chemical compositions is very small