基于均匀设计和主成分分析的粘结NdFeB永磁体制备工艺优化研究  

Preparation Crafts Optimizing of Bonded NdFeB Permanent Magnet Based on Uniform Design and Principal Component Analysis

作　　者：储林华[1] 查五生[1] 刘锦云[1] 王向中[1] 周晓庆[1] 刘桂明[1] 

机构地区：[1]西华大学材料科学与工程学院,四川成都610039

出　　处：《稀有金属》2009年第2期191-195,共5页Chinese Journal of Rare Metals

基　　金：四川省教育厅重点项目(2004A110)资助

摘　　要：基于粘结NdFeB永磁体制备工艺优化实验,针对普通反向传播神经网络(BPNN)方法在预报建模中普遍存在"过拟合"和泛化能力差的问题,从优化实验方案、减少输入层节点数两个角度,结合均匀设计软件和主成分分析方法,提高训练样本的分布均匀性、"主动"改善网络结构,建立了一个粘结NdFeB永磁体制备工艺优化的2-5-3型BPNN预测模型。研究结果表明,改进的BP神经网络模型对Br,Hc j及(BH)m预测的相对误差的最大值分别为1.83%,1.28%和1.53%,较之传统的模型,泛化能力显著提高,网络预测也比较稳定,具有很好的实用性。Based on the optimized preparation process of bonded NdFeB magnets, a 2-5-3 BPNN prediction model was built for the bonded NdFeB magnets , which aimed at the problem of the overfitting and generalization error of the general BP neural network method for model-building and forecasting, combined uniform design to improve the uniformity of training sample with principal component analysis to reduce nodes of input layer so as to improve the network structure. The results showed that the maximum relative errors of the improved BPNN model for Br, Hcj and （BH）m were 1.83%, 1.28% and 1.53% respectively, compared with the conventional model, the generalization ability was improved greatly and the forecast results were more reliable.

关 键 词：均匀设计 主成分分析 BP神经网络 粘结NdFeB永磁体 泛化 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]