基于优化的GRNN和BP神经网络的磁滞曲线拟合对比分析  

Contrast analysis of hysteresis curve fitting between optimized GRNN and BP neural network

作　　者：何汉林[1] 孟爱华[1] 祝甲明[1] 宋红晓 

机构地区：[1]杭州电子科技大学机械工程学院,浙江杭州310018 [2]杭州浙大精益机电技术工程有限公司,浙江杭州310000

出　　处：《机电工程》2013年第1期116-120,共5页Journal of Mechanical & Electrical Engineering

基　　金：国家自然科学基金资助项目(50905051);浙江省自然科学基金资助项目(Y1080004);浙江省重点科技创新团队资助项目(2010R50003)

摘　　要：针对超磁致伸缩材料(GMM)的磁滞非线性,运用广义回归神经网络(GRNN)和前馈BP神经网络分别对GMM的磁滞回线进行非线性逼近,通过网络的训练、预测,与Jiles-Atherton(J-A)模型进行了对比,分析了两种神经网络的逼近效果,给GMM的运用起到了很好的指导作用。其中,在GRNN神经网络中,由于所取数据有限,为了扩大样本容量,采取交叉验证方法对GRNN神经网络进行了训练,采用循环算法找出了最佳的径向基函数扩展系数SPREAD,并对传统GRNN神经网络进行了优化。研究结果表明:优化后的GRNN神经网络对于磁滞回线的预测精度明显高于BP神经网络。Aiming at the nonlinear hysteresis curve of the giant magnetostrictive material （GMM）, the generalized regression neural network （GRNN）and feed-forward BP neural network were applied to approach it. With the training and prediction of the networks, as well as comparing with the Jiles-Atherton（J-A）model, the approaching effect of the networks was analyzed, which guides the applying of the GMM well. Between them, the GRNN was trained by cross-validation method in order to enlarge the sample capacity. The best radial basis function expansion coefficient（SPREAD）was found out using circulation, and the conventional GRNN was optimized. The results indicate that the accuracy on the hysteresis curve predicted by optimized GRNN is obviously higher than the one done by BP.

关 键 词：超磁致伸缩材料 广义回归神经网络 BP神经网络 磁滞曲线拟合 

分 类 号：TH39[机械工程—机械制造及自动化] TM1[电气工程—电工理论与新技术]