Application Of BP Neural Network And SVR In Investigation For Influence Of Elements On Magnetic Properties Of (Nd,Pr) FeB Permanent Magnet

Based on summarizing and analyzing the influences of alloy compositions on the (Nd,Pr)FeB permanent magnet properties, the Bayesian regularization BP neural network and support vector machines regression have been established between alloy compositions and magnetic properties, the influences of Pr, Co and Zr on magnetic properties has been predicted and researched.For enhancing the generalization of the BP neural network, it has been trained by the way of weighted detecting method and clustering multiple based on the Bayesian regularization BP neural network. The results show this model's generalization is better. The relative errors between measured value and predicted value of Br are confined to about 2% and those of Hcj, (BH)m to 5%.Support Vector Machine (SVM), compared with BP network, has a remarkable advantage of dealing with high-dimensional nonlinear regression problem of small samples. Because SVM is based on the statistical learning theory. Theε-SVR and LS-SVR have been introduced to predict the influences of alloy elements on magnetic properties in this paper. The parameters of model have been optimized with high predictive accuracy and generalization ability simultaneously. They have been confirmed by testing samples. Compared withε-SVR, the forecast precision of LS-SVR is higher. The mean absolute percent errors (MAPE) between measured value and predicted value of Br, Hcj and (BH)m are 0.58%,1.64% and 1.87%.In the small samples, the influences of interaction among elements on magnets magnetic properties have been predicted by LS-SVR, which is compared with that of modified BP network. 3-dimension figures and contour lines of content-properties have been obtained by the LS-SVR model. Analysis of those shows that the total of Nd and Pr affects greatly the magnetic properties, which is identical with the prediction of modified BP network. At the same time, the magnetic properties which have been predicted by LS-SVR exist extremum under the Co-Zr interaction, which is consistent with the theoretical analysis of the alloy elements on magnetic properties. The predicted results of LS-SVR are more reliable and clear than those of modified BP network. The adding elements range is optimized by LS-SVR, Pr is at8% ~ 10%, Co is at1.8% ~ 2.5%, Zr is at1% ~ 1.5%. According to the results, the (Nd0.2Pr0.8)10.5Fe80.5Co2Zr1B6 alloy has been obtained in this paper, which is prepared to bonded magnet of excellent properties: Br=0.662T, Hcj=616kA/m1, (BH)m=74.0kJ/m3. The predicted results agree well with experimental results. Therefore, under small samples situation, the SVR is feasible and can greatly shorten the experiment circle for the study of alloy composition on the magnetic properties. The LS-SVR is an efficient and reliable method in prediction of magnetic properties.