CAJ | 학술논문

针对一般BP神经网络泛化能力差,在Bayesian正则化BP神经网络的基础上,运用加权检验、"表决网"等方法的思路训练网络,并通过主成分分析方法对输入数据进行降维,建立了磁粉制备工艺(淬速度和晶化退火温度)、合金成分与磁性能之间的BPNN(back propagation network)预测模型.结果表明:该模型泛化能力较高,预测的Br相对误差在2%左右、Hcj和(BH)max都在5%以内,且每次预测的相对误差平均值波动不超过1%.
침대일반BP신경망락범화능력차,재Bayesian정칙화BP신경망락적기출상,운용가권검험、"표결망"등방법적사로훈련망락,병통과주성분분석방법대수입수거진행강유,건립료자분제비공예(쉬속도화정화퇴화온도)、합금성분여자성능지간적BPNN(back propagation network)예측모형.결과표명:해모형범화능력교고,예측적Br상대오차재2%좌우、Hcj화(BH)max도재5%이내,차매차예측적상대오차평균치파동불초과1%.
The (Nd2Fe14B/α-Fe) permanent magnetic property prediction model was bulit by taking magnetic particle preparation processes(spinning speed and annealing temperature) and alloy components as network input, the magnetic properties as output. For enhancing the model's ability of generalization it was trained by the way of weighted detecting method and clustering multiple based on the Bayesian-regularization BP neural network. The input data was analyzed the principal components for reducing its dimension.The results show that this model's generalization is better. The relative error between the measured value and predicted value of Br is confined to about 2% and that of Hcj、(BH)max to 5%. And the average of the relative error fluctuates within 1% in every prediction.