人工神经网络预测镁铝水滑石纳米晶的差示扫描量热法的吸热值

PREDICTION OF DIFFERENTIAL SCANNING CALORIMETRY ENDOTHERMIC VALUE OF Mg, AIHYDROTALCITE NANOCRYSTALLINE USING ARTIFICIAL NEURAL NETWORK

常压下采用液相法制备了高纯度的镁铝水滑石纳米晶体试样。基于镁铝水滑石的差示扫描量热法(differentialscanningcalorimetry,DSC)吸热值和水滑石纳米晶体的制备工艺因素(包括:反应时间,反应温度,以及原料NaOH,MgCl2,Na2CO3,和NaAlO2的添加浓度)之间的非线性关系,建立三层结构的反向误差传播(back-propagation,BP)网络。为克服由于BP神经网络隐含层的存在,增加训练的工作量,以及当输入的参变量较多时,导致误差曲面过于复杂,出现数据冗余的缺点,尝试了采用主成分分析,对输入变景首先进行预处理。同时,在迭代公式中附加动量项和变换收敛步长,以加快网络训练速度,提高模型的预测能力与相对准确性。计算结果表明:使用改进的BP神经网络模型,对镁铝水滑石纳米晶体的DSC吸热值有较好的预测能力。

Mg, Al-hydrotalcite nanocrystalline samples of high purity were synthesized by one-step liquid reaction method at atmospheric pressure. A three layer structure back-propagation （BP） network model based on the non-linear relationship between the differential scanning calorimetry （DSC） endothermic value of the Mg, Al-hydrotalcite and the technological factors, such as reaction time, reaction temperature, and amount of raw material NaOH, MgCl2, Na2CO3, and NaAlO2 added was established, Moreover, in order to accelerate the convergence rate and avoid non-convergence, the momentum terms were introduced. Also, the variable learning speed was adopted and main component analysis was first applied to the input variables. The results show that the improved BP neural networks model is very efficient for the prediction of the Mg, Al-hydrotalcite DSC endothermic value.