利用BP神经网络预测BT20钛合金的流动应力

Prediction of flow stress of BT20 titanium alloy by BP neural network

为建立BT20钛合金(Ti6Al2Zr1MolV)的流动应力预测模型,通过热压缩试验获得其流动应力曲线,并对BP神经网络的算法进行改进,实现BT20钛合金的流动应力的准确预测。研究表明:采用人工神经网络(ANNs)预测流动应力不需要考虑材料特性,有效地避免了传统经验或回归本构模型由于假设和简化带来的误差,并且神经网络具有很强的非线性离散数据处理能力,选取合适的网络模型(主要是隐层数及隐层单元数),输入足够的样本数据对神经网络进行训练即可获得令人满意的预测精度;采用含两个中间层,网络结构为3×16×14×1的改进BP神经网络模型能够较为准确的预测BT20钛合金的流动应力,计算效率较高,该预测模型可作为其塑性成形过程有限元模拟的本构关系。

To establish the prediction model of flow stress of BT20 titanium alloy （Ti6Al2Zr1Mo1V）, the flow stress curves have been obtained by hot compression experiments and the algorithm of BP neural network has been improved, based on which the flow stress of BT20 alloy has been precisely predicted. The investigation results show that artificial neural networks （ANNs） can predict flow stress without considering material characteristic, which effectively avoids the error caused by some hypotheses and simplifications of conventional empirical or regressive constitutive models. Because ANNs have strong ability in treating nonlinear discrete data, satisfactory prediction precision can be reached on condition that proper network models （mainly including the number of hide layers and neurons）are selected and enough sample data are input to train neural networks. The improved BP neural networks with the network structure of 3×16×14×1 in this paper can predict flow stress of BT20 alloy quite accurately with high calculation efficiency and the prediction model may be used as constitutive relationship for FEM simulation of BT20 alloy during plastic forming.