基于迁移学习与残差网络的冰形图像预测

Ice shape image prediction based on transfer learning and residual network

神经网络冰形预测方法中,风洞试验结冰数据具有较高精度的冰形特征,但试验成本昂贵、获得数据较少,未能得到充分利用,而大多数是针对数值计算数据开展研究。为此,提出了一种结合迁移学习和残差网络的图像化预测方法,以翼型截面图像和结冰工况参数为输入,二维冰形图像为输出,建立深度神经网络预测模型,实现高精度二维冰形预测。通过大量数值计算数据获得预训练模型,再加入少量风洞试验数据进行微调,实现冰形预测。结果表明,所提出的方法可以预测较高精度的二维翼型结冰图像,大部分冰形特征参数与风洞试验冰形的相对误差保持在15%以内。

Among the ice shape prediction methods of neural networks,the ice shape data of wind tunnel experiment has high-precision ice shape features,but the experiment cost is expensive and the data obtained is limited so it has not been fully utilized,and most of the research is focused on numerical calculation data.To this end,an image prediction method combining transfer learning with residual network is proposed.A deep neural network prediction model is established by taking airfoil cross section image and icing condition parameters as input and 2D ice shape image as output,which realizes the high-precision 2D ice shape prediction.This method obtains the pre-training model through a large amount of numerical calculation data,and then fine-tunes it with a small amount of wind tunnel experiment data to achieve ice shape prediction.The results show that the proposed method can predict high-precision 2D airfoil icing images,and most of the relative error between prediction and wind tunnel experiment data in ice shape feature parameters is kept within 15%.