基于机器学习的单孔绝热气膜冷却效率分布预测研究

Prediction of Adiabatic Film Cooling Efficiency Distribution of Single Hole Based on Machine Learning

关键词:  机器学习  气膜冷却  绝热  冷却效率  神经网络  分布预测

Key words:  Machine learning  Film cooling  Adiabatic  Cooling efficiency  Neural network  Distribution prediction

摘要:

为了研究利用机器学习技术对绝热冷却效率进行直接预测的方法及特点，搭建了基于上采样卷积神经网络的机器学习模型，生成了用于训练以及验证的数值模拟数据集，使用了监督式学习的方法对模型进行了训练。训练使用反向传播算法和基于随机梯度下降的Adam优化算法，输入模型的参数包括吹风比、主流湍流强度、喷射角、孔形状、孔尺寸，模型输出为绝热冷却效率云图。模型的预测结果显示，基于上采样卷积神经网络的模型在回归预测问题上性能表现良好（测试集像素点绝对误差在0.05左右），同时，给出了此类卷积网络的训练意见。研究表明，对于云图式的回归预测目标，卷积神经网络预测结果可靠，灵活性高，有良好的工程应用价值。

Abstract:

In order to study the method and characteristics of applying machine learning to predict the adiabatic film cooling efficiency of single film cooling hole， a machine learning model based on up-sampling convolutional neural network （CNN） was built. The numerical simulation data used for training and verification was generated. The model was trained by supervised learning method. The training uses back propagation algorithm and Adam optimizer based on stochastic gradient descent. The blowing ratio， mainstream turbulence intensity， injection angle， hole shape， and size of the hole were set as inputs for the model. The contour of adiabatic film cooling efficiency was considered as output for the model. The prediction results of the model show that the model based on the up-sampling convolutional neural network performs well on the prediction problem （absolute error of pixels is about 0.05 in test set）. Besides， training suggestions for networks of this kind were given. The study show that for contour regression objective， it is flexible for CNN to generate reliable predictions， hence this method has better engineering application value.