引入物理约束的航空发动机燃烧室温度场预测模型

Prediction model for aero-engine combustor temperature field with physical constraints

通过入口流动参数快速获得航空发动机燃烧室截面的温度场,可以帮助研究人员快速了解发动机内燃烧室的燃烧状态,这对于航空发动机燃烧室的设计和优化较为重要。本文提出了一种基于深度学习方法的航空发动机燃烧室温度场快速预测方案,文中构建的含有注意力模块的双支路网络模型,测试集温度场相对偏差均值为0.64%,可以高精度地实现燃烧室温度场的预测目标;而在远离学习工况的区域,仅由均方差训练的注意力网络模型会面临性能退化的问题,通过引入物理损失函数项,注意力网络模型的预测性能得到显著改善,温度场相对误差均值的平均值降低了48.4%。相较于传统的卷积网络模型和全连接网络模型,引入物理损失函数项训练的注意力网络模型在学习数据区间内外都有着更好的预测表现。

Quickly obtaining the temperature field of the aero-engine combustor cross-section through the inlet flow parameters can help researchers quickly understand the combustion state of the combustor in the en⁃gine,which is more important for the design and optimization of the aero-engine combustor.This paper proposed a fast prediction scheme for the temperature field of an aero-engine combustor based on a deep learning ap⁃proach.The dual-path network model with an attention module constructed in the article achieved 0.64%average relative deviation in the temperature field of the test sets,which can achieve high-precision prediction of the com⁃bustor temperature field.However,in the districts far from the learning conditions,the attention network model trained only with mean square error faces performance degradation.By introducing the physical loss function,the predictive performance of the attention network model is significantly improved,resulting in a 48.4%reduction in the average relative error of the temperature field.Compared with the traditional convolutional network model and fully-connected network,the attention network trained with the physical loss function has better prediction per⁃formance both inside and outside the learning edge.