Prediction of aerodynamic characteristics of compressor blade profile based on deep learning
-
摘要:
采用了数值模拟与机器学习相结合的方式对压气机双圆弧叶型流场气动参数预测开展了研究。对双圆弧叶型进行参数化批量建模,通过计算流体力学进行数值模拟,将数值模拟的模型数据与气动性能的映射提供给多层神经网络(MLP)和卷积神经网络(CNN)进行学习,分别对预测模型的准确率进行了测试比较。研究发现:通过深度学习的方式可以有效的对压气机内部流场气动参数进行准确预测,该模型预测的压力系数误差率小于0.2%,总压损失系数误差率小于1.2%,并证明CNN在气动参数预测的精度上优于传统全连接神经网络。
-
关键词:
- 机器学习 /
- 深度学习 /
- 卷积神经网络 /
- 计算流体力学(CFD) /
- 压气机叶型
Abstract:A combination of numerical simulation and machine learning was used to investigate the prediction of aerodynamic coefficients in the flow field of a double-circular-arc leaf shape of a compressor. Parametric batch modeling of the double-arc impeller shape was carried out, and numerical simulation was performed by computational fluid dynamics. The mapping of model data from numerical simulation to aerodynamic performance was provided to multilayer neural network (MLP) and convolutional neural network (CNN) for learning, and the accuracy of the prediction models was tested and compared respectively. It was found that the accurate prediction of the impeller mechanical internal flow field aerodynamic coefficients can be effectively performed by deep learning, and the error rate of the model predicted pressure coefficient was less than 0.2% and the error rate of loss coefficient was less than 1.2%, proving that CNN was better than traditional fully connected neural network in the accuracy of aerodynamic coefficient prediction.
-
表 1 ResNet-34 结构
Table 1. ResNet-34 architecture
层名 输出大小 参数结构 Conv1 143×218 7×7,64,步长 2 Conv2 72×109 3×3 最大池化,步长 2 $\left[ {\begin{array}{*{20}{l}} {3 \times 3,}&{64} \\ {3 \times 3,}&{64} \end{array}} \right] \times 3$ Conv3 36×55 $\left[ {\begin{array}{*{20}{l}} {3 \times 3,}&{128} \\ {3 \times 3,}&{128} \end{array}} \right] \times 4$ Conv4 18×28 $\left[ {\begin{array}{*{20}{l}} {3 \times 3,}&{256} \\ {3 \times 3,}&{256} \end{array}} \right] \times 6$ Conv5 9×14 $\left[ {\begin{array}{*{20}{l}} {3 \times 3,}&{512} \\ {3 \times 3,}&{512} \end{array}} \right] \times 3$ Output 1×1 平均池化, 1-D 全连接 表 2 对比测试结果
Table 2. Comparison of test results
比较方法 Cp $ \omega $ Erms Er/% Erms Er/% 方法1 0.1504 2.872 0.2475 10.45 方法2 0.2792 5.332 0.3263 13.78 ResNet-18 0.0097 0.184 0.0293 1.256 ResNet-50 0.0126 0.240 0.0313 1.343 ResNet-34 0.0091 0.173 0.278 1.192 -
[1] 张伟伟,寇家庆,刘溢浪. 智能赋能流体力学展望[J]. 航空学报,2021,42(4): 26-71.ZHANG Weiwei,KOU Jiaqing,LIU Yilang. Prospect of artificial intelligence empowering fluid mechanics[J]. Acta Aeronautica et Astronautica Sinica,2021,42(4): 26-71. (in Chinese) [2] 张伟伟,朱林阳,刘溢浪,等. 机器学习在湍流模型构建中的应用进展[J]. 空气动力学学报,2019,37(3): 444-454.ZHANG Weiwei,ZHU Linyang,LIU Yilang,et al. Progresses in the application of machine learning in turbulence modeling[J]. Acta Aerodynamica Sinica,2019,37(3): 444-454. (in Chinese) [3] 李義進,宋紅超,周帥,等. 人工智能在计算流体力学中的应用[J]. 航空动力,2018(1): 57-59.LI Yijin,SONG Hongchao,ZHOU Shuai,et al. The application of artificial intelligence on computational fluid dynamics[J]. Aerospace Power,2018(1): 57-59. (in Chinese) [4] 陈海昕,邓凯文,李润泽. 机器学习技术在气动优化中的应用[J]. 航空学报,2019,40(1): 52-68.CHEN Haixin,DENG Kaiwen,LI Runze. Utilization of machine learning technology in aerodynamic optimization[J]. Acta Aeronautica et Astronautica Sinica,2019,40(1): 52-68. (in Chinese) [5] FONT B,WEYMOUTH G D,NGUYEN V T,et al. Deep learning of the spanwise-averaged Navier-Stokes equations[J]. Journal of Computational Physics,2021,434(10): 110199.1-110199.25. [6] 王述之,战林浩,曹博超,等. 基于直接数值模拟数据和神经网络的湍流封闭模型构建[J]. 水动力学研究与进展:A辑,2020,35(2): 141-154.WANG Shuzhi,ZHAN Linhao,CAO Bochao,et al. Turbulence closure model based on neural network and direct numerical simulation data[J]. Chinese Journal of Hydrodynamics: Series A,2020,35(2): 141-154. (in Chinese) [7] 尹宇辉, 杨普, 张宇飞. 基于机器学习的湍流建模研究[R]. 杭州: 中国力学大会, 2019. [8] 陈海,钱炜祺,何磊. 基于深度学习的翼型气动系数预测[J]. 空气动力学学报,2018,36(2): 294-299.CHEN Hai,QIAN Weiqi,HE Lei. Aerodynamic coefficient prediction of airfoils based on deep learning[J]. Acta Aerodynamica Sinica,2018,36(2): 294-299. (in Chinese) [9] ZHANG Y, SUNG W J, MAVRIS D. Application of convolutional neural network to predict airfoil lift coefficient[R]. Kissimmee, US: AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, 2018. [10] 何磊,钱炜祺,汪清,等. 机器学习方法在气动特性建模中的应用[J]. 空气动力学学报,2019,37(3): 470-479.HE Lei,QIAN Weiqi,WANG Qing,et al. Applications of machine learning for aerodynamic characteristics modeling[J]. Acta Aerodynamica Sinica,2019,37(3): 470-479. (in Chinese) [11] 刘凌君,周越,高振勋. 基于神经网络的翼型气动力计算和反设计方法[J]. 气体物理,2018,3(5): 41-47.LIU Lingjun,ZHOU Yue,GAO Zhenxun. Aerodynamic force calculation and inverse design for airfoil based on neural network[J]. Physics of Gases,2018,3(5): 41-47. (in Chinese) [12] 姚琳. 机器学习在翼型优化问题中的应用[J]. 航天标准化,2020(1): 43-47.YAO Lin. Application of machine learning in airfoil optimization[J]. Aerospace Standardization,2020(1): 43-47. (in Chinese) [13] SEKAR V,ZHANG M,SHU C,et al. Inverse design of airfoil using a deep convolutional neural network[J]. AIAA Journal,2019,57(3): 993-1003. doi: 10.2514/1.J057894 [14] 廖鹏,姚磊江,白国栋,等. 基于深度学习的混合翼型前缘压力分布预测[J]. 航空动力学报,2019,34(8): 1751-1758.LIAO Peng,YAO Leijiang,BAI Guodong,et al. Prediction of hybrid airfoil leading edge pressure distribution based on deep learning[J]. Journal of Aerospace Power,2019,34(8): 1751-1758. (in Chinese) [15] SEKAR V,KHOO B C. Fast flow field prediction over airfoils using deep learning approach[J]. Physics of Fluids,2019,31(5): 057103.1-057103.14. [16] 艾子健,秦国良,陈雪飞,等. 翼型叶片精确参数化建模的研究及应用[J]. 流体机械,2015,43(8): 36-40.AI Zijian,QIN Guoliang,CHEN Xuefei,et al. Research and application on accurate parametric modeling for aerofoil blade[J]. Fluid Machinery,2015,43(8): 36-40. (in Chinese) [17] AUNGIER R H. Axial-flow compressors-a strategy for aerodynamic design and analysis[M]. New York: ASME Press, 2003. [18] FOTTNER L. Test cases for computation of internal flows in aero engine components[R]. AGARD Advisory Report No. 275, 1990. [19] STARK U,HOHEISEL H. The combined effect of axial velocity density ratio and aspect ratio on compressor cascade performance[J]. Journal of Engineering for Power,1981,103(1): 247-255. [20] LECUN Y,BOTTOU L,BENGIO Y,et al. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE,1998,86(11): 2278-2324. doi: 10.1109/5.726791 [21] GLOROT X, BENGIO Y. Understanding the difficulty of training deep feedforward neural networks[R]. Sardinia, Italy: the 13th International Conference on Artificial Intelligence and Statistics, 2010. [22] ELMAN J L. Learning and development in neural networks: the importance of starting small[J]. Cognition,1993,48(1): 71-99. doi: 10.1016/0010-0277(93)90058-4 [23] GOODFELLOW I, WARDE-FARLEY D, MIRZA M, et al. Maxout networks[R]. Atlanta, US: International Conference on Machine Learning, 2013. [24] KRIZHEVSKY A,SUTSKEVER I,HINTON G E. ImageNet classification with deep convolutional neural networks[J]. Advances in Neural Information Processing Systems,2012,25: 1097-1105. [25] BOUREAU Y L, BACH F, LECUN Y, et al. Learning mid-level features for recognition[R]. San Francisco, US: IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2010. [26] HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[R]. Las Vegas, US: IEEE Conference on Computer Vision and Pattern Recognition, 2016.