Shock Wave Intelligent Prediction Method for Hypersonic Vehicle
-
摘要: 高超声速飞行器激波位置的准确预测能够有效提升数值模拟的精度和效率。一方面, 对高超声速飞行器激波附近网格进行正交和加密处理, 可有效提升数值计算精度; 另一方面, 使用高超声速飞行器激波位置对计算网格进行修正, 能够加速CFD计算收敛过程。提出了一种基于机器学习的高超声速飞行器激波智能预测方法, 对典型高超声速飞行器外形进行激波位置的高效准确预测。首先, 针对典型高超声速飞行器外形和典型飞行状态, 使用数值模拟方法获得收敛的流场, 并采用基于Mach数等值线的激波提取方法, 从流场中判别激波面并提取构成激波面的关键点位置, 形成训练数据; 然后采用有监督学习算法, 学习关键点位置, 并利用二次曲线沿流向拟合关键点形成初步的激波线族; 最后, 基于剖面压力云图, 构造基于投影压力图像的智能预测神经网络, 对初步形成的激波线族进行修正, 并获得三维激波面。大量的实验结果表明, 激波预测模型能够对高超声速飞行器激波位置做出准确预测, 预测的激波面与CFD数值计算结果中提取的激波面误差在10-4量级。Abstract: Accurate prediction of shock wave position of hypersonic aircrafts can effectively improve the accuracy and efficiency of computational fluid dynamics (CFD) simulation. On the one hand, orthogonalization and densification of the grid near the shock wave of the hypersonic vehicle can effectively improve the numerical accuracy. On the other hand, using the shock wave position of the hypersonic vehicle to correct the computational grid can speed up the CFD convergence process. A shock wave intelligent prediction method for hypersonic vehicles based on machine learning was proposed, which could efficiently and accurately predict the shock position of the typical hypersonic aircraft shape. Firstly, for the typical hypersonic vehicle shape and typical flight state, numerical methods were used to obtain a convergent flow field. Secondly, the shock wave extraction method based on Mach number contour was used to identify the shock wave surface from the flow field and extract the key points that constitute the shock wave to form training data. After that, the supervised learning method was used to predict the positions of these key points and the quadratic curve was used to fit these key points along the flow direction to form a preliminary shock line family. Finally, based on the typical pressure profile, an image-based neural network was constructed to correct the preliminary shock line family and obtain the three-dimensional shock surface. A large number of experimental results show that the shock wave prediction model can effectively predict the shock wave position of the hypersonic vehicle, and the error between the reconstructed shock wave surface and the extracted shock surface from the CFD results is in the order of 10-4.
-
Key words:
- numerical simulation /
- CFD /
- shock wave /
- machine learning /
- neural network
-
表 1 基于Mach数等值线的激波面提取算法
Table 1. Shock wave extraction algorithm based on Mach number contour
input: u,v,w,ρ,g,xk,yk,Mainput output: xshock,yshock,zshock 1 for any grid point P(i, j, k)do 2 $v_{\mathrm{F}}=\sqrt{u^2+v^2+w^2}$ 3 $v_{\text {sound }}=\sqrt{\gamma \times p / \rho}$ 4 ${Ma}(i, j, k)=\frac{v_{\mathrm{F}}}{v_{\text {sound }}}$ 5 Mashock=Mainput×0.99 6 for any grid point P(i, j, k)do 7 if Ma(i, j, k)≤Mashockand Ma(i, j, k+1)>Mashock then 8 put P(i, j, k) into set S 9 for Pi, j, k∈S do 10 di, j, k=di, j, k-1+d(Pi, j, k, Pi, j, k-1) 11 dshock=di, j, k+$\frac{d_{i, j, k+1}-d_{i, j, k}}{{Ma}(i, j, k+1)-M a(i, j, k)}$×
(Mashock-Ma(i, j, k))12 xshock=x(i, j, 1)+$\frac{d_{\text {shock }}}{d_{i, j, k}}$×(x(i, j, k)-x(i, j, 1)) 13 yshock=y(i, j, 1)+$\frac{d_{\text {shook }}}{d_{i, j, k}}$×(y(i, j, k)-y(i, j, 1)) 14 zshock=z(i, j, 1)+$\frac{d_{\text {shook }}}{d_{i, j, k}}$×(z(i, j, k)-z(i, j, 1)) 表 2 不同外形的MSEave与MAEave
Table 2. MSEave and MAEave of different shapes
shape MSEave MAEave blunt cone 1.482 6×10-4 7.652 7×10-3 double cone 2.184 2×10-4 1.088 1×10-2 lifting body 2.382 8×10-4 1.145 3×10-2 表 3 浮点运算次数计算结果
Table 3. Calculation results of the number of floating point operations
model FLOPs shape-net 4.0×109 inlet-net 1.54×105 concatenated-net 1.89×109 shock-net 5.89×109 表 4 预测的压力场与CFD计算得到的压力场误差
Table 4. Error between the predicted pressure field and the pressure field calculated by CFD
shape MSEave MAEave blunt cone 1.655 7×10-10 7.394 2×10-9 double cone 1.724 1×10-10 8.566 4×10-9 double ellipsoid 2.732 8×10-10 1.299 1×10-8 lifting body 1.868 2×10-10 8.711 9×10-9 表 5 不同外形的MSEave
Table 5. MSEave of different shapes
shape MSEave MAEave blunt cone 1.185 7×10-4 5.815 5×10-3 double cone 1.687 3×10-4 8.622 7×10-3 lifting body 1.778 5×10-4 8.760 5×10-3 表 6 双椭球外形的MSEave与MAEave
Table 6. MSEave and MAEave of double ellipsoid shape
shape MSEave MAEave double ellipsoid 2.519 1×10-4 1.156 9×10-2 -
[1] Landau L D, Lifshitz E M. Fluid mechanics[M]. 2nd ed. Amsterdam: Elsevier, 1987. [2] Nath G, Dutta M, Chaurasia S. Exact solution for isothermal flow behind a shock wave in a self-gravitating gas of variable density in an azimuthal magnetic field[J]. Journal of Engineering Physics and Thermophysics, 2020, 93(5): 1247-1254. doi: 10.1007/s10891-020-02228-y [3] Wu Z N, Xu Y Z, Wang W B, et al. Review of shock wave detection method in CFD post-processing[J]. Chinese Journal of Aeronautics, 2013, 26(3): 501-513. doi: 10.1016/j.cja.2013.05.001 [4] 姜维, 杨云军, 陈河梧. 带减阻杆高超声速飞行器外形气动特性研究[J]. 实验流体力学, 2011, 25(6): 28-32, 53. doi: 10.3969/j.issn.1672-9897.2011.06.006Jiang W, Yang Y J, Chen H W. Investigations on aerodynamics of the spike-tipped hypersonic vehicles[J]. Journal of Experiments in Fluid Mechanics, 2011, 25(6): 28-32, 53 (in Chinese). doi: 10.3969/j.issn.1672-9897.2011.06.006 [5] Liu Y, Lu Y T, Wang Y Q, et al. A CNN-based shock detection method in flow visualization[J]. Computers & Fluids, 2019, 184: 1-9. [6] 彭俊, 胡宗民, 姜宗林. 基于机器学习预测激波相互作用位置的研究[C]. 第十一届全国流体力学学术会议论文集. 深圳: 中国力学学会, 2020: 939.Peng J, Hu Z M, Jiang Z L. Research on prediction of shock interaction position based on machine learning[C]. The 11th National Academic Conference on Fluid Mechanics. Shenzhen: Chinese Society of Theoretical and Applied Mechanics, 2020: 939(in Chinese). [7] Beck A D, Zeifang J, Schwarz A, et al. A neural network based shock detection and localization approach fordiscontinuous Galerkin methods[J]. Journal of Computational Physics, 2020, 423: 109824. doi: 10.1016/j.jcp.2020.109824 [8] Monfort M, Luciani T, Komperda J, et al. A deep learning approach to identifying shock locations in turbulent combustion tensor fields[A]. Schultz T, Özarslan E, Hotz I. Modeling, Analysis, and Visualization of Anisotropy[M]. Cham: Springer, 2017: 375-392. [9] Kossaczká T, Ehrhardt M, Günther M. Enhanced fifth order WENO shock-capturing schemes with deep learning[J]. Results in Applied Mathematics, 2021, 12: 100201. doi: 10.1016/j.rinam.2021.100201 [10] Paciorri R, Bonfiglioli A. Accurate detection of shock waves and shock interactions in two-dimensional shock-capturing solutions[J]. Journal of Computational Physics, 2020, 406: 109196. doi: 10.1016/j.jcp.2019.109196 [11] GrzywińskiS. Research of vibroacoustic disturbances influence on the shock wave detection process[C]. 25th International Conference on Engineering Mechanics 2019. Svratka, 2019: 129-132. [12] Zheng L, Lawlor B, Katko B J, et al. Image processing and edge detection techniques to quantify shock wave dynamics experiments[J]. Experimental Techniques, 2021, 45(4): 483-495. doi: 10.1007/s40799-020-00415-3 [13] Yao Z C, Zhang H, Zhang X J, et al. Analysis on the influence mechanism of conical shock wave on pulsed laser forward detection[J]. Optik, 2019, 183: 783-791. doi: 10.1016/j.ijleo.2019.02.093 [14] Gettle G L, Homer Jr V H. Acoustic/shock wave attenuating assembly: US, 5225622[P]. 1993-07-06. [15] 张涛, 郑晓刚, 汤祎麒, 等. 基于非轴对称吻切技术的三维激波逆向乘波设计[J]. 航空科学技术, 2020, 31(11): 35-46. https://www.cnki.com.cn/Article/CJFDTOTAL-HKKX202011006.htmZhang T, Zheng X G, Tang Y Q, et al. Inverse waverider design for 3D shock wave based on non-axisymmetric osculating cones method[J]. Aeronautical Science and Technology, 2020, 31(11): 35-46 (in Chinese). https://www.cnki.com.cn/Article/CJFDTOTAL-HKKX202011006.htm [16] 涂国华, 袁湘江, 陆利蓬. 激波捕捉差分方法研究[J]. 应用数学和力学, 2007, 28(4): 433-440. doi: 10.3321/j.issn:1000-0887.2007.04.008Tu G H, Yuan X J, Lu L P. Developing shock-capturing difference methods[J]. Applied Mathematics and Mechanics, 2007, 28(4): 433-440 (in Chinese). doi: 10.3321/j.issn:1000-0887.2007.04.008 [17] 刘君, 邹东阳, 徐春光. 基于非结构动网格的非定常激波装配法[J]. 空气动力学学报, 2015, 33(1): 10-16. https://www.cnki.com.cn/Article/CJFDTOTAL-KQDX201501002.htmLiu J, Zou D Y, Xu C G. An unsteady shock-fitting technique based on unstructured moving grids[J]. Acta Aerodynamica Sinica, 2015, 33(1): 10-16 (in Chinese). https://www.cnki.com.cn/Article/CJFDTOTAL-KQDX201501002.htm [18] 宗文刚, 邓小刚, 张涵信. 双重加权实质无波动激波捕捉格式[J]. 空气动力学学报, 2003, 21(2): 218-225. https://www.cnki.com.cn/Article/CJFDTOTAL-KQDX200302013.htmZong W G, Deng X G, Zhang H X. Double weighted essentially non-oscillatory shock-capturing schemes[J]. Acta Aerodynamica Sinica, 2003, 21(2): 218-225 (in Chinese). https://www.cnki.com.cn/Article/CJFDTOTAL-KQDX200302013.htm [19] 张玉东, 傅德薰, 马延文, 等. 高精度非定常激波装配法[J]. 计算物理, 2007, 24(5): 533-536. https://www.cnki.com.cn/Article/CJFDTOTAL-JSWL200705006.htmZhang Y D, Fu D X, Ma Y W, et al. A high-order unsteady shock-fitting scheme[J]. Chinese Journal of Computational Physics, 2007, 24(5): 533-536(in Chinese). https://www.cnki.com.cn/Article/CJFDTOTAL-JSWL200705006.htm [20] Payedimarri A B, Concina D, Portinale L, et al. Prediction models for public health containment measures on COVID-19 using artificial intelligence and machine learning: a systematic review[J]. International Journal of Environmental Research and Public Health, 2021, 18(9): 4499. [21] Ji M, Liu Y M, Zhao M D, et al. Use of machine learning algorithms to predict the understandability of health education materials: development and evaluation study[J]. JMIR Medical Informatics, 2021, 9(5): e28413. [22] 徐军, 丁宇新, 王晓龙. 使用机器学习方法进行新闻的情感自动分类[J]. 中文信息学报, 2007, 21(6): 95-100. https://www.cnki.com.cn/Article/CJFDTOTAL-MESS200706016.htmXu J, Ding Y X, Wang X L. Sentiment classification for Chinese news using machine learning methods[J]. Journal of Chinese Information Processing, 2007, 21(6): 95-100 (in Chinese). https://www.cnki.com.cn/Article/CJFDTOTAL-MESS200706016.htm [23] 薛晓娟, 贾元华, 王晟由, 等. 基于机器学习的铁路建设间接经济效益预测[J]. 武汉理工大学学报, 2021, 43(8): 43-50. https://www.cnki.com.cn/Article/CJFDTOTAL-WHGY202108007.htmXue X J, Jia Y H, Wang S Y, et al. Research on indirect economic benefits of railway construction based on machine learning[J]. Journal of Wuhan University of Technology, 2021, 43(8): 43-50 (in Chinese). https://www.cnki.com.cn/Article/CJFDTOTAL-WHGY202108007.htm [24] Moon J, Jung S, Park S, et al. Machine learning-based two-stage data selection scheme for long-term influenza forecasting[J]. Computers, Materials & Continua, 2021, 68(3): 2945-2959. [25] dos Santos Luciano A C, Picoli M C A, Duft D G, et al. Empirical model for forecasting sugarcane yield on a local scale in Brazil using Landsat imagery and random forest algorithm[J]. Computers and Electronics in Agriculture, 2021, 184: 106063.