Temperature Tomography for Combustion Field Based on Hierarchical Vision Transformer and Multi-scale Features Merging
-
摘要: 可调谐二极管激光吸收光谱层析成像(TDLAT)是一种重要的光学非侵入式燃烧诊断技术，可实现燃烧场2维横截面气体温度和浓度等流场参数分布的重建。该文将视觉Transformer(ViT)与多尺度特征融合引入TDLAT领域，研究有限数量测量数据与整个测量空间温度分布的非线性映射，提出基于级联ViT与多尺度特征融合的燃烧场温度层析成像网络(HVTMFnet)。该网络提取并融合TDLAT测量数据的局部-全局相关特征，实现整个测量空间的层次化温度分布重建。仿真实验与实际TDLAT系统实验均表明，HVTMFnet重建图像的质量优于现有的基于卷积神经网络(CNN)和基于残差网络的温度层析成像方案。与基于CNN的温度层析成像方案相比，HVTMFnet的重建误差能够降低49.2%～72.1%。
-
关键词:
- 可调谐二极管激光吸收光谱 /
- 层析成像 /
- 温度重建 /
- 视觉Transformer
Abstract: Tunable Diode Laser Absorption Tomography (TDLAT) is an important non-intrusive combustion diagnostic technology, which can be used to reconstruct two-dimensional cross-sectional distributions of flow-field parameters such as gas temperature and concentration in the combustion field. In this paper, Vision Transformer (ViT) and multi-scale features merging are introduced into TDLAT to study the nonlinear mapping between a limited number of measurement data and the temperature distribution in the entire tomographic filed. Temperature tomography network (HVTMFnet) is proposed based on the hierarchical Vision Transformer (ViT) and Multi-scale Features merging. By extracting and merging the local and global correlation characteristics of TDLAT measurement data, HVTMFnet reconstructs the hierarchical temperature distribution in the entire tomographic field. Both simulations and lab-scale experiments with TDLAT system show that HVTMFnet retrieves better-quality temperature images than existing temperature tomography schemes based on Convolutional Neural Network (CNN) and residual network. In comparison to the temperature tomography scheme based on CNN, HVTMFnet reduces the reconstruction error by 49.2%～72.1%. -
表 1 实验中HVTMFnet模型的参数设置
${H_{m - 1}} \times {W_{m - 1}} \times {C_{m - 1}}$ ${H_{{\text{patch}}}}$ ${W_{{\text{patch}}}}$ $N_{{\text{patch}}}^m$ ${D_m}$ $D_{{\text{FFN}}}^m$ $N_{{\text{layer}}}^m$ $N_{{\text{head}}}^m$ $d_{{\text{head}}}^m$ ${H_m} \times {W_m} \times {C_m}$ m=1 4×8×2 2 2 8 16 64 2 2 8 2×4×16 m=2 2×4×16 2 2 2 128 512 6 8 16 1×2×128 
下载: 导出CSV
表 2 实验中采用的数据集参数设置
$ {T_{{\text{min}}}} $ $ {X_{{\text{min}}}} $ $ {u_d} $ $ {v_d} $ $ {\beta _d} $ $ {\mu _{x,d}} $ $ {\mu _{y,d}} $ $ {\sigma _d} $ 300 K 0.01 U(300,600) U(0.10,0.11) U(0.70,1) U(34,65) U(34,65) U(10,25)
下载: 导出CSV
表 3 4种网络模型训练时间及重建时间的对比
H-CNN HTT-ResNet PI-CNN HVTMFnet 训练时间(min) 0.81 2.67 2.86 6.10 重建时间(s) 2.23×10–6 7.24×10–6 5.45×10–5 7.83×10–6
下载: 导出CSV
-
[1] CAI Weiwei and KAMINSKI C F. A tomographic technique for the simultaneous imaging of temperature, chemical species, and pressure in reactive flows using absorption spectroscopy with frequency-agile lasers[J]. Applied Physics Letters, 2014, 104(3): 034101. doi: 10.1063/1.4862754 [2] JANG H and CHOI D. Similarity analysis for time series-based 2D temperature measurement of engine exhaust gas in TDLAT[J]. Applied Sciences, 2020, 10(1): 285. doi: 10.3390/app10010285 [3] HUANG Jianqing, ZHAO Jianan, and CAI Weiwei. Compressing convolutional neural networks using POD for the reconstruction of nonlinear tomographic absorption spectroscopy[J]. Computer Physics Communications, 2019, 241: 33–39. doi: 10.1016/j.cpc.2019.03.020 [4] 蔡天赋, 李明玉, 靳一, 等. 基于Landweber迭代算法的欠采样恢复数字预失真技术[J]. 电子与信息学报, 2021, 43(11): 3166–3173. doi: 10.11999/JEIT201051CAI Tianfu, LI Mingyu, JIN Yi, et al. An under-sampling restoration digital predistortion technique based on Landweber iteration algorithm[J]. Journal of Electronics &Information Technology, 2021, 43(11): 3166–3173. doi: 10.11999/JEIT201051 [5] 刘云, 薛盼盼, 李辉, 等. 基于深度学习的关节点行为识别综述[J]. 电子与信息学报, 2021, 43(6): 1789–1802. doi: 10.11999/JEIT200267LIU Yun, XUE Panpan, LI Hui, et al. A review of action recognition using joints based on deep learning[J]. Journal of Electronics &Information Technology, 2021, 43(6): 1789–1802. doi: 10.11999/JEIT200267 [6] HUANG Jianqing, LIU Hecong, DAI Jinghang, et al. Reconstruction for limited-data nonlinear tomographic absorption spectroscopy via deep learning[J]. Journal of Quantitative Spectroscopy and Radiative Transfer, 2018, 218: 187–193. doi: 10.1016/j.jqsrt.2018.07.011 [7] 王明, 向鹏, 祁建民, 等. 基于改进卷积神经网络的激光吸收光谱层析成像[J]. 应用激光, 2021, 41(4): 890–901. doi: 10.14128/j.cnki.al.20214104.890WANG Ming, XIANG Peng, QI Jianmin, et al. Reconstruction for tunable diode laser absorption tomography based on convolutional neural networks[J]. Applied Laser, 2021, 41(4): 890–901. doi: 10.14128/j.cnki.al.20214104.890 [8] SI Jingjing, LI Guoliang, CHENG Yinbo, et al. Hierarchical temperature imaging using pseudoinversed convolutional neural network aided TDLAS tomography[J]. IEEE Transactions on Instrumentation and Measurement, 2021, 70: 4506711. doi: 10.1109/TIM.2021.3110282 [9] 司菁菁, 付庚宸, 程银波, 等. 基于层次化离散与残差网络的可调谐二极管激光吸收光谱层析成像[J]. 电子与信息学报, 2022, 44(7): 2539–2546. doi: 10.11999/JEIT210160SI Jingjing, FU Gengchen, CHENG Yinbo, et al. Tunable diode laser absorption tomography based on hierarchical discretization and residual network[J]. Journal of Electronics &Information Technology, 2022, 44(7): 2539–2546. doi: 10.11999/JEIT210160 [10] VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[C]. The 31st International Conference on Neural Information Processing Systems, Long Beach, USA, 2017: 6000–6010. [11] DOSOVITSKIY A, BEYER L, KOLESNIKOV A, et al. An image is worth 16x16 words: Transformers for image recognition at scale[C/OL]. The 9th International Conference on Learning Representations, 2021. [12] HENDRYCKS D and GIMPEL K. Gaussian error linear units (GELUs)[J]. arXiv: 1606.08415, 2016. [13] BAO Yong, ZHANG Rui, ENEMALI G, et al. Relative entropy regularized TDLAS tomography for robust temperature imaging[J]. IEEE Transactions on Instrumentation and Measurement, 2020, 70: 4501909. doi: 10.1109/TIM.2020.3037950 [14] WEI Chuyu, SCHWARM K K, PINEDA D I, et al. Physics-trained neural network for sparse-view volumetric laser absorption imaging of species and temperature in reacting flows[J]. Optics Express, 2021, 29(14): 22553–22566. doi: 10.1364/OE.427730 [15] KINGMA D P and BA J. Adam: A method for stochastic optimization[C]. The 3rd International Conference on Learning Representations, San Diego, USA, 2015. -
图(8) / 表(3)
计量
- 文章访问数: 563
- HTML全文浏览量: 359
- PDF下载量: 175
- 被引次数: 0
下载:
