TY - JOUR
T1 - High-speed planar imaging of OH radicals in turbulent flames assisted by deep learning
AU - Guo, Hao
AU - Zhang, Wei
AU - Nie, Xiangyu
AU - Dong, Xue
AU - Sun, Zhiwei
AU - Zhou, Bo
AU - Wang, Zhenkan
AU - Richter, Mattias
PY - 2022/3/1
Y1 - 2022/3/1
N2 - High-speed planar imaging of key combustion species, like hydroxyl radicals (OH), is crucial for understanding the complex chemistry–turbulence interactions in turbulent flames. However, conducting high-speed (kHz) diagnostics is challenging due to the requirements on advanced optical system, including both fast lasers and cameras. In this paper, we report a computational imaging method to artificially achieve higher diagnosing rates based on experimental data at relatively low rates. Sequencies of planar laser-induced fluorescence (PLIF) of OH recorded at 100 kHz in a turbulent flame were first down sampled to 50 kHz, 33.3 kHz and 20 kHz, respectively, and then used as a data source to train several networks. The accuracies of the models were assessed by comparing the predicted images with those from laser measurements. It was found that, among the models tested, convolutional long short-term memory network (CONV-LSTM) can provide the best predictions and is reliable in predicting consecutive images with higher repetition rates. The model can also generate consecutive OH-PLIF images at 200 kHz based on the 100 kHz experimental data. This work sheds light on the hybrid of deep learning-based computational methods with conventional high-speed laser diagnosing techniques, which can potentially increase the temporal resolution of planar optical measurement in turbulent flows, and significantly reduce the computational timing.
AB - High-speed planar imaging of key combustion species, like hydroxyl radicals (OH), is crucial for understanding the complex chemistry–turbulence interactions in turbulent flames. However, conducting high-speed (kHz) diagnostics is challenging due to the requirements on advanced optical system, including both fast lasers and cameras. In this paper, we report a computational imaging method to artificially achieve higher diagnosing rates based on experimental data at relatively low rates. Sequencies of planar laser-induced fluorescence (PLIF) of OH recorded at 100 kHz in a turbulent flame were first down sampled to 50 kHz, 33.3 kHz and 20 kHz, respectively, and then used as a data source to train several networks. The accuracies of the models were assessed by comparing the predicted images with those from laser measurements. It was found that, among the models tested, convolutional long short-term memory network (CONV-LSTM) can provide the best predictions and is reliable in predicting consecutive images with higher repetition rates. The model can also generate consecutive OH-PLIF images at 200 kHz based on the 100 kHz experimental data. This work sheds light on the hybrid of deep learning-based computational methods with conventional high-speed laser diagnosing techniques, which can potentially increase the temporal resolution of planar optical measurement in turbulent flows, and significantly reduce the computational timing.
U2 - 10.1007/s00340-021-07742-2
DO - 10.1007/s00340-021-07742-2
M3 - Article
AN - SCOPUS:85125261850
SN - 0946-2171
VL - 128
JO - Applied Physics B: Lasers and Optics
JF - Applied Physics B: Lasers and Optics
IS - 3
M1 - 52
ER -