Tomographic Imaging and Deep Learning based Reconstruction of Burner Flames

This paper presents a tomographic and deep learning (DL) technique for the three-dimensional (3-D) reconstruction of burner flames. Two-dimensional (2-D) flame images are obtained using a tomographic imaging system from different directions around the burner. A flame data augmentation technique using a morphological operator is used to generate the complete training and testing datasets. The simultaneous algebraic reconstruction technique (SART) is used to generate the ground truth, i.e., flame cross-sectional datasets. A DL method based on a convolutional neural network (CNN) is employed for the reconstruction of the flame cross- and longitudinal sections. The CNN parameters are optimized through a trial-and-error approach as well as simulation. The CNN is constructed using a machine learning (ML) hardware accelerator i.e., a tensor processing unit to perform faster reconstruction. The proposed model is evaluated using the 2-D flame images obtained on a lab-scale gas-fired test rig under different operation conditions. Results obtained from the experiments suggest that the proposed strategy can accurately and faster reconstruct the flame cross- and longitudinal sections.