Study On Absorption Spectral Tomography Based On Deep Learning

Fuel combustion is the main source of energy in the world today.Combustion diagnosis,which has great significance to improve combustion efficiency and reduce atmospheric pollutant emissions,is one of the most important methods to improve and optimize the combustion process.Tunable Diode Laser Absorption Tomography（TDLAT）is an important optical combustion diagnostic method,which has advantages of non-intrusive,high sensitivity and rapid response,and can realize the reconstruction of the gas temperature and concentration distribution in the cross-section of combustion device.The measurement space provided by some combustion diagnosis scenarios is limited,which limits the amount of projection data that can be obtained.In this case,existing TDLAT reconstruction algorithms have problems such as poor noise resistance and high computational cost.Deep Learning（DL）is a method of representing learning for a large amount of data,which has the advantage of self-learning data features,and can quickly realize the nonlinear mapping process between the measured data and the reconstructed data.Combining with the latest theories of deep learning,a TDLAT reconstruction algorithm based on Convolutional Neural Networks（CNN）and a TDLAT reconstruction algorithm based on multi-task learning are studied in this paper,which can realize the rapid and accurate reconstruction of flame parameter distributions with different characteristics.The main work is as follows:1.Aiming at the problem that it is difficult for the existing TDLAT field reconstruction algorithm to reconstruct the gas temperature distribution rapidly and accurately when the projection data is insufficient,the TNNAT reconstruction algorithm based on CNN is studied and improved.The basic structure and optimization method of CNN for TDLAT reconstruction is established.The quantitative analysis of the integral absorbance data of the input network and the temperature data of the output network is carried out,and the importance of its synchronous preprocessing to improve the reconstruction accuracy of the algorithm is analyzed.A step-by-step learning mode is proposed by using the characteristics of smooth distribution of flame parameters.The influence of CNN configuration on the reconstruction accuracy of the algorithm is analyzed,and the optimal CNN configuration is established.2.As the combustion diagnosis scenes is complex and variable,the applicability of the improved algorithm is verified under different characteristic flames and multiple beam arrangements,also,a set of planar flame furnace temperature measurement system which is used to verify the improved algorithm is built.The improved algorithm is compared with the existing CNN-based TDLAT reconstruction algorithm on Gaussian phantom.The results show that the improved algorithm has higher reconstruction accuracy and robustness on different characteristic flames than existing algorithm.In the absence of noise,the improved algorithm’s average reconstruction error for the temperature distribution of 1432 different Gaussian phantom is only 0.24%.The algorithm is further verified on turbulence methane plume,the results show that the improved algorithm can better reconstruct the temperature distribution in time series,and the reconstruction process of a single group of data can reach 0.048ms at the earliest The algorithm is also validated under 10 different beam arrangements,which show good adaptability.The temperature measurement experiment of flat flame burner is carried out,the experimental results show that the improved algorithm can reconstruct the two-dimensional temperature distribution of the flame section of a planar flame furnace under different combustion states.3.A algorithm based on multi-task learning to reconstruct temperature and H₂O concentration distribution synchronously is proposed for the first,which realize the flame cross-section temperature and H₂O concentration distribution under different combustion conditions.The temperature and H₂O concentration synchronous reconstruction algorithm based on multi-task learning is studied.This algorithm can not only reconstruct the two gas parameters synchronously,but also save the calculation cost and has stronger anti-noise ability.The section of the flame z=1.5cm on the flat flame burner was measured using proposed algorithm,and the reconstructed image clearly reflects the profile of the flame temperature and H₂O concentration distribution.