Tomographic measurements of temperature and species concentration via deep learning and wavelength modulated absorption spectroscopy

Abstract

To achieve the 2-D nonintrusive measurements of temperature and species concentration in the combustion field, a new framework, combining calibration-free wavelength modulation absorption spectroscopy (CF-WMS) with a designed convolutional neural network (CNN), was developed. The principle of the CF-WMS, along with the architecture of the CNN net, the training, and the performance of the network, has been investigated. The region of interest was discretized into 24 × 24 pixels2, and 48 probing beams with six targeted frequencies were used to verify the feasibility of the designed CNN with WMS 2f/1f signal for temperature and species concentration reconstruction. 20 000 samples of temperature and water vapor concentration distributions are randomly fabricated, featuring three randomly positioned Gaussian distributions. Reconstructed images of the phantoms agreed well with the original distributions with the relative error of about 5.0%–9.2% and 8.0%–12.4% using 17 000 training datasets with different beam arrangements for temperature and H2O species, respectively. Several representative beam arrangements with a limited number have been examined and compared. The beam arrangement BA2 and BA3 exhibited better performance than BA1 with average errors of about 5% and 8% for temperature and H2O species, respectively. Such a method can provide an effective way to achieve spatially and temporally resolved, real-time, in situ monitoring in practical combustion environments.