Retrieving soot volume fraction fields for laminar axisymmetric diffusion flames using convolutional neural networks. (1st February 2021)

• Record Type: Journal Article
• Authors: Rodríguez, A.; Escudero, F.; Cruz, J.J.; Carvajal, G.; Fuentes, A.
• Is Part Of: Fuel. Volume 285(2021)
• Publication Date: 2021-02-01
• Subjects: Soot diagnostics -- CoFlame code -- Synthetic images -- LOSA technique -- Ill-posed problem -- Artificial neural networks; Fuel -- Periodicals; Coal -- Periodicals; Coal; Fuel; Periodicals; 662.6
• DOI: 10.1016/j.fuel.2020.119011
• Languages: English
• ISSNs: 0016-2361

Abstract

Graphical abstract: Highlights: Numerical framework for simulating experimental measurements for axisymmetric flames. The framework enables generation of rich datasets for reference and measured signals. A Convolutional Neural Network (CNN) is used to retrieve soot volume fraction fields. CNN outperforms traditional inversion methods based on numerical deconvolution. CNN trained with synthetic data can retrieve the fields from real experimental data. Abstract: Typical procedures for estimating soot volume fraction distribution in laboratory flames require solving ill-posed inverse problems to recover the fields from convoluted signals that integrate light extinction from soot particles along the line-of-sight of a photo-detector. Classical deconvolution methods are highly sensitive to noise and the choice of tunable regularization parameters, which prevents obtaining consistent estimations even for the same reference flame settings. This paper presents a novel approach based on Convolutional Neural Networks (CNNs) for estimating the soot volume fraction fields from 2D images of line-of-sight attenuation (LOSA) measurements in coflow laminar axisymmetric diffusion flames. Using a set of reference synthetic soot volume fraction fields of canonical flames and their corresponding projected LOSA images, we trained a CNN for reconstructing soot fields from images representing the data captured by a camera. Experimental results show that the proposed CNN approach outperforms classical deconvolution methods when reconstructing the flame spatial soot distribution from noisy images of LOSA.