Automatic Classification of Soot Propensity in Flames Using Image Processing and Machine Learning

Abstract

This paper describes and evaluates different ap- proaches for determining the soot propensity of an axisymmetric laminar diffusion flame with respect to its smoke point. Using a training dataset of images of flames captured in laboratory conditions and labeled according to Line-Of-Sight Attenuation (LOSA) measurements, we trained different classifiers for deter- mining the soot propensity of flames using images in a testing dataset. We evaluate a traditional classifier based on Support Vector Machine (SVM), combined with dimensionality reduction techniques and compare the results with different architectures of Convolutional Neural Networks (CNNs). Experimental results in terms of classification performance and inference time set the proposed classifiers as a promising solution for non-invasive and low-cost instrumentation for characterizing the status of combustion flames in industrial settings.