Detection of slight variations in combustion conditions with machine learning and computer vision

When monitoring combustion conditions, detecting minor variations, which may be complex even for the human eye, is critical for providing a fast response and correcting deviations. The aim of this study is to detect slight variations in combustion conditions by developing a flame monitoring system using machine learning and computer vision techniques applied to color images. Predictive models are developed for fuel blends with different heating values. The predictive models classify the combustion equivalence ratio based on multiple conditions, using a mean step size of 0.10 between states, a lower value than previously reported in related studies. Three machine learning algorithms are used for each fuel blend: logistic regression, support vector machine, and artificial neural network (multilayer perceptron). These models are fed the statistical, geometrical, and textural features extracted from the color images of the flames. The classification achieves accuracies from 0.78 to 0.97 in the detection of slight variations in the combustion conditions for all heating values. Thus, the monitoring system developed in this study is a promising alternative for implementation on an industrial scale and quick detection of changes in combustion conditions