Data-driven Analysis of Turbulent Flame Images

TL;DR

This paper employs CNNs to classify unburned pockets in turbulent premixed flames with CO2 addition, achieving high accuracy and aiding understanding of transient flame phenomena.

Contribution

It introduces a CNN-based method for classifying unburned pockets in turbulent flames, enhancing flame characterization during transient events.

Findings

CNN achieved over 85% accuracy across all flame conditions.

Unburned pockets are effectively identified using deep learning.

CO2 addition influences the classification accuracy.

Abstract

Turbulent premixed flames are important for power generation using gas turbines. Improvements in characterization and understanding of turbulent flames continue particularly for transient events like ignition and extinction. Pockets or islands of unburned material are features of turbulent flames during these events. These features are directly linked to heat release rates and hydrocarbons emissions. Unburned material pockets in turbulent CH$_4$/air premixed flames with CO$_2$ addition were investigated using OH Planar Laser-Induced Fluorescence images. Convolutional Neural Networks (CNN) were used to classify images containing unburned pockets for three turbulent flames with 0%, 5%, and 10% CO$_2$ addition. The CNN model was constructed using three convolutional layers and two fully connected layers using dropout and weight decay. The CNN model achieved accuracies of 91.72%, 89.35% and 85.80% for the three flames, respectively.