A novel AI-based combustion diagnostic technology for the identification of chemical source information via flame images: Fuel type and reaction condition

Abstract

The improvement of combustion diagnostics was an important basis for developing combustion-related research and technologies. Based on the flame image as the carrier, this work proposed a novel combustion diagnostic technique for the identification of chemical source information through the analysis of abundant flame features contained in the image using a random forest algorithm. Five types of flames under 16 reaction conditions were adopted as experimental cases to examine the accuracy of the proposed technique. Results indicated that the random forest algorithm could well “learn” the correlation mechanism between flame features and chemical source information, and identify the fuel type and reaction condition of unknown flame according to this mechanism. For identifying the fuel type, the accuracy on ethylene, 20 % ethyl acetate, 40 % ethyl acetate, 20 % ethyl butyrate, and 40 % ethyl butyrate doped flames reached 86.2 %, 84.4 %, 100 %, 63.6 %, and 64.3 %, respectively. For identifying the reaction condition, the mean absolute error, max error, and mean absolute percentage error between the predicted O2 content and the actual value were 0.96 %, 2.53 %, and 3.46 %, respectively. Moreover, the feature analysis result indicated that the importance degree of each feature was not equivalent between the different identification assignments. The flame features such as rectangularity and the variance of G and B in the blue region were the key features for identifying the fuel type. However, the variance of R and R/B of the blue region played a significant role in identifying the reaction condition.