Online fuel tracking by combining principal component analysis and neural network techniques

This paper presents a novel approach to the online tracking of pulverized fuel during combustion. A specially designed flame detector containing three photodiodes is used to derive multiple signals covering a wide spectrum of flame radiation from the infrared to ultraviolet regions through the visible band. Various flame features are extracted from the time and frequency domains. A back-propagation neural network is deployed to map the flame features to an individual type of fuel. The neural network has incorporated principal component analysis to reduce the complexity of the network and hence its training time. Experimental tests were conducted on a 0.5 MW_th combustion test facility using eight different types of coal. Results obtained demonstrate that the approach is effective for the online identification of the type of fuel being fired under steady combustion conditions, and the average success rate is 93.4%.