Intelligent system for monitoring and stoichiometric optimization of combustion

This research work describes an approach for recognition of the actual state of a fossil fuels combustion process respect to its stoichiometry, through a multi-layer feedforward artificial neural network with backpropagation training algorithm. The input/output patterns are made up the statistical moments of the probability distribution function and the principal components of the electromagnetic radiation signals emitted by the flame. A solid state optical detector and a Labview data acquisition program are employed to captured and store the signals. Then, a processing is made with Matlab to extract information from them, in order to integrate the training patterns. Once the neural network has been properly trained, is performed a test process to assess its generalization capability, using new data sets, that the training algorithm has never seen before. With this system, we have achieved a perfect recognition in four flame states, finding that the signals, which actually are used solely to determine either presence or absence of the flame, contain information that can be extracted and analyzed to help to keep the process as closely as possible to the stoichiometric conditions.