In aircraft engines or in some industrial cameras happen a number of combustion reactions in extreme conditions that make these in very aggressive environments for research. There is too complicated to study and control of these processes given the high pressures and temperatures prevailing in these scenarios.

However, a group of scientists investigating UC3M how to improve measurements at these locations. Esteban García-Cuesta, under the direction of professors Anthony J. Ines de Castro and M. Galvan, the Departments of Physics and Informatics of UC3M, respectively, working in a multidisciplinary project for the recovery of physical properties in combustion processes, specifically the temperature of the flame. This property is very important because it contains all the chemical information of the reaction and, knowing it, could determine the overall state of combustion and change automatically depending on the needs.

Castro and his team made computer simulations of optical techniques, in particular the spectra in the infrared emission of the gases emitted in these reactions, such as carbon dioxide (CO2), water vapor, carbon monoxide (CO) or oxides of nitrogen (NOx). These compounds, present major emission bands in the infrared region of the spectrum, have been selected by the investigators as parameters to measure. Spectral information is related mathematically to the profiles of flame temperature, so that its measure enables the retrieval of this information within the combustion chamber itself. However, “the conversion of these spectra in the temperature profile shows a nonlinear relationship, so it is not a trivial problem – says Professor Castro – because different temperatures may lead to similar spectra, making the solution is unique, “concludes the researcher.

Researchers have applied a type of neural network (multilayer perceptron) To make the investment in thermal spectral data. As the spectral measurements are made at high resolution the amount of calculation required for processing and use, technical calls machine learningIs high and may even decrease performance. According to Castro, “we need a lot of spectral information to measure the spectrum well, but information overload is not suitable for a neural network and we must develop techniques for the selection or feature extraction to reduce the number of input parameters without losing information” . Therefore, the researchers apply the Training or learning of the network as “an intelligent selection of information, in this case wavelengths to extract the physical information that interests us and avoid redundant.’s how des the information network is very important and this point is our research, “adds the researcher. The next step facing scientists is to measure spectra in real systems. So far, measurements have achieved quite approximate theoretical temperature (about 3 degrees Kelvin tighter in the hottest area of the flame). These computational studies aim to determine the feasibility of data conversion techniques.

The temperature monitoring with the use of so-called feedback systems allow automatic control of combustion processes. The latter received the information from the combustion reaction, including the state you are in and fed back into the system. Also have the ability to change variables such as inflow of combustible gases, thus obtaining better control over the overall process. This is very important to reduce pollutant emissions and maximize energy losses.

The study “Multilayer perceptron as inverse model in a ground-based remote sensing temperature retrieval problem “ has been published in the journal Engineering Applications of Artificial Intelligence by Esteban García Cuesta, Ines M. Galván, Antonio J. De Castro, researchers UC3M.

More
Office of Scientific Information UC3M