Infrared remote sensing to measure the temperature of flames

In airplane engines or in some industrial combustion chambers, combustion reactions take place in extreme conditions of heat and pressure, making these environments too harsh for investigation. Nevertheless, a research group from the Universidad Carlos III of Madrid (UC3M) seeks new methods to improve sensing and data collection in such places.

Esteban García-Cuesta, supervised by Antonio J. de Castro and Inés M. Galván, from the departments of physics and Information Technology (IT) of the UC3M respectively, works in a multidisciplinary project for the recovery of physical properties of combustion processes, more specifically the temperature of the flame. This property is very important since it represents all the chemical information about the reaction. Hence, knowing the temperature, the global status of the reaction could be determined and modified automatically when required.

Castro and his team use computer simulations of optical techniques, in particular the infrared emission spectroscopy of the gases produced in the reaction, such as carbon monoxide (CO), or nitrogen oxides (NOx). These compounds, that have clear infrared emission spectral bands, have been selected by the researchers as the parameters to be measured. The spectral information is mathematically related to the temperature of the flame, allowing for its calculation inside a combustion chamber. However, there is the added complication that the translation of the spectra into a temperature profile exhibits a non linear relationship which Professor Castro considers a significant problem, since different temperatures could have very similar spectra, hence producing multiple answers rather than an unique value for the temperature.

Neural networks

The research team applied the multilayer perceptron neural network to translate the spectral data in to thermal data. Since the spectral measurements are taken in high definition, the required calculations for their interpretation using “machine learning” techniques is very high which reduces the performance. Professor Castro stated that although a lot of spectral data is required to evaluate the spectrum accurately, the excess of information is not suitable for a neural network approach. Hence new filters and extraction characteristic techniques must be developed to allow for a reduction in the number of parameters considered while maintaining the accuracy of the information. For this reason, the researchers apply the networks training as an intelligent process of information selection, in this case wavelengths, in order to extract the physical information required and avoid redundancy. Professor Castro underlines that the way in which the information is presented to the neural network is crucial, and at this precise point is where their research centres.

For next stage. these scientists plan to tackle the measurement of spectra in real systems. So far, they have managed to determine theoretical measurements of temperatures with accurate results (about 3 degrees Kelvin more precise at the hottest point of the flame). These computational studies aim to determine the viability of such techniques for the conversion of data.

This temperature monitoring, used in conjunction with feedback systems, could achieve automatic control of the combustion processes. These later systems would receive the information about the combustion reaction, analyse the status of the reaction and feed it back to the system, while possessing the capability to change variables such as the flow of gases and fuel involved, thereby grasping control over the whole process. This would be key for the reduction in the contamination produced and to attain an increased efficiency of the reaction.

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