Characterization of soot properties in coflow laminar axisymmetric diffusion flames using image processing and machine learning
Rodríguez Barreda, Alonso Antonio
MetadatosMostrar el registro completo del ítem
Traditional non-invasive optical techniques for estimating soot volume fraction and soot temperature in coflow laminar axisymmetric diffusion flames require solving ill-posed inverse problems from convoluted signals. Applying these techniques using inversion and regularization makes them highly susceptible to experimental noise and the choice of adjustable regularization parameters, which can prevent obtaining consistent and accurate estimations. This thesis proposes replacing these techniques with a machine learning model that implements the inversion step in the estimation of these soot properties. We develop a framework to generate physically-grounded simulated signals of reference soot fields and their corresponding convoluted projections in the camera plane, thus enabling a supervised learning approach to train our machine learning model. We focus on line-of-sight attenuation measurements for the characterization of the soot volume fraction field and Broadband Emission measurements for the characterization of the soot temperature field, as these techniques represent a promising approach to low-cost characterization of their respective soot property in flames. Using our framework, we generate datasets for each of the two soot properties to train a model based on U-Net, a fully convolutional neural network previously used in similar inversion problems. We then compare the performance of our models over the synthetic dataset versus traditional techniques and perform a final validation over data obtained during experimental campaigns. These results show that our machine learning models outperform traditional inversion techniques when processing noisy measurements, especially in areas of interest for soot formation, such as the flame center and its path of maximum soot concentration along the flame. The resilience to noise shown by machine learning models makes them attractive for implementing low-cost techniques to characterize soot properties in flames using experimental equipment of different quality, representing a promising research avenue.