Towards low-cost soot pyrometry in laminar flames using broadband emission measurements and Artificial Neural Networks

This paper presents a low-cost approach based on Artificial Neural Networks (ANNs) for retrieving fields of soot temperature in laminar flames from broadband soot emission signals captured with a color camera. Using a framework to generate numerical simulations of soot temperature fields in laminar flames and their corresponding projections to the camera plane, we generated large datasets for designing and training different ANN models to infer the relationships between the reference temperature fields and the emission measurements captured with the camera. Experiments over simulated datasets show that properly trained ANNs outperform traditional onion-peeling deconvolution techniques used for retrieving soot temperature from emission signals, delivering accurate temperature estimations that are close to the ones obtained with the more sophisticated modulated absorption/emission techniques that require a much more complex experimental setup. We also show that ANNs trained with simulated data can provide consistent and accurate temperature fields from emission measurements taken in real experimental campaigns using both commercial and industrial-grade color cameras.