3-D soot temperature and volume fraction reconstruction of afterburner flame via deep learning algorithms

Abstract

3-D soot temperature and volume fraction reconstruction of afterburner flame has been based on spectrally resolved measurement methods and inverse problem theory. So far, the Convolutional Neural Networks (CNN) has been widely used in solving inverse problems to realize real-time reconstruction. However, the convolution and pooling operation can cause erroneous accumulation when dealing with high precision and high mesh density 3-D reconstruction applications. To address this problem, the autoencoder long-short-term-memory (LSTM) neural network and the synthetic dataset considering the interior and exterior parameters of camera are adopted to realize rapid, accurate, and simultaneous reconstructions of the soot temperature and volume fraction of afterburner flame. This method, named StfNet-3D by the authors, is investigated regarding its reconstruction accuracy, noise immunity, and computational costs by comparing that with the CNN, LSTM, and traditional 3-D reconstruction method (3D TVR) through simulations and real-time experiments.