Variational Auto-Encoder for combustion analysis
For the analysis on the complex combustion field, we apply DNN (Deep Neural Network) technologies to obtain the abstract feature together with its temporal variation. Homeomorphlism mapping from observable space to latent sapce is available with VAE (Variational Auto-Encoder). This way visualizes the trajectory of the dyanamics on a reduced dimension space. Increasing the number of NN hidden layer achieves higher level of abstraction and we expect more effective dimension reduction to analyse the comprex combustion field.
Fig. 1 Schematic of DAE that extract essential feature and their locus on feature space from time-varying combustion field data.
[1] Tanabe, M., Reduced Dimension Analysis on Combustion Oscillation in a Model Rocket Combustor Using a Deep Neural Network, Trans. JSASS, Aerospace Technology Japan, vol. ists-31, 2018.
[2] Motohashi, K., et. al., Analysis method on flame with acoustic forcing using neural network, 6th Pacific-Asia Conference on Mechanical Engineering, Manila, 2017.
This project is supported by JSPS KAKENHI Grant