Reduced Dimension Analysis on Combustion Oscillation in a Model Rocket Combustor Using a Deep Neural Network

Abstract

The similarity between POD (Proper Orthogonal Analysis) and NN (Neural Network) is explained and an example of NN to perform reduced dimension analysis on a combustion oscillation problem is presented. The dimension reduction procedure by Snapshot-POD is shown to be expressed by a three layer AE (Auto-Encoder). Based on this, a DAE (Deep Auto-Encoder), consisting of a four layer encoder and four layer decoder is tested. The encoder has layers of 128-32-8-2 neurons and the decoder has the ones of 2-8-32-128 neurons in its layers. The DAE reduces the dimension of the input data into two, which is the number of the encoder output variables. As a reference, a POD that takes the first and the second mode neglecting higher modes are employed to reduce the dimension into two. A URANS simulated time varying temperature, heat release, and pressure distributions of CVRC (Continuously Variable Resonance Combustor) are analyzed by the POD and DAE. As a result, the 2D data from DAE and POD agreed well. It was confirmed that the dimension reduction performance and the resulting amount of information was almost consistent. By analyzing mode maps, the ability to identify the modes for pumping up the oscillation is demonstrated.