Degree Type

Thesis

Date of Award

2019

Degree Name

Master of Science

Department

Mechanical Engineering

Major

Mechanical Engineering

First Advisor

Soumik Sarkar

Abstract

Different cyber-physical systems involving sequential data require accurate frameworks for predicting the state of the system leading to effective monitoring. If the framework is explanatory, the insights provided by the explanations can improve scientific understanding of the system. Detecting the transition to an impending instability is important to initiate effective control in a combustion system. Building robust frameworks is important in this context.

As one of the early applications of characterizing instability in a combustion system using Deep Neural Networks, we train our proposed deep convolutional neural network (CNN) model on sequential image frames extracted from hi-speed flame videos by inducing instability in the system following a particular protocol- varying the acoustic length. We leverage the sound pressure data to define a non-dimensional instability measure used for applying an inexpensive but noisy labeling technique for training. We attempt to detect the onset of instability in a transient dataset where instability is induced by a different protocol. With the continuous variation of the control parameter, we can successfully detect the critical transition to a state of high combustion instability demonstrating the robustness of our proposed detection framework, which is independent of the combustion inducing protocol.

We propose another model which considers the temporal correlations. The model can explain the contribution of each image in the input sequence for generating a single prediction label without compromising on the accuracy. After encoding the images of the input sequence using 2D convolutional neural network and long short term memory recurrent neural network, we capture the global temporal structure by using a temporal attention mechanism. The attention weights highlight the significant image frames that are most relevant for each prediction.

We demonstrate the performance of our models in a problem where explainability and robustness have not been explored like this before. This can lead to better understanding and efficient control.

Copyright Owner

Tryambak Gangopadhyay

Language

en

File Format

application/pdf

File Size

52 pages

Share

COinS