Interpretable Deep Learning for Monitoring Combustion Instability

Abstract

Transitions from stable to unstable states occurring in dynamical systems can be sudden leading to catastrophic failure and huge revenue loss. For detecting these transitions during operation, it is of utmost importance to develop an accurate data-driven framework that is robust enough to classify stable and unstable scenarios. In this paper, we propose deep learning frameworks that show remarkable accuracy in the classification task of combustion instability on carefully designed diverse training and test sets. We train our model with data from a laboratory-scale combustion system showing stable and unstable states. The dataset is multimodal with correlated data of hi-speed video and acoustic signals. We develop a labeling mechanism for sequences by implementing Kullback-Leibler Divergence on the time-series data. We develop deep learning frameworks using 3D Convolutional Neural Network and Long Short Term Memory network for this classification task. To go beyond the accuracy and to gain insights into the predictions, we incorporate attention mechanism across the time-steps. This aids in understanding the time-periods which contribute significantly to the prediction outcome. We validate the insights from a domain knowledge perspective. By exploring inside the accurate black-box models, this framework can be used for the development of better detection frameworks in different dynamical systems.