Detection of Precursors of Thermoacoustic Instability in a Swirled Combustor Using Chaotic Analysis and Deep Learning Models

Abstract

This paper investigates the role of chaotic analysis and deep learning models in combustion instability predictions. To detect the precursors of impending thermoacoustic instability (TAI) in a swirled combustor with various fuel injection strategies, a data-driven framework is proposed in this study. Based on chaotic analysis, a recurrence matrix derived from combustion system is used in deep learning models, which are able to detect precursors of TAI. More specifically, the ResNet-18 network model is trained to predict the proximity of unstable operation conditions when the combustion system is still stable. The proposed framework achieved state-of-the-art 91.06% accuracy in prediction performance. The framework has potential for practical applications to avoid an unstable operation domain in active combustion control systems and, thus, can offer on-line information on the margin of the combustion instability.