Multivariate Signal Prediction Using Attention Transfer Learning for Early Detection of Thermoacoustic Instabilities

Abstract

The prediction and control of thermoacoustic instability play a crucial role in combustion systems. This paper presents a novel methodology that employs attention transfer learning for the prediction of multivariate thermoacoustic signals, including acoustic pressure and heat release rate. By leveraging attention transfer learning, the proposed model can be generalized to thermoacoustic signal prediction under a variety of conditions with little additional training resources. Experimental data from various conditions in annular combustors are employed to evaluate the predictive performance of the model. The results demonstrate its ability to accurately predict future pressure signals over a wide-ranging time horizon while maintaining consistent dynamics, thus providing valuable information for active control. The proposed model is expected to be an effective step toward a data-driven solution for the early detection of thermoacoustic instability.