Early detection of thermoacoustic instability in a solid rocket motor: A generative adversarial network approach with limited data

Thermoacoustic instability (TAI) poses a significant challenge to the development of solid rocket motor (SRM) and detecting early warning signals (EWS) for TAI is crucial. Deep learning-based approaches hold promise for reliable EWS. However, existing EWS for TAI often lack timeliness and universality, while the shortage of training datasets increasingly hinders their practicality in industrial combustors. To address these challenges, this study introduced the Wasserstein generative adversarial network (WGAN) algorithm, combined with the random convolutional kernel transform (RCKT). By leveraging only a single set of real SRM data, WGAN＋ RCKT enables the synthesis of data with three distinct typical dynamic states, and the synthetic data exhibits convincing fidelity and diversity. By utilizing synthetic datasets, a prediction model is constructed based on the multiclass classification deep neural network. This model achieves convincing timeliness in producing EWS of a full-scaled SRM by probabilistically monitoring the stable state, vicinity of the Hopf bifurcation point, and unstable state. Furthermore, the prediction model is applicable beyond SRM systems, as evidenced by its successful implementation in the Rijke tube. This confirms its satisfactory universality, highlighting its ability to extract essential dynamic properties shared by different TAI systems. Our findings affirm the feasibility and tremendous potential of utilizing WGAN+RCKT in scenarios with limited real data, presenting a promising approach for detecting advanced and universal EWS for TAI in industrial combustors.