Combustion mode time-series prediction in rotating detonation combustor based on predictor of encoder transformer

Achieving real-time prediction of combustion modes in rotating detonation combustor (RDC) is of great significance for the development of high-performance rotating detonation propulsion system. A novel neural network based on Transformer architecture is proposed in this study, namely, RDC–PET (Rotating Detonation Combustor–Predictor of Encoder Transformer), which predicts the combustion mode in the RDC. Experimental datasets were built with high-speed images of RDC under different operating conditions, where RP-3 aviation kerosene was used as the fuel and a binary oxygen–nitrogen mixture (40% oxygen by volume) was used as the oxidizer. The influence of different hyperparameters, including the number of encoder layers, input sequence length, and attention heads, on the model performance is investigated. Different embedding methods were applied to compare their effectiveness in predicting combustion modes of the RDC. The findings reveal that Physical Embedding, which utilizes an attention mechanism to capture the correlations between physical variables, enhances predictive accuracy. The prediction results for single-wave, counter-two-wave, and unstable combustion modes are analyzed in depth. Results show that RDC–PET accurately predicts the propagation behavior of detonation waves with different quantities under typical combustion modes of the RDC and captures the evolutionary processes such as wave collision, splitting, collapse, and merging. In summary, RDC–PET demonstrates excellent performance in predicting mode switch in RDC, thereby providing early warning of combustion instability in RDC and consequently allowing time for active control in engineering applications. RDC–PET has potential to serve as a general prediction model.