Reconstruction of the flame nonlinear response using deep learning algorithms

This paper demonstrates the ability of the neural network trained on frequency-sweeping signals with different amplitudes to reconstruct the flame nonlinear response. The neural network architecture consists of a decreasing sequence increasing dimension model and a sequence model; the latter one uses the long short-term memory (LSTM) and encoder of Transformer. Results show that the neural network trained using the proposed sweeping method with limited training data can reconstruct realistic signals over the envisaged range of frequencies and amplitudes. The nonlinear flame responses obtained by the neural network are further embedded into the closed-loop thermoacoustic feedback to quantify the reconstruction performance of sequence signals. It is demonstrated that the neural network can accurately capture the evolution of the limit cycle. This paper has also compared the effect of different types and sizes of datasets on trained neural networks model; the results show that models trained with our proposed datasets perform better. For small-size datasets, LSTM performs significantly better than the encoder of Transformer. The encoder of Transformer is more suitable for large-size datasets.