Early detection of thermoacoustic instability in a turbulent hydrogen-methane combustor via deep learning of recurrence plots

We extend our previous work on early detection of thermoacoustic instability by mapping short pressure time series to unbinarized recurrence plots and training a residual neural network (ResNet-18) on these two-dimensional topological images. While our earlier studies demonstrated that recurrence-plot patterns inherently encode incipient instability signatures, here we assess the generalizability of this approach in a turbulent multi-nozzle hydrogen-methane combustor. Four pressure datasets were recorded at distinct equivalence ratios, with the combustor length tuned to induce either a supercritical or subcritical Hopf bifurcation. A lightweight ResNet-18 model is trained on recurrence plots generated from three of the four datasets, and its performance is then evaluated on the fourth (hold-out) dataset. We find that the model gives accurate early warning indicators of thermoacoustic instability for the previously unseen dataset, outperforming conventional methods. These results demonstrate that a single data-driven precursor, once trained via our deep-learning framework, can be deployed for robust early detection of thermoacoustic instability across a range of operating regimes and bifurcation types.