Forecasting thermoacoustic instabilities in liquid propellant rocket engines using multimodal Bayesian deep learning

We present a method that combines multiple sensory modalities in rocket thrust chamber to predict impending thermoacoustic instabilities with uncertainties. This is accomplished by training an autoregressive Bayesian neural network model forecasts the future amplitude of dynamic pressure time series, using sensor measurements (injector pressure/ temperature measurements, static pressure, high-frequency OH* chemiluminescence measurements) and flow rate control signals as input. The validated experimental data from representative cryogenic research chamber. nature our algorithms allows us work dataset whose size restricted expense each run, without making overconfident extrapolations. find networks are able accurately forecast evolution anticipate instability events on unseen runs 500 milliseconds advance. compare predictive accuracy models different combinations inputs. signal particularly informative. also use technique integrated gradients interpret influence inputs prediction. negative log-likelihood points test indicates prediction uncertainties well-characterized simulating failure event results dramatic increase epistemic component uncertainty, would be expected when encounters unfamiliar, out-of-distribution