Bayesian Recurrent Neural Networks for Monitoring Rotorcraft Icing from Aeroacoustics Time-Series Data

Motivated by the need for reliable on-board and real-time rotor blade icing detection with quantified uncertainty, we demonstrate the use of Bayesian recurrent neural networks, particularly the long short-term memory (LSTM), to predict time-series aerodynamic performance metrics from measurements of aeroacoustic time-series. The neural network models are trained using simulation data generated from physics-based models. To help improve robustness against measurement error in real life, we augment the data targets with artificial noise corruption. The training is done by minimizing the negative evidence lower bound (ELBO), taking a mean-field variational inference approach that finds independent Gaussian distribution most closely approximates the true Bayesian posterior. Promising performance is observed in the test set, where Bayesian LSTM achieved lower prediction errors compared to a Monte Carlo dropout version of the LSTM model.