Correlation Effects in Bayesian Neural Networks for Computational Aeroacoustics Ice Detection

In-flight rotor icing presents a serious problem in the operation of rotorcraft in cold climates, as complex ice shapes can significantly degrade the aerodynamic performance and handling characteristics of rotorcraft. Reliable real-time detection of ice formation is thus a critical enabling technology in improving rotorcraft safety. In this paper, we continue our previous work to explore a novel approach towards developing a real-time in-flight ice detection system using computational aeroacoustics and Bayesian neural networks (BNNs). We focus on our use of BNNs constructed from the simulated aeroacoustics dataset to enable rapid predictions of aerodynamic performance indicators together with quantified uncertainty. Specifically, we investigate the effectiveness and tradeoffs among several approximate Bayesian inference techniques for training BNNs: (Gaussian) mean-field variational inference (MFVI), full-covariance variational inference (FCVI), and Stein variational gradient descent (SVGD). We find the correlations in weight uncertainty to be low for this application and with our current BNN setup, although MFVI (which ignores correlations altogether) noticeably under-predicts the variance. SVGD is both computationally fast and captures non-Gaussian and correlation structures, appearing to be a well-suited method for the situation currently considered.