Uncertainty Informed Anomaly Scores with Deep Learning: Robust Fault Detection with Limited Data

Abstract

Quantifying the predictive uncertainty of a model is an important ingredient in data-driven decision making. Uncertainty quantification has been gaining interest especially for deep learning models, which are often hard to justify or explain. Various techniques for deep learning based uncertainty estimates have been developed primarily for image classification and segmentation, but also for regression and forecasting tasks. Uncertainty quantification for anomaly detection tasks is still rather limited for image data and has not yet been demonstrated for machine fault detection in PHM applications. In this paper we suggest an approach to derive an uncertaintyinformed anomaly score for regression models trained with normal data only. The score is derived using a deep ensemble of probabilistic neural networks for uncertainty quantification. Using an example of wind-turbine fault detection, we demonstrate the superiority of the uncertainty-informed anomaly score over the conventional score. The advantage is particularly clear in an ”out-of-distribution” scenario, in which the model is trained with limited data which does not represent all normal regimes that are observed during model deployment.