Detecting Failure in Jet Engines Using Uncertainty-based Changepoint Anomaly Detection

Abstract

Anomaly detection in Prognostic and Health Management (PHM) domain by exploiting the information from deep learning uncertainty is presented in this paper. The behaviour of the uncertainty is monitored by cumulative sum (CUSUM) anomaly detection to detect abrupt changes in uncertainty, which translates the transition from healthy state to deterioration state. A probabilistic Long Short-Term Memory (LSTM) neural network is employed to predict the Remaining Useful Life (RUL) sequence distributions of engineered system.  A case study of turbofan engines prognostic is presented to demonstrate the ability of this method. The proposed technique shows excellent result in term of Root Mean Square Error (RMSE) measure between ground truth anomaly and predicted anomaly and good result in scoring metric that evaluates the combination of early and accuracy of anomaly detection compared to the ground truth.