A Bi-Modular Auto Encoder-Based Unsupervised Degradation Detection Methodology for Remaining Useful Life Prediction

Remaining Useful Lifetime (RUL) is a component of Predictive Maintenance (PdM) that aims at predicting failure times of complex systems. Modern methodologies rely on analyzing the entire time-series of a system to predict its RUL. However, in run-to-failure data where there is no indication of when the system developed degradation, these approaches fall short because there is no discrimination between samples that denote typical operational behavior, and those that denote abnormal behavior. Thus, these methods’ RUL predictions fall short in scenarios where the system may be operating under typical operating conditions. This paper proposes a novel bi-modular degradation detection methodology for RUL prediction. First, we train an Auto Encoder to learn the normal operational behavior of a system to discriminate between samples that may be indicative of normal and abnormal operating behavior. Second, a regression model is trained on samples denoting abnormal behavior to predict RUL. Experiments and analysis on the Auto Encoder’s latent space demonstrate that our methodology learns important properties of the data that lead to better RUL results.