Providence - a Deep Learning Framework for Time-to-Event Prediction

Prognostic Health Monitoring (PHM) and Condition Based Maintenance (CBM) are fields with robust bodies of research which have, in recent years, shown promise to be transformed by deep learning. A catalyst for the recent surge in time-to-event research has been the tandem advancement and widespread deployment of sensing technologies alongside maturation of deep learning techniques for time series. Consequently, deep learning for time-to-event modeling has gained in both popularity and effectiveness. Multivariate time series and sensor data have become a cornerstone of failure prognostics. Prior work has shown that traditional methods struggle to capture the complex nature of the underlying dynamical system, the relationships between sensor signals, the degradation of the system over time, multiple failure modes, and the often very rare event of failure. Recurrent Neural Networks (RNNs) have been used widely in remaining useful life (RUL) research, effectively modeling the temporal nature of the problem and complex relationships between sensor signals. Still, capturing the degradation of a dynamic system requires a domain-specific solution, as provided in the Weibull Time-to-Event RNN (WTTE-RNN). Building upon these advances, we introduce Providence, a neural network framework for generating Weibull time-to-event (TTE) estimates by solving a sequence-to-sequence learning problem. By learning a Weibull distribution, we translate the problem from one of forecasting to one of multivariate distribution fitting. We perform this task on a multi-device dataset and produce Weibull predictions per-device. By learning per-device, we avoid the issue of source distribution variance across multiple devices. With the Weibull distribution, we can predict a TTE for an arbitrary event; herein, we focus on device failures. We benchmark performance of the Providence framework with the publically available NASA Turbofan and Backblaze hard disk drive datasets. Additional experiments find that Transformers with temporal attention are able to learn distributions across an entire fleet. Finally, we demonstrate the efficacy of alternative approach to the RUL problem and champion it for its high interpretability.