Prognostics with Variational Auto-encoder by Generative Adversarial Learning
Prognostics predicts future performance progression
and remaining useful life (RUL) of in-service systems
based on historical/contemporary data. One of the challenges for prognostics is the development of methods which are capable of handling real-world uncertainties that typically lead to inaccurate predictions. To alleviate the impacts of uncertainties and to achieve accurate degradation trajectory and RUL predictions, a novel sequence-to-sequence predictive model is proposed based on a variational auto-encoder (VAE) that is trained with Generative Adversarial Networks (GAN). A Long Short-Term Memory (LSTM) network and Gaussian mixture model are utilized as building blocks so that the model is capable of providing probabilistic predictions. Correlative and monotonic metrics are applied to identify sensitive features in the degradation progress, in order to reduce the uncertainty induced by raw data. Then, the extracted features are concatenated with one-hot health
state indicators as training data for the model to learn the end-of-life (EoL) without the need for prior knowledge of failure thresholds. The performance of the proposed model is validated by health monitoring data collected from real-world aero-engines, wind turbines, and lithium-ion batteries. The results demonstrate that significant performance improvement can be achieved in long-term degradation progress and RUL prediction tasks.