Time-Dependent Survival Neural Network for Remaining Useful Life Prediction

摘要

Remaining useful life (RUL) prediction has been a topic of practical interest in many fields involving preventive intervention, including manufacturing, medicine and healthcare. While most of the conventional approaches suffer from censored failures arising and statistically circumscribed assumptions, few attempts have been made to predict RUL by developing a survival learning machine that explores the underlying relationship between time-varying prognostic variables and failure-free survival probability. This requires a purely data-driven prediction approach, devoid of any a survival model and all statistical assumptions. To this end, we propose a time-dependent survival neural network that additively estimates a latent failure risk and performs multiple binary classifications to generate prognostics of RUL-specific probability. We train the neural network by a new survival learning criterion that minimizes the censoring Kullback-Leibler divergence and guarantees monotonicity of the resulting probability. Experiments on four datasets demonstrate the great promise of our approach in real applications.