Weibull Variational Autoencoder for Remaining Useful Life Prediction

Abstract

Remaining useful life (RUL) prediction is a critical technology for preventing unexpected failures and reducing maintenance costs in modern industrial systems. However, traditional model‐based approaches are limited by the need for explicit mathematical modeling of degradation mechanisms, while data‐driven methods often require large‐scale datasets and lack interpretability. To address these challenges, this study proposes a Weibull Variational Autoencoder (WVAE). The WVAE is designed to learn probabilistic characteristics grounded in the Weibull distribution from failure history data and predict failure times accordingly. A composite loss function combining mean squared error (MSE) with negative log‐likelihood is employed to jointly ensure predictive accuracy and distributional fidelity, while Monte Carlo Dropout–based inference is used to quantify uncertainty and provide confidence intervals. Simulated datasets incorporating three types of failures, including infant mortality, random, and wear‐out failures, as well as multiple levels of noise, were constructed to reflect diverse system characteristics and industrial conditions. Evaluation results demonstrate that, compared with several benchmarking models, the WVAE produces predictions statistically consistent with historical failure distributions while maintaining stable forecasting performance. Furthermore, by directly estimating the parameters of the Weibull distribution, the WVAE provides statistically interpretable predictions that can be trusted by domain experts. The model is further validated on the Backblaze hard drive field‐reliability dataset, confirming its applicability to real‐world failure data with diverse distributional characteristics.