State-space recurrent neural networks for predictive analytics and latent state estimation

Abstract

This paper presents a framework to predict the remaining life (RL) of degrading systems under sensor condition monitoring. By integrating state-space modeling with stochastic recurrent neural networks, our approach efficiently processes condition-monitoring time-series data and models systems’ latent degradation states. We propose a stochastic model that captures dependencies among latent degradation states, sensor outputs, and RL in a causally coherent manner and utilizes stochastic neural networks to navigate the inherent uncertainties of system dynamics. To enhance the interpretability of RL estimation and latent state modeling, we propose interpretable regularization terms. These terms are incorporated into the loss function to optimize both the prediction precision of estimating remaining life and latent states and control the monotonic behavior of their estimates, thereby improving the model’s overall performance and interpretability. Our methodology is validated through numerical experiments and comparison with benchmark models, demonstrating its potential to improve predictive maintenance strategies by effectively estimating the remaining life and monitoring the state of latent degradation over time.