Uncertainty-Aware Remaining Useful Life Prediction for Electromechanical Relays with Bayesian Lstm Networks

Abstract

Predictive maintenance of electromechanical relays remains challenging: deterministic remaining-useful-life (RUL) estimates are unreliable for relays that fail spontaneously or under resistive loads, and give the operator no indication of when a prediction can be trusted. This paper addresses this gap with a Bayesian long short-term memory (LSTM) network that predicts not a single RUL value but a full predictive distribution. Monte-Carlo dropout is combined with an output head that estimates an input-dependent (heteroscedastic) variance, so that the model produces RUL prediction intervals, separates aleatoric from epistemic uncertainty, and derives a near-failure probability as an operator-facing early-warning signal. On the Phoenix Contact Relay (PCR) run-to-failure dataset the approach attains a relative-RUL RMSE of 0.17 at an empirical 95% interval coverage, and the near-failure probability rises before failure even for spontaneously degrading relays. A per-load analysis shows accurate, well-calibrated predictions on the inductive loads and attributes the remaining error and slight over-confidence to the abruptly failing resistive loads.