Modeling Non-monotonic Stochastic Degradation via Filter-Augmented Nonlinear Regression

We propose a Recursive Filter-Augmented Nonlinear Regression (RFNR) model to analyze degradation in PEM fuel cells, where long-term decay is often obscured by “recoverable" performance shocks. Unlike standard models, the RFNR framework decouples the structural nonlinear trend from stochastic, event-driven transients via a recursive component. Model parameters are estimated using nonlinear least squares, with robust inference provided by HAC covariance estimation to address non-stationary noise and autocorrelation. Using experimental data, we show that isolating the degradation signal reduces estimation bias and improves the interpretability of health indicators, providing a scalable tool for real-time reliability monitoring in systems subject to irregular operational fluctuations.