I-GLIDE: Input Groups for Latent Health Indicators in Degradation Estimation

TL;DR

The paper tackles remaining-useful-life prognostics in multi-sensor systems by crafting interpretable health indicators (HIs) and embedding uncertainty quantification into their construction. It introduces I-GLIDE, a subsystem-group, multi-head architecture that learns domain-specific latent representations and adapts RaPP-based HI metrics per sensor group, employing MC dropout and VAEs to separate aleatoric and epistemic uncertainty. Through extensive experiments on CMAPSS and MILL datasets, I-GLIDE achieves state-of-the-art RUL performance with a simple RF regressor and provides clearer diagnostics of degradation pathways. The work bridges anomaly detection and prognostics, offering a principled, uncertainty-aware approach to degradation modeling and pointing to future directions in causal and physics-informed HI designs.

Abstract

Accurate remaining useful life (RUL) prediction hinges on the quality of health indicators (HIs), yet existing methods often fail to disentangle complex degradation mechanisms in multi-sensor systems or quantify uncertainty in HI reliability. This paper introduces a novel framework for HI construction, advancing three key contributions. First, we adapt Reconstruction along Projected Pathways (RaPP) as a health indicator (HI) for RUL prediction for the first time, showing that it outperforms traditional reconstruction error metrics. Second, we show that augmenting RaPP-derived HIs with aleatoric and epistemic uncertainty quantification (UQ) via Monte Carlo dropout and probabilistic latent spaces- significantly improves RUL-prediction robustness. Third, and most critically, we propose indicator groups, a paradigm that isolates sensor subsets to model system-specific degradations, giving rise to our novel method, I-GLIDE which enables interpretable, mechanism-specific diagnostics. Evaluated on data sourced from aerospace and manufacturing systems, our approach achieves marked improvements in accuracy and generalizability compared to state-of-the-art HI methods while providing actionable insights into system failure pathways. This work bridges the gap between anomaly detection and prognostics, offering a principled framework for uncertainty-aware degradation modeling in complex systems.

Figures (5)

• Figure 1: I-GLIDE Architecture Framework -- A: Subsystem-specific encoder-decoder heads learn distinct latent representations, fused into a shared latent space $z$ via reconstruction loss (trained on healthy data). B: HIs are extracted using RaPP metrics González-Muñiz_Díaz_Cuadrado_García-Pérez_2022 and UQ Kingma_Welling_2013 over full trajectories. C: Aggregated HIs are used to predict RUL, trained via a Random Forest (RF) regressor $\mathcal{F}$.
• Figure 2: AE HI trajectories for Engine 1 for the monolithic architecture. We can observe that the HIs model a degradation, but cannot distinguish sub-system components. We only show the SAP metric for the encoder HIs because NAP shows extreme values.
• Figure 3: $\text{I-GLIDE}_\text{AE}$ HI trajectories for Engine 1, comparing degradation effects on HPC and Turbine. Latent encoder HIs (a) show rising HPC degradation and reduced Turbine HIs due to cross-component effects. System-wide latent $z$ HIs trends are in (b). Epistemic uncertainty (c) rises sharply for HPC as degradation progresses, remaining stable for the Turbine until late-cycle HPC interference. UQ confirms causal cross-component effects without confusing intrinsic health states.
• Figure 4: $\text{I-GLIDE}_\text{VAE}$ HI trajectories for Engine 1, comparing degradation effects of all sub-system. Latent encoder SAP HIs (a) show rising HPC degradation and a decreased of HIs from other sub-systems. In b) we see that both HIs seem to follow a system-wide latent $z$ HIs trends are in (c). Epistemic uncertainty (d) We show the VAE uncertainties for two sub-systems: HPC and turbine. We observe that the turbine value has an overall decreasing trend whereas HPC increases near the end. The uncertainties values are not normalized between groups which can be a hindrance, that is why a meta regressor $\mathcal{F}$ can learn how to interpret them automatically..
• Figure 5: Turbofan Engine Schematic – showing the different sub-systems that are represented in the C-MAPSS dataset Saxena_Goebel_Simon_Eklund_2008.