Source-linked AI summary
I-GLIDE: Input Groups for Latent Health Indicators in Degradation Estimation
Lucas Thil, Jesse Read, Rim Kaddah, Guillaume Doquet
TL;DR
Existing health indicators can be sensitive to uncertainty and may not separate subsystem degradation, complicating RUL estimation in complex systems. I-GLIDE constructs subsystem-specific indicators with multi-head autoencoders and uncertainty-aware latent representations. The paper reports state-of-the-art C-MAPSS performance using a simple Random Forest regressor and lower standard deviations in expected accuracy.
Problem
Autoencoder reconstruction-error HIs can be sensitive to noise and epistemic uncertainty and often lack granularity to distinguish subsystem-specific degradation.
Source: §1 · p. 1Method
I-GLIDE uses dedicated encoder-decoder pairs for subsystem-specific latent representations, with uncertainty-aware HI construction and adapted RaPP metrics.
Source: §1 · p. 2, §3.2 · p. 6, §3.3 · p. 7Results
I-GLIDE reports state-of-the-art C-MAPSS performance using a Random Forest regressor and lower standard deviations in expected accuracy than other models.
Source: §5 · p. 11, §6 · p. 14Takeaways & Limitations
By linking HI quality to RUL accuracy and separating subsystem trends, the framework supports targeted fault diagnosis and maintenance planning.
Source: §5 · p. 12Both evaluated datasets model exponential degradation, which may not represent real systems with linear or piecewise trends, complex noise, or heterogeneous failure modes; the architecture also assumes monotonic degradation.
Source: §5 · p. 12
Abstract
from arXiv · showhide
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.
1 Introduction
The paper targets noise-sensitive health indicators that do not isolate subsystem-specific degradation. It proposes I-GLIDE, a multi-head autoencoder framework that integrates uncertainty into health-indicator construction and reports improved RUL prediction.
1 Introduction
- Existing autoencoder reconstruction-error HIs can be sensitive to noise and epistemic uncertainty and lack granularity for separating subsystem degradation.Source: §1 · p. 1
- I-GLIDE uses multi-head encoders and decoders to isolate degradation patterns across subsystems such as fans and high-pressure compressors.Source: §1 · p. 2
- I-GLIDE is reported to achieve state-of-the-art RUL prediction, outperforming established HI methods on NASA C-MAPSS and MILL NASA degradation data.The benchmark includes RaPP and Monte Carlo dropout-based uncertainty estimation [9].Source: §1 · p. 2
- The framework integrates aleatoric and epistemic uncertainty into latent-HI construction to improve RUL estimation while retaining explainability.Source: §1 · p. 2
2 Background and Related Works
The background traces RUL estimation from early usage- and feature-based approaches to learned health indicators, latent-space refinement, and uncertainty-aware modeling. It highlights the need to distinguish localized degradation while preserving system-level coherence.
2.1 RUL Prognostics and Health Indicators
- Complex-system component degradation can cascade into disruptions, costs, and safety risks, motivating timely RUL estimates for proactive maintenance.Source: §2.1 · pp. 2–3
2.2 Evolution of HI Extraction
- HI extraction evolved from handcrafted and physics-based features to autoencoders and latent-space metrics, but interpretability and subsystem-specific trends remain challenges.RaPP latent-space metrics were reported to outperform classical reconstruction errors, while holistic aggregation can obscure subsystem trends.Source: §2.2 · p. 3, Latent-Space HI Refinement · p. 4
2.3 Uncertainty-Aware Subsystem Modeling
- Uncertainty methods distinguish data noise from model ambiguity, but uncertainty quantification is often post-hoc and subsystem-aware HI applications remain underexplored.Source: §2.3 · p. 4
2.4 Monotonicity and Degradation Dynamics
- RUL is modeled as declining with state of health, while subsystem interactions can create non-stationary dynamics that call for localized indicators consistent with system-wide behavior.Source: §2.4 · p. 4
3 Proposed Approach: I-GLIDE
I-GLIDE combines sensor-group encoder-decoder heads in a shared latent space, then constructs group-specific RaPP health indicators with uncertainty estimates. The resulting indicators are evaluated for RUL estimation using a meta-regressor.
3.1 Uncertainty Quantification
- The method quantifies epistemic uncertainty from scalar reconstruction errors across Monte Carlo samples; aleatoric uncertainty is undefined for a vanilla AE with a deterministic latent space.The VAE architecture supports joint uncertainty estimation.Source: §3.1 · p. 5
3.2 Architecture
- Subsystem-specific encoder-decoder heads are trained on healthy data and fused into a shared latent space, with I-GLIDEAE and I-GLIDEVAE using deterministic and Gaussian latent spaces, respectively.The shared latent space is learned through reconstruction loss.Source: §3.2 · pp. 5–6, Figure 1 · p. 6
3.3 Adapting Domain-Specific Latent Space Health Indicators
- For each sensor group, I-GLIDE computes SAP as the Euclidean distance between original and reconstructed activations across layers, and derives NAP from group-specific activation-distance vectors.The NAP calculation uses centered training-set distances and their singular-value decomposition.Source: §3.3 · pp. 6–7
- Group-specific indicators preserve anomalies in individual sensor groups without dilution by nominal signals from other groups, mapping HIs to physical groups for root-cause analysis.Dedicated encoders and decoders calculate SAP and NAP for each group while sharing latent space z.Source: §3.3 · pp. 6–7
3.4 Final set of HIs
- The HI set combines group-level SAP and NAP, shared latent-space RaPP metrics, and group-level aleatoric and epistemic uncertainties.The shared latent-space SAP metric is derived from all data rather than individual groups.Source: §3.4 · p. 7, §3.3 · p. 7
- A meta-regressor F is trained for RUL estimation, and the evaluation also compares the indicators with the González RaPP set of latent-space SAP and NAP metrics.The González indicator set contains ϵNAPLS and ϵSAPLS.Source: §3.4 · p. 7
4 Experiments
The experiments assess whether uncertainty-augmented latent HIs and subsystem-specific grouping improve downstream RUL prediction. They evaluate on C-MAPSS and MILL, comparing HI architectures through random-forest RUL prediction.
4 Experiments
- The experiments first test whether uncertainty-augmented latent HIs improve RUL estimation, then whether subsystem disentanglement further improves it.The stated aim is to understand degradation mechanisms from data with minimal domain knowledge.Source: §4 · pp. 7–8
4.1 Datasets
- C-MAPSS training focuses on RUL ≤80 timesteps and testing follows the established protocol with RUL ≤125; MILL evaluation spans three degradation phases.MILL training treats wear ≤0.20 as healthy; tests include complete, moderate, and severe degradation phases.Source: §4.1 · p. 8
- C-MAPSS and MILL are used to evaluate the predictive capability of the constructed HIs using RMSE for RUL prediction.C-MAPSS contains simulated jet-engine run-to-failure trajectories; MILL records tool-wear progression during milling experiments.Source: §4.1 · p. 8
4.2 Experimental Methodology
- The protocol compares single-timestep HIs by downstream RUL prediction with a random-forest regressor, then benchmarks subsystem-specific I-GLIDE against monolithic AE/VAE counterparts.It bypasses classical HI metrics such as monotonicity and trendability.Source: §4.2 · pp. 8–9
4.3 Results
- HImono reduces average C-MAPSS RMSE by 22.95% for AEs and 28.44% for VAEs compared with HIGonzález.HImono also outperforms HIGonzález across all four C-MAPSS subsets.Source: §4.3 · p. 10, Table 4 · p. 10
- I-GLIDE reduces VAE-based RMSE by 39.03% and standard deviation by 56.07%; its average RMSE on FD004 is 14.19.The comparison covers C-MAPSS subsets; FD004 is identified as the most complex subset.Source: §4.3 · p. 10, Table 6 · p. 11
- I-GLIDEAE HIs achieve the lowest RMSE across all three MILL degradation phases, while VAE gains are subtler.The phases are complete trajectories, moderate degradation, and severe degradation.Source: §4.3 · p. 10, Table 5 · p. 11
5 Discussion
I-GLIDE HIs match or exceed prior state-of-the-art on three of four C-MAPSS benchmarks, while the discussion highlights subsystem-specific degradation patterns and limitations in the evaluation scope.
5 Discussion
- I-GLIDE HIs match or exceed prior state-of-the-art on three of four C-MAPSS benchmarks, and show lower standard deviations than competing models.The comparison covers RUL prediction; the authors argue lower variability supports more robust prediction.Source: §5 · p. 11, Table 7 · p. 12
- For Engine 1, I-GLIDE separates rising HPC degradation from turbine effects while shared latent trends retain system-wide degradation information.Epistemic uncertainty rises for HPC as degradation progresses and remains stable for the turbine until late-cycle HPC interference.Source: §5 · pp. 11–12, Figure 3 · p. 13
- Traditional monotonicity, trendability, and prognosability scores can poorly reflect RUL prediction accuracy, whereas subsystem trends support targeted diagnosis and maintenance planning.The authors link HI quality directly to RUL prediction accuracy as a decision-relevant measure.Source: §5 · p. 12
- Both evaluated datasets model exponential degradation, while real-world noise profiles and heterogeneous failure modes may require tailored adaptations.The authors also note that the architecture assumes monotonous degradation, limiting its ability to model recovery phases.Source: §5 · p. 12
- Subsystem groupings rely on domain heuristics, so poorly defined sensor groups could bias latent representations and undermine HI interpretability.The discussion identifies recovery phases, such as transient improvements, as another boundary for the monotonic-degradation assumption.Source: §5 · pp. 12–13
6 Conclusion and Future Work
I-GLIDE combines subsystem-specific HIs with a shared latent representation, and the conclusion reports strong C-MAPSS performance using a simple Random Forest regressor. The authors propose longer observation windows, expert-annotated latent visualizations, and graph neural networks as future directions.
6 Conclusion and Future Work
- The work establishes a prognostics benchmark for RaPP-derived HIs, reports improved predictive capabilities with uncertainty quantification, and introduces dedicated encoder-decoder pairs for subsystem representations.The subsystem-specific representations share a latent space to retain global system dynamics.Source: §6 · p. 13
- I-GLIDE HIs achieve state-of-the-art C-MAPSS performance using a simple Random Forest regressor, while high-quality HIs paired with simple models outperform deep learning on raw data.The discussion frames this result as evidence that HI quality matters for prognostics success.Source: §6 · p. 14
- Longer observation windows may better resolve slow degradation signatures and transient noise, while uncertainty-specific t-SNE visualizations with expert annotations could connect latent clusters to physical stages.These are proposed directions for aligning HI trajectories with real-world failure timelines and domain knowledge.Source: §6 · p. 14
- Graph neural networks trained on fused HIs are proposed to model causal subsystem interactions and scale prognostics to systems with complex interdependencies.The example interaction is turbine-to-compressor wear propagation.Source: §6 · p. 14
7 Appendix
The appendix illustrates subsystem-specific HI and epistemic-uncertainty trajectories and describes probabilistic methods for uncertainty quantification.
7 Appendix
- For Engine 1, HPC HIs rise while other-subsystem HIs decline; epistemic uncertainty rises near the end for HPC but trends downward for turbine.Uncertainty values are not normalized between groups, so a meta-regressor F can learn how to interpret them.Source: Figure 4 · p. 18
7.1 Uncertainty Quantification
- MC dropout approximates Bayesian inference by sampling posterior weights during inference, while Monte Carlo sampling approximates the predictive integral over weights and latent variables.For regression, the likelihood is Gaussian, with encoder-derived mean and variance.Source: Bayesian Frameworks for Uncertainty · pp. 18–19
- I-GLIDEAE's subsystem-specific groups vary in size, affecting encoder bottlenecks; sparsity is retained notably in the second layer to keep the architecture similar to a vanilla AE.The appendix also lists model architecture and training hyperparameter tables.Source: Table 10 · p. 19, Table 8 · p. 19, Table 9 · p. 19
- VAEs model aleatoric uncertainty with a stochastic latent space and decoder-weight dropout for epistemic uncertainty; vanilla AEs cannot isolate aleatoric uncertainty from deterministic latents.VAE training combines reconstruction loss with KL-divergence regularization.Source: Variational Autoencoders for Aleatoric… · p. 20