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

AI summary

Overview

• Research area: Machine learning for prognostics and health management (PHM) — specifically the construction of Health Indicators (HIs) from multi-sensor data for Remaining Useful Life (RUL) prediction in complex engineering systems.
• Technical level: Intermediate. The paper assumes familiarity with autoencoders (AEs), variational autoencoders (VAEs), latent spaces, and uncertainty quantification (UQ), but its core claims are about architecture design and benchmarking rather than new mathematics.
• Scope: The paper proposes I-GLIDE, a multi-head autoencoder framework that builds subsystem-specific latent Health Indicators (with uncertainty estimates) and benchmarks them on the NASA C-MAPSS turbofan dataset and the MILL NASA tool-wear dataset.

What This Paper Is About

RUL prediction depends on the quality of Health Indicators, but existing autoencoder-based HIs are sensitive to noise and epistemic uncertainty, and they tend to blend all sensor signals into one global representation that cannot separate degradation in different subsystems (e.g., fan versus high-pressure compressor). The paper asks whether (a) RaPP (Reconstruction along Projected Pathways) metrics can serve as HIs, (b) adding aleatoric and epistemic uncertainty quantification makes those HIs more robust, and (c) splitting sensors into functionally coherent groups produces HIs that are both more accurate and more interpretable.

Key Contributions

1. Systematic analysis of existing AE-derived HIs. The authors identify and characterize limitations of prior HIs, notably vulnerability to noise and an inability to isolate subsystem-level degradation.
2. First adaptation of RaPP as a health indicator for RUL prediction, with a claim that RaPP-derived HIs outperform traditional reconstruction-error metrics.
3. Uncertainty-aware HI construction, integrating aleatoric uncertainty (via a probabilistic latent space) and epistemic uncertainty (via Monte Carlo dropout) as core HI components rather than post-hoc refinements, improving RUL prediction robustness.
4. The I-GLIDE framework, a multi-head autoencoder in which each encoder-decoder pair targets a distinct sensor group. The paper presents two instantiations, I-GLIDE_AE and I-GLIDE_VAE, which produce subsystem-specific HIs while sharing a common latent space to preserve system-wide coherence.

Main Findings

• RaPP plus UQ beats the prior RaPP baseline. On C-MAPSS, AE-based HI_mono reduced RMSE by 22.95% on average versus HI_González, with reported gains such as FD002 (15.71 vs 22.91) and FD003 (8.07 vs 12.03). VAE-based HI_mono achieved a 28.44% average RMSE improvement. The same direction of improvement was observed on MILL.
• Subsystem grouping improves robustness. I-GLIDE_AE produced a 39.96% reduction in the standard deviation of RMSE across C-MAPSS subsets from AE-based HIs. For VAEs, I-GLIDE_VAE reduced RMSE by 39.03% and standard deviation by 56.07%, resolving the instability seen in monolithic VAEs on FD002/FD003.
• Best single-subset results on C-MAPSS (Table 4). I-GLIDE_AE scored 9.47 (FD001), 16.18 (FD002), 8.29 (FD003), 12.32 (FD004), average 11.57. I-GLIDE_VAE scored 12.33, 16.76, 8.5, 11.4, average 12.25.
• Averages over 10 runs (Table 6). I-GLIDE_AE averaged 14.82 ± 2.19 RMSE versus 16.27 ± 3.65 for HI_mono and 20.63 ± 4.40 for HI_González. I-GLIDE_VAE averaged 14.87 ± 1.75.
• New state-of-the-art claim on FD004. The paper reports that I-GLIDE's average results on FD004, the most complex C-MAPSS subset, reach an RMSE of 14.19 using only a Random Forest regressor.
• Results on MILL (Table 5). I-GLIDE_AE-driven HIs achieved the lowest RMSE in every wear scenario: 13.64 (wear 0.0–0.70), 14.37 (0.20–0.70), 16.17 (0.50–0.70), compared with 16.14 / 16.47 / 16.25 for AE HI_mono and 23.78 / 24.34 / 22.33 for AE HI_González. VAE gains on MILL were subtler.
• Comparison against deep learning baselines (Table 7). I-GLIDE with a Random Forest reports 9.47 / 16.18 / 8.29 / 12.32 (AE) and 12.33 / 16.76 / 8.5 / 11.4 (VAE), compared with e.g. Probabilistic RUL CNN at 12.42 / 13.72 / 12.16 / 15.95 and CNN-LSTM at 11.17 / – / 9.99 / –. The authors state their HIs match or exceed prior state of the art on three of four C-MAPSS benchmarks.
• Independence from model complexity. A simple Random Forest regressor was used for all HI-to-RUL mapping, which the authors interpret as evidence that HI quality, not model complexity, drives prognostics performance.
• Interpretability through causal cross-component effects. For Engine 1 (FD001), the monolithic HI showed weak latent-space sensitivity to subsystem dynamics, while I-GLIDE displayed a clear upward trend in the HPC encoder HI and an abrupt shift in the turbine HI as degradation propagated. Epistemic uncertainty rose sharply for HPC while remaining stable for the turbine until late-cycle HPC interference.
• Critique of traditional HI metrics. The authors argue monotonicity, trendability and prognosability often produce misleading scores that correlate poorly with actual RUL prediction, and instead judge HI quality directly by downstream RUL accuracy.

Methodology in Plain English

The researchers treat an autoencoder as a compressor: it squeezes sensor readings into a small latent representation and then tries to reconstruct the original input. Because it is trained only on healthy data, it reconstructs healthy behavior well and degraded behavior poorly, so the size of the discrepancy becomes a Health Indicator. Rather than measuring discrepancy in the raw input space, the authors use RaPP, which compares the internal activations (and latent representations) of the encoder for the original input and its reconstruction. Two RaPP variants are computed per sensor group: SAP, the Euclidean distance between original and reconstructed activations across layers, and NAP, a normalized version using the singular value decomposition of the centered distance matrix.

Rather than feeding all sensors into one encoder, I-GLIDE gives each sensor group its own encoder-decoder pair. On C-MAPSS the groups are Fan, LPC, HPC, Core, Pressure Turbine, and Other. The per-group encoders feed a shared latent space so that the whole system stays coherent while individual subsystems can be examined separately. The same setup is instantiated twice: once with a deterministic latent (I-GLIDE_AE) and once with a Gaussian latent (I-GLIDE_VAE).

Uncertainty is computed by running the network multiple times. With Monte Carlo dropout, repeated passes let the authors separate epistemic uncertainty (variation when decoder weights are fixed) from aleatoric uncertainty (variation when the latent is fixed). Because a vanilla autoencoder has a deterministic latent, aleatoric uncertainty cannot be isolated there, which is offered as a reason to use VAEs. The final HI set per group combines SAP, NAP, latent-space SAP and NAP, and the two uncertainty terms; a monolithic variant without sensor grouping and the earlier González RaPP set are used for comparison.

To evaluate, the authors deliberately skip classical HI quality scores and instead train a Random Forest regressor to map HIs to RUL, measuring error in RMSE. A single timestep is used per HI. On C-MAPSS the training samples are restricted to RUL ≤ 80 and testing uses R_early = 125 to follow prior protocols. On MILL, healthy training is defined as wear ≤ 0.20, with evaluation on complete trajectories and on moderate (wear > 0.20) and severe (wear > 0.50) degradation.

Why This Matters

Impact on research. The paper provides what it describes as the first prognostics benchmark for HIs generated via RaPP methods, and it argues for judging HI quality by downstream RUL accuracy rather than by traditional monotonicity or trendability scores. It also repositions uncertainty quantification from an optional add-on to a core part of HI construction, and it demonstrates that subsystem-aware architectures can beat deep learning baselines when paired with a simple regressor.

Real-world applications:

• Aerospace engine maintenance, where C-MAPSS-derived HIs could help schedule service before turbofan subsystem failure.
• Manufacturing tool wear monitoring, using acoustic, vibration and current signals during milling to track wear and plan tool changes.
• Condition-based maintenance programs in general, where interpretable subsystem-level HIs support root-cause analysis rather than opaque alarms.
• Decision support for safety- or cost-critical assets, where uncertainty estimates indicate when a health assessment should not be trusted.

Industry relevance. The work is co-affiliated with Safran Tech, an aerospace manufacturer, and was supported by the French government under the "France 2030" program within the JNI3 project at the SystemX Technological Research Institute. Its central practical claim — that high-quality HIs paired with simple models outperform deep learning on raw data — is directly relevant to industrial deployments where model simplicity, explainability and robustness matter more than architectural novelty.

Future Directions

• Temporal modeling. Extending observation windows could better resolve slow degradation signatures and transient noise, aligning HI trajectories with real-world failure timelines.
• Latent-space interpretation. Coupling uncertainty-specific t-SNE visualizations with expert annotations could map latent clusters onto physical degradation stages, bridging data-driven results with domain knowledge.
• Causal subsystem modeling. Using architectures such as graph neural networks on fused HIs could disentangle degradation propagation, for example turbine-to-compressor wear, and scale prognostics to systems with complex interdependencies.
• Addressing stated limitations. The authors note that both C-MAPSS and MILL model exponential degradation, that the architecture assumes strictly monotonous degradation (so it cannot represent recovery phases), and that sensor groupings rely on domain heuristics that could bias the latent representations when poorly defined. They also want to formalize methods for interpreting causal relationships between HIs and identifying noise patterns in degradation signals.

Target Audience

This paper is most useful to researchers and practitioners in prognostics and health management, industrial AI, and reliability engineering who work on RUL prediction from multi-sensor time series. It will particularly interest readers designing autoencoder-based health indicators, those evaluating uncertainty quantification in predictive maintenance pipelines, and engineers in aerospace or manufacturing who need interpretable, subsystem-level diagnostics rather than black-box RUL outputs. Readers without background in autoencoders, latent variable models or uncertainty decomposition will need supplementary reading, since the architecture and UQ derivations assume that vocabulary.

Authors’ 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.