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

Description

I-GLIDE is a project aimed at reproducing the C-MAPSS experiment using our architecture and the RaPP metrics (reconstruction along projected pathways). The dataset used is C-MAPSS, which is a popular dataset for predictive maintenance and prognostics.

Key Features

• Health Indicators (HI) Extraction: We extract health indicators from our Autoencoder (AE)-based model.
• RUL Prediction: The extracted HIs are fed into a Random Forest (RF) to predict the Remaining Useful Life (RUL).
• Architecture Choices: You can choose between different architectures, including AE, VAE, I-GLIDE AE, and I-GLIDE VAE.

Fig. 1. I-GLIDE Architecture Framework

• A: Subsystem-specific encoder-decoder heads learn distinct latent representations, which are fused into a shared latent space $z$ via reconstruction loss (trained on healthy data).
• B: Health Indicators (HIs) are extracted using RaPP metrics [González-Muñiz et al., 2022] and Uncertainty Quantification (UQ) [Kingma & Welling, 2013] over full trajectories.
• C: Aggregated HIs are used to predict Remaining Useful Life (RUL), trained via a Random Forest (RF) regressor $( \mathcal{F} )$.

Dataset

The C-MAPSS dataset is used for this experiment. For each trajectory, the last point is represented with rmse_last.

How to Run

1. Navigate to the /cmapss directory.
2. Choose the Jupyter notebook corresponding to the architecture you want to use:
   • AE
   • VAE
   • I-GLIDE AE
   • I-GLIDE VAE
3. Follow the instructions in the notebook to run the experiment.

Results

Below is a comparison of different sets of HIs extracted from various architectures to predict the RUL on the C-MAPSS test dataset using a Random Forest for $\mathcal{F}$. The best models are highlighted. RMSE was used as a metric over the RUL estimation for the last trajectories point available in C-MAPSS test set.

Table: Model Performance Across C-MAPSS Subsets, Average performance and standard deviation

Model, HI Set	FD001	FD002	FD003	FD004	Avg.	Improvement
AE, HI_González [1]	19.00 ±4.78	25.69 ±4.19	18.38 ±6.18	19.46 ±2.46	20.63 ±4.40	--
AE, HI_mono	13.14 ±2.50	20.35 ±3.46	13.87 ±5.07	17.73 ±3.56	16.27 ±3.65	+21.13% / +17.15%
I-GLIDE_AE, HI_groups	12.11 ±2.72	22.01 ±2.88	10.23 ±1.85	14.92 ±1.31	14.82 ±2.19	+8.94% / +39.96%
VAE, HI_González [1]	34.13 ±3.71	31.05 ±1.89	27.25 ±2.58	25.23 ±2.03	29.42 ±2.55	--
VAE, HI_mono	27.19 ±5.97	22.81 ±2.86	24.64 ±5.26	22.89 ±1.82	24.38 ±3.98	+17.10% / -55.83%
I-GLIDE_VAE, HI_groups	15.32 ±2.08	18.83 ±1.51	11.12 ±2.29	14.19 ±1.11	14.87 ±1.75	+39.03% / +56.07%

Note:

• Underline (represented here as bold): Outperforms methods without HIs.
• Improvement: Average improvement over the previous row.
• Bold: Best results per subset

Table: Comparison of HI Sets, best models taken

HI Extractor	HI Set for $\mathcal{F}(.)$	FD001	FD002	FD003	FD004
AE	$\text{HI}_\text{González}$ [1]	11.43	22.91	12.03	16.78
AE	$\text{HI}_\text{mono}$	10.53	15.71	8.07	14.35
VAE}	$\text{HI}_\text{González}$ [1]	27.56	28.62	24.36	22.33
VAE	$\text{HI}_\text{mono}$	18.77	19.44	15.59	19.81
$\text{I-GLIDE}_\text{AE}$	$\text{HI}_\text{groups}$	9.47	16.18	8.29	12.32
$\text{I-GLIDE}_\text{VAE}$	$\text{HI}_\text{groups}$	12.33	16.76	8.5	11.4

Table: Comparison of I-GLIDE Method for HI Extraction

Model	FD001	FD002	FD003	FD004
MLP [2]	37.56	80.03	37.39	77.37
CNN [2]	18.45	30.29	19.82	29.16
CNN-LSTM [3]	11.17	-	9.99	-
MS-DCNN [4]	11.44	19.35	11.67	22.22
VAE + RNN [5]	11.44	24.12	14.88	26.54
MLE(4X)+CCF [6]	11.57	18.84	11.83	20.78
RVE [5]	13.42	14.92	12.51	16.37
Probabilistic RUL CNN [7]	12.42	13.72	12.16	15.95
I-GLIDE_AE + RF (ours, best)	9.47	16.18	8.29	12.32
I-GLIDE_VAE + RF (ours, best)	12.33	16.76	8.5	11.4

Note: The best results for each subset are highlighted in bold.

Citation:

@inproceedings{thil2025glide,
  title={I-GLIDE: Input Groups for Latent Health Indicators in Degradation Estimation},
  author={Thil, Lucas and Read, Jesse and Kaddah, Rim and Doquet, Guillaume},
  booktitle={Joint European Conference on Machine Learning and Knowledge Discovery in Databases},
  pages={395--411},
  year={2025},
  organization={Springer}
}