💻 computer science

Continuous Degradation Representation Learning for Remaining Useful Life Prediction

The paper proposes CoDeR, a continuous degradation representation learning framework that integrates a Dual-axis Coupled Transformer Encoder and a Time-Aware Contrastive Learning module to overcome the limitations of fragmented representations in existing deep learning models, thereby achieving more accurate Remaining Useful Life predictions for aero-engines.

Original authors: Wei Li, Ruitao Ning, Shilin He, Bo Li, Zhidong Zhang

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Wei Li, Ruitao Ning, Shilin He, Bo Li, Zhidong Zhang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a race car engine slowly lose its power as it races toward the finish line. The goal of "Remaining Useful Life" (RUL) prediction is to guess exactly how many laps are left before the engine sputters and stops. For a long time, computer scientists have tried to teach machines to make this guess by looking at the engine's sensors—like a dashboard full of gauges measuring heat, pressure, and speed.

However, the authors of this paper, Wei Li and his team from North University of China, noticed a problem with how most computers were doing this. They found that existing methods often treated the engine's life like a broken puzzle, looking at snapshots of data that were disconnected from one another. It's like trying to understand a movie by looking at random, scattered frames; you might see a car, then a cloud, then a tire, but you miss the smooth story of the race. These "fragmented" views made it hard for the computer to see the continuous, steady slide of the engine getting older.

To fix this, the team built a new system they call CoDeR (Continuous Degradation Representation Learning). Think of CoDeR as a super-smart detective that doesn't just look at the clues; it understands the story behind them.

The Detective's Two Superpowers

CoDeR uses two main tricks to solve the mystery of the engine's life:

The Dual-Axis Coupled Transformer Encoder (DCTE): Imagine the engine's sensors as a giant grid of dancers. Some dancers move in time (temporal), showing how the engine changes second by second. Others move in space (spatial), showing how different parts of the engine (like the fan or the fuel pump) rely on each other. Old methods tried to watch the time dancers and the space dancers separately, then mashed the results together. CoDeR, however, uses a "Dual-axis" approach. It watches the dancers together, understanding how a change in the fuel pump right now affects the pressure a second later. This allows it to capture the full, coupled dance of the engine's degradation.
Time-Aware Contrastive Learning (TACL): This is the secret sauce. Imagine you are arranging photos of the engine on a wall. Old methods might just group "good" engines together and "bad" engines together, but they might accidentally mix up the ones that are in the middle of the race. CoDeR uses a special rule: it knows that a photo taken at lap 100 should look very similar to the photo at lap 101, and very different from the photo at lap 500. It forces the computer to arrange the data so that the "story" flows smoothly. It's like ensuring that if you walk down a hallway of engine photos, the images change gradually, not in jerky jumps. This keeps the "degradation trajectory" (the path of the engine getting older) continuous and easy to follow.

The Results: A Smoother Ride

The team tested CoDeR on famous engine datasets called C-MAPSS and N-CMAPSS. These datasets contain data from hundreds of simulated engines running under different conditions.

On the C-MAPSS test: The team ran 10 separate trials to make sure the results weren't just luck. They found that CoDeR consistently beat the older methods. For example, on the hardest test set (FD004), the old method had an error (RMSE) of 16.71, while CoDeR lowered it to 15.68. In the "Score" metric, which punishes bad guesses heavily, CoDeR dropped the score from 1291 down to 1023. Lower is better!
On the N-CMAPSS test: This dataset is even trickier because it mimics real, messy flight conditions with changing speeds and altitudes. Here, CoDeR achieved an error (RMSE) of 3.41, beating the next best method (CPI-GAN) which had 3.82.

The authors also looked at the "feature maps" (visualizations of how the computer sees the data). They found that CoDeR created a much smoother, more organized path for the engine's life, whereas the older methods left the data scattered and messy. This suggests that by keeping the degradation story continuous, the computer can make much more reliable predictions.

What CoDeR is NOT

It is important to note what this paper does not claim. The authors explicitly state that they are not solving the problem of "real-time" deployment yet. They admit that their system, which uses a complex Transformer architecture, is computationally expensive. It's like having a brilliant detective who takes a long time to solve the case; they haven't figured out how to make the detective work fast enough for a live, split-second decision in a real factory just yet. They also didn't test this on actual physical engines in a hangar; the results are based on simulations and benchmark datasets.

The Bottom Line

The paper suggests that by teaching computers to see the engine's life as a continuous, flowing story rather than a collection of broken snapshots, we can predict when an engine will fail more accurately. While the system still needs to be made faster for real-world use, the authors' experiments show that this "continuous" approach is a significant step forward in keeping our skies safe.

Technical Summary: Continuous Degradation Representation Learning for Remaining Useful Life Prediction

1. Problem Statement

Predicting the Remaining Useful Life (RUL) of aero-engines is critical for aviation safety and maintenance scheduling. While deep learning has become the dominant paradigm for RUL prediction, existing approaches face two primary limitations:

Fragmented Representations: Current deep learning models often learn discontinuous or fragmented degradation representations from multivariate sensor data. They struggle to accurately model the progressive, continuous nature of degradation trajectories, often relying on separate temporal-spatial feature extraction followed by fusion, which may miss fine-grained interactions.
Disordered Feature Spaces: Even with powerful backbone networks, samples from different degradation stages may exhibit weak ordering or overlapping distributions in the feature space. This lack of structural alignment with the continuous degradation process hinders accurate RUL estimation.
Limitations of Existing Contrastive Learning: While contrastive learning has been used to improve feature separability, most methods rely on discrete positive-negative pair construction. This imposes uniform separation among samples and fails to capture the continuous, progressive nature of degradation trajectories, where temporal proximity should correlate with feature similarity.

2. Methodology: The CoDeR Framework

The authors propose CoDeR (Continuous Degradation Representation learning), a framework designed to learn temporally coherent degradation representations. The architecture consists of three main stages: data preparation, joint learning, and RUL prediction.

2.1 Dual-axis Coupled Transformer Encoder (DCTE)

To address the need for modeling complex dependencies in multivariate sensor data, the authors introduce the DCTE. Unlike standard models that process temporal and spatial dimensions separately, DCTE jointly models them:

Parallel Positional Encoding: The model injects positional embeddings along both the temporal axis ( $P_t$ ) and the sensor (spatial) axis ( $P_s$ ) to preserve sequential order in both dimensions.
Intra-axis Modeling: Symmetric Transformer Encoders ($Enct$ and $Encs$) process temporal dynamics and sensor-wise dependencies independently within their respective dimensions.
Bidirectional Cross-axis Attention: To capture complementarity, the model employs bidirectional cross-attention. It projects spatial representations into the temporal domain and vice versa using learnable linear mappings ( $P_{s \to t}$ and $P_{t \to s}$ ), allowing dynamic information exchange at each layer.
Feature Fusion: The final joint representation is formed by fusing the temporal and spatial outputs, which are then flattened for downstream regression.

2.2 Time-Aware Contrastive Learning (TACL)

To enforce continuity in the learned feature space, the authors propose the TACL module, which integrates temporal priors into contrastive learning:

Sampling Strategy:
- Positive Samples: Selected from the local temporal neighborhood of an anchor sample (excluding the anchor itself).
- Negative Samples: Selected from the remaining sequence, divided into consecutive sub-intervals to ensure coverage of distant degradation stages.
Time-Distance-Weighted Objective: Unlike standard contrastive losses, TACL introduces adaptive weights based on temporal distance.
- Positive Weights ( $\alpha_i$ ): Decrease as the temporal distance between the anchor and positive sample increases, encouraging stronger attraction for temporally proximal samples.
- Negative Weights ( $\beta_j$ ): Increase with temporal distance, strengthening the repulsion of samples from distant degradation stages.
Loss Function: The contrastive loss is defined as a weighted sum where the attraction and repulsion forces vary continuously with temporal distance, ensuring smooth transitions in the embedding space.

2.3 Joint Optimization

The framework employs a multi-objective loss function that combines:

Regression Loss ( $L_{MSE}$ ): Mean Squared Error for RUL prediction.
Continuity Constraint ( $L_{TACL}$ ): The time-aware contrastive loss.
The total loss is $L = L_{MSE} + \mu \cdot \frac{1}{|B|} \sum L^{(i)}_{TACL}$ , where $\mu$ balances prediction accuracy and temporal consistency.

3. Key Contributions

The paper claims the following contributions:

CoDeR Framework: A novel framework for aero-engine RUL prediction that explicitly aims to improve the continuity of degradation representations learned from multivariate sensor data.
DCTE Architecture: A Dual-axis Coupled Transformer Encoder that models dependencies across both temporal and feature dimensions simultaneously, enabling more informative feature extraction.
TACL Module: A Time-Aware Contrastive Learning module that incorporates temporal progression priors into contrastive learning, encouraging degradation representations with improved temporal continuity in the feature space.
Empirical Validation: Demonstrations on C-MAPSS and N-CMAPSS benchmarks showing competitive prediction accuracy and more ordered, continuous degradation representations compared to existing methods.

4. Experimental Results

The authors evaluated CoDeR on the C-MAPSS (FD001–FD004) and N-CMAPSS (DS02) benchmark datasets.

Performance on C-MAPSS:
- CoDeR consistently outperformed the DCTE baseline (without TACL) and various state-of-the-art methods (including CNNs, LSTMs, and other Transformer variants like MACPTD, MSMT-FDA, and DAST) across all subsets.
- On the challenging multi-condition subsets (FD002 and FD004), CoDeR achieved the lowest RMSE and Score values among compared methods.
- Ablation Studies: Removing the cross-attention mechanism, spatial branch, or temporal branch in DCTE degraded performance, with the temporal branch having the largest impact. Replacing TACL with standard InfoNCE loss resulted in significant performance drops and less structured feature distributions.
Generalization on N-CMAPSS:
- CoDeR achieved the best overall RMSE (3.41) compared to other methods on the N-CMAPSS dataset, which features more realistic, high-frequency, and dynamically varying operating conditions.
- The model maintained stable prediction behavior across different engine units and flight phases.
Feature Visualization:
- t-SNE visualizations revealed that CoDeR produces smoother, more compact, and continuous feature trajectories compared to the scattered and overlapping distributions of baseline models.
- Attention map analysis showed that DCTE successfully learned meaningful spatiotemporal feature selection patterns, focusing on pressure-related variables, fan/core speeds, and fuel-flow ratios.

5. Significance and Limitations

Significance:
The paper argues that accurate RUL prediction requires not just informative feature extraction but also a well-structured representation space that reflects the continuous physical process of degradation. By coupling temporal and spatial modeling (DCTE) and enforcing temporal continuity via weighted contrastive learning (TACL), CoDeR addresses the limitations of discrete modeling and uniform separation in existing methods. The results suggest that structuring the latent feature space to mirror the continuous degradation process leads to more reliable and robust RUL estimates, particularly under complex operating conditions.

Limitations:
The authors modestly acknowledge several limitations:

Computational Cost: The Transformer-based architecture and the TACL module introduce additional optimization overhead and computational costs, which may challenge real-time deployment.
Future Work: The authors state that future work will focus on improving computational efficiency to facilitate practical industrial deployment.

The paper does not claim to solve all RUL prediction challenges or propose applications beyond the aero-engine domain, nor does it speculate on future implications beyond the stated goal of industrial deployment efficiency.