🤖 AI

An Unsupervised Framework for Dynamic Health Indicator Construction and Its Application in Rolling Bearing Prognostics

This paper proposes an unsupervised framework that constructs a dynamic health indicator for rolling bearing prognostics by combining a skip-connection-based autoencoder for expert-free feature extraction with an HI-generating module that explicitly models temporal dependencies to capture degradation trends.

Original authors: Tongda Sun, Chen Yin, Huailiang Zheng, Yining Dong

Published 2026-05-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tongda Sun, Chen Yin, Huailiang Zheng, Yining Dong

Original paper licensed under CC BY 4.0 (http://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 trying to predict when an old car engine is going to break down. You have a lot of noisy data coming from sensors—vibrations, sounds, and heat readings. The goal is to create a single "Health Indicator" (HI), like a dashboard warning light, that tells you exactly how sick the engine is and when it will fail.

The problem with most current methods is that they are like a mechanic who looks at the engine only at this exact second. They might see a vibration and say, "It's bad," but they forget that the engine was getting worse slowly over the last hour. They miss the story of the breakdown.

This paper proposes a new, smarter way to build that "Health Indicator" using a computer system that learns on its own (unsupervised learning). Here is how it works, broken down into simple steps:

1. The "Noise-Canceling" Feature Finder (The Skip-connection Autoencoder)

The Problem: Raw sensor data is messy. It's like trying to hear a whisper in a crowded stadium. Traditional methods often throw away important details while trying to clean up the noise.
The Solution: The authors built a special digital filter called a Skip-connection-based Autoencoder.

The Analogy: Imagine you are trying to summarize a long, complex movie. A normal summary might just list the main characters. But this new system is like a director who keeps a "backstage pass" (the skip connection). It looks at the big picture and the tiny details simultaneously. It connects the early scenes of the movie directly to the final summary.
The Result: It strips away the background noise and keeps the most important "clues" about the machine's health, creating a clean, low-dimensional map of the machine's condition without needing a human expert to tell it what to look for.

2. The "Time-Traveling" Health Score (The Dynamic HI)

The Problem: Most health indicators are static snapshots. They tell you the health score now, but they don't remember the score from yesterday. If the score jumps up and down randomly, it's hard to predict when the machine will actually die.
The Solution: The authors added a "Time-Traveling" module.

The Analogy: Think of a normal health score as a single photo of a runner. A Dynamic Health Indicator is like a video of the runner. It doesn't just show where the runner is; it knows how fast they were running five seconds ago and predicts where they will be next.
How it works: The system has a built-in "prediction engine" inside it. While it creates the health score, it constantly asks itself: "Based on the last few seconds of data, what should the score be right now?" If the actual score doesn't match the prediction, it adjusts the score to make it fit the pattern better. This forces the health indicator to follow a smooth, logical path that respects the machine's history.

3. The "One-Way Street" Rule (Monotonicity)

The Problem: Machines don't get better as they age; they only get worse. However, noisy data can sometimes make a health score look like it's improving for a split second, which confuses the prediction.
The Solution: The system is given a strict rule: The score can only go up (or down, depending on how you measure it), never back.

The Analogy: Imagine a sandglass. Once the sand falls, it can't go back up. The system forces the health indicator to behave like that sandglass, ensuring it represents a steady, irreversible decline. This makes the prediction much more reliable.

What Did They Find?

The researchers tested this new system on real rolling bearings (the kind used in factories and wind turbines) using two different datasets. They compared their method against many other popular techniques.

Better Storytelling: Their "Dynamic Health Indicator" was much better at showing the smooth, steady decline of the machine's health compared to others, which were often jagged or erratic.
Better Crystal Ball: Because the health score was so smooth and respected the timeline, the system could predict when the machine would fail much more accurately. In some tests, it predicted the failure time with over 99% accuracy, whereas other methods were much more off.
No Expert Needed: The whole system learned from the raw data automatically. You don't need a human engineer to manually pick out which vibrations matter; the computer figures it out.

The Bottom Line

This paper introduces a framework that treats machine health not as a series of isolated snapshots, but as a continuous, evolving story. By using a special "memory-keeping" filter and forcing the health score to follow a logical timeline, they created a tool that is significantly better at predicting when a machine will break down, allowing for smarter maintenance before a disaster happens.

Technical Summary: An Unsupervised Framework for Dynamic Health Indicator Construction and Its Application in Rolling Bearing Prognostics

1. Problem Statement

Rolling bearings are critical components in industrial machinery, yet their failure under harsh operating conditions leads to unexpected breakdowns and economic losses. Predictive maintenance relies heavily on constructing a Health Indicator (HI) to represent degradation, predict remaining useful life (RUL), and estimate failure times.

Existing HI construction methods face two primary limitations:

Reliance on Expert Knowledge: Traditional Physics-based HIs (PHIs) and Virtual HIs (VHIs) often depend on manual feature engineering (e.g., RMS, kurtosis, entropy) or the fusion of specific statistical features. This approach requires domain expertise, lacks generalization, and may introduce human bias.
Neglect of HI-Level Temporal Dependence: While some deep learning methods extract features from raw signals, they often treat HI construction as a static task, deriving the indicator solely from current observations. They fail to explicitly model the HI-level temporal dependence—the correlation between the current HI state and its past values. Since degradation is a progressive process, ignoring this dynamic relationship limits the HI's ability to represent degradation trends accurately and hinders reliable prognostics.

2. Methodology

The authors propose a novel unsupervised framework for constructing a Dynamic Health Indicator (Dynamic HI). The framework operates in two sequential stages, eliminating the need for manual feature engineering while explicitly modeling temporal dynamics.

Stage 1: Degradation Feature Learning via Skip-connection-based Autoencoder (SkipAE)

Objective: To map raw vibration signals into a low-dimensional, representative Degradation Feature Space (DFS) without expert intervention.
Architecture: A SkipAE is employed as the backbone. Unlike standard autoencoders, SkipAE utilizes skip connections to link encoder layers directly to corresponding decoder layers.
Mechanism: This architecture preserves multi-level and multi-scale information during the encoding and reconstruction process, mitigating information loss and the vanishing gradient problem. It extracts essential degradation features ( $z_k$ ) directly from raw signals ( $x_k$ ).

Stage 2: Dynamic HI Generation with Inner HI-Prediction Block

Objective: To construct the HI ( $y_k$ ) from the learned features while guaranteeing HI-level temporal dependence.
Architecture: A lightweight HI-generating module maps features $z_k$ to a 1D HI series. Crucially, this module integrates an inner HI-prediction block.
Mechanism:
- Temporal Modeling: The prediction block forecasts the current HI state ( $\hat{y}_k$ ) based on a window of past states ( $\{y_i\}_{k-L}^{k-1}$ ). The input is decomposed into trend and seasonal components to enhance prediction accuracy.
- Loss Function: The framework optimizes a combined loss function ( $L_{combined}$ $L_{co mbin e d}$ ) comprising:
  1. Reconstruction Loss ( $L_{rec}$ ): Ensures the HI retains information from the original degradation features.
  2. Prediction Loss ( $L_{pre}$ ): Minimizes the error between the actual HI and the predicted HI, explicitly enforcing HI-level temporal dependence.
  3. Monotonic Loss ( $L_{mono}$ ): Enforces a monotonic increasing trend to align with the physical nature of degradation.
Training: The system is trained unsupervised using stochastic gradient descent. The HI series used for the prediction target is updated iteratively during training.

3. Key Contributions

Skip-connection-based Feature Learning: The proposal of a SkipAE module to automatically learn comprehensive, multi-scale degradation features from raw signals, removing the dependency on manual feature engineering.
Dynamic HI Construction with Temporal Guarantee: The introduction of a novel HI-generating module that integrates an inner prediction block. This explicitly models and guarantees HI-level temporal dependence, distinguishing the proposed Dynamic HI from traditional static HIs.
Unsupervised Framework for Prognostics: The development of a complete unsupervised framework that ensures the constructed HI possesses high representability, monotonicity, and predictability, verified through extensive experiments on two benchmark datasets.

4. Experimental Results

The framework was evaluated on two datasets: the IEEE PHM 2012 dataset (PRONOSTIA platform) and a self-collected HIT-B dataset (accelerated life test). The study included three tasks: same-machine same-condition, same-machine different-condition, and different-machine different-condition.

HI Construction Performance: The proposed method outperformed comparison methods (including RMS, P-Entropy, ISOMAP, KPCA, CEEMDAN, VAE, and SSAE) across all metrics: Monotonicity (Mon), Trendability (Tred), Robustness (Rob), and Hybrid Scale (HS). Notably, it achieved the highest scores in Mon and Tred, indicating superior ability to reflect the degradation trend and resist noise.
Prognostic Performance: When used for degradation trend prediction, the Dynamic HI demonstrated significantly lower Root Mean Square Error (RMSE) and higher Predictability Index (Pred) compared to other methods. The average Pred value exceeded 0.9, whereas other methods ranged lower.
Ablation Studies: Removing the SkipAE, the inner HI-prediction block, or the HI-generating module resulted in performance degradation across all metrics. This confirmed that the skip connections are vital for feature representativeness, and the inner prediction block is essential for capturing temporal dynamics and improving prognostic accuracy.

5. Significance and Claims

The paper claims that the proposed unsupervised framework addresses the critical gap in existing HI construction methods by explicitly modeling the HI-level temporal dependence. By ensuring that the current HI state is explicitly correlated with its past values, the constructed Dynamic HI better captures the inherent dynamic nature of the degradation process.

The authors assert that this approach leads to:

Superior Representability: The HI more accurately reflects the true degradation trend of rolling bearings.
Enhanced Predictability: The explicit modeling of temporal dependence significantly improves the accuracy of future degradation predictions and RUL estimation.
Robustness: The method performs well across different operating conditions and machine types without requiring re-engineering of features.

The study concludes that the proposed Dynamic HI facilitates more reliable predictive maintenance decisions by providing a health indicator that is not only monotonic and robust but also inherently predictable due to its dynamic construction.