Rethinking Remaining Useful Life Prediction with Scarce Time Series Data: Regression Under Indirect Supervision

Jiaxiang Cheng, Yipeng Pang, Guoqiang HuIJCAI 2025 · DOI

TL;DR

This paper tackles the challenge of predicting Remaining Useful Life (RUL) when data is scarce and the target variable (RUL) isn't directly measured but inferred from a dependent variable. They introduce "parameterized static regression," a novel approach that uses individual data points instead of sequences, inherently handling data scarcity and capturing temporal dependencies through a "parametrical rectification" process. This method shows competitive performance, especially in low-data scenarios, offering a more robust solution for real-world industrial prognostics.

What Problem Does This Paper Address?

Predicting the Remaining Useful Life (RUL) of industrial equipment is crucial for proactive maintenance and preventing failures. Traditional supervised learning methods for RUL prediction rely on having direct measurements of the target variable (RUL) for training. However, in many real-world scenarios, RUL is not directly observed. Instead, it's indirectly supervised, meaning the RUL is a function of another observable variable, like time or a degradation indicator.

A major hurdle in these indirect supervision scenarios is data scarcity. Time series data for RUL prediction is often sparse and irregularly sampled. Existing methods that use sequential time series inputs (like RNNs or LSTMs) struggle with this scarcity, often requiring interpolation. Interpolation, however, can introduce significant biases, especially with very limited data. This paper aims to bridge this gap by developing a method that can accurately predict RUL even with scarce, indirectly supervised time series data, without relying on interpolation.

Methods

The core contribution is the parameterized static regression framework. Unlike traditional temporal regression that uses sequences of data points, this method treats each data point independently for initial regression. This "static" approach inherently handles data scarcity because it doesn't require a contiguous sequence of observations.

To incorporate temporal dependencies, which are crucial for RUL prediction, they introduce a parametrical rectification (PR) process. After an initial static regression provides "posteriori estimates," the PR process uses these estimates and the underlying labeling function (used during training) to approximate a parametric function. This effectively "rectifies" the static predictions to capture the temporal evolution of RUL.

The paper also proposes a novel batch training technique specifically for indirect supervision. This technique involves sampling data from time series that share the same labeling function. This targeted sampling helps prevent overfitting and improves training efficiency, especially with large datasets.

Two types of models are explored for the static regression part: an Auto-Encoder Regressor (AER) for dimensionality reduction and feature extraction, and a Residual Gated Units Regressor (ResGUR).

Datasets

The authors evaluate their proposed method on two widely used turbofan engine degradation simulation datasets:

  • CMAPSS (C-MAPSS): This dataset represents regular single-sample time series (RSTS) data.
  • N-CMAPSS: This dataset represents regular multi-sample time series (RMTS) data.

They simulate data scarcity by reducing the percentage of samples available in the training and/or testing sets, allowing them to test their method under various low-data conditions.

Key Insights and Ideas

  • Static Regression as a Foundation for Scarcity: The paper highlights that a static regression approach, which uses individual data points, is fundamentally better suited for handling data scarcity than sequential models that require interpolation.
  • Parametrical Rectification Bridges the Gap: The PR process is a clever way to inject temporal information back into static regression predictions, allowing the model to learn from the underlying RUL degradation patterns without needing dense sequential data.
  • Labeling Function Matters: The choice of parametric labeling function (linear vs. non-linear, like Weibull) significantly impacts RUL prediction performance, underscoring the importance of understanding the underlying degradation process.
  • Targeted Batching for Efficiency: The proposed batch training strategy is crucial for making indirect supervision learning tractable and efficient, especially in large-scale industrial datasets.

Key Results

  • The proposed parameterized static regression method demonstrates competitive performance compared to existing methods, particularly in scenarios with high data scarcity.
  • The PR process significantly enhances the prediction performance of static regression models, showing notable improvements even under 90% data scarcity.
  • Experiments show that the non-linear labeling function (Weibull) leads to significantly better performance than a linear function.
  • The proposed batch training technique is effective, with optimal performance achieved by sampling a small fraction (around 2%) of the total available data for training.
  • The method shows superior performance with both train and test sets experiencing scarcity, outperforming prior art in certain scenarios.

Limitations and Future Work

  • Data Scarcity Impact on Inference: While the paper addresses data scarcity during training, it notes that in practical scenarios, scarcity can occur during both training and inference. The paper's approach is designed to handle this.
  • Uncertainty Quantification: The paper mentions that prior work lacked uncertainty quantification, and while their method shows improved performance, further exploration into quantifying prediction uncertainty could be beneficial.
  • Generalizability to Different Scarcity Patterns: The paper focuses on simulated scarcity. Future work could explore how the method performs with naturally occurring, irregular data sampling patterns in real-world industrial settings.
  • Exploration of More Complex Labeling Functions: While non-linear functions were explored, investigating even more sophisticated or adaptive labeling functions could further improve RUL prediction accuracy.

Summary updated

BibTeX
@inproceedings{cheng2025rethinking,
  author = {Jiaxiang Cheng and Yipeng Pang and Guoqiang Hu},
  title = {{Rethinking Remaining Useful Life Prediction with Scarce Time Series Data: Regression Under Indirect Supervision}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/1018},
}

Loading related content...

Figure