A Multi-Granularity Supervised Contrastive Framework for Remaining Useful Life Prediction of Aero-Engines

Zixuan He, Ziqian Kong, Zhengyu Chen, Yuling Zhan, Zijun Que, Zhengguo Xu

TL;DR

This paper introduces a novel "Multi-Granularity Supervised Contrastive" (MGSC) framework to improve Remaining Useful Life (RUL) prediction for aero-engines. By aligning feature representations of similar RUL states in a structured way, MGSC significantly boosts prediction accuracy, overcoming limitations of traditional regression-based methods.

What Problem Does This Paper Address?

Predicting the Remaining Useful Life (RUL) of aero-engines is crucial for safety and maintenance. Current data-driven approaches primarily rely on regression, often using Mean Squared Error (MSE) as the sole loss function. While effective for capturing the magnitude of error, this approach overlooks the underlying structure of the learned feature space. Specifically, it doesn't explicitly encourage features from similar RUL states to be close together, leading to less discriminative representations. Furthermore, existing supervised contrastive learning methods, which excel at structuring feature spaces, face challenges with RUL prediction due to the need for very large batch sizes to balance positive and negative samples, a constraint often impractical due to hardware limitations. This paper aims to bridge this gap by developing a contrastive learning framework that effectively structures the feature space for RUL prediction while mitigating batch size and sample imbalance issues.

Methods

The core innovation is the Multi-Granularity Supervised Contrastive (MGSC) strategy. This approach combines two levels of supervised contrastive learning:

  • Coarse-grained contrast: Uses "Health Status" (HS) labels, which represent broader operational phases (e.g., normal, degradation). This allows for a larger classification scale, enabling preliminary feature space regularization even with smaller batch sizes and helping to balance positive and negative samples.
  • Fine-grained contrast: Uses the actual RUL labels within each HS class. This further refines the feature space by aligning embeddings of samples with very similar RULs, ensuring ordered clusters.

To integrate these contrastive objectives with the RUL regression task, a multi-phase training strategy is proposed. This involves sequentially training with coarse-grained contrast, then fine-grained contrast, and finally performing RUL regression. The network architecture includes an encoder (CNN-LSTM), a projector (used during contrastive training), and regression layers.

Datasets

The experiments were conducted on the CMAPSS dataset, a widely used benchmark for aero-engine RUL prediction. This dataset comprises four subsets (FD001, FD002, FD003, FD004), simulating full life-cycle degradation under various operating conditions and fault models.

Key Insights and Ideas

  • Feature space structure matters for RUL: Simply minimizing regression error isn't enough. Structuring the feature space so that similar RUL states are clustered together significantly improves prediction accuracy.
  • Multi-granularity is key: Combining broad (HS) and specific (RUL) labels in contrastive learning provides a powerful way to regularize the feature space. The coarse-grained approach handles large-scale alignment and batch size issues, while the fine-grained approach refines it for precise RUL prediction.
  • Addressing practical constraints: The MGSC framework is designed to overcome the limitations of large batch sizes and sample imbalance inherent in applying supervised contrastive learning to RUL prediction.
  • Intuitive visualization: The t-SNE visualizations of the encoder's feature space clearly demonstrate how MGSC creates well-structured, ordered manifolds compared to traditional regression, making the degradation trend more apparent.

Key Results

The proposed MGSC framework, when applied to baseline models like CNN-LSTM, CNN, and LSTM, consistently improved RUL prediction accuracy across all four CMAPSS datasets. For instance, the CNN-LSTM model with MGSC achieved significant reductions in RMSE and improvements in the scoring function compared to the baseline without MGSC. The paper highlights that the MGSC strategy effectively creates an "ideal manifold structure" in the feature space, leading to better alignment of embeddings for similar RUL labels.

Limitations and Future Work

The authors acknowledge that FD004, with its complex operating conditions and failure modes, presented a more challenging case, with one metric showing a slight decrease. They suggest that this might be due to the negative contrastive effect in such complex scenarios.

Future work includes exploring other multi-granularity contrast learning methods and contrast methods specifically tailored for time-series data. The authors also plan to apply the MGSC framework to more advanced RUL predictive tasks.

Summary updated

Abstract

Accurate remaining useful life (RUL) predictions are critical to the safe operation of aero-engines. Currently, the RUL prediction task is mainly a regression paradigm with only mean square error as the loss function and lacks research on feature space structure, the latter of which has shown excellent performance in a large number of studies. This paper develops a multi-granularity supervised contrastive (MGSC) framework from plain intuition that samples with the same RUL label should be aligned in the feature space, and address the problems of too large minibatch size and unbalanced samples in the implementation. The RUL prediction with MGSC is implemented on using the proposed multi-phase training strategy. This paper also demonstrates a simple and scalable basic network structure and validates the proposed MGSC strategy on the CMPASS dataset using a convolutional long short-term memory network as a baseline, which effectively improves the accuracy of RUL prediction.
BibTeX
@misc{he2024multi,
  author = {Zixuan He and Ziqian Kong and Zhengyu Chen and Yuling Zhan and Zijun Que and Zhengguo Xu},
  title = {{A Multi-Granularity Supervised Contrastive Framework for Remaining Useful Life Prediction of Aero-Engines}},
  year = {2024},
  eprint = {2411.00461},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.00461},
}
Figure