WuHZ

Advanced Engineering Informatics • Volume 74 • 2026 • Article 104685

Transferable Snippet Augmentation Network (TSAN)

Engine-specific degradation prediction via trend-grouped fine-tuning, using transferable snippets from similar degradation stages to improve forecasting accuracy under limited data.

Degradation predictionTransfer learningTrend groupingTemporal forecastingEGTM predictionAero-engine PHMTransferable snippet augmentation
Read the paperOpen DOICopy BibTeX

Authors

Haoze Wu; Shisheng Zhong; Minghang Zhao; Yongjian Zhang; Xuyun Fu; Song Fu

Affiliations

  • a. School of Mechatronics Engineering, Harbin Institute of Technology, Harbin 150001, China
  • b. Department of Mechanical Engineering, Harbin Institute of Technology, Weihai 264209, China
  • c. Weihai Key Laboratory of Intelligent Operation and Maintenance, Harbin Institute of Technology, Weihai 264209, China

Core idea

TSAN shifts personalization from one-model-per-engine to group-wise fine-tuning over shared degradation tendencies. Instead of adapting to each engine independently, it reuses transferable degradation snippets from engines in similar life stages, reducing maintenance overhead while improving prediction accuracy and stability.

TSAN at a glance

A two-stage framework: train one base predictor on all engines, then cluster degradation-trend representations and fine-tune specialized predictors for each trend group.

Trend-grouped fine-tuning workflow of TSAN

Fig. 2: TSAN replaces per-engine fine-tuning with trend-grouped fine-tuning based on transferable snippets.

Problem

Per-engine fine-tuning is expensive to maintain and weak when a new engine has only a small amount of historical data.

Key idea

Share short degradation snippets across engines when they exhibit similar deterioration tendencies.

Representation

TSAN learns latent degradation-trend embeddings and assigns samples to trend groups with Gaussian mixture modeling.

Benefit

Fewer specialized models are needed, while prediction accuracy and robustness both improve.

Overview

Abstract

Accurate prediction of key gas path performance parameters is important for aero-engine maintenance planning. A unified model trained on all engines often misses engine-specific behavior, while fine-tuning one model per engine is costly and unreliable for data-scarce engines. TSAN addresses this by moving adaptation from the engine level to the degradation-trend level. It first trains a base model on all engines, then embeds and clusters degradation snippets into trend groups, and finally fine-tunes specialized predictors for each group. Experiments show 36.0%, 35.7%, and 34.1% improvements over baselines in MAE, MRE, and RMSE, along with better stability and fewer maintained models.

Paper details

Title
Engine-specific degradation prediction of aviation engines via transferable snippet augmentation: A trend-grouped fine-tuning perspective
Journal
Advanced Engineering Informatics
Volume
74 (2026)
Article
104685
Online
15 April 2026

Keywords & concepts

TSANAviation engine gas path performanceTemporal forecastingTransfer learningTrend-grouped fine-tuningEGTMDegradation representationGaussian mixture modelTransferable snippet augmentationFleet-scale maintenance

Method

TSAN combines representation learning, trend grouping, and grouped fine-tuning. The base model learns degradation-aware hidden states, then transferable snippets with similar latent trends are grouped to build specialized predictors.

1. Train a shared base model

All engines contribute to one base predictor so the model can learn common operational and degradation dynamics.

2. Learn trend representations

A representation module encodes degradation snippets and aligns them with the predictive hidden state for physically meaningful trend embeddings.

3. Cluster into trend groups

Gaussian mixture modeling groups snippets with similar degradation tendencies, enabling soft assignment and inference on new samples.

4. Fine-tune by trend, not by engine

At inference time, the most similar trend group is selected and its specialized model is used for prediction.

Comparison of EGTM change processes and transferable snippets

Fig. 5: Different engines and different life stages can share locally similar degradation snippets, supporting transfer beyond engine identity.

Perspective shift

From engine-level adaptation to degradation-trend-level adaptation.

Data reuse

Useful segments from other engines become transferable support for a target engine.

Deployment

Model maintenance becomes lighter because one predictor can serve a trend group rather than a single engine.

Results

TSAN consistently outperforms the baseline engine-level fine-tuning strategy and other sequence models, especially on engines with limited historical data.

Model count

53% fewer

Average number of fine-tuned models drops from the engine perspective to the trend-group perspective.

Main gains

36.0% / 35.7% / 34.1%

Relative improvements in MAE, MRE, and RMSE over the best fine-tuning baseline.

Stability

+18.0% / +11.1% / +15.2%

Lower standard deviation across engines on MAE, MRE, and RMSE indicates more stable prediction.

Three representative degradation-trend groups

Fig. 8: Representative early-life, mid-life, and late-life degradation trend groups identified by TSAN.

Benchmark comparison across 7 methods

Fig. 9: TSAN achieves the best overall error profile across engines and metrics among seven compared methods.

Figure gallery

A few representative visuals from the paper that show the grouped fine-tuning idea, transferable snippets, and benchmark results.

TSAN grouped fine-tuning workflow

Trend-grouped fine-tuning

TSAN organizes transferable snippets into degradation-trend groups before selecting a specialized predictor.

Comparison of EGTM change processes

Transferable snippets

Local trajectory similarity exists across engines and across life stages of the same engine, supporting cross-engine transfer.

Visualization of TSAN trend groups

Trend-group visualization

Distinct EGTM ranges correspond to interpretable early-, mid-, and late-life degradation stages.

TSAN benchmark comparison

Benchmark results

TSAN yields the strongest MAE, MRE, and RMSE performance among the evaluated methods.

Citation

If this work is useful, please cite the paper.

@article{wu2026tsan,
  title   = {Engine-specific degradation prediction of aviation engines via transferable snippet augmentation: A trend-grouped fine-tuning perspective},
  author  = {Wu, Haoze and Zhong, Shisheng and Zhao, Minghang and Zhang, Yongjian and Fu, Xuyun and Fu, Song},
  journal = {Advanced Engineering Informatics},
  volume  = {74},
  pages   = {104685},
  year    = {2026},
  doi     = {10.1016/j.aei.2026.104685},
  url     = {https://doi.org/10.1016/j.aei.2026.104685}
}

DOI

https://doi.org/10.1016/j.aei.2026.104685

Author share link

https://authors.elsevier.com/c/1mxnI5FA1kHGCw

Contact

  • Shisheng Zhong: zhongss#hit.edu.cn
  • Yongjian Zhang: zhangyj#hitwh.edu.cn