Scholarlink

<p>Aero-engine health management requires accurate health state (HS) assessment and remaining useful life (RUL) prediction. Single-task deep learning often limits performance, while dual-task learning faces task conflict and negative transfer. To overcome these challenges, we propose a multi-sensor joint prediction framework with a gating-enhanced progressive layered extraction (PLE) model that balances task-specific features. A physics-informed uncertainty loss enables dynamic parameter optimization. To address high-dimensional input complexity, we design a bidirectional feature extraction architecture combining multi-headed self-attention (MHSA), bidirectional long short-term memory (BiLSTM), and recursive hidden information transfer (HIT). Health indicators (HI) are constructed via weighted multi-sensor, fusion based on Pearson's correlation, and degradation stages are validated using dynamic time warping (DTW). Experiments on the C-MAPSS and N-CMAPSS datasets demonstrate that the proposed framework achieves state-of-the-art performance in joint HS and RUL prediction, exhibiting significant advantages particularly under complex operating conditions. Compared with single-task models, the joint framework not only improves predictive accuracy but also reduces computation time, demonstrating its feasibility for real-time prognostics health management (PHM) applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Joint prediction of aero-engine health based on multi-sensor feature fusion and expert gated augmentation network

  • 期刊论文

  • 发表日期：

    2026年1月18日

查看全文
  • Yufeng Zhang,
  • Hongfei Zhan

摘要

Aero-engine health management requires accurate health state (HS) assessment and remaining useful life (RUL) prediction. Single-task deep learning often limits performance, while dual-task learning faces task conflict and negative transfer. To overcome these challenges, we propose a multi-sensor joint prediction framework with a gating-enhanced progressive layered extraction (PLE) model that balances task-specific features. A physics-informed uncertainty loss enables dynamic parameter optimization. To address high-dimensional input complexity, we design a bidirectional feature extraction architecture combining multi-headed self-attention (MHSA), bidirectional long short-term memory (BiLSTM), and recursive hidden information transfer (HIT). Health indicators (HI) are constructed via weighted multi-sensor, fusion based on Pearson's correlation, and degradation stages are validated using dynamic time warping (DTW). Experiments on the C-MAPSS and N-CMAPSS datasets demonstrate that the proposed framework achieves state-of-the-art performance in joint HS and RUL prediction, exhibiting significant advantages particularly under complex operating conditions. Compared with single-task models, the joint framework not only improves predictive accuracy but also reduces computation time, demonstrating its feasibility for real-time prognostics health management (PHM) applications.