Scholarlink

<p>The enhanced self-tuning on-board real-time model (eSTORM), developed by NASA, is a cornerstone in aero-engine health monitoring and has significantly improved engine safety through real-time monitoring. However, its reliance on model accuracy can lead to deviations in health monitoring, often undetectable in idealized simulations. To address these issues, this paper proposes an improved method that combines a time series prediction model with fine-tuning. A long short-term memory (LSTM) model is introduced to capture temporal features and enhance prediction accuracy. In addition, a multi-head attention (MHA) mechanism replaces multiple artificial neural networks (ANNs), overcoming the mismatch issue caused by untrained ANNs with new data. Additionally, an offline layer-wise fine-tuning (LWFT) method is employed to address data incompleteness in previously uncovered flight envelope regions. Results show that, compared to the original eSTORM, the improved method reduces MAXE, MAE, and RMSE for the most challenging temperature parameter by 56 %, 52.8 %, and 53.84 %, respectively. This work not only extends the internationally recognized NASA eSTORM framework but also provides practical solutions for military and civil aero-engine health monitoring 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

Enhanced eSTORM for time series prediction and fine-tuning in aero-engine health monitoring

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Cheng Chen,
  • Ren-jun Chen,
  • Qian-gang Zheng,
  • Hai-bo Zhang

摘要

The enhanced self-tuning on-board real-time model (eSTORM), developed by NASA, is a cornerstone in aero-engine health monitoring and has significantly improved engine safety through real-time monitoring. However, its reliance on model accuracy can lead to deviations in health monitoring, often undetectable in idealized simulations. To address these issues, this paper proposes an improved method that combines a time series prediction model with fine-tuning. A long short-term memory (LSTM) model is introduced to capture temporal features and enhance prediction accuracy. In addition, a multi-head attention (MHA) mechanism replaces multiple artificial neural networks (ANNs), overcoming the mismatch issue caused by untrained ANNs with new data. Additionally, an offline layer-wise fine-tuning (LWFT) method is employed to address data incompleteness in previously uncovered flight envelope regions. Results show that, compared to the original eSTORM, the improved method reduces MAXE, MAE, and RMSE for the most challenging temperature parameter by 56 %, 52.8 %, and 53.84 %, respectively. This work not only extends the internationally recognized NASA eSTORM framework but also provides practical solutions for military and civil aero-engine health monitoring applications.