Scholarlink

<p>This study proposes an explainable machine learning framework based on SHAP (SHapley Additive exPlanations) for predicting the remaining useful life (RUL) of turbine systems. Using the NASA C-MAPSS dataset, we evaluate eight machine learning algorithms and find that LightGBM achieves the best performance, with a root mean square error (RMSE) of 36.68, mean absolute error (MAE) of 26.53, and coefficient of determination (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation>) of 0.713. SHAP analysis identifies sensor measurements 11 and 15, along with operational cycles (time_in_cycles), as the most influential features for RUL prediction. The analysis further reveals critical threshold values (approximately 0.5) for key sensors and highlights significant nonlinear interaction effects among them. This work not only enhances prediction accuracy but also offers interpretable insights to support decision-making in predictive maintenance, thereby demonstrating significant potential for health management in fluid dynamic systems.</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

Application of SHAP-based explainable machine learning in remaining useful life prediction for aircraft engine systems

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Chunhong Yuan,
  • Yule Cai,
  • Zixin Zhang

摘要

This study proposes an explainable machine learning framework based on SHAP (SHapley Additive exPlanations) for predicting the remaining useful life (RUL) of turbine systems. Using the NASA C-MAPSS dataset, we evaluate eight machine learning algorithms and find that LightGBM achieves the best performance, with a root mean square error (RMSE) of 36.68, mean absolute error (MAE) of 26.53, and coefficient of determination ( \(R^2\) ) of 0.713. SHAP analysis identifies sensor measurements 11 and 15, along with operational cycles (time_in_cycles), as the most influential features for RUL prediction. The analysis further reveals critical threshold values (approximately 0.5) for key sensors and highlights significant nonlinear interaction effects among them. This work not only enhances prediction accuracy but also offers interpretable insights to support decision-making in predictive maintenance, thereby demonstrating significant potential for health management in fluid dynamic systems.