Scholarlink

Aero-engine RUL prediction, crucial for flight safety and economic maintenance, has seen data-driven methods like LSTM and Transformer gain prominence. However, existing studies based on single-condition datasets struggle with the complex, varying conditions of real flights. This paper introduces a novel RUL prediction method combining causal association and dynamic weighting by large language models (LLMs). It decomposes the information entropy between operating conditions and sensor data into redundant, unique, and synergistic parts, supplements feature information via correlation analysis, and converts these into natural language for LLM-generated dynamic weights. An improved Transformer encoder captures dynamic degradation features, fusing with LLM weights through gated residual fusion. The training involves main model pre-training and joint fine-tuning with physical constraints and regularization. Experiments on C-MAPSS datasets show our method’s superiority, achieving RMSE of 17.19 and Score of 1642.64 on FD002, outperforming other methods.

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

Causal Correlation-Driven Dynamic Weighting with Large Language Models for Aero-Engine Remaining Useful Life Prediction

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ying Liu,
  • Shuai Xu

摘要

Aero-engine RUL prediction, crucial for flight safety and economic maintenance, has seen data-driven methods like LSTM and Transformer gain prominence. However, existing studies based on single-condition datasets struggle with the complex, varying conditions of real flights. This paper introduces a novel RUL prediction method combining causal association and dynamic weighting by large language models (LLMs). It decomposes the information entropy between operating conditions and sensor data into redundant, unique, and synergistic parts, supplements feature information via correlation analysis, and converts these into natural language for LLM-generated dynamic weights. An improved Transformer encoder captures dynamic degradation features, fusing with LLM weights through gated residual fusion. The training involves main model pre-training and joint fine-tuning with physical constraints and regularization. Experiments on C-MAPSS datasets show our method’s superiority, achieving RMSE of 17.19 and Score of 1642.64 on FD002, outperforming other methods.