Scholarlink

Accurate prediction of the Remaining Useful Life (RUL) is essential for turbofan engine health management. However, deep learning methods often fail to effectively model the complex relationships between sensor data and operational conditions, resulting in limited interpretability and practical utility. To address these challenges, we propose the Spatial-Temporal Fusion Conditional Variational Autoencoder (STFCVAE), a novel framework that integrates physics-informed sensor grouping with attention-based encoding to extract degradation features. Our approach employs a CVAE architecture enhanced by an exponentially weighted triplet loss, ensuring a structured and discriminative latent representation. Additionally, a two-level reconstruction decoder improves both signal reconstruction and model interpretability. Experiments on the C-MAPSS dataset show that STF-CVAE outperforms existing methods in both RUL prediction accuracy and interpretability while providing degradation trends through its interpretable latent space.

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

Conditional Variational Autoencoder for Remaining Useful Life Prediction with Spatial-Temporal Feature Fusion

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yuzhe Hao,
  • Yilin Wang,
  • Tao Chen,
  • Yongsheng Yang,
  • Yuanxiang Li

摘要

Accurate prediction of the Remaining Useful Life (RUL) is essential for turbofan engine health management. However, deep learning methods often fail to effectively model the complex relationships between sensor data and operational conditions, resulting in limited interpretability and practical utility. To address these challenges, we propose the Spatial-Temporal Fusion Conditional Variational Autoencoder (STFCVAE), a novel framework that integrates physics-informed sensor grouping with attention-based encoding to extract degradation features. Our approach employs a CVAE architecture enhanced by an exponentially weighted triplet loss, ensuring a structured and discriminative latent representation. Additionally, a two-level reconstruction decoder improves both signal reconstruction and model interpretability. Experiments on the C-MAPSS dataset show that STF-CVAE outperforms existing methods in both RUL prediction accuracy and interpretability while providing degradation trends through its interpretable latent space.