Scholarlink

<p>Accurate remaining useful life (RUL) prediction is essential for condition-based maintenance in modern manufacturing systems. While Transformer-based models have demonstrated strong sequence modeling capability, RUL prediction requires degradation-consistent representation learning across multiple temporal resolutions. This study investigates multi-scale degradation-aware feature representation within a Transformer framework and proposes a Multi-scale Feature Dual-Aspect Self-Attention Transformer (MFDAST). The proposed approach integrates multi-resolution statistical feature extraction and a monotonic Health Indicator (HI) as a regularization mechanism to encourage stable degradation modeling. Genetic algorithm (GA)-based hyperparameter tuning is further applied to explore parameter interactions under controlled settings. Experiments conducted on the FEMTO-ST (PRONOSTIA) bearing degradation dataset demonstrate consistent performance improvements over representative RNN-based and baseline Transformer models under identical experimental configurations, achieving up to 50% reduction in RMSE under specific operating conditions. An additional ablation study on the C-MAPSS turbofan engine dataset further validates the generality of the proposed design across different RUL prediction scenarios. The results confirm that multi-scale degradation representation is the primary source of performance gains, while HI modeling and GA optimization provide incremental improvements in prediction accuracy and training stability. These findings highlight the importance of degradation-aware multi-resolution representation for RUL prediction and provide practical insights for designing Transformer-based prognostic models in industrial 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

Innovative transformer-driven remaining useful life (RUL) prediction enhanced by adaptive multi-scale feature engineering

  • 期刊论文

  • 发表日期：

    2026年4月2日

查看全文
  • Lydia Hsiao-Mei Lin,
  • Fang-Kai Ting,
  • Shu-Han Liao,
  • Chia-Hsun Hsieh,
  • Chia-Hsuan Ho,
  • Simon Hung-Yi Lu,
  • Richard Tzong-Han Tsai

摘要

Accurate remaining useful life (RUL) prediction is essential for condition-based maintenance in modern manufacturing systems. While Transformer-based models have demonstrated strong sequence modeling capability, RUL prediction requires degradation-consistent representation learning across multiple temporal resolutions. This study investigates multi-scale degradation-aware feature representation within a Transformer framework and proposes a Multi-scale Feature Dual-Aspect Self-Attention Transformer (MFDAST). The proposed approach integrates multi-resolution statistical feature extraction and a monotonic Health Indicator (HI) as a regularization mechanism to encourage stable degradation modeling. Genetic algorithm (GA)-based hyperparameter tuning is further applied to explore parameter interactions under controlled settings. Experiments conducted on the FEMTO-ST (PRONOSTIA) bearing degradation dataset demonstrate consistent performance improvements over representative RNN-based and baseline Transformer models under identical experimental configurations, achieving up to 50% reduction in RMSE under specific operating conditions. An additional ablation study on the C-MAPSS turbofan engine dataset further validates the generality of the proposed design across different RUL prediction scenarios. The results confirm that multi-scale degradation representation is the primary source of performance gains, while HI modeling and GA optimization provide incremental improvements in prediction accuracy and training stability. These findings highlight the importance of degradation-aware multi-resolution representation for RUL prediction and provide practical insights for designing Transformer-based prognostic models in industrial applications.