Scholarlink

<p>Remaining useful life (RUL) forecasting underpins intelligent condition monitoring and predictive maintenance on the shop floor. Yet real-world deployment is often limited by unreliable early-warning and unstable prognostic accuracy, due to long-horizon temporal dependencies, noise, and scarce run-to-failure labels. CET–MSA (Convolutionally-Enhanced Temporal Multi-Head Self-Attention) is introduced as a plug-and-play block that inserts lightweight, neighborhood-aware convolutions into efficient self-attention. Different from prior attention-based RUL models that primarily use attention as a temporal reweighting mechanism, CET–MSA (i) preconditions the key/value streams via temporal local-context enhancement to suppress non-stationary and impulsive disturbances and (ii) applies pyramidal key/value reduction to enable scalable long-horizon modeling with bounded latency and memory. The local path enforces temporal consistency and suppresses impulsive disturbances, while the global path preserves scalable long-horizon modeling via pyramidal down-sampling with optional locality constraints. To mitigate label scarcity and domain shift across operating conditions, a degradation-aware transfer strategy is further adopted to pre-train on sources sharing degradation mechanisms and fine-tune on short-sequence targets. On NASA C-MAPSS (FD001–FD004), CET–MSA consistently reduces the safety-aware Score and RMSE; against a 3D-block attention baseline it improves average Score by 27.8% and RMSE by 3.86%. Throughput, per-sequence latency, and peak memory are also reported to highlight deployability on resource-constrained shop-floor hardware. These results indicate that CET–MSA enables noise-robust and efficient RUL prediction, providing actionable inputs for maintenance scheduling and asset-health management in intelligent manufacturing.</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

Noise-robust remaining useful life prediction for intelligent condition monitoring via convolutionally-enhanced temporal self-attention

  • 期刊论文

  • 发表日期：

    2026年6月27日

查看全文
  • Yutao Wang,
  • Jia’e Zhang,
  • Yaru Qin

摘要

Remaining useful life (RUL) forecasting underpins intelligent condition monitoring and predictive maintenance on the shop floor. Yet real-world deployment is often limited by unreliable early-warning and unstable prognostic accuracy, due to long-horizon temporal dependencies, noise, and scarce run-to-failure labels. CET–MSA (Convolutionally-Enhanced Temporal Multi-Head Self-Attention) is introduced as a plug-and-play block that inserts lightweight, neighborhood-aware convolutions into efficient self-attention. Different from prior attention-based RUL models that primarily use attention as a temporal reweighting mechanism, CET–MSA (i) preconditions the key/value streams via temporal local-context enhancement to suppress non-stationary and impulsive disturbances and (ii) applies pyramidal key/value reduction to enable scalable long-horizon modeling with bounded latency and memory. The local path enforces temporal consistency and suppresses impulsive disturbances, while the global path preserves scalable long-horizon modeling via pyramidal down-sampling with optional locality constraints. To mitigate label scarcity and domain shift across operating conditions, a degradation-aware transfer strategy is further adopted to pre-train on sources sharing degradation mechanisms and fine-tune on short-sequence targets. On NASA C-MAPSS (FD001–FD004), CET–MSA consistently reduces the safety-aware Score and RMSE; against a 3D-block attention baseline it improves average Score by 27.8% and RMSE by 3.86%. Throughput, per-sequence latency, and peak memory are also reported to highlight deployability on resource-constrained shop-floor hardware. These results indicate that CET–MSA enables noise-robust and efficient RUL prediction, providing actionable inputs for maintenance scheduling and asset-health management in intelligent manufacturing.