Scholarlink

<p>Accurate remaining useful life (RUL) prediction for rolling bearings hinges on reliable signal decomposition and a robust health indicator (HI)-to-RUL mapping. This paper proposes an integrated framework that couples an improved whale optimization algorithm (IWOA)-driven adaptive variational mode decomposition (VMD) in a convolutional neural network (CNN) with bidirectional long short-term memory (BiLSTM), dubbed as CNN-BiLSTM (CBL) predictor. Unlike existing VMD-based approaches that rely on manual parameter tuning or unconstrained optimization, the proposed method employs an entropy-guided objective function driven by the IWOA to joint search under reconstruction-error constraints. This enables the method to generate stable and physically meaningful intrinsic mode functions (IMFs) while filtering out unreliable components. Based on the decomposed components, we construct a lightweight self-attention health indicator (SA-HI) to better reflect degradation progression than traditional statistics. Extensive experiments on the XJTU-SY bearing dataset show that SA-HI achieves better monotonicity, trendility, and robustness than several existing factors, and the overall framework delivers more accurate and stable RUL predictions under leave-one-bearing-out validation and cross-condition tests, outperforming several mainstream deep learning baselines.</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

A self-attention health indicator and CNN-BiLSTM based framework for remaining useful life prediction of rolling bearings

  • 期刊论文

  • 发表日期：

    2026年6月3日

查看全文
  • Rong Yuan,
  • Fan Wu,
  • Yulin Li,
  • Yang Ou

摘要

Accurate remaining useful life (RUL) prediction for rolling bearings hinges on reliable signal decomposition and a robust health indicator (HI)-to-RUL mapping. This paper proposes an integrated framework that couples an improved whale optimization algorithm (IWOA)-driven adaptive variational mode decomposition (VMD) in a convolutional neural network (CNN) with bidirectional long short-term memory (BiLSTM), dubbed as CNN-BiLSTM (CBL) predictor. Unlike existing VMD-based approaches that rely on manual parameter tuning or unconstrained optimization, the proposed method employs an entropy-guided objective function driven by the IWOA to joint search under reconstruction-error constraints. This enables the method to generate stable and physically meaningful intrinsic mode functions (IMFs) while filtering out unreliable components. Based on the decomposed components, we construct a lightweight self-attention health indicator (SA-HI) to better reflect degradation progression than traditional statistics. Extensive experiments on the XJTU-SY bearing dataset show that SA-HI achieves better monotonicity, trendility, and robustness than several existing factors, and the overall framework delivers more accurate and stable RUL predictions under leave-one-bearing-out validation and cross-condition tests, outperforming several mainstream deep learning baselines.