Scholarlink

<p>Predicting the remaining life of bearings is critical to preventing equipment failure and downtime. Traditional single models struggle to account for the characteristics of different wear stages and have limited nonlinear modeling capabilities. A novel framework is proposed to a multistage prediction framework: first, vibration signals are divided into overlapping time–frequency windows and a small graph structure is constructed. Using GCN, samples are classified into stable periods, stable wear periods, and severe wear periods, achieving stage-specific feature aggregation. Subsequently, for stable periods and stable wear periods, a bidirectional GRU (KMBG) with fused multi-head attention and KAN modules is employed, to enhance key feature extraction and nonlinear mapping capabilities; for the severe wear stage, an RNN model is used; finally, the segmented KMBG–KMBG–RNN (KKR) scheme is employed to output the final RUL. Using the XJTU-SY bearing full-lifetime dataset as an example, experimental results show that compared to the standard GRU, KMBG reduces RMSE by 2.9% and MAE by 1.6%; compared to Bi-GRU, RMSE and MAE were reduced by 0.9% and 1.2%, respectively; the KKR model achieved optimal performance with an average RMSE of 12.4 and MAE of 9.4. Experiments demonstrate that this method exhibits excellent prediction accuracy and generalization capability under various operating conditions, providing a reliable reference for industrial intelligent maintenance.</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

Stage-adaptive remaining life forecast of rolling bearings via small-graph convolution

  • 期刊论文

  • 发表日期：

    2025年9月2日

查看全文
  • Tianming Xu,
  • Dong Jiang,
  • Zhengsong Xu

摘要

Predicting the remaining life of bearings is critical to preventing equipment failure and downtime. Traditional single models struggle to account for the characteristics of different wear stages and have limited nonlinear modeling capabilities. A novel framework is proposed to a multistage prediction framework: first, vibration signals are divided into overlapping time–frequency windows and a small graph structure is constructed. Using GCN, samples are classified into stable periods, stable wear periods, and severe wear periods, achieving stage-specific feature aggregation. Subsequently, for stable periods and stable wear periods, a bidirectional GRU (KMBG) with fused multi-head attention and KAN modules is employed, to enhance key feature extraction and nonlinear mapping capabilities; for the severe wear stage, an RNN model is used; finally, the segmented KMBG–KMBG–RNN (KKR) scheme is employed to output the final RUL. Using the XJTU-SY bearing full-lifetime dataset as an example, experimental results show that compared to the standard GRU, KMBG reduces RMSE by 2.9% and MAE by 1.6%; compared to Bi-GRU, RMSE and MAE were reduced by 0.9% and 1.2%, respectively; the KKR model achieved optimal performance with an average RMSE of 12.4 and MAE of 9.4. Experiments demonstrate that this method exhibits excellent prediction accuracy and generalization capability under various operating conditions, providing a reliable reference for industrial intelligent maintenance.