Scholarlink

<p>The full-lifecycle degradation data of rolling bearings exhibit large volume, strong heterogeneity, pronounced nonlinearity, and marked non-stationarity. Conventional deep learning approaches for remaining useful life prediction suffer from limited feature representation and poor modeling of temporal dependencies, leading to suboptimal accuracy and limited robustness. To overcome these limitations, this study proposes a dual-stream iTransformer–LSTM architecture for remaining useful life prediction of rolling bearings. First, intrinsic mode functions are extracted from raw vibration signals via empirical mode decomposition, followed by the construction of a multidimensional degradation feature set through combined time–frequency domain analysis. Subsequently, iTransformer and LSTM modules are employed to extract local temporal variations and long-term dependencies, respectively, and a cross-attention mechanism is introduced to facilitate feature fusion. Finally, the Kolmogorov–Arnold Network is employed for high-dimensional feature mapping, enhancing nonlinear representation and improving predictive performance. Experimental validation on the IEEE PHM2012 benchmark dataset demonstrates that, compared to several existing prediction methods, the proposed method reduces the mean absolute error by 5.8 to 50.5% and the root mean square error by 7.32 to 49.33% across different bearing samples, thereby confirming the effectiveness and feasibility of the approach.</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

iTransformer-LSTM dual-stream architecture for rolling bearing remaining useful life prediction

  • 期刊论文

  • 发表日期：

    2025年9月10日

查看全文
  • Zhigang Chen,
  • Mengyao Shi,
  • Yanxue Wang,
  • Longqiao Chen

摘要

The full-lifecycle degradation data of rolling bearings exhibit large volume, strong heterogeneity, pronounced nonlinearity, and marked non-stationarity. Conventional deep learning approaches for remaining useful life prediction suffer from limited feature representation and poor modeling of temporal dependencies, leading to suboptimal accuracy and limited robustness. To overcome these limitations, this study proposes a dual-stream iTransformer–LSTM architecture for remaining useful life prediction of rolling bearings. First, intrinsic mode functions are extracted from raw vibration signals via empirical mode decomposition, followed by the construction of a multidimensional degradation feature set through combined time–frequency domain analysis. Subsequently, iTransformer and LSTM modules are employed to extract local temporal variations and long-term dependencies, respectively, and a cross-attention mechanism is introduced to facilitate feature fusion. Finally, the Kolmogorov–Arnold Network is employed for high-dimensional feature mapping, enhancing nonlinear representation and improving predictive performance. Experimental validation on the IEEE PHM2012 benchmark dataset demonstrates that, compared to several existing prediction methods, the proposed method reduces the mean absolute error by 5.8 to 50.5% and the root mean square error by 7.32 to 49.33% across different bearing samples, thereby confirming the effectiveness and feasibility of the approach.