Scholarlink

<p>Many studies have focused on remaining useful life (RUL) prediction via continual learning. This approach involves learning degradation patterns from different datasets in multi-stages to achieve reliable and accurate results. However, existing research faces two key issues: catastrophic forgetting (losing prior knowledge) and plasticity loss (diminished capacity to learn new information) after multi-stage learning. To simultaneously mitigate both issues in RUL prediction, we introduce TFGI, a framework that integrates adaptive bidirectional feature fusion, dynamic plasticity preservation, and time-frequency joint learning, which is designed with inherent parallelism to support efficient execution on high-performance computing (HPC) platforms. In TFGI, the input data are preprocessed by a time-frequency joint learning mechanism. This mechanism exploits the butterfly-operation parallelism of FFT and enhances the model’s sensitivity to multi-stage degradation patterns. Unlike existing methods that focus solely on regularization or replay strategies, TFGI introduces a temporal-aware gating mechanism. This mechanism adaptively balances bidirectional feature extraction, BiGRU captures short-term local degradation via gated memory, while Informer isolates long-term global precursors via ProbSparse attention. Furthermore, we implement a dynamic plasticity preservation mechanism to mitigate plasticity loss during training using a continual backpropagation algorithm. Its neuron-wise utility computations are trivially data-parallel, which makes the training scalable across GPU devices. Our framework improves RUL prediction accuracy when compared with several state-of-the-art (SOTA) methods on a rolling bearing run-to-failure dataset. Its parallel dual-path design, combined with O(L ln L) complexity attention, makes it a promising candidate for real-time RUL prediction on HPC systems, aligning well with the scope of supercomputing research.</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

TFGI: a continual learning framework for remaining useful life prediction based on temporal fusion model

  • 期刊论文

  • 发表日期：

    2026年6月24日

查看全文
  • Haonan Yang,
  • Jianchao Tang,
  • Shaowu Yang

摘要

Many studies have focused on remaining useful life (RUL) prediction via continual learning. This approach involves learning degradation patterns from different datasets in multi-stages to achieve reliable and accurate results. However, existing research faces two key issues: catastrophic forgetting (losing prior knowledge) and plasticity loss (diminished capacity to learn new information) after multi-stage learning. To simultaneously mitigate both issues in RUL prediction, we introduce TFGI, a framework that integrates adaptive bidirectional feature fusion, dynamic plasticity preservation, and time-frequency joint learning, which is designed with inherent parallelism to support efficient execution on high-performance computing (HPC) platforms. In TFGI, the input data are preprocessed by a time-frequency joint learning mechanism. This mechanism exploits the butterfly-operation parallelism of FFT and enhances the model’s sensitivity to multi-stage degradation patterns. Unlike existing methods that focus solely on regularization or replay strategies, TFGI introduces a temporal-aware gating mechanism. This mechanism adaptively balances bidirectional feature extraction, BiGRU captures short-term local degradation via gated memory, while Informer isolates long-term global precursors via ProbSparse attention. Furthermore, we implement a dynamic plasticity preservation mechanism to mitigate plasticity loss during training using a continual backpropagation algorithm. Its neuron-wise utility computations are trivially data-parallel, which makes the training scalable across GPU devices. Our framework improves RUL prediction accuracy when compared with several state-of-the-art (SOTA) methods on a rolling bearing run-to-failure dataset. Its parallel dual-path design, combined with O(L ln L) complexity attention, makes it a promising candidate for real-time RUL prediction on HPC systems, aligning well with the scope of supercomputing research.