Scholarlink

<p>Accurate Remaining Useful Life (RUL) prediction for Proton Exchange Membrane Fuel Cells (PEMFCs) is critically challenged by incomplete monitoring data and domain shifts. To address these issues, this paper proposes a novel few-shot transfer learning framework based on Time-Decay Self-attention Transformer (TS-transformer). Firstly, a hybrid imputation strategy combining linear interpolation and a Long Short-Term Memory (LSTM) network is used to reconstruct missing data. Secondly, Variational Mode Decomposition (VMD) combined with Mutual Information-Maximum Mean Discrepancy (MI-MMD) is utilized to suppress reconstruction noise and extract robust domain-invariant features. Finally, according to the electrochemical relaxation degradation laws of PEMFCs, a time-decay self-attention mechanism is embedded into the Transformer architecture. By incorporating dedicated mathematical constraints, this mechanism enables the model to adaptively prioritize recent degradation states, thereby overcoming the inherent lack of temporal inductive bias in the standard Transformer model. Experimental results demonstrate that high-precision prediction is achieved with only 10% of the target domain samples. Under severe data incompleteness scenarios involving 75% discrete intermittent missingness and 35% continuous segment missingness, the proposed model reduces RUL prediction errors by 52.66%–82.11% compared with existing state-of-the-art baselines. These findings provide a robust solution for PEMFC Prognostics and Health Management (PHM) under few-shot and incomplete data conditions.</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 physics-guided time-decay self-attention transformer-based few-shot transfer learning for PEMFC remaining useful life prediction under data incompleteness

  • 期刊论文

  • 发表日期：

    2026年5月22日

查看全文
  • Xinya Qiu,
  • Jianfang Jia,
  • Xiaoqiong Pang,
  • Jie Wen,
  • Yuanhao Shi,
  • Jianchao Zeng

摘要

Accurate Remaining Useful Life (RUL) prediction for Proton Exchange Membrane Fuel Cells (PEMFCs) is critically challenged by incomplete monitoring data and domain shifts. To address these issues, this paper proposes a novel few-shot transfer learning framework based on Time-Decay Self-attention Transformer (TS-transformer). Firstly, a hybrid imputation strategy combining linear interpolation and a Long Short-Term Memory (LSTM) network is used to reconstruct missing data. Secondly, Variational Mode Decomposition (VMD) combined with Mutual Information-Maximum Mean Discrepancy (MI-MMD) is utilized to suppress reconstruction noise and extract robust domain-invariant features. Finally, according to the electrochemical relaxation degradation laws of PEMFCs, a time-decay self-attention mechanism is embedded into the Transformer architecture. By incorporating dedicated mathematical constraints, this mechanism enables the model to adaptively prioritize recent degradation states, thereby overcoming the inherent lack of temporal inductive bias in the standard Transformer model. Experimental results demonstrate that high-precision prediction is achieved with only 10% of the target domain samples. Under severe data incompleteness scenarios involving 75% discrete intermittent missingness and 35% continuous segment missingness, the proposed model reduces RUL prediction errors by 52.66%–82.11% compared with existing state-of-the-art baselines. These findings provide a robust solution for PEMFC Prognostics and Health Management (PHM) under few-shot and incomplete data conditions.