【摘要】 有效的健康指标（HI）和预测方法是评估质子交换膜燃料电池（PEMFC）剩余使用寿命（RUL）的关键，而在工况复杂的动态负载下直接获取HI是具有挑战性的。此外，传统的深度学习方法受限于一维数据结构的分析模式，难以准确预测PEMFC中包含多种变化模式的老化趋势。因此，本文提出了一种应对动态工况的预测框架，针对动态负载进行分级处理以得到具有老化趋势的伪动态数据，随后提出了一种经验模态分解、功率谱密度和能量分析相结合的方法（EPE），从分级负载中提取了反映动态工况老化趋势的HI。此外在预测方法上，采用了TimesNet网络将一维HI时间序列转换到二维空间并估计了RUL。结果表明，提取的HI可有效表征PEMFC的老化趋势。在RUL估计方面，TimesNet相比BiLSTM和CNN-BiLSTM的误差分别下降了61.8%和25%。更多还原

【Abstract】 Effective health indicators(HI) and predictive methods are crucial for assessing the remaining useful life(RUL) of proton exchange membrane fuel cell(PEMFC). However, directly obtaining HIs under complex dynamic load conditions poses significant challenges. Traditional deep learning methods, constrained by their one-dimensional data analysis framework, struggle to accurately predict the aging trends of PEMFCs, which encompass various modes of degradation. To address this, our study proposes a novel prediction framework tailored for dynamic operating conditions. This framework involves a hierarchical processing of dynamic loads to generate pseudodynamic data that encapsulates aging trends. Subsequently, a method that combines empirical mode decomposition(EMD), power spectral density(PSD), and energy analysis(EA) is proposed to extract HIs that reflect aging trends under dynamic conditions. Furthermore, in terms of prediction methods, the TimesNet network is utilized to transform one-dimensional HI time series into two-dimensional space and estimate RUL. The results demonstrate that the extracted HIs effectively capture the aging trends of PEMFCs. In terms of RUL estimation, TimesNet reduces the prediction error by 61.8% compared to bidirectional long-short term memory(BiLSTM) and by 25% compared to bidirectional long short-term memory with convolutional neural network(CNN-BiLSTM).更多还原