A novel health indicator for PEMFCs based on VAE-LSTM-attention and mahalanobis distance with long-term remaining useful life prediction

Proton exchange membrane fuel cells (PEMFCs) are a highly promising technology for power generation; however, their large-scale deployment is still hindered by durability. Prognostics and health management methods provide an effective way to extend the lifespan of PEMFC systems. However, constructing a reliable health indicator (HI) for assessing the health state of PEMFCs remains challenging, particularly under dynamic operating conditions. To address this issue, we propose a new health assessment algorithm based on a variational autoencoder (VAE). The encoder is implemented as an attention-equipped long short-term memory (LSTM-Attention) module, in combination with a Mahalanobis distance (MD). First, we select representative degradation features using correlation metrics. Then, aging features are embedded into a compressed latent degradation space via the encoder network. An HI is then derived from the MD between the evolving latent degradation features and a reference feature representing the initial healthy state. Finally, the unscented particle filter (UPF) algorithm estimates the long-term remaining useful life (RUL). The proposed method is validated using static and dynamic aging datasets from the FCLAB benchmark. Results show that the HI effectively captures the latent degradation patterns of PEMFCs and outperforms traditional voltage-based indicators. Moreover, by integrating multi-sensor data, the method exhibits strong adaptability under dynamic operating conditions, reducing the impact of load changes on degradation assessment. The long-term RUL predictions also offer valuable insights for predictive maintenance of PEMFC systems. • A VAE-LSTM-Attention framework is developed to extract latent degradation features. • A novel HI defined based on the Mahalanobis distance is introduced. • The proposed HI captures latent and comprehensive degradation patterns. • An HI degradation model and UPF algorithm are integrated for long-term RUL prediction.