Abstract: Addresses the challenge of traditional physical models inadequately capturing nonlinear and time-varying degradation under complex operating conditions. A WP-LSTM-Attention method for predicting the Remaining Useful Life (RUL) of hollow cathodes under time-varying temperature control was developed. Firstly, A nonlinear time-varying Wiener Process (WP) model was established based on the hollow cathode’s emitter evaporation failure mechanism. Secondly, Long Short-Term Memory (LSTM) neural networks were employed to extract multi-source time-series features. The Multi-Head Attention Mechanism(MHAM) under temperature control was utilized to assign weights to degradation-sensitive features, enabling prediction of Wiener model parameters and degradation trends. Optimized parameter back propagation through a joint loss function achieved physics-data dual-driven updates. Finally, The Probability Density Function (PDF) of the RUL was derived using the first hitting time concept. The results demonstrated that, compared to single physical or data-driven models, the proposed model achieved a 22.81% reduction in Root Mean Square Error (RMSE) and an 18.11% reduction in Mean Absolute Error (MAE). These results validated the model’s effectiveness and accuracy for LaB6 hollow cathode life prediction in few-sample scenarios. The conclusion demonstrates that integrating physics with data-driven approaches significantly enhances prediction accuracy and robustness. This research provides a novel methodological pathway and technical support for lifetime prediction of LaB6 hollow cathodes.

Key words: hollow cathode, remaining useful life prediction, long short-term memory neural networks, multi-head attention mechanism, Wiener process