Symbolic deep learning based prognostics for dynamic operating proton exchange membrane fuel cells

Fuel cell (FC) is a promising alternative energy source in a wide range of applications. Due to the unsatisfactory durability performance, FC has not yet been widely used. Prognostics and health management (PHM) has been demonstrated to be an effective solution to enhance the FC durability performance by predicting FC degradation characteristics and adopting health condition based control and maintenance. As the primary task of PHM, prognostics seeks to estimate the remaining useful life (RUL) of FC as early and accurately as possible. However, when FC faces dynamic operating conditions, its degradation characteristics are often hidden in the complex system dynamic behaviors, which makes prognostics challenging. To address this issue, a hybrid prognostics approach is proposed in this paper. Specifically, the health indicator of FC is extracted using a degradation behavior model and sliding-window model identification method. Subsequently, a symbolic-based long short-term memory networks (LSTM) is used to predict the health indicator degradation trend and estimate the RUL. The experimental and simulation results show that the proposed model is able to describe the dynamic behavior of the FC stack voltage and the extracted health indicator show a significant degradation trend. Moreover, health indicator prediction and RUL estimation performance can be improved by deploying the proposed symbolic-based LSTM prognostics model. The proposed approach provides a prognostic horizon approaching 50% of the FC life-cycle, and the average relative accuracy of estimated RUL is close to 90%.