Toward low-data and real-time PEMFC diagnostic: Multi-sine stimulation and hybrid ECM-informed neural network

Abstract

Proton exchange membrane fuel cell (PEMFC) faults often occur abruptly, during which the multi electrochemical processes within the cell exhibit distinct changes in behavior. Consequently, real-time monitoring of multi-electrochemical processes (MEP) information is crucial for diagnosing PEMFC faults. However, existing methods struggle to balance real-time performance, interpretability, and low training data requirements, significantly limiting their reliability and feasibility for PEMFC fault diagnosis in practical scenarios. To address these issues, a novel hybrid ECM-Informed Neural Network (V-ECM) is proposed to efficiently characterize the states of the PEMFC internal electrochemical reaction, which effectively integrates real-time performance with interpretability while reducing dependence on large training datasets. The key innovations include: (1) Real-time capability—Designing a multi-sine excitation signal using the Distribution of Relaxation Time (DRT), which significantly accelerates signal acquisition speed of the electrochemistry-related response voltage. (2) High interpretability and low data dependency—Integrating physical constraints of electrochemical processes from a mechanism-based equivalent circuit model (ECM) into the loss function, to guide feature learning in the deep neural network, which contribute to improving interpretability and reducing reliance on training data. Compared to existing state-of-the-art PEMFC fault diagnosis methods, the proposed method offers high-precision, high-efficiency fault diagnosis, promising a viable solution for real-time PEMFC fault diagnosis in practical applications.