Fault Root Cause Diagnosis Based on Causal Discovery and Uncertainty Awareness for Enabling the Reliability of Proton Exchange Membranes

ABSTRACT The health state of the proton exchange membrane (PEM) is affected by external environmental variables and directly determines the reliability of PEM fuel cells (PEMFC). Faults caused by abnormal conditions must be diagnosed in time to maintain stable operation of PEM. Existing methods mainly classify stack‐level fault types and hardly provide an effective basis for implementing precise fault‐tolerant control. This work proposes a diagnosis framework integrating variable causal structure and uncertainty awareness without requiring fault data for training. First, prior‐constrained causal discovery identifies causal relations among variables, and high‐frequency resistance (HFR) of PEM is estimated from external signals. An uncertainty‐aware graph neural network is constructed based on the causal structure to learn the health distribution of each variable. Finally, a health index is defined according to the deviation between observed values and predicted health distribution for anomaly localization. Experimental results of four fault tests show that the average anomaly detection rate of the proposed method is about 95% with an average false alarm rate below 0.01%. The estimated HFR and health indices follow physically consistent trends under different faults. Meanwhile, this method can output anomaly propagation paths and severity, providing strong support for timely online maintenance of PEM.