A novel method of EIS application in online fault diagnosis of high-temperature PEMFC with CNN

Electrochemical impedance spectroscopy (EIS) can be utilized for online diagnosis of proton-exchange membrane fuel cells (PEMFC) based on machine learning methods. This work proposed a novel fault diagnosis strategy based on online EIS detection in a high-temperature PEMFC stack and convolutional neural network (CNN). This method works by mapping impedances into a 2D matrix as input of CNN and adding current as an additional feature in the fully connected layers of CNN. Three different fault cases, namely different types of faults, mixed faults, and different severities of faults are investigated to evaluate diagnosis performance. The diagnosis performance is much higher than the support vector machine and k-nearest neighbor models using equivalent circuit model parameters as features. The robustness of the proposed diagnosis model was analyzed with 0–10 % noise. Results show that the robustness of the CNN model can be poor when different levels of faults are collected for each fault severity within the defined thresholds. Using noisy data to augment datasets for the model training can achieve high robustness for all three cases and mitigate the overfitting problem. This work aims to detect faulty conditions in PEMFCs onboard and hopes to help improve PEMFC lifetime via prognostic and health management.