Intelligent health states recognition of fuel cell by cell voltage consistency under typical operating parameters

In vehicular fuel cell, the change of operating parameters (pressure, temperature, humidity) may lead to health problem, which is a key parameter for fuel cell system shutdown. In this study, the health state of the proton exchange membrane fuel cell is recognized by considering several typical operating parameters. The cell voltage consistency (spatial fluctuation degree) is used to characterize the health state of fuel cell. Specifically, the health state of the minimum cell voltage is also considered. The process of health states labeling is achieved with the non-parametric statistics and unsupervised learning methods by calculating the threshold values for health evaluation indexes. Moreover, a variety of feature selection methods are applied to select the features which have relatively significant on health of fuel cell for improving the efficiency of health recognition. In addition, the random forest algorithm is used to identify the health state of based on the results of feature selection. The main results show that the relatively optimal features are temperature, current, cathode stoichiometry and pressure, respectively. Furthermore, the accuracy rate of random forest algorithm achieves to 95.04%. The effectiveness of the proposed methods is validated under operation condition of low current density and various temperatures by the results of dynamic loading experiments. The presented method of health recognition can be used to health management of fuel cell vehicle.

Automated summary

This study focuses on recognizing the health state of proton exchange membrane fuel cells by considering operating parameters, with the aim of improving the efficiency of health management in fuel cell vehicles. By using a combination of non-parametric statistics, unsupervised learning, and feature selection methods, the study successfully identifies key features leading to better health recognition and achieves a high accuracy rate of 95.04% using the random forest algorithm. Additionally, the effectiveness of the proposed method is validated through dynamic loading experiments under various operating conditions.