Abstract: Fault diagnosis has been considered as a critical solution for improving the reliability and durability of proton exchange membrane fuel cell(PEMFC), while researches regarding early fault diagnosis are still at its infancy and urgently needed in applications where timely maintenance and control are required, such as fuel cell vehicle. Considering limited early fault diagnosis methods, a fault diagnosis approach using PEMFC magnetic field data is proposed. A PEMFC numerical model is developed firstly to investigate the correlation between PEMFC states and magnetic field variation. On such basis, a PEMFC magnetic field detection system is built, from which collected magnetic field is analysed by a convolutional neural network (CNN) to verify its early diagnosis effect at various PEMFC faults, including flooding, dehydration at different fault levels. Results indicate that with magnetic field data analysed by CNN, different PEMFC faults at various levels can be on-line identified and early detected. This research verifies the feasibility of using magnetic field data for PEMFC fault diagnosis, which is of vital significance to further perform PEMFC fault diagnosis to improve the reliability and durability of PEMFC.