In this paper, a data-driven fault diagnostic strategy is designed for proton exchange membrane fuel cell systems. In order to take the system dynamics and spatial in homogeneity into account, the individual cell voltages measured in a sliding diagnosis window are considered integrally as a diagnostic observation. In the proposed diagnostic approach, a time-series analysis tool, named shapelet transform, is used to extract the discriminative features from the diagnostic observations. The classification tool, named sphere shaped multiclass support vector machine, is then carried out in the feature space in order to realize both fault detection and fault isolation. The approach is validated on the experimental testbench. The performance of the approach is evaluated and the necessity of considering both dynamic behaviors and the spatial inhomogeneity is highlighted.