Accurate online health assessment of fuel cell systems is a key for the timely mitigation and maintenance actions to be taken in order to maximise reliability of operation and useful life span of the cells. The majority of approaches rely on occasional probing of the system with small-amplitude signals around an operating point. The responses are then used to create either a parametric or a non-parametric model of the linearised system dynamics. However, during the probing session, the measurements might be corrupted with random noise and disturbances. Consequently, the evaluated parameters, being points on the impedance curve, parameters of the equivalent circuit models or the distribution of relaxation times, contain some uncertainty. That fact is largely ignored in the state of the art techniques, meaning that only mean value estimates are taken into account in the further analysis. In this paper we use a non-parametric two-sample Kolmogorov-Smirnov test to detect a change in the internal condition by evaluating changes at each frequency point on the Nyquist curve. Moreover, we show that in some cases it is even possible to isolate the fault origin from the pattern of detected changes. The applicability of the approach is demonstrated on the detection of water management faults of an industrial proton exchange membrane fuel cell system.