Although a variety of design and control strategies have been proposed to improve the performance of polymer electrolyte membrane (PEM) fuel cell systems, temporary faults in such systems still might occur during operations due to the complexity of the physical process and the functional limitations of some components. The development of an effective condition monitoring system that can detect these faults in a timely manner is complicated by the operating condition variation, the significant variability/uncertainty of the fuel cell system, and the measurement noise. In this research, we propose a model-based condition monitoring scheme that employs the Hotelling T-2 statistical analysis for fault detection of PEM fuel cells. Under a given operating condition, the instantaneous load current, the temperature and fuel/gas source pressures of the fuel cell are measured. These measurements are then fed into a lumped parameter dynamic fuel cell model for the establishment of the baseline under the same operating condition for comparison. The fuel cell operation is simulated under statistical sampling of parametric uncertainties with specified statistics (mean and variance) that account for the system variability/uncertainty and measurement noise. This yields a group of output voltages (under the same operating condition but with uncertainties) as the baseline. Fault detection is facilitated by comparing the real-time measurement of the fuel cell output voltage with the baseline voltages by employing the Hotelling T-2 statistical analysis. The baseline voltages are used to evaluate the output T-2 statistics under normal operating condition. Then, with a given confidence level the upper control limit can be specified. Fault condition will be declared if the T-2 statistics of real-time voltage measurement exceeds the upper control limit. This model-based robust condition monitoring scheme can deal with the operating condition variation, various uncertainties in a fuel cell system, and measurement noise. Our analysis indicates that this scheme has very high detection sensitivity and can detect the fault conditions at the early stage. (c) 2006 Elsevier B.V. All rights reserved.