To solve the problem of noise elimination in fault feature extraction of sensor signal and describing fault propagation under model uncertainty, this article presents a novel fault diagnosis approach based on empirical mode decomposition (EMD) and directed factor graph (DFG). The EMD method is used to decompose the sensor output signal into a number of intrinsic mode function (IMF) components, a block energy criterion based on the signal samples between two adjacent zero-crossings of IMF is proposed to distinguish the useful signal from noise. Directed factor graph is used to model the cause-effect relations between system components, and as the basis for fault diagnosis through probabilistic reasoning under the model uncertainty. A power supply module of a spacecraft power system is provided as case study to show the feasibility and validity of the proposed method.