The gear box is a key component of the electric multiple unit(EMU), as its performance directly influences the reliability and security of the EMU. Since the gear box has a complex structure and may be excited by many factors in actual operation, it is difficult to establish a proper dynamic model. This paper introduces a time-series algorithm for detecting the gear box fault that is data-driven, which saves the trouble of establishing a dynamic model and is suitable for online fault prediction. By measuring the acceleration signal from the gear box using the wireless micro electro-mechanism sensor(MEMS), the time-series model and the fault sensitive parameters (FSP) are defined by the model′s regression parameters. The averages of the FSP are significantly different between fault and fault free tests. The fault can be detected by comparing the averages of FSP and t test of the hypothesis testing. This practice has proved that the method introduced in this paper can identify the early fault of the EMU′s gear box online, and its application can be of great value.