A pattern recognition method based on ensemble empirical mode decomposition (EEMD) is proposed for the non-stationary features of output signal in the fiber-optic intrusion monitoring system. The system based on the principle of Mach-Zehnder interferometer and four single-mode optical fibers in the cable are utilized to build up the distributed crosstalk sensor, by which the real-time detection of abnormal events can be realized. The vibration signals are decomposed into a series of intrinsic mode functions (IMF) using the EEMD algorithm with self-adaptability. According to the characteristics of the various vibration signal intensities, a method using the EEMD energy entropy to eliminate the disturbance of non-intrusion events is proposed. Double support vector machine is built to identify the intrusion type. The experimental results illustrate that this method can evidently get rid of the non-intrusion disturbance and effectively discern different intrusion events such as fence-climbing, cableknocking and other signals. The correct recognition rate in average is greater than 92%. What′s more, the alarm rate is increased and the false alarm rate is reduced in the system.