In the fiber optic perimeter security system, which is based on Mach-Zehnder interferometer, the human intrusion and partial environmental noise can cause the fiber vibration. Distinguishing intrusion from environmental events without reducing the efficiency is a key requirement for any perimeter intrusion detection systems. In this paper, an signal classification system is presented for detection and recognition. This system compares event signal features, which are based on frequency domain and time domain respectively. Firstly, this research analyzes data extraction system based on human engineering. This model preprocesses time-domain data according to the characteristics of human behavior. After an event detection and feature extraction proce-dure, a classification algorithm applies K-means clustering method and the cosine similarity. In this method, the combination of signal in time and frequency domain is used to classify and improve the anti disturbance ability of the system and reduce the false alarm rate.