To accurately and rapidly recognize and classify different kinds of sensing events in distributed fiber optic vibration sensing system, a time-frequency based hybrid feature extraction algorithm has been proposed. In the algorithm, a zero crossing rate based time domain feature vector and a wavelet packet energy based frequency domain feature vector are used as the feature description of the given sensing event. Then, the feature vectors are classified by radial basis function neural network classifier. A series of experimental results show that the vibrations can be accurately recognized from the noise with high efficiency. Specifically, the average identification rate of 94. 5% is achieved and the recognition response time can be limited in 0. 3 s.