In the application of fiber optic perimeter security systems, realizing intelligent vibration sensing requires both accurately identifying specific types of sensing events and providing targeted processing solutions for such events. In this paper, we propose a signal feature-extraction algorithm that includes multidimensional time information features, combining these features in a convolutional long short-term deep neural network (CLDNN) that identifies and classifies specific vibration-sensing events. First, we stack and intercept the collected optical fiber sensing event information to obtain a broad picture containing multidimensional time characteristics of the sensing event. Next, we input these collected data into the CLDNN structure. We define five distinct types of events or signals for our identification and classification experiments: knocking, crashing, waggling, kicking, and no intrusion whatsoever. Experimental results show that the proposed algorithm can effectively recognize and classify these five types of signals with an average recognition rate of over 96% and a recognition response time that can be limited to 0.006 s.