Phase-sensitive optical time-domain reflectometer (φ-OTDR) is widely used for safety monitoring of large-scale civil objects, by which external vibrations along the sensing fiber can be detected. It has to be noticed that the category of vibration signals should be accurately distinguished for many real applications. At present, an extensively approach of signal recognition is deep convolutional neural network (DCNN). In the work, the one-dimensional DCNN (1D-DCNN) is applied to recognize different sound-induced vibrations based on their time-domain intensity signals detected by an amplitude-demodulated φ-OTDR system. It is turned out that the DCNN successfully shows the capability of recognizing walking, rock drill, explosion, hand hammer, car siren, and background noises with a high accuracy. Additionally, the 1D time-domain intensity vectors are rearranged into 2D matrices and the 2D-DCNN is accordingly employed to identify these vibration signals. The confusion matrices demonstrate that the 1D-DCNN has a higher average recognition accuracy to identify the concerned sounds with respect to the 2D-DCNN.