Manually identifying fast radio burst (FRB) events from the massive candidates by a human is a laborious and time-consuming task. It's an unsustainable working mode for the constantly growing volume of observation data. In this paper, we present a method of FRB candidates classification based on convolutional neural networks (CNN). First, we build training and test sets with real observation data and simulated FRBs. Second, a two-input deep convolutional neural network model is constructed, trained and optimized, and the FRB candidate classifier is obtained. Then, the effectiveness and performance of the classifier are tested and verified by using single pulses from pulsar. Experiment results show that this method can quickly and accurately identify single pulse events from candidates, which greatly improves the processing speed and efficiency of FRB candidates.