In order to improve the accuracy of the power quality disturbance signals classification, firstly, the one-dimensional time series power quality disturbance signals are reconstructed into the multidimensional phase space by phase-space reconstruction theory, thereby obtaining the disturbance signals trajectory and projecting to the two-dimensional plane, and forming a two-dimensional images. Then these images are binarized to reduce the amount of data of the signal and highlight the trajectory profile. Finally, the convolution neural network is used to extract the features of trajectory images and achieve the classification and identification of the corresponding disturbance signals. Simulation experiments are carried out under the Caffe framework of convolution neural network, showing that the method has high recognition accuracy and good ability to resist noise. This work is supported by National Natural Science Foundation of China (No. 51767017 and No. 51267012).