In view of the large amount of data and low accuracy rate of the existing power quality disturbance signal recognition methods, a new recognition method of power quality disturbance signals is proposed by extracting the features from compressed sensing sparse vectors. In this method, the original signals are sampled to obtain the measurement signals based on the theory of compressive sensing, and the sparse vectors are obtained by the orthogonal matching pursuit algorithm of ?1-minimization. Then the features of the maximum, the second maximum, root mean square, standard deviation, kurtosis and margin factor are extracted as the inputs of the neural networks, and the power quality disturbance signal recognitions are realized. According to 6 kinds of typical power quality disturbance signals, the simulation experiments are conducted. The simulation results show that the data size of the proposed recognition method for feature extraction is greatly reduced with only 30 instead of 1024 for the existing methods. As a result of the realization of a very small amount of data to retain the original all useful feature information, the proposed method is more promising to improve the recognition accuracy. Compared with the widely-used wavelet transform recognition method, the average accuracy rate of this proposed method is as high as 98.71%, which is much higher than 92.86% of the wavelet transform method. This work is supported by National Natural Science Foundation of China (No. 51307144).