For the accurate detection and separation of increasingly serious harmonic pollution in power system, a harmonic analysis method based on sparse blind source separation is proposed. Firstly, two channel observed signals are sampled from power network with different delay time, and a mathematical model of power quality harmonic blind separation is established. Then, the short-time Fourier transform of the observed signals is made, and the points in frequency scatter diagram are clustered by the method of point density based arc circle clustering to estimate the mixing matrix. Finally, the different frequency harmonics are separated by solving L1 norm minimized problem. The test results of simulation signal and real signal of metro power system show that, the proposed method can accurately separate harmonics with an obvious advantage in computation efficiency and separation of high-order harmonics of low amount.