The successful control of the active power filter requires an accurate current reference. Based on the adaptive interference canceling theory， an adaptive least mean squares（LMS） algorithm employing the fuzzy step size is presented for the fast detection of harmonic current in power networks. By analyzing the unfavorable factors of the LMS adaptive algorithm for harmonic detection and to establish the fuzzy inference system， the mean square error of variance and input-output signals correlation function are selected as parameters for suitable automatic adjustment of the step size. The method proposed can not only provide fast convergence and noise restraint for the tracking of fundamental and harmonic components from distorted signals， but also maintain a fairly high detection accuracy. The results of computer simulations and physical experiments verify the reliability and feasibility of the method proposed.