The invention provides an online low-current earth fault positioning method and positioning device based on a BP neural network and by means of a multi-source data fusion algorithm. Due to the fact that terminals arranged on different positions on a circuit accurately catch zero-sequence current transient signals of zero-sequence currents which exceed a starting value before a previous period of a starting value and after four periods of the starting value, signal characteristics of a fault transient signals are extracted by means of a Prony algorithm, a wavelet packet algorithm, an HHT algorithm and a fractal algorithm, the characteristics extracted by the algorithms are analyzed by means of the BP neural network, and a zone where a fault occurs is determined. The positioning device comprises a terminal and a main station, wherein the terminal is arranged on an overhead line tower or in a cable ring main unit, the input end of the terminal receives a zero-sequence current signal synthetized by a phase current signal of a CT secondary side of a power distribution circuit including an overhead line and a cable, the terminal is connected with the main station through optical cable communication or mobile communication, and the main station is arranged in a transformer substation or a dispatch center, comprises an optical cable communication module and a mobile communication module and receives a signal sent by the terminal. The method and the device are mature in technology and high in reliability.