The invention discloses an evoked potential extracting method based on random gradient adaptive filtering. Potential signals repeatedly stimulated are collected through multiple potential measuring systems and comprise mixed signals of spontaneous potentials, spontaneous potentials obtained after stimulation is applied, and evoked potentials, all the mixed signals are superposed for averaging, and an average mixed signal is obtained; the stimulation application time serves as a starting point of counting of extracting time, a section of signal with the length being N is extracted from the average mixed signal to serve as a first main signal, a signal obtained by delaying the first main signal by one time interval serves as a first reference signal, the first main signal and the first reference signal are input into LMS adaptive filters, and first output signals of all channels are obtained; then, the first output signals serve as second reference signals, a section of spontaneous potential with the length being N serves as a second main signal, the second reference signals and the second main signal are put into the LMS adaptive filters, and evoked potentials of all the channels are obtained. By means of the evoked potential extracting method, the evoked potentials high in quality are obtained under the background that adequate priori knowledge is not available, and the number of repeated stimulation times is effective reduced.