In order to solve the problem that the signal is submerged in periodic narrow-band interference in the partial discharge measurement site, an improved variable step size least mean square (LMS) adaptive filter algo-rithm applied to the transformer partial discharge online monitoring system is proposed in this paper. By constructing a new filter function combined with the actual situation of the AD chip range adaptive ad-justment step size, it solves the traditional The LMS algorithm needs to match the order and step size, and has the disadvantages of poor convergence and easy divergence. The special case where the delay is found to be zero in the test is discussed and analyzed, which provides a high signal-to-noise ratio adaptive filter design. For reference. The change of the initial iteration step length of the new method is compared with the iter-ation process of the traditional fixed-step LMS algorithm, which proves that the new method has the ad-vantages of fast convergence and not easy to diverge. Finally, through the transformer partial line moni-toring device built in the laboratory, the different effects of the traditional LMS algorithm and the algorithm in this paper under the measured data are compared and analyzed, and the signal-to-noise ratio (SNR), mean square error (MSE) and waveform similarity coefficient (NCC) verifies the superiority of the new method.