A large amount of studies have shown that compared with sweep frequency response analysis (SFRA), impulse frequency response (IFRA) is feasible for online condition monitoring of power transformer winding deformation, but nowadays even offline IFRA still has many drawbacks, e.g., the features of the injected impulse is not standardized and the sensitivity, reliability and reproducibility of IFRA is in question. In this paper, the features of impulses suitable for IFRA is extracted,and a novel denoising method based on threshold wavelet denoising is designed to extract appropriate IFRA response curves. Three levels of axial displacements were experimentally simulated on a distribution transformer. Both the IFRA system and a commercial SFRA analyzer were applied to obtain the frequency responses of the trans-former at each fault level. Both the IFRA and SFRA measurement are conducted twice with two different connection schemes. Correlation Coefficient and Euclidean Distance are employed for the interpretation of IFRA frequency responses. The reliability, reproducibility and sensitivity of the developed IFRA system is verified through experimental studies.