As radial impact signals are often composed of rich frequency components continuously distributed in a wide region, accurately extracting the characte ristic frequencies of vibration signals of a rotor system under radial impact is of great significance for fault diagnosis. The customary Hilbert-Huang transform (HHT) cannot be directly used to decompose such signals into several mono-frequency intrinsic mode functions (IMFs). Because of the existence of multi-frequency IMFs, the estimations of instantaneous frequencies for these signals are usually inaccurate. A new feature extraction method is developed in this study, i.e., the bandpass filtering HHT (BF-HHT), which can accurately extract the characteristic frequencies of radial impact signals. Firstly, frequencies with main power are collected through calculating power spectrum densities of the signals′. Secondly, a set of bandpass filters with centering frequencies selected from the first step are designed to filter the original signals. Finally, the-time frequency characteristics are extracted from the filtered signals by HHT. Both the customary HHT and the BF-HHT are used to analyze the signals collected from rotor test bench and industrial sites. Comparisons between the results from both the methods show that, the BF-HHT can be used to extract the characteristic frequencies of radial impact signals, meanwhile, it can explore the instantaneous characteristic frequencies.