Abstract:Several time—frequency representations including time varying autoregressive model (TVAR), short time Fourier transform (STFT), Wigner—Ville distribution (WVD), Choi—Williams distribution (CWD), continuous wavelet transform (CWT) and Hilbert—Huang transform (HHT) were compared to the respect of their achievable time and frequency resolution, cross-term suppression and computational load. Those methods were taken to analyze an AM—FM signal simulating the vibration of a running-up rotor. Comparing the results, it could be concluded that TVAR with the best comprehensive performance to such kind of signal. Consequently, TVAR was applied to process a real running-up vibration signal of a rotor with STFT as a benchmark. The result showes TVAR has potential to process nonstationary signals with superior feature extraction ability, and is fairly noise insensitive. In conclusion, TVAR provides a superior approach for time—frequency analysis and fault diagnosis of rotation machine under nonstationary conditions.