The time-frequency analysis ability for non-stationary signals was compared between parametric model based Time-Varying Autoregressive model(TVAR) and non-parametric time-frequency methods such as the SFTF and the CWD.The three methods had been evaluated with respect to a set of simulated signals and the signals collected form a rotation machine test rig during the speed-up process.The performances of TVAR in both the time-frequency resolution and the suppression of cross term are all excellent.TVAR is suitable for non-stationary signals analysis of rotation machine.And the results effectively discover the characteristics and fault features of signals.TVAR is effective to fault diagnosis of rotation machine under nonstaionary conditions.