Time-varying autoregressive model and its application to nonstationary vibration signal analysis

Time-varying autoregressive model (TVAR) is investigated and applied to analyze the signals collected from a rotation machine test rig under nonstationary conditions. TVAR is an improved autoregressive model with coefficients evolving with signal statistical characteristics. The performances in time-frequency analysis are compared between TVAR and some traditional methods by analyzing some frequency modulation (FM) signals. It is shown that TVAR has high resolutions, no cross terms and is insensitive to noises, etc. Using TVAR nonstationary signals collected in the continuously varying speed process under normal or fault states are analyzed. The results show that TVAR excels at disposing nonstationary signals and has a superior feature extracting ability; TVAR is an effective method for fault diagnosis and modal analysis under non-stationary conditions.