Good features are critical for both fault diagnosis and prognosis of gearbox. Most traditional features are effective only when the gearbox is under stationary operating condition. Some features which are modified based on the traditional features are no longer sensitive to load changes and remain sensitive to fault propagation. However, it is only permit the load changing in a very small range. In engineering applications, some machines usually work in a non-stationary operating condition; both speed and load are varying over a wide range (e.g. wind turbine). So, in order to solve this dilemma, we propose using the energy ratio between residual signal and deterministic periodic signal which is separated by autoregressive model as the condition indicator for fault diagnosis and prognosis. The effectiveness of this feature is demonstrated and compared to other traditional features using two run-to-failure data sets of gearbox collected in laboratory.
