NONSTATIONARY MODELING OF VIBRATION SIGNALS FOR MONITORING THE CONDITION OF MACHINERY

Most previously developed algorithms used for machinery diagnostics and condition monitoring are based on the assumption of signal stationarity. Little attention has been paid to the time-dependent details of vibration signals. In this paper, non-stationary modelling of vibration signals is proposed. An auto-regressive model and the corresponding processes of feature extraction and condition monitoring are investigated in detail. Orthogonal transformation is then introduced for both feature compression and the establishment of a statistical model of the template representing the normal condition. Lastly, the monitoring measure is determined by using similarity analysis of pattern recognition theory. An example for the vibration monitoring of a hydraulic piston pump shows how the proposed method can be used. © 1990.