Application of Time-Frequency Distributions to the Blind Source Separation of Mechanical Fault Signals

In this study, blind source separation (BSS) method was applied to separate the multi-channel fault vibration signals generated by a rotor. As the signals were non-stationary, an algorithm based on spatial time-frequency distributions was applied to the experimental vibration signals to obtain the non-stationary vibration sources that were mutually independent. Further, AR modeling estimates of these sources were calculated with BURG method. A neural network was applied to the AR modeling parameters to perform the fault classification. The separation results of an experiment on a rotor's multi-fault show that this method is feasible for fault diagnosis.