Probabilistic inter-disturbance interval estimation for bearing fault diagnosis

We describe a new method for detecting characteristic bearing fault signatures from accelerometer vibration data based on a probabilistic model of the fault signal generation process. It is common to assume that single-point bearing defects cause periodic disturbances in bearing vibration signals, but this assumption may not be valid in practice. Our new method is less sensitive to departures from periodicity, such as fault disturbance amplitude and timing variations, than standard spectral or autocorrelation-based approaches. We demonstrate the utility of our method by distinguishing among inner race, outer race, and rolling element faults in a bearing fault test rig. Our method is significantly better than standard techniques at detecting rolling element (ball) faults.