The Optimization of Vibration Data Analysis for the Detection and Diagnosis of Incipient Faults in Roller Bearings

The rolling element bearing is a key component of many machines. Accurate and timely diagnosis of its faults is critical for proactive predictive maintenance. The research described in this thesis focuses on the development of techniques for detecting and diagnosing incipient bearing faults. These techniques are based on improved dynamic models and enhanced signal processing algorithms.

Various common fault detection techniques for rolling element bearings are reviewed in this work and a detailed experimental investigation is described for several selected methods. Envelope analysis is widely used to obtain the bearing defect harmonics from the spectrum of the envelope of a vibration signal. This enables the detection and diagnosis of faults, and has shown good results in identifying incipient faults occurring on the different parts of a bearing. However, a critical step in implementing envelope analysis is to determine the frequency band that contains the signal component corresponding to the bearing fault (the one with highest signal to noise ratio). The choice of filter band is conventionally made via manual inspection of the spectrum to identify the resonant frequency where the largest change has occurred. In this work, a spectral kurtosis (SK) method is enhanced to enable determination of the optimum envelope analysis parameters, including the filter band and centre frequency, through a short time Fourier transform (STFT). The results show that the maximum amplitude of the kurtogram indicates the optimal parameters of band pass filter that allows both outer race and inner race faults to be determined from the optimised envelope spectrum. A performance evaluation is carried out on the kurtogram and the fast kurtogram, based on a simulated impact signal masked by different noise levels. This shows that as the signal to noise ratio (SNR) reaches as low as -5dB the STFT-based kurtogram is more effective at identifying periodic components due to bearing faults, and is less influenced by irregular noise pulses than the wavelet based fast kurtogram.

A study of the accuracy of rolling-bearing diagnostic features in detecting bearing wear processes and monitoring fault sizes is presented for a range of radial clearances. Subsequently, a nonlinear dynamic model of a deep groove ball bearing with five degrees of freedom is developed. The model incorporates local defects and clearance increments in order to gain the insight into the bearing dynamics. Simulation results indicate that the vibrations at fault characteristic frequencies exhibit significant variability for increasing clearances. An increased vibration level is detected at the bearing characteristic frequency for an outer race fault, whereas a decreased vibration level is found for an inner race fault. Outcomes of laboratory experiments on several bearing clearance grades, with different local defects, are used herein for model validation purposes.

The experimental validation indicates that the envelope spectrum is not accurate enough to quantify the rolling element bearing fault severity adequately. To improve the results, a new method has been developed by combining a conventional bispectrum (CB) and modulation signal bispectrum (MSB) with envelope analysis. This suppresses the inevitable noise in the envelope signal, and hence provides more accurate diagnostic features. Both the simulation and the experimental results show that MSB extracts small changes from a faulty bearing more reliably, enabling more accurate and reliable fault severity diagnosis. Moreover, the vibration amplitudes at the fault characteristic frequencies exhibit significant changes with increasing clearance. However, the vibration amplitude tends to increase with the severity of an outer race fault and decrease with the severity of an inner race fault. It is therefore necessary to take these effects into account when diagnosing the size of a defect.