Bearing vibration and oil debris signal enhancement for machinery condition monitoring.

摘要

Vibration signal and lubricant oil condition are two major sources of information for machine health condition monitoring. Though vibration signal is an indirect indicator of machine conditions, it contains very rich information. On the other hand, the lubricating oil analysis provides a direct indicator of machine health conditions. The joint use of the two sources of information would compensate for their limitations and thus better maintenance actions can be expected. However, this alone is not sufficient since the two sources are often severely contaminated by background and machine interference noises. Using such contaminated data without careful de-noising will inevitably cause misleading maintenance decisions and hence premature machine failure as well as lost productivity. As such, this thesis addresses the de-noising issues for both vibration and oil condition signals. Due to different natures of the vibration signals and signals measured through oil debris monitoring sensors, different approaches will be developed in this study for the enhancement of the two types of signals. In de-noising vibration signals, this research focuses on bearings since they are one of the most vulnerable and frequently used components in rotating machinery. The results obtained based on bearings could be applied to other rotating machine components with some modifications.; Wavelet transform, in particular the Gabor wavelet transform, has been used for de-noising impulsive signals measured from faulty bearings. However, it has been a challenging task to select proper wavelet parameters. This work introduces a method to guide the selection process by a smoothness index (SI). The SI is defined as the ratio of the geometric mean to the arithmetic mean of the wavelet coefficient moduli of the vibration signal. For the signal contaminated by Gaussian white noise, we have shown that the modulus of the wavelet coefficients follows Rician distribution. Based on this observation, we then prove that the SI converges to a constant number (0.8455...) in the absence of mechanical faults or for very low signal to noise ratio. This result provides a dimensionless SI upper bound corresponding to the most undesirable case. We have also shown that the SI value decreases in the presence of impulses with properly selected parameters.; However, this approach is based on the assumption that the most impulsive components of the measured vibration are due to the faults. This assumption may not be valid in general. On the other hand, the proposed method requires a global search for the minimum SI for all combinations of wavelet parameters in the chosen discretized ranges which is a computationally demanding task. In addition, through bandpass filtering the signal, the in-band noise with frequency content in the range covered by the daughter wavelet is not eliminated. As a result, the performance of the wavelet filter based de-noising method deteriorates as the background noise intensity increases.; To mitigate the above difficulties, a novel scale selection method is proposed. In this approach we incorporated our knowledge of the resonance frequency excitation phenomenon in the scale selection algorithm. Furthermore, to improve the efficiency of the method, spectral subtraction is applied prior to wavelet transform. The proposed spectral subtraction method leads to improvements in both the final result of the process and the capability of the wavelet filter based de-noising method for lower SNR vibration signals. The proposed joint spectral subtraction and wavelet de-noising method has been successfully tested using experimental data.; For the oil condition signals, the main issue is that the oil debris sensor is not only sensitive to the metal debris or particles but the structural vibrations as well. The weak signals of small particles are often concealed in the vibration signals. This either causes false alarm (since the shape of a particle signal resembles th