Rotating machinery monitoring: Feature extraction, signal separation, and fault severity evaluation

Machine health monitoring systems (HMS) have been implemented for detection and diagnosis of different machinery faults to improve safety and reliability and cut down cost. However, the accuracy and timeliness of many such systems are still inadequate due to various noises and the interference of multiple signals. Therefore, many such systems tend to be either sluggish in fault detection or over-vigilant leading to false alarms. In addition, the literature has shown a lack of dependable and consistent fault severity measures for different defects. The existing fault severity measures often require comparisons with reference data. However, the reference data are not always available particularly for new machines or new operations. This obviously undermines their usefulness. This often causes either over-vigilance or slow actions, hence further hindering the diagnosis process. To provide accurate and timely detection and diagnosis decisions, this study focuses on three important subjects of fault diagnosis, i.e., fault detection, fault isolation, and fault severity assessment. To improve the accuracy of fault detection, this study proposes three approaches: (1) A joint continuous wavelet transform (CWT) and autocorrelation method for extracting periodic fault features from a noisy signal mixture; (2) A kurtosis-based wavelet thresholding algorithm for de-noising and for detecting impulsive fault features; and (3) A generalized differentiation method for in-line oil-debris signature identification. The joint CWT and autocorrelation method can effectively remove fake (or spurious) impulses and hence improve the accuracy in detecting true fault signatures. Unlike many existing de-noising methods, the kurtosis-based wavelet thresholding algorithm can reduce both Gaussian noise and spurious impulses. Early detection of incipient fault is therefore possible due to the substantially purified signal. Different from the fault signatures in rotating machine elements, the oil-debris signature does not appear periodically but has a unique shape. Based on the generalized differentiation method, a signature extractor and an inflexion point locator are devised to extract in-line oil-debris signatures. The inherent low-pass filter property of two operators enables the algorithm to extract the oil-debris signature from the noisy signal without applying extra filters. To identify multiple faults from a single data set from a rotating machine, a single-channel signal separation method is proposed. Since the number of sources is not a priori information, the algorithm is suited for applications to mechanical fault diagnosis in which the number of fault sources is unknown. To develop a reliable and consistent fault severity measure, the Lempel-Ziv complexity has been exploited and applied to the output from the CWT of the signal. A normalized Lempel-Ziv complexity gives a non-dimensional value between zero and one [0, 1], close to zero for a pure sinusoidal signal and one for a white Gaussian noise. The main objective is to develop a non-dimensional fault severity measure that is sensitive to fault severity and can be used in different applications without major modifications. Another objective is that the fault severity measure should be developed such that it can be used alone without relying on another set of reference data. The proposed techniques for de-noising, fault detection and isolation, and fault severity have been evaluated in our lab or using data from the literature. The results have shown that the proposed methods can effectively purify signals, detect faults from vibration and oil-debris signals, separate multiple faults, and properly evaluate bearing fault severity