Acoustic emissions diagnosis of rotor-stator rubs using the KS statistic

https://doi.org/10.1016/S0888-3270(03)00050-5Get rights and content

Abstract

Acoustic emission (AE) measurement at the bearings of rotating machinery has become a useful tool for diagnosing incipient fault conditions. In particular, AE can be used to detect unwanted intermittent or partial rubbing between a rotating central shaft and surrounding stationary components. This is a particular problem encountered in turbines used for power generation. For successful fault diagnosis, it is important to adopt AE signal analysis techniques capable of distinguishing between various types of rub mechanisms. It is also useful to develop techniques for inferring information such as the severity of rubbing or the type of seal material making contact on the shaft.

It is proposed that modelling the cumulative distribution function of rub-induced AE signals with respect to appropriate theoretical distributions, and quantifying the goodness of fit with the Kolmogorov–Smirnov (KS) statistic, offers a suitable signal feature for diagnosis. This paper demonstrates the successful use of the KS feature for discriminating different classes of shaft-seal rubbing.

Introduction

Measurement of high frequency acoustic emissions (AE) has become a viable technique in the condition monitoring of many types of rotating machinery [1], [2], [3], [4], [5]. In real operational machinery it is often only practical to take AE measurements from non-rotating members, at or on the bearing housing. Consequently, AE signals originating from the rotating shaft will incur significant attenuation across the transmission path to an AE receiver attached at the bearing housing. This can be related to inhomogenities and scatterers within the structure, reflections at acoustic boundaries, interference and attenuation effects across the bearing interfaces. Moreover, the AE signal will be further coloured by the characteristic frequency response of the AE transducer itself. In light of these factors, interpretation of the AE signals is not trivial and often departs from the classic AE signal model [6], [7].

In recent years various signal processing and pattern recognition techniques have been successfully applied to AE signals for diagnosing the severity and location of defects in various types of rotating machinery. Notably artificial neural networks (ANN) [8] and clustering [4], [9] have been adopted for AE signal classification. Regardless of the classification engine employed, it is invariably identification of the key resolving features or descriptors within the AE signal that is paramount for successful classification. Generally, the size of the feature vector chosen depends upon the specific application and recognition requirements. Previous studies [10] employed relatively large feature vectors for the AE signal classification problem. However, for discriminating between different classes of rub signatures, it was considered useful to define a feature vector with a minimum number of parameters.

Typical features extracted from AE signatures in condition monitoring include peak or total energy, standard deviation, median, AE counts, rms voltage and duration. However, these are all related to absolute energy levels of the measured waveform or rely upon pre-set amplitude thresholds. As such, the quantities exhibit considerable variability from one bearing measurement to the next and are thus extremely dependent upon factors such as background noise, in addition to AE transducer positioning and coupling. Consequently, it is believed that such features are not ideal for AE waveform classification, especially in cases where several measurement positions are required. Alternative features more related to the amplitude statistics of the measured AE waveforms and independent of absolute energy levels have also been considered. Notably, the fourth statistical moment known as kurtosis and the ratio of peak to rms voltage known as crest factor have been applied for condition monitoring in rotating machinery. However, laboratory tests conducted as part of this research have indicated both of these quantities to be unsuitable for classification.

In contrast, it has been shown that modelling an AE signal as an autoregressive stochastic process, as described by Melton [11] and later Mba [4], can provide good AE classification results. However, the use of AR coefficients as signal features approximating the shape of the signal has some disadvantages. Primarily, it is always necessary to determine the number of AR coefficients necessary to adequately represent each AE signal. Although numerous algorithms exist for determining the model order, it should be noted that the classification results could be sensitive to the AR model order. Secondly, it was evident from this study that AE signal classification using AR model coefficients was severely impaired when the measured AE signals was modulated by small levels of background acoustic noise.

In light of this discussion, it is postulated that a robust AE signal feature based upon amplitude statistics and independent of absolute energy levels or pre-defined thresholds, and less effected by pre-signal processing, can be a useful addition to AE signal classification. This paper proposes that the standard Kolmogorov–Smirnov statistic can provide such an AE waveform feature parameter for classifying different types of rubbing in rotating machinery. To demonstrate this, classification results are presented from rub experiments conducted on a journal bearing test rig that rotates at 1500 rev/min (rpm). In addition, classification performance of this technique was tested by modulating the measured AE signals with background noise taken from bearings of an operational 550 MW turbine unit.

Section snippets

Acoustic emission and rubbing in rotating machinery

Fundamentally, a light frictional rub between the central shaft and surrounding stationary components, such as the seals within a turbine, will cause microscopic perturbation and a transient release of broadband strain energy referred to as Acoustic Emissions. Although originating from a different mechanical process, this wave motion is in practice very similar to the wave energy that propagates from microscopic cracks within solid structures. A number of reasons can be identified for the onset

Theory

Fig. 1 shows an example AE signal measured at the bearing housing of a test rig whilst rotating at 1500 rpm. This signature was a result of simulating a partial rub on the shaft with a steel seal fixture. Clearly, the AE burst shows some resemblance to the shape of a classic AE waveform produced by crack propagation or a Hsu–Nielson source, although it is not possible to identify individual extensional and flexural wave modes as can be often obtained by the Modal AE approach in thin plates. A

Experimental set-up

A Physical Acoustic Corporation wideband (WD) piezoelectric sensor was employed for this work with a range of 100–1000 kHz. Pre-amplification ranged from 40 to 60 dB. The signal output from the pre-amplifier was connected (i.e. via BNC/coaxial cable) directly to a commercial data acquisition card that occupies one of the ISA slots within a Pentium host PC. This AEDSP acquisition card provided up to 8 MHz sampling rate and incorporated 16-bit precision giving a dynamic range of more than 85 dB.

Results and discussion

Using the aforementioned system, AE signals were recorded for different types of partial shaft-seal rubbing. Primarily signals from the partial rubbing of steel and brass seal fixtures were taken. Secondly, partial rub signals were taken from three steel seal fixtures exhibiting different states of wear. Finally, rub signatures were taken from three seal fixtures with faces of different geometric shape. This section presents results pertaining to the use of the KS statistic to classify the

Conclusions

This paper introduces the use of the Kolmogorov–Smirnov (KS) test statistic as a useful signal descriptor in AE analysis. The KS results presented via hierarchical dendrograms indicate the potential of this statistic in classification of partial rubbing on shafts of fast rotating machinery. Moreover, the success of KS classification has been shown when the measured AE rub signals were modulated by background noise from a real operational 500-MW turbine.

The KS technique offers a different and

References (18)

  • A. Choudhury et al.

    Application of acoustic emission techniques for the detection of defects in rolling element bearings

    Tribology International

    (2000)
  • P.D. McFadden, J.D. Smith, Acoustic Emission transducers for vibration monitoring of bearings at low speeds, Report...
  • I. Sato

    Rotating machinery diagnosis with acoustic emission techniques

    Electrical Engineering, Japan

    (1990)
  • T. Holroyd et al.

    The use of acoustic emission for machine condition monitoring

    British Journal of Non-Destructive Testing

    (1992)
  • D. Mba et al.

    Condition monitoring of low-speed rotating machinery using stress wavesPart 1 and Part 2

    Proceedings of the Institute of Mechanical Engineers

    (1999)
  • C.B. Board, Stress wave analysis of turbine engine faults, Aerospace Conference, IEEE Proceedings (Cat. No.00TH8484)...
  • W. Venkatesan, Detection and modeling of acoustic emissions for fault diagnostics, IEEE, 0-8186-7576-4/96,...
  • C. Li

    On-line bearing localized defect detection by pattern recognition analysis

    ASME Transactions , Journal of Engineering for Industry

    (1989)
  • K. Ono, Q. Huang, Pattern recognition analysis of acoustic emission signals, Progress in Acoustic Emission, VII,...
There are more references available in the full text version of this article.

Cited by (0)

View full text