This paper describes the performance of parametric spectral estimation techniques when used to generate frequency spectra from vibration signals taken from low speed rotating equipment. The results show a consistently good performance for vibration signals which represent a variety of different operating conditions. Noise may be in the form of an unwanted sinusoid related to the main power source, vibrations from nearby machinery or vibrations from various components within or closely associated with the machinery being monitored.
Original Description:
Original Title
INSIGHT MachineConditionMonitoringPartII The Effects of Noise in the Vibration Signal
This paper describes the performance of parametric spectral estimation techniques when used to generate frequency spectra from vibration signals taken from low speed rotating equipment. The results show a consistently good performance for vibration signals which represent a variety of different operating conditions. Noise may be in the form of an unwanted sinusoid related to the main power source, vibrations from nearby machinery or vibrations from various components within or closely associated with the machinery being monitored.
This paper describes the performance of parametric spectral estimation techniques when used to generate frequency spectra from vibration signals taken from low speed rotating equipment. The results show a consistently good performance for vibration signals which represent a variety of different operating conditions. Noise may be in the form of an unwanted sinusoid related to the main power source, vibrations from nearby machinery or vibrations from various components within or closely associated with the machinery being monitored.
CONDITION MONITORING.
MACHINE CONDITION MONITORING:
PART 2 - THE EFFECTS OF NOISE IN THE VIBRATION
SIGNAL
CK Mechefske
CONDITION monrTORING
The use of parametric model-based spectral estimation techniques to allow analysis of vibration signals inthe frequency domain has proven
successful in several applications. In each ofthese situations, however, the vibration signal from the machinery being monitored has been kept
{free from noise external tothe vibration generating process. This was done in order to test the procedure under optimum conditions, While the
Parametric spectral estimation procedures are theoretically capable ofrejectingalarge noise component and have been tested using sine waves in
noise, there have been no tests to date which measure the technique’ performance characteristics when used with vibration signals from rotating
‘machinery contaminated with various amounts of noise. This paper describes the performance of parametric spectral estimation techniques when
used to generate frequency specrafrom vibration signals taken from low speed rotating equipment. Various spectral esimation techniques are
used to extract the frequency information from vibration signals contaminated with a range of noie levels. The results show a consistently good
‘performance for vibration signals which represent a variety of diferent operating conditions. Comparisons with standard FFT -based frequency
spectra are also included.
introduction
‘Tn many applications of machine condition monitoring the amount
‘of noise inthe vibration signal used to generate the frequency spectra
is in some way indicative of the condition of the equipment being
‘monitored. In other cases, the noise component obscures the
characteristic defect information being sought and must be removed
in order to gain access to the defect related signal component.
Understanding how the frequency spectral estimation routine being
used, whether fast Fourier transform-based or parametric model
based, reacts to the noise present in the vibration signal is essential
to being able to extract meaningful information from the frequency
spectra
‘All vibration signals will contain some noise, The noise maybe in
the form of an unwanted sinusoid related to the main power source,
vibrations from nearby machinery, vibrations from various
‘components within or closely associated with the machinery being
‘monitored, of, as is frequently the ease, a random component from
4 known or unknown source. The frequency spectra estimation
procedure must allow the unwanted components of the vibration
signal to be separated from the components which contain the
‘machinery condition information
Before this can take place, of course, the determination of what
the useful part of the signal and what isthe noise component must
bbe carried out. For example, the random component must not be
automaticaly assumed to be useless noise. In many cases by
‘monitoring the overall level of the random component an indication
‘of machinery deterioration may be gained. For a new monitoring
Dr Chris Mechefske received degrees
from the University of Waterloo,
Canada (Honours Bachelor of Earth
Sciences) and the University of Guelph,
Canada (Honours Bachelor of Science
= Engineering). In 1993 he received a
PAD in Mechanical Engineering from
Monash University, Melbourne,
‘Australia. Other and future research
inmerests include: pattern recognition,
artificial intelligence, mechatronics,and
‘process control and automation.
‘Machine Condition Monitoring: Part I - Optimum Vibration Signal
Lengths appeared in the British Journal of NDT, Vol. 35 No, 9,
September, 1998,
874
situation this process of determining the useful pat ofthe signal may
‘be quite involved and require vibration signal samples tobe taken at
several points on and around the machinery. In some cases the
‘unwanted portion of the vibration signal may not be removed but
instead must be noted, and then’ taken into consideration when
analysing the frequency spectra.
‘This paper details a study conducted to further the understanding
of parametric spectral estimation when used to generale frequency
spectra for use in the field of machine condition monitoring. The
effects of various amounts of noise in the vibration signal are
investigated. Plots showing frequency spectra generated from
vibration signals containing varying amounts of noise are compared,
Vibration signals representing sinusoids in noise and low speed
(S100 RPM) rolling element bearing vibration signals in noise are
used to calculate both FFT-based and parametric model-based
frequency spect,
2._Parametric Spectral Estimation
‘Parametric special estimation involves three steps. These steps
have been fly outlined in previous publications") so only a brief
description will be given here, Figst an appropriate model type is
selected based largely on the type of vibration signal being
‘considered. Vibration signals from various sources are more
ficiently and accurately modelled using different model types.
Next, the model is computed using one of a number of available -
algorithms. Again the choice depends on the type of vibration signal
being considered and what the model is being used for. Finally the
frequency spectrum is calculated. The last step is completely
‘dependent on the modelling procedure used in the second step.
21 Selection of the Appropriate Model Type
‘One of the advantages of the model-based approach to spectral
estimation is the higher resolution achievable with model-besed
techniques over that achievable with traditional FFT-based
techigues especially for short data sets\24) ‘The resolution
improvement is dependent on the ability to fit « model witha small
umber of parameters. A poor choice of model type will result in a
Jarge number of parameters being necessary 10 accurately model the
signal generating process. Therefore the correct choice of model
type is of utmost importance.
Before the decision regarding the best type of model to se can be
‘made, there should be consideration of all other information at hand.
In this way, selection ofan exact model forthe process thet generated
British Joumal of NOT‘the data samples, or at leat @ model that i « good approximation of
‘the actual underying proces, may be made.) In the ease ofroling